chore: version packages

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

.changeset/fullscreen-tab-switch.md

@@ -1,5 +0,0 @@
----
-'aicodeman': patch
----
-
-Switching to a fullscreen Claude tab is about ten times faster. Claude keeps that conversation itself and tmux holds no scrollback for it, so instead of downloading and replaying a 1 MB tail of old screen redraws on every switch, the browser loads just the current screen (a few KB), the same as a page load already did. Tabs whose pane does keep scrollback are unchanged.

.changeset/history-notice-lazy.md

@@ -1,5 +0,0 @@
----
-'aicodeman': patch
----
-
-The "Showing the most recent 1.0 MB of this session" notice no longer appears on every tab switch. It shows up only when you scroll to the top of a terminal, leaves when you scroll back down, and stays closed for that tab once you dismiss it. Sessions with nothing more to load (fullscreen Claude, whose history lives in Claude itself) never show it, and when there is more, it states the scrollback line count instead of an inflated byte figure. `GET /api/v1/sessions/:id/terminal` reports the new `paneHistoryLines` field.

.changeset/notification-history-duration.md

@@ -1,5 +0,0 @@
----
-'aicodeman': minor
----
-
-Notifications stay as long as you want. Settings → Notifications has a "Toast display time" and a "Browser notification display time" (seconds, per device; the defaults stay 3s and 8s).

.claude-plugin/marketplace.json

@@ -10,7 +10,7 @@
       "name": "codeman",
       "source": "./plugins/codeman",
       "description": "Drive Codeman from inside a Claude Code session: spawn worker sessions, prompt them, wait for them, read their answers, clean up. Acts only inside a Codeman-managed session.",
-      "version": "1.40.0",
+      "version": "1.42.0",
       "author": {
         "name": "Ark0N",
         "url": "https://github.com/Ark0N"

.github/CONTRIBUTING.md

@@ -57,7 +57,7 @@ Expect `test:browser`/`test:mobile`/`test:perf` to fail where the machine cannot
 
 The browser suite also runs nightly (and on demand) in `.github/workflows/browser-suite.yml`; it is informational, not a gate.
 
-If you add a test that binds a port, bind port 0 (`new WebServer(0, …)` + `server.boundPort`, or `listen({ port: 0 })` + `address().port`), or use `app.inject()` when no socket is needed; `test/test-ports-guard.test.ts` fails a `WebServer` built on any other port. Mobile tests (`test/mobile/**`, via `createTestServer(PORT)`) keep the fixed-port convention in `test/mobile/README.md` for now, because that helper caches servers by port. Never 3000.
+If you add a test that binds a port, bind port 0 (`new WebServer(0, …)` + `server.boundPort`, or `listen({ port: 0 })` + `address().port`), or use `app.inject()` when no socket is needed; mobile tests call `createTestServer()` and read `server.boundPort`. `test/test-ports-guard.test.ts` fails a `WebServer` built on any other port and a raw `listen` on a fixed one. Never 3000.
 
 Tests are tmux-safe by design: under vitest, the tmux layer becomes an in-memory mock, so tests cannot touch real sessions.
 

.gitignore

@@ -84,6 +84,23 @@ tmp/
 # added there). No trailing slash so it still matches the symlink, not just dirs.
 /public
 
+# Video never goes to GitHub (docs/media-assets.md). The only tracked moving
+# images are the GIFs a public page embeds, in docs/images/; masters, promo cuts,
+# drafts and captures live in the local media library pr/media/ (ignored above).
+# The promo branch tracks its Remotion render inputs, so those stay addable.
+# A deliberate exception elsewhere takes `git add -f`.
+*.mp4
+*.webm
+*.mov
+*.m4v
+*.mkv
+*.avi
+*.gif
+!/docs/images/*.gif
+!/scripts/remotion/public/**/*.mp4
+!/scripts/remotion/public/**/*.webm
+!/scripts/remotion/public/**/*.gif
+
 # Opt-in gesture overlay runtime assets: large MediaPipe wasm + model (~27 MB)
 # fetched at build/install by scripts/fetch-gesture-assets.mjs, kept out of git.
 # (The gesture bundle itself, gesture-codeman.js, IS tracked — built from
@@ -107,7 +124,9 @@ todo.md
 @fix_plan.md
 readme-preview.mjs
 
-# Uploaded images land here under each session working dir (runtime artifact)
+# Prompt uploads land here under each session working dir (runtime artifact);
+# .claude-images/ is where they landed before the move.
+.codeman-uploads/
 .claude-images/
 
 # Local-LLM harness smoke-test config (real IPs/keys) — see the .example.json

.prettierignore

@@ -28,3 +28,5 @@ scripts/remotion/
 
 # Hand-maintained; Prettier escapes underscores in glob paths and corrupts paragraphs.
 CLAUDE.md
+src/web/CLAUDE.md
+docs/releasing.md

AGENTS.md

@@ -12,6 +12,6 @@ Quick pointers:
 
 - Type check: `tsc --noEmit` · Lint: `npm run lint` · Format: `npm run format:check`
 - Tests: `npm test` (the CI gate, safe to run bare) or `npm test -- test/<file>.test.ts` for one file
-- Route tests use `app.inject()`; new tests needing a socket bind port 0 (`new WebServer(0, …)` + `boundPort`), except mobile tests, which keep `createTestServer(PORT)` for now
+- Route tests use `app.inject()`; new tests needing a socket bind port 0 (`new WebServer(0, …)` + `boundPort`); mobile tests use `createTestServer()` and read `server.boundPort`
 - Branch off `master` for all work; Conventional Commit-style messages (`fix(mobile): ...`)
 - Never commit secrets or local state from `~/.codeman/`

CHANGELOG.md

@@ -1,5 +1,137 @@
 # aicodeman
 
+## 1.42.0
+
+### Minor Changes
+
+- 1238316: ### Thanks
+  - @JDProfresh for five PRs: prompt uploads that take documents (#621, from the plan in #553), a video Codeman cannot take refused before it uploads (#622), a prompt-cache readout kept to the CLIs that report one and a statusline report with many cache-miss causes no longer thrown away (#623), and two fixes for Claude's workspace trust dialog in a brand-new folder: idle waits that hold until the dialog is answered (#634) and a dialog that gets answered every time (#635).
+  - @Randalix for Codex prompts that really get submitted, sent as one bracketed paste and a separate Enter (#628), and for the careful measurements in #627 that found the cause.
+  - @ludovicfourrage, a first-time contributor, for reporting (#629) and fixing (#632) the nested tmux status bar colour that leaked into live output after a history reload.
+  - @opticon454 for three PRs: a one-time "what's new" note after each update (#618), Previous and Next Session as rebindable shortcuts that refuse a chord another action already uses (#617), and a DeepSeek install from CLI management that ends with a DeepSeek you can run (#616).
+
+  **Git worktrees for a case (#625).** Give every agent on a repository its own checkout. Pick a case whose folder is a git repository, choose **Git worktrees…** in the case picker (**Worktrees** on the phone's case sheet), type a name and pick the CLI: Codeman creates `<case>/.claude/worktrees/<name>` on a new branch `worktree-<name>` and opens a session there, in a tab named after the worktree. **Open** starts another session in a worktree you already have. Removing one runs plain `git worktree remove`, never `--force`, so uncommitted work stays where it is, and Close Session can kill the session and remove its worktree in one go. The layout is the one Claude Code's own `claude --worktree` uses, so each sees the other's worktrees. To keep the main checkout clean, Codeman lists the worktrees folder and the files it writes for sessions in the repository's local `.git/info/exclude`; your `.gitignore` is never touched.
+
+  **Codex sessions turn red when Codex asks you something (#624).** Codex's command and edit approvals, its question dialogs and the trust prompt on a new repository are read straight off the screen and raised the way a Claude permission prompt is: red tab, desktop notification, push, webhook, the phone's NEEDS YOU row and an Approvals Inbox card that tells you to answer in the terminal. An agent driving a codex worker can now wait for that with `until=blocked`.
+
+  **Prompt uploads take documents (#621).** Attach, drop or paste a PDF, a Word, PowerPoint or Excel file, a zip, or a text file (txt, md, csv, tsv, json, log, yaml, xml) and its path lands on your prompt, ready to send, like a photo or a video does. Every file is checked against its actual bytes on the way in, so a program renamed to `.pdf` is refused, and text files must be UTF-8. The saved name keeps a readable part of the original: `Q3 report.pdf` is saved as `paste-...-Q3-report.pdf`. The button in the Compose and paste dialogs now says Attach.
+
+  **Previous and Next Session are rebindable (#617).** Previous Session (Alt/Option+[) and the alternate Next Session (Alt/Option+]) are now ordinary rows in App Settings > Shortcuts, so you can rebind or switch them off like everything else. Capturing a chord another action already uses is refused with a message naming that action (Alt+1 to 9, Ctrl/Shift+Enter and Ctrl/Cmd+V stay reserved), and re-enabling or resetting a shortcut runs the same check. A chord you rebound no longer also types into the terminal: a rebind to Alt+J used to switch tabs and send ESC j to the session too. Built-in defaults like Ctrl+L still reach the CLI.
+
+  **A note after each update (#618).** The first time a browser loads a newer Codeman, a toast lists that release's headline changes, with a **See what's new** button that opens the full release notes. It shows once per version, and a browser that has never opened Codeman before just records the version and stays quiet, so the first note you will see is the one for the release after this one. The headlines are taken from the changelog when the release is cut.
+
+  **Multi-user mode: only admins change server-wide settings.** In multi-user mode any signed-in regular user could rewrite the server's settings (feature switches, run mode, case order, the default model for cron and the orchestrator, the app config), start a self-update, rotate the tunnel QR code, clear stale session state, and sign every user out with Revoke All, admins included. They could also read the Deepgram API key an admin saved. Those writes and operations now answer 403 to a regular user, Revoke All signs them out of their own logins only, and their settings read leaves out the admin's key and case names. In the browser a regular user's per-device settings still save, a change to a server-wide setting shows one notice that only an admin can change it, instead of pretending it was saved, and their own run mode and respawn presets are kept on their device. Single-user mode is unchanged.
+
+  **Fixes.**
+  - **Codex submits long prompts (#628).** Prompts sent through the API, `codeman agent send`, the MCP server, cron or the agent skill go into the pane as one bracketed paste followed by a separate Enter, so Codex no longer leaves them sitting unsent in its composer. A single character (a menu answer like `3` or `y`) is still typed, so numbered dialogs keep picking the option you sent, and a trailing `;` survives, so `find . -exec rm {} \;` arrives whole.
+  - **New Claude sessions in an untrusted folder (#634, #635).** An agent waiting for a fresh session to go idle no longer gets its answer while the workspace trust dialog is still up, so its first prompt can no longer press Enter on the highlighted "No, exit" and kill the session it just started. And the dialog is now answered every time: a screen read that came too soon after the previous one used to be dropped, which now and then left a new session sitting on the dialog. Deleting a session while the dialog is up also stops Codeman reading its dead pane for seconds afterwards.
+  - **Nested tmux colours after a history reload (#632).** Opening a remote or Docker case, or a shell running tmux inside it, no longer hands the nested tmux status bar's green background to whatever the agent prints next.
+  - **The codeman skill with arguments (#642).** The skill works again when Claude Code loads it with arguments ("use codeman to start 5 claude workers"). Claude Code swaps a skill's `$1` and `$2` for words of your request, which broke the pasted helpers; they now use `${1}` and `${2}`, and a cached copy written by the broken version is rewritten on the next run.
+  - **DeepSeek installs ready to run (#616).** Installing the terminal profile finds the `pnpm` next to `dsh` even when the server's own PATH lacks `~/.local/bin`, and the DeepSeek install command adds `pnpm` when the machine has none.
+  - **Unsupported videos refused up front (#622).** An Android `.3gp` or an `.mkv` gets "Unsupported file type" right away, instead of uploading up to a gigabyte and failing at the end.
+  - **Statusline reports (#623).** The prompt-cache readout is stored only for CLIs whose statusline reports one (Claude today), like the model badge, and a Claude report with more than 20 cache-miss causes keeps the first 20 instead of being refused, which used to blank the footer and the plan-usage chip.
+
+  **Fixes applied while landing.** Removing a worktree that is, sits inside or holds a Docker case's workspace is refused before git runs; it would have deleted the folder a container still mounts (#625). Codex's trust prompt also reads as blocked in the variant reached through its "Update available" screen, which ends in "esc back" (#624). The DeepSeek install only adds `pnpm` when none is on PATH, so a machine with a corepack `pnpm` keeps it instead of the whole install failing (#616). Three statements in the upload docs were corrected: only a local agent and a Docker case share the uploads folder (remote sessions are refused), GIFs keep their frames instead of being re-encoded, and a developer note no longer names a release that never shipped (#621). The update note skips the Thanks list and group labels like "Fixes.", still works for a release with no headlines, and is translated for zh-CN (#618).
+
+  **Docs.** The README and its Chinese translation caught up with 1.41.0 and 1.41.1 and show a sharper hero GIF, the Codeman Skill got a wiki page and two GIFs linked from both READMEs, CLAUDE.md was split into the root file, `src/web/CLAUDE.md` and `docs/releasing.md`, the release steps now include the MCP registry publish, and a new note says where video and GIF masters live (not on GitHub).
+
+## 1.41.1
+
+### Patch Changes
+
+- ### Thanks
+  - @timkjr for four PRs: shells that reload from tmux's rendered history instead of replaying old full-screen repaints (#601), the full scrollback of a remote SSH session read from the remote tmux (#602) with no rows lost at the join (#609), and a phone shell Paste that inserts a multi-line block instead of running it line by line (#600).
+  - @aakhter for three PRs: Codex panes that page their transcript with the wheel and touch (#527, fixes #508), new and spawned tabs that land in tab groups (#528), and a Focus section of shortcuts at the top of the vertical rail (#582).
+  - @JDProfresh for three PRs: Claude Code's prompt-cache state per session (#607, step 1 of #554), the phone Attach button with camera-roll video uploads (#606, part two of #553), and an upload that no longer stalls the server on a dead network mount (#605).
+  - @Randalix for clickable `~/`, relative and hard-wrapped file paths plus a folder grid for browsing a whole folder of renders (#604, closes #603).
+  - @opticon454 for three PRs: a Working folder row with a Copy button in Session Options (#593), tab colour, pin and image watcher surviving a restart (#591), and Link Existing saying why it refused a case (#590).
+  - @ideaCompany, a first-time contributor, for rendering local HTML files as real pages in the File Viewer (#547), and for the careful security write-up behind it.
+  - @zomby1, a first-time contributor, for OSC 52 clipboard support (#611).
+  - @Hugo-DDT, a first-time contributor, for two PRs: MCP server sync reading Gemini's bare `url`, OpenCode's V2 layout and `opencode.jsonc` (#599, fixes #586), and DeepSeek permission prompts that no longer offer Approve/Deny buttons that cannot work (#610, from #598).
+
+  **HTML files open as pages, and markdown opens rendered from the terminal (#547).** Click an `.html` path an agent printed (a test report, a coverage page, a dashboard) and the File Viewer now shows it as a real page, with its own scripts, styles, images and data files from its folder, instead of raw source. A **Page** pill switches to the source and its Edit button. The page runs in a sandbox with an opaque origin: it cannot read Codeman's cookies, use Codeman's API or read files outside its own folder. Network access is open, so a dashboard that loads Chart.js or fonts from a CDN works as it would opened from disk. Broad folders (`/`, home, `/tmp`, a drive's mount point) are refused, so put a report in a folder of its own. And a `.md` or `.markdown` path clicked in the terminal now opens the rendered markdown viewer everywhere (main terminal, tiles, split pane, touch), instead of the live log viewer that showed "streaming"; the MD pill still gives the source. Logs, `.txt`, `.json` and code keep the live log viewer.
+
+  **`codeman mcp`: drive Codeman from any MCP client (#615).** A stdio Model Context Protocol server with seven tools (list, start, send, wait, read, interrupt, delete), so Claude Code, Codex, Gemini CLI, Cursor or Claude Desktop can start worker sessions in any installed agent CLI, prompt them, wait for them and read what they did. It is a client of the existing HTTP API with the same guards as `codeman agent`: it refuses control characters in input, never acts on the session it runs in, and by default only touches sessions it started. Setup lines for Claude Code, JSON clients and Codex are in the README and the wiki page "Driving Codeman From An Agent".
+
+  **The codeman skill for every agent CLI (#613).** With **Codeman Agent Skill** on (App Settings > Agents & CLIs > All agent CLIs, still off by default), Codex, Gemini CLI, OpenCode, Antigravity, Grok Build, DeepSeek Harness and OMP sessions now get the same `codeman` skill Claude Code had, in `.agents/skills/codeman`, so an agent in any mode knows how to start and manage worker sessions. `codeman skill install --mode <cli>` installs it by hand.
+
+  **Tiles: move them by hand, or tear one off into its own window (#584).** With the camera gesture overlay on, pinch a tile and carry it onto another tile to swap them, onto an empty cell to move it, or carry a tab onto the grid to add it. With the new **Detach Tiles** setting (App Settings > Appearance, off by default), drag a tile's header out of the window and it pops out into its own window; drop it on another Codeman window's tiles and it moves there. Known limits: all Codeman windows in one browser share one saved tile layout, a drag between two browser profiles is not seen by the source window, and a drag-out held for more than a few seconds can be blocked as a pop-up.
+
+  **Focus section on the vertical rail (#582).** "Add to Focus" in a row's menu (or right-click a web tab) pins a shortcut to that tab in a small collapsible section at the top of the rail. The tab itself stays where it is.
+
+  **Tabs land in groups (#528).** A tab group's menu has **New session**, which starts a session straight into that group. A session an agent spawns lands right after its parent; App Settings > Appearance > Tabs > **Spawned tabs follow their parent** (off by default) keeps it moving with the parent.
+
+  **Folder grid and more clickable paths (#604).** Paths written as `~/...`, relative to the session folder, or broken across terminal lines are now clickable in the terminal and the Response Viewer, and folders are too: a folder opens a grid of its files (renders, screenshots, reports) that you can page through with the arrow keys.
+
+  **Phone Attach button and videos (#606).** On a phone the toolbar has an Attach button that opens the system file sheet directly, without raising the keyboard, and camera-roll videos (mp4, mov, webm, up to 1 GB) upload into `.codeman-uploads/` like images.
+
+  **OSC 52 clipboard (#611).** A program in a session that sets the clipboard with OSC 52 (Claude Code's `/copy`, a tmux yank, vim) now puts the text on your browser's clipboard, in the main terminal and in every tile. Read requests are refused, and a replayed buffer never copies.
+
+  **Prompt-cache readout (#607).** Claude sessions show whether their prompt cache is warm (`cache:until HH:MM`) or cold in the footer and in Session Options > Respawn, restored after a restart, with a reason when a launch will never report it. Display only, nothing changes behaviour.
+
+  **Faster tab switches to fullscreen Claude (#588, #583).** Switching back to a fullscreen Claude tab now loads only its current screen (a few KB) instead of replaying a 1 MB tail of old redraws (about 1.3 to 1.7 s saved per switch). The "Showing the most recent 1.0 MB" bar now appears only when there really is older history and you scroll to the top looking for it, gives a line count, and stays closed once you close it.
+
+  **Fixes.**
+  - **Shells reload cleanly (#601).** A shell tab, tile or split pane reloads from tmux's rendered history, so a prompt no longer appears four or five times and lines no longer go missing after a resize or a full-screen program.
+  - **Remote SSH sessions keep their scrollback (#602, #609).** Loading a remote session's full history reads it from the remote tmux in one bounded ssh round trip, where the local pane kept only a fraction (48 of 279 lines in one measurement, none after a reattach).
+  - **Codex pages with the wheel and touch (#527).** A Codex pane on its alternate screen pages its transcript with PageUp/PageDown from the first wheel notch or swipe, in the main terminal and in tiles, and never pages into an open permission dialog.
+  - **openpyxl spreadsheets preview (#612).** An `.xlsx` written by openpyxl with a cell comment, a table or a chart now opens in the File Viewer instead of failing.
+  - **Phone shell Paste (#600).** A multi-line paste in a shell is inserted as one bracketed paste and waits for Enter, instead of running each line as it arrives.
+  - **Restart keeps tab colour, pin and image watcher (#591).**
+  - **Session Options shows the working folder (#593)**, with a Copy button and a label that says when the folder is on a remote host or mounted into a Docker container.
+  - **Link Existing says why (#590)**, for example "Invalid case name format", instead of "Invalid request body".
+  - **Uploads on a dead mount (#605).** A prompt upload into a workspace on a network mount that stopped answering is refused at once instead of tying up a server thread.
+  - **MCP server sync (#599).** Gemini servers with a bare `url` sync as HTTP, OpenCode's V2 `mcp.servers` layout and `opencode.jsonc` are read.
+  - **DeepSeek approvals (#610).** DeepSeek Harness permission prompts show "Answer in the terminal" in the Approvals Inbox, the phone NEEDS YOU strip and push notifications, instead of Approve/Deny buttons the server always refused.
+
+  **Fixes applied while landing.** Remote SSH and Docker tabs keep their scrollback on a tab switch (they were about to lose it with #588). No colour bleed where remote history meets the live screen (#602, #609). A bare `~` in a Docker case no longer lists the host's home folder, and shell prompts (`user@host:~/proj$`) no longer turn into broken folder links (#604). Spawned sibling tabs keep their creation order (#528). An OSC 52 copy over plain HTTP gives keyboard focus back to the terminal (#611). A refused video upload answers instead of hanging, and the zh-CN "Attach" label of the tile grid is back (#606). The Working folder row fits on phones, with zh-CN text (#593). OpenCode sync no longer stalls on a server named `servers` (#599). Moving a tile to another window no longer overwrites that window's saved layout (#584). HTML pages: a `~/` link to an `.html` file renders as a page instead of falling back to source, `/tmp` is refused even when `TMPDIR` points elsewhere, symlinks cannot get around the hidden-file and file-type rules, workspace confinement is re-checked on every request, media in a page can seek, and the service worker never caches a page (#547). Codex users of `codeman mcp`: the config now forwards the session id, so its self-protection works (#615).
+
+## 1.41.0
+
+### Minor Changes
+
+- 87f1c9c: ### Thanks
+  - @Randalix for `codeman agent` (#557), the session verbs (`ls`, `spawn`, `send`, `wait`, `read`, `interrupt`, `rm`) for agents in every CLI mode, and for moving every test server onto an ephemeral port (#570), which finishes #440. Thanks also for reporting and fixing git's clone errors on non-English hosts (#568, shipped as #572 with your commit).
+  - @opticon454 for five PRs: MCP server sync for GitHub Copilot CLI (#581), an Apply button that saves Settings without closing them (#565), configurable toast and browser-notification display times (#564), in-document links in rendered markdown that scroll to their heading (#563), and npm-based CLI installs that work when the system npm prefix is root-owned (#562).
+  - @JDProfresh for three PRs: pasted and uploaded files moving into a hidden, self-ignoring `.codeman-uploads/` folder (#574, after your #553 proposal), local echo that keeps painting on phones when the view sits above the bottom (#576), and upload failures that say why they failed (#578).
+  - @aakhter for session search on the vertical tab rail (#580), built on the sidebar's existing filter instead of a second one, and for keeping alerted tabs visible during a search.
+
+  ![Codeman tile grid: six live agents powering on and off with the CRT tile animation](https://raw.githubusercontent.com/Ark0N/Codeman/08694b5862534b2b1224e9e272ad9858c787ebef/release-1.41/tiles-crt-stats-800.gif)
+
+  **Tile Animations (#571).** The tile grid from 1.40.0 can now open with a show. App Settings has a new **Animations** section (right after Appearance) that holds every animation setting: the Entrance Theme (moved out of Appearance), the new **Tile Animations** row, and a button that opens the animation lab. Tile styles: **CRT** (each tile switches on as a hot line in a diagonal wave, and switches off to a line and a dot), **Fly from tab** (each tile grows out of its session tab and flies back into it), **Deal** (dealt out of the Tiles button like cards), **Beam down**, **Cascade**, **Pop**, **Soft** and **None**. A styled tile plays in two beats: the frame enters in the tile style, then its screen powers on in the terminal style of your Entrance Theme. Picking a theme presets a matching tile style, and a new **Launch** theme flies the tiles out of their tabs. Off by default (the grid keeps its quick fade), per device, nothing moves under reduced motion, and the animations never cost an extra PTY resize. The terminal pane's Boot entrance style is gone; a saved Boot falls back to off.
+
+  **Tiles open the sessions you care about, and keep your layout.** With no grid arranged yet, the Tiles button (or `Ctrl+Shift+G`) now fills the grid with the sessions that are working first (most recently started first), then the ones waiting on you (a permission prompt or a finished turn you have not seen), then the most recently used, instead of the tabs in strip order, which opened the oldest ones. The session you are on is still always included. Once you arrange a grid (which session sits where, empty cells, the tile count, divider sizes, the focused and zoomed tile), that layout is remembered in this browser and the Tiles button brings it back exactly, however the grid was closed, and so does a reload while it was open. A session that has gone since frees its place, which the same ranking fills. A grid larger than the window opens in full with the focused tile zoomed, instead of being trimmed (and losing the rest of your layout).
+
+  **The wheel scrolls Claude inside a tile (#577).** In a tile or the split's second pane, a Claude session on its fullscreen renderer now scrolls its own conversation with the mouse wheel, exactly like the main terminal: the wheel goes to Claude as mouse reports, aimed at that tile's session and computed from the tile's own screen. Before, the wheel did nothing or scrolled the stale frames left over from loading the tile. Shift+wheel (scroll local history) works in every tile on Windows and Linux too; it was dead there.
+
+  **`codeman agent`: session verbs for every CLI (#557).** Agents in any mode (Codex, OpenCode, Gemini, Pi and the rest, not just Claude) can now drive other Codeman sessions from the command line: `codeman agent ls | spawn | send | wait | read | interrupt | rm`. It is a thin client over the existing session API: every call names the session that made it, `wait` blocks on a signal (`--until stop,exit`) or a literal output marker (`--match`), and exit codes say what happened (`0` ok, `1` error, `2` timeout, `3` exited, `4` refused). Ids shorter than 8 characters are refused, so a stray `rm 9` can never pick a session at random, and `rm` never deletes the session it runs in. See the README section "`codeman agent`" and the wiki page Driving Codeman From An Agent. This is phase 1 of #445.
+
+  **Settings: Apply (#565).** Next to Save, an Apply button saves the same way but keeps Settings open. Switching on MCP server sync makes its Preview and Sync usable straight away, and CLI management's add, enable and disable work without closing and reopening Settings.
+
+  **MCP server sync reaches GitHub Copilot CLI (#581).** With MCP server sync on (App Settings, off by default), Copilot CLI's `~/.copilot/mcp-config.json` now takes part like the agent CLIs' own files: its servers are copied to the others and theirs to it, additively, with the previous file kept as `.codeman-bak`. Copilot joins only when it is installed or already has that file, a server switched off in Copilot is never copied, and `COPILOT_HOME` is followed. Copilot is a sync target only, not a new run mode.
+
+  **Notifications stay up longer if you want (#564).** Settings → Notifications has a Toast display time and a Browser notification display time (1 second to 5 minutes, per device; the defaults stay 3 s and 8 s).
+
+  **CLI logos on tabs can be switched off (#569).** App Settings → Appearance → Tabs → **CLI Logos on Tabs** hides the agent logo on every tab surface (header strip, rails, sidebar, phone chips, the desktop home list) on this device. On by default. Tile headers, split pane headers and the Run menus keep their logos.
+
+  **Search sessions on the vertical rail (#580).** The vertical tab rail has a **Search sessions** box at the top: type part of a name and the rail narrows to the tabs that match (a web tab by its title), across every group, collapsed ones included, without touching your groups, their collapse or the tab order. A tab with an alert stays visible even when its name does not match, so a prompt waiting on you is never filtered away. Escape or × clears it, you can still drag a found tab into a group, and nothing is saved. The sidebar's filter box shares the same filter: a tab with an alert stays visible there too, and in the by-case tab layout a case with no match now hides.
+
+  **Closing a tab is instant.** Closing a session used to take half a second or more before the tab went away. The tab, tile or split pane now goes (and the next session is selected) the moment you click, while the server shuts the session down in the background, and the server side is faster too (about 450 ms down to 200-260 ms for a Claude session): it no longer sleeps fixed intervals, no longer freezes for about 70 ms per close on a synchronous tmux call, and scans for a session's subagents once instead of once per subagent. If the server refuses the close, the tab comes back where it was with the error. This also fixes a bug where a failed close still said "Session closed" while the session kept running.
+
+  **Fixes.**
+  - **Clone errors on non-English hosts (#572, from #568).** Cloning a repository as a case now classifies a failed clone correctly whatever the host's language: a missing branch or tag is "does not exist on the remote" (400) and a missing repository is a 404, instead of a generic 422 with git's German (or any other) error text. Git runs with `LC_ALL=C` for clones and repo status, so the repo status card's error text is English on every host as well.
+  - **Links within a markdown file (#563).** A link to another heading of the same document (`[Install](#installation)`) in the File Viewer or Response Viewer scrolls to that heading instead of doing nothing. Headings get GitHub-style slugs, repeated titles are numbered, and non-ASCII headings work.
+  - **npm CLI installs on a root-owned prefix (#562).** Installing an npm-based CLI from Settings (DeepSeek's `dsh`, pi, ...) no longer fails with EACCES when the system node keeps its global prefix under `/usr`: the install goes to `~/.local`, where Codeman already looks for CLIs. A prefix you set yourself, or one you can write to, is left alone, including when Codeman runs under `npm run`.
+  - **Uploads go to a hidden `.codeman-uploads/` folder (#574).** Images you paste or upload into a prompt are saved in `<workspace>/.codeman-uploads/` instead of `.claude-images/`. The folder ignores itself in git (it carries a `.gitignore` of `*`), stays hidden in the Files panel, and is cleaned up as before: files older than 7 days in an hourly sweep, and the folder when the last session in that workspace closes. The old `.claude-images/` folder gets nothing new and is still swept and removed during 1.41.x. An upload to a remote (SSH) session is now refused with a clear message, since the file would land on the Codeman host where the remote agent cannot read it.
+  - **Typing on a phone after a tab switch (#576, fixes #575).** With local echo on (the default on touch devices), text typed while the terminal sat above the bottom was buffered but never painted, so the keyboard looked dead. The view sits there after every tab switch and after the keyboard closes. The overlay now paints whenever the prompt row is on screen, and hides only when you scroll the prompt out of view.
+  - **Upload failures say why (#578).** When a prompt image upload fails, the toast shows the server's reason (for example a rate limit) instead of only "1 failed".
+  - **The npm page shows the English README.** npmjs.com had been rendering the Chinese README, because npm picks the package's readme from an unsorted file match at publish time. The publish now moves `README.zh-CN.md` aside while it runs (the file and every link to it stay as they are), and the package description and homepage say what Codeman is and point at getcodeman.com.
+  - **New cases ask for clickable file paths.** The CLAUDE.md generated into a new case asks the agent to report every file it created as a full absolute path, which Codeman turns into a link that opens the File Viewer, and mentions the codeman skill for starting and managing worker sessions.
+
+  **For contributors (#570).** Every in-process test server binds an ephemeral port, the mobile suite included, and the port guard now also refuses raw listeners on a fixed port, so two test runs on one machine never collide.
+
+  **Fixes applied while landing.** zh-CN translations for the two new notification display-time settings and for the new Animations section. A session whose name matches an interface word ("Lab", "New session") is no longer translated in the tab strip when the interface is in Chinese. Plus test and doc cleanups left over from review.
+
 ## 1.40.0
 
 ### Minor Changes

CLAUDE.md

@@ -50,35 +50,9 @@ The production server caches static files for 1 year, `immutable` (`maxAge: '1y'
 
 Uses [Semantic Versioning](https://semver.org/) (`MAJOR.MINOR.PATCH`) via `@changesets/cli`. What SemVer actually covers (the CLI, documented env vars, **and the HTTP/SSE API under `/api/v1`**: endpoint paths, response envelope, `errorCode` values and SSE event names are public/stable; on-disk state, internal TS modules, and experimental features are internal/unstable) is defined in `docs/versioning-policy.md`. Third-party integration surfaces are documented in `docs/extending-codeman.md`. Security reporting + known limitations live in `.github/SECURITY.md`.
 
-When user says "COM":
+When the user says **COM** (`COM` = patch, `COM minor`, `COM major`): **Read [`docs/releasing.md`](docs/releasing.md) and follow every step in order**; never improvise a release from memory. It covers the changeset, `npm run version-packages`, the `**Version**` line below, the branch check and explicit-path commit, deploy, the getcodeman.com badge, watching BOTH CI workflows, the MCP registry publish, and the Discussions announcement.
 
-1. **Determine bump type**: `COM` = patch (default), `COM minor` = minor, `COM major` = major
-2. **Create a changeset file** (no interactive prompts). Write a `.md` file in `.changeset/` with a random filename:
-
-   ```bash
-   cat > .changeset/$(openssl rand -hex 4).md << 'CHANGESET'
-   ---
-   "aicodeman": patch
-   ---
-
-   Detailed description of ALL changes since last release (not just the most recent commit — review full git log since last version tag)
-   CHANGESET
-   ```
-
-   Replace `patch` with `minor` or `major` as needed. Include `"xterm-zerolag-input": patch` on a separate line if that package changed too.
-
-3. **Consume the changeset**: `npm run version-packages` (auto-bumps `package.json` files, updates `CHANGELOG.md`, runs `npm install --package-lock-only`, and verifies lockfile sync via `scripts/check-lockfile-sync.mjs` — all in one command; never hand-edit `CHANGELOG.md` or `package-lock.json` versions)
-4. **Sync CLAUDE.md version**: Update the `**Version**` line below to match the new version from `package.json`
-5. **Commit and deploy**: verify the branch first (`git branch --show-current`), then stage EXPLICIT paths — never `git add -A`, which has swept another session's WIP into a release. `git status --short` and account for every line before committing:
-   `git add <paths> && git commit -m "chore: version packages" && git push && npm run build && systemctl --user restart codeman-web`
-6. **Refresh the getcodeman.com version badge**: the landing page's status bar carries the release version (`v<x.y.z> · getcodeman.com · MIT`), so it goes stale on every release if nobody bumps it. The site source and its deploy script are maintained outside this repository, on the maintainer's machine only; follow the local site handbook there, which also covers the numbers strip and `sitemap.xml` refresh that belong in the same pass. Poll production (`curl -s https://getcodeman.com/ | grep v<x.y.z>`) before calling it done, since the edge lags a deploy by up to a minute. Not applicable to contributor clones — skip it and say so.
-7. **Wait for CI**: after `git push`, TWO workflows fire per master push — `CI` and `Release` (the npm publish + GitHub release). List both runs for the pushed commit with `gh run list --commit $(git rev-parse HEAD) --json databaseId,workflowName` and watch EACH with `gh run watch <id> --exit-status`. Confirm both pass before considering the release done (`gh run list -L 1` returns only one of the two).
-
-8. **Announce it in Discussions**: every release gets a post in the [Announcements](https://github.com/Ark0N/Codeman/discussions/categories/announcements) category, shaped like #418 and #302: the short version first (one bold lead-in per theme, features before fixes, what it does for the user rather than how it works), contributor @-mentions inline where their work is described (a mention notifies them, and a contributor reposting is the cheapest reach this project has), a link to the release, and the Thanks names at the end. Casual first-person voice, no em-dashes, humanizer pass when the skill is available. A same-day follow-on patch (1.28.1 after 1.28.0) is folded into the previous post as an edit, never a second thread. Post it with `gh api graphql -F body=@<file> -f title='Codeman <x.y.z>: <hook>' -f repo=R_kgDOQ-SMDg -f cat=DIC_kwDOQ-SMDs4DCHZE -f query='mutation($repo:ID!,$cat:ID!,$title:String!,$body:String!){createDiscussion(input:{repositoryId:$repo,categoryId:$cat,title:$title,body:$body}){discussion{number url}}}'` (the repo's node id and its Announcements category id; `pinDiscussion` does not exist in the API, so pinning stays a click in the UI). This step exists because announcements stopped at 1.18 (#302) while ten releases shipped with nobody notified; #418 is the backfill covering 1.19.0 to 1.28.1, and release notes only count as content once they reach a surface people are subscribed to.
-
-CI runs `npm run check:lockfile` on every push/PR, so lockfile drift fails the build even if the `version-packages` script is bypassed.
-
-**Version**: 1.40.0 (must match `package.json`)
+**Version**: 1.42.0 (must match `package.json`)
 
 ## Project Overview
 
@@ -98,30 +72,13 @@ Codeman is a Claude Code session manager with web interface and autonomous Ralph
 
 | Task                           | Command                                                                                                                                                                                                        |
 | ------------------------------ | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| Terminal dashboard             | `codeman tui` (`--list` prints the numbered list and exits, `codeman tui <n>` attaches to row n; both short-circuit before any screen setup). Needs a TTY; without a server it starts attach-only. `docs/tui.md` |
-| Dev with TLS                   | `npx tsx src/index.ts web --https`                                                                                                                                                                             |
-| Override window title hostname | `npx tsx src/index.ts web --title-hostname <name>` (default: `os.hostname()` — `codeman:<name>` is used for tab title, title-flash, and OS desktop notification prefix)                                        |
-| Bind a non-loopback host       | `npx tsx src/index.ts web --host 0.0.0.0` (or `-H`; env `CODEMAN_HOST`; default `127.0.0.1`). Without `CODEMAN_PASSWORD` it **starts but warns loudly** — see Common Gotchas + `docs/security-architecture.md` |
-| Mount under a reverse-proxy sub-path | `npx tsx src/index.ts web --base-url /codeman` (env `CODEMAN_BASE_URL`; default `/`). Normalized in `src/config/base-path.ts` (`''` = root). See Reverse-proxy base path below + `docs/wiki/Remote-Access.md` |
-| Continuous typecheck           | `tsc --noEmit --watch`                                                                                                                                                                                         |
-| Watch-mode test                | `npm run test:watch -- test/<file>.test.ts` (runs the CI gate's config; pass a file to narrow it)                                                                                                               |
-| Test coverage                  | `npm run test:coverage`                                                                                                                                                                                        |
-| Dead-code sweep                | `npm run knip` (config in `config/knip.json`, passed via `--config`)                                                                                                                                          |
 | Rebuild gesture overlay        | `npm run build:gesture` (esbuild `packages/gesture-control/src/codeman/entry.ts` → `src/web/public/gesture/gesture-codeman.js`; commit the result)                                                             |
 | Regenerate the CLI catalogue   | `npm run generate:cli-catalog` (`--check` to fail on drift). Rewrites `config/clis.stock.json` **and** the marked block in `install.sh` from `stock.ts`. ⚠ **Commit both.** They are what `install.sh` and the Docker agent image read, since neither can import TypeScript; `test/cli-catalog-sync.test.ts` and a CI `--check` step fail if either goes stale. See `docs/cli-registry.md` |
 | Build the docker agent image   | `node scripts/build-agent-image.mjs --no-cache` (builds `codeman/agent:base` from `docker/agent.Dockerfile`; prerequisite for Docker cases; `--engine`/`--image`). ⚠ **Always `--no-cache`** — a plain rebuild re-uses the cached `npm install -g` layer and silently keeps the CLIs frozen at their original versions, which once shipped a BROKEN codex while reporting success. See `docs/docker-cases.md` |
-| Gesture playground             | `npm run dev` **in** `packages/gesture-control/` (standalone vite demo, fake tabs)                                                                                                                             |
-| Check public-asset formatting  | `npm run check:public-assets` (prettier-checks `src/web/public/**` text assets; `scripts/check-public-assets.mjs`)                                                                                             |
-| Frontend JS syntax check       | `npm run check:frontend-syntax` (`scripts/check-frontend-syntax.mjs`; runs in CI)                                                                                                                              |
 | Browser-test exclusion check   | `npm run check:browser-excludes` (`scripts/check-browser-test-excludes.mjs`; runs in CI, <1s). Fails if a test importing playwright/puppeteer is still collected by `config/vitest.ci.config.ts`; add it to `BROWSER_TEST_GLOBS` in `config/test-suites.ts` |
 | Pre-push hook                  | Installed by `npm install` (`scripts/git-hooks.mjs`, via postinstall): runs the static CI checks (~10-40s) before `git push`. Skip once: `CODEMAN_SKIP_PREPUSH=1 git push`. Skips itself with a notice when HEAD is not the pushed commit or the tree has uncommitted changes the checks would read. Marker-owned, so a hand-written `pre-push` is never overwritten; installs ONLY into the repo's own `<git-common-dir>/hooks` (worktree-safe; a `core.hooksPath` elsewhere, e.g. a global one, is left alone) |
-| Excluded-suite runners         | `npm run test:browser` · `npm run test:mobile` · `npm run test:perf` · `npm run test:all` (everything, environmental failures included) — see Testing                                                           |
-| Production start               | `npm run start`                                                                                                                                                                                                |
 | Production logs                | `journalctl --user -u codeman-web -f`                                                                                                                                                                          |
 | Detached server                | `codeman web -d` (`--status`, `--stop`; pidfile+log at `dataPath('web.pid'/'web.log')`). ⚠ Refuses to start a 2nd server on one data dir — see Instance isolation                                              |
-| Install/remove the service     | `codeman service install` / `status` / `uninstall` (systemd user unit on Linux, LaunchAgent on macOS; names from `config/service-names.ts`)                                                                     |
-| Dependency doctor              | `codeman doctor` (alias `check-deps`; `--json`, `--category core\|office\|other`). Probes Node/Claude CLI/tmux/LibreOffice/MS Office against `config/dependency-registry.ts`; engine is pure given an injectable `ProbeHost` |
-| Multi-user accounts            | `codeman users add <name>` / `passwd <name>` / `list` / `rm <name>` (writes `~/.codeman/users.json`, mode 0600; see Multi-user mode)                                                                            |
 
 **CI**: `.github/workflows/ci.yml` (push to master/main + PRs, Node 22) runs two jobs: **(1)** `check:lockfile`, `typecheck`, `lint`, `check:frontend-syntax`, `check:browser-excludes`, `format:check`, then a **server boot smoke test** (`tsx src/index.ts web --port 3151` must answer `/api/status` within 30s); **(2)** the **unit/integration test suite** via `npm run test:ci` (`config/vitest.ci.config.ts` — excludes the browser-driven `test/mobile/**` suite, `perf-*` benchmarks, and the Playwright tests in `BROWSER_TEST_GLOBS`; globs live in `config/test-suites.ts`), followed by the **`packages/xterm-zerolag-input` package tests** (a bare `npx vitest run` in that directory; its vitest is hoisted by the root `npm ci`, so no separate install, and `npm test` at the root does NOT run them). `npm test` runs this same config, so local green == CI green. Tests are tmux-safe in CI: `TmuxManager` no-ops all shell commands under `VITEST` (see Testing). A third workflow, `wiki-sync.yml`, fires only on master pushes touching `docs/wiki/**` and mirrors that directory to the GitHub wiki (browser edits to the wiki are overwritten by the next sync, so fix pages via `docs/wiki/`). `browser-suite.yml` runs `npm run test:browser` (minus codex-predictive-echo) nightly and on demand via `workflow_dispatch`; it is informational and never gates a push or PR.
 
@@ -155,29 +112,7 @@ Codeman is a Claude Code session manager with web interface and autonomous Ralph
 
 ### Core Files (by domain)
 
-| Domain           | Key files                                                                                              | Notes                                                                                            |
-| ---------------- | -------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------- |
-| **Entry**        | `src/index.ts`, `src/cli.ts`, `daemon-control`, `service-installer`, `config/service-names`, `cli-style` | The last three back `web -d` / `service install`; `cli-style` is the shared palette/table/spinner/confirm kit |
-| **TUI**          | `src/tui/`: `tui-app` ★ + `tui-client` (the only IO) over a pure core (`-model`, `-layout`, `-render`, `-keys`, `-ansi`, `-composer`, `-approvals`, `-digest`, `-sse`, `-types`) | `codeman tui`, a CLIENT of the server, never a second brain. Design doc: `docs/tui-plan.md`; user guide `docs/tui.md` |
-| **DeepSeek**     | `src/utils/deepseek-cli-resolver.ts`, `src/deepseek-status-shim.ts`, `src/deepseek-web-server.ts` (background `dsh web`, not a session) | `dsh` is a PROFILE LAUNCHER, not an agent; read `docs/deepseek-integration.md` first |
-| **Session**      | `src/session.ts` ★, `session-manager`, `session-auto-ops`, `session-cli-builder`, `session-task-cache`, `session-order` (pure), `session-pty-exit-breaker`, `session-trust-dialog` (pure), `usage-limit-patterns`, `usage-telemetry`; `src/services/unified-session-service.ts` | Pure/unit-tested helpers are split out of `session.ts` on purpose                                 |
-| **Mux**          | `src/mux-interface.ts`, `src/mux-factory.ts`, `src/tmux-manager.ts` ★, `src/proc-tree.ts` (pure, bounded descendant walk)                                  |                                                                                                  |
-| **Respawn**      | `src/respawn-controller.ts` ★ + 4 helpers (`-adaptive-timing`, `-health`, `-metrics`, `-patterns`)      | Read `docs/respawn-state-machine.md` first                                                       |
-| **Ralph**        | `src/ralph-tracker.ts` ★, `src/ralph-loop.ts` + 5 helpers (`-config`, `-fix-plan-watcher`, `-plan-tracker`, `-stall-detector`, `-status-parser`) | Read `docs/ralph-wiggum-guide.md` first                                                          |
-| **Orchestrator** | `src/orchestrator-loop.ts`, `-planner`, `-verifier`                                                    | Read `docs/orchestrator-loop-architecture.md` first                                              |
-| **Cron**         | `src/cron/cron-service.ts`, `cron-time.ts` (pure next-run math), `cron-input.ts`                        | Read `docs/cron-discovery.md` first. Distinct from legacy `ScheduledRun` (`/api/scheduled`)      |
-| **Agents**       | `src/subagent-watcher.ts` ★, `team-watcher`, `bash-tool-parser`, `transcript-watcher`, `workflow-run-watcher` | `workflow-run-watcher` is STANDALONE and never touches `subagent-watcher`                        |
-| **AI**           | `src/ai-checker-base.ts`, `ai-idle-checker.ts`, `ai-plan-checker.ts`                                   |                                                                                                  |
-| **Tasks**        | `src/task.ts`, `task-queue.ts`, `task-tracker.ts`                                                      |                                                                                                  |
-| **State**        | `src/state-store.ts`, `run-summary.ts`, `session-lifecycle-log.ts`, `intent-store.ts`, `tab-layout.ts` (pure model) + `-service` (sole mutation boundary) + `-persistence` + `-legacy-order`                                     |                                                                                                  |
-| **Infra**        | `src/hooks-config.ts`, `push-store`, `tunnel-manager`, `image-watcher`, `file-stream-manager`, `remote-hosts` + `remote-reconnect` + `remote-wake` (IO: `dgram`/`net`/`child_process`), `docker-hosts` + `docker-export` | Remote/docker case overlays; see Key Patterns                                                    |
-| **Web tabs**     | `src/webview-store.ts`, `webview-capabilities.ts`, `src/web/webview-proxy.ts` (pure), `src/web/routes/webview-routes.ts` | Dashboard URLs as tabs; NOT a SessionMode                                                        |
-| **Search**       | `src/search-service.ts`                                                                                | Pure in-memory core for `GET /api/search`                                                        |
-| **Attachments**  | `src/attachment-registry.ts`, `attachment-magic`, `generated-artifact-attachments`, `session-attachment-history`, `document-preview-cache`, `document-thumbnailer`, `document-conversion-limiter`, `config/attachment-guard` | See Key Patterns                                                                                 |
-| **Plan**         | `src/plan-orchestrator.ts`, `src/prompts/*.ts`, `src/templates/` (`claude-md.ts` + `case-template.md`) | `templates/` holds the CLAUDE.md scaffold generated into new cases                                |
-| **Web**          | `src/web/server.ts` ★, `sse-events.ts`, `routes/*.ts` (one module per domain + barrel; `session-routes.ts` ★), `route-helpers.ts`, `ports/*.ts`, `middleware/auth.ts`, `schemas.ts`, `self-update.ts`, `plan-usage-latest.ts`, `ws-connection-registry.ts`, `heic-jpeg-converter.ts` + `heic-jpeg-worker.ts` |                                                                                                  |
-| **Frontend**     | `src/web/public/app.js` (core) + the modules listed in the Frontend load order + `sw.js` (+ `voice-pcm-worklet.js`, fetched from JS, not in the load order) | See Frontend section for the load order, which is authoritative                                  |
-| **Types**        | `src/types/index.ts` (barrel) → domain files; also `src/types.ts` root re-export                    | See `@fileoverview` in index.ts                                                                  |
+**Read before touching a subsystem**: respawn → `docs/respawn-state-machine.md`; Ralph → `docs/ralph-wiggum-guide.md`; orchestrator → `docs/orchestrator-loop-architecture.md`; cron → `docs/cron-discovery.md`; DeepSeek (`dsh` is a PROFILE LAUNCHER, not an agent) → `docs/deepseek-integration.md`; TUI (`src/tui/`, a CLIENT of the server, never a second brain; only `tui-app` + `tui-client` do IO over a pure core) → `docs/tui-plan.md`. Pure helpers are split out of `session.ts` and the other large files on purpose, so tests exercise the shipped code: keep new logic that way.
 
 ★ = Large, central file (>50KB) — read its `@fileoverview` first. All files have `@fileoverview` JSDoc — read that before diving in. Discovery aid: `grep -l '@fileoverview' src/web/routes/*.ts` lists all route modules; same grep works for `src/types/`, `src/web/public/*.js`.
 
@@ -187,42 +122,41 @@ Codeman is a Claude Code session manager with web interface and autonomous Ralph
 
 **Utilities**: `src/utils/` — re-exported via index. Key: `CleanupManager`, `LRUMap` (⚠ NOT in the barrel — import from `./utils/lru-map.js` directly), `StaleExpirationMap`, `BufferAccumulator`, `stripAnsi`, `Debouncer`, `KeyedDebouncer`. Also: `claude-cli-resolver`/`opencode-cli-resolver`/`codex-cli-resolver`/`gemini-cli-resolver`/`antigravity-cli-resolver`/`pi-cli-resolver`/`grok-cli-resolver`/`deepseek-cli-resolver`/`omp-cli-resolver` (CLI path resolution, one per `SessionMode`, all nine sharing the lookup chain in `cli-executable-resolver`: server PATH, then that CLI's install dirs, then an interactive login shell LAST, since it is the only step that spawns anything and it is what finds nvm/Homebrew installs under a service manager's minimal PATH; ⚠ `pi-`, `grok-` and `deepseek-cli-resolver` additionally probe the binary's identity, since `pi` is a generic name, `grok` has npm squatters, and Debian ships an unrelated `dsh`), `file-query` (⚠ Files-panel search matcher, glob-by-two-pointer, never RegExp), `string-similarity` (fuzzy matching), `regex-patterns` (ANSI/token/spinner patterns), `assertNever` (exhaustive checks), `token-validation` (auth tokens), `nice-wrapper` (process priority), `shell-resolver` (⚠ resolves a real login shell for `mode: 'shell'`; the literal string `$SHELL` used to be expanded by the SERVER's shell, which is empty in a container), `event-loop-monitor` (a sync `execSync` freezes the port while the process stays alive, leaving no trace), `dependency-checker` + `dependency-report` (the `codeman doctor` probe engine, registry in `config/dependency-registry.ts`).
 
-### Data Flow
-
-1. Session spawns `claude --dangerously-skip-permissions` via node-pty
-2. PTY output buffered, ANSI stripped, parsed for JSON messages
-3. WebServer broadcasts to SSE clients at `/api/events`
-4. State persists to `~/.codeman/state.json` via StateStore
-
 ### Key Patterns
 
-**Input**: `session.writeViaMux()` for programmatic/curl input via tmux `send-keys -l` + `send-keys Enter`, single-line only. Interactive **browser** input goes through a durable **exactly-once** layer: a stable `clientId` + monotonic per-session `seq` persisted to localStorage until the server ACKs, so a dropped link cannot lose or double-deliver a prompt. `ws-connection-registry.ts` supersedes only same-TAB reconnects, so two tabs on one session coexist. → [architecture-invariants#input-delivery-and-ws-resilience](docs/architecture-invariants.md#input-delivery-and-ws-resilience)
+**Input**: `session.writeViaMux()` for programmatic/curl input via one tmux bracketed paste (`paste-buffer -p`) + `send-keys Enter`, single-line only. A single character is typed with `send-keys -l` instead, because selection dialogs ignore a pasted digit. Interactive **browser** input goes through a durable **exactly-once** layer: a stable `clientId` + monotonic per-session `seq` persisted to localStorage until the server ACKs, so a dropped link cannot lose or double-deliver a prompt. `ws-connection-registry.ts` supersedes only same-TAB reconnects, so two tabs on one session coexist. → [architecture-invariants#input-delivery-and-ws-resilience](docs/architecture-invariants.md#input-delivery-and-ws-resilience)
 
-**Agent wait primitives**: bounded long-polls: `GET /api/sessions/:id/wait`, `GET .../wait-output` (literal substring, **never** regex) and `wait`/`waitTimeout` on `POST .../input`. Registry `session-wait-registry.ts` (pure), bounds `config/agent-wait.ts`. ⚠️ A timeout is a 200 (`wait.timedOut`). ⚠️ `stop`/`blocked` exist for `claude` and `deepseek` ONLY (rule lives in `hooksAvailableForMode()`): explicit request elsewhere is a 400. ⚠️ Send-and-wait registers the waiter BEFORE the write; teardown must `notifySignal('exit')` BEFORE `cancelAll()`; hangup abort listens on `reply.raw` (guarded by `writableFinished`), never `req.raw`; liveness comes from `isPaneDead`, never `session.pid`. ⚠️ Signals are edge-triggered with no history: gather fan-outs via send-and-wait or `wait-output` markers. Packaged as the `skills/codeman` skill (`codeman skill install`, plugin marketplace, or injection behind `agentSkillEnabled`, SYNCED, default OFF): injection is add-only, marker-owned (`applyAgentSkill`), refuses symlinks, and refreshes a marker-owned user-level copy (`refreshUserAgentSkill`). → [architecture-invariants#agent-wait-primitives](docs/architecture-invariants.md#agent-wait-primitives), `docs/api-reference.md`
+**Agent wait primitives**: bounded long-polls: `GET /api/sessions/:id/wait`, `GET .../wait-output` (literal substring, **never** regex) and `wait`/`waitTimeout` on `POST .../input`. Registry `session-wait-registry.ts` (pure), bounds `config/agent-wait.ts`. ⚠️ A timeout is a 200 (`wait.timedOut`). ⚠️ `stop` exists for `claude` and `deepseek` ONLY (rule lives in `hooksAvailableForMode()`), `blocked` also for a CLI declaring `capabilities.workDetect.blockedLine` (codex, read off the screen; `blockedSignalAvailable()`): explicit request elsewhere is a 400. ⚠️ Send-and-wait registers the waiter BEFORE the write; teardown must `notifySignal('exit')` BEFORE `cancelAll()`; hangup abort listens on `reply.raw` (guarded by `writableFinished`), never `req.raw`; liveness comes from `isPaneDead`, never `session.pid`. ⚠️ Signals are edge-triggered with no history: gather fan-outs via send-and-wait or `wait-output` markers. Packaged as the `skills/codeman` skill (`codeman skill install`, plugin marketplace, or injection behind `agentSkillEnabled`, SYNCED, default OFF, into the directory each CLI scans: `capabilities.agentSkillDir`, `.claude/skills` for claude, `.agents/skills` for the others, none for pi): injection is add-only, marker-owned (`applyAgentSkill`), refuses symlinks, and refreshes a marker-owned user-level copy (`refreshUserAgentSkill`). → [architecture-invariants#agent-wait-primitives](docs/architecture-invariants.md#agent-wait-primitives), `docs/api-reference.md`
+
+**Agent CLI (`codeman agent …`, `src/cli-agent.ts`)**: the command-line client for the wait primitives and the session verbs, for agents the skill does not reach (pi, shell, no `jq`; `seedAgentSessionPreamble` covers every local session whose CLI takes the skill, `capabilities.agentSkillDir`; every pane still exports `CODEMAN_MUX`/`CODEMAN_SESSION_ID`/`CODEMAN_API_URL`). It is a CLIENT: no new route, no second transport (the plan's "Socket API" non-goal stands), every call carries `X-Codeman-Parent-Session`, only `spawn`'s quick-start carries `X-Codeman-Agent-Origin` (pinned in test/cli-agent.test.ts), auth (`src/codeman-credentials.ts`, shared with `codeman attach` and the TUI) from `CODEMAN_PASSWORD` or the data dir's `.env` (the same reader `codeman attach` uses — `readCodemanEnvFile`). Verbs return exit codes (`0` ok / `1` error / `2` timeout / `3` exited / `4` refused) instead of calling `process.exit`, so `test/cli-agent.test.ts` drives every verb against a recording fake transport plus the real `httpRequest` against a local server. ⚠️ Three guards live in code, not prose, and the tests pin them: refuse without `CODEMAN_MUX=1`+`CODEMAN_API_URL` (never guess a URL); `send` transmits printable text + `\r` only (`inputRefusal`: any byte < 0x20 or DEL refuses BEFORE the transport — opencode's `Ctrl+C` is `app_exit`), ESC exists solely as `interrupt` with no `\r`; `rm` fails closed (`deleteRefusal`: empty id, self shorter than 8, prefix match in EITHER direction, re-checked on the resolved id). ⚠️ The routes accept FULL ids only — an 8-char prefix is a 404 (measured live) — while `ls` prints prefixes, so `resolveSessionId` maps a short id through `GET /api/v1/sessions` and refuses an ambiguous one. It also refuses any id under 8 characters (exit 4, on every verb, before any request), the same floor as the server's `PARENT_SESSION_ID_MIN_PREFIX`, so `rm 9` can never pick an arbitrary session. ⚠️ A 400 for `--until stop` on a hook-less mode is passed through verbatim, never papered over: the marker path is the answer there. `spawn` on a readiness timeout exits 2 and LEAVES the session for inspection (the skill's `spawn_worker` deletes it; the CLI's caller is often a human); its stdout is the id ALONE (prose on stderr) so `SID=$(…)` works. ⚠️ Two server answers that look like success and are not: `delivered:false` without `duplicate` (the bytes went into a dead pane — the field exists so a client says "restart" instead of "wait longer") and `wait.ended:true, signal:null` (the worker died during a `stop`/`idle` wait: the registry satisfies only waiters that listed `exit`, then cancels the rest) — both exit 3, checked BEFORE the happy paths in `waitExitCode`. → README "`codeman agent`", `docs/api-reference.md`
+
+**MCP server (`codeman mcp`, `src/mcp-server.ts`)**: a Model Context Protocol server on stdio for any MCP client, a CLIENT of the HTTP API like `codeman agent` (reuses its `httpRequest`, `inputRefusal`, `buildSendBody`; no new route). Seven tools: `list_sessions`, `start_session`, `send_prompt`, `wait_for_session`, `read_output`, `interrupt_session`, `delete_session`. ⚠️ Dual-era: answers the legacy `initialize` handshake (2024-11-05 to 2025-11-25) AND the 2026-07-28 revision (per-request `_meta` with `protocolVersion` + `clientCapabilities`, `server/discover`, `resultType: "complete"` on every modern result); a modern `tools/list`/`server/discover` MUST carry `ttlMs` + `cacheScope` (Claude Code 2.1.296 drops every tool without them: measured). ⚠️ stdout is protocol only; logs go to stderr. ⚠️ The URL comes from `--url`/`CODEMAN_API_URL`, never a guess, and unconfigured it still lists tools (registries introspect) while every call answers how to configure it. ⚠️ Prompt/interrupt/delete only sessions the PROCESS started (in-memory allowlist, so a restart fails closed) or, inside a pane, the calling session's own children, unless `--allow-any-session`; never its own session. ⚠️ `read_output` returns text ONLY, no `structuredContent` (a client that prefers it showed the model metadata instead of the output: measured). Waits default 45 s (Codex's MCP tool timeout is 60 s). Registry manifest `config/mcp-registry/server.json` (`mcpName` in package.json, version kept by `scripts/sync-plugin.mjs`). Tests: `test/mcp-server.test.ts`.
 
 **Agent-created case marker** (`src/agent-case-marker.ts`): a case dir that `POST /api/quick-start` CREATES for an agent-driven spawn (signal: the preamble's `X-Codeman-Agent-Origin` header / `agentOrigin` field, else a resolved `parentSessionId`) gets `.codeman-agent-case.json`, published as `agentCreated` on `GET /api/cases`; `GET /api/cases/agent-created` is the cleanup listing (`inUse`, `modifiedAt`) behind Add Case → Manage. ⚠️ Only the branch that CREATES the directory may write it: never label a linked case, cloned repo or pre-existing path (it drives a recursive delete). ⚠️ Reading is total: anything but a well-formed v1 marker reads as not agent-created. ⚠️ Removal stays on `DELETE /api/cases/:name` (the ONE recursive-delete path), and the sweep excludes `inUse` cases. ⚠️ Changing the preamble's headers requires bumping `CODEMAN_PREAMBLE`. → [architecture-invariants#agent-created-case-marker](docs/architecture-invariants.md#agent-created-case-marker)
 
-**Agent preamble cache GC**: the §0 preamble seeded per claude session (`$XDG_CACHE_HOME/codeman-agent-<id>.sh`) is now REMOVED with the session (`removeAgentSessionPreamble` from `_doCleanupSession`, `killMux` only — a detach leaves the session recoverable and its agent would come back to a loader whose file we deleted) and swept at boot (`pruneAgentSessionPreambles(this.sessions.keys())`, once, after restore, so every session this instance owns is in the keep set). Nothing removed them before: 236 leftovers measured on a working machine, the oldest three weeks old. ⚠️ The sweep needs BOTH guards — never a live session's file at any age (the two-line loader reads it mid-run), and `AGENT_PREAMBLE_MAX_AGE_MS` (7d) of age on top, which is what keeps ANOTHER instance's sessions (whose ids this process cannot see) out of the blast radius. Losing one is degradation, not breakage: the §0 fallback block rewrites it. Tests live with the seed's in `test/agent-skill.test.ts`.
+**Agent preamble cache GC**: the §0 preamble seeded per local session whose CLI takes the skill (`capabilities.agentSkillDir`) (`$XDG_CACHE_HOME/codeman-agent-<id>.sh`) is now REMOVED with the session (`removeAgentSessionPreamble` from `_doCleanupSession`, `killMux` only — a detach leaves the session recoverable and its agent would come back to a loader whose file we deleted) and swept at boot (`pruneAgentSessionPreambles(this.sessions.keys())`, once, after restore, so every session this instance owns is in the keep set). Nothing removed them before: 236 leftovers measured on a working machine, the oldest three weeks old. ⚠️ The sweep needs BOTH guards — never a live session's file at any age (the two-line loader reads it mid-run), and `AGENT_PREAMBLE_MAX_AGE_MS` (7d) of age on top, which is what keeps ANOTHER instance's sessions (whose ids this process cannot see) out of the blast radius. Losing one is degradation, not breakage: the §0 fallback block rewrites it. Tests live with the seed's in `test/agent-skill.test.ts`.
 
 **Idle detection**: Multi-layer (completion message → AI check → output silence → token stability). See `docs/respawn-state-machine.md`.
 
-⚠️ **A `❯` sighting is NOT the end of a turn, and neither is silence.** Claude redraws the composer all through a turn, and its working line (`✻ Actualizing… (13m 23s · …)`) is invisible to `SPINNER_PATTERN`, keyword lists and the raw stream. `_confirmIdle()` (session.ts) requires the pane to go quiet AND the SCREEN (`capturePaneText()` + the working-line pattern) to agree; a sustained run of repaints (`session-activity.ts`) marks a turn as started. ⚠️ The composer glyph and working line are per-CLI registry DATA (`capabilities.workDetect`), never Claude constants; a CLI declaring neither falls back to Claude's pair. ⚠️ `workingLine` is config-supplied and runs on the PTY hot path, so it must compile through `compileVersionRegex()` in BOTH the schema refine and `_workingLinePattern()` (ReDoS guard; null, not throw). → [architecture-invariants#idle-detection-composer-glyph-and-working-line](docs/architecture-invariants.md#idle-detection-composer-glyph-and-working-line)
+⚠️ **A `❯` sighting is NOT the end of a turn, and neither is silence.** Claude redraws the composer all through a turn, and its working line (`✻ Actualizing… (13m 23s · …)`) is invisible to `SPINNER_PATTERN`, keyword lists and the raw stream. `_confirmIdle()` (session.ts) requires the pane to go quiet AND the SCREEN (`capturePaneText()` + the working-line pattern) to agree; a sustained run of repaints (`session-activity.ts`) marks a turn as started. ⚠️ The workspace-trust dialog's `❯` (on `No, exit`) arms the check too, so the same capture holds idle off while `isTrustDialogScreen()` matches inside `TRUST_DIALOG_WINDOW_MS`: an agent that waited for idle and prompted the dialog confirmed `No, exit` and killed the pane. ⚠️ The composer glyph and working line are per-CLI registry DATA (`capabilities.workDetect`), never Claude constants; a CLI declaring neither falls back to Claude's pair. ⚠️ `workingLine` is config-supplied and runs on the PTY hot path, so it must compile through `compileVersionRegex()` in BOTH the schema refine and `_workingLinePattern()` (ReDoS guard; null, not throw). → [architecture-invariants#idle-detection-composer-glyph-and-working-line](docs/architecture-invariants.md#idle-detection-composer-glyph-and-working-line)
 
 ⚠️ **A turn that ENDED waiting for its own workers is working, not idle.** Claude closes such a turn with `✻ Waiting for 1 dynamic workflow to finish` (background agents / ultracode) and resumes by itself; `capabilities.workDetect.awaitingLine` makes the idle probe count it as work. ⚠️ Claude never redraws that row, so it stays on screen after the workers finish: test it ONLY as the newest column-0 row above the composer (`isAwaitingWorkers()`), never pane-wide and never on the stream. → [architecture-invariants#idle-detection-composer-glyph-and-working-line](docs/architecture-invariants.md#idle-detection-composer-glyph-and-working-line)
 
+⚠️ **A codex dialog is BLOCKED, read off the screen, and only ever an ALERT.** Codex has no hook for its approval, question and trust dialogs, so `capabilities.workDetect.blockedLine` names their chrome and `web/screen-blocked.ts` raises what a `permission_prompt` hook raises, as an `answerInTerminal` item. ⚠️ Never key a keystroke off that match (the agent writes most of the pane), and never hand its item a `capture` (the inbox's sweep is Claude-shaped; the session's 2 s re-read owns the clear). A blocked screen beats the working line. → [architecture-invariants#screen-read-blocked-dialogs-codex](docs/architecture-invariants.md#screen-read-blocked-dialogs-codex)
+
 ⚠️ **A quiet pane is not always a pane that wants you.** A CLI can declare an optional `capabilities.workDetect.watchingLine` (a monitor, background shell or cloud hand-off it is still running); the idle probe reads it into `Session.watching` and `notePrompt()` opens that idle item ALREADY acknowledged, so no surface alerts. Only `idle` is eligible, and the label is pane-derived and prompt-injectable, so a pattern must anchor on chrome only that CLI draws. → [architecture-invariants#the-watching-signal-a-quiet-pane-that-is-not-waiting-for-you](docs/architecture-invariants.md#the-watching-signal-a-quiet-pane-that-is-not-waiting-for-you). Tests: `test/session-watching.test.ts`, `test/watching-no-alert.test.ts`.
 
 **An exited agent in a live pane** (`paneExit`, #446): panes use `remain-on-exit on`, so `/exit` leaves a pane, session and pid that look alive; `TmuxManager.startPaneExitWatcher()` publishes `SessionState.paneExit` via `session:updated`. ⚠️ Never set `status: 'error'` or null the `pid` for it; the field is TRI-STATE (absent = UNKNOWN, never alive, scoped by `Session.paneExitApplies`); an absent `#{pane_dead_status}` is not 0; a path that starts a command in a pane must clear the record AND persist. A clean exit is CLOSED via `cleanupSession()` (`pane-exit-sweep.ts`): only an explicit numeric status 0 with no signal, confirmed by 2 reads, with no start/attach in flight (`paneLifecycleInFlight`) and not within 10 s of one (a startup error keeps its row); a crashed agent keeps its row. → [architecture-invariants#an-exited-agent-in-a-live-pane-paneexit](docs/architecture-invariants.md#an-exited-agent-in-a-live-pane-paneexit)
 
 **Dead-pane respawn resume pin** (`_buildRespawnPaneOptionsWithResumePin()`, session.ts): recovering a dead pane, like a custom-model `restartCli()`, must pin the conversation or claude refuses the reused `--session-id`. The pin takes the first transcript-backed candidate (chain tail, launch seed, own id), never `_claudeSessionId`, adds nothing when none is backed, and is never applied to remote or docker sessions. → [architecture-invariants#dead-pane-respawn-the-resume-pin](docs/architecture-invariants.md#dead-pane-respawn-the-resume-pin)
 
-**Workspace-trust dialog auto-accept** (`session-trust-dialog.ts`, pure): Claude Code's per-directory trust dialog is always answered yes, or the session is stuck. ⚠️ Match the compacted SCREEN (`compactScreenText()`, all whitespace removed), never the stream (tmux sends words joined by cursor-forwards, not spaces). ⚠️ Never answer with a blind `\r` (newer versions highlight "No, exit" first): `trustDialogNextKey()` returns ONE key per re-read frame (arrow, then Enter only once `❯` is on the trust option), and the LAST marked option wins. ⚠️ All three guards must hold: startup window `TRUST_DIALOG_WINDOW_MS` (90s), two-marker match (`isTrustDialogScreen`), attempt cap `TRUST_DIALOG_MAX_ATTEMPTS` (6). ⚠️ Read `capturePaneText()`; only a direct-PTY session falls back to a SHORT buffer tail. ⚠️ The scan must schedule its own next read (`_trustDialogTimer`, cleared in `_clearAllTimers()`), not rely on PTY output. → [architecture-invariants#workspace-trust-dialog-auto-accept](docs/architecture-invariants.md#workspace-trust-dialog-auto-accept)
+**Workspace-trust dialog auto-accept** (`session-trust-dialog.ts`, pure): Claude Code's per-directory trust dialog is always answered yes, or the session is stuck. ⚠️ Match the compacted SCREEN (`compactScreenText()`, all whitespace removed), never the stream (tmux sends words joined by cursor-forwards, not spaces). ⚠️ Never answer with a blind `\r` (newer versions highlight "No, exit" first): `trustDialogNextKey()` returns ONE key per re-read frame (arrow, then Enter only once `❯` is on the trust option), and the LAST marked option wins. ⚠️ All three guards must hold: startup window `TRUST_DIALOG_WINDOW_MS` (90s), two-marker match (`isTrustDialogScreen`), attempt cap `TRUST_DIALOG_MAX_ATTEMPTS` (6). ⚠️ Read `capturePaneText()`; only a direct-PTY session falls back to a SHORT buffer tail. ⚠️ The scan must schedule its own next read (`_trustDialogTimer`, cleared in `_clearAllTimers()`), not rely on PTY output, and a throttled read is DEFERRED onto that timer, never dropped (the dialog's paint is often the pane's last output); a stopped session never scans. → [architecture-invariants#workspace-trust-dialog-auto-accept](docs/architecture-invariants.md#workspace-trust-dialog-auto-accept)
 
 **Process-tree walks are bounded** (`proc-tree.ts`, pure): `collectDescendants(pid, byParent)` is the ONE descendant traversal, fed by one cached `ps -eo pid=,ppid=` snapshot (`refreshProcSnapshot()` in tmux-manager.ts: async, in-flight-shared, and ANY error discards the result rather than caching a truncated `ps`). ⚠️ Never walk a process tree with per-node `pgrep` or unbounded recursion (the unbounded version took a machine down): the walk must terminate on cycles, cap depth (`PROC_WALK_MAX_DEPTH`) and node count (`PROC_WALK_MAX_NODES`), and never spawn anything. ⚠️ Keep it in its own module so the test exercises the shipped code, and report truncation through `onTruncated` naming both caps, never silently. → [architecture-invariants#process-tree-walks-are-bounded](docs/architecture-invariants.md#process-tree-walks-are-bounded)
 
 **Auto-resume on usage limit** (opt-in per session, top of the Respawn tab): when Claude halts on a subscription limit, `usage-limit-patterns.ts` (pure, unit-tested) parses the reset time and `SessionAutoOps` arms a timer for reset+2min, then sends Esc + `continue`. ⚠️ Respawn cycles are blocked while paused (`isLimitPaused` guard in `onIdleDetected`), which is what prevents `/clear` from wiping the paused conversation. Claude-mode only. → [architecture-invariants#auto-resume-on-usage-limit](docs/architecture-invariants.md#auto-resume-on-usage-limit)
 
-**Plan-usage chip** (`showPlanUsageLimits`, per-device: desktop default **ON**, handhelds OFF): resolve DISPLAY only through `planUsageChipEnabled()` (settings-ui.js). The same setting is the server-side COLLECTION switch, read fresh by `readPlanUsageTelemetryEnabled()` (hooks-config.ts) at every claude create/respawn. ⚠️ An ABSENT key reads as ON in the reader; `GET /api/settings` must never write. ⚠️ A save sends `showPlanUsageLimits` ONLY when it flips the chip on that device (`planUsageCollectionFlip()`), or a phone switches collection off for every desktop. Claude data comes from the statusLine exporter, injected as an EPHEMERAL `claude --settings` flag (`resolveStatusLineCliCommand`), never written to disk, WRAPPING a user's own statusLine, posting to `POST /api/status-telemetry`. Codex comes from a read-only `account/rateLimits/read` poll (main bucket only). → [architecture-invariants#plan-usage-chip-statusline-telemetry](docs/architecture-invariants.md#plan-usage-chip-statusline-telemetry), `docs/usage-limits-display-plan.md`
+**Plan-usage chip** (`showPlanUsageLimits`, per-device: desktop default **ON**, handhelds OFF): resolve DISPLAY only through `planUsageChipEnabled()` (settings-ui.js). The same setting is the server-side COLLECTION switch, read fresh by `readPlanUsageTelemetryEnabled()` (hooks-config.ts) at every claude create/respawn. ⚠️ An ABSENT key reads as ON in the reader; `GET /api/settings` must never write. ⚠️ A save sends `showPlanUsageLimits` ONLY when it flips the chip on that device (`planUsageCollectionFlip()`), or a phone switches collection off for every desktop. Claude data comes from the statusLine exporter, injected as an EPHEMERAL `claude --settings` flag (`resolveStatusLineCliCommand`), never written to disk, WRAPPING a user's own statusLine, posting to `POST /api/status-telemetry`. Codex comes from a read-only `account/rateLimits/read` poll (main bucket only). The statusline's `prompt_cache` object (Claude Code v2.1.251+) is kept by `StatusTelemetrySchema` (every documented scalar field, null-tolerant; `parsePromptCache()` in usage-telemetry.ts normalizes only the stored subset) and shown in the footer as `cache:until HH:MM` / `cache:cold` (only when the user has no statusline of their own) and in Session Options → Respawn (`#promptCacheStatus`, fed by `Session.promptCache` in `toState()` plus the change-only `session:promptCache` SSE event, worded by the pure `CodemanPromptCache.format`): warm-until/cold, the re-cache stake, the hit rate, the last miss's causes, and a `Context compacted at HH:MM` note after a compaction or a tool-result clearing (stamped on the first warm report whose `recache_tokens_if_cold` is null after one that carried it, never from `expected_rebuilds`, which Claude Code bumps only at the next request; NEVER an alert). ⚠️ `Session.setPromptCache()` stamps `compactedAt` onto a NEW object (never mutating the parsed report) and the route broadcasts the STORED object; the last report is restored on restart (constructor `promptCache`, like `displayModel`) and the readout is time-aware (a passed `expiresAt` reads cold, not stale-warm). It is the source of truth for the cache TTL, never the plan tier. → [architecture-invariants#plan-usage-chip-statusline-telemetry](docs/architecture-invariants.md#plan-usage-chip-statusline-telemetry), `docs/usage-limits-display-plan.md`
 
 **Orchestrator**: State machine that turns a user goal into a phased plan and drives it to completion: `idle → planning → approval → executing → verifying → (replanning) → completed/failed`. `OrchestratorLoop` (engine) delegates plan generation to `orchestrator-planner` and per-phase verification gates to `orchestrator-verifier`, executing phases via team agents/`task-queue`. State persists under the `orchestrator` key in `state.json`. Distinct from Ralph (single-session autonomous loop) — orchestrator coordinates multi-phase, multi-agent execution. See `docs/orchestrator-loop-architecture.md`.
 
@@ -236,7 +170,7 @@ Codeman is a Claude Code session manager with web interface and autonomous Ralph
 
 **Docker Compose deployment** (`docker/`): Codeman runs in a container and spawns Docker cases as **SIBLING** containers via the host socket, never nested. `resolveDockerDaemonMountSource()` maps HOME bind sources into the daemon's namespace (`CODEMAN_DOCKER_HOST_HOME`); `CODEMAN_CASES_PATH` makes workspaces resolve to the same absolute path on both sides. ⚠️ `CODEMAN_CASES_PATH` must move every consumer: resolve it only via `config/cases-dir.ts`. ⚠️ `.dockerignore` matches whole paths: keep `**/.env` or `docker/.env` secrets ship in the image. ⚠️ Long-form binds create missing sources ROOT-OWNED: `Start-Codeman.sh` pre-creates them, and `docker/entrypoint.sh` (root, `cap_add: [CHOWN, DAC_OVERRIDE, KILL, SETGID, SETUID]` against `cap_drop: ALL`, `KILL` for tini; pinned by the test) fixes ownership then drops to `PUID:PGID` via `setpriv`, never re-owning foreign dirs. ⚠️ Append `/opt/codeman-cli` and `~/.local/bin` to `PATH`, never prepend. ⚠️ `server.Dockerfile`, the compose file and `.env.example` feed the self-updater's environment gate (`docs/docker-self-update.md`). → [architecture-invariants#docker-compose-deployment](docs/architecture-invariants.md#docker-compose-deployment), `docs/docker-compose.md`
 
-**CLI registry** (`src/config/cli-registry/`): every run mode is a `CliEntry` (discovery, launch argv template, env handling, `capabilities`, and the `overlays` behind remote/docker pane commands). **No code outside `stock.ts` may branch on a CLI id**: use a capability field or a NAMED PROFILE (`profiles.ts`); `test/cli-registry-no-id-branching.test.ts` and `test/frontend-cli-no-id-branching.test.ts` enforce it. ⚠️ Config holds typed argv tokens, never shell text; literals are validated at LOAD time and a bad one rejects the whole entry. ⚠️ Keep `external`, `hooks` and `altScreen` independent; never derive one from another. ⚠️ Config regexes (`discovery.version.regex`, `capabilities.workDetect.workingLine`, `capabilities.workDetect.watchingLine`, `capabilities.workDetect.awaitingLine`, `capabilities.modelDetect.screenLine`) must compile through `compileVersionRegex()`. ⚠️ `privilegedParams[].param` names a LAUNCH PARAM, not the legacy `<Mode>Config` field (bridged only by `launch.legacyConfigAliases`); a wrong name silently clamps nothing. ⚠️ Resolve the registry AT CALL TIME, never in a module-level const. ⚠️ Remote claude/omp arms of `buildRemoteLaunchCommand` are not covered by the pane-command golden. `~/.codeman/clis.json` overrides entries. It is WRITTEN only by the opt-in CLI management routes (`cliManagementEnabled`, default OFF; `/api/clis`, `cli-registry-routes.ts`), and only through `mutateRegistryFile()` in `registry-writer.ts`, which serializes mutations and refuses (409) a file that does not parse or has group/world permission bits rather than overwriting it. Importing the registry still writes nothing. → [architecture-invariants#cli-registry](docs/architecture-invariants.md#cli-registry), `docs/cli-registry.md`
+**CLI registry** (`src/config/cli-registry/`): every run mode is a `CliEntry` (discovery, launch argv template, env handling, `capabilities`, and the `overlays` behind remote/docker pane commands). **No code outside `stock.ts` may branch on a CLI id**: use a capability field or a NAMED PROFILE (`profiles.ts`); `test/cli-registry-no-id-branching.test.ts` and `test/frontend-cli-no-id-branching.test.ts` enforce it. ⚠️ Config holds typed argv tokens, never shell text; literals are validated at LOAD time and a bad one rejects the whole entry. ⚠️ Keep `external`, `hooks` and `altScreen` independent; never derive one from another. ⚠️ Config regexes (`discovery.version.regex`, `capabilities.workDetect.workingLine`, `capabilities.workDetect.watchingLine`, `capabilities.workDetect.awaitingLine`, `capabilities.workDetect.blockedLine`, `capabilities.modelDetect.screenLine`) must compile through `compileVersionRegex()`. ⚠️ `privilegedParams[].param` names a LAUNCH PARAM, not the legacy `<Mode>Config` field (bridged only by `launch.legacyConfigAliases`); a wrong name silently clamps nothing. ⚠️ Resolve the registry AT CALL TIME, never in a module-level const. ⚠️ Remote claude/omp arms of `buildRemoteLaunchCommand` are not covered by the pane-command golden. `~/.codeman/clis.json` overrides entries. It is WRITTEN only by the opt-in CLI management routes (`cliManagementEnabled`, default OFF; `/api/clis`, `cli-registry-routes.ts`), and only through `mutateRegistryFile()` in `registry-writer.ts`, which serializes mutations and refuses (409) a file that does not parse or has group/world permission bits rather than overwriting it. Importing the registry still writes nothing. → [architecture-invariants#cli-registry](docs/architecture-invariants.md#cli-registry), `docs/cli-registry.md`
 
 **External CLI modes (OpenCode, Codex, Gemini, Antigravity, Pi, Grok, DeepSeek, OMP)**: `isExternalCliMode()` in `session.ts` gates Claude-specific behavior off (Ralph tracker, BashToolParser, token parsing, ❯ readiness; readiness is output stabilization); work detection is per-CLI `capabilities.workDetect` data, not this gate. All eight **require tmux, no direct PTY fallback** (secrets go via socket-scoped `tmux setenv`, never the command line). ⚠️ `run*()` in `session-ui.js` MUST unwrap the `{success,data}` envelope. ⚠️ **Codex uses predictive write-through echo, never the buffer overlay**: `_predictHookOnData` must never `return` (wire path stays byte-identical), and flushed text and a bracketed paste must go out as separate delayed writes. ⚠️ **Pi**: no bypass flag, never invent one; `approveProjectTrust` executes repo code, so it is in the clamp's **materialize** branch; never wire `--api-key`. ⚠️ **Grok**: `alwaysApprove` is stripped for non-granted owners (only-if-sent). ⚠️ **DeepSeek**: the agent is a PROFILE (Run gates on `isDeepSeekRunnable()`); the permission switch is the `DSH_PERMISSION_MODE` env var, so `clampEnvOverridesForOwner()` must DROP `DSH_PERMISSION_MODE`, `DSH_HOME` and `DEEPSEEK_BASE_URL` for non-granted owners; `hooksAvailableForMode()` is per-SESSION for it (pass `sessionHookOptions(session)`) and is never a stand-in for `mode === 'claude'`; answers come from `deepseek-transcript.ts`, paired by header `cwd` + boot window, never newest-mtime. ⚠️ **OMP**: `OMP_AUTH_BROKER_URL`/`_TOKEN` are clamped the same way. → [architecture-invariants#external-cli-modes-opencode-codex-gemini-antigravity-pi-grok-deepseek-omp](docs/architecture-invariants.md#external-cli-modes-opencode-codex-gemini-antigravity-pi-grok-deepseek-omp)
 
@@ -246,11 +180,7 @@ Codeman is a Claude Code session manager with web interface and autonomous Ralph
 
 ⚠️ **llama-swap endpoints** (one model at a time): the apply routes check `GET /running` and return `requiresConfirmation` before evicting a model another live session uses; `confirmedSwap` and `confirmedContext` are SEPARATE flags and must stay so. Claude alone gets a context floor (`CLAUDE_MIN_SAFE_CONTEXT_TOKENS`); context is parsed from `/running`'s `cmd`, never trusted from `/props`. Backend log lines come from llama-swap's `/api/events` `upstream` source, never `/logs`. → [architecture-invariants#custom-model-endpoint-profiles](docs/architecture-invariants.md#custom-model-endpoint-profiles)
 
-**MCP server sync** (opt-in, `mcpSyncEnabled`, SYNCED, default OFF; `src/mcp-sync.ts`, `GET`/`POST /api/mcp-sync`): copies each installed, enabled CLI's user-level MCP servers into the others. It is the ONE subsystem that writes another CLI's REAL user config (`~/.claude.json`, `~/.codex/config.toml`, `~/.gemini/*`, opencode's), which is why it is opt-in and admin-only in multi-user mode (both verbs 403 for a non-admin, and the Settings group is hidden for them). Where each CLI keeps the file is registry data, `capabilities.mcpConfig` (`{ path, format, relocation? }`), never a branch on the id. ⚠️ ADDITIVE only: a name already defined, in any shape, is never edited or removed (a different same-name definition is a reported conflict), and a server switched off in its own CLI is never copied. ⚠️ Never write a file that did not parse; re-parse the NEW text and require every added server to read back before the tmp+rename (written through a symlink, previous file kept as `<file>.codeman-bak`, one apply at a time, else 409). ⚠️ A file that receives copied `env`/`headers` (secrets) is left `0600`, and so is the backup. ⚠️ Responses carry server NAMES only, never env values, headers or file text: a parse failure is reported by line and column (`describeMcpSyncError`), never the parser's own message (smol-toml and V8 both quote source). ⚠️ `mcpConfig.relocation` names the env var the CLI reads to move its file (`CLAUDE_CONFIG_DIR`, `CODEX_HOME`, `XDG_CONFIG_HOME`, `GEMINI_CLI_HOME`), resolved from the SERVER env at call time; a relative value reports the target `skipped`, never a guessed write, and a per-session `envOverrides` relocation is not followed. Tests must pass `home` (which drops the `process.env` default) or clear those vars first. → `docs/cli-registry.md` (MCP server sync), `docs/api-reference.md`, `docs/wiki/Settings-Reference.md`
-
-**Run launch synchronization**: the Run entrypoint holds an in-flight lock and disables `#runBtn` for the whole launch (≥500ms) so a double click cannot create duplicate `w<n>-<case>` sessions; `_ensureCreatedSessionVisible()` runs before `selectSession()` and `_onSessionCreated()` stays an idempotent upsert, so POST-first and SSE-first both render exactly one tab. ⚠️ **Closing has the mirror-image race**: `closeSession()` must read `wasActive` BEFORE its `await` and announce the delete via `_closingSessions`, and `_onSessionDeleted` skips the active-session handoff for ids in that set; never read `activeSessionId` after the fact. The fallback picks the first `sessionOrder` entry still in `sessions`. ⚠️ The close is OPTIMISTIC: the tab goes and the next one is selected BEFORE the DELETE is sent (upserts skip ids in `_closingSessions`), and only a delete the server refused (checked with a GET) puts the row back; never await the server before hiding the tab. Tests: `test/session-close-fallback.test.ts`. → [architecture-invariants#run-launch-synchronization](docs/architecture-invariants.md#run-launch-synchronization)
-
-**Session lineage lines** (tab → tab it spawned, `sessionLineageLines`, per-device, desktop default ON): a create request may name its spawner via a `parentSessionId` body field or the `X-Codeman-Parent-Session` header; `resolveParentSessionId()` (route-helpers.ts) resolves it (exact id or unique ≥8-char prefix, live, visible, same owner) and ⚠️ anything unresolvable is DROPPED, never a 400. Rides `toState()`, no new SSE event. ⚠️ Rendering is a LAYER on the existing SVG pass (`_appendLineageConnectionLines` at the tail of `_updateConnectionLinesImmediate()`) and draws EVERY family, emphasizing the selected tab's (what it spawned, and the family it was spawned into: `lineage-family--focus`, thicker, opaque, drawn last; lanes stay in strip order), so a selection change must redraw (`_updateActiveTabImmediate`, gated on `_lineageTotalEdges`), and again on the strip's size `transitionend` (the active tab widens for ~150ms after the redraw). Geometry is pure in `computeLineageTree()`: one rounded orthogonal tree per spawning tab, every route starting at the PARENT (siblings share the trunk exactly), routed ONLY through the gaps between tab rows plus a spine left of every row, so it never crosses a tab or the terminal. ⚠️ That room is reserved in CSS by `.session-tabs.lineage-tree` (`_syncLineageGutter()`, called before the wrap is measured), keyed on whether ANY family exists, never on the selection, or every tab switch would resize the header and the PTY. Colors are keyed on the SPAWNING tab, claimed in strip order for every family, and memoized (never by draw index or selection order). ⚠️ Desktop only (z-index vs the fixed mobile header). ⚠️ Paths must keep `data-agent-id="lineage:<childId>"` (the entrance animation queries it); skip endpoints scrolled out of the strip. → [architecture-invariants#session-lineage-lines-tab--tab-it-spawned](docs/architecture-invariants.md#session-lineage-lines-tab--tab-it-spawned)
+**MCP server sync** (opt-in, `mcpSyncEnabled`, SYNCED, default OFF; `src/mcp-sync.ts`, `GET`/`POST /api/mcp-sync`): copies each installed, enabled CLI's user-level MCP servers into the others. It is the ONE subsystem that writes another CLI's REAL user config (`~/.claude.json`, `~/.codex/config.toml`, `~/.gemini/*`, opencode's), which is why it is opt-in and admin-only in multi-user mode (both verbs 403 for a non-admin, and the Settings group is hidden for them). Where each CLI keeps the file is registry data, `capabilities.mcpConfig` (`{ path, format, relocation? }`), never a branch on the id. ⚠️ Tools that sync but are not run modes (GitHub Copilot CLI, the first) are plain data in `MCP_SYNC_ONLY_TOOLS` (`src/mcp-sync-targets.ts`), take part only when installed or already configured, and their companion `settings.json` (`disabledMcpServers`) is read, never written; a test must clear their relocation vars too. ⚠️ ADDITIVE only: a name already defined, in any shape, is never edited or removed (a different same-name definition is a reported conflict), and a server switched off in its own CLI is never copied. ⚠️ Never write a file that did not parse; re-parse the NEW text and require every added server to read back before the tmp+rename (written through a symlink, previous file kept as `<file>.codeman-bak`, one apply at a time, else 409). ⚠️ A file that receives copied `env`/`headers` (secrets) is left `0600`, and so is the backup. ⚠️ Responses carry server NAMES only, never env values, headers or file text: a parse failure is reported by line and column (`describeMcpSyncError`), never the parser's own message (smol-toml and V8 both quote source). ⚠️ `mcpConfig.relocation` names the env var the CLI reads to move its file (`CLAUDE_CONFIG_DIR`, `CODEX_HOME`, `XDG_CONFIG_HOME`, `GEMINI_CLI_HOME`, and `COPILOT_HOME` for Copilot), resolved from the SERVER env at call time; a relative value reports the target `skipped`, never a guessed write, and a per-session `envOverrides` relocation is not followed. Tests must pass `home` (which drops the `process.env` default) or clear those vars first. → `docs/cli-registry.md` (MCP server sync), `docs/api-reference.md`, `docs/wiki/Settings-Reference.md`
 
 **Auto-named sessions** (`autoNameSessions`, SYNCED, default OFF): a placeholder tab (`w3-myapp`) takes its first real prompt as a title in the `<prefix>: <title>` form, so the case identity and `w<n>` counter survive. Ownership is `SessionState.nameSource` (`placeholder` | `auto` | `manual`; the `name` setter / `PUT /api/sessions/:id/name` makes it `manual`, never touched again). ⚠️ `applyAutoName()` flips to `auto` even if the string is unchanged, so only the FIRST titled prompt names the tab. ⚠️ Only user input counts: `SessionWriteOptions.fromUser` is set by the browser WS path and `POST /api/sessions/:id/input` ONLY; any new user-input path must set it (and the send-key Shift+Enter path must call `trackUserInput()`). ⚠️ The pure tracker (`session-auto-name.ts`) sits on the raw keystroke stream with an explicit rule per key; add a rule for any new key class. ⚠️ `nameSource` also decides `--name`: only a `manual` name is pinned on the claude CLI (`Session.cliPinnedName`), since `--name` is also the `/resume` title; a rename appends a `custom-title` row to a LOCAL, non-docker transcript, and a same-name PUT is a no-op (never flips to `manual`). Tests: `test/session-auto-name.test.ts`. → [architecture-invariants#auto-named-sessions-first-prompt--tab-title](docs/architecture-invariants.md#auto-named-sessions-first-prompt--tab-title)
 
@@ -258,13 +188,13 @@ Codeman is a Claude Code session manager with web interface and autonomous Ralph
 
 **Unified session list**: `GET /api/sessions/unified` merges live sessions, persisted state, lifecycle-log history and transcript files into one deduped list (pure core `src/services/unified-session-service.ts`), backing the Cmd+K Session Manager, pinning and cross-device tab order (`PUT /api/session-order`, `src/session-order.ts`). ⚠️ Transcript history is THREE stores (`~/.claude/projects`, `~/.omp/agent/sessions`, `~/.codex/sessions`), folded via the `claudeSessionId → Codeman id` alias map (not Claude-only despite the name). ⚠️ `resumeId` is set by a SCANNER row only, never a live session; every surface that re-projects these rows (phone overview included) must carry it through, or a tap silently starts a second conversation. → [architecture-invariants#unified-session-list-and-session-manager](docs/architecture-invariants.md#unified-session-list-and-session-manager)
 
-**Owner tab layouts** (`tab-layout*.ts` + `GET`/`PUT /api/tab-layout`): named tab GROUPS over the flat strip, scoped per owner (`@single` when multi-user is off), persisted as `tabLayouts` in state.json. The frontend reads AND edits it (`tab-layout-browser.js` + the grouped-rail block in app.js): the vertical rail draws the owner's groups as collapsible sections (collapse is per-device localStorage), and with no groups or a failed read the rail is the flat list. Groups are created, renamed, reordered and deleted, and rows moved between them, from the row/group menus (Shift+F10 on a header too) and by pointer drag in the grouped rail. ⚠️ Grouping is a render layer only: `sessionOrder`, Alt+N and every other order consumer still read the server-projected session order, and a grouped row's markup is the flat row's markup. ⚠️ Only the GROUPED rail is an ARIA tree (`role=tree`, headers owning `role=group`s, one roving `tabindex=0`); the strip, sidebar and flat rail stay `tablist`/`tab`. ⚠️ Every browser write is a named operation through ONE serialized `PUT /api/tab-layout` at a time (`createEditCoordinator`): a 409 replays the operations onto the server's layout and retries (bounded), and an SSE reload is deferred while a write is in flight. Never PUT the layout from anywhere else in the frontend (the `pagehide` keepalive in `_persistPendingTabLayoutEdits` is the one deliberate exception). ⚠️ `TabLayoutService` is the single mutation boundary (one completed server action = at most one versioned write); never write layout state from a route or manager directly. ⚠️ The layout PROJECTS onto `PUT /api/session-order` via `tab-layout-legacy-order.ts`; change both sides together. ⚠️ Reconciliation is gated on a SUCCESSFUL restore (`markRestorationComplete`/`assertDeletionReady()`): a failed restore must leave the layout untouched or live tabs get pruned. → [architecture-invariants#owner-tab-layouts](docs/architecture-invariants.md#owner-tab-layouts)
+**Owner tab layouts** (`tab-layout*.ts` + `GET`/`PUT /api/tab-layout`): named tab GROUPS over the flat strip, scoped per owner (`@single` when multi-user is off), persisted as `tabLayouts` in state.json. The frontend reads AND edits it (`tab-layout-browser.js` + the grouped-rail block in app.js): the vertical rail draws the owner's groups as collapsible sections (collapse is per-device localStorage), and with no groups or a failed read the rail is the flat list. Groups are created, renamed, reordered and deleted, and rows moved between them, from the row/group menus (Shift+F10 on a header too) and by pointer drag in the grouped rail. A ref may carry `focus: true` ("Add to Focus" in the row/web-tab menu, the named `setFocus` operation): the rail then draws a SHORTCUT to it in a pinned, collapsible Focus section (`_renderTabFocus()`), its own element in `#tabRail` ABOVE `#sessionTabs`, so it shows on the flat rail too and is never a row, tree item, Alt+N slot, lineage anchor or drag target. ⚠️ Every copy of a ref (`copyRef` in tab-layout-browser.js, `parseRef`/`mapRef` in tab-layout.ts) must carry `focus` like `placement`: the browser PUTs whole layouts, so a dropped field is erased on the next save by whoever saves. ⚠️ Grouping is a render layer only: `sessionOrder`, Alt+N and every other order consumer still read the server-projected session order, and a grouped row's markup is the flat row's markup. ⚠️ Only the GROUPED rail is an ARIA tree (`role=tree`, headers owning `role=group`s, one roving `tabindex=0`); the strip, sidebar and flat rail stay `tablist`/`tab`. ⚠️ Every browser write is a named operation through ONE serialized `PUT /api/tab-layout` at a time (`createEditCoordinator`): a 409 replays the operations onto the server's layout and retries (bounded), and an SSE reload is deferred while a write is in flight. Never PUT the layout from anywhere else in the frontend (the `pagehide` keepalive in `_persistPendingTabLayoutEdits` is the one deliberate exception). ⚠️ Placing a NEW session in a group rides its create request (`tabGroupId` on `POST /api/sessions` / `/api/quick-start`) into `TabLayoutService.sessionCreated()`, which registers the session inside the owner lock and places it in the same write; never place a created session with a follow-up PUT. A spawned child lands after its parent; whether it keeps FOLLOWING it until moved by hand is the synced `spawnedTabsFollowParent` setting (default OFF: pinned `manual` in the creating write), read fresh in `sessionCreated()`. ⚠️ A legacy order PUT marks a child `manual` ONLY when the requested order splits it from its parent's block (`applyLegacySessionRank`), and the browser never echoes create/delete through `PUT /api/session-order`: the server already re-projected and broadcast it. Web-tab OPENNESS stays per device. ⚠️ `TabLayoutService` is the single mutation boundary (one completed server action = at most one versioned write); never write layout state from a route or manager directly. ⚠️ The layout PROJECTS onto `PUT /api/session-order` via `tab-layout-legacy-order.ts`; change both sides together. ⚠️ Reconciliation is gated on a SUCCESSFUL restore (`markRestorationComplete`/`assertDeletionReady()`): a failed restore must leave the layout untouched or live tabs get pruned. → [architecture-invariants#owner-tab-layouts](docs/architecture-invariants.md#owner-tab-layouts)
 
 **Hook events**: Claude Code hooks trigger via `/api/hook-event` (`permission_prompt`, `elicitation_dialog`, `elicitation_complete`, `elicitation_response`, `idle_prompt`, `stop`, `teammate_idle`, `task_completed`, `prompt_submitted`); see `src/hooks-config.ts` and `docs/claude-code-hooks-reference.md`. ⚠️ Every claude session installs the hooks block into its workspace (add-only merge) from every create path and from `restoreMuxSessions()`, gated by `workspaceHooksEnabled` (SYNCED, default ON). ⚠️ Route that decision through `applyWorkspaceHooks`, never call `ensureCodemanHooks` at a new site, or the setting silently stops applying. ⚠️ An AskUserQuestion / plan-selection dialog arrives as `permission_prompt` (RED alert), not `elicitation_dialog` (MCP elicitation). → [architecture-invariants#hook-events-and-workspace-hook-installation](docs/architecture-invariants.md#hook-events-and-workspace-hook-installation)
 
 **Reboot restore** (`src/reboot-restore.ts` pure + `web/reboot-restore-registry.ts` + `routes/reboot-restore-routes.ts` + `reboot-restore-ui.js`): after a host reboot kills every pane, Codeman holds an IN-MEMORY plan of the destroyed sessions and a banner offers to rebuild them. ⚠️ The heuristic only decides whether to ASK. ⚠️ Rebuild is TAKE-then-build (entries leave the plan before the first `await`, single-flighted per owner). ⚠️ Re-check grant, workspace and already-live at click time, reading already-live FRESH per entry; key confinement on the entry's OWNER, never the caller. ⚠️ Rebuilt sessions come back disarmed: no respawn/Ralph, `rearmAutoResumeSchedule: false`, and pass `nameSource` through. ⚠️ Undo a failed rebuild with `discardPartiallyBuiltSession()`, NEVER `cleanupSession()`. Claude-mode only, never remote/docker. Tests: `test/reboot-restore.test.ts`. → [architecture-invariants#reboot-restore](docs/architecture-invariants.md#reboot-restore)
 
-**Approvals Inbox** (`approvalsInboxEnabled`, SYNCED, default OFF; the store and answer endpoints run regardless): `web/approval-inbox.ts` is an in-memory, claude-only queue fed by `/api/hook-event`, at most ONE item per session, answered via `POST /api/approvals/:id/answer` through `writeViaMux` (menu answers never carry `\r`). ⚠️ Accept `option` digits ONLY if they match options parsed from a fresh RE-CAPTURE of the pane; a dialog no longer on screen is a 409. ⚠️ Resolve permission/question items only via the pane-verified `verifyStillAnswerable()`; the heuristic `working` signal alone may resolve `idle` items only. ⚠️ `applyCapture()` is ADD-ONLY for `options`. ⚠️ Viewing ACKNOWLEDGES an idle item (never resolves it) and only a human selection does: app-made selections pass `selectSession(id, { auto: true })`; `_ackDelivery` spends the IDLE alert only. ⚠️ `handleInit` seeds tab alerts from `GET /api/approvals` REGARDLESS of the setting. → [architecture-invariants#approvals-inbox](docs/architecture-invariants.md#approvals-inbox)
+**Approvals Inbox** (`approvalsInboxEnabled`, SYNCED, default OFF; the store and answer endpoints run regardless): `web/approval-inbox.ts` is an in-memory queue fed by `/api/hook-event` (answerable items claude-only; DeepSeek's and codex's screen-read ones are alert-only `answerInTerminal`), at most ONE item per session, answered via `POST /api/approvals/:id/answer` through `writeViaMux` (menu answers never carry `\r`). ⚠️ Accept `option` digits ONLY if they match options parsed from a fresh RE-CAPTURE of the pane; a dialog no longer on screen is a 409. ⚠️ Resolve permission/question items only via the pane-verified `verifyStillAnswerable()`; the heuristic `working` signal alone may resolve `idle` items only. ⚠️ `applyCapture()` is ADD-ONLY for `options`. ⚠️ Viewing ACKNOWLEDGES an idle item (never resolves it) and only a human selection does: app-made selections pass `selectSession(id, { auto: true })`; `_ackDelivery` spends the IDLE alert only. ⚠️ `handleInit` seeds tab alerts from `GET /api/approvals` REGARDLESS of the setting. → [architecture-invariants#approvals-inbox](docs/architecture-invariants.md#approvals-inbox)
 
 **Read My Mind intent profiles** (`readMyMindEnabled`, SYNCED, default OFF; `docs/readmymind-plan.md`): per-CASE profiles (goals + recent prompts) keyed by owner + realpath(workingDir), captured from the transcript (`transcript:user_prompt`), never the input paths; the listener must stay inside `startTranscriptWatcher()`'s `if (!watcher)` block. Store `src/intent-store.ts` → `intents.json`, ⚠️ written 0600 tmp+rename and never fed to `/api/search` (prompts carry secrets). Routes in `readmymind-routes.ts` (ownership via `findSessionOrFail` WITH `req`); predictor = pure `readmymind-context.ts` + IO in `readmymind-collectors.ts` + `readmymind-predictor.ts`, claude-only, one in flight per session (409). ⚠️ Suggestions render via value/`textContent` ONLY and nothing auto-sends, ever. Frontend `readmymind-ui.js`. → [architecture-invariants#read-my-mind-intent-profiles](docs/architecture-invariants.md#read-my-mind-intent-profiles)
 
@@ -274,19 +204,9 @@ Codeman is a Claude Code session manager with web interface and autonomous Ralph
 
 **Circuit breakers**: the Ralph breaker prevents respawn thrashing (`CLOSED` → `HALF_OPEN` → `OPEN`; reset via `/api/sessions/:id/ralph-circuit-breaker/reset`). **Distinct: the PTY-exit breaker** (`session-pty-exit-breaker.ts`) trips after repeated rapid PTY exits and blocks auto-restarts. ⚠️ It resets ONLY via an explicit `{clearBreaker:true}` body on `POST /api/sessions/:id/interactive`; the frontend's auto-reattach in `selectSession()` sends no body and must never clear it. → [architecture-invariants#circuit-breakers-ralph--pty-exit](docs/architecture-invariants.md#circuit-breakers-ralph-and-pty-exit)
 
-**Full-scrollback replay**: `GET /api/sessions/:id/terminal?full=1` returns the whole tmux scrollback ALONE (`source='mux-full-history'`), superseding the byte buffer. First load of each non-shell TUI session requests it (`_fullHistoryLoaded`); Shell selection and drop recovery use a bounded 1 MiB `?tail=`, and a Shell scroll-to-top pulls a bounded `?full=1&tail=` window (a window no longer than the browser's buffer is skipped before the downgrade guard, so it never marks the session exhausted); the unbounded pull stays behind **Load full history**. A Shell split-pane Pane B has its own copy of the bounded pull against its own xterm (`TerminalTile._pullHistory`, terminal-tile.js); keep the two in step. → [architecture-invariants#split-pane-sessions](docs/architecture-invariants.md#split-pane-sessions) ⚠️ The capture ends with a RELATIVE cursor move back to the pane's caret (never `CUP`), so no line-deleting transform may run over it; those skips key on `isFullCapture`, never on `?full=1` alone. ⚠️ A re-pull must never shrink the buffer (`_replayWouldShrinkBuffer()`). ⚠️ `captureCols`/`captureRows` are absent when no frame was positioned: test `Number.isFinite`, never truthiness. ⚠️ A frame dropped at the 128 KiB render cap MUST be recovered, and the recovery verifies itself: `_scheduleDroppedOutputRecovery` re-arms (bounded by `DROP_RECOVERY_MAX_ATTEMPTS`) while `_onSessionNeedsRefresh` reports no repaint, but never after a capture-fetch `'deadline'`. → [architecture-invariants#full-scrollback-replay](docs/architecture-invariants.md#full-scrollback-replay)
+**Full-scrollback replay**: `GET /api/sessions/:id/terminal?full=1` returns the whole tmux scrollback ALONE (`source='mux-full-history'`), superseding the byte buffer. First load of each non-shell TUI session requests it (`_fullHistoryLoaded`); Shell selection, drop recovery, clear and a tile's load use `SHELL_LOAD_QUERY` (constants.js), a bounded window of tmux's RENDERED history (`?full=1&tail=1MiB&alt=recording&lines=10000`), never the raw byte recording, whose full-screen repaints replayed at another pane height push copies of their top row into scrollback; the route reports a client `lines=` bound below the history ceiling that left lines in tmux as `truncationReason: 'tail'`, and `alt=recording` falls back to the visible frame over the tailed byte recording when the pane is in the alternate screen (a nested tmux in a remote or Docker case keeps no tmux history; vim/less take it too while open). A Shell scroll-to-top pulls a bounded `?full=1&tail=` window (a window no longer than the browser's buffer is skipped before the downgrade guard, so it never marks the session exhausted); the unbounded pull stays behind **Load full history**. A Shell split-pane Pane B has its own copy of the bounded pull against its own xterm (`TerminalTile._pullHistory`, terminal-tile.js); keep the two in step. → [architecture-invariants#split-pane-sessions](docs/architecture-invariants.md#split-pane-sessions) ⚠️ The capture ends with a RELATIVE cursor move back to the pane's caret (never `CUP`), so no line-deleting transform may run over it; those skips key on `isFullCapture`, never on `?full=1` alone. ⚠️ A re-pull must never shrink the buffer (`_replayWouldShrinkBuffer()`). ⚠️ `captureCols`/`captureRows` are absent when no frame was positioned: test `Number.isFinite`, never truthiness. ⚠️ A frame dropped at the 128 KiB render cap MUST be recovered, and the recovery verifies itself: `_scheduleDroppedOutputRecovery` re-arms (bounded by `DROP_RECOVERY_MAX_ATTEMPTS`) while `_onSessionNeedsRefresh` reports no repaint, but never after a capture-fetch `'deadline'`. → [architecture-invariants#full-scrollback-replay](docs/architecture-invariants.md#full-scrollback-replay)
 
-**Split-pane sessions** (`showSplitButton`, header button, default OFF, desktop-only at 1180px (width alone, so a wide Android tablet clears it), per-device): a second live session ("Pane B") beside the active one, in a `TerminalTile` (terminal-tile.js; the picker, divider and auto-collapse stay in terminal-split.js) with its own xterm + WebSocket, resizable via a draggable divider. Pane B reconnects after a drop, sends input through the exactly-once queue over its own socket (`_registerInputSocket`), has clickable paths and image paste, and owns its geometry (no 40x10 floor, `zc` columns adopted, font changes call `tile.fit()`). ⚠️ Only typed input enters that persisted queue: xterm's query replies are dropped and focus/mouse reports go out ephemeral. ⚠️ App-level terminal actions find their pane through `_focusedPane()` (the terminal focused last), never `this.terminal`. Still plainer than the primary pane (no local-echo overlay, CJK IME textarea or touch handlers), though it wires the same keyCode-229 soft-keyboard controller (Android autocorrect, #441/#541), and NOT persisted across reloads. The tile grid (below) reuses `TerminalTile`, and the two are never open together. → [architecture-invariants#split-pane-sessions](docs/architecture-invariants.md#split-pane-sessions)
-
-**Tile grid** (`showTileGridButton`, header Tiles button + `Ctrl+Shift+G`, default ON on desktop and OFF on handhelds and touch-primary tablets (primary pointer coarse, `getDefaultSettings()`; a touchscreen laptop keeps ON), desktop-only at 1180px, per-device; tile-grid.js, design `docs/tile-grid-plan.md`): 1 to 6 live sessions side by side (the cap is ONE constant, `TILE_GRID_MAX` in constants.js, owner decision; the layout table still covers 7 to 9, unreachable), each a `TerminalTile` with a header (state dot, harness logo, name, model, menu, zoom, ×; no +, owner decision), laid out by count (`CodemanTileGrid`, constants.js) with draggable column/row dividers, zoom (tmux-style), an Attach overlay, and per-device persistence (`codeman:tile-grid`, ids only, restored INSIDE `handleInit` in place of the single-view select, so the main terminal never loads on that page load). While the grid is open the main terminal is PARKED: `activeSessionId` is the focused tile's session, and every main-terminal path that would write, fetch, resize or reconnect stands aside through `_tilesOwnTerminal()` (the WebGL long-task observer included). ⚠️ Every capture a tile fetches (initial, reconnect refresh, `{t:'r'}`, history pull) goes through ONE `TileLoadQueue` (concurrency 1, a deadline covering the body), because each is a synchronous tmux call on the server. ⚠️ Only a USER-initiated pick of a non-tiled session (or `leaveTiles`, a followed link) leaves the grid; an `auto` selection never collapses it, and every app-driven fallback (close, delete, restore) picks a tile. A caller of `closeTileGrid({ reselect: false })` must null `activeSessionId` before any `_cleanupPreviousSession`, or the parked terminal's stale content is saved as a snapshot. ⚠️ The grid and the split are never open together. ⚠️ Tile chords go through `tileShortcutFor` and are swallowed in every xterm key handler BEFORE the Shift+Enter gate; the toggle follows `showTileGridButton` (owner decision: OFF makes the chord inert, it passes through like any unbound key). ⚠️ The Tiles button's click and Ctrl+Shift+G are ONE function, `toggleTileGrid` (owner decision): they open the grid AT ONCE with the remembered count (`codeman:tile-count`, default 6, at most `_tileGridLimit()`) of `tileGridOpenSet` (constants.js: the stored grid, else an open split's two, else the tabs in order, the active one focused; trimmed or filled by `tileGridSetForCount`, a stored grid re-formed into its cells by `reformTileCells`), never a menu; right-click is the 2 / 4 / 6 count menu (`openTileCountMenu`, owner decision 10), which owns its Escape. The button has no native title: its hover card (`_installTileGridHint`, `#tileGridHint`) says it, is its `aria-describedby` (always present, hidden, kept current) and hides in the capture phase on any press, click or right-click, so the menu never opens beside it. ⚠️ The open, the count menu and the toggle's close animate opacity and transform only (the close leaves an inert cloned still copy until the single view's selection settles, at most 700 ms); nothing moves under `prefers-reduced-motion`. ⚠️ Opening builds one tile terminal per frame (`_connectTilesPaced`), so `openTileGrid` returns before the terminals exist: focus is handed over in `_connectTile` (`focusOnConnect`). ⚠️ A divider drag reflows locally per frame and sends ONE resize per affected tile at pointer-up. ⚠️ The grid is CELLS (owner: an empty cell can be any cell): `grid.cells` (id or `null`) is the one source of truth, `grid.ids` a derived getter, never written; the shape still comes from the tile count, a shape change goes through `fitTileCells`, and the stored `ids` carry the cells with `null` holes. ⚠️ A tile moves (its header dragged or its tab dropped onto another tile or an empty cell, `Ctrl+Shift+Arrows`) ONLY through `_reorderTiles`: never a remount, reconnect or reload, and since sizes belong to the cells only a tile whose cell size changed fits; off while a tile is zoomed. The header drag carries its own type, never text, and is not `draggedTabId`; the header focuses on click, never on press, so a cancelled drag changes nothing. The arrow chords skip text fields. ⚠️ Sessions Run from this tab join the open grid (`_joinTileGridFromRun`, called from `_ensureCreatedSessionVisible`); sessions created elsewhere never do. Such a tile connects before its pane exists, the server drops that resize and spawns at 120x40, and Run's own resize measures the parked terminal (nothing), so the chrome refresh calls `tile.paneStarted()` when the session's pid appears or changes. ⚠️ An agent that exited in a live pane (`paneExit`) cannot be re-attached in place (both attach routes refuse while the pane's tmux client runs, and say so in a 200 envelope): its tile shows the exit and points at Close session. ⚠️ Each header (a tile's, both split panes': Pane A gets one only while the split is open, fitted through `sendResize`/`syncTermLine truncated
-
-**Terminal touch gestures: link taps and text selection**: on touch devices xterm's linkifier and SelectionService never see the gesture, so both are driven explicitly (terminal-ui.js). ⚠️ A tap activates the link under it through the SAME provider as the hover linkifier (`_terminalLinkAtPoint`), synchronously inside `touchend` (keeps the user gesture `window.open` needs) and BEFORE any mouse report; the caret's logical line (`_tapIsOnCaretLine`) and TUI-owned rows (`_isActionableMobileTerminalTap`) keep their meaning. ⚠️ Gate on the caret line, never on tap intent (a shell calls every tap `'input'`). ⚠️ Long-press selects via xterm's public `select()`; keep the three guards: suppress the compat mouse pair after `touchend`, the bounded focus guard + `contextmenu` suppression for the platform long-press, and no closing `terminal.focus()` on phones. Tests: `test/terminal-touch-tap.test.ts`. → [architecture-invariants#terminal-touch-gestures-link-taps-and-text-selection](docs/architecture-invariants.md#terminal-touch-gestures-link-taps-and-text-selection)
-
-**Auto Copy (copy-on-select)** (`autoCopySelection`, per-device, default OFF): a finished terminal selection lands on the clipboard with no keystroke. ⚠️ Copy at the END of a gesture, never in `onSelectionChange` (per-cell); it only arms `_autoCopyPending` and a document-level `mouseup` flushes. ⚠️ The flush must be SYNCHRONOUS in the handler (both clipboard paths need user activation); never defer it to a timer. ⚠️ Touch needs its own calls from `_endTouchSelectionGesture()`/`_selectTouchSelectionLine()` (no mouseup arrives). ⚠️ Unlike `copyTerminalSelection()`, never clear the selection or focus the terminal; restore prior focus. Guards are pure in `decideAutoCopy()` (constants.js, 1M-char cap, refused not truncated). Tests: `test/terminal-auto-copy.test.ts`. → [architecture-invariants#auto-copy-copy-on-select](docs/architecture-invariants.md#auto-copy-copy-on-select)
-
-**Ctrl+V paste trap** (`image-input.js`): `Ctrl+V` routes through `_handleImagePaste()`, which focuses a hidden `contenteditable` trap and reads the clipboard from the paste event landing there; images upload and their paths are typed in, text goes through `terminal.paste()` so bracketed-paste markers survive. ⚠️ **The trap must consume exactly ONE paste event** (Firefox delivers two per keypress: the `execCommand('paste')` event and the keydown's default action); the one-shot flag lives on the trap, never on a browser check. ⚠️ Do not remove the `execCommand('paste')` call: on some mobile engines it is the only route into the trap, and the trap is the only place image blobs are read. Tests: `test/image-paste-trap.test.ts`. → [architecture-invariants#terminal-paste-ctrlv](docs/architecture-invariants.md#terminal-paste-ctrlv)
-
-**Terminal scrollback strip + wheel/touch forwarding**: codex/claude/gemini get the FULL strip (alt-screen, `3J`, mouse DECSETs); tmux-backed opencode gets the MIDDLE strip (alt-screen toggles + mouse DECSETs, `3J` kept); every other tmux-backed mode (shell/antigravity/pi/grok/deepseek/omp) gets the NARROW strip (alt-screen toggles only). Table: `CliCapabilities.altScreen` JSDoc, pinned in test/claude-scrollback-strip.test.ts. ⚠️ Gated on `useMux`: direct-PTY sessions must keep the alt screen. Wheel and touch forward to the CLI for **claude ≥ 2.1.187 ONLY, and only while it has mouse tracking on** (`cliMouseTracking`: fullscreen claude sets it, its default inline renderer does not and scrolls locally like codex); ⚠️ never re-add codex without a fresh measurement (it ignores SGR wheel reports). ⚠️ `getClaudeCliVersion()` must never cache a FAILED probe. ⚠️ Hand-report clicks only while the CLI has mouse tracking on: `_shouldReportMouseToCli()` gates all three report sites on `cliMouseTracking` (from `_recordStrippedMouseMode()`, session.ts), or a plain shell prints the reports as literal text. ⚠️ A `TerminalTile` (grid tile, split Pane B) pages the wheel and reports clicks through these SAME gates with itself as the target (its terminal, its session, never `activeSessionId`); never copy the mode lists into terminal-tile.js. Read `_logScrollRouting()` before diagnosing a scroll report. → [architecture-invariants#terminal-scrollback-strip-flavors-and-wheeltouch-forwarding](docs/architecture-invariants.md#terminal-scrollback-strip-flavors-and-wheeltouch-forwarding)
+**Terminal scrollback strip + wheel/touch forwarding**: codex/claude/gemini get the FULL strip (alt-screen, `3J`, mouse DECSETs); tmux-backed opencode gets the MIDDLE strip (alt-screen toggles + mouse DECSETs, `3J` kept); every other tmux-backed mode (shell/antigravity/pi/grok/deepseek/omp) gets the NARROW strip (alt-screen toggles only). Table: `CliCapabilities.altScreen` JSDoc, pinned in test/claude-scrollback-strip.test.ts. ⚠️ Gated on `useMux`: direct-PTY sessions must keep the alt screen. Wheel and touch forward to the CLI for **claude ≥ 2.1.187 ONLY, and only while it has mouse tracking on** (`cliMouseTracking`: fullscreen claude sets it, its default inline renderer does not and scrolls locally like codex); ⚠️ never re-add codex without a fresh measurement (it ignores SGR wheel reports). ⚠️ `getClaudeCliVersion()` must never cache a FAILED probe. ⚠️ Hand-report clicks only while the CLI has mouse tracking on: `_shouldReportMouseToCli()` gates all three report sites on `cliMouseTracking` (from `_recordStrippedMouseMode()`, session.ts), or a plain shell prints the reports as literal text. ⚠️ A `TerminalTile` (grid tile, split Pane B) forwards the wheel as SGR reports, pages it and reports clicks through these SAME gates with itself as the target (its terminal, its session, never `activeSessionId`), and scrolls Shift+wheel itself (xterm makes it a horizontal no-op); never copy the mode lists into terminal-tile.js. Read `_logScrollRouting()` before diagnosing a scroll report. → [architecture-invariants#terminal-scrollback-strip-flavors-and-wheeltouch-forwarding](docs/architecture-invariants.md#terminal-scrollback-strip-flavors-and-wheeltouch-forwarding)
 **Detached start + service install**: `codeman web -d` relaunches the same entry script `detached:true` (setsid); `nohup` is not what makes it survive. ⚠️ Both `-d` and `service install` must REFUSE when a server is already up on this data dir (pidfile + `/api/status` probe), or a second instance attaches to the first one's live sessions. ⚠️ Never report success not observed: poll `/api/status` until the child answers or dies. `--stop` must verify the pid still looks like Codeman (`ps -o command=`) before signalling. Unit/label names live only in `config/service-names.ts`. `service install` bakes the installing shell's PATH into the unit and never writes `CODEMAN_PASSWORD` into it. → [architecture-invariants#detached-start-and-service-install](docs/architecture-invariants.md#detached-start-and-service-install)
 
 **Self-update** (App Settings → System → Updates): in-app updater for git-clone installs under a supervisor (`systemd`, `launchd`, `launchd-daemon`, `docker-compose`, else `none`). The work runs in a DETACHED `scripts/self-update.sh` writing `update-status.json`, polled across the restart; pure helpers in `src/web/self-update.ts`. ⚠️ Compose: the restart kills the script, so nothing may be appended after the `restarting` marker; the repo must stay a host bind mount over `/opt/codeman` and the image must keep devDependencies + toolchain. ⚠️ `evaluateEnvironmentGate()` refuses releases that change `server.Dockerfile`/`docker-compose.yaml` or add `.env.example` keys, re-evaluated on `POST /api/system/update`; unknowns fail OPEN, but the exit-to-restart needs `--restart-by-exit 1` (`CODEMAN_RESTART_BY_EXIT=1` only in the Compose file). ⚠️ Keep the agent CLIs in `server.Dockerfile` pinned. → [docs/docker-self-update.md](docs/docker-self-update.md), [architecture-invariants#self-update](docs/architecture-invariants.md#self-update)
@@ -295,29 +215,31 @@ Codeman is a Claude Code session manager with web interface and autonomous Ralph
 
 **Attachments** (live external document references; all wiring in `file-routes.ts`): a **registry** maps a stable `attachmentId` to a realpath-resolved, extension-allowlisted absolute path, so browser requests never carry arbitrary absolute paths. ⚠️ The **magic-link scanner** (`codeman://attach?...` in terminal output) is **prompt-injectable**, so its scan path is force-confined to the session workspace; a hostile prompt could otherwise exfiltrate arbitrary host files over SSE. The security gate is an extension **allowlist**, not a blocklist. `document-conversion-limiter.ts` caps converter spawns globally: without it, N large docs detected at once fork N multi-minute processes, which is a resource-exhaustion vector. → [architecture-invariants#attachments](docs/architecture-invariants.md#attachments)
 
-**File-path links (terminal + chat)**: a path an agent prints is clickable on BOTH surfaces and opens the file-preview overlay. ⚠️ ONE pattern (`FILE_PATH_LINK_PATTERN` / `absoluteFilePathPattern()` in constants.js) feeds the xterm link provider AND `_linkifyFilePaths()`, a fresh instance per call (`lastIndex`). The chat linkifier walks TEXT NODES with DOM APIs, never rebuilds sanitized markup as a string. ⚠️ An out-of-workspace path goes through the ATTACHMENT routes (`POST /api/sessions/:id/attachments` with `notify: false`), never by widening `file-content`/`file-raw` or `file-stream-manager`'s `tail -f` allowlist. ⚠️ `TEXT_ATTACHMENT_EXTENSIONS` IS `EDITABLE_EXTENSIONS` (never a second list), and widening READ must never widen RUN: `html`/`htm`/`svg` stay download-only, other text is inert `text/plain`+`nosniff`. Media extensions are single-sourced in `attachment-registry.ts`. ⚠️ **XLSX previews parse in the BROWSER**, never on the server: `spreadsheet-preview.js` fetches the raw route with `?preview=true` (413 above `MAX_XLSX_BROWSER_PREVIEW_BYTES`, 10 MB) and hands the bytes to `spreadsheet-preview-worker.js`, the only place the pinned `exceljs`/`fflate` vendor bundles load (never on page load). `admitXlsx()` caps the ZIP before ExcelJS runs, counting what ExcelJS will EXPAND as well as what it reads (a merge costs its area, a `<col>` past 16384 is refused, validations are never parsed via `ignoreNodes`, and defined names are never expanded: the worker stubs `_definedNames.model`). ⚠️ The INDEX a row or sheet claims is bounded too, since ExcelJS allocates and walks up to it: a `<row r>` outside 1-1048576 and an `xl/workbook.xml` `<sheet sheetId>` above `LIMITS.maxSheetId` (65535) are refused, and `sendTile` reuses the merges read at load (`mergesById`), never `sheet.model`, which rebuilds the whole sheet. ⚠️ The COLUMN index costs the same way (a row's cells live at `_cells[col - 1]`, so one `XFD` cell makes ExcelJS's `eachRow`/`eachCell`/`hasValues` visit 16,384 slots per row): `worksheetMetadata` builds each sheet's row and cell index from the keys that exist (`Object.keys` of `_rows` and `_cells`, skipping falsy and `Null`-type cells as `eachCell({ includeEmpty: false })` does), reads row heights in that pass and merges from `sheet._merges`, and `sendTile` reads cells from that index (`populatedRowsById`); never call ExcelJS's dense `eachRow`/`eachCell` there. `parseThemePalette` returns the default palette for a theme above 64 KB (64 * 1024 characters), since its patterns are quadratic on unclosed tags. ⚠️ Merges are capped per sheet (`LIMITS.maxMergesPerSheet`, 2,000) AND workbook-wide (`LIMITS.maxMerges`, 10,000), refused as `merge-limit` before ExcelJS loads, since ExcelJS's `_mergeCellsInternal` checks each merge against every earlier one on its sheet (quadratic). Every `<numFmt>` in `xl/styles.xml` is read with `readTagAttributes()` and refused (`number-format`) when its decoded `formatCode` is over 255 characters or has a `[` after its last `]`: ExcelJS's `isDateFmt` rescans to the end of the code per unclosed `[` once per numeric cell, and the code is echoed into the notice bar. ⚠️ `formatCellValue` caps every cell's display text at `LIMITS.maxCellTextChars` (1,000, ellipsis, never splitting a surrogate pair) for every value shape (rich text, hyperlink text, formula source, errors), since structured clone copies each tile cell's whole string to the page and the load timeout no longer covers tiles. `richTextPrefix` also visits at most `maxCellTextChars + 1` runs, not only keeps that much text: an empty `<r/>` adds nothing, so a shared string of a million empty runs was walked whole per cell, per tile (56 s a tile). ⚠️ Every part ExcelJS parses (all but the pure-bytes `xl/media/<name>.<ext>`) is also charged per START TAG against `LIMITS.maxElements` (2,000,000, refused as `element-limit`): ExcelJS builds an object per element in sharedStrings, styles, comments, drawings, VML and tables too, and empty stored blocks pad a deflate stream past the ratio cap, so 8.3M empty `<r/>` runs in a 375 KB file cost 570 MB of heap. ⚠️ Admission keys every entry on the name ExcelJS will SEE (`excelJsEntryName()`: JSZip's `.`/`..`/empty-segment resolution, then one leading `/` stripped), refuses two entries that land on one name, and treats anything matching ExcelJS's UNANCHORED `xl/worksheets/sheet<N>.xml` as a worksheet, so `/xl/...` or `xl/./...` cannot skip a counter, and ExcelJS parses only a STORE-only archive rebuilt from the entries admission inflated (`buildAdmittedArchive()`), never the fetched bytes; cell text goes through `textContent`, formulas are never evaluated. Bumping either package or editing the worker/core changes `SPREADSHEET_ASSET_VERSION`, which `npm run check:public-assets` pins. xls/ods stay download-only. → [architecture-invariants#file-path-links-terminal--response-viewer](docs/architecture-invariants.md#file-path-links-terminal--responseLine truncated
+**Prompt uploads** (`POST /api/sessions/:id/paste-image`: clipboard paste, drop, the Compose Attach key and the phone toolbar's Attach button; client wiring in `image-input.js` and `keyboard-accessory.js`): a pick is written to `<workspace>/.codeman-uploads/` and its path lands on the prompt as unsent text. The folder is IN the workspace because that is the only path identical for a local and a container agent, FLAT because `<workspace>/.codeman/` would BE the data dir when the workspace is the home directory, and self-ignoring through a `.gitignore` of `*` written ONCE with `wx` (a file already there is the user's and is never overwritten; the user's own ignore file is never touched). ⚠️ The directory names live ONLY in `paste-image-gc.ts` (`UPLOAD_DIR_NAMES`): the pre-move `.claude-images/` receives nothing but stays readable for one release, so the GC sweep and the delete cleanup (through `uploadDirs()`) and the image watcher's ignore filter (reading the names) cover BOTH; retire the legacy entry there, never by hand elsewhere. ⚠️ `uploadDirs()` is async and probes the workspace BOUNDED first (`probePathKind()`, the #516 rule: an `unknown` path is skipped, never touched; the sweep keeps the stall cap and the delete passes `pastCap`), then returns only REAL directories (lstat, so a planted symlink is never one), none that is or contains `getDataDir()` (one strictly inside it only holds uploads and is listed), and none for a remote session, since both consumers delete; the check is made at list time, and the same-user race after it is accepted. Images are buffered in memory under `MAX_PASTE_IMAGE_BYTES` (50MB) after a client-side canvas re-encode; a DOCUMENT (pdf, docx, pptx, xlsx, zip and the text family txt/md/csv/tsv/json/log/yaml/yml/xml; the curated `ALLOWED_DOCUMENT_EXTS`, never `SUPPORTED_ATTACHMENT_EXTENSIONS` or `EDITABLE_EXTENSIONS`, which carry sh/py/js/html) is admitted by its name's extension alone, buffered like an image, judged on its bytes by the same fail-closed `imageMagicMatchesExt` (`%PDF-`, `PK\x03\x04` for the OOXML trio and zip, UTF-8 without a NUL byte for text: a SIGNATURE or ENCODING check, never a structure or content check, so generic zip bytes pass as `.docx` and a script renamed `.txt` is text; the sh/py/js/html exclusion is a SCOPE choice by extension, keeping the server from storing what it would serve or run, not a content filter: the agent can already write any of those) and saved as `paste-<ts>-<hex>-<stem><ext>` with a sanitized stem (`[A-Za-z0-9._-]`, 40 chars, a wholly non-ASCII name leaves no stem; the `paste-` prefix is what the GC matches). A VIDEO is never buffered: `?kind=video` streams it to disk under `MAX_PASTE_VIDEO_BYTES` (1GB, env `CODEMAN_MAX_PASTE_VIDEO_BYTES`), judges the container on the first 12 bytes as they arrive (ISO BMFF / QuickTime atom, EBML for webm, Ogg for ogv; a mislabeled body is drained unwritten from there, never parked on disk first) and unlinks on any failure, a dead transfer included. ⚠️ The flag is the ONLY way onto that path and admits video types alone (the registry's `VIDEO_ATTACHMENT_EXTENSIONS`, single-sourced: mp4/webm/mov/m4v/ogv, so an uploaded video is one the file viewer can also play; 3gp/mkv are out, since a phone's camera roll produces neither and neither previews), so an image can never ride the larger cap through memory; without it the route is byte-identical to the image-only version. Busboy takes its size limit before the part is seen, which is why the choice rides the URL and not the file. ⚠️ A video draws on its OWN per-(IP, session) rate bucket (`consumeVideoToken`, 3/min) beside the image bucket (`consumePasteToken`, 30/min): at 1GB each, 30/min would be 30GB of disk a minute, so the lower rate bounds it to 3GB; the image path keeps the 30/min bucket and is unchanged without the flag. A remote (SSH) session is refused (400) before any disk touch: its `workingDir` is the remote path, so the file would land on THIS host where the agent cannot read it (Docker is fine, the workspace is bind-mounted at the same absolute path). ⚠️ A live session's workspace can still sit on a network mount that stopped answering, so the route probes it BOUNDED (`probePathKind(session.workingDir, { pastCap: true })`, the #516 rule) before the first disk touch and refuses an `unknown` mount through `describeUnknownPath()` (OPERATION_FAILED), the way session create does, so a mount already unreachable cannot freeze a libuv worker on every upload (a mount that dies after the probe is the same residual race create accepts). ⚠️ Dispatch mirrors the client: an `image/*` MIME takes the IMAGE path whatever the name says (a GIF named `animation.txt` stays a GIF), a document extension takes the document branch, and anything else takes the image path exactly as before, where the HEIC sniff still recovers a HEIC under a generic name; a foreign extension with no image MIME gets its `Unsupported file type` 400 only AFTER that sLine truncated
+
+**File-path links (terminal + chat)**: a path an agent prints is clickable on BOTH surfaces and opens the file-preview overlay. ⚠️ ONE pattern (`FILE_PATH_LINK_PATTERN` / `absoluteFilePathPattern()` in constants.js) feeds the xterm link provider AND `_linkifyFilePaths()`, a fresh instance per call (`lastIndex`). Both call `findFilePathLinks()`, which adds `~/`, relative and folder paths without overlapping a rooted match or a URL; a relative path resolves against `workingDir` (`resolveLinkedFilePath`), `~/` only server-side on the session's host (attachment route), and `terminalPathLine()` joins a path a program hard-wrapped short of the edge. ⚠️ A folder link opens the **folder grid**: `POST /api/sessions/:id/folders` guards + GRANTS the folder (`folder-listing.ts`, own registry, never the 200-capped attachment registry), files come by bare name via `…/folders/:folderId/files/:name/*` through `resolveServableAttachmentPath` plus the same extension/regular-file gate on the symlink's resolved target, so no browser request carries a path; `~/` is refused in a Docker case (the container's home is not this host's). The chat linkifier walks TEXT NODES with DOM APIs, never rebuilds sanitized markup as a string. ⚠️ An out-of-workspace path goes through the ATTACHMENT routes (`POST /api/sessions/:id/attachments` with `notify: false`), never by widening `file-content`/`file-raw` or `file-stream-manager`'s `tail -f` allowlist. ⚠️ `TEXT_ATTACHMENT_EXTENSIONS` IS `EDITABLE_EXTENSIONS` (never a second list), and widening READ must never widen RUN: `html`/`htm`/`svg` stay download-only, other text is inert `text/plain`+`nosniff`. ⚠️ The ONE place HTML renders is `/html-view/<cap>/...` (`html-view-routes.ts`, `html-view-capabilities.ts`): `POST /api/sessions/:id/html-view` mints a capability for the file's own directory, every response carries `Content-Security-Policy: sandbox` WITHOUT `allow-same-origin` (opaque origin even as a top-level tab, so the page can never read Codeman or drive its API) and with NO fetch directive (open network, owner decision 2026-10-13: never restrict it or add `allow-same-origin` without the owner), and only non-dot, asset-allowlisted files under that directory are served (GET/HEAD only; capabilities revoked with the web-tab ones on logout). That tree is served recursively, so minting is refused for a hidden directory or a broad root (`isBroadHtmlViewRoot`); the preview renders by default and the **Page** pill (or a refused mint) falls through to the source view and its Edit button. Never serve HTML same-origin anywhere else. Tests: `test/routes/html-view-routes.test.ts`. Media extensions are single-sourced in `attachment-registry.ts`. ⚠️ **XLSX previews parse in the BROWSER**, never on the server; the admission limits that keep a hostile workbook from freezing the worker live in `src/web/CLAUDE.md` (Spreadsheet preview). → [architecture-invariants#file-path-links-terminal--response-viewer](docs/architecture-invariants.md#file-path-links-terminal--response-viewer)
 
 **Filesystem path picker** (Link Existing "Browse" + the mobile keyboard's `📁 Path` key): lazy one-directory browsing via `GET /api/filesystem/browse`, with `GET /api/filesystem/preview` for the tapped file. Inserts the path **without** Enter, so the prompt is never submitted; the sibling `⌫ All` key clears only the unsent prompt and must never send the agent's `/clear`. ⚠️ This is a **second file-serving surface and inherits neither the attachment confinement nor its ownership scoping** — it allowlists Home, `CASES_DIR`, `/mnt/d` and `CODEMAN_FILE_PICKER_ROOTS`, blocks sensitive trees, and rejects symlink escapes **after** `realpath`. ⚠️ The optional `sessionId` is an ownership boundary that must be `canAccessOwned`-checked by hand (it does not go through `findSessionOrFail`), and in multi-user mode a non-admin gets only their own `userSpacePath` as a root: per-user spaces live INSIDE `homedir()`, so a `Home` root exposes every other user's workspace. Previews go through the same global conversion limiter, and Markdown/TXT/JSON are served as inert `text/plain`. → [architecture-invariants#filesystem-path-picker](docs/architecture-invariants.md#filesystem-path-picker)
 
 **File Viewer edit mode** (issue #212): the file-preview overlay edits workspace text files in place — `GET .../file-content?edit=1` + `PUT /api/sessions/:id/file-content`, policy in `src/config/file-editing.ts`. This is a **third file surface and the only one that WRITES**: read-path confinement (realpath + workspace + ownership) plus sensitive/blocked/`.git` denies and an extension **allowlist**; writes are `wx`-temp + rename (no `O_CREAT` anywhere = edit-in-place is structural); optimistic concurrency via sha256 `baseHash` → 409. ⚠️ `edit=1` never truncates and the client must never save a plain-preview buffer (the 500-line truncation would silently delete the rest). ⚠️ CRLF/UTF-8 guards: EOL re-applied server-side, non-UTF-8 refused via round-trip compare. → [architecture-invariants#file-viewer-edit-mode](docs/architecture-invariants.md#file-viewer-edit-mode), `docs/file-viewer-edit-plan.md`
 
-**File Viewer text view: rendered markdown + Lines/Wrap toggles** (`_renderFilePreviewText()` in panels-ui.js): a `.md`/`.markdown` opens RENDERED by default with an `MD` pill back to source; the plain-text view has `Lines` (CSS-counter gutter) and `Wrap` toggles. ⚠️ ONE markdown pipeline: the viewer calls `_renderMarkdown(text, { breaks: false })` (marked + the DOMPurify allowlist, the Response Viewer's; chat keeps the default `breaks: true`, a file must not turn every hard wrap into a `<br>`) and binds the Response Viewer's click delegate (`_bindResponseViewerInteractions`) on the preview body for code-copy buttons and path links; never a second parser or handler. ⚠️ The document is built inside a `<template>` (a detached div with `innerHTML` set starts fetching every `<img src>` before the rewrite), then `_rebaseFilePreviewMarkdownRefs()` points relative images at the workspace-confined `file-raw` under the document's directory and root-relative ones under the workspace root (never a widened route; a failed load degrades to alt text), after `decodeURIComponent`ing the ref and dropping `?query`/`#fragment` (marked percent-encodes destinations, and the route encodes again), and turns workspace links into `a.rv-path` carrying `data-session-id` for the delegate (the linkifier's absolute paths get it too), stripping the `target` marked gave them. ⚠️ A preview opened by attachment id under a bare file name (attachment cards, history drawer: the registry keeps no relative path) has no directory, so its workspace refs degrade (images to alt text, links to their text), never resolve against the workspace root. ⚠️ The container carries `data-i18n-skip` or the translator rewrites the document's prose. ⚠️ Toggles are per-device localStorage keys (`codeman:filePreview*`), never `SettingsUpdateSchema`; Lines/Wrap are class flips on the ONE `<pre>`, with rules scoped `.file-preview-body > pre.file-preview-text` so they never leak into the document's code blocks. Markdown fetches `lines=10000` (the route ceiling), other text keeps 500. ⚠️ `md` stays OUT of `FILE_PREVIEW_EXTENSIONS`: only an IN-WORKSPACE `.md` path clicked in the TERMINAL keeps the tail viewer (live follow); the Files panel, chat paths, out-of-workspace terminal paths and attachment cards all reach `openFilePreview()` and render it. Tests: `test/file-preview-markdown.test.ts`. → [architecture-invariants#file-viewer-text-view-rendered-markdown-and-text-toggles](docs/architecture-invariants.md#file-viewer-text-view-rendered-markdown-and-text-toggles)
-
 **Files panel search** (COD-236, the `q` param on `GET /api/sessions/:id/files`): `compileFileQuery()` (`utils/file-query.ts`, pure) compiles the query into a predicate the server-side walk prunes with; a query returns a FLAT match list and the walk recurses past non-matching directories. An empty, whitespace-only or overlong (`MAX_QUERY_LENGTH`, 256) query compiles to `null`, keeping the default tree response byte-identical. ⚠️ **Never compile a glob into a RegExp** (`*a*a*a…` backtracks and freezes the event loop for the whole server): `globMatch()` is a two-pointer wildcard walk. → [architecture-invariants#files-panel-search](docs/architecture-invariants.md#files-panel-search)
 
 **Git status indicator** (`showGitStatus`, per-device, default OFF; `src/git-workspace-status.ts`, `routes/git-status-routes.ts`, `git-status-ui.js`): the bottom-bar indicator and its panel read `GET /api/sessions/:id/git-status` and `/git-diff`, read-only and offline (it never fetches and never writes). ⚠️ git never runs on a repository at or inside a Docker case workspace (walk-up, scan and diff alike, and a cached repository list is re-checked against the CURRENT Docker roots), since a container could plant a clean filter that runs on the host; remote and Docker sessions answer `unsupported`. ⚠️ `git-diff` takes `repo`/`path`/`kind` only as keys matched against the workspace's own repository list (`findWorkspaceRepo()`) and that repository's current status, never as paths. ⚠️ Keep `--no-optional-locks`, `core.fsmonitor=false` and `log.showSignature=false` on every call and `--no-ext-diff --no-textconv` on diffs; clean filters still run, which is why the Docker rule exists. ⚠️ An enclosing repository at `$HOME` or above is ignored (`isUnrelatedAncestor()`), and is identified by one cached `rev-parse` before any full status runs. Everything git supplies renders via `textContent`. Tests: `test/git-workspace-status.test.ts`, `test/routes/git-status-routes.test.ts`, and `test/git-status.browser.test.ts` (browser suite, not in the gate).
 
-**Raw file bodies are streamed and range-aware**: `file-raw`, the attachments `/raw` route and `GET /api/download` share `sendFileBody()`, advertise `Accept-Ranges: bytes` and answer `Range` with `206` + `Content-Range` (single-range, parser in `src/web/http-range.ts`); without it `<video>` cannot seek. The size cap (`MAX_FILE_DOWNLOAD_BYTES`, default 2GB, env `CODEMAN_MAX_DOWNLOAD_BYTES`, `0` = unlimited) is a sanity bound, not memory protection; never reintroduce a whole-file buffer. ⚠️ Bodies go out via `reply.hijack()`, so `sendRawStream` must copy the status onto `reply.raw` by hand or a partial body ships as `200`. ⚠️ Closing the preview must pause and unload media (`_stopFilePreviewMedia`), since a detached `HTMLMediaElement` keeps playing. → [architecture-invariants#raw-file-bodies-streamed-and-range-aware](docs/architecture-invariants.md#raw-file-bodies-streamed-and-range-aware)
+**Raw file bodies are streamed and range-aware**: `file-raw`, the attachments `/raw` route, `GET /api/download` and the rendered-HTML route `GET /html-view/<cap>/...` share `sendFileBody()`, advertise `Accept-Ranges: bytes` and answer `Range` with `206` + `Content-Range` (single-range, parser in `src/web/http-range.ts`); without it `<video>` cannot seek. The size cap (`MAX_FILE_DOWNLOAD_BYTES`, default 2GB, env `CODEMAN_MAX_DOWNLOAD_BYTES`, `0` = unlimited) is a sanity bound, not memory protection; never reintroduce a whole-file buffer. ⚠️ Bodies go out via `reply.hijack()`, so `sendRawStream` must copy the status onto `reply.raw` by hand or a partial body ships as `200`. ⚠️ Closing the preview must pause and unload media (`_stopFilePreviewMedia`), since a detached `HTMLMediaElement` keeps playing. → [architecture-invariants#raw-file-bodies-streamed-and-range-aware](docs/architecture-invariants.md#raw-file-bodies-streamed-and-range-aware)
 
 **Ultracode / workflow-run visualization** (opt-in, default OFF): the Workflow tool writes a completion artifact only at run *end*, so live in-flight runs exist solely as transcript dirs. `workflow-run-watcher.ts` therefore synthesizes ACTIVE runs from transcripts until the completion artifact appears and supersedes them. It is **STANDALONE** and deliberately never imports or touches `subagent-watcher.ts`, despite reading the same tree. Two independent toggles: `showUltracodeAgents` (docked panel) and `ultracodeFloatingWindows` (floating windows); the watcher starts if **either** is on. → [architecture-invariants#ultracode--workflow-run-visualization](docs/architecture-invariants.md#ultracode-and-workflow-run-visualization)
 
 **Clone a repository as a case** (issue #236, Add Case → **Clone Repo**): `POST /api/cases/clone` clones synchronously into the caller's case space (bounded by `GIT_CLONE_TIMEOUT_MS`, no job store); `POST /api/cases/clone-preflight` checks anonymous cloneability and lists refs. Core in `src/git-clone.ts`. ⚠️ **The URL is a code-execution surface**: refuse every `::` form and a leading `-`, spawn only argv arrays with `--` before operands. ⚠️ Stay non-interactive (`gitNonInteractiveEnv()`) or the open request hangs; never collect credentials, refuse `user:password@` URLs. ⚠️ Timeout kills the process GROUP, remove the destination only if this attempt created it, and repo contents win over scaffolding (existing `CLAUDE.md` kept, hooks merged, repo `.claude/settings*` warned about). → [architecture-invariants#clone-a-repository-as-a-case](docs/architecture-invariants.md#clone-a-repository-as-a-case)
 
+**Git worktrees of a case** (case picker → **Git worktrees…**; `src/git-worktrees.ts`, `routes/worktree-routes.ts`): a local repository case keeps worktrees at `<repo>/.claude/worktrees/<name>` on `worktree-<name>`, Claude Code's own `--worktree` layout. ⚠️ Only that managed path is ever created or removed, every component `lstat`ed and realpath-checked (a symlinked `.claude` is refused). ⚠️ git never runs for a Docker or remote case, nor at or inside a Docker workspace (`worktree add` runs hooks and filters on the host). ⚠️ Remove is `git worktree remove` WITHOUT `--force`, refused while a live session is inside; branch delete is `-d`, never `-D`. ⚠️ Sessions start through the existing create paths (quick-start `worktree`, or `POST /api/sessions` `workingDir`), never a new spawn path. → [architecture-invariants#git-worktrees-of-a-case](docs/architecture-invariants.md#git-worktrees-of-a-case)
+
 **Cross-session search**: `GET /api/search` federates an in-memory search over session metadata, run-summary events, and attachment-history entries. The pure core `searchSources()` does substring matching with hard per-type caps: **no regex (so no ReDoS) and no filesystem reads (so no traversal)**. The server-private `externalPath` is never read. PAST sessions (#261) come from `session-history-index.ts`, a capped snapshot of the unified list filled **outside** the request path (`/api/sessions/unified` publishes it; a stale one is rebuilt fire-and-forget), that indirection is what keeps the no-fs property. ⚠️ The snapshot is stored UNSCOPED with a per-row owner and MUST be re-filtered through `canAccessOwned()` on read; history rows carry `jumpTo.kind:'resume-session'`, since a closed session has no tab to select. → [architecture-invariants#cross-session-search](docs/architecture-invariants.md#cross-session-search)
 
 **Web tabs** (dashboard URLs as tabs): a saved URL renders as a tab beside sessions, **NOT a `SessionMode`**, and is proxied through Codeman's own origin (`/webview/<cap>/...`). ⚠️ The proxy is NOT an API surface: its capability-based auth exemption stays fenced to safe methods and non-route paths (`test/webview-auth-exemption.test.ts`). ⚠️ Iframes omit `allow-same-origin` unless `trusted`, and `Authorization`/`codeman_session` are stripped upstream in both modes. ⚠️ **Egress guard**: link-local and cloud-metadata targets are refused at save time, by a sync hostname check at each connect (IP literals skip DNS), AND on the resolved address (`webview-egress.ts`); use the `undici` package's own `fetch` + `Agent`, never Node's global fetch. ⚠️ Loopback links in agent output auto-open as proxied web tabs (`openLinkThroughWebTabIfLoopback`), but never auto-route `*.localhost` (prompt-injectable DNS). Capabilities are revoked on logout (`revokeOwner`). → [architecture-invariants#web-tabs](docs/architecture-invariants.md#web-tabs), `docs/web-tabs.md`
 
-**Multi-user mode** (opt-in `--multiuser` / `CODEMAN_MULTIUSER=1`, OFF by default): named users with scrypt-hashed passwords in `~/.codeman/users.json`. Gated everywhere by `isMultiUserMode()`; when OFF, behavior is byte-identical to single-user because every scoping helper short-circuits. ⚠️ **Not a security boundary at the agent layer**: every session still runs as the SAME OS account. This separates WORKSPACES; it does not sandbox users (Docker cases are the isolation story). Ownership threads through `Session.owner` and is enforced in `findSessionOrFail`, list endpoints, SSE routing (fail-closed), WS, search, and file-preview. → [architecture-invariants#multi-user-mode](docs/architecture-invariants.md#multi-user-mode), `docs/multi-user-plan.md`
+**Multi-user mode** (opt-in `--multiuser` / `CODEMAN_MULTIUSER=1`, OFF by default): named users with scrypt-hashed passwords in `~/.codeman/users.json`. Gated everywhere by `isMultiUserMode()`; when OFF, behavior is byte-identical to single-user because every scoping helper short-circuits. ⚠️ **Not a security boundary at the agent layer**: every session still runs as the SAME OS account. This separates WORKSPACES; it does not sandbox users (Docker cases are the isolation story). Ownership threads through `Session.owner` and is enforced in `findSessionOrFail`, list endpoints, SSE routing (fail-closed), WS, search, and file-preview. ⚠️ `settings.json` is ONE server-wide file, so every write to it is admin-only (`PUT /api/settings`, `/api/execution/model-config`, `/api/cases/order`, plus `PUT /api/config`: 403 `FORBIDDEN` before parsing), a non-admin's quick-start/Ralph start skips `lastUsedCase`, its GET hides `voiceSettings.apiKey`/`caseOrder`/`lastUsedCase` from a non-admin (`settingsForNonAdmin`), and the browser skips those writes for a non-admin via `_settingsReadOnly()` (settings-ui.js; also true once a write came back 403 `FORBIDDEN`, since `/api/me` can be slow or fail; their run mode and respawn presets stay per-device, `_applyOwnSettings`) (`test/routes/system-routes-settings-admin-only.test.ts`). System ops are admin-only too (`/api/system/update`, `/api/system/span-displays`, `/api/tunnel/qr/regenerate`, `/api/cleanup-state`), and a non-admin's `/api/auth/revoke` revokes only their own logins (`test/routes/system-routes-system-ops-admin-only.test.ts`). → [architecture-invariants#multi-user-mode](docs/architecture-invariants.md#multi-user-mode), `docs/multi-user-plan.md`
 
 **Away digest**: `GET /api/away-digest` aggregates what happened while you were away from the lifecycle log, run-summary events, live sessions, token stats, and recent subagents. Pure aggregator in `web/away-digest.ts`. ⚠️ Returns `{success:true,digest}`, a legacy raw-ish shape consistent with the other raw GET handlers in `system-routes.ts`; frontend and tests read `.digest`. → [architecture-invariants#away-digest](docs/architecture-invariants.md#away-digest)
 
@@ -327,76 +249,14 @@ Codeman is a Claude Code session manager with web interface and autonomous Ralph
 
 ### Frontend
 
-Frontend JS modules have `@fileoverview` with `@dependency`/`@loadorder` tags. Load order: `constants.js`(1) → `i18n.js`(1.5) → `mobile-handlers.js`(2) → `voice-input.js`(3) → `notification-manager.js`(4) → `keyboard-accessory.js`(5) → `input-cjk.js`(5.5) → `mobile-ime-preview.js`(5.52) → `terminal-keycode229-recovery.js`(5.55) → `sanitize-html.js`(5.6) → `tab-layout-browser.js`(5.9) → `app.js`(6) → `tab-rail-resize.js`(6.5) → `terminal-ui.js`(7) → `terminal-tile.js`(7.4) → `terminal-split.js`(7.5) → `tile-grid.js`(7.6) → `respawn-ui.js`(8) → `ralph-panel.js`(9) → `orchestrator-panel.js`(9.5) → `cron-ui.js`(9.7) → `settings-ui.js`(10) → `panels-ui.js`(11) → `readmymind-ui.js`(11.3) → `ultracode-panel.js`(11.5) → `approvals-ui.js`(11.6) → `reboot-restore-ui.js`(11.65) → `admin-ui.js`(11.7) → `session-ui.js`(12) → `host-wake-ui.js`(12.2) → `webview-tabs.js`(12.5) → `mobile-overview.js`(12.55) → `home-sessions.js`(12.56) → `git-status-ui.js`(12.57) → `entrance-animations.js`(12.6) → `ralph-wizard.js`(13) → `api-client.js`(14) → `subagent-windows.js`(15) → `ultracode-windows.js`(15.5) → `session-lineage.js`(15.6) → `image-input.js`(16) → `spreadsheet-preview.js`(16.5). `i18n.js` translates static + newly inserted application DOM while skipping terminal/response/file/user-name surfaces; `input-cjk.js` handles CJK IME composition via an always-visible textarea below the terminal (`window.cjkActive` blocks xterm's onData). `terminal-keycode229-recovery.js` forwards a committed `input` event that xterm's `_inputEvent` guard drops (Chrome-on-Android soft keyboards send `composed: true` after a keydown), and only when xterm emitted no canonical data for that keystroke. ⚠️ **That decision is settled at the NEXT keydown as well as on its own zero-delay timer** (#441): the drain runs from xterm's custom key handler, which fires BEFORE xterm processes that key, so a soft keyboard that commits the last character and sends Enter in one InputConnection transaction puts the character on the wire ahead of the `\r`. On the timer alone that character is not merely late, it is LOST: xterm emits the `\r` first and bumps the canonical counter past the candidate's snapshot, so the candidate stands down (measured, `hell\r` where the user typed `hello`). The edit-based diff described below is settled first at that keydown, so the drain decides with the evidence the timer had and Enter's textarea clear cannot turn the pending line into DELs. Ordering is pinned by `test/terminal-keycode229-recovery.browser.test.ts`, which the CI gate does NOT run. The same module replaces xterm's `_handleAnyTextareaChanges` (an append-only `newValue.replace(oldValue, '')` diff) with an edit-based one, so an Android autocorrect on space (delete + insert) reaches the PTY once instead of duplicating the line; ⚠️ that diff is settled at the NEXT keydown, before xterm handles that key, because xterm clears the textarea for Enter and a pending diff would then send one DEL per character ahead of the submitted line (except a composition xterm finalizes synchronously at that key, which stays xterm's). ⚠️ Every xterm that takes keyboard input wires its OWN controller from this module: the primary pane (terminal-ui.js `initTerminal()`) and each `TerminalTile` (terminal-tile.js `_createKeyCode229Recovery()`, so every grid tile and the split's Pane B, which a wide Android tablet reaches), each on its own textarea, composition helper and session; in both, `handleKeyEvent` must run ABOVE the custom key handler's keyCode-229 early return, and `notifyCanonicalData` sits in the onData lambda, never in the send path the recovered bytes also take. `mobile-ime-preview.js` (iOS WebKit only) paints the text an IME is composing: an iOS IME commit is routed into the local-echo overlay through the ordinary printable/paste branch and then `_transferMobileImeCommitToLocalEcho`, and without local echo the preview clears only on output parsed AFTER the commit (or its 2 s fallback). ⚠️ It watches keydown in the capture phase on `terminal.element`, never on the textarea, because xterm finalizes the composition and emits the commit in its own capture listener on the textarea.
-
-**Entrance animations** (`entrance-animations.js`, all OFF by default): opt-in animations for tabs, terminal, windows and connection lines, chosen via `data-tab-anim` / `data-term-anim` / `data-win-anim` / `data-line-anim` on `<html>`; the default `legacy` theme short-circuits every hook. ⚠️ Tabs and lines are destroyed mid-animation on re-render, so re-apply to the fresh element by id with a negative `animation-delay` (resume, never restart). ⚠️ Terminal-pane styles may animate only transform / opacity / clip-path (anything else resizes the PTY via FitAddon); `blur` is the ONE sanctioned `filter` exception, do not generalise it. ⚠️ Line glow lives in `--line-glow` so blur keyframes interpolate. Persisted per-device in `codeman:*Anim` localStorage keys, never in `SettingsUpdateSchema`; lab at `?animlab=1`. Test: `test/entrance-animations.test.ts`. → [architecture-invariants#entrance-animations](docs/architecture-invariants.md#entrance-animations)
-
-**Mobile tab strip scrolling** (issue #257): under 768px the tab strip scrolls horizontally, so the active tab must be kept reachable. `_updateActiveTabImmediate()` reveals it via `computeTabScrollLeft()` (constants.js, rect math on the strip's own `scrollLeft`, never `scrollIntoView()`, which scrolls the document under the fixed header); `_fullRenderSessionTabs()` must restore `scrollLeft` across rebuilds and re-reveal only when the active tab changed (`_lastRenderedActiveTabId`) or, grouped by state, the ACTIVE tab moved to another state band (`_noteActiveTabBand()`, both render paths, scrolling row only; another tab's move never scrolls). ⚠️ The phone-block `min-width` on `.session-tab.active .tab-name` keeps the tab's centre off the gear/close icons, sized for numberless tabs 10+ (floor 40px); do not shrink it. ⚠️ Never reintroduce hoisting the active session to the front of the strip. Tests: `test/mobile-tab-tap-zones.test.ts`, `test/tab-triage.test.ts`. → [architecture-invariants#mobile-tab-strip-scrolling](docs/architecture-invariants.md#mobile-tab-strip-scrolling)
-
-**Session list layout: header strip or left sidebar** (`sessionListLayout`, default `header`; per-device via `displayKeys`, also in `SettingsUpdateSchema`): the list can move into a collapsible `<aside>` (Alt+B, `toggleSessionSidebar`) or, via `tabOrientation`, a resizable vertical `#tabRail` (desktop/tablet only). ⚠️ There is ONE `#sessionTabs`, MOVED between hosts, never a second list: `applySessionListLayout()` runs first, then `applyTabOrientation()`, both BEFORE `applyTabWrapSettings()`, and both arm/disarm `_startSidebarRichClock()`. ⚠️ Axis decisions use `_isVerticalTabList()`, never `isSessionSidebarActive()` alone. ⚠️ Rich rows share one gate, `isRichTabRows()`; rich CSS pairs sidebar+rail with comma-grouped selectors, never `:is()`, and card rules stay rail-scoped. ⚠️ Rail sort (`tabRailSort`, default `activity`) is the flex `order` property only, never a DOM reorder; the arrow-key walk alone follows computed `order`. ⚠️ Leaving sidebar mode clears `_sidebarFilter`; the handheld overlay drawer is `inert` when closed, the docked rail never. → [architecture-invariants#session-list-layout-header-strip-vs-left-sidebar](docs/architecture-invariants.md#session-list-layout-header-strip-vs-left-sidebar)
-
-**Tab layouts** (`tabArrangement`, per-device, default `classic`; `state`/`case`/`ledger` are opt-in; Discussion #426): `state` (option C) splits the list into needs you / waiting / working / idle (idle is `quiet`: its heading stays as the row's anchor but draws no label; `tabStateOrder: 'urgent-last'` flips the order, needs you at the bottom): rows in the desktop header strip with left-aligned labels in a measured column (the brand sits over the strip's corner, so the rows after the first start under "Codeman"), sections in the flat rail and the sidebar, inline dividers on the tablet strip, group order with no headings on phones. `case` (option A) wraps each case's tabs in one `.tab-cluster` box and hides the `-<case>` of a generated name (`.tab-name-case`). `ledger` (option B) is a CSS-only grid on the desktop header strip. `classic` is the old strip. Pure cores: `CodemanTabTriage` and `CodemanTabClusters` (constants.js). ⚠️ `state` is the flex `order` property plus `aria-hidden` heading/break elements reconciled in place by `_syncTabTriageChrome()` after BOTH render paths, never a DOM reorder (a state change is an incremental pass). ⚠️ `case` boxes ARE a DOM grouping, so the incremental path rebuilds whenever `_tabClusterLayout().key` changes; membership follows `_mobileOverviewCaseFor()` like the home screens. ⚠️ Drag only reorders within a group or box (`_isTabDropAcrossGroups()`). ⚠️ Named groups in the vertical rail win over both groupings. ⚠️ `classic` must leave no trace (no headings, no inline order, no boxes, no name split), and the ledger must not change the markup. Tests: `test/tab-triage.test.ts`, `test/tab-clusters.test.ts`. → [architecture-invariants#tab-layouts-by-state-by-case-ledger](docs/architecture-invariants.md#tab-layouts-by-state-by-case-ledger)
-
-**Header stats styles** (`headerStatsStyle`, per-device, desktop only, default `compact`; Discussion #426 option G): `tiles` / `compact` (two pills, a ring beside every value) / `classic` (as before), keyed on `data-header-stats` (pre-paint + `applyHeaderStatsStyle()`), forced to `classic` below 768px and in a solo window. ⚠️ The clustered styles MOVE `#connectionIndicator` into `#headerSystemStats` and `#planUsageChip` right after it; `classic` moves them back to comment anchors. ⚠️ WS joins the pill only while System Stats is shown. ⚠️ The extra parts (stat rings, tile words, plan rings, meters) are always rendered and hidden by default in CSS, which is what keeps `classic` unchanged; tile words are derived beside the connection descriptor, never added to it (`test/connection-indicator.test.ts` pins its shape). ⚠️ A tile is a three-row grid (label, value, bar) with pixel line-heights and the bundled JetBrains Mono; never lay the bar over a fixed-height tile (a taller system mono put the value on top of it). The WS tile's grid lives on the inner `.connection-tile`, because JS writes the indicator's `display` inline. Tests: `test/header-stats-style.test.ts`. → [architecture-invariants#header-stats-styles](docs/architecture-invariants.md#header-stats-styles)
-
-**Phone overview home screen** (`mobile-overview.js`, per-device `mobileOverviewEnabled`, default ON): under 600px the "C" logo shows NEEDS YOU / CURRENT / PAST SESSIONS instead of the welcome overlay, branched in `showWelcome()`/`hideWelcome()` via width-driven `shouldUseMobileOverview()`. ⚠️ The container ships `hidden` and only this module removes it: never give `.mobile-overview` a bare `display` rule (desktop does not load `mobile.css`). ⚠️ The split Run button must carry the toolbar's own classes (`btn-toolbar btn-run mode-<backend>` / `btn-run-gear`) and mobile.css must set no `background`/`color` on it; row status must mirror the session-tab alert language. PAST rows resume through the shared `resumeHistorySession()`. Status pills carry `data-i18n-skip`. → [architecture-invariants#phone-overview-home-screen](docs/architecture-invariants.md#phone-overview-home-screen)
-
-**Desktop home tab rail** (`home-sessions.js`, desktop only): the welcome overlay's left gutter carries the open tabs as a rail docked flush left, full height, in overview order, each row showing `created … · <state> <duration>` from `_mobileOverviewSince()`; state classification is reused from mobile-overview.js (so it loads after it). ⚠️ The number badge is the Alt+1..9 tab-strip index, never renumber it to row position. ⚠️ The width gate lives in two places that must stay equal: `HOME_SESSIONS_MIN_WIDTH` (1180) and a `max-width: 1179px` media query. ⚠️ `.home-sessions[hidden]` must re-assert `display: none`. ⚠️ Size all children in `em` off the one `clamp()` knob, never `rem`/px. Age stamps tick in place (`_tickHomeSessionsTimes()`), never by re-render. Test: `test/home-sessions.test.ts`. → [architecture-invariants#desktop-home-tab-rail](docs/architecture-invariants.md#desktop-home-tab-rail)
-
-**Home-screen session order** (`CodemanSessionOrder` in constants.js, pure): BOTH home screens (phone overview, desktop rail) must order rows through this ONE comparator. Rank `needs` → `error` → `waiting` → `working` → `idle` → `done`. ⚠️ The tiebreak flips: states a session is still IN sort oldest-first, states it has STOPPED sort newest-first. ⚠️ The running group keys off `lastSubmitAt`, never `lastActivityAt` (a working pane repaints constantly). ⚠️ A 0 stamp means unknown and sorts last within its state. Final tiebreak is `orderIndex`, so the list never shuffles. The tab strip itself is NOT sorted by this. Test: `test/session-overview-order.test.ts`. → [architecture-invariants#home-screen-session-order](docs/architecture-invariants.md#home-screen-session-order)
-
-**Welcome "Resume Conversation" list** (terminal-ui.js): `loadHistorySessions()` fetches once and caches the corpus on `_historyAll`/`_historyCases`; every subsequent view (filter box, sort select, expand, the periodic refresh in panels-ui.js) goes through `_renderHistoryList()`, so never append rows to `#historyList` directly or re-fetch to re-sort. ⚠️ The box height is **class-driven**: expanding the list without `.history-list.expanded` leaves the collapsed `max-height` in place and just deepens a scroll well, which is the bug #260 reported (35 sessions in a ~4-row box). ⚠️ The A–Z sort keys off `_historyRowLabel()`, the SAME string the row renders (`name || firstPrompt || path`), most rows are transcript-backed and have no session name, so sorting on `name` alone silently does nothing. ⚠️ A filter implies expansion, and `_renderSearch()` hides `#historyHeader` (title + controls) as one unit while a search is active. Tests: `test/history-list-controls.test.ts`.
-
-**Command palette + shortcut registry**: `Ctrl/Cmd/Alt+K` opens the session palette; shortcuts live in a rebindable registry (`DEFAULT_SHORTCUTS`/`getShortcutRegistry()`/`matchesShortcutEvent()` in app.js, overrides in `settings.shortcutOverrides`). ⚠️ Palette-chord keys must ALSO be swallowed in `attachCustomKeyEventHandler` (terminal-ui.js) or xterm writes the control byte into the PTY. ⚠️ The global Escape handler (app.js) calls EVERY close method on every Escape, in the capture phase, so a close method that does more than hide (the palette and Session Manager restore focus) must return early when its overlay is not open. ⚠️ `saveAppSettings()` rebuilds settings from the DOM, so keys edited elsewhere (`shortcutOverrides`, `showTokenCount`, `showCost`) need explicit `_prev` carry-over. ⚠️ **Smart copy (`Ctrl+C`)**: with no selection it must `return true` without `preventDefault()` or the interrupt is lost; keep `copyTerminalSelection` out of `SHORTCUT_ACTIONS`. The gate tests the CLEANED selection (`CodemanCopySelection.clean`: trailing padding, plus a LEADING margin only up to the width the CLI declares in `capabilities.transcriptGutter`, never one derived from the pane); the strip is not idempotent, so clean once and pass the RAW selection on, and leave Alt+drag column selections untouched. → [architecture-invariants#command-palette-and-shortcut-registry](docs/architecture-invariants.md#command-palette-and-shortcut-registry)
-
-**Per-device vs synced settings**: the `displayKeys` set in settings-ui.js is a **client-side merge policy**, not a wire filter. A display key seeds from the server only when localStorage has no value for it, which is what prevents one device overwriting another; `showPlanUsageLimits` is additionally `delete`d from the incoming payload outright. Separately, `SettingsUpdateSchema` is `.strict()` and simply **does not declare** `skin`, `showFileViewerButton`, `showCronButton`, `webglRendererEnabled`, `localEchoEnabled`, `cjkInputEnabled`, or `extendedKeyboardBar`, so sending one of those is a validation error. The rest (`showResponseViewer`, `showPlanUsageLimits`, `language`, and most `show*` keys) ARE in the schema and do persist server-side; they are per-device by client policy only. ⚠️ Adding a new per-device setting means deciding **both** questions: membership in `displayKeys`, and presence in the schema.
-
-**Settings surface** (`#appSettingsModal` + `#sessionOptionsModal` + `#createCaseModal`): one `set-*` language shared through a single `:is(...)` id scope in styles.css. App Settings' rail is a table of contents over ONE scrolling document (`switchSettingsTab` scrolls); Session Options and Add Case really switch (`switchOptionsTab` / `switchCaseModalTab`), and their larger per-modal size blocks are the design, not drift. ⚠️ **The load/save contract is `getElementById` by id**: renaming or dropping a control id silently stops it loading or saving. ⚠️ The Session Options "Session" entry still keys off `context` (label-only rename). ⚠️ Add Case keeps its legacy `.form-row` markup via an adapter; every `<details>` there needs `.set-adv-chev` plus both marker suppressions. ⚠️ Model cards and the effort segment are views over hidden `<select>`s, which stay the source of truth. ⚠️ `.modal-tabs*` classes are retired; `admin-ui.js` needs `.set-rail-items` + `.set-doc` to survive any restructure. Guard: `test/app-settings-structure.test.ts`. → [architecture-invariants#settings-surface-app-settings-session-options-add-case](docs/architecture-invariants.md#settings-surface-app-settings-session-options-add-case)
-
-**Header button visibility**: most header controls are opt-in and hidden by a marker class (`btn-multimonitor--hidden`, `btn-response-viewer-header--hidden`, `btn-file-viewer--hidden`, `btn-cron--hidden`, and the desktop-gated `btn-split--hidden` / `btn-tile-grid--hidden`, which also need a `@media (max-width: 1179px)` backstop) that `applyHeaderVisibilitySettings()` (settings-ui.js) toggles after settings load; the multi-monitor button is instead stripped at render by `renderIndexHtml`. ⚠️ Hiding must go through the marker class: the base rules are `display:inline-flex !important`, so an inline style cannot override them. Current desktop default is WS/CPU/MEM + File Viewer + Tiles (1180px and wider) + gear, with the token chip and lifecycle-log button OFF. ⚠️ New header controls must not leak onto phones; `test/mobile-header-buttons-policy.test.ts` is the static guard. → [architecture-invariants#header-button-visibility-multi-monitor-response-viewer-file-viewer-cron](docs/architecture-invariants.md#header-button-visibility-multi-monitor-response-viewer-file-viewer-cron)
+**Frontend invariants live in [`src/web/CLAUDE.md`](src/web/CLAUDE.md)**: module load-order notes and every `src/web/public/` UI subsystem (tab strip, sidebar, rail and tab layouts, tile grid, split pane, lineage lines, home screens, settings surface, skins and fonts, mobile keyboard/toolbar/composer, terminal touch, copy and paste, file-preview markdown, XLSX preview limits, z-index stack, connection-loss UI, SSE watchdog, keyboard shortcuts incl. the Ctrl+W owner decision). Claude Code loads it automatically when the Read tool opens a file under `src/web/`. ⚠️ Reading files through Bash (`cat`, `sed`, `grep`) does NOT load it, so **Read `src/web/CLAUDE.md` before editing anything in `src/web/public/`** if it is not already in context.
 
 **Gesture control** (camera hand-tracking overlay, opt-in, default OFF): `CODEMAN_GESTURE=1` makes the feature *available*; `gestureControlEnabled` turns it on. The bundle is injected by `renderIndexHtml` only when enabled, which is why that method is `async` and reads settings with `readSettings(true)` (a fresh read: a post-save reload lands inside the 2s cache TTL and would otherwise render the pre-toggle state). **Source lives in `packages/gesture-control/`; edit there, run `npm run build:gesture`, and commit the regenerated bundle** because dev serves the committed bundle with no runtime bundler. The MediaPipe wasm + model are fetched separately and gitignored. ⚠️ Keep `MP_VERSION` in `fetch-gesture-assets.mjs` in sync with `@mediapipe/tasks-vision`. → [architecture-invariants#gesture-control-the-source-package](docs/architecture-invariants.md#gesture-control-the-source-package)
 
-**Terminal font weight** (`terminalFontWeight` / `terminalFontWeightBold`, per-device, default = xterm's own `normal`/`bold`): Claude Code's markdown bold is a bare `ESC[1m`, so the weight step is its only cue. `CodemanTerminalFont.resolveWeights()` (constants.js, pure) resolves each slot against **its own** xterm default. ⚠️ The `@font-face` for `fonts/jetbrains-mono-variable.woff2` must stay declared `100 800` (the browser synthesizes from the descriptor, not the file); narrowing it silently makes the setting a no-op. ⚠️ A live save must reach both echo overlays (`refreshFont()`) and open Agent Teams panes. ⚠️ Leave `_awaitTerminalFont()` untouched. Test: `test/terminal-font-weight.test.ts`. → [architecture-invariants#terminal-font-weight](docs/architecture-invariants.md#terminal-font-weight)
-
-**Theme skins / branding / i18n**: `skin` selects a palette via `data-skin` on `<html>`, applied by an **inline pre-paint script** in `index.html` reading `localStorage['codeman:skin']` to avoid a flash of wrong theme. ⚠️ A skin is **four things that must stay in sync**, and missing any one degrades silently: the `html[data-skin="…"]` token block in `styles.css`, the xterm ANSI palette in `terminal-ui.js`, the pre-paint allowlist, and the Settings picker (both in `index.html`). `test/skin-themes.test.ts` is the static guard. Light skins additionally need `color-scheme: light` and xterm `minimumContrastRatio: 4.5`, and `applyTerminalSkin()` must call the local-echo overlay's `refreshFont()` because it caches the terminal fg/bg. `displayName` changes user-facing browser branding only and must NEVER rename npm package, CLI, API, storage, CSS, or protocol identifiers. `language` (`en`/`zh-CN`) keeps English as the canonical source so live switching stays reversible. User display names flow through `textContent`/attribute APIs and the server title's HTML escaper, never `innerHTML`. → [architecture-invariants#theme-skins](docs/architecture-invariants.md#theme-skins)
-
-**Foldable settings identity**: responsive layout is width-driven via `MobileDetection.getDeviceType()`, but the localStorage namespace uses `MobileDetection.isHandheldDevice()` so an unfolded Android foldable keeps `codeman-app-settings-mobile`. ⚠️ Do not switch per-device settings namespaces from instantaneous viewport width: a posture-triggered WebView reload would lose opt-in UI. Regression profile: `OPPO Find N5 (unfolded)` in `test/mobile/devices.ts`. → [architecture-invariants#foldable-settings-identity](docs/architecture-invariants.md#foldable-settings-identity)
-
-**Folding devices: a fold is not a keyboard, and dialogs avoid the hinge**: ⚠️ in `handleViewportResize()`, a visual-viewport resize that changes the WIDTH is a shape change (rotation, fold) and must never be read as the keyboard; it re-baselines instead, or `keyboardVisible` latches with no keyboard. ⚠️ `init()` must seed `lastViewportWidth`. ⚠️ With the keyboard up, a shape change baselines to `window.innerHeight`, never the shrunk visual height. ⚠️ The hinge is reserved via `--fold-inline-end`/`--fold-block-end` (0px when unfolded): each overlay fold rule must re-state its own gutter, a base gutter overridden by a later `@media` block needs its own fold restatement there on a zero base, and dialogs use physical sides (left/top segment) in every language. Guard: `test/foldable-layout.test.ts`. → [architecture-invariants#folding-devices](docs/architecture-invariants.md#folding-devices)
-
-**WebGL renderer toggle** (`webglRendererEnabled`, per-device): the GPU-stall watchdog's sticky `codeman-webgl-disabled` marker survives page loads and is cleared only by an explicit OFF→ON save or `?webgl=force`. `?nowebgl` forces the DOM renderer per-load. → [architecture-invariants#webgl-renderer-toggle](docs/architecture-invariants.md#webgl-renderer-toggle)
-
-**Shell keyboard accessory bar + one-shot Ctrl** (`keyboard-accessory.js`): a shell-mode session swaps the mobile accessory bar for terminal controls; `setMode()` records `extendedKeyboardBar` as the base layout and `refreshForActiveSession()` resolves base-vs-shell. ⚠️ Ctrl is a one-shot modifier applied in `terminal.onData` (after `shouldSuppressTerminalQueryResponse`, before every send path), and must skip `isTerminalFocusOrMouseReport()` chunks. ⚠️ It must disarm on use, second tap, any other accessory key, session switch, keyboard dismissal and layout swap. ⚠️ `_handleCjkInput()` must apply it too (the CJK textarea bypasses onData). Mapping: `ctrlByteFor()`. ⚠️ mobile.css's light-skin repaint must keep excluding `.accessory-btn:not(.armed)` or the armed state is invisible. → [architecture-invariants#shell-keyboard-accessory-bar-and-one-shot-ctrl](docs/architecture-invariants.md#shell-keyboard-accessory-bar-and-one-shot-ctrl)
-
-**Mobile prompt composer** (`keyboard-accessory.js`): the agent bars' Paste key is **Compose**, a native multiline dialog where only **Send** submits (the shell bar keeps plain Paste). Opening it adopts the whole terminal prompt (`_takePendingLocalEcho`), erasing the flushed prefix with backspaces counted in code points. ⚠️ Drafts are per-session and in memory only (`_composerDrafts`), never persisted (prompts carry secrets). ⚠️ Delivery is a hand-built bracketed-paste frame via `_sendInputAsync` WITHOUT `useMux`, then a separate delayed Enter WITH it; never `terminal.paste()`, and the frame must never take the mux fallback (it strips newlines). ⚠️ `_composerMaxLength` must stay derived from `MAX_INPUT_LENGTH` minus the markers, or an oversized frame wedges the durable queue. ⚠️ The composer overlay needs its own gutter restatement after the fold rules. Test: `test/mobile-prompt-composer.test.ts`. → [architecture-invariants#mobile-prompt-composer](docs/architecture-invariants.md#mobile-prompt-composer)
-
 **PTY and browser terminal geometry** (#464): a browser terminal whose width differs from the PTY's garbles Claude's redraws, so `syncTerminalGeometry()` (terminal-ui.js) is the ONE function that may resize the main terminal (never a bare `fitAddon.fit()`), a font change is a geometry change, and the fit is withheld wherever the SIGWINCH is. ⚠️ Resize is answered with `Session.ptyGeometry`, and a client adopts its COLUMNS only (never rows); `ptyGeometry` is null without a live pane. → [architecture-invariants#pty-and-browser-terminal-geometry](docs/architecture-invariants.md#pty-and-browser-terminal-geometry)
 
-**Terminal resilience**: a replay clear is the queued in-stream `\x1bc` in `_resetTerminalForReplay()`, never `reset()`/`clear()` (queued bytes fuse into the snapshot); the renderer watchdog `_kickRenderer()` reads xterm privates, pinned by `test/xterm-private-api.test.ts` against the resolved lockfile version; every terminal capture fetch has a deadline that covers the BODY (`_fetchTerminalCapture`). → [architecture-invariants#terminal-resilience-replay-clears-renderer-liveness-fetch-deadlines](docs/architecture-invariants.md#terminal-resilience-replay-clears-renderer-liveness-fetch-deadlines)
-
-**WebSocket output-gap reconcile** (`_wsOutputGapSession`, app.js): output frames carry no sequence number, so an unintentional WS close while SSE stays up marks the session and the next open reconciles. ⚠️ The marker is cleared only once a repaint actually happened (`_markTerminalBufferReconciled()`, never in a `finally`). → [architecture-invariants#websocket-output-gap-reconcile](docs/architecture-invariants.md#websocket-output-gap-reconcile)
-
 **Service worker precache** (`sw.js` + `scripts/build.mjs`): `BUILD_ID` and `HASHED_ASSETS` are build-generated and the build THROWS unless each declaration appears exactly once; `caches.match` must pass `ignoreSearch: true` because `cacheBustAssets` appends `?v=` to hashed names. → [architecture-invariants#service-worker-precache-and-cache-key](docs/architecture-invariants.md#service-worker-precache-and-cache-key)
 
-**Dismissing the on-screen keyboard** (`terminal-ui.js`): two gestures blur the terminal's hidden textarea. (1) `_installMobileKeyboardDismiss()`, a document `touchend` that must never fire inside `#terminalContainer` or on a control (`MOBILE_KEYBOARD_DISMISS_EXEMPT_SELECTOR`, via `closest()`; roving-tabindex items sit at `tabindex=-1`, so the grouped rail's are listed as `[role="treeitem"]`). (2) In `_handleMobileTerminalTap`, a second tap on inert `content` blurs; the prompt row keeps focus-then-position. ⚠️ A scroll also ends in `touchend`: both classifiers must share one threshold (`TAP_THRESHOLD` reads `MOBILE_KEYBOARD_DISMISS_TAP_SLOP`), and multi-touch is never a tap. ⚠️ CI cannot see the only test for (1): run `npm run test:mobile -- test/mobile/keyboard.test.ts` by hand and diff the FAIL list against master. → [architecture-invariants#dismissing-the-on-screen-keyboard](docs/architecture-invariants.md#dismissing-the-on-screen-keyboard)
-
-**Phone toolbar: Enter replaces Shell** (post-1.8.0): inside `@media (max-width: 599px)` `btn-shell` is `display:none` and `btn-enter` takes its slot (`order: 4`); starting a shell moved into the Run dropdown (`Terminal / Shell` → `setRunMode('shell')` → `run()` → `runShell()`, button label "Run SH"). `runMode` is `z.string().max(20)` server-side, so new modes need no schema change. Desktop and tablet keep the green Run Shell button unchanged.
-
-⚠️ **`sendEnterKey()` MUST go through `terminal._core.coreService.triggerDataEvent('\r', true)`** — not `sendInput()`, and never a raw POST to `/api/sessions/:id/input`. `localEchoEnabled` defaults to `MobileDetection.isTouchDevice()`, so on every phone the characters you type are buffered in the `LocalEchoOverlay` and have **never reached the PTY**; the `onData` Enter branch in terminal-ui.js is what flushes `pendingText` first and only then sends `\r` (after an 80ms delay so text lands first). Sending a bare `\r` submits an empty line and strands the typed text on screen, so the button looks dead. Replaying the keypress reuses the overlay flush, the flushed-offset cleanup and the ordering instead of reimplementing them. `KeyboardAccessory.sendKey()` is for escape sequences (arrows/Esc) and is the WRONG template to copy for input.
-
-⚠️ **Skin overrides outrank plain class rules.** `styles.css` nests its skin block inside `html:not([data-skin="og"]) { … }`, so a bare `.btn-toolbar` rule in there resolves to specificity **(0,2,1)** and beats a `.btn-toolbar.btn-x` rule **(0,2,0)** in `mobile.css` regardless of load order. Toolbar-button colors set from mobile.css therefore need `!important` — that is why mobile.css leans on it so heavily. Symptom: only your `!important` properties land and everything else silently renders in generic toolbar grey.
-
-**Connection-loss UI** (`computeConnectionLossUi()` in constants.js, writer `_updateConnectionLossUi()` in app.js): the service worker serves the cached app shell, so an unreachable server (phone off the tailnet, VPN down, server stopped) used to render a normal-looking empty dashboard whose only tell was the 8px header dot, which reads as "no sessions", not "no connection". Two surfaces now: a full-screen **overlay** while no server state has loaded this page load (nothing behind it is worth preserving), and a non-blocking **banner** once it has (the terminal scrollback stays readable). ⚠️ A **2.5s grace** is load-bearing: a COM deploy restarts the server and SSE is back in ~200ms, and a banner on every deploy trains the user to ignore it. `navigator.onLine === false` skips the grace, since that is never a blip. Retry re-arms SSE **and** the terminal WS (`planWsReconnect` can 'give-up', and the SSE backoff caps at 30s).
-
-**SSE staleness watchdog** (`computeSseStale()` in constants.js, `_checkSseStale()` + a 5s interval in app.js): an `EventSource` can stop delivering without erroring, so the client forces a reconnect when nothing arrives. ⚠️ The server keepalive must stay the named `sse:heartbeat` event (`cleanupDeadClients()`, sse-stream-manager.ts), never an SSE comment, which `EventSource` cannot observe; its no-op client listener must stay registered. ⚠️ Judge staleness only while `connected` and online (the loop breaker). ⚠️ The liveness stamp lives inside `addListener`. ⚠️ Clear the interval only at the top of `connectSSE()`, or intervals stack. → [architecture-invariants#sse-staleness-watchdog](docs/architecture-invariants.md#sse-staleness-watchdog)
-
-**Z-index layers** (keep new overlays consistent with this stack): local echo overlay (7; with local echo on it also draws the iOS IME composition preview, as an underlined tail after its pending text via `setComposition`), iOS IME composition preview span when local echo is off (6 inside `.xterm-helpers`, whose own z-index 5 is its EFFECTIVE layer, so it sits UNDER the overlay and must never be used while the overlay shows text), terminal touch-selection bar (900, below floating agent windows), subagent windows + split picker menu + Tiles count menu (1000), plan agents (1100), mobile/tablet fixed header (1200), modals on ≤768px (1300, must beat the fixed header), log viewers (2000), connection-loss overlay (2500), image popups (3000), response viewer (5000, backdrop 4999), file-preview overlay (5100, must outrank the response viewer that launches it), toasts/path picker (10000+), custom-model center-status banner (10001; its `[hidden]` must re-assert `display: none` or `dismiss()` leaves an invisible click-blocker), custom-model swap-confirm/context-warning modals (10010). → [architecture-invariants#z-index-layers](docs/architecture-invariants.md#z-index-layers)
-
-**Respawn presets**: `solo-work` (3s/60min), `subagent-workflow` (45s/240min), `team-lead` (90s/480min), `ralph-todo` (8s/480min), `overnight-autonomous` (10s/480min).
-
-**Keyboard shortcuts**: Escape (close), Ctrl+? (shortcut overlay), Ctrl/Cmd/Alt+K (session palette), Ctrl+Tab (next), Alt+[/] (prev/next tab), Alt+1-9 (switch tab), Ctrl+Shift+{/} (move tab left/right), Shift+Enter or Ctrl+Enter (newline), Ctrl+C (copy selection, else interrupt) / Ctrl+Shift+C (copy, never interrupts), Ctrl+L (clear), Ctrl+Shift+R (restore size), Ctrl+Shift+V (voice input), Ctrl/Cmd +/- (font), Shift+Wheel (local scrollback when mouse passthrough is active), Shift+drag (start a selection in a stripped-DECSET pane, where xterm's own Shift branch is unreachable and a Shift+drag used to select nothing; `_installShiftDragSelection`), right-click (copy the selection, the mintty/PuTTY convention, since xterm paints into a canvas and the native menu has no Copy for it; with nothing selected the native menu is left alone). Rebindable via the registry. ⚠️ **Ctrl+W is deliberately NOT bound** (owner decision): it is delete-word in every shell and agent CLI, and as Close Session it killed sessions with no confirm; `close-session` keeps `bindings: []` and stays bindable (`test/ctrl-w-never-closes.test.ts`).
-
 ### Security
 
 **Full model: [`docs/security-architecture.md`](docs/security-architecture.md)** (network binding, auth pipeline, the tunnel caveat, file-serving hardening, supply-chain, instance isolation, recommended setups). **Layer-by-layer detail with the history behind each: [architecture-invariants#security-layers](docs/architecture-invariants.md#security-layers).**
@@ -433,10 +293,10 @@ One module per domain in `src/web/routes/` (plus a barrel; `ls src/web/routes/`
 - **API endpoint**: Types in `src/types/` domain file, route in `src/web/routes/*-routes.ts`. Return the `ApiResponse` envelope (`{ success: true, data }`; errors via `createErrorResponse()` with proper status code). Validate with Zod schemas in `schemas.ts`.
 - **SSE event**: Add to `src/web/sse-events.ts` + `SSE_EVENTS` in `constants.js`, emit via `broadcast()`, handle in `app.js` (`addListener(`)
 - **Session setting**: Add to `SessionState`, include in `session.toState()`, call `persistSessionState()`
-- **App setting**: decide per-device vs synced first. Per-device keys go in the `displayKeys` set in settings-ui.js and must NOT be added to `SettingsUpdateSchema` (it is `.strict()`). ⚠️ Anything in `PUT /api/settings` that acts on a setting (the `toggleService` watcher calls) must resolve from **`merged`** (persisted + incoming), never from the raw request body: a partial PUT omits keys it doesn't intend to change, and `body.x ?? default` turns every omission into "apply the default" and silently resets live services. Pinned by `test/routes/system-routes-settings-partial-put.test.ts`.
+- **App setting**: decide per-device vs synced first. Per-device keys go in the `displayKeys` set in settings-ui.js and must NOT be added to `SettingsUpdateSchema` (it is `.strict()`). ⚠️ Anything in `PUT /api/settings` that acts on a setting (the `toggleService` watcher calls) must resolve from **`merged`** (persisted + incoming), never from the raw request body: a partial PUT omits keys it doesn't intend to change, and `body.x ?? default` turns every omission into "apply the default" and silently resets live services. Pinned by `test/routes/system-routes-settings-partial-put.test.ts`. ⚠️ A synced setting is admin-only to write in multi-user mode (the PUT 403s a non-admin, whose browser does not send it): a new route that writes `settings.json` takes the same `isMultiUserMode() && !requireAdmin(req, reply)` gate.
 - **Hook event**: Add to `HookEventType`, add hook in `hooks-config.ts:generateHooksConfig()`, update `HookEventSchema`
 - **Mobile feature**: Add to relevant singleton, guard with `MobileDetection.isMobile()`. New header buttons must stay off phones (`test/mobile-header-buttons-policy.test.ts`).
-- **New test**: a test that constructs `WebServer` itself binds port 0 (`new WebServer(0, …)`, then `server.boundPort`; enforced by `test/test-ports-guard.test.ts`), and a raw Fastify/`ws` server listens on `port: 0` and reads `address().port`. Mobile tests (`test/mobile/**`, via `createTestServer(PORT)`) keep the fixed-port convention in `test/mobile/README.md` until the follow-up sweep, because that helper caches servers by port. Route tests use `app.inject()` (no port needed); see `test/routes/_route-test-utils.ts`.
+- **New test**: a test that constructs `WebServer` itself binds port 0 (`new WebServer(0, …)`, then `server.boundPort`), and a raw `http`/`net`/Fastify/`ws` server listens on `0` and reads `address().port`; mobile tests get theirs from `createTestServer()` and read `boundPort` too. `test/test-ports-guard.test.ts` refuses a fixed port in either. Route tests use `app.inject()` (no port needed); see `test/routes/_route-test-utils.ts`.
 
 **Validation**: Zod v4 (different API from v3). Define schemas in `schemas.ts`, use `.parse()`/`.safeParse()`.
 
@@ -446,6 +306,8 @@ All in `~/.codeman/`: `state.json` (sessions, settings, respawn, orchestrator, c
 
 **Generated top-level dirs** (all gitignored — don't edit or commit): `dist/` (esbuild output), `out/`, `coverage/`, `test-results/`, `tmp/`, `screenshots-echo-diag/`. The committed gesture bundle (`src/web/public/gesture/gesture-codeman.js`) IS tracked, but its runtime wasm/model assets (`src/web/public/gesture/wasm/`, `*.task`) are fetched and gitignored.
 
+**Videos and GIFs**: only GIFs a public page embeds are tracked (`docs/images/`), and `.gitignore` blocks `*.mp4`/`*.webm`/`*.mov`/`*.gif` everywhere else. Masters, promo cuts, drafts and captures live in the local, gitignored media library `pr/media/` (guide `pr/media/README.md`, generated `INVENTORY.md` via `node pr/media/inventory.mjs`); never leave a render in `tmp/` or `out/`. A GIF leaving `docs/images/` needs a hotlink check first (articles embed raw `master` URLs). → `docs/media-assets.md`
+
 ## Testing
 
 **`npm test` is the gate and is safe to run bare** — it runs `config/vitest.ci.config.ts`, exactly what CI runs, so local green means CI green.
@@ -465,7 +327,7 @@ npm run test:perf      # wall-clock benchmarks — need an otherwise idle machin
 npm run test:all       # literally everything; fails ~87 tests on a clean master here, which is why it is not the default
 ```
 
-⚠️ **`npm test` cannot see those suites**, so a change touching mobile/gesture/terminal-render behaviour needs the matching runner by hand — diff its FAIL list against master rather than reading it as pass/fail. That blind spot is what let two semantically-conflicting PRs merge green (see the on-screen-keyboard note above). The browser suite (not mobile or perf) also runs nightly in `.github/workflows/browser-suite.yml`, which catches a regression after it lands, not before.
+⚠️ **`npm test` cannot see those suites**, so a change touching mobile/gesture/terminal-render behaviour needs the matching runner by hand — diff its FAIL list against master rather than reading it as pass/fail. That blind spot is what let two semantically-conflicting PRs merge green (see the on-screen-keyboard note in `src/web/CLAUDE.md`). The browser suite (not mobile or perf) also runs nightly in `.github/workflows/browser-suite.yml`, which catches a regression after it lands, not before.
 
 ⚠️ **A file filter must match the runner.** `npm test -- test/mobile/keyboard.test.ts` matches nothing and exits GREEN having run zero tests, because the gate's config excludes that path — an excluded file needs its own runner (`npm run test:mobile -- <file>`, `npm run test:browser -- <file>`, `npm run test:perf -- <file>`). Vitest treats "no files matched a filter" as success, so read the file count, not just the colour.
 
@@ -475,7 +337,7 @@ Raw `npx vitest` skips the config (and with it `setup.ts`); always use `npm test
 
 **Tmux safety**: under vitest (`VITEST`), `TmuxManager` no-ops ALL shell commands (`IS_TEST_MODE` in `src/tmux-manager.ts`), docker IO is no-op'd likewise, and `Session` spawns an echo PTY (`TEST_PTY_SCRIPT`) instead of attaching tmux. `test/setup.ts` gives each file a temp `HOME`/`USERPROFILE` and strips `CODEMAN_PASSWORD`/`CODEMAN_USERNAME`, `CODEMAN_GESTURE` and `CODEMAN_INSTANCE`/`CODEMAN_DATA_DIR`/`CODEMAN_TMUX_SOCKET` (pinned by `test/test-env-isolation.test.ts`). ⚠️ `CODEMAN_DATA_DIR` overrides the temp HOME, so never drop its strip; strip `CODEMAN_INSTANCE` in the setup file, never in a hook (captured at first import). ⚠️ Delete case trees only via `safeRmHomeTree()`. ⚠️ Raw `npx vitest` without `--config` skips `setup.ts` and its isolation. → [architecture-invariants#test-isolation-tmux-docker-and-home](docs/architecture-invariants.md#test-isolation-tmux-docker-and-home)
 
-**Ports**: a test that constructs `WebServer` binds port 0 and reads `server.boundPort`; `test/test-ports-guard.test.ts` fails any other port outside its shrink-only legacy list. A raw Fastify/`ws` server listens on `port: 0` and reads `address().port`. The mobile suite keeps fixed ports (3150+, per `test/mobile/README.md`) until the follow-up sweep, since `createTestServer()` caches servers by port. Never 3000 (the live instance).
+**Ports**: every in-process test server binds port 0 (the one exception, `test/codex-predictive-echo.test.ts`, starts a separate lab server process on 3222). A test that constructs `WebServer` reads `server.boundPort` (mobile tests: `createTestServer()`, then `server.boundPort`); a raw `http`/`net`/Fastify/`ws` server listens on `0` and reads `address().port`. `test/test-ports-guard.test.ts` fails a `WebServer` built on any other port and a raw `listen` on a number or `…PORT` constant. A test that needs a port nothing listens on binds 0, reads it and closes (`closedPort()` in `test/daemon-control.test.ts`), never "the server's port + 1". Never 3000 (the live instance).
 
 ⚠️ **Browser tests can pass vacuously on mobile input paths.** Two traps, both hit on 2026-07-27 while fixing the phone Enter button: **(1)** driving input with `app.sendInput('…')` writes PAST the `LocalEchoOverlay`, so `pendingText` stays empty and any overlay bug is invisible — type with `page.keyboard.type()` instead; **(2)** headless Chromium reports `MobileDetection.isTouchDevice()` **false even with `hasTouch: true`**, so `_localEchoEnabled` is off and the local-echo branch never executes. Force it (`app._localEchoEnabled = true`) or the test proves nothing. Assert on real state (`app._localEchoOverlay.pendingText`, plus `tmux -L codeman capture-pane -p -t <pane>` for what actually reached the PTY), not on HTTP 200.
 
@@ -493,7 +355,7 @@ curl -sk https://localhost:3000/api/subagents | jq # Background agents
 cat ~/.codeman/state.json | jq                     # Persisted state
 ```
 
-Legacy screenshots (deprecated): `GET/POST /api/screenshots` still read and write `~/.codeman/screenshots/`, but the upload page that fed them is gone, they log a one-time deprecation warning, and they are removed in a later MAJOR, after at least one MINOR release that carries the warning (`docs/versioning-policy.md`). To hand a file to an agent use `POST /api/sessions/:id/paste-image`.
+Legacy screenshots (deprecated): `GET/POST /api/screenshots` still read and write `~/.codeman/screenshots/`, but the upload page that fed them is gone, they log a one-time deprecation warning, and they are removed in a later MAJOR, after at least one MINOR release that carries the warning (`docs/versioning-policy.md`). To hand a file to an agent use `POST /api/sessions/:id/paste-image` (see Prompt uploads).
 
 ## Performance & Limits
 

README.md

@@ -24,7 +24,7 @@
 </p>
 
 <p align="center">
-  <img src="docs/images/tiles-crt-stats-20261010.gif" alt="Codeman tile grid: six live agents (DeepSeek Harness, Claude Code, Pi, Codex, OpenCode and a shell) powering on and off with the CRT animation, with the live header strip showing CPU, memory and Claude plan usage" width="800">
+  <img src="docs/images/tiles-crt-stats-1600-20261010.gif" alt="Codeman tile grid: six live agents (DeepSeek Harness, Claude Code, Pi, Codex, OpenCode and a shell) powering on and off with the CRT animation, with the live header strip showing CPU, memory and Claude plan usage" width="800">
 </p>
 
 **Codeman** is a self-hosted mission control for AI coding agents. It spawns Claude Code, OpenCode, Codex, Antigravity, Gemini, Pi, Grok, DeepSeek Harness, or OMP inside persistent tmux sessions, streams the real terminal to any browser, and keeps agents productive after you walk away: it re-prompts on idle, resumes when a usage limit resets, runs scheduled jobs, and shows every background agent working in real time.
@@ -46,6 +46,7 @@ The installer asks before every system change, and re-running the same line upda
 - **Truly phone-friendly** - a [touch-optimized terminal](#mobile-optimized-web-ui) with instant local echo, QR login, swipe navigation, and push notifications
 - **Runs while you sleep** - [idle detection + respawn cycling](#respawn-controller) and auto-resume when a subscription limit resets, for 24+ hour unattended runs
 - **See your agents think** - [live floating windows](#live-agent-visualization) for every subagent and teammate, with real-time transcripts
+- **Agents can drive it too** - the [codeman skill](#the-agent-skill-start-here) (for every agent CLI), [`codeman agent`](#codeman-agent--session-to-session-verbs-in-every-cli-mode) and an [MCP server](#mcp-server-codeman-mcp-drive-codeman-from-any-mcp-client) let one agent start workers in any CLI, prompt them, wait for them and collect their answers
 - **Nothing gets lost** - tmux persistence across restarts and network drops, exactly-once input delivery, full-scrollback replay
 - **Self-hosted and private** - loopback-only by default, MIT licensed, no telemetry, runs entirely on your machine
 
@@ -212,6 +213,8 @@ The most responsive AI coding agent experience on any phone. Full xterm.js termi
 
 - **Keyboard accessory bar** — `/init`, `/clear`, `/compact` quick-action buttons above the virtual keyboard; destructive commands require a double-press to confirm, so you never fire one by accident; on Codex sessions the bar also shows `⇧←` / `⇧→` (Shift+Left / Shift+Right: edit the last queued message / return through the prompt stack)
 - **Dedicated Enter button** — replays the keypress through the terminal, so text buffered by local echo is flushed first rather than stranded
+- **Attach button**: the picture icon beside the mic opens the system photo and file sheet without raising the keyboard. Photos and camera-roll videos (mp4, mov, webm, up to 1 GB) upload into the session's folder and their path lands on your prompt, unsent
+- **Paste in a shell**: a multi-line paste goes in as one block and waits for your Enter, instead of running line by line as it arrives
 - **Swipe navigation & smart keyboard handling** — swipe left/right to switch sessions; toolbar and terminal shift up when the keyboard opens (`visualViewport` API)
 - **Built for phones** — safe-area insets for notch and home indicator, 44px touch targets, bottom-sheet case picker, native momentum scrolling; on a folding phone (iPhone Duo) dialogs stay clear of the hinge, and opening or closing the device is never mistaken for the keyboard
 
@@ -264,18 +267,18 @@ Hit start — Codeman spawns the CLI via a real PTY and streams it to your brows
 
 ### 3. Read the dashboard
 
-- **Tabs (top)** — one per session. `Alt+1`-`9` to jump, `Ctrl+Tab` for next, drag to reorder (tab order syncs across your devices). Prefer a list? **App Settings → Appearance → Tabs** moves it into a left sidebar with a filter box (`Alt+B` collapses it) or a vertical rail whose rows sort by activity: blocked on you first, then longest running, then most recently quiet.
+- **Tabs (top)** — one per session. `Alt+1`-`9` to jump, `Ctrl+Tab` for next, drag to reorder (tab order syncs across your devices). Prefer a list? **App Settings → Appearance → Tabs** moves it into a left sidebar with a filter box (`Alt+B` collapses it) or a vertical rail whose rows sort by activity: blocked on you first, then longest running, then most recently quiet. The rail also has a **Search sessions** box, named **tab groups** (a group's menu starts a **New session** straight into it), and a **Focus** section of shortcuts to the tabs you keep coming back to.
 - **Terminal (center)** — a real `xterm.js` terminal; full TUIs render correctly. Type directly and press **Enter** to send. `Shift+Enter` inserts a newline.
 - **Side panels** — Respawn, Orchestrator, Cron, Subagents, Settings (toggled from the toolbar).
 
 ### 4. Talk to the agent
 
 - **Type prompts** straight into the terminal — input is delivered exactly-once even across reconnects (a dropped link never loses or double-sends a prompt).
-- **Paste or drag-and-drop images** directly into the session.
+- **Paste, drop or attach images, videos and documents** (PDF, Office files, zip and plain text) directly into the session. They are saved in a hidden `.codeman-uploads/` folder in the workspace (it ignores itself in git and is swept after 7 days), and the file's path lands on your prompt, unsent.
 - **Voice input** — `Ctrl+Shift+V` (Deepgram Nova-3, or this machine's Claude Code login with no API key; auto-silence stop).
-- **Attachments** — register external files/docs and preview Office/PDF inline; any file path an agent prints is clickable, in the terminal and in the chat view.
+- **Attachments** — register external files/docs and preview Office/PDF inline; any file path an agent prints is clickable, in the terminal and in the chat view, including `~/` and relative paths and paths broken across lines. An `.html` report opens as a real page, Markdown opens rendered, and a folder opens as a grid of its files.
 - **When it needs you** — the tab turns yellow (waiting for input) or red (a question is blocking). The **Approvals Inbox** _(opt-in)_ queues every pending prompt across sessions, answerable from the header bell or the phone home screen, and 🧠 **Read My Mind** _(opt-in)_ drafts your next prompt from the case's goals and recent work.
-- **Copy what you see** — `Shift+drag` selects text even while the CLI owns the mouse, right-click copies it, and Auto Copy _(opt-in)_ copies a selection the moment you release it.
+- **Copy what you see** — `Shift+drag` selects text even while the CLI owns the mouse, right-click copies it, and Auto Copy _(opt-in)_ copies a selection the moment you release it. A program that copies with OSC 52 (Claude Code's `/copy`, a tmux yank, vim) puts the text on your browser's clipboard too.
 
 ### 5. Make it autonomous
 
@@ -294,7 +297,7 @@ Hit start — Codeman spawns the CLI via a real PTY and streams it to your brows
 
 ### 7. Operate & maintain
 
-- **App Settings** — model, effort, permission startup mode, theme/skin, terminal font family and weight, entrance animations, notifications, display toggles, per-CLI options, a synced custom display name, and per-device English/Simplified Chinese UI language.
+- **App Settings** — model, effort, permission startup mode, theme/skin, terminal font family and weight, entrance and tile animations (their own **Animations** section), notifications (with adjustable toast and browser-notification display times), display toggles, per-CLI options, a synced custom display name, and per-device English/Simplified Chinese UI language. **Apply** saves without closing Settings.
 - **Run it in the background** — `codeman web -d` detaches from your shell (`--status`, `--stop`); `codeman service install` makes it a systemd user unit / macOS LaunchAgent that survives reboots. Both verify the server actually answers before reporting success, and both refuse to start a second server on one data dir. See [Keep it running in the background](#quick-start---installation).
 - **Self-update** — git-clone installs update in place from **App Settings → System → Updates**.
 - **Deploy your own changes** — see [Development](#development).
@@ -387,7 +390,14 @@ Run **20 parallel sessions** with full visibility — real-time xterm.js termina
   <img src="docs/images/tile-grid-20261009.gif" alt="Tile grid: the Tiles button opens six live sessions side by side (DeepSeek Harness, Claude Code, Codex, a shell, OpenCode and Pi), and a second click returns to a single session" width="800">
 </p>
 
-Watch and drive up to **six sessions side by side** in one window. Click **Tiles** in the header (or press `Ctrl+Shift+G`) and your sessions open as a grid of live terminals: every tile takes your keystrokes and shows its agent's logo, model and state in its header. Right-click **Tiles** to choose 2, 4 or 6 tiles, and drag a tile by its header to move it. Click **Tiles** again to return to a single session; the grid is remembered for next time. Desktop only (a window about 1180px wide or more). Full guide: [Tile Grid](docs/wiki/Tile-Grid.md).
+Watch and drive up to **six sessions side by side** in one window. Click **Tiles** in the header (or press `Ctrl+Shift+G`) and your sessions open as a grid of live terminals: every tile takes your keystrokes and shows its agent's logo, model and state in its header. Right-click **Tiles** to choose 2, 4 or 6 tiles, and drag a tile by its header to move it. Click **Tiles** again to return to a single session. Desktop only (a window about 1180px wide or more). Full guide: [Tile Grid](docs/wiki/Tile-Grid.md).
+
+- **Opens what matters**: with no grid arranged yet, it fills the tiles with the sessions that are working first, then the ones waiting on you, then the most recently used, and always includes the one you are on.
+- **Keeps your layout**: once you arrange a grid (which session sits where, empty cells, tile count, divider sizes, the focused and zoomed tile), the Tiles button brings it back exactly, and so does a page reload while it was open.
+- **Tile animations** _(opt-in)_: **App Settings → Animations → Tile Animations** makes the grid open with a show: **CRT** (each tile switches on as a hot line, as in the GIF at the top), **Fly from tab**, **Deal**, **Beam down**, **Cascade**, **Pop** or **Soft**. Nothing moves under reduced motion.
+- **Scrolls like the main terminal**: the mouse wheel scrolls a fullscreen Claude conversation inside its tile, a Codex tile pages its transcript, and `Shift+wheel` scrolls the tile's local history.
+- **Detach Tiles** _(beta, opt-in)_: with **App Settings → Appearance → Detach Tiles** on, drag a tile's header out of the browser and it pops out into its own window; drop it on another Codeman window's tiles and it moves there.
+- **Hands-free**: with the [gesture overlay](#more-features) on, pinch a tile and carry it onto another tile to swap them, onto an empty cell to move it, or carry a tab onto the grid to add it.
 
 ### Persistent Sessions
 
@@ -464,10 +474,15 @@ PTY Output → 16ms Server Batch → DEC 2026 Wrap → SSE → Client rAF → xt
 - **Gesture control** _(opt-in)_ — a MediaPipe hand-tracking overlay to grab/drag session windows and pinch buttons, hands-free. Enable with `CODEMAN_GESTURE=1` + App Settings → Terminal & Input
 - **Multi-monitor span** _(macOS)_ — one click opens a browser window maximized across all displays, so floating agent/gesture panels can cross the physical seam
 - **File Viewer button** _(opt-in)_ — a header button that toggles the built-in file browser panel with one tap; enable under App Settings → Header & Panels → Header buttons
+- **HTML reports open as pages**: click an `.html` path an agent printed (a test report, a coverage page, a dashboard) and the File Viewer shows it as a real page, with the scripts, styles, images and data files from its folder; the **Page** pill switches to the source and its Edit button. The page runs sandboxed with an opaque origin, so it cannot read Codeman's cookies, call Codeman's API or read files outside its own folder. Network access stays open, so a dashboard that loads Chart.js or fonts from a CDN works as it would from disk. Broad folders (`/`, your home, `/tmp`) are refused, so give a report a folder of its own. A `.md` path clicked in the terminal opens rendered too
+- **Clickable paths and the folder grid**: paths written as `~/…`, relative to the session folder, or broken across terminal lines are clickable in the terminal and the Response Viewer, and so are folders: a folder opens a grid of its files (renders, screenshots, reports) that you page through with the arrow keys
+- **OSC 52 clipboard**: a program in a session that sets the clipboard with OSC 52 (Claude Code's `/copy`, a tmux yank, vim) puts the text on your browser's clipboard, in the main terminal and in every tile. Clipboard read requests are refused, and replaying a buffer never copies
+- **MCP server sync** _(opt-in)_: copies the user-level MCP servers each installed CLI has (Claude, Codex, Gemini, OpenCode, Antigravity, and GitHub Copilot CLI) into the others' own config files. Additive only, a server you switched off is never copied, **Preview** shows the changes first, and the previous file is kept as `.codeman-bak`. App Settings → Agents & CLIs; admin only in multi-user mode
 - **CJK / IME input** — full composition support for Chinese / Japanese / Korean, with Ctrl- and Alt-modified navigation keys passed through to the CLI
 - **Plan usage in the header** — live Claude subscription usage (the 5-hour and weekly windows) from a statusline exporter Codeman hands to `claude` at spawn and never writes into your settings files, plus Codex limits from its own app-server; per device, on for desktops and off for phones
-- **Session list, your way** — the header strip, a left sidebar with a filter box, or a vertical rail whose detailed rows carry created and state stamps and sort by activity; the phone home screen and the desktop home rail use the same order
-- **Terminal looks** — seven skins, four of them light, per-device font family and weight (the bundled JetBrains Mono covers weights 100 to 800), and opt-in entrance animations for tabs, agent windows, the terminal pane and connection lines
+- **Prompt-cache readout**: each Claude session shows whether its prompt cache is warm (`cache:until HH:MM`) or cold, in the footer (when you have no statusline of your own) and in Session Options → Respawn, with the cost of a re-cache and a note after a compaction. Display only; it survives a restart
+- **Session list, your way** — the header strip, a left sidebar with a filter box, or a vertical rail whose detailed rows carry created and state stamps and sort by activity; the phone home screen and the desktop home rail use the same order. On the rail, **Search sessions** narrows the list to the tabs that match, across every group, and a tab with an alert stays visible even when its name does not. Named **tab groups** collapse, reorder and take drags; a group's menu has **New session**, and a worker an agent spawns lands right after its parent (**Spawned tabs follow their parent**, off by default, keeps it moving with the parent). **Add to Focus** pins a shortcut to any tab in a section at the top of the rail, and **CLI Logos on Tabs** can be switched off per device
+- **Terminal looks** — seven skins, four of them light, per-device font family and weight (the bundled JetBrains Mono covers weights 100 to 800), and opt-in animations for tabs, agent windows, the terminal pane, connection lines and [tile-grid tiles](#tile-grid), all in App Settings → Animations (with a lab to try them)
 - **OS notifications & hostname-aware titles** — desktop alerts and tab titles are prefixed `codeman:<host>` so multi-host setups stay unambiguous
 
 ---
@@ -499,6 +514,7 @@ Point a case at another machine and run the agent **there**, over SSH, with the
 - **Shared sessions**: several clients can attach the same remote session at different window sizes without clamping each other; discovery shows a "shared" badge with the client count.
 - **Injection-safe**: every ssh command line flows through a single shell-escaping builder, and host/path/identity fields are schema-guarded.
 - **Files too**: previews, downloads and text reads in a remote case go over the same ssh connection (one `realpath` + `stat` probe, then a streamed `cat`, `Range` seeking included), so a clicked path opens the file on the machine the agent is on. Nothing is copied to the Codeman host; editing and Office previews answer a clear 400 instead of a misleading 404.
+- **Full scrollback**: loading a remote session's history reads it from the remote tmux in one bounded ssh round trip, so a session keeps its whole scrollback after a reattach or a tab switch, with no rows lost where the history meets the live screen.
 
 Set it up under **New Case → Remote** (host, user, identity file, optional jump host). Full design: [`docs/remote-sessions.md`](docs/remote-sessions.md).
 
@@ -669,7 +685,7 @@ These run for **every** request — before auth, even on the default no-password
 ### Input, files & headers
 
 - **Schema-validated inputs** — every API body is checked with Zod v4 schemas; a `CLAUDE_CODE_*` / `OPENCODE_*` / `CODEX_*` / `ANTIGRAVITY_*` / `GEMINI_*` / `GOOGLE_*` / `PI_*` / `GROK_*` / `XAI_*` / `DSH_*` / `DEEPSEEK_*` / `OMP_*` env-prefix allowlist gates which settings each CLI can receive, and the keys that could redirect a CLI's traffic (base URLs, config homes) are clamped for non-admin users
-- **Path containment** — file routes `realpath` before boundary checks (no TOCTOU); `..`, absolute paths, and symlinks resolving outside the working dir are rejected. Caps: 10 MB text preview / 2 GB raw & download (`CODEMAN_MAX_DOWNLOAD_BYTES`; bodies stream and answer `Range` requests, so the cap is a sanity bound rather than memory protection); `/api/download` blocklists sensitive paths (`.env`, `*credentials*`, `~/.ssh/`, `.aws/credentials`). SVG/HTML is served `octet-stream` + `nosniff` + attachment so it downloads rather than executes
+- **Path containment** — file routes `realpath` before boundary checks (no TOCTOU); `..`, absolute paths, and symlinks resolving outside the working dir are rejected. Caps: 10 MB text preview / 2 GB raw & download (`CODEMAN_MAX_DOWNLOAD_BYTES`; bodies stream and answer `Range` requests, so the cap is a sanity bound rather than memory protection); `/api/download` blocklists sensitive paths (`.env`, `*credentials*`, `~/.ssh/`, `.aws/credentials`). SVG/HTML is served `octet-stream` + `nosniff` + attachment so it downloads rather than executes. The one place HTML renders is the File Viewer's page view: a separate `/html-view/<capability>/` address scoped to the file's own folder, served under `Content-Security-Policy: sandbox` without `allow-same-origin`, so the page has an opaque origin and can never read Codeman's cookies or drive its API; hidden folders and broad roots (`/`, home, `/tmp`) are refused, and the capabilities are revoked on logout
 - **Security headers** — `Content-Security-Policy` (`default-src 'self'`, every exception enumerated), `X-Content-Type-Options: nosniff`, `X-Frame-Options: SAMEORIGIN`, HSTS over HTTPS, and CORS reflected **only** for `localhost` / `127.0.0.1` / `::1`
 
 ### Supply chain & isolation
@@ -703,7 +719,7 @@ The web UI remains the primary surface; see **[docs/tui.md](docs/tui.md)** for t
 
 | Shortcut                        | Action                                                        |
 | ------------------------------- | ------------------------------------------------------------- |
-| `Ctrl/Cmd+W`                    | Kill active session                                           |
+| `Ctrl+W`                        | Not a Codeman shortcut: it reaches the terminal (delete the previous word). **Close Session** has no default key and can be bound in App Settings → Shortcuts |
 | `Ctrl/Cmd/Option+K`             | Find open session or start a new one                          |
 | `Ctrl/Cmd+Tab`                  | Next session                                                  |
 | `Alt/Option+[` / `Alt/Option+]` | Previous / next session                                       |
@@ -716,6 +732,10 @@ The web UI remains the primary surface; see **[docs/tui.md](docs/tui.md)** for t
 | `Ctrl/Cmd+L`                    | Clear terminal                                                |
 | `Ctrl+Shift+R`                  | Restore terminal size                                         |
 | `Ctrl+Shift+V`                  | Toggle voice input                                            |
+| `Ctrl+Shift+G`                  | Open or close the tile grid                                   |
+| `Alt+Shift+Arrows`              | Focus the tile to the left, right, above or below             |
+| `Ctrl+Shift+Arrows`             | Move the focused tile one place (into an empty cell, or swap) |
+| `Alt+Shift+Enter`               | Zoom the focused tile, or restore the grid                    |
 | `Ctrl/Cmd +` / `-`              | Font size                                                     |
 | `Ctrl/Cmd+?`                    | Keyboard help                                                 |
 | `Shift+Enter`                   | Insert newline (sent to terminal)                             |
@@ -733,23 +753,23 @@ For AI agents and automation that control Codeman without a browser: an agent th
 
 ### The agent skill (start here)
 
-Everything in this section also ships as a **Claude Code skill** in [`skills/codeman`](skills/codeman/SKILL.md). Install it once and you never paste API docs into a prompt again. You ask for what you want in plain English, and the agent already sitting inside a Codeman session loads the recipes and drives the API itself.
+Everything in this section also ships as an **agent skill**, [`skills/codeman`](skills/codeman/SKILL.md), in the `SKILL.md` format Claude Code, Codex, Gemini CLI, OpenCode, Grok, OMP and most other agent CLIs read. Install it once and you never paste API docs into a prompt again. You ask for what you want in plain English, and the agent already sitting inside a Codeman session loads the recipes and drives the API itself. Full guide, with every install route, the recordings and what it will and will not do on its own: [The Codeman Skill](https://github.com/Ark0N/Codeman/wiki/Codeman-Skill).
 
 <p align="center">
-  <a href="docs/images/codeman-skill-20261010.png"><img src="docs/images/codeman-skill-20261010.gif" alt="A real codeman skill run: one plain-English request to a lead session, three Claude Code workers opening as new tabs, and lineage lines from the lead to every worker" width="900"></a>
+  <a href="docs/images/codeman-skill-crt-20261010.png"><img src="docs/images/codeman-skill-crt-20261010.gif" alt="A real codeman skill run: one short prompt typed into Claude Code, the tile grid powering on, then a DeepSeek Harness worker on a local qwen model and a Claude Code worker powering on as new tiles, with lineage lines from the lead to both" width="900"></a>
 </p>
 
 #### Step 1: install it
 
 | How            | Command                                                    | Scope                                                                                      |
 | -------------- | ---------------------------------------------------------- | ------------------------------------------------------------------------------------------ |
-| Skills CLI     | `npx skills add Ark0N/Codeman --skill codeman -g`          | Global, works for any skills-aware agent                                                   |
+| Skills CLI     | `npx skills add Ark0N/Codeman --skill codeman -g`          | Global, for any skills-aware agent ([skills.sh](https://skills.sh)); the way to give it to pi |
 | Claude Code plugin | `/plugin marketplace add Ark0N/Codeman` then `/plugin install codeman@codeman` | Global, through Claude Code's plugin manager; `/plugin update codeman` follows releases. Pick this OR a `codeman skill install`, not both: a Claude Code with both lists the skill twice (`codeman` and `codeman:codeman`) |
 | Bundled CLI    | `codeman skill install`                                    | Global (`~/.claude/skills/codeman`), for npm installs that never cloned the repo            |
-| Bundled CLI    | `codeman skill install --case <name>`                      | One case only                                                                              |
-| Web UI         | App Settings → Agents & CLIs → Claude → **Agent Skill**    | Auto-injects into each case on Claude session create (`agentSkillEnabled`, SYNCED, default off) |
+| Bundled CLI    | `codeman skill install --case <name> [--mode <mode>]`      | One case only; `--mode codex` (or any other run mode) writes it where that CLI looks for skills |
+| Web UI         | App Settings → Agents & CLIs → All agent CLIs → **Codeman Agent Skill** | Auto-injects into each case on session create, where that session's CLI looks: `.claude/skills/` for Claude Code, `.agents/skills/` for Codex, Gemini CLI, OpenCode, Antigravity, Grok, DeepSeek and OMP; not pi, whose project skills trigger a trust prompt (`agentSkillEnabled`, SYNCED, default off) |
 
-`codeman skill uninstall [--case <name>]` reverses the CLI installs, and never touches a `skills/codeman` you wrote yourself.
+`codeman skill uninstall [--case <name>]` reverses the CLI installs (with `--case`, every Codeman-managed copy in that case, whichever CLI's directory it sits in), and never touches a `skills/codeman` you wrote yourself.
 
 #### Step 2: ask for things
 
@@ -763,9 +783,19 @@ That is the entire interface. No curl, no endpoint names, no session ids. These
 | _"Have a claude worker on `refactor-auth` summarize `src/session.ts`, then close it."_     | Spawns, runs the readiness ladder (first-run trust dialog included), send-and-wait, reads the clean transcript answer, deletes. |
 | _"Watch session w4 and tell me if it gets stuck on a permission prompt."_                  | Blocks on the `blocked` signal and surfaces the question to **you**. It never answers another session's prompt itself. |
 
+One short line is enough. Here a Claude Code lead gets `use codeman to start 5 claude workers, one per TODO item`, and five real Claude Code workers power on, shown side by side in the tile grid:
+
+<p align="center">
+  <img src="docs/images/codeman-skill-grid-20261015.gif" alt="A real codeman skill run in the tile grid: the lead Claude Code session powers on as the only tile, the prompt use codeman to start 5 claude workers, one per TODO item is typed, then five Claude Code worker sessions power on as new tiles one after another with the CRT animation, and all six agents work at once" width="900">
+</p>
+
 #### Step 3: nothing
 
-The agent deletes every session it started. Watch the tabs appear and disappear in the dashboard while it works.
+The agent deletes every session it started. Watch the tabs appear and disappear in the dashboard while it works. With eight workers, each one gets its own tab, and a lineage line runs back to the lead that spawned it:
+
+<p align="center">
+  <img src="docs/images/codeman-skill-tabs-20261015.gif" alt="A real codeman skill run in the normal view: the prompt use codeman to start 8 claude workers, one per TODO item is typed, eight worker tabs power on left to right while lineage lines draw from the lead to each one, then clicking a worker tab shows a live Claude Code worker receiving its task" width="900">
+</p>
 
 #### A real run, start to finish
 
@@ -802,7 +832,7 @@ Every recipe in there was verified against a live server, and the comments recor
 - **It self-gates.** Outside a Codeman session (`CODEMAN_MUX` unset) the skill refuses to act and does not guess an API URL, so a global install costs an unrelated Claude Code session nothing.
 - **It is deliberately conservative.** Unprompted, it may only spawn sessions, prompt them, and delete ones **it created in that same conversation, by exact id**, through a fail-closed guard that refuses to delete the agent's own session. Deleting a case (which erases a real directory of your code), bulk kills, respawn/ralph/cron/orchestrator changes and settings writes all require you to ask, naming the target.
 
-⚠️ Turning `agentSkillEnabled` back off **does not remove already-injected copies** (a create-time sweep would yank the skill out from under other live sessions sharing that `.claude/` dir). Remove them per case with `codeman skill uninstall --case <name>`.
+⚠️ Turning `agentSkillEnabled` back off **does not remove already-injected copies** (a create-time sweep would yank the skill out from under other live sessions sharing that workspace). Remove them per case with `codeman skill uninstall --case <name>`.
 
 ---
 
@@ -827,7 +857,7 @@ When a CLI runs in a Codeman-managed session, these environment variables are se
 4. **Response envelope.** Most endpoints return `{ "success": true, "data": … }` (errors: `{ "success": false, "error", "errorCode" }`). A few legacy GETs return bare bodies — **handle both** (`body.data ?? body`).
 5. **`/api/v1/*`** is a stable alias of `/api/*`.
 6. **Wait instead of polling, and don't treat a timeout as an error.** The wait endpoints answer with HTTP `200` and `wait.timedOut: true` when nothing happened in time, so loop over short waits (60s is the default) rather than issuing one long call, because tunnels cut idle connections. `wait.timeoutMs` tells you the timeout the server actually applied after clamping (600s ceiling).
-7. **Only `claude` and `deepseek` sessions emit `stop` and `blocked`.** Those two come from hooks (Claude Code's own, and the DeepSeek Harness status bridge); `shell` and the other external CLIs (opencode/codex/gemini/antigravity/pi/grok/omp) accept only `idle`, `working` and `exit`. Asking for `stop` explicitly on those is a `400`; omitting `until` is always safe. ⚠️ On a `shell` session `idle` fires **once**, at startup, and never again, so send-and-wait there can only time out; synchronize hook-less sessions with a `wait-output` marker.
+7. **Only `claude` and `deepseek` sessions emit `stop`; `codex` emits `blocked` too.** Those come from hooks (Claude Code's own, and the DeepSeek Harness status bridge), except codex's `blocked`, which Codeman reads off its screen while an approval, question or trust dialog is up (answer it in the terminal). `shell` and the other external CLIs (opencode/gemini/antigravity/pi/grok/omp) accept only `idle`, `working` and `exit`, and codex adds `blocked`. Asking for `stop` explicitly on those is a `400`; omitting `until` is always safe. ⚠️ On a `shell` session `idle` fires **once**, at startup, and never again, so send-and-wait there can only time out; synchronize hook-less sessions with a `wait-output` marker.
 8. **Nothing reports "ready", so wait for it explicitly.** A new session answers `{"signal":"exit","immediate":true}` (that means *not started*, not *crashed*) until its PID exists, and a `claude` worker in a fresh case then sits on the CLI's trust dialog. Prompt it there and the wait resolves on `idle` in ~2s looking exactly like a finished turn, while the text sits stuck in the dialog. Recipe 2b below is the sequence that avoids it.
 
 ### Recipes
@@ -944,6 +974,74 @@ codeman tui --list                       #      numbered session list (plain tex
 codeman tui 3                            #      attach to session 3 of that list
 ```
 
+### `codeman agent` — session-to-session verbs in every CLI mode
+
+The skill above drives the API with `curl` and `jq` from a preamble Codeman seeds per session. Where it does not reach (a `pi` or `shell` session, a machine without `jq`), the same environment is still there (`CODEMAN_MUX=1`, `CODEMAN_SESSION_ID`, `CODEMAN_API_URL` are exported into every pane), so `codeman agent` packages the same verbs as commands. It is a thin client over the endpoints listed under [API](#api): no new route, no new transport, auth and ownership unchanged. One line in a case's `AGENTS.md` is enough: *"other sessions: `codeman agent --help`"*.
+
+```bash
+codeman agent ls                                          # sessions; * marks this one
+SID=$(codeman agent spawn scratch-1 --mode claude)        # quick-start + wait for the composer where the mode has a ready mark
+codeman agent send "$SID" 'review src/, then say DONE' --until stop,exit --timeout 300000   # --wait = default signal set
+codeman agent read "$SID"                                 # last answer (as the server reads it for that mode)
+codeman agent read "$SID" --tail 3000                     # terminal tail, ANSI stripped (every mode)
+codeman agent send "$SID" 'run the tests, then print WORKDONE followed by _4711'   # hook-less modes (opencode, pi, …): ask for the marker in halves …
+codeman agent wait "$SID" --match WORKDONE_4711           # … and wait on the joined form, which the prompt's echo never contains
+codeman agent interrupt "$SID"                            # a bare ESC, conversation intact
+codeman agent rm "$SID"                                   # any session except this one
+```
+
+Rules the commands enforce rather than document: they refuse outside a Codeman session and never guess a URL; `send` transmits printable text plus Enter only (a control byte such as `Ctrl+C` is `app_exit` in opencode — ESC exists solely as `interrupt`, which never appends Enter); `rm` refuses an empty id, an unprovable self id and a prefix match in either direction. Ids may be the 8-character prefixes `ls` prints (resolved through the list; an ambiguous prefix refuses, anything shorter than 8 characters refuses on every verb). A prompt that starts with `-` goes after `--` (`send "$SID" -- "- fix the bug"`). The echo of the prompt you sent is output too: a `--match` marker that appears verbatim in the prompt matches at once, before the worker has done anything, so the prompt asks for it in halves. A remote session whose host is asleep answers a fire-and-forget `send` with `buffered` (Codeman wakes the host and types the prompt once the pane is back) or `dropped` (over the wake buffer's cap, nothing will be typed: exit `1`). Exit codes: `0` delivered/matched/signal, `1` error, `2` timeout, `3` the worker exited or the wait ended without an answer (`delivered:false`, `ended:true`), `4` refused. `spawn` prints the id alone on stdout (prose goes to stderr), so `SID=$(…)` captures exactly the id. `--json` prints the envelope's `data` for every verb. `--until stop` on a mode without hook signals is the server's 400, passed through — the marker path (`--match`) is the answer there, exactly as for the skill.
+
+### MCP server (`codeman mcp`): drive Codeman from any MCP client
+
+`codeman mcp` is a [Model Context Protocol](https://modelcontextprotocol.io) server on stdio, so any MCP client (Claude Code, Claude Desktop, Codex, Gemini CLI, Cursor, ...) can start Codeman sessions in any installed agent CLI, prompt them, wait for them, read their answers and clean up, from inside Codeman or from outside it. Like `codeman agent`, it is a client of the HTTP API: no new route, auth and ownership unchanged.
+
+```bash
+# Claude Code
+claude mcp add codeman -e CODEMAN_API_URL=https://127.0.0.1:3000 -- codeman mcp
+```
+
+```json
+{
+  "mcpServers": {
+    "codeman": {
+      "command": "codeman",
+      "args": ["mcp"],
+      "env": { "CODEMAN_API_URL": "https://127.0.0.1:3000" }
+    }
+  }
+}
+```
+
+```toml
+# Codex (~/.codex/config.toml): raise the 60 s default tool timeout for long waits
+[mcp_servers.codeman]
+command = "codeman"
+args = ["mcp"]
+env = { CODEMAN_API_URL = "https://127.0.0.1:3000" }
+env_vars = ["CODEMAN_MUX", "CODEMAN_SESSION_ID"]  # Codex passes MCP servers a minimal env; these let it know the Codeman session it runs in
+tool_timeout_sec = 120
+```
+
+The JSON form is what Claude Desktop, Cursor and Gemini CLI take. Add `"CODEMAN_PASSWORD"` to `env` when the server has a password (or let it read the data dir's `.env`). Without the `codeman` command on PATH (a git checkout run without the installer), use `node /path/to/Codeman/dist/index.js mcp`.
+
+| Tool | Does |
+| ---- | ---- |
+| `list_sessions` | Every session you can see, with its run mode, status, folder, and whether these tools may prompt it |
+| `start_session` | A new worker in a case (a fresh scratch folder, or an existing case) in any installed CLI; waits until it can take a prompt |
+| `send_prompt` | Types a one-line prompt and presses Enter; optionally waits for the end of the turn |
+| `wait_for_session` | Waits for a signal (`stop`, `blocked`, `idle`, `working`, `exit`) or a literal marker in the output |
+| `read_output` | The last answer, the whole conversation, or the last bytes of the terminal |
+| `interrupt_session` | One Escape: stops the current turn, keeps the conversation |
+| `delete_session` | Closes a worker (its case folder stays) |
+
+Rules the server enforces rather than documents:
+
+- **It never guesses a server.** The URL comes from `CODEMAN_API_URL` or `--url`; without one it still starts and lists its tools, and every call says how to configure it.
+- **Only its own sessions.** Prompting, interrupting and deleting are limited to sessions this server started (and, run inside a Codeman session, that session's own workers). Listing, reading and waiting work on everything. `codeman mcp --allow-any-session` lifts the limit, for a user who wants an assistant to drive their existing tabs. It never acts on the session it runs in.
+- **Waits fit the client.** A wait defaults to 45 s (Codex cuts MCP tools off at 60 s by default) and is capped at 600 s; `timed_out` / `still_working` means call again.
+- **Both protocol eras.** It answers the 2026-07-28 revision (per-request `_meta`, `server/discover`) and the `initialize` handshake of 2024-11-05 to 2025-11-25.
+
 ### Hooks (events flowing _back_ to Codeman)
 
 Codeman registers Claude Code hooks that `POST /api/hook-event` (`permission_prompt`, `idle_prompt`, `stop`, `task_completed`, …) so the dashboard reacts in real time. This endpoint is auth-exempt on loopback but, under a managed tunnel, requires the `X-Codeman-Hook-Secret` header (read it from `$CODEMAN_HOOK_SECRET_FILE`). You normally don't call this by hand — Codeman wires it up — but it's how the autonomy layers "see" what the agent is doing.

README.zh-CN.md

@@ -24,7 +24,7 @@
 </p>
 
 <p align="center">
-  <img src="docs/images/tiles-crt-stats-20261010.gif" alt="Codeman 平铺视图：六个实时智能体（DeepSeek Harness、Claude Code、Pi、Codex、OpenCode 和一个 shell）以 CRT 动画开启与关闭，顶部实时显示 CPU、内存和 Claude 套餐用量" width="800">
+  <img src="docs/images/tiles-crt-stats-1600-20261010.gif" alt="Codeman 平铺视图：六个实时智能体（DeepSeek Harness、Claude Code、Pi、Codex、OpenCode 和一个 shell）以 CRT 动画开启与关闭，顶部实时显示 CPU、内存和 Claude 套餐用量" width="800">
 </p>
 
 > 本文档由英文版 [`README.md`](README.md) 翻译而来。如有出入，以英文版为准。
@@ -48,11 +48,12 @@ codeman web
 - **真正的手机友好**：[触控优化的终端](#移动端优化的-web-ui)，即时本地回显、二维码登录、滑动导航与推送通知
 - **睡觉时也在跑**：[空闲检测 + 重生循环](#重生控制器respawn-controller)，订阅限额重置后自动续跑，支持 24 小时以上的无人值守运行
 - **看见智能体在想什么**：每个子智能体和团队成员都有[实时浮动窗口](#实时智能体可视化)，附带实时活动记录
+- **智能体也能驱动它**：[codeman 技能](#智能体技能从这里开始)（适用于所有智能体 CLI）、[`codeman agent`](#codeman-agent每种-cli-模式都能用的会话间命令) 和一个 [MCP 服务器](#mcp-服务器codeman-mcp从任意-mcp-客户端驱动-codeman)，让一个智能体在任意 CLI 里拉起工作会话、给它们发提示、等它们完成并收回答案
 - **什么都不会丢**：tmux 让会话挺过重启和断网，输入精确一次送达，完整的回滚缓冲区回放
 - **自托管、私有**：默认仅环回、MIT 许可、无遥测，完全运行在你自己的机器上
 
 <p align="center">
-  <img src="docs/images/codeman-tour-20260724.png" alt="Codeman 仪表盘导览：按项目分组的会话标签页、一键 Run 启动新智能体、页头实时用量" width="900">
+  <img src="docs/images/codeman-tour-20261010.png" alt="Codeman 仪表盘导览：按项目分组的会话标签页、一键 Run 启动新智能体、页头实时用量" width="900">
 </p>
 
 ---
@@ -63,12 +64,13 @@ codeman web
 curl -fsSL https://getcodeman.com/install | bash
 ```
 
-该脚本会在缺失时自动安装 Node.js、tmux 和一套构建工具链（node-pty 没有 Linux 预编译包，需要从源码编译），把 Codeman 克隆到 `~/.codeman/app` 并完成构建。几点须知：
+该脚本会在缺失时自动安装 Node.js、tmux 和一套构建工具链（node-pty 没有 Linux 预编译包，需要从源码编译），把 Codeman 克隆到 `~/.codeman/app` 并完成构建。它会先看看机器上已有什么，最多问三个问题，然后无人值守地完成所有工作，最后给出访问 URL 和一个供手机扫描的二维码。几点须知：
 
-- **先询问，后改动。** 所有系统级改动（安装软件包、下载 AI CLI）都会先征求确认；结束时的菜单可选择：直接在本终端运行、安装为后台服务（systemd/launchd，开机自启），或暂不启动。不选就不会有任何后台进程。
-- **怎么访问，由你决定。** 安装器提供三种到达仪表盘的方式：**Tailscale**（环回绑定，由 `tailscale serve` 代理，得到带真实证书的 `https://<机器名>.<tailnet>.ts.net`，用你的 tailnet 当登录，无需密码）、**局域网内任意设备**（`0.0.0.0`，会提示设置一个强烈推荐的密码），或**仅本机**（`127.0.0.1`，最安全）。绑定网络却跳过密码需要显式确认，并以醒目警告收尾。高亮的默认项反映机器上已有的状态（已在用 Tailscale 时默认 Tailscale，重跑时沿用现有绑定），直接回车绝不会引入新软件。手动运行的 `codeman web` 仍默认仅环回。
-- **重跑即更新。** 再次运行同一条命令即可原地更新已完成的安装：`~/.codeman/app` 中的本地改动会被 stash（绝不丢弃），运行中的服务会自动重启并校验。若首次安装中途失败，重跑会继续完成完整的安装流程。也可以使用 `install.sh update` 与 `install.sh uninstall`。
-- **CI / 无终端环境：** 没有终端时，涉及系统改动的步骤会带着说明中止，而不是静默执行；在自动化场景设置 `CODEMAN_NONINTERACTIVE=1` 即可批准这些步骤。
+- **三个问题，全部在开头问完。** 仪表盘怎么访问、（可选）这台机器在你的 tailnet 里叫什么，以及是否把 Codeman 作为后台服务运行（systemd/launchd，开机自启；直接回车即为「是」）。所有需要你出面的事，包括对所有缺失软件包的一次性同意、一次 sudo 密码和 Tailscale 登录，都在构建之前完成，所以编译期间你可以走开。
+- **怎么访问，由你决定。** 安装器提供三种到达仪表盘的方式：**Tailscale**（环回绑定，由 `tailscale serve` 代理，得到带真实证书的 `https://<机器名>.<tailnet>.ts.net`，用你的 tailnet 当登录，无需密码）、**局域网内任意设备**（`0.0.0.0`，会提示设置一个强烈推荐的密码），或**仅本机**（`127.0.0.1`，最安全）。绑定网络却跳过密码需要显式确认，并以醒目警告收尾。高亮的默认项反映机器上已有的状态（已在用 Tailscale 时默认 Tailscale，重跑时沿用现有绑定），直接回车绝不会引入新软件。如果你节点上的 `:443` 已被别的应用占用，Codeman 会改用 `https://<机器名>.<tailnet>.ts.net/codeman` 或第二个端口，而不是把它替换掉。手动运行的 `codeman web` 仍默认仅环回。
+- **名字由你来定。** 默认情况下，URL 用的是这台机器在 tailnet 里现有的名字。第二个问题答「是」会把机器改名为 `codeman-<hostname>`（这也会改掉它在 SSH 里的名字，所以默认是「否」）；之后也可以用 `install.sh name` 来改。
+- **重跑即更新。** 再次运行同一条命令即可原地更新已完成的安装：`~/.codeman/app` 中的本地改动会被 stash（绝不丢弃），运行中的服务会自动重启并校验。若首次安装中途失败，重跑会继续完成完整的安装流程。`install.sh status` 会再次打印 URL 和二维码；另外还有 `install.sh update`、`install.sh tailscale` 和 `install.sh uninstall`。
+- **给急性子的参数。** `curl -fsSL https://getcodeman.com/install | bash -s -- --tailscale --service` 直接在命令行上回答这些问题（还有 `--lan`、`--local`、`--run`、`--no-start`、`--name <n>`、`--port <n>`、`--yes`）。**CI / 无终端环境：** 没有终端时，涉及系统改动的步骤会带着说明中止，而不是静默执行；在自动化场景设置 `CODEMAN_NONINTERACTIVE=1` 即可批准这些步骤。
 
 你至少需要安装一个 AI 编程 CLI —— [Claude Code](https://docs.anthropic.com/en/docs/claude-code)、[OpenCode](https://opencode.ai)、[Codex](https://developers.openai.com/codex/cli)、[Antigravity](https://antigravity.google)、[Gemini CLI](https://github.com/google-gemini/gemini-cli)、[Pi](https://pi.dev)、[Grok Build](https://github.com/xai-org/grok-build)、[DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) 或 [OMP](https://github.com/can1357/oh-my-pi)（任意组合均可；自 Google 面向消费者停售后，Gemini CLI 仅限企业版，Antigravity 是其继任者）。安装器会自动检测这九个中已安装的任意一个；若一个都没有，会给出一个菜单让你安装其中任意一个（DeepSeek 除外，它的 npm 包只装一个启动器，没有可运行的 profile），也可以选择跳过、稍后自行安装。安装完成后：
 
@@ -213,6 +215,8 @@ Codeman 依赖 tmux，因此 Windows 用户需要 [WSL](https://learn.microsoft.
 
 - **键盘配件栏** —— 在虚拟键盘上方提供 `/init`、`/clear`、`/compact` 快捷按钮；破坏性命令需双击确认，绝不误触；在 Codex 会话上还会显示 `⇧←` / `⇧→`（Shift+Left / Shift+Right：编辑上一条排队的消息 / 在提示栈里回退）
 - **独立的 Enter 按钮** —— 以按键方式回放，先冲刷本地回显缓冲的文本，不会让内容滞留在屏幕上
+- **附件按钮**：麦克风旁边的图片图标会直接打开系统的照片与文件面板，不会弹出键盘。照片和相册里的视频（mp4、mov、webm，最大 1 GB）会上传到会话所在的文件夹，文件路径落在你的提示里，但不会自动发送
+- **在 shell 里粘贴**：多行粘贴会作为一整块插入并等你按 Enter，而不是一边到达一边逐行执行
 - **滑动导航与智能键盘处理** —— 左右滑动切换会话；键盘弹出时工具栏与终端整体上移（`visualViewport` API）
 - **为手机而生** —— 刘海与 Home 指示条的安全区适配、44px 触控目标、底部抽屉式 case 选择器、原生惯性滚动；折叠屏手机（iPhone Duo）上对话框会避开铰链，开合设备也绝不会被误判成键盘弹出
 
@@ -221,7 +225,7 @@ codeman web --https
 # 在手机上打开：https://<你的IP>:3000
 ```
 
-> `localhost` 走纯 HTTP 即可。从其他设备访问时请使用 `--https`，或使用 [Tailscale](https://tailscale.com/)（推荐）：安装器可以替你配好（在网络访问提示处选择 **Tailscale**，或在已有安装上运行 `bash ~/.codeman/app/install.sh tailscale`）。这样你会得到带真实证书的 `https://<你的机器>.<tailnet>.ts.net`：只对你的 tailnet 可见、无需密码，手机上的 PWA 安装和推送通知也都能用。
+> `localhost` 走纯 HTTP 即可。从其他设备访问时请使用 `--https`，或使用 [Tailscale](https://tailscale.com/)（推荐）：安装器可以替你配好（在网络访问提示处选择 **Tailscale**，或在已有安装上运行 `bash ~/.codeman/app/install.sh tailscale`）。这样你会得到带真实证书的 `https://<你的机器>.<tailnet>.ts.net`：只对你的 tailnet 可见、无需密码，手机上的 PWA 安装和推送通知也都能用。安装器最后会给出这个 URL 和一个可扫描的二维码，之后随时可以用 `bash ~/.codeman/app/install.sh status` 再次打印。
 
 ### 安全的二维码认证
 
@@ -265,18 +269,18 @@ codeman service install           # systemd/launchd 服务：重启后自动回
 
 ### 3. 读懂仪表盘
 
-- **标签（顶部）** —— 每个会话一个。`Alt+1`–`9` 跳转，`Ctrl+Tab` 下一个，拖拽排序（标签顺序会跨设备同步）。更喜欢列表？**App Settings → Appearance → Tabs** 可以把它挪进左侧边栏（带筛选框，`Alt+B` 折叠）或一条竖向导轨，导轨的行按活动状态排序：先是等你处理的，然后是跑得最久的，最后是刚刚安静下来的。
+- **标签（顶部）** —— 每个会话一个。`Alt+1`–`9` 跳转，`Ctrl+Tab` 下一个，拖拽排序（标签顺序会跨设备同步）。更喜欢列表？**App Settings → Appearance → Tabs** 可以把它挪进左侧边栏（带筛选框，`Alt+B` 折叠）或一条竖向导轨，导轨的行按活动状态排序：先是等你处理的，然后是跑得最久的，最后是刚刚安静下来的。导轨上还有一个 **Search sessions** 搜索框、命名的**标签分组**（分组菜单里的 **New session** 会直接在该组里新建会话），以及一个 **Focus** 区，放着你常回去的那些标签的快捷方式。
 - **终端（中央）** —— 真实的 `xterm.js` 终端；完整 TUI 正常渲染。直接输入并按 **Enter** 发送。`Shift+Enter` 插入换行。
 - **侧边面板** —— Respawn、Orchestrator、Cron、Subagents、Settings（从工具栏切换）。
 
 ### 4. 与智能体对话
 
 - **直接在终端输入提示** —— 即使跨越重连，输入也是精确一次送达（连接中断绝不会丢失或重复发送提示）。
-- **粘贴或拖放图片**，直接进入会话。
+- **粘贴、拖放或附加图片、视频和文档**（PDF、Office 文件、zip 和纯文本），直接进入会话。它们保存在工作区里一个隐藏的 `.codeman-uploads/` 文件夹中（它在 git 里会忽略自己，7 天后被清理），文件路径会落在你的提示里，但不会自动发送。
 - **语音输入** —— `Ctrl+Shift+V`（Deepgram Nova-3，或者直接用这台机器的 Claude Code 登录、不需要任何 API key；自动静音停止）。
-- **附件** —— 注册外部文件/文档，并内联预览 Office/PDF；智能体打印出的任何文件路径都可以点击，终端里和对话视图里都行。
+- **附件** —— 注册外部文件/文档，并内联预览 Office/PDF；智能体打印出的任何文件路径都可以点击，终端里和对话视图里都行，包括 `~/` 路径、相对路径和被折成多行的路径。`.html` 报告会以真实网页打开，Markdown 会渲染后打开，文件夹则以其文件的网格打开。
 - **需要你的时候** —— 标签会变黄（等待输入）或变红（有个问题挡住了它）。**审批收件箱（Approvals Inbox）**（可选启用）把所有会话里等着你的提示排成一个队列，可以从页头的铃铛或手机首页直接作答；🧠 **Read My Mind**（可选启用）会根据这个 case 的目标和最近的工作替你起草下一条提示。
-- **看到什么就能复制什么** —— `Shift+拖动` 在 CLI 接管了鼠标时也能选中文本，右键复制选中内容，自动复制（Auto Copy，可选启用）在松开鼠标的瞬间就复制。
+- **看到什么就能复制什么** —— `Shift+拖动` 在 CLI 接管了鼠标时也能选中文本，右键复制选中内容，自动复制（Auto Copy，可选启用）在松开鼠标的瞬间就复制。用 OSC 52 复制的程序（Claude Code 的 `/copy`、tmux 的 yank、vim）也会把文本放进你浏览器的剪贴板。
 
 ### 5. 让它自主运行
 
@@ -295,7 +299,7 @@ codeman service install           # systemd/launchd 服务：重启后自动回
 
 ### 7. 运维与维护
 
-- **App Settings** —— 模型、effort、权限启动模式、主题/皮肤、终端字体与字重、入场动画、通知、显示开关、各 CLI 的专属选项，以及跨设备同步的自定义显示名称和按设备保存的英文/简体中文界面语言。
+- **App Settings** —— 模型、effort、权限启动模式、主题/皮肤、终端字体与字重、入场与平铺动画（有自己的 **Animations** 分区）、通知（toast 和浏览器通知的显示时长可调）、显示开关、各 CLI 的专属选项，以及跨设备同步的自定义显示名称和按设备保存的英文/简体中文界面语言。**Apply** 按钮保存设置但不关闭设置窗口。
 - **让它在后台运行** —— `codeman web -d` 脱离你的 shell（`--status`、`--stop`）；`codeman service install` 把它装成 systemd 用户单元 / macOS LaunchAgent，重启后自动回来。两者都会先确认服务器真正应答再报告成功，也都拒绝在同一个数据目录上启动第二个服务器。见[让它在后台一直运行](#快速开始--安装)。
 - **自更新** —— git-clone 安装可在 **App Settings → System → Updates** 中原地更新。
 - **部署你自己的改动** —— 见[开发](#开发)。
@@ -388,7 +392,14 @@ WATCHING → IDLE DETECTED → SEND UPDATE → /clear → /init → CONTINUE →
   <img src="docs/images/tile-grid-20261009.gif" alt="平铺网格：点击平铺按钮，六个实时会话并排打开（DeepSeek Harness、Claude Code、Codex、一个 shell、OpenCode 和 Pi），再点击一次即回到单个会话" width="800">
 </p>
 
-在一个窗口里并排查看和操作最多**六个会话**。点击顶栏的**平铺**按钮（或按 `Ctrl+Shift+G`），会话会以实时终端网格的形式打开：每个窗格都能直接接收键盘输入，并在窗格标题栏中显示智能体的图标、模型和状态。右键单击**平铺**可选择 2、4 或 6 个窗格，按住窗格标题栏拖动即可移动窗格。再次点击**平铺**即回到单个会话，网格会被记住，下次直接恢复。仅限桌面端（窗口宽度约 1180px 以上）。完整说明：[Tile Grid](docs/wiki/Tile-Grid.md)（英文）。
+在一个窗口里并排查看和操作最多**六个会话**。点击顶栏的**平铺**按钮（或按 `Ctrl+Shift+G`），会话会以实时终端网格的形式打开：每个窗格都能直接接收键盘输入，并在窗格标题栏中显示智能体的图标、模型和状态。右键单击**平铺**可选择 2、4 或 6 个窗格，按住窗格标题栏拖动即可移动窗格。再次点击**平铺**即回到单个会话。仅限桌面端（窗口宽度约 1180px 以上）。完整说明：[Tile Grid](docs/wiki/Tile-Grid.md)（英文）。
+
+- **先开重要的**：还没有排好网格时，它会先放正在工作的会话，然后是等你处理的，再是最近用过的，并且总会包含你当前所在的会话。
+- **记住你的布局**：一旦你排好网格（哪个会话在哪、空着的格子、窗格数、分隔条大小、聚焦和放大的窗格），平铺按钮会原样恢复它，网格打开时刷新页面也一样。
+- **平铺动画**（可选启用）：**App Settings → Animations → Tile Animations** 让网格华丽登场：**CRT**（每个窗格像老显像管一样以一道亮线开机，就是顶部 GIF 里那样）、**Fly from tab**、**Deal**、**Beam down**、**Cascade**、**Pop** 或 **Soft**。开启「减少动态效果」时一切静止。
+- **滚动和主终端一样**：鼠标滚轮会在窗格里滚动全屏 Claude 的对话，Codex 窗格会翻页查看它的记录，`Shift+滚轮` 滚动窗格的本地历史。
+- **分离窗格（Detach Tiles）**（测试版，可选启用）：打开 **App Settings → Appearance → Detach Tiles** 后，把窗格标题栏拖出浏览器，它就弹出成独立窗口；拖到另一个 Codeman 窗口的窗格上，它就移到那边。
+- **徒手操作**：开启[手势叠加层](#更多特性)后，捏住一个窗格拖到另一个窗格上即可交换，拖到空格子上即可移动，把一个标签拖进网格即可加入。
 
 ### 持久化会话
 
@@ -423,7 +434,7 @@ codeman web --title-hostname dev-box       # codeman:dev-box（用于覆盖嘈
 ### 标签提醒（Tab Alerts）
 
 <p align="center">
-  <img src="docs/images/tab-alerts-glow-20260815.gif" alt="会话标签：一个普通的活动标签，旁边是黄色的等待输入标签和红色的需要决定标签，都带着呼吸式光晕" width="900">
+  <a href="docs/images/codeman-tab-states-20261010.png"><img src="docs/images/codeman-tab-states-20261010.gif" alt="标签状态（带标注）：一个绿色描边在转的工作中标签，一个被智能体的问题挡住的红色标签（问题显示在它下方），以及一个回合已完成的黄色标签，两个提醒标签都在呼吸闪烁" width="900"></a>
 </p>
 
 每个标签一眼就能看出状态。运行中的会话保持绿色状态点。会话停下来等待输入时，标签变**黄**：稳定的描边、着色的背景、黄色的点，上面叠一层缓慢的呼吸光晕。当权限提示或提问**挡住**了智能体，标签变**红**，脉动更快。底色永远不会闪灭，所以哪怕只瞥一眼（或截一张图）也能读到真实状态；标签被选中时描边依然可见，页面刷新后会从服务端重新装载待处理的提醒，因此一个被挡住的会话绝不可能藏在一个看起来正常的标签后面。
@@ -449,8 +460,11 @@ PTY 输出 → 16ms 服务端批处理 → DEC 2026 包裹 → SSE → 客户端
 ## 更多特性
 
 - **后台守护进程与服务安装** —— `codeman web -d` 以脱离终端的方式运行服务器，带 pid 文件、`~/.codeman/web.log` 和经过校验的启动（它会轮询到服务器应答为止，所以端口冲突绝不会被当成成功）；`codeman service install` 写入一个 systemd 用户单元（Linux）或 LaunchAgent（macOS），并把你 shell 的 PATH 一并写进去，这样 nvm 或 Homebrew 装的 `node`、`tmux` 和 `claude` 才真的找得到。机密永远不会写进单元文件
+- **设置里的诊断**：**App Settings → System → Diagnostics → Run checks** 会在服务器上运行 `codeman doctor`，列出 Node、tmux、每个智能体 CLI 以及可选的 Office 工具，附带版本、路径和安装提示。除了 `PATH`，它还会查找每个 CLI 惯常的安装目录（`~/.local/bin`、`~/.npm-global/bin` 之类），所以即使服务只拿到一个最小的 `PATH`，大多数安装也能被找到。多用户模式下仅管理员可用
 - **自更新** —— systemd/launchd 管理下的 git-clone 安装可在 **App Settings → System → Updates** 中原地更新：它会检测最新发行版，自动暂存（stash）脏工作树，并在服务重启期间流式展示构建进度（npm 安装会被报告为不可更新）
+- **底栏里的 Git 状态**：默认关闭（**App Settings → Header & Panels → Bottom bar → Git status**，按设备生效）。底栏右侧的一个小指示器让你一眼看清当前会话所在仓库的状态：`● 3` 表示未提交的文件，`↑ 2` 表示未推送的提交，`⚠` 表示合并冲突，`✓` 表示全部已提交并推送。点击它会打开一个可拖动的窗口，列出已暂存、未暂存、未跟踪和有冲突的文件（按折叠的文件夹分组，关掉该设置则显示为平铺列表）以及未推送的提交；**点击文件即可查看它的 diff**（新文件全部显示为新增，已删除的文件全部显示为删除），**Open file** 可直接跳到查看器。一个包含多个项目的文件夹，会为向下两层内找到的每个仓库各给一个可折叠分区，打开之前全部折叠。只读且离线（Codeman 绝不 fetch，也绝不改动仓库）；Docker 和远程会话不显示
 - **把 GitHub 仓库克隆成 case** —— 在 **Add Case → Clone Repo** 里粘贴一个仓库 URL，Codeman 会把它克隆到 `~/codeman-cases/<name>` 并注册为普通 case，随时可以跑智能体。输入时它会预检 URL（告诉你能否匿名克隆，并为可选的分支/标签字段提供仓库真实的分支与标签），从 URL 里填好 case 名，还让你选 Run 按钮该用哪个 CLI。支持 `https://` 的公开仓库；Codeman 绝不收集或保存凭据
+- **在自定义文件夹里创建 case**：在 **Add Case → Create New** 里勾选 **Create in a custom folder**，选一个父文件夹（可用 Browse 浏览）和一个名字，Codeman 就会在那里而不是 `~/codeman-cases` 里搭建新 case。目标必须是一个新的或空的文件夹；系统目录、你的主目录本身、`~/.ssh` 这类凭据目录，以及 Codeman 自己的数据目录都会被拒绝。多用户模式下仅管理员可用
 - **多 CLI** —— 每个会话可选 **Claude Code**、**OpenCode**、**Codex**、**Antigravity**、**Gemini**、**Pi**、**Grok**、**DeepSeek Harness** 或 **OMP**；环境变量前缀自动隔离（`CLAUDE_CODE_*`、`OPENCODE_*`、`CODEX_*`、`ANTIGRAVITY_*`、`GEMINI_*`/`GOOGLE_*`、`PI_*`、`GROK_*`/`XAI_*`、`DSH_*`/`DEEPSEEK_*` 与 `OMP_*`）。详见 [`docs/opencode-integration.md`](docs/opencode-integration.md)、[`docs/pi-integration.md`](docs/pi-integration.md)、[`docs/grok-integration.md`](docs/grok-integration.md)、[`docs/deepseek-integration.md`](docs/deepseek-integration.md) 与 [`docs/omp-integration.md`](docs/omp-integration.md)
 - **自定义模型端点**（1.29.0 新增，目前仅 HTTP API）—— 让某个会话的 CLI 指向任意 OpenAI 兼容端点，而不是它自己的官方后端：本地的 llama.cpp、llama-swap、Ollama 或 vLLM 机器，也可以是 Azure AI Foundry、OpenRouter 这类云端网关。端点只需保存一次（`POST /api/model-endpoints`，模型列表从它的 `/v1/models` 自动发现），再应用到会话（`POST /api/sessions/:id/custom-model`），CLI 就会在原地重启并接上该端点。Claude、OpenCode、Pi、Grok 与 OMP 已实测通过；Codex、Gemini 与 DeepSeek 存在已记录的缺口，Antigravity 没有可用机制。工具栏选择器是下一步。详见 [`docs/custom-model-endpoints.md`](docs/custom-model-endpoints.md)
 - **Web 标签页** —— 把 Grafana、Uptime Kuma、一个 Vite 开发服务器或任何仪表盘 URL 作为标签页打开在会话旁边（Run 下拉菜单 → **Web / URL** → **Add URL**）。仪表盘通过 Codeman 自己的源代理，因此 `http://` 目标在手机上走 HTTPS 也能用、走隧道也能用；单页应用能在自己的路径上正常路由，页面自己重载后也能自行恢复。智能体打印出的 `localhost` 链接会自动以 Web 标签页打开。详见 [`docs/web-tabs.md`](docs/web-tabs.md)
@@ -462,10 +476,15 @@ PTY 输出 → 16ms 服务端批处理 → DEC 2026 包裹 → SSE → 客户端
 - **手势控制** _(可选)_ —— 一个 MediaPipe 手部追踪叠加层，可徒手抓取/拖动会话窗口并捏合按钮。用 `CODEMAN_GESTURE=1` + App Settings → Terminal & Input 启用
 - **多显示器横跨** _(macOS)_ —— 一键打开一个横跨所有显示器最大化的浏览器窗口，让浮动的智能体/手势面板可以跨越物理拼接缝
 - **文件查看器按钮** _(可选)_ —— 页头新增一个按钮，一键切换内置文件浏览器面板；在 App Settings → Header & Panels → Header buttons 中启用
+- **HTML 报告以网页打开**：点击智能体打印出的 `.html` 路径（测试报告、覆盖率页面、仪表盘），文件查看器会把它显示成一个真实网页，并带上同一文件夹里的脚本、样式、图片和数据文件；**Page** 胶囊按钮可切到源码及其 Edit 按钮。页面在沙箱里以不透明源运行，所以它读不到 Codeman 的 cookie、调不了 Codeman 的 API，也读不了自己文件夹以外的文件。网络访问保持开放，因此从 CDN 加载 Chart.js 或字体的仪表盘和直接从磁盘打开时一样能用。范围过大的文件夹（`/`、你的主目录、`/tmp`）会被拒绝，所以请给报告一个单独的文件夹。在终端里点击的 `.md` 路径也会渲染后打开
+- **可点击的路径与文件夹网格**：写成 `~/…`、相对于会话文件夹或被折成多行的路径，在终端和回复查看器（Response Viewer）里都能点击，文件夹也可以：点击文件夹会打开它的文件网格（渲染图、截图、报告），用方向键翻看
+- **OSC 52 剪贴板**：会话里的程序用 OSC 52 设置剪贴板时（Claude Code 的 `/copy`、tmux 的 yank、vim），文本会进入你浏览器的剪贴板，主终端和每个平铺窗格都一样。读取剪贴板的请求一律拒绝，回放缓冲区也绝不会触发复制
+- **MCP 服务器同步**（可选启用）：把每个已安装 CLI 的用户级 MCP 服务器（Claude、Codex、Gemini、OpenCode、Antigravity，以及 GitHub Copilot CLI）复制进其他 CLI 自己的配置文件。只增不改，你关掉的服务器绝不会被复制，**Preview** 会先列出改动，原文件保留为 `.codeman-bak`。位于 App Settings → Agents & CLIs；多用户模式下仅管理员可用
 - **CJK / 输入法支持** —— 完整支持中文 / 日文 / 韩文的组合输入，Ctrl、Alt 修饰的导航键也会原样透传给 CLI
 - **页头里的套餐用量** —— 页头实时显示 Claude 订阅用量（5 小时窗口与每周窗口），数据来自 Codeman 在拉起 `claude` 时临时交给它的 statusline 导出器，绝不会写进你的设置文件；Codex 的限额则来自它自己的 app-server。按设备生效：桌面默认开，手机默认关
-- **会话列表，随你摆** —— 页头横条、带筛选框的左侧边栏，或一条竖向导轨，导轨的详细行带有创建时间与状态时长并按活动状态排序；手机首页和桌面首页导轨用的是同一套顺序
-- **终端外观** —— 七套皮肤（其中四套浅色）、按设备保存的字体与字重（内置的 JetBrains Mono 覆盖 100 到 800 的字重），以及可选启用的入场动画，覆盖标签、智能体窗口、终端面板和连接线
+- **提示缓存状态**：每个 Claude 会话都会显示它的提示缓存是热的（`cache:until HH:MM`）还是冷的，显示在页脚（前提是你没有自己的 statusline）和 Session Options → Respawn 中，并给出重新缓存的代价，压缩上下文后还会附一条说明。只做显示，重启后依然保留
+- **会话列表，随你摆** —— 页头横条、带筛选框的左侧边栏，或一条竖向导轨，导轨的详细行带有创建时间与状态时长并按活动状态排序；手机首页和桌面首页导轨用的是同一套顺序。导轨上的 **Search sessions** 会把列表筛到名字匹配的标签，跨所有分组；带提醒的标签即使名字不匹配也始终可见。命名的**标签分组**可以折叠、重排和拖入拖出；分组菜单里有 **New session**，智能体拉起的工作会话会落在它的父会话后面（**Spawned tabs follow their parent**，默认关闭，开启后它会跟着父会话移动）。**Add to Focus** 会在导轨顶部的一个区里为任意标签放一个快捷方式，**CLI Logos on Tabs** 可以按设备关闭
+- **终端外观** —— 七套皮肤（其中四套浅色）、按设备保存的字体与字重（内置的 JetBrains Mono 覆盖 100 到 800 的字重），以及可选启用的动画，覆盖标签、智能体窗口、终端面板、连接线和[平铺网格的窗格](#平铺网格)，全部位于 App Settings → Animations（附带一个试玩用的动画实验室）
 - **操作系统通知与主机名感知标题** —— 桌面提醒与标签标题以 `codeman:<host>` 为前缀，使多主机配置不再含糊
 
 ---
@@ -480,7 +499,7 @@ PTY 输出 → 16ms 服务端批处理 → DEC 2026 包裹 → SSE → 客户端
 - **默认加固** —— 非 root、`--cap-drop ALL`、`no-new-privileges`、PID/内存上限，绝不使用 `--privileged` 或 docker socket；**密封（sealed）** 配置（不注入主机凭据、关闭网络）只需一个开关。
 - **无感认证、凭据隔离** —— 主机上的 Claude / Codex / Antigravity / Gemini / OpenCode / Pi / Grok / OMP 登录在容器内开箱即用：凭据在启动时以只读种子方式复制注入，onboarding/信任提示已预先答复，不会弹出登录向导。容器保留自己的副本，绝不回写主机的凭据存储；跨边界共享的只有对话转录，导出文件也绝不包含机密。
 - **挂到你已经在跑的容器上** —— 在 Docker 面板勾选 **Attach to an existing container**，就能把 case 链接到一个现成容器，而不是新建一个。Codeman 只 `exec` 进去，绝不启动、停止、重启或删除它；一个被接管的容器可以在不同目录下支撑多个 case，**复制一个已有 case** 会用同一容器上的兄弟 case 预填表单。多用户模式下仅管理员可用，因为容器的挂载属于启动它的人。
-- **迁移到另一台机器** —— 把容器的完整环境（工具链 + 工作区）导出为可移植的 `.tar.gz`，在另一台机器上导入到新案例即可继续。
+- **迁移到另一台机器** —— 把容器的完整环境（工具链 + 工作区）导出为可移植的 `.tar.gz`，在另一台机器上用 `docker load` 载入，再导入到一个新案例即可继续。
 - **持久耐用** —— Codeman 重启后重连会回到同一个存活的智能体；容器停止/重启后则从绑定挂载的转录恢复对话。
 
 前置条件：只需 Docker（或 Podman）。智能体基础镜像会在首次使用时自动构建，构建进度实时显示在 UI 中（也可用 `node scripts/build-agent-image.mjs` 预构建）。完整指南：[`docs/docker-cases.md`](docs/docker-cases.md)。
@@ -497,6 +516,7 @@ PTY 输出 → 16ms 服务端批处理 → DEC 2026 包裹 → SSE → 客户端
 - **共享会话**：多个客户端可以以不同窗口尺寸同时附着同一个远程会话而互不挤压；发现列表会显示带客户端计数的「shared」徽标。
 - **注入安全**：所有 ssh 命令行都经由单一的 shell 转义构建器生成，主机/路径/身份文件字段均有模式校验。
 - **文件也行**：远程 case 里的预览、下载和文本读取走同一条 ssh 连接（一次 `realpath` + `stat` 探测，然后流式 `cat`，支持 `Range` 拖动进度），所以点一个路径打开的就是智能体所在那台机器上的文件。什么都不会复制到 Codeman 主机；编辑和 Office 预览会明确返回 400，而不是一个误导性的 404。
+- **完整的回滚历史**：加载远程会话的历史时，会通过一次有上限的 ssh 往返直接从远程 tmux 读取，所以重新附着或切换标签之后，会话依然保有完整的回滚缓冲区，历史与实时屏幕衔接处也不会丢行。
 
 在 **New Case → Remote** 中配置（主机、用户、身份文件、可选跳板机）。完整设计：[`docs/remote-sessions.md`](docs/remote-sessions.md)。
 
@@ -649,7 +669,7 @@ Codeman 默认用 `--dangerously-skip-permissions` 启动会话，因此 Web UI
 
 ### 网络与访问
 
-- **默认仅环回** —— 绑定 `127.0.0.1`，仅可从本机访问，因此「无密码」默认配置开箱即安全。在未设置 `CODEMAN_PASSWORD` 的情况下绑定非环回主机会*启动但打印一条醒目警告*，并给出三个具体修复方案（设置密码、环回 + 一个带认证的隧道，或用 `--allow-unauthenticated-network` 显式确认）
+- **默认仅环回** —— 服务器程序本身绑定 `127.0.0.1`，仅可从本机访问，因此「无密码」默认配置开箱即安全（引导式安装器会询问网络访问方式，并替你配置绑定和密码）。在未设置 `CODEMAN_PASSWORD` 的情况下绑定非环回主机会*启动但打印一条醒目警告*，并给出三个具体修复方案（设置密码、环回 + 一个带认证的隧道，或用 `--allow-unauthenticated-network` 显式确认）
 - **可选认证，真实会话** —— 通过 `CODEMAN_USERNAME`（默认 `admin`）/ `CODEMAN_PASSWORD` 的 HTTP Basic 认证。成功后签发一个不透明的 256 位 `codeman_session` cookie（`randomBytes(32)`）—— 服务端校验，而非客户端签名，因此无法离线伪造（24h TTL、自动延长、设备上下文审计日志）
 - **按 IP 速率限制** —— 失败 10 次 → `429` 并带 `Retry-After`（15 分钟衰减）。即便攻击者在同一 IP 上猛攻，有效 cookie 或正确密码也能*立即*恢复 —— 这很重要，因为所有隧道流量共享同一个环回 IP。二维码认证有自己独立的限制器
 - **可配置的权限模式**：`--dangerously-skip-permissions` 只是默认值。**App Settings → Agents & CLIs → Claude → Startup Mode** 可以把新会话切换为 Anthropic 的分类器护栏 `auto` 模式（低打扰，需要 Claude Code 2.1.207+）、`normal` 提示模式，或一份显式的允许工具列表。多用户模式下，未获授权的用户会被强制为 `auto`，shell 会话与跳过权限需要按用户显式授权
@@ -667,7 +687,7 @@ Codeman 默认用 `--dangerously-skip-permissions` 启动会话，因此 Web UI
 ### 输入、文件与响应头
 
 - **模式校验的输入** —— 每个 API 请求体都用 Zod v4 模式检查；一个 `CLAUDE_CODE_*` / `OPENCODE_*` / `CODEX_*` / `ANTIGRAVITY_*` / `GEMINI_*` / `GOOGLE_*` / `PI_*` / `GROK_*` / `XAI_*` / `DSH_*` / `DEEPSEEK_*` / `OMP_*` 环境变量前缀允许列表把控每个 CLI 能接收哪些设置，而那些能把 CLI 流量改道的键（base URL、配置目录）对非管理员用户会被钳制
-- **路径限定** —— 文件路由在边界检查前先 `realpath`（无 TOCTOU）；`..`、绝对路径、以及解析到工作目录之外的符号链接都会被拒绝。上限：10 MB 文本预览 / 2 GB 原始与下载（`CODEMAN_MAX_DOWNLOAD_BYTES`；响应体是流式的并支持 `Range` 请求，所以这个上限只是合理性边界，不是内存保护）；`/api/download` 对敏感路径（`.env`、`*credentials*`、`~/.ssh/`、`.aws/credentials`）做黑名单。SVG/HTML 以 `octet-stream` + `nosniff` + attachment 提供，因此会被下载而非执行
+- **路径限定** —— 文件路由在边界检查前先 `realpath`（无 TOCTOU）；`..`、绝对路径、以及解析到工作目录之外的符号链接都会被拒绝。上限：10 MB 文本预览 / 2 GB 原始与下载（`CODEMAN_MAX_DOWNLOAD_BYTES`；响应体是流式的并支持 `Range` 请求，所以这个上限只是合理性边界，不是内存保护）；`/api/download` 对敏感路径（`.env`、`*credentials*`、`~/.ssh/`、`.aws/credentials`）做黑名单。SVG/HTML 以 `octet-stream` + `nosniff` + attachment 提供，因此会被下载而非执行。HTML 唯一会被渲染的地方是文件查看器的网页视图：一个独立的 `/html-view/<capability>/` 地址，只覆盖该文件所在的文件夹，响应带 `Content-Security-Policy: sandbox` 且不含 `allow-same-origin`，因此页面是不透明源，永远读不到 Codeman 的 cookie，也驱动不了它的 API；隐藏文件夹和范围过大的根目录（`/`、主目录、`/tmp`）会被拒绝，退出登录时这些凭证也会被吊销
 - **安全响应头** —— `Content-Security-Policy`（`default-src 'self'`，每个例外都逐条列举）、`X-Content-Type-Options: nosniff`、`X-Frame-Options: SAMEORIGIN`、HTTPS 下的 HSTS，以及**仅**对 `localhost` / `127.0.0.1` / `::1` 反射的 CORS
 
 ### 供应链与隔离
@@ -701,7 +721,7 @@ Web UI 仍是主要界面；完整指南见 **[docs/tui.md](docs/tui.md)**（英
 
 | 快捷键                          | 动作                                                     |
 | ------------------------------- | -------------------------------------------------------- |
-| `Ctrl/Cmd+W`                    | 杀掉当前会话                                             |
+| `Ctrl+W`                        | 不是 Codeman 快捷键：它会到达终端（删除前一个词）。**Close Session** 默认没有按键，可在 App Settings → Shortcuts 里绑定 |
 | `Ctrl/Cmd/Option+K`             | 查找已打开的会话或新建一个                               |
 | `Ctrl/Cmd+Tab`                  | 下一个会话                                               |
 | `Alt/Option+[` / `Alt/Option+]` | 上一个 / 下一个会话                                      |
@@ -714,6 +734,10 @@ Web UI 仍是主要界面；完整指南见 **[docs/tui.md](docs/tui.md)**（英
 | `Ctrl/Cmd+L`                    | 清屏                                                     |
 | `Ctrl+Shift+R`                  | 恢复终端尺寸                                             |
 | `Ctrl+Shift+V`                  | 切换语音输入                                             |
+| `Ctrl+Shift+G`                  | 打开或关闭平铺网格                                       |
+| `Alt+Shift+方向键`              | 聚焦左、右、上、下方的窗格                               |
+| `Ctrl+Shift+方向键`             | 把聚焦的窗格移动一格（移进空格子，或与邻居交换）         |
+| `Alt+Shift+Enter`               | 放大聚焦的窗格，或恢复网格                               |
 | `Ctrl/Cmd +` / `-`              | 字体大小                                                 |
 | `Ctrl/Cmd+?`                    | 键盘帮助                                                 |
 | `Shift+Enter`                   | 插入换行（发送到终端）                                   |
@@ -731,19 +755,23 @@ Web UI 仍是主要界面；完整指南见 **[docs/tui.md](docs/tui.md)**（英
 
 ### 智能体技能（从这里开始）
 
-这一节的所有内容也打包成了一个 **Claude Code 技能**，位于 [`skills/codeman`](skills/codeman/SKILL.md)。装一次，就再也不用把 API 文档粘进提示词。你用大白话说想要什么，已经坐在 Codeman 会话里的智能体会自己加载配方并驱动 API。
+这一节的所有内容也打包成了一个**智能体技能**，位于 [`skills/codeman`](skills/codeman/SKILL.md)，采用 Claude Code、Codex、Gemini CLI、OpenCode、Grok、OMP 以及大多数其他智能体 CLI 都能读取的 `SKILL.md` 格式。装一次，就再也不用把 API 文档粘进提示词。你用大白话说想要什么，已经坐在 Codeman 会话里的智能体会自己加载配方并驱动 API。完整说明（所有安装方式、录屏，以及它会自己做什么、不会做什么）：[The Codeman Skill](https://github.com/Ark0N/Codeman/wiki/Codeman-Skill)（英文）。
+
+<p align="center">
+  <a href="docs/images/codeman-skill-crt-20261010.png"><img src="docs/images/codeman-skill-crt-20261010.gif" alt="一次真实的 codeman 技能运行：在 Claude Code 里输入一句简短的提示，平铺网格开机，然后一个跑在本地 qwen 模型上的 DeepSeek Harness 工作会话和一个 Claude Code 工作会话作为新窗格开机，主会话到两者之间画出派生连线" width="900"></a>
+</p>
 
 #### 第 1 步：安装
 
 | 方式             | 命令                                                       | 范围                                                                                       |
 | ---------------- | ---------------------------------------------------------- | ------------------------------------------------------------------------------------------ |
-| Skills CLI       | `npx skills add Ark0N/Codeman --skill codeman -g`          | 全局，任何支持技能的智能体都能用                                                           |
+| Skills CLI       | `npx skills add Ark0N/Codeman --skill codeman -g`          | 全局，任何支持技能的智能体都能用（[skills.sh](https://skills.sh)）；给 pi 装技能就用这个方式 |
 | Claude Code 插件 | `/plugin marketplace add Ark0N/Codeman`，然后 `/plugin install codeman@codeman` | 全局，通过 Claude Code 自带的插件管理器；`/plugin update codeman` 跟随新版本。与 `codeman skill install` 二选一：两者都装会让技能出现两次（`codeman` 和 `codeman:codeman`） |
 | 内置 CLI         | `codeman skill install`                                    | 全局（`~/.claude/skills/codeman`），给那些从 npm 安装、从未克隆过仓库的用户                |
-| 内置 CLI         | `codeman skill install --case <name>`                      | 仅一个 case                                                                                |
-| Web UI           | App Settings → Agents & CLIs → Claude → **Agent Skill**    | 每次在某个 case 创建 Claude 会话时自动注入（`agentSkillEnabled`，跨设备同步，默认关闭）    |
+| 内置 CLI         | `codeman skill install --case <name> [--mode <mode>]`      | 仅一个 case；`--mode codex`（或任何其他运行模式）会把它写到那个 CLI 查找技能的位置          |
+| Web UI           | App Settings → Agents & CLIs → All agent CLIs → **Codeman Agent Skill** | 每次创建会话时自动注入到 case 里该会话的 CLI 查找技能的位置：Claude Code 是 `.claude/skills/`，Codex、Gemini CLI、OpenCode、Antigravity、Grok、DeepSeek 和 OMP 是 `.agents/skills/`；pi 除外，因为它的项目级技能会触发信任提示（`agentSkillEnabled`，跨设备同步，默认关闭） |
 
-`codeman skill uninstall [--case <name>]` 可以撤销 CLI 安装，并且绝不会碰你自己写的 `skills/codeman`。
+`codeman skill uninstall [--case <name>]` 可以撤销 CLI 安装（带 `--case` 时，删除该 case 里所有由 Codeman 管理的副本，不论它在哪个 CLI 的目录里），并且绝不会碰你自己写的 `skills/codeman`。
 
 #### 第 2 步：开口要
 
@@ -757,9 +785,19 @@ Web UI 仍是主要界面；完整指南见 **[docs/tui.md](docs/tui.md)**（英
 | _「让一个 claude 工作会话在 `refactor-auth` 上总结 `src/session.ts`，然后关掉它。」_  | 拉起、走完就绪阶梯（包括首次运行的信任对话框）、发送并等待、读取干净的 transcript 答案、删除。 |
 | _「盯着会话 w4，如果它卡在权限提示上就告诉我。」_                                      | 阻塞在 `blocked` 信号上，并把问题交给**你**。它绝不会替另一个会话回答提示。                    |
 
+一句简短的话就够了。下面这个 Claude Code 主会话收到 `use codeman to start 5 claude workers, one per TODO item`，五个真实的 Claude Code 工作会话依次开机，并排显示在平铺网格里：
+
+<p align="center">
+  <img src="docs/images/codeman-skill-grid-20261015.gif" alt="一次真实的 codeman 技能运行（平铺网格）：主 Claude Code 会话作为唯一的窗格开机，输入提示 use codeman to start 5 claude workers, one per TODO item，然后五个 Claude Code 工作会话以 CRT 动画依次作为新窗格开机，六个智能体同时工作" width="900">
+</p>
+
 #### 第 3 步：没有了
 
-智能体会删掉它启动的每一个会话。你可以在仪表盘里看着标签出现又消失。
+智能体会删掉它启动的每一个会话。你可以在仪表盘里看着标签出现又消失。起八个工作会话时，每个都有自己的标签，还有一条派生连线连回启动它的主会话：
+
+<p align="center">
+  <img src="docs/images/codeman-skill-tabs-20261015.gif" alt="一次真实的 codeman 技能运行（普通视图）：输入提示 use codeman to start 8 claude workers, one per TODO item，八个工作会话标签从左到右依次开机，主会话到每个标签之间画出派生连线，然后点开一个工作会话标签，看到一个正在接收任务的真实 Claude Code 工作会话" width="900">
+</p>
 
 #### 一次真实的运行，从头到尾
 
@@ -796,7 +834,7 @@ deleted 9f2d8e5f, aff9c691, be9f1f15    标签消失
 - **它会自我门禁。** 不在 Codeman 会话里（`CODEMAN_MUX` 未设置）时，技能拒绝动作，也不去猜 API 地址，所以全局安装对无关的 Claude Code 会话没有任何代价。
 - **它刻意保守。** 未经提示，它只会拉起会话、给它们发提示，并删除**它在同一段对话里自己创建的**会话（按精确 id，经由一个拒绝删除智能体自身会话的失败即关闭守卫）。删除 case（会抹掉一个真实的代码目录）、批量杀会话、改动 respawn/ralph/cron/orchestrator 以及写设置，都需要你开口并指名目标。
 
-⚠️ 把 `agentSkillEnabled` 关回去**不会删掉已经注入的副本**（在创建时做清扫，会把技能从共用同一个 `.claude/` 目录的其他活动会话脚下抽走）。要删就按 case 删：`codeman skill uninstall --case <name>`。
+⚠️ 把 `agentSkillEnabled` 关回去**不会删掉已经注入的副本**（在创建时做清扫，会把技能从共用同一个工作区的其他活动会话脚下抽走）。要删就按 case 删：`codeman skill uninstall --case <name>`。
 
 ---
 
@@ -821,7 +859,7 @@ deleted 9f2d8e5f, aff9c691, be9f1f15    标签消失
 4. **响应信封。** 多数端点返回 `{ "success": true, "data": … }`（错误：`{ "success": false, "error", "errorCode" }`）。少数遗留 GET 返回裸响应体 —— **两种都要处理**（`body.data ?? body`）。
 5. **`/api/v1/*`** 是 `/api/*` 的稳定别名。
 6. **用等待代替轮询，别把超时当成错误。** 等待类端点在没等到事情发生时也以 HTTP `200` 加 `wait.timedOut: true` 应答，所以要循环调用短等待（默认 60 秒），而不是发一个超长的调用：隧道会掐断空闲连接。`wait.timeoutMs` 告诉你服务端钳制之后真正采用的超时（上限 600 秒）。
-7. **只有 `claude` 与 `deepseek` 会话会发出 `stop` 与 `blocked`。** 这两个来自 hook（Claude Code 自己的，以及 DeepSeek Harness 的状态桥接）；`shell` 与其他外部 CLI（opencode/codex/gemini/antigravity/pi/grok/omp）只接受 `idle`、`working` 与 `exit`。在这些模式上显式索要 `stop` 会得到 `400`；不传 `until` 则永远安全。⚠️ `shell` 会话的 `idle` 只在启动时触发**一次**，此后再也不会，所以在那里用「发送并等待」只能等到超时：没有 hook 的会话请用 `wait-output` 标记来同步。
+7. **只有 `claude` 与 `deepseek` 会话会发出 `stop`；`codex` 还会发出 `blocked`。** 它们来自 hook（Claude Code 自己的，以及 DeepSeek Harness 的状态桥接），codex 的 `blocked` 除外：批准、提问或信任对话框出现时，Codeman 从它的屏幕上读出这个信号（请在终端里回答）。`shell` 与其他外部 CLI（opencode/gemini/antigravity/pi/grok/omp）只接受 `idle`、`working` 与 `exit`，codex 另外接受 `blocked`。在这些模式上显式索要 `stop` 会得到 `400`；不传 `until` 则永远安全。⚠️ `shell` 会话的 `idle` 只在启动时触发**一次**，此后再也不会，所以在那里用「发送并等待」只能等到超时：没有 hook 的会话请用 `wait-output` 标记来同步。
 8. **没有任何东西会报告「就绪」，得自己显式等。** 新会话在 PID 出现之前一律回答 `{"signal":"exit","immediate":true}`（意思是*还没启动*，不是*崩了*），而全新 case 里的 `claude` 工作会话接着会停在 CLI 的信任对话框上。此时给它发提示，等待会在约 2 秒后因 `idle` 解除，看上去和一个跑完的回合一模一样，而文本其实卡在对话框里。下面的配方 2b 就是避开它的顺序。
 
 ### 常用配方
@@ -933,6 +971,76 @@ codeman tui --list                       #      带编号的会话列表（管
 codeman tui 3                            #      附着到该列表里的第 3 个会话
 ```
 
+### `codeman agent`：每种 CLI 模式都能用的会话间命令
+
+上面的技能靠 Codeman 为每个会话写入的一段前导脚本，用 `curl` 和 `jq` 驱动 API。在它够不着的地方（`pi` 或 `shell` 会话、没有 `jq` 的机器），同样的环境依然存在（每个面板里都导出了 `CODEMAN_MUX=1`、`CODEMAN_SESSION_ID`、`CODEMAN_API_URL`），所以 `codeman agent` 把同一批动词打包成了命令。它只是 [API](#api) 一节所列端点之上的一层薄客户端：没有新路由，没有新的传输方式，认证和归属规则不变。在 case 的 `AGENTS.md` 里写一行就够了：*"other sessions: `codeman agent --help`"*。
+
+```bash
+codeman agent ls                                          # 列出会话；* 标记当前这个
+SID=$(codeman agent spawn scratch-1 --mode claude)        # quick-start，并在该模式有就绪标记时等到输入框出现
+codeman agent send "$SID" 'review src/, then say DONE' --until stop,exit --timeout 300000   # --wait = 默认信号集
+codeman agent read "$SID"                                 # 最后一条回答（按服务端对该模式的读取方式）
+codeman agent read "$SID" --tail 3000                     # 终端尾部，已去掉 ANSI（所有模式）
+codeman agent send "$SID" 'run the tests, then print WORKDONE followed by _4711'   # 没有 hook 的模式（opencode、pi 等）：让它分两半输出标记……
+codeman agent wait "$SID" --match WORKDONE_4711           # ……然后等拼接后的形式，提示本身的回显里永远不会出现它
+codeman agent interrupt "$SID"                            # 只发一个 ESC，对话保持完整
+codeman agent rm "$SID"                                   # 除当前会话外的任何会话
+```
+
+这些规则由命令强制执行，而不只是写在文档里：在 Codeman 会话之外它们拒绝运行，也绝不猜测 URL；`send` 只发送可打印文本加 Enter（像 `Ctrl+C` 这样的控制字节在 opencode 里就是 `app_exit`，所以 ESC 只以 `interrupt` 的形式存在，而且它从不追加 Enter）；`rm` 拒绝空 id、无法证明不是自己的 id，以及任一方向的前缀匹配。id 可以用 `ls` 打印的 8 字符前缀（通过列表解析；有歧义的前缀会被拒绝，短于 8 个字符的 id 在所有动词上都会被拒绝）。以 `-` 开头的提示要放在 `--` 之后（`send "$SID" -- "- fix the bug"`）。
+
+你发出的提示的回显也算输出：如果 `--match` 标记原样出现在提示里，它会立刻匹配，工作会话还什么都没做，所以提示要让它分两半输出。主机正在休眠的远程会话，对不等待结果的 `send` 会回答 `buffered`（Codeman 唤醒主机，等面板恢复后再敲入提示）或 `dropped`（超出唤醒缓冲区的上限，什么都不会敲入：退出码 `1`）。退出码：`0` 已送达/已匹配/收到信号，`1` 错误，`2` 超时，`3` 工作会话已退出或等待在没有回答的情况下结束（`delivered:false`、`ended:true`），`4` 被拒绝。`spawn` 在 stdout 上只打印 id（说明文字走 stderr），所以 `SID=$(…)` 恰好拿到 id。`--json` 让每个动词打印响应信封里的 `data`。在没有 hook 信号的模式上使用 `--until stop`，会原样透传服务端的 400：那里的答案是标记路径（`--match`），和技能完全一样。
+
+### MCP 服务器（`codeman mcp`）：从任意 MCP 客户端驱动 Codeman
+
+`codeman mcp` 是一个走 stdio 的 [Model Context Protocol](https://modelcontextprotocol.io) 服务器，因此任何 MCP 客户端（Claude Code、Claude Desktop、Codex、Gemini CLI、Cursor 等）都能在任意已安装的智能体 CLI 里启动 Codeman 会话、给它们发提示、等它们完成、读取回答并清理，无论是在 Codeman 内部还是外部。和 `codeman agent` 一样，它是 HTTP API 的客户端：没有新路由，认证和归属规则不变。
+
+```bash
+# Claude Code
+claude mcp add codeman -e CODEMAN_API_URL=https://127.0.0.1:3000 -- codeman mcp
+```
+
+```json
+{
+  "mcpServers": {
+    "codeman": {
+      "command": "codeman",
+      "args": ["mcp"],
+      "env": { "CODEMAN_API_URL": "https://127.0.0.1:3000" }
+    }
+  }
+}
+```
+
+```toml
+# Codex（~/.codex/config.toml）：为长时间等待调高默认 60 秒的工具超时
+[mcp_servers.codeman]
+command = "codeman"
+args = ["mcp"]
+env = { CODEMAN_API_URL = "https://127.0.0.1:3000" }
+env_vars = ["CODEMAN_MUX", "CODEMAN_SESSION_ID"]  # Codex 只给 MCP 服务器一个最小环境；这两个变量让它知道自己运行在哪个 Codeman 会话里
+tool_timeout_sec = 120
+```
+
+JSON 形式适用于 Claude Desktop、Cursor 和 Gemini CLI。服务器设置了密码时，在 `env` 里加上 `"CODEMAN_PASSWORD"`（或者让它读取数据目录里的 `.env`）。如果 PATH 上没有 `codeman` 命令（没走安装器的 git 检出），改用 `node /path/to/Codeman/dist/index.js mcp`。
+
+| 工具 | 作用 |
+| ---- | ---- |
+| `list_sessions` | 你能看到的每个会话，附带运行模式、状态、文件夹，以及这些工具是否可以给它发提示 |
+| `start_session` | 在一个 case 里（全新的临时文件夹或已有的 case）用任意已安装的 CLI 起一个新工作会话；等到它能接收提示为止 |
+| `send_prompt` | 输入一行提示并按 Enter；可选择等到这一回合结束 |
+| `wait_for_session` | 等待一个信号（`stop`、`blocked`、`idle`、`working`、`exit`）或输出里的一个字面标记 |
+| `read_output` | 最后一条回答、整段对话，或终端的最后若干字节 |
+| `interrupt_session` | 一次 Escape：停止当前回合，保留对话 |
+| `delete_session` | 关闭一个工作会话（它的 case 文件夹会保留） |
+
+这些规则由服务器强制执行，而不只是写在文档里：
+
+- **它绝不猜测服务器。** URL 来自 `CODEMAN_API_URL` 或 `--url`；没有配置时它照样启动并列出工具，每次调用都会说明该如何配置。
+- **只动自己的会话。** 发提示、中断和删除只限于这个服务器自己启动的会话（以及它运行在某个 Codeman 会话里时，该会话自己的工作会话）。列出、读取和等待对所有会话都可用。`codeman mcp --allow-any-session` 会解除这个限制，适合想让助手驱动自己现有标签的用户。它永远不会对自己所在的会话下手。
+- **等待时长适配客户端。** 一次等待默认 45 秒（Codex 默认在 60 秒时切断 MCP 工具），上限 600 秒；返回 `timed_out` / `still_working` 时再调用一次即可。
+- **两个协议时代都支持。** 它既能应答 2026-07-28 版本（每个请求带 `_meta`、`server/discover`），也能应答 2024-11-05 到 2025-11-25 的 `initialize` 握手。
+
 ### Hook（事件*回流*到 Codeman）
 
 Codeman 会注册 Claude Code hook，它们 `POST /api/hook-event`（`permission_prompt`、`idle_prompt`、`stop`、`task_completed` 等），让仪表盘实时响应。该端点在环回上免认证，但在受管隧道下需要 `X-Codeman-Hook-Secret` 头（从 `$CODEMAN_HOOK_SECRET_FILE` 读取）。通常你不需要手动调用它 —— Codeman 会自动接好 —— 但自主层正是靠它「看见」智能体在做什么。

config/clis.stock.json

@@ -238,8 +238,8 @@
       },
       "install": {
         "command": {
-          "linux": "npm install -g @deepseek-ai/dsh",
-          "darwin": "npm install -g @deepseek-ai/dsh"
+          "linux": "npm install -g @deepseek-ai/dsh && (command -v pnpm >/dev/null || npm install -g pnpm)",
+          "darwin": "npm install -g @deepseek-ai/dsh && (command -v pnpm >/dev/null || npm install -g pnpm)"
         },
         "npmPackage": "@deepseek-ai/dsh",
         "docsUrl": "https://github.com/deepseek-ai/deepseek-harness",

config/mcp-registry/server.json

@@ -0,0 +1,43 @@
+{
+  "$schema": "https://static.modelcontextprotocol.io/schemas/2025-12-11/server.schema.json",
+  "name": "io.github.Ark0N/codeman",
+  "title": "Codeman",
+  "description": "Start, prompt and read AI coding agent sessions (Claude Code, Codex, Gemini CLI and more) in Codeman",
+  "version": "1.42.0",
+  "websiteUrl": "https://getcodeman.com",
+  "repository": {
+    "url": "https://github.com/Ark0N/Codeman",
+    "source": "github"
+  },
+  "packages": [
+    {
+      "registryType": "npm",
+      "identifier": "aicodeman",
+      "version": "1.42.0",
+      "transport": {
+        "type": "stdio"
+      },
+      "packageArguments": [
+        {
+          "type": "positional",
+          "value": "mcp"
+        }
+      ],
+      "environmentVariables": [
+        {
+          "name": "CODEMAN_API_URL",
+          "description": "Where your Codeman server runs, for example https://127.0.0.1:3000. Never guessed.",
+          "isRequired": true,
+          "format": "string"
+        },
+        {
+          "name": "CODEMAN_PASSWORD",
+          "description": "Codeman's HTTP Basic password, when the server has one.",
+          "isRequired": false,
+          "isSecret": true,
+          "format": "string"
+        }
+      ]
+    }
+  ]
+}

config/test-suites.ts

@@ -19,9 +19,12 @@
  * why these are a runnable suite (`npm run test:browser`) rather than skipped.
  */
 export const BROWSER_TEST_GLOBS = [
+  'test/tmux-capture-sgr.browser.test.ts',
   'test/tab-rail-resize.browser.test.ts',
   'test/tab-activation.browser.test.ts',
   'test/tab-layout-editing.browser.test.ts',
+  'test/tab-rail-search.browser.test.ts',
+  'test/tab-focus.browser.test.ts',
   'test/session-sidebar-ux.browser.test.ts',
   'test/session-options-responsive.browser.test.ts',
   'test/inline-rename.test.ts',
@@ -34,6 +37,7 @@ export const BROWSER_TEST_GLOBS = [
   'test/codex-predictive-echo.test.ts', // also needs a real codex binary
   'test/split-pane-terminal.browser.test.ts',
   'test/terminal-tile-scroll.browser.test.ts',
+  'test/terminal-touch-paging.browser.test.ts',
   'test/shift-enter-keypress.browser.test.ts',
   'test/key-tester.browser.test.ts',
   'test/webhook-settings.browser.test.ts',
@@ -45,6 +49,7 @@ export const BROWSER_TEST_GLOBS = [
   'test/spreadsheet-preview.browser.test.ts',
   'test/mobile-ime-preview.browser.test.ts',
   'test/run-mode-menu-scroll.browser.test.ts',
+  'test/markdown-anchor-links.browser.test.ts',
   'test/fullscreen-tab-switch-capture.browser.test.ts',
 ];
 

docs/api-reference.md

@@ -43,8 +43,9 @@ payload return `{ "success": true, "data": {} }`.
 
 > Non-JSON endpoints are exempt from the envelope: `GET /api/sessions/:id/file-raw`,
 > `GET /api/sessions/:id/tail-file` (SSE), `GET /api/download`,
-> `GET /api/screenshots/:name`, `GET /q/:code` (QR redirect), and the
-> `GET /ws/sessions/:id/terminal` WebSocket upgrade.
+> `GET /api/screenshots/:name`, `GET /q/:code` (QR redirect),
+> `GET /html-view/:cap/*` (a [rendered HTML file](#rendered-html-files) and its
+> assets), and the `GET /ws/sessions/:id/terminal` WebSocket upgrade.
 >
 > **Deprecated:** `POST /api/screenshots`, `GET /api/screenshots` and
 > `GET /api/screenshots/:name` keep working but log a one-time warning on first
@@ -77,16 +78,108 @@ the HTTP status.
 |-------------|------|---------|
 | `INVALID_INPUT` | 400 | Malformed request / failed validation |
 | `UNAUTHORIZED` | 401 | Authentication required or failed |
+| `FORBIDDEN` | 403 | Signed in, but not allowed (e.g. a non-admin calling an admin-only route in multi-user mode) |
+| `PASSWORD_CHANGE_REQUIRED` | 403 | Multi-user: the user must change their password before anything else |
 | `NOT_FOUND` | 404 | Resource does not exist |
+| `USER_NOT_FOUND` | 404 | Multi-user: no user with that name |
 | `SESSION_BUSY` | 409 | Session is busy |
 | `CONFLICT` | 409 | Conflicts with current state (e.g. already running) |
 | `ALREADY_EXISTS` | 409 | Resource already exists |
+| `USER_EXISTS` | 409 | Multi-user: a user with that name already exists |
+| `LAST_ADMIN` | 409 | Multi-user: refusing to demote, disable or delete the last enabled admin |
 | `OPERATION_FAILED` | 422 | Well-formed but could not be completed |
 | `RATE_LIMITED` | 429 | Too many requests |
 | `INTERNAL_ERROR` | 500 | Unexpected server error |
 
 Adding a new error code is non-breaking; removing or renaming one is a major change.
 
+## Cron jobs
+
+Saved jobs and their launch history are separate from the legacy `/api/scheduled`
+duration-bounded loops. Use `/api/v1/cron/...` in external clients; `/api/cron/...`
+is the unversioned alias. These routes use the response envelope above; the table
+lists the value inside `data` on success.
+
+| Method | Path | Request body | Response `data` |
+| --- | --- | --- | --- |
+| GET | `/api/v1/cron/jobs` | None | `CronJob[]` |
+| POST | `/api/v1/cron/jobs` | Full job definition below | `{ job: CronJob }` |
+| GET | `/api/v1/cron/jobs/:id` | None | `CronJob` |
+| PUT | `/api/v1/cron/jobs/:id` | Partial job definition | `{ job: CronJob }` |
+| DELETE | `/api/v1/cron/jobs/:id` | None | `{}` |
+| PUT | `/api/v1/cron/jobs/:id/enabled` | `{ enabled: boolean }` | `{ job: CronJob }` |
+| POST | `/api/v1/cron/jobs/:id/run` | None | `{ run: CronJobRun, activeAgents: number }` |
+| GET | `/api/v1/cron/jobs/:id/runs` | None | `CronJobRun[]` |
+| GET | `/api/v1/cron/runs` | None | `CronJobRun[]` |
+
+### Job request fields
+
+The create body requires `name`, `agentType`, `workingDir`, `promptMode`,
+`inputMode`, `scheduleType`, `enabled`, and `concurrencyPolicy`. Additional fields
+are required according to the selected prompt and schedule:
+
+| Field | Type / validation |
+| --- | --- |
+| `name` | String, 1–200 characters |
+| `agentType` | A supported session mode (including `shell`) |
+| `workingDir` | Existing, allowed working-directory path |
+| `launchCommand` | Optional single-line string, at most 2000 characters; for shell jobs |
+| `promptMode` | `inline_text` or `prompt_file_path` |
+| `promptText` | Required for `inline_text`; nonempty single-line string, at most 100000 characters |
+| `promptFilePath` | Required for `prompt_file_path`; absolute path inside `workingDir` to a regular file, at most 1 MiB, read when the job fires |
+| `inputMode` | `paste` or `typed` |
+| `scheduleType` | `once`, `interval`, `daily`, or `weekly` |
+| `runAt` | Required for `once`; positive integer Unix timestamp in milliseconds |
+| `intervalMinutes` | Required for `interval`; integer from 1 to 525600 |
+| `dailyTime` | Required for `daily`; `HH:MM` in server-local time |
+| `weeklyDays` | Required for `weekly`; 1–7 weekday integers, 0 (Sunday) through 6 (Saturday) |
+| `weeklyTime` | Required for `weekly`; `HH:MM` in server-local time |
+| `enabled` | Boolean |
+| `concurrencyPolicy` | `warn_only` or `skip_if_same_agent_running`; scheduled runs only |
+| `autoClosePreviousSession` | Optional boolean, default `true`; ignored for `once` |
+| `notes` | Optional string, at most 2000 characters |
+
+`PUT /jobs/:id` accepts any subset of these fields, then validates the merged job.
+When changing `promptMode` or `scheduleType`, supply the fields the new mode needs.
+`Run Now` works even when the job is disabled, bypasses the scheduled concurrency
+policy, and does not change the schedule. `activeAgents` counts live sessions of
+the same agent type, excluding sessions created by this job.
+For recurring jobs with `autoClosePreviousSession` enabled (the default), `Run Now`
+also closes the previous run's session before launching, even if it is still working.
+
+### Job and run response fields
+
+`CronJob` contains the request fields plus server-maintained `id`, optional
+`owner` (multi-user mode), `createdAt`, `updatedAt`, `lastRunAt`, `nextRunAt`,
+`lastStatus`, `lastDueKey`, and optional `completedOnce`. Times are Unix
+milliseconds; `lastRunAt`, `nextRunAt`, `lastStatus`, and `lastDueKey` can be `null`.
+`lastDueKey` is an opaque internal duplicate-launch guard, not a stable API format.
+
+`CronJobRun` contains `id`, `cronJobId`, nullable `sessionId` and `sessionName`,
+`startedAt`, nullable `finishedAt`, `status`, optional `errorMessage`,
+`triggerType` (`scheduled` or `manual_run_now`), and nullable `createdSessionUrl`.
+Run times are also Unix milliseconds. Status is one of `created`,
+`session_started`, `prompt_sent`, `failed`, or `skipped`.
+
+Prompt delivery continues asynchronously after session launch, so `Run Now` can
+return `session_started` before the prompt is sent. Read run history for subsequent
+updates, but do not assume a terminal status will follow: if the session is closed
+during the readiness wait or the server restarts before delivery, the run can remain
+`session_started` indefinitely with `finishedAt: null`.
+`finishedAt` refers to the launch/prompt-delivery attempt, **not completion
+of the agent's task**; `prompt_sent` does not prove that the task succeeded.
+
+In multi-user mode, list/history endpoints filter to accessible jobs. An unknown
+or inaccessible job returns `NOT_FOUND`. Job creation and updates can return
+`403 FORBIDDEN` for a working directory outside the owner's workspace or a shell /
+launch-command job without the required privilege grant. Invalid definitions or
+working directories return `INVALID_INPUT`; launch/delivery failures are recorded
+on the run, so inspect its `status` and `errorMessage` even after an HTTP success.
+
+See [Cron Jobs](wiki/Cron-Jobs.md) for the UI, scheduling, and prompt-file rules.
+See the [complete cron guide](cron-guide.md) for the `cron:runCreated` and
+`cron:runUpdated` SSE events.
+
 ## Long-polling (agent wait)
 
 Three calls block until something happens instead of answering immediately. They
@@ -117,15 +210,17 @@ auto-retried by several clients (silently doubling the polling load), `504` is w
 a genuine tunnel failure looks like, and `204` cannot carry `waitedMs` / `status` /
 `limitPaused`. Reserve error handling for the four codes in the table below.
 
-**2. `stop` and `blocked` fire only for `claude` sessions.** Both come from Claude
-Code hooks, and no other mode installs them: `shell` runs no agent, and the external
-CLIs (`opencode`, `codex`, `gemini`, `antigravity`, `pi`) render their own TUIs and post
-no hooks. For every non-`claude` mode only `idle`, `working` and `exit` are
-accepted, and of those only `exit` is dependable: see the caveats under
-[Signals](#signals) before building on `idle`. Requesting `stop` or `blocked`
-**explicitly** on such a session is a
-`400`; omitting `until` never fails, the server just drops them from the default set
-and echoes the narrowed set back as `wait.until`. Three more places hooks can go
+**2. `stop` fires only for `claude` (and `deepseek`) sessions, `blocked` also for
+`codex`.** Both come from Claude Code hooks, and no other mode installs them: `shell`
+runs no agent, and the external CLIs render their own TUIs and post no hooks
+(`deepseek`'s status bridge reports the same two, see its own docs). The one exception
+is `blocked` on `codex`: Codeman reads codex's approval, question and trust dialogs off
+its screen and raises `blocked` for them (see [Signals](#signals)). For every other mode
+only `idle`, `working` and `exit` are accepted, and of those only `exit` is
+dependable: see the caveats under [Signals](#signals) before building on `idle`.
+Requesting a signal the session can never emit **explicitly** is a `400` (`stop` on
+`codex` included); omitting `until` never fails, the server just drops them from the
+default set and echoes the narrowed set back as `wait.until`. Three more places hooks can go
 missing even in `claude` mode: a **Docker case** needs
 `CODEMAN_DOCKER_BRIDGE_HOOKS=1`, since a container cannot reach a loopback-bound
 Codeman (without it, only `idle` / `working` / `exit` work); a **remote-SSH
@@ -151,8 +246,8 @@ contract is a **marker unique to each call** (`MARK="DONE_$RANDOM"`, send
 |--------|--------|--------------------|
 | `idle` | the session's own `idle` event | `claude`: yes, on ❯-prompt detection after activity. `shell`: **once only**, ~500 ms after start, and never again. External CLIs: not guaranteed (they render their own TUIs and readiness is output stabilization) |
 | `working` | the session's own `working` event | `claude` only in practice (spinner and work-keyword detection are Claude output formats) |
-| `stop` | the Claude Code `stop` hook, the definitive end-of-turn signal | `claude` only |
-| `blocked` | a `permission_prompt` or `elicitation_dialog` hook | `claude` only, and rarer than it looks: see below |
+| `stop` | the Claude Code `stop` hook, the definitive end-of-turn signal | `claude` (and `deepseek`) only |
+| `blocked` | a `permission_prompt` or `elicitation_dialog` hook; on `codex`, a dialog read off its screen | `claude`, `deepseek` and `codex`; on `claude` rarer than it looks: see below |
 | `exit` | no process is behind the session | every mode |
 
 `stop` is the signal to orchestrate on where it exists; `idle` is a heuristic
@@ -162,9 +257,9 @@ the wait promptly instead of burning the caller's whole timeout on something tha
 can no longer happen). On a `claude` worker, prefer an explicit `until=stop,exit`
 once the session is up: the default set's `idle` also resolves on a spinner pause,
 and on a fresh session the **startup** `idle` (emitted when the CLI first comes up)
-can land inside your first wait window and report a turn that never ran. Measured:
-a session parked on the trust dialog emits no *further* `idle`, so it is the
-startup transition, not the dialog, that produces the false success below.
+can land inside your first wait window and report a turn that never ran. That
+startup `idle` is held while Claude's workspace-trust dialog is on screen (inside
+the 90 s startup window), because the dialog's own ❯ is not a composer.
 
 ⚠️ **`exit` means "nothing is running", which includes "not started yet".** The
 server answers from `pid === null` plus a mux-layer pane-death probe, and that
@@ -189,6 +284,17 @@ asking the user a question. So `until=stop,blocked,exit` is a reasonable belt on
 long turn, but a worker that never comes back is far more likely to be working than
 blocked, and polling `blocked` alone will sit at its timeout.
 
+**`blocked` on `codex` is read off the screen**, because codex has no hooks and is
+launched with its approvals on unless `codexConfig.dangerouslyBypassApprovals` is set.
+A codex pane showing a command, edit or network approval, a question (`request_user_input`)
+or the startup trust prompt raises `blocked` (plus the red alert, push and webhook a Claude
+`permission_prompt` raises, alert-only: answer it in the terminal), and the session reads
+`idle` while it is up. Unlike a hook, this is a level the server keeps reading, so a wait
+for `blocked` that arrives while the dialog is already up answers at once
+(`immediate: true`); a wait that did not ask for `blocked` is unaffected. It clears 2 to 4 s
+after the dialog leaves the screen. `stop` stays a `400` on codex: nothing on a screen says a
+turn ended.
+
 ⚠️ **On a `shell` session, only `exit` and marker-matching are dependable.** A shell
 session emits its one `idle` at startup and then stays `status: "idle"` forever,
 whatever the pane is doing, so it never emits a *transition*. Since send-and-wait
@@ -202,11 +308,11 @@ instead. The same caution applies to the external CLIs.
 Nothing here reports "the agent is ready for a prompt", and no combination of
 `until`/`fresh` synthesizes one. A freshly created session reads as `exit` (above),
 and a `claude` worker in a brand-new case comes up on the CLI's **trust dialog**,
-which contains a ❯ prompt of its own. Send-and-wait posted at that moment types the
-prompt into the dialog, where the `\r` never gets past it, while the session's
-startup `idle` lands inside the wait window: the wait resolves on `idle` in a
-couple of seconds with `timedOut: false`, which looks exactly like a completed
-turn.
+which contains a ❯ of its own. A prompt posted at that moment is typed into the
+dialog, and its `\r` confirms the highlighted `No, exit`: the CLI quits and the
+pane dies, so send-and-wait resolves on `exit`. Codeman holds the session's `idle`
+while the dialog is on screen, so waiting for `idle` before prompting no longer
+clears too early, but that still does not make `idle` a readiness signal.
 
 The reliable sequence is: poll `GET /api/v1/sessions/:id` until `.data.pid` is
 non-null, then `wait-output` for the composer's own marker (`bypass`, the status
@@ -432,7 +538,7 @@ read the timeout as "the worker is wedged" and kill a session that was working f
 
 | `errorCode` | HTTP | When |
 |-------------|------|------|
-| `INVALID_INPUT` | 400 | unknown `until` / `wait` token; `stop` or `blocked` requested explicitly on a mode that installs no hooks (the message names the mode); `regex=` on `/wait-output`; `match` outside 1 to 200 chars; a non-numeric `timeout` |
+| `INVALID_INPUT` | 400 | unknown `until` / `wait` token; `stop` or `blocked` requested explicitly on a mode that can never emit it (`stop` without hooks, `blocked` without hooks or a screen-read dialog; the message names the mode); `regex=` on `/wait-output`; `match` outside 1 to 200 chars; a non-numeric `timeout` |
 | `NOT_FOUND` | 404 | no such session, or one this caller does not own |
 | `SESSION_BUSY` | 409 | this session's waiter cap is full |
 | `RATE_LIMITED` | 429 | a per-owner or process-wide waiter cap is full. Retry later; the session you named is not the problem |
@@ -456,20 +562,120 @@ client that opens many concurrent waits against one session will still hit the c
 
 What a session's terminal shows, for a client to replay: `data.terminalBuffer`,
 with `source` (`mux-visible`, `mux-full-history` or `history`), `truncated`,
-`truncationReason`, `fullSize`, and `captureCols`/`captureRows` when the pane's
-geometry was read. `paneHistoryLines` (present whenever the body is a pane capture)
-is the number of scrollback rows tmux holds above the visible frame, which is the
-most a `full=1` request can add. `truncated` describes the byte stream instead: for a
-pane with `paneHistoryLines: 0` (a fullscreen CLI in the alternate screen) the bytes a
-`tail` cut dropped are earlier repaints that no request returns. The capture runs
-synchronous tmux calls on the server; the `Server-Timing` header reports `capture`,
-`prepare` and `total`.
+`truncationReason`, `fullSize`, `historyLinesCut` (a `lines=` bound left older
+lines in tmux, so the remainder is more than `fullSize - retainedBytes`), and
+`captureCols`/`captureRows` when the pane's geometry was read. `paneHistoryLines`
+(present when the body is a pane capture whose geometry was read, like
+`captureCols`/`captureRows`, and absent for a remote SSH session, whose local pane
+cannot see the remote history) is the number of scrollback rows tmux holds above the
+visible frame, which is the most a `full=1` request can add. `truncated` describes
+the byte stream instead: for a pane with `paneHistoryLines: 0` (a fullscreen CLI in
+the alternate screen) the bytes a `tail` cut dropped are earlier repaints that no
+request returns. The capture runs synchronous tmux calls on the server; the
+`Server-Timing` header reports `capture`, `prepare` and `total`.
 
 | Query | Meaning |
 |---|---|
 | `full=1` | tmux's scrollback, not only the visible frame (`source: 'mux-full-history'`), ending with a relative cursor move back to the pane's caret. |
 | `tail=<bytes>` | Keep the newest `<bytes>` of the result (`truncationReason: 'tail'` when it cut). |
-| `lines=<n>` | With `full=1` only: read at most `<n>` lines of tmux history above the visible frame. An integer of at least 1, clamped to the configured history limit; absent or malformed, the whole limit (100,000 lines by default), as before. `truncated` and `truncationReason` describe byte cuts only, not this bound. Without it a full capture reads all of that history before `tail` cuts it, so a client that keeps a fixed number of lines (the tile grid sends its xterm's scrollback plus its rows) should send it. |
+| `lines=<n>` | With `full=1` only: read at most `<n>` lines of tmux history above the visible frame. An integer of at least 1, clamped to the configured history limit; absent or malformed, the whole limit (100,000 lines by default), as before. When a bound below the configured limit left older lines in tmux (its `history_size` is deeper), `truncated` is true with `truncationReason: 'tail'`; a byte cap still reports `'capped'`. Without it a full capture reads all of that history before `tail` cuts it, so a client that keeps a fixed number of lines (the tile grid sends its xterm's scrollback plus its rows) should send it. |
+| `alt=recording` | With `full=1` only: if the pane is in the alternate screen (a nested tmux, where tmux keeps no history, or any full-screen app such as vim or less while it is open), answer with the visible frame over the byte recording instead (`source: 'mux-visible'`, `tail` applied as without `full=1`). Shells' automatic loads send it. |
+
+## The `codeman agent` CLI (client over these endpoints)
+
+`codeman agent ls|spawn|send|wait|read|interrupt|rm` (`src/cli-agent.ts`) is the command-line client for the endpoints above, for agents the skill does not reach (a pi or shell session, a machine without `jq`). It adds no route: `spawn` is `POST /api/v1/quick-start` (+ `wait-output` on the mode's `capabilities.composerReadyMark` from the CLI registry, where it declares one), `send` is `POST …/input` with `clientId`+`seq` (and `wait`/`waitTimeout` for `--wait` / `--until <signals>`; `delivered:false` without `duplicate` and `wait.ended` both exit 3 — the CLI never reports a dead worker as done), `wait` is `GET …/wait` (`--until`) or `GET …/wait-output` (`--match`, `from=buffer` by default), `read` is `GET …/last-response` or `GET …/terminal?tail=`, `interrupt` is `POST …/input` with a bare `\u001b`, `rm` is `DELETE …/sessions/:id`. A fire-and-forget `send` to a sleeping wake-on-LAN host reads the route's `buffered` (own line, exit 0) and `dropped` (exit 1: the chunk is gone). An id may be the 8-character form `ls` prints, resolved through `GET /api/v1/sessions`; anything shorter refuses before any request, the same floor as `PARENT_SESSION_ID_MIN_PREFIX`. Every call carries `X-Codeman-Parent-Session`; only `spawn`'s quick-start carries `X-Codeman-Agent-Origin: codeman-agent-cli` (the agent-scratch label must never reach a request that cannot create the case directory). Basic auth comes from `CODEMAN_PASSWORD` or the data dir's `.env`. Server-side error codes are shown verbatim (`INVALID_INPUT: until=stop …` on a hook-less mode is not hidden); exit codes are `0` ok, `1` error, `2` timeout, `3` the session exited, `4` refused by a client-side guard. See the README section "`codeman agent`" for the guards and `test/cli-agent.test.ts` for the pinned behaviour.
+
+`codeman mcp` (`src/mcp-server.ts`) is a Model Context Protocol server on stdio over the same endpoints, for MCP clients: `list_sessions` (`GET /api/v1/sessions`), `start_session` (`POST /api/v1/quick-start` with `X-Codeman-Agent-Origin: codeman-mcp`, then `wait-output` on the mode's ready mark), `send_prompt` (`POST …/input` with `wait`/`waitTimeout`), `wait_for_session` (`GET …/wait` or `…/wait-output`), `read_output` (`GET …/last-response` or `…/terminal?tail=`), `interrupt_session` and `delete_session`. It adds no route. The URL comes from `CODEMAN_API_URL` or `--url`, never a guess; prompting, interrupting and deleting are limited to sessions the process started (or the calling session's own children) unless it runs with `--allow-any-session`. See the README section "MCP server" and `test/mcp-server.test.ts`.
+
+## Prompt uploads (`POST /api/v1/sessions/:id/paste-image`)
+
+A `multipart/form-data` body with one `image` part. The file is written into the
+session's workspace as `<workingDir>/.codeman-uploads/paste-<ms>-<hex>.<ext>`, and
+`data` carries `path` and `filename` for the client to type the path into the
+prompt. The folder is Codeman's own: hidden, created on first use with a
+`.gitignore` containing `*` (written once, never over a file already there), and
+cleaned up the way pasted images always were: `paste-*` files older than 7 days
+go in an hourly sweep, and the folder goes when the last session of that
+workspace is killed. Uploads made before this release sit in `.claude-images/`;
+that folder receives nothing new, and is swept and removed the same way for one
+release. A remote (SSH) session answers 400, since the file would land on the
+Codeman host under a path the remote agent cannot read. A Docker session of an
+owned case is fine, its workspace is bind-mounted at the same absolute path; an
+adopted container (`owned: false`) mounts nothing, so its agent can open the file
+only if the container itself exposes that host path.
+
+The optional query flag `?kind=video` selects a second path for a camera-roll
+video: the file streams to disk instead of being buffered, under a far larger
+per-file cap (`MAX_PASTE_VIDEO_BYTES`, 1 GB by default, env
+`CODEMAN_MAX_PASTE_VIDEO_BYTES`), and only video container types are accepted
+(mp4, webm, mov, m4v, ogv, single-sourced from the file viewer's playable set).
+The container is checked from the first bytes as
+the body arrives; a file whose bytes do not match its declared type is unlinked
+and answered 415, one over the cap is unlinked and answered 413, and a transfer
+that dies midway leaves nothing behind. Because one video can be 1 GB,
+`?kind=video` draws on its own rate limit, 3 uploads per minute per (IP, session),
+beside the 30 per minute the image path keeps; without the flag the route is
+byte-identical to the image-only version, same 50 MB image cap and same 30/min
+bucket. The response shape is unchanged (`path`, `filename`).
+
+Beyond images, the route also accepts documents a user hands an agent to read,
+admitted by the name's extension (the Content-Type a phone attaches to a `.md`
+or `.csv` is unreliable): the office formats `.pdf`, `.docx`, `.pptx`, `.xlsx`,
+`.zip` and the text family `.txt`, `.md`, `.csv`, `.tsv`, `.json`, `.log`,
+`.yaml`, `.yml`, `.xml`. A document is buffered under the same 50 MB image cap,
+its bytes are checked against the extension (a magic-number signature, or UTF-8
+with no NUL byte for text) and answered 415 on a mismatch, and it is saved under
+a sanitized stem of its name, so `Q3 report.pdf` is findable as
+`paste-<ms>-<hex>-Q3-report.pdf`. A named type that is neither an image nor a
+document (a script, an HTML page, a `.sh`) is refused 400 `Unsupported file
+type`. That refusal is a scope choice, not a security property: `file-raw`
+already serves non-media text inert (`text/plain`, `nosniff`), so this just
+keeps the server from storing what it would serve or an agent could run, and
+html and svg are the ones deliberately kept out.
+
+## Rendered HTML files
+
+An `.html`/`.htm` file renders as a page in a sandbox, in two steps
+(`src/web/routes/html-view-routes.ts`).
+
+**`POST /api/v1/sessions/:id/html-view`** with `{ "path": "<file>" }` (absolute,
+relative to the session's workspace, or `~/` for this host's home, which a Docker
+case refuses with `400` since its `~` is the container's; a non-empty string of at most 4096 characters
+without NUL or line breaks, and no other keys) answers `data: { url }`, a
+`/html-view/<cap>/<file>` URL. The capability grants GET access to the files under
+that file's own folder, so a page's relative CSS, scripts, images and data files
+resolve. It is held in memory only (a restart invalidates it), has a rolling TTL
+(`CODEMAN_WEBVIEW_CAPABILITY_TTL_MS`, default 12 h), is reused for the same folder,
+and is revoked on logout, admin logout and user deletion. Refusals:
+
+- `400`: an invalid body, a remote (SSH) session, or a file that is not HTML (the
+  symlink-resolved name counts too);
+- `404`: no such file;
+- `403`: a file outside the workspace that the attachment guard blocks (a blocked
+  tree, or `attachmentConfineToWorkspace` on), a file in a hidden directory (inside
+  the workspace, one below the workspace root), or a file directly in a folder too
+  broad to serve recursively: `/`, the home folder, `/tmp`, `/var/tmp`, `$TMPDIR`,
+  the cases and user-space roots, a drive's mount point (`/mnt/d`), or any folder
+  above one of those (`/home`, `/Users`).
+
+**`GET /html-view/:cap/*`** serves the page and its assets. It needs no cookie: the
+capability in the path is the credential, because the page runs in an opaque origin
+that never sends one. It is not under `/api`, so it has no `/api/v1` alias and no
+envelope. `GET` and `HEAD` only; a CORS preflight on a live capability is answered
+with `Access-Control-Allow-Methods: GET, HEAD`. What it serves: files under the
+capability's folder whose path (requested and symlink-resolved) has no dot segment
+and whose extension is on an asset allowlist (html, css, js/mjs, json, map, txt,
+csv, md, svg and common image, font, wasm, audio, video and PDF types), never
+outside the folder, never a blocked tree, and, for a folder outside the workspace,
+not while `attachmentConfineToWorkspace` is on. Every response carries
+`Content-Security-Policy: sandbox allow-scripts allow-forms allow-popups allow-modals allow-downloads`
+as its whole policy (never `allow-same-origin`, so the page cannot read Codeman's
+document, cookies or API, even opened as a top-level tab; and no fetch directive, so
+by the owner's decision the page may load from and send requests to any host), `Cache-Control: no-store`,
+`Referrer-Policy: no-referrer` and `Access-Control-Allow-Origin: *`. A `Range`
+request is answered with `206` + `Content-Range` (`416` past the end), like
+`file-raw`. Every refusal is a plain-text `404 Not found`; with a password set, an
+unknown or revoked capability is the auth hook's `401` instead.
 
 ## Session lineage (`parentSessionId`)
 
@@ -522,18 +728,67 @@ most 64 characters, but treat it as untrusted text. A `statusline` or `screen` v
 persisted and restored after a server restart until the next report replaces it; a
 `config` value is read again at every pane start, attach and relaunch instead.
 
+## Prompt cache (`promptCache`)
+
+Claude-mode session state (`GET /api/v1/sessions`) carries the main conversation's
+prompt-cache state as Claude Code last reported it on the statusline (v2.1.251+), with a
+change-only `session:promptCache` event (`{ sessionId, ...promptCache }`):
+
+```json
+"promptCache": { "warm": true, "ttl": "1h", "expiresAt": 1791653039000, "recacheTokensIfCold": 45000, "misses": 2, "expectedRebuilds": 1, "hitRatio": 0.91, "lastMissCauses": ["ttl_expired_1h"], "compactedAt": 1791649439872 }
+```
+
+| Field                 | Meaning                                                                                                   |
+| --------------------- | --------------------------------------------------------------------------------------------------------- |
+| `warm`                | Whether the cached prefix is still within its TTL. The only field always present.                         |
+| `ttl`                 | `5m` or `1h`; absent when Claude reported neither.                                                         |
+| `expiresAt`           | Epoch MILLISECONDS when the prefix goes cold (the statusline reports seconds). Absent once cold.           |
+| `recacheTokensIfCold` | Tokens the next request re-writes if the cache has gone cold. Absent right after a compaction.             |
+| `misses`              | Requests that re-processed content the cache already held.                                                |
+| `expectedRebuilds`    | Cache rebuilds that followed a compaction or a clearing of old tool results.                               |
+| `hitRatio`            | Cache read tokens as a fraction of all input tokens this session, 0 to 1.                                  |
+| `lastMissCauses`      | Claude Code's diagnosis of the last miss (e.g. `ttl_expired_1h`, `tools_changed`); untrusted text.         |
+| `compactedAt`         | Epoch MILLISECONDS when the context was compacted or its old tool results cleared: the first warm report without `recacheTokensIfCold` after one that had it. Derived, not reported; dropped at the next request. |
+
+The whole object is absent until the first report, and restored after a server restart
+(so the readout is not blank until the pane's next render) rather than recomputed; a value
+whose `expiresAt` has passed reads as cold. The statusline re-runs when a warm cache hits
+`expiresAt`, so the cold flip arrives on its own. A post-`/clear` conversation switch drops
+it. Shown in Session Options → Respawn and, for sessions without a statusline of their own,
+as a `cache:until HH:MM` / `cache:cold` group in the terminal footer.
+
+Beside it, `statusLineExporter` says whether this launch carries the statusline exporter
+that sends the report, decided when the pane starts: `injected` (it does), `collection-off`
+(Plan Usage was off at launch), `workspace-statusline` (the workspace's
+`.claude/settings.local.json` sets its own `statusLine`, which Codeman never replaces) or
+`remote` (remote and Docker sessions). Absent when unknown.
+
+## Placing a new session in a tab group (`tabGroupId`)
+
+`POST /api/v1/sessions` and `POST /api/v1/quick-start` accept an optional
+`tabGroupId` (1 to 100 characters): the id of a group in the caller's tab layout
+(`GET /api/v1/tab-layout`). The session is created as usual and placed at the end
+of that group in the same layout write; the response then also carries the
+committed layout as `tabLayout`. An id the caller's layout does not contain (a
+deleted group, another user's group) is ignored and the session lands where it
+otherwise would, never a `400`. A malformed value (`null`, an empty string, a
+non-string) is a `400 INVALID_INPUT`. Requests without the field are unchanged.
+
 ## Approvals Inbox
 
 Cross-session queue of prompts waiting on a human (permission dialogs,
-AskUserQuestion questions, idle prompts). Claude-mode sessions only; items are
-in-memory (a server restart drops them; the next prompt re-fires the hook).
+AskUserQuestion questions, idle prompts). Answerable items come from Claude-mode
+sessions only; DeepSeek Harness (its status bridge) and Codex (dialogs read off its
+screen) open **alert-only** items with `answerInTerminal: true`. Items are
+in-memory (a server restart drops them; the next prompt re-fires the hook, and a
+codex dialog still on screen is read again after the pane re-attaches).
 Design: [`approvals-inbox-plan.md`](approvals-inbox-plan.md).
 
 - `GET /api/v1/approvals` → `{ approvals: ApprovalItem[] }`, oldest first,
   ownership-scoped in multi-user mode. `ApprovalItem`: `{ id, sessionId,
   sessionName, kind: 'permission'|'question'|'idle', createdAt, toolName?,
   toolSummary?, message?, cwd?, context?, options?: {n, label}[],
-  acknowledgedAt? }`. `context` is the ANSI-stripped visible pane frame;
+  acknowledgedAt?, answerInTerminal? }`. `context` is the ANSI-stripped visible pane frame;
   `options` is present only when the dialog's numbered choices parsed
   confidently; `acknowledgedAt` marks an item a human has already looked at
   (see `/viewed` below) and tells clients not to re-arm its tab alert. Listing
@@ -547,8 +802,9 @@ Design: [`approvals-inbox-plan.md`](approvals-inbox-plan.md).
   (sends the digit; accepted only when `n` is among the item's parsed
   `options`), or `{ action: 'text', text }` (idle prompts only; submits the
   line as a prompt). `404 NOT_FOUND` when the item is no longer pending,
-  `409 CONFLICT` when the dialog left the screen or another actor answered
-  first, `422 OPERATION_FAILED` when the session refused input.
+  `400 INVALID_INPUT` for an `answerInTerminal` item (`<CLI> approvals must be
+  answered in the terminal`), `409 CONFLICT` when the dialog left the screen or
+  another actor answered first, `422 OPERATION_FAILED` when the session refused input.
 - `POST /api/v1/approvals/:id/dismiss` removes the item without keystrokes.
 - `POST /api/v1/approvals/session/:sessionId/viewed` → `{ sessionId,
   acknowledged: itemId | null }`. Marks the session's pending **idle** item as
@@ -792,6 +1048,70 @@ Both routes accept two optional query parameters, which the UI sends from its pe
 
 Lists are capped (300 files and 50 commits per repository) while the counts stay exact. A branch with no upstream reports the commits no remote has (`HEAD --not --remotes`); a repository with no remote reports `unpushedCount: 0`, since there is nothing to push to. Concurrent polls of one folder share a single git invocation; `?fresh=1` (what the panel's Refresh button and opening the panel send) skips the short-lived caches, though it still joins a computation already running.
 
+## Git worktrees of a case
+
+A local case whose folder is the top of a git repository can keep **worktrees** at `<case>/.claude/worktrees/<name>`, each on its own branch, so several agents on one repository never share a checkout. This is the layout Claude Code's own `claude --worktree <name>` uses (branch `worktree-<name>`), so either tool finds the other's worktrees. Remote (SSH) and Docker cases are refused with `400 INVALID_INPUT` before git runs, and so is a local case whose folder is at or inside a Docker case workspace: `git worktree add` runs the repository's post-checkout hook and checkout filters on this machine, and a container can write there. A case folder that is not a repository, is a bare repository, or sits below the repository's top level is `400` too. The case resolves the way quick-start resolves it (a linked case is admin-only in multi-user mode, everything else is the caller's own case space; another user's case is `404`), and a case folder on a mount that does not answer is `422 OPERATION_FAILED`, never `404`.
+
+`GET /api/cases/:name/worktrees` lists `git worktree list` for the case's repository:
+
+```json
+{
+  "caseName": "myrepo",
+  "repoRoot": "/home/me/codeman-cases/myrepo",
+  "worktreesDir": "/home/me/codeman-cases/myrepo/.claude/worktrees",
+  "truncated": false,
+  "worktrees": [
+    { "path": "/home/me/codeman-cases/myrepo", "head": "6910cb15…", "branch": "main", "detached": false, "bare": false,
+      "locked": false, "prunable": false, "main": true, "isCase": true, "managed": false, "sessionCount": 0, "sessions": [] },
+    { "path": "/home/me/codeman-cases/myrepo/.claude/worktrees/fix-login", "branch": "worktree-fix-login",
+      "main": false, "isCase": false, "managed": true, "name": "fix-login",
+      "sessionCount": 1, "sessions": [{ "id": "2a568379-…", "name": "w1-fix-login" }] }
+  ]
+}
+```
+
+`managed` is `true` only for `<repoRoot>/.claude/worktrees/<name>`: those are the ones Codeman starts sessions in and removes; any other worktree of the repository is listed for information. `sessionCount` counts live sessions of every owner running in that worktree (a session belongs to the deepest worktree holding its folder), `sessions` lists the ones the caller may see. `lockReason`/`prunableReason` appear when git reports them.
+
+`POST /api/cases/:name/worktrees` with `{ name, branch?, base? }` creates one (`201`, `{ worktree: { name, path, branch, base, head, excludeAdded, warnings } }`):
+
+- `name`: 1 to 64 characters of `A-Z a-z 0-9 . _ -`, starting with a letter or digit, no `..`, not ending in `.` or `.lock`. `400` otherwise.
+- `branch`: the NEW branch, default `worktree-<name>`. Letters, digits, `.`, `_`, `-`, `/`, then `git check-ref-format --branch` must accept it unchanged. An existing branch is `409 ALREADY_EXISTS` (never reset), and so is a name whose folder already exists.
+- `base`: where the branch starts, default `HEAD` (the case's current commit): a branch, tag or commit, resolved before anything is written (`400` when it is not a commit, which includes an empty repository).
+- `excludeAdded` lists the lines appended to the repository's local `info/exclude` (in the real common git dir, never a tracked `.gitignore`) when git did not already ignore them: `/.claude/worktrees/`, so the main checkout does not show the worktrees as untracked, `/.claude/settings.local.json`, the hooks file Codeman writes for every claude session, which would otherwise make `git worktree remove` refuse, and the agent skill copy (`/.claude/skills/codeman/`, `/.agents/skills/codeman/`: every skills folder the CLI registry names) that sessions get while `agentSkillEnabled` is on, for the same reason.
+- A symlinked `.claude` or `.claude/worktrees` directory is refused (`400`): a repository cannot redirect the worktree outside itself.
+
+`DELETE /api/cases/:name/worktrees/:worktree[?deleteBranch=true]` removes a managed worktree with `git worktree remove` **without `--force`**: git refuses a worktree with modified or untracked files (or a locked one), and that refusal is the answer, `409 CONFLICT` with git's own words in `error`. A live session at or inside the worktree is `409` as well. With `deleteBranch=true` the branch is deleted with `git branch -d`, never `-D`, so an unmerged branch is kept: the response is still `200`, with `branchDeleted: false` and git's reason in `branchError`. Response: `{ removed: { name, path, branch, branchDeleted, branchError? } }`. A folder git does not list as a worktree of this repository, or a symlink in its place, is `404`/`400`. A worktree that is, sits inside, or holds a Docker case workspace is `400` too (git would run inside a folder the container writes, and the removal would delete that workspace under it). A managed worktree whose folder is gone (deleted by hand; the list reports it `prunable`) is still removable: git then only drops its record, which is what frees the name for a new worktree. A session never starts in one.
+
+**Starting a session in a worktree** goes through the normal create paths: `POST /api/quick-start` takes `worktree: "<name>"` next to `caseName` and runs the session (and the hooks, the agent skill and `modelOverride`) in the worktree folder; the response adds `worktree` and `workingDir`. It is refused for remote and Docker cases (`400`), for a worktree that is not a registered, managed one (`404`), and for a case that does not exist (`404`: a worktree launch never scaffolds a case). `POST /api/sessions` takes the worktree path as `workingDir` like any other folder.
+
+## Settings
+
+`~/.codeman/settings.json` is one file for the whole server: the synced App Settings (feature switches, the plan-usage collection switch `showPlanUsageLimits`, `runMode`, `respawnPresets`, `tunnelEnabled`, ...), plus `modelConfig`, `caseOrder` and `lastUsedCase`. Reads are open to every signed-in user; in multi-user mode a non-admin reads it without what is the admin's (`voiceSettings.apiKey`, the Deepgram key an admin saved, and the admin's case names in `caseOrder` and `lastUsedCase`). **In multi-user mode every write is admin only:** for a non-admin the routes below answer `403 FORBIDDEN` before the body is even validated, so nothing is written, no watcher is toggled and no tunnel is started. Single-user mode is unaffected.
+
+| Method | Path                          | Body                                 | Notes |
+| ------ | ----------------------------- | ------------------------------------ | ----- |
+| `GET`  | `/api/settings`               | none                                 | The stored settings object. Never writes. For a non-admin in multi-user mode, without `voiceSettings.apiKey`, `caseOrder` and `lastUsedCase`. |
+| `PUT`  | `/api/settings`               | partial settings (strict schema)     | Merged into the stored file; services (watchers, tunnel, tmux history limit) reconcile to the merged result. `403 FORBIDDEN` for a non-admin in multi-user mode. A tunnel enable without `CODEMAN_PASSWORD` is refused separately (`403 OPERATION_FAILED`, resend with `acknowledgeUnauthTunnel: true`). |
+| `GET`  | `/api/execution/model-config` | none                                 | `modelConfig` (cron's and the plan orchestrator's default model). |
+| `PUT`  | `/api/execution/model-config` | `{ defaultModel?, agentTypeOverrides?, ... }` | Replaces `modelConfig`. `403 FORBIDDEN` for a non-admin in multi-user mode. |
+| `PUT`  | `/api/cases/order`            | `{ order: string[] }`                | The one case order every user's `GET /api/cases` is sorted by. `403 FORBIDDEN` for a non-admin in multi-user mode. |
+| `PUT`  | `/api/config`                 | `{ pollIntervalMs?, defaultTimeoutMs?, maxConcurrentSessions?, respawn? }` (strict) | The process-wide app config in `state.json`. `403 FORBIDDEN` for a non-admin in multi-user mode. |
+
+A non-admin's `POST /api/quick-start` and `POST /api/ralph-loop/start` work as before but do not record their case as `lastUsedCase` (a server-wide key).
+
+### System operations
+
+The same multi-user rule covers the operations that act on the whole server. For a non-admin:
+
+| Method | Path                          | Multi-user non-admin |
+| ------ | ----------------------------- | -------------------- |
+| `POST` | `/api/system/update`          | `403 FORBIDDEN` (the update restarts the server under every user) |
+| `POST` | `/api/system/span-displays`   | `403 FORBIDDEN` (opens a window on the server's desktop) |
+| `POST` | `/api/tunnel/qr/regenerate`   | `403 FORBIDDEN` (it clears every user's outstanding QR login code) |
+| `POST` | `/api/cleanup-state`          | `403 FORBIDDEN` (drops every user's stale session records) |
+| `POST` | `/api/auth/revoke`            | Revokes only the caller's own login sessions: the bare form (`{}`) signs them out everywhere, and a `sessionToken` is revoked only when it is theirs. For an admin, and in single-user mode, the bare form still revokes every login session |
+| `GET`  | `/api/tunnel/info`            | `authSessions` lists only the caller's own logins |
+
 ## CLI management
 
 Read and write the CLI registry (`docs/cli-registry.md`). Every **write** route answers `403 FORBIDDEN` while `cliManagementEnabled` is off (the default), and for a non-admin in multi-user mode. A write that would overwrite a `clis.json` which does not parse, or which has group/world permission bits, is refused with `409 CONFLICT` and a message naming the fix; the file is left untouched.
@@ -817,10 +1137,10 @@ Copies MCP servers between the agent CLIs' own user-level config files (`docs/cl
 Result (`data`):
 
 - `applied` — `false` for the dry run.
-- `targets[]` — one per enabled CLI that declares an MCP config: `id`, `label`, `file`, `status`, `error?`, `servers` (names it already has), `added` (names added, or that would be), `skipped` (names its dialect cannot express, e.g. SSE for Codex and Antigravity).
+- `targets[]` — one per enabled CLI that declares an MCP config, plus GitHub Copilot CLI (`id: "copilot"`, a sync-only target that is not a run mode): `id`, `label`, `file`, `status`, `error?`, `servers` (names it already has), `added` (names added, or that would be), `skipped` (names its dialect cannot express, e.g. SSE for Codex and Antigravity).
   - `status`: `ok`; `absent` (not installed and no config file, so not read or created); `skipped` (the CLI's relocation env var, e.g. `CODEX_HOME`, is set to a relative path in the server's environment, so its file cannot be located safely and is neither read nor written); `unreadable` (the file exists but cannot be parsed safely, so it is not written); `failed` (a read or write error, the file may be unchanged).
   - `error` says why a target is not `ok`. A parse failure is reported by position only (`not valid TOML (line 3, column 21)`, `not valid JSON`), never with text from the file.
-  - `file` honours each CLI's own relocation env var as the server process sees it (`CLAUDE_CONFIG_DIR`, `CODEX_HOME`, `XDG_CONFIG_HOME`, `GEMINI_CLI_HOME`); see `docs/cli-registry.md`.
+  - `file` honours each CLI's own relocation env var as the server process sees it (`CLAUDE_CONFIG_DIR`, `CODEX_HOME`, `XDG_CONFIG_HOME`, `GEMINI_CLI_HOME`, and `COPILOT_HOME` for the sync-only Copilot CLI); see `docs/cli-registry.md`.
 - `conflicts[]` — names defined differently by different CLIs. Existing definitions are kept; the first CLI's is copied where the name is missing.
 - `disabled[]` — names left out because every definition is switched off in its own CLI (codex `enabled = false`, opencode `enabled: false`, antigravity `disabled: true`).
 - `unsupported[]` — labels of enabled agent CLIs with no known MCP config file (nothing is guessed).

docs/architecture-invariants.md

@@ -2,7 +2,7 @@
 
 Implementation detail extracted from `CLAUDE.md` so that file stays small enough to load into every session cheaply. Most sections are the original paragraphs, verbatim, including the version history and PR references that explain *why* each rule exists; newer ones are written here first and summarized back into `CLAUDE.md` as a short rule plus a pointer.
 
-`CLAUDE.md` keeps the short form of each rule plus a pointer to the section here. Read the pointer first; come here when you need the mechanism, the file names, or the history behind a constraint.
+`CLAUDE.md` keeps the short form of each rule plus a pointer to the section here; for the frontend (`src/web/public/`) that short form lives in `src/web/CLAUDE.md`, which the root file points to. Read the pointer first; come here when you need the mechanism, the file names, or the history behind a constraint.
 
 ---
 
@@ -77,7 +77,7 @@ OMP's registry entry declares `privilegedParams: []` and a launch spec that only
 
 Further detail: ⚠️ **Auto-reconnect revives ONLY when the durable remote tmux session is verifiably still alive** (`remoteTmuxSessionAlive()`, a `has-session` probe over ssh, #355): a clean agent exit (Ctrl-C, Ctrl-D, `exit`) tears that session down, and `isPaneDead()` cannot tell it from a transport drop, so the watcher used to relaunch a FRESH agent after every clean exit (claude only looked fine because its `|| --resume` fallback masked it). An unreachable host answers `undefined`, which also means do not revive. ⚠️ `has-session` prints NOTHING on success, so the probe is classified by EXIT STATUS (`classifyRemoteAliveExit`: 0 alive, ssh's 255 or a timeout unknown, anything else gone); reading stdout classified every live session as gone and silently disabled transport-drop reconnects. The answer is cached per session and forgotten whenever the pane is seen alive again, or a stale `true` from one transport drop would revive the next clean exit.
 
-File reads in a remote case are the second ssh surface (see Remote SSH cases below). The reason writes, office previews and thumbnails are deliberately unsupported over ssh is that no remote file is ever copied onto the server's disk. ssh children are bounded because terminal output in a remote session is written on the remote host, so a prompt-injected agent there can print hundreds of `codeman://attach` links.
+File reads in a remote case are the second ssh surface (see Remote SSH cases below). The reason writes, office previews and thumbnails are deliberately unsupported over ssh is that no remote file is ever copied onto the server's disk. ssh children are bounded because terminal output in a remote session is written on the remote host, so a prompt-injected agent there can print hundreds of `codeman://attach` links. The third surface is a remote session's full-history load, which reads the scrollback from the remote tmux (`src/remote-pane-history.ts`) through the same limiter, with a bounded slot wait (`maxWaitMs`) since the local capture is its fallback.
 
 ### Remote SSH cases
 
@@ -119,7 +119,7 @@ Spawn commands are pinned as literal strings in `test/cli-registry-spawn-golden.
 
 ### Input delivery and WS resilience
 
-**Input**: `session.writeViaMux()` for programmatic/curl input — tmux `send-keys -l` (literal) + `send-keys Enter`. Single-line only (fire-and-once). Interactive **browser** input goes through a durable **exactly-once** layer: each frame carries a stable `clientId` + monotonic per-session `seq`, persisted to localStorage until the server ACKs (`{t:'ia',seq}` over WS, or HTTP 2xx), so a dropped link/reconnect can't lose or double-deliver a prompt. A duplicate is ACKed as `{t:'ia',seq,dup:true,last}` so a client whose persisted counter rolled back below the server's watermark can lift itself out instead of typing into a silently dead terminal (`docs/reliable-input-delivery.md`). **WS resilience** (#149): the upgrade URL carries `cid = clientId + ':' + perTabNonce`, and `ws-connection-registry.ts` supersedes only same-TAB reconnects (two tabs on one session coexist; input frames keep the bare `clientId` for seq dedup); reconnects back off exponentially (attempts preserved across `_connectWs`), and the header connection chip renders from a real `_wsState` lifecycle (`connecting`/`connected`/`fallback`/`reconnecting`/`disconnected`).
+**Input**: `session.writeViaMux()` for programmatic/curl input — the text as one tmux bracketed paste (`paste-buffer -p`, so Codex cannot fold the Enter into a paste burst) + `send-keys Enter`. A single character is typed with `send-keys -l` instead (`submitsAsKeys()`): Claude Code's and Codex's selection dialogs ignore a pasted `3`, and the Enter then confirms the highlighted option. Single-line only (fire-and-once). Interactive **browser** input goes through a durable **exactly-once** layer: each frame carries a stable `clientId` + monotonic per-session `seq`, persisted to localStorage until the server ACKs (`{t:'ia',seq}` over WS, or HTTP 2xx), so a dropped link/reconnect can't lose or double-deliver a prompt. A duplicate is ACKed as `{t:'ia',seq,dup:true,last}` so a client whose persisted counter rolled back below the server's watermark can lift itself out instead of typing into a silently dead terminal (`docs/reliable-input-delivery.md`). **WS resilience** (#149): the upgrade URL carries `cid = clientId + ':' + perTabNonce`, and `ws-connection-registry.ts` supersedes only same-TAB reconnects (two tabs on one session coexist; input frames keep the bare `clientId` for seq dedup); reconnects back off exponentially (attempts preserved across `_connectWs`), and the header connection chip renders from a real `_wsState` lifecycle (`connecting`/`connected`/`fallback`/`reconnecting`/`disconnected`).
 
 ### Per-session env overrides: exact-key allowlist and CLAUDE_CONFIG_DIR
 
@@ -141,9 +141,9 @@ Spawn commands are pinned as literal strings in `test/cli-registry-spawn-golden.
 
 **Signal availability is decided by MODE, not by `isExternalCliMode()`.** `stop` and `blocked` come from Claude Code hooks, so `hooksAvailableForMode()` is true only for `claude`: `shell` is not an external-CLI mode but installs no hooks either, so `until=stop` on a shell session is a guaranteed unresolvable wait dressed up as a timeout. The behavior split is deliberate and must survive refactors: an EXPLICIT request for an unavailable signal is a 400 naming the mode, while the DEFAULT set (`stop,idle,exit`) silently drops them and echoes the narrowed set back as `wait.until`, because omitting the parameter must never 400. Signal quality is not uniform either: `stop` is definitive (Claude Code says the turn is over), `idle` is inferred from output stabilization plus prompt detection and can flap mid-turn when a spinner pauses, which is why `stop` is the documented default to orchestrate on and `idle` is the fallback for sessions that emit no hooks. ⚠️ **`idle` being ACCEPTED for a mode does not mean it ever FIRES there.** `startShell()` emits exactly one `idle` on a 500ms readiness timer and nothing afterwards, so a shell session sits at `status:'idle'` no matter what its pane is doing; since send-and-wait and `fresh=1` both require a TRANSITION, both can only time out on a shell worker (measured: a default `wait` on `sleep 4` burned its full 25s). The ❯-prompt and spinner detection that drives the real `working`/`idle` cycle is Claude's output format, so hook-less modes synchronize with `wait-output` markers or `exit`, and the docs must say so rather than listing `idle` as "available" and letting the reader infer it is usable. ⚠️ **Signals are edge-triggered with no history**, and that is a real orchestration limit: a signal that fires while no waiter is registered is gone, unobservable by any later wait variant (`until=stop` after the turn ended just times out, `fresh` or not — measured, R2-A). Documented client patterns must therefore register the waiter before the event can fire (send-and-wait) or gather on latched `wait-output` markers with `from=buffer`; the skill's fan-out flow was rewritten accordingly, and "fire-and-forget N prompts, then gather signal-waits sequentially" must never be documented again. The durable fix, a server-side latched last-signal-per-turn, is deferred with Part 3 of `docs/agent-control-plan.md`. ⚠️ Relatedly, the route corrects liveness that `SessionStatus` cannot express: `currentSignalFor()` answers `exit` whenever `pid === null` (exited, detached, or created-and-never-started) **or the mux pane is dead**, because `Session` parks a DEAD PTY at `_status = 'idle'` and trusting the status would answer the default wait `{signal:'idle', immediate:true}` for a crashed worker while `until=exit` blocked forever on an event that already happened. ⚠️ **`pid` alone cannot carry liveness for a tmux-backed session**: that pid is the local `tmux attach` CLIENT, so a worker exiting inside its pane leaves `pane_dead=1` with the client alive and `pid` never goes null — the `pid === null` branch is unreachable in the normal configuration (unit tests exercise it because `MockSession` sets `pid` by hand; only a live instance showed the gap). Liveness is therefore probed at the mux layer: `workerIsDead()` consults `mux.isPaneDead(muxName)` with a ~750 ms per-pane cache, ONLY on blocking waits (measured: 0 tmux execs across 100 non-wait input POSTs — the browser hot path pays nothing), plus a refcounted 3 s `watchForDeadWorker` interval so a worker dying while a wait is parked resolves it in ~3 s instead of burning the timeout. The probe fails SAFE by construction ("cannot tell" is never "dead": non-mux sessions, a missing or throwing `isPaneDead`, all return false). On send-and-wait, a "successful" `send-keys` into a dead pane additionally overrides `delivered` to `false` and rolls the dedup seq back (`undoOnFailure`), because the bytes went nowhere and a retry against a restarted worker must not be refused as a duplicate. The cost is that a just-created session reads as `exit`, which the wire docs must spell out as "not started yet"; the fix lives at the route rather than in `signalForStatus()` because the registry deliberately holds no `Session` reference. ⚠️ Two `claude`-mode cases still lose hooks for reasons outside the registry: a Docker case cannot reach a loopback-bound Codeman without `CODEMAN_DOCKER_BRIDGE_HOOKS=1`, and a remote-SSH case runs the agent on another host whose hooks may never reach this server. Bounds are env-overridable (`CODEMAN_WAIT_MAX_MS`, `CODEMAN_WAIT_DEFAULT_MS`, `CODEMAN_WAIT_MAX_PER_SESSION`, `CODEMAN_WAIT_MAX_PER_OWNER`, `CODEMAN_WAIT_MAX_TOTAL`, `CODEMAN_WAIT_BUFFER_SCAN_BYTES`) and each is clamped to a hard bound so a typo degrades to the default instead of disabling the protection; they are internal tuning knobs like the rest of `src/config/`, NOT part of the SemVer-covered env-var surface in `versioning-policy.md`. Tests: `tesLine truncated
 
-Further detail: `stop`/`blocked` now fire for **`claude` and `deepseek`** (`shell` installs no hooks either); asking for one explicitly on any other mode is a 400 and the default set silently drops them. `deepseek` qualifies because the DeepSeek Harness TUI REPORTS idle/working/blocked to its supervisor and Codeman is that supervisor (`deepseek-status-shim.ts`), so its signals are definitive rather than inferred, and `hooksAvailableForMode()` in `session-wait-registry.ts` is the one place that rule lives (see the External CLI modes section for the per-session `statusReporting` wrinkle and the two dsh traps baked into the skill preamble: boot `idle` landing ~300 ms before the composer paints, and `sendwait` asking for `wait:"stop,exit"`). That is also why `deepseek` is the one non-claude mode the skill drives like claude: `spawn_workers alpha beta:deepseek` is a mixed fleet in one call, and `sendwait`/`last_text` need no variant.
+Further detail: `stop`/`blocked` now fire for **`claude` and `deepseek`** (`shell` installs no hooks either); asking for one explicitly on any other mode is a 400 and the default set silently drops them. `deepseek` qualifies because the DeepSeek Harness TUI REPORTS idle/working/blocked to its supervisor and Codeman is that supervisor (`deepseek-status-shim.ts`), so its signals are definitive rather than inferred, and `hooksAvailableForMode()` in `session-wait-registry.ts` is the one place that rule lives (see the External CLI modes section for the per-session `statusReporting` wrinkle and the two dsh traps baked into the skill preamble: boot `idle` landing ~300 ms before the composer paints, and `sendwait` asking for `wait:"stop,exit"`). That is also why `deepseek` is the one non-claude mode the skill drives like claude: `spawn_workers alpha beta:deepseek` is a mixed fleet in one call, and `sendwait`/`last_text` need no variant. ⚠️ **`blocked` alone has a second source**: a CLI with no hook for it can declare the dialogs it blocks on (`capabilities.workDetect.blockedLine`, codex), and `blockedSignalAvailable()` (hooks OR that field) is the rule for that one signal; `stop` stays hooks-only, since nothing on a screen says a turn ended. A screen-read `blocked` is a LEVEL the server keeps reading (`Session.screenBlocked`), so the wait route answers it immediately (`currentSignalForWait()`) when the caller asked for `blocked`, and only then: such a pane is also `idle`, and `until=idle` must keep answering `idle`. See [Screen-read blocked dialogs](#screen-read-blocked-dialogs-codex).
 
-**Packaging as the `skills/codeman` agent skill.** The primitives are packaged as the `skills/codeman` agent skill, installable three ways: via `codeman skill install [--case <name>]` / `skill uninstall`; as a Claude Code plugin from the repo's own marketplace (`/plugin marketplace add Ark0N/Codeman` + `/plugin install codeman@codeman`, see the root-files note in CLAUDE.md); or auto-injected into a case's `.claude/skills/` on Claude session create behind `agentSkillEnabled` (SYNCED, default OFF). ⚠️ Plugin skills are namespaced, so a Claude Code holding BOTH the plugin and a user-level or per-case copy lists the skill twice, `codeman` and `codeman:codeman` (measured 2026-09-14: neither shadows the other, both work); the docs tell users to pick one route.
+**Packaging as the `skills/codeman` agent skill.** The primitives are packaged as the `skills/codeman` agent skill, installable three ways: via `codeman skill install [--case <name>]` / `skill uninstall`; as a Claude Code plugin from the repo's own marketplace (`/plugin marketplace add Ark0N/Codeman` + `/plugin install codeman@codeman`, see the root-files note in CLAUDE.md); or auto-injected into a case on session create behind `agentSkillEnabled` (SYNCED, default OFF). ⚠️ WHERE it lands is registry data, `capabilities.agentSkillDir` (absent with `agentSkillInjection: true` = `.claude/skills`): `.claude/skills/codeman` for claude, `.agents/skills/codeman` for every other CLI that takes it (pi does not: a project skill triggers its trust prompt). `agentSkillDir` is a CLOSED enum (`AGENT_SKILL_DIRS`, schema.ts) because `clis.json` can set it and it decides where the server writes. The per-session preamble is seeded for every local session whose CLI takes the skill, and `codeman skill uninstall --case` removes the copy from every distinct registry directory (`registryAgentSkillDirs()`, cli.ts). ⚠️ Plugin skills are namespaced, so a Claude Code holding BOTH the plugin and a user-level or per-case copy lists the skill twice, `codeman` and `codeman:codeman` (measured 2026-09-14: neither shadows the other, both work); the docs tell users to pick one route.
 
 Injection is ADD-ONLY at create, marker-owned (`applyAgentSkill` in `hooks-config.ts` never touches an unmarked user copy) and refuses symlinks (this repo's own `.claude/skills/codeman` is a symlink to the source, which the injector must never write through). ⚠️ Claude Code loads a same-named USER-LEVEL skill (`~/.claude/skills/codeman`, written once by `codeman skill install` with no `--case`) over the per-case copy, and nothing used to refresh it: a stale Aug-9 user copy shadowed every fresh injection (2026-08-14: agents ran the old recipes, spawned workers serially and lost their lineage arcs), so session create now also refreshes a marker-owned user copy (`refreshUserAgentSkill`; refresh-only, never installs, foreign/symlink refused).
 
@@ -157,7 +157,7 @@ Session create additionally pre-seeds the skill's §0 preamble cache (`seedAgent
 
 **Plan-usage chip** (`showPlanUsageLimits`, per-device: desktop default **ON** since 1.9.3, handhelds OFF) renders compact Claude and Codex provider rows. Claude Code (v2.1.80+) pipes a JSON blob to a configured `statusLine.command` on each render; on Pro/Max it carries a `rate_limits` object (`five_hour`/`seven_day` windows only — no Opus weekly field — each `{used_percentage 0-100, resets_at epoch-SECONDS}`). ⚠️ **Injected as an EPHEMERAL `claude --settings` CLI flag at spawn (2026-09-07), never written to disk** — `resolveStatusLineCliCommand()`/`ensureStatusLineExporterScript()` in `hooks-config.ts` (`generateStatusLineCommand()`/`applyStatusLineConfig()` remain, but only as the legacy disk-write self-heal path: a workspace an older Codeman build touched gets its stale `.claude/settings.local.json` entry stripped the first time a session starts there again). The exporter WRAPS a user's own real statusline (`findEffectiveUserStatusLineCommand()`, walking Claude Code's own settings precedence) rather than replacing it, and POSTs the `rate_limits` blob to `POST /api/status-telemetry`. That route (auth-exempt like `/api/hook-event` — localhost-only, hook-secret-gated whenever auth is active, COD-91) parses via `usage-telemetry.ts` (pure, unit-tested), broadcasts SSE `session:statusTelemetry` (de-duped per session by `telemetrySignature` since the statusline fires on every assistant message), and returns a compact plain-text footer for the exporter to **print-through** (foreground POST in the no-wrap branch so its own stdout becomes the footer, printing NOTHING on failure, `curl -sfk` plus `|| true`, since a bare brand word on the statusline is what discussion #405 opened with; backgrounded — `>/dev/null 2>&1 </dev/null &`, closing stdin too — only in the wrap branch, where the user's own command owns the footer; `curl --max-time 5` bounds a hung, not just refused, Codeman). Main Codex subscription usage comes from the signed-in host CLI's read-only app-server `account/rateLimits/read` request at startup and every 5 minutes; `usage-telemetry.ts` selects only the main `codex` bucket (never model-specific buckets such as Spark), maps whatever 5-hour/7-day windows it supplies, and omits the provider row when unavailable. Credentials stay inside the CLI and no auth material is sent to the browser. `plan-usage-latest.ts` merges both process-wide sources and replays them in the SSE init snapshot (`getLightState`) so `#planUsageChip` renders immediately on page load/reconnect. `planUsageChipEnabled()` remains the single resolver behind the checkbox and chip visibility (DISPLAY only) — the SAME `showPlanUsageLimits` setting also doubles as the server-side telemetry COLLECTION switch, read FRESH from `settings.json` by `readPlanUsageTelemetryEnabled()` at every claude session create/respawn (`TmuxManager.createSession`/`respawnPane`), never cached, with no per-session field and no per-request wire field — applies uniformly across every claude-creation path (interactive Run, cron, Ralph Loop API, quick-start) and survives a Codeman restart by construction (nothing per-session to lose). ⚠️ An ABSENT key reads as ON, the same way an absent `workspaceHooksEnabled` does: the desktop chip already shows as on for an install that never touched the setting, and the exporter posts only to this Codeman over loopback. Resolving the default in the reader is what keeps `GET /api/settings` a plain read. It briefly reconciled the key on first read (persisting `true` when absent), but `readJsonConfig()` answers `{}` for ANY read failure, not only ENOENT, and every page load hits that route, so one unlucky read replaced the whole settings file with a one-key file; pinned by `test/routes/system-routes-settings-get-plan-usage-default.test.ts`. ⚠️ The client sends `showPlanUsageLimits` in a settings save ONLY when that save FLIPS the chip relative to what the device had (`planUsageCollectionFlip()` in settings-ui.js): the chip defaults OFF on handhelds, so sending it on every save let a phone saving its font size persist `false` and switch collection off for every desktop, whose chip then went stale with no error anywhere. An explicit toggle on any device still writes the switch. Injection covers LOCAL tmux-spawned claude sessions only: the non-tmux direct-PTY fallback (`Session.startInteractive` when tmux is unavailable) and the remote/docker pane builders do not carry the flag. Registry-gated on `getCli(mode)?.capabilities.statusLineTelemetry` rather than a hardcoded mode string. **Distinct from auto-resume** (which reacts to the Claude limit _message_; this proactively shows live percentages). Design: `docs/usage-limits-display-plan.md`. Tests: `test/usage-telemetry.test.ts`, `test/codex-plan-usage.test.ts`, `test/plan-usage-chip.test.ts`, `test/plan-usage-latest.test.ts`, `test/hooks-config.test.ts` (statusline exporter script + `readPlanUsageTelemetryEnabled`), `test/statusline-cli-flag.test.ts`.
 
-Further detail: the handheld OFF default comes from the mobile block in `getDefaultSettings()`. `planUsageChipEnabled()` backs the two call sites that must never disagree (the App Settings checkbox and the chip's visibility). The absent-key-reads-as-ON rule mirrors `readWorkspaceHooksEnabled()`.
+Further detail: the handheld OFF default comes from the mobile block in `getDefaultSettings()`. `planUsageChipEnabled()` backs the two call sites that must never disagree (the App Settings checkbox and the chip's visibility). The absent-key-reads-as-ON rule mirrors `readWorkspaceHooksEnabled()`. The same statusline payload carries a `prompt_cache` object since Claude Code v2.1.251 (`warm`, `ttl`, `expires_at` in epoch seconds, `recache_tokens_if_cold`, `misses`, `expected_rebuilds`, `hit_ratio`, `last_miss_at`, `last_miss_cause`); `StatusTelemetrySchema` lists every documented scalar field null-tolerant and keeps `causes` as `unknown` (the object gains fields across CLI versions and several are documented as possibly null, so a stricter shape would 400 the whole post and blank the footer — the recurring `.optional()`/null trap), while `parsePromptCache()` (usage-telemetry.ts) normalizes only the subset Codeman stores, so a hostile localhost payload cannot bloat state. The footer shows it as `cache:until HH:MM` (server-local clock, never a countdown: the footer only re-renders on statusline triggers) while warm and `cache:cold` otherwise. Claude re-runs the statusline when `expires_at` passes, so the footer flips to cold on its own. ⚠️ A user's OWN statusline wins the footer (the exporter then posts in the background and prints theirs), so the state ALSO lives on `Session.promptCache` (`setPromptCache()`, in `toState()` for the Session Options → Respawn readout `#promptCacheStatus`, worded by the pure `CodemanPromptCache.format` in constants.js, which also shows the hit rate, the last miss's `causes`, and — after a compaction (`/compact` or Claude Code's own idle compaction) or a tool-result clearing, the first warm report whose `recache_tokens_if_cold` is null after one that carried it — a `Context compacted at HH:MM` note that is NEVER an alert) and the route broadcasts `session:promptCache` (the STORED object, so the compaction stamp rides along) only when the report CHANGES (the statusline fires per assistant message). ⚠️ `setPromptCache()` returns a NEW object when it stamps `compactedAt`, never mutating the parser's output, which is what keeps broadcasting the stored object meaningful; the stamp needs a KNOWN prior `recache_tokens_if_cold` (a first report already inside the null window, such as a mid-session restart's, gets none), is carried while the window holds and dropped at the next request, and never reads `expected_rebuilds`: Claude Code's tracker clears the window and counts the expected rebuild in that same next request, so no single report carries both (2.1.294-2.1.296), and `adoptClaudeSessionId()` drops the whole report on a conversation switch so a post-`/clear` pane never shows the old prefix's warm-until. The last report is restored on a Codeman restart (the constructor's `promptCache`, set directly so it is not re-stamped, like `displayModel`) so the readout is not blank until the pane's next render; the readout is time-aware, so a value whose `expiresAt` has passed reads as cold rather than as a stale warm (and `parsePromptCache` validates the converted-to-ms expiry against the Date range, so a huge `expires_at` cannot render `NaN:NaN`). ⚠️ A launch without the exporter never reports, so `TmuxManager._resolveStatusLine()` records why on the mux record at every create and respawn (`MuxSession.statusLine`, published as `SessionState.statusLineExporter`: `injected`, `collection-off`, `workspace-statusline`, `remote`) and the readout names the cause instead of "not reported yet" forever. It is a spawn-time fact: flipping Plan Usage on later changes nothing until the next launch. ⚠️ This object, not the plan tier, is the source of truth for the TTL: Claude Code requests the hour only for a subscription within plan usage, and drops to five minutes on an API key, a cloud provider or extra-usage credits.
 
 ### Cron jobs
 
@@ -209,13 +209,13 @@ Further detail: the `<prefix>: <title>` form (`w3-myapp: fix the login redirect`
 
 ### Full-scrollback replay
 
-**Full-scrollback replay** (COD-164/#148, reworked for #205): `GET /api/sessions/:id/terminal?full=1` returns the ENTIRE tmux scrollback (capture-pane `-e -S -<lines>` bounded by the configured history limit, explicit `maxBuffer` from the terminal-history config, early byte-cap before normalization, CRLF-normalized for shell panes). On success the capture is returned ALONE (`source='mux-full-history'` — it supersedes the byte buffer; no duplication). The first load of each non-shell TUI session per page requests `full=1` (`_fullHistoryLoaded` Set in app.js — the old one-shot `_initialFullBufferLoad` flag was consumed by whichever tab auto-selected, leaving every other TUI tab one frame of history). Shell sessions instead load a bounded 1 MiB `?tail=` window on every selection and automatic drop recovery: a 100k-line shell capture can be tens of MiB, and automatically parsing it makes tab-switch latency scale with the entire session. Shell FULL history is explicit-button-only; reaching the top during an ordinary wheel/touch gesture must not reset xterm and replay the multi-megabyte capture on its main thread. That gesture instead pulls a BOUNDED window of it (`?full=1&tail=<TERMINAL_TAIL_SIZE>`, so a capture beyond 1 MiB comes back `truncationReason: 'tail'` and the banner offers the rest), and skips the replay when the window holds no more rows than the browser already has. ⚠️ **That skip runs BEFORE the downgrade guard, and it never touches the banner state**: the guard reads "smaller than the browser" as "tmux has nothing more to give", true of an unbounded capture and false of a window cut at the tail size, so a bounded window that reached it marked the session `exhausted` and removed **Load full history** while tmux still held the rest; and re-labelling the banner from a skipped payload would call a terminal that holds all of a Load full history pull "the most recent 1 MiB". A skipped window that came back truncated puts the session on the 60 s cooldown (`_fullHistoryRepullUseless`), because it can never reach anything older than the browser shows and `tail` is applied AFTER the server's synchronous `capture-pane` of the whole history, so each ask costs every client an event-loop stall (about 0.3 s at 100k lines); an untruncated one keeps the 4 s cooldown, since it is all of tmux's history and the next burst can add to it. Declining the gesture outright was a dead end: after any burst a shell pane held about one screen of browser scrollback, and the button renders only once a replay was truncated, so a young shell tab had no way back to history tmux still held. Other modes may still re-pull `full=1` at the TOP, and pressing **Load full history** forces the request for any recoverably truncated session (`_maybeRefetchFullHistory`, 4s per-session gesture cooldown, in-flight + tab-switch guards, viewport position held across the replay); Shell full pulls are not retained in the tab cache, so the next switch stays bounded. Chunked replay enqueues 32 KiB pieces across safe yields, appends an xterm parse marker, then releases the live-output gate; output arriving after that release stays ordered behind the snapshot, and the marker callback supplies accurate parse timing. ⚠️ **How the load ENDS depends on where the payload came from**, and `_bufferLoadFinishOpts` (app.js) is the one place that decides it for all four fetch-and-write paths. A payload built from the server's accumulated byte history is current up to the response, so the events queued during the load already appear in it and stay DISCARDED; replaying them would duplicate output, most visibly Ink's cursor-up redraws. A pane capture (`mux-visible` or `mux-full-history`) is current only up to CAPTURE time, so `_finishBufferLoad` replays the queue from the response's own arrival timestamp (`since`) and the pre-capture events stay dropped. ⚠️ **A path that then restores a scroll position must re-take the sticky-scroll baseline** (`_syncStickyScrollBaseline`): the replay runs inside `chunkedTerminalWrite` before its promise resolves, with the terminal freshly reset, so `batchTerminalWrite` samples `_wasAtBottomBeforeWrite` as true and the next `flushPendingWrites` would scroll to the bottom over the restore. ⚠️ The cutoff is a client-side timestamp and the server broadcasts on a batch timer (8ms WebSocket, 16-50ms SSE), so a batch pending when the capture ran arrives after the response and replays although the capture holds it — bounded by one batch interval, and closable only server side by flushing that batch before the capture. Tests for the three: `test/terminal-flush-budget.test.ts` pins which sources flush, `test/terminal-buffer-flush.test.ts` pins the `since` cutoff and the baseline re-take, and `test/capture-load-window.browser.test.ts` drives both against a live server. Live output is separately one-chunk-in-flight: xterm's callback releases each 32/64 KiB write before the next is submitted, keeping the remainder in the Line truncated
+**Full-scrollback replay** (COD-164/#148, reworked for #205): `GET /api/sessions/:id/terminal?full=1` returns the ENTIRE tmux scrollback (capture-pane `-e -S -<lines>` bounded by the configured history limit, explicit `maxBuffer` from the terminal-history config, early byte-cap before normalization, CRLF-normalized for shell panes). On success the capture is returned ALONE (`source='mux-full-history'` — it supersedes the byte buffer; no duplication). The first load of each non-shell TUI session per page requests `full=1` (`_fullHistoryLoaded` Set in app.js — the old one-shot `_initialFullBufferLoad` flag was consumed by whichever tab auto-selected, leaving every other TUI tab one frame of history). Shell sessions instead load a bounded window of tmux's RENDERED history on every selection, automatic drop recovery, clear and tile load (`SHELL_LOAD_QUERY` in constants.js: `?full=1&tail=1MiB&lines=10000&alt=recording`): a 100k-line shell capture can be tens of MiB, and automatically parsing it makes tab-switch latency scale with the entire session (`lines=` bounds the synchronous capture-pane, about 8 ms per 1000 lines). They never replay the raw byte recording ahead of the visible frame any more: it holds full-screen repaints (an ssh'd tmux, `less`, any TUI) drawn for the pane height of the moment, and replayed into a terminal of another height each repaint's last newline scrolls its top row into scrollback, so one line tmux holds once showed up dozens of times. Because `lines=` is shallower than xterm's scrollback, the route reports a cut: the capture's `display-message` query also reads `#{history_size}` (`PaneCaptureOptions.capturedHistoryLines`, the same field the route reports as `paneHistoryLines`), and a full capture whose CLIENT bound (`lines=` below the configured `tmuxHistoryLimit`) left lines in tmux comes back `truncated`, `truncationReason: 'tail'`. ⚠️ Only a client bound below the ceiling counts: a request without `lines=` is bounded by the ceiling itself, and on tmux before 3.7 a live pane can hold more than a lowered ceiling, which no request can read. ⚠️ **A pane in the alternate screen keeps no tmux history**, and a nested tmux client (remote SSH and Docker cases, or `tmux`/`ssh -t host tmux` run by hand) holds its pane there, so its full capture is one screen. The same query reads `#{alternate_on}` (`capturedAlternateOn`), and a request carrying `alt=recording` (only `SHELL_LOAD_QUERY`) is then answered the old way, the visible frame over the tailed byte recording, which still carries the scrollback; agents, scroll pulls and **Load full history** never send it. `#{alternate_on}` cannot tell a nested tmux from vim, `less` or `htop`, so a local shell with one of those open takes the same fallback until it exits (the pre-fix load, with its repaint duplicates). Shell FULL history is explicit-button-only; reaching the top during an ordinary wheel/touch gesture must not reset xterm and replay the multi-megabyte capture on its main thread. That gesture instead pulls a BOUNDED window of it (`?full=1&tail=<TERMINAL_TAIL_SIZE>`, so a capture beyond 1 MiB comes back `truncationReason: 'tail'` and the banner offers the rest), and skips the replay when the window holds no more rows than the browser already has. ⚠️ **That skip runs BEFORE the downgrade guard, and it never touches the banner state**: the guard reads "smaller than the browser" as "tmux has nothing more to give", true of an unbounded capture and false of a window cut at the tail size, so a bounded window that reached it marked the session `exhausted` and removed **Load full history** while tmux still held the rest; and re-labelling the banner from a skipped payload would call a terminal that holds all of a Load full history pull "the most recent 1 MiB". A skipped window that came back truncated puts the session on the 60 s cooldown (`_fullHistoryRepullUseless`), because it can never reach anything older than the browser shows and `tail` is applied AFTER the server's synchronous `capture-pane` of the whole history, so each ask costs every client an event-loop stall (about 0.3 s at 100k lines); an untruncated one keeps the 4 s cooldown, since it is all of tmux's history and the next burst can add to it. Declining the gesture outright was a dead end: after any burst a shell pane held about one screen of browser scrollback, and the button renders only once a replay was truncated, so a young shell tab had no way back to history tmux still held. Other modes may still re-pull `full=1` at the TOP, and pressing **Load full history** forces the request for any recoverably truncated session (`_maybeRefetchFullHistory`, 4s per-session gesture cooldown, in-flight + tab-switch guards, viewport position held across the replay); Shell full pulls are not retained in the tab cache, so the next switch stays bounded. Chunked replay enqueues 32 KiB pieces across safe yields, appends an xterm parse marker, then releases the live-output gate; outpLine truncated
 
-**The partial-history notice is lazy and pane-gated** (`computeHistoryTruncationNotice()` in constants.js, `_renderHistoryTruncationBanner()` / `_setHistoryNoticeRevealed()` in app.js). `truncated` on a terminal response measures the server's BYTE stream, which is not what `full=1` can return: for a pane with no tmux scrollback (fullscreen claude starts in the alternate screen and never writes to the main grid, so `#{history_size}` is 0) the bytes a tail cut drops are old repaint frames. Measured on prod 2026-10-10, every claude pane: the tab-switch tail cut a 1.2 to 6.2 MB stream to 1 MiB, the local scrollback that tail replays into is repeats of the same frames (one pane: 450 non-blank rows, 44 unique, one row 33 times), and `full=1` returned 1 to 5 KB, so **Load full history** could only end in the downgrade refusal while the banner promised "4.8 MB more". The capture therefore reads `#{history_size}` in the SAME `display-message` as the cursor (no extra tmux call), the route returns it as `paneHistoryLines` (absent when the body is not a capture), and a reported 0 hides the notice in every state; a positive count replaces the byte gap in the message. ⚠️ Key the decision on `history_size`, never on `alternate_on`: a shell running `less` is in the alternate screen with its whole history still retained (measured, 99,972 rows) and `full=1` returns all of it. The notice is also LAZY: it appears only after a wheel/touch gesture reaches the top of the browser buffer and the pull that gesture started has settled (`_maybeLoadMoreHistoryOnScroll`), goes away when a downward gesture is back at live output or the tab switches (`selectSession` nulls `_historyNoticeRevealedFor` before repainting), and its × is sticky per session until reload (`_historyNoticeDismissed`). Tests: `test/history-truncation-notice.test.ts`, `test/history-notice-reveal.test.ts`.
+**The partial-history notice is lazy and pane-gated** (`computeHistoryTruncationNotice()` in constants.js, `_renderHistoryTruncationBanner()` / `_setHistoryNoticeRevealed()` in app.js). `truncated` on a terminal response measures the server's BYTE stream, which is not what `full=1` can return: for a pane with no tmux scrollback (fullscreen claude starts in the alternate screen and never writes to the main grid, so `#{history_size}` is 0) the bytes a tail cut drops are old repaint frames. Measured on prod 2026-10-10, every claude pane: the tab-switch tail cut a 1.2 to 6.2 MB stream to 1 MiB, the local scrollback that tail replays into is repeats of the same frames (one pane: 450 non-blank rows, 44 unique, one row 33 times), and `full=1` returned 1 to 5 KB, so **Load full history** could only end in the downgrade refusal while the banner promised "4.8 MB more". The capture therefore reads `#{history_size}` in the SAME `display-message` as the cursor (no extra tmux call), the route returns it as `paneHistoryLines` (absent when the body is not a capture, and for a remote SSH session, whose local pane hosts the remote tmux client and always reports 0 while `full=1` reads the remote history), and a reported 0 hides the notice in every state; a positive count replaces the byte gap in the message. ⚠️ Key the notice decision on `history_size`, never on `alternate_on`: a shell running `less` is in the alternate screen with its whole history still retained (measured, 99,972 rows) and `full=1` returns all of it. The one place `alternate_on` decides is a shell's AUTOMATIC load (`alt=recording`, see Full-scrollback replay): there the question is whether the capture holds the scrollback a nested tmux client keeps in its byte recording, not whether a pull can return more. The notice is also LAZY: it appears only after a wheel/touch gesture reaches the top of the browser buffer and the pull that gesture started has settled (`_maybeLoadMoreHistoryOnScroll`), goes away when a downward gesture is back at live output or the tab switches (`selectSession` nulls `_historyNoticeRevealedFor` before repainting), and its × is sticky per session until reload (`_historyNoticeDismissed`). Tests: `test/history-truncation-notice.test.ts`, `test/history-notice-reveal.test.ts`.
 
-**A tab switch to a pane with no tmux scrollback takes `full=1`, not the 1 MiB tail** (`_paneHistoryLines` + `_notePaneHistory()` in app.js). Every terminal response (select, refresh, re-pull) records `paneHistoryLines` per session, and `selectSession` serves a non-shell switch from `full=1` whenever the last report was 0. For such a pane that capture IS the visible frame (0.9 to 5.6 KB on prod) and matches what a page load shows, while the tail was 1 MiB of old repaints. Measured 2026-10-10 with real prod payloads through the real client in chromium: median switch 1.26 to 1.66 s down to 0.10 to 0.19 s, parse ~1.1 to 1.5 s down to 2 to 3 ms; server time 136 to 1070 ms (tail) vs 89 to 330 ms (full). An empty local buffer is also what lets `_maybePageCliTranscript` send a gesture to the CLI's own transcript. ⚠️ Only a reported 0 counts: a response without the field FORGETS the entry (unknown is never empty), so an older server or a byte-history fallback keeps the tail; a pane that starts keeping history (claude switched to its inline view) is learned from the next response and goes back to the tail. ⚠️ Shells never take this path. ⚠️ The refresh path (`_onSessionNeedsRefresh`) keeps its downgrade guard unchanged: a server-triggered refresh must not wipe local rows the user may be reading. Tests: `test/pane-history-memory.test.ts`, `test/fullscreen-tab-switch-capture.browser.test.ts` (browser suite).
+**A tab switch to a pane with no tmux scrollback takes `full=1`, not the 1 MiB tail** (`_paneHistoryLines` + `_notePaneHistory()` in app.js). Every terminal response (select, refresh, re-pull) records `paneHistoryLines` per session, and `selectSession` serves a non-shell switch from `full=1` whenever the last report was 0. For such a pane that capture IS the visible frame (0.9 to 5.6 KB on prod) and matches what a page load shows, while the tail was 1 MiB of old repaints. Measured 2026-10-10 with real prod payloads through the real client in chromium: median switch 1.26 to 1.66 s down to 0.10 to 0.19 s, parse ~1.1 to 1.5 s down to 2 to 3 ms; server time 136 to 1070 ms (tail) vs 89 to 330 ms (full). An empty local buffer is also what lets `_maybePageCliTranscript` send a gesture to the CLI's own transcript. ⚠️ Only a reported 0 counts: a response without the field FORGETS the entry (unknown is never empty), so an older server or a byte-history fallback keeps the tail; a pane that starts keeping history (claude switched to its inline view) is learned from the next response and goes back to the tail. ⚠️ Shells never take this path. ⚠️ Remote and Docker sessions never take this path: their local pane hosts a nested tmux client in the alternate screen, so it reports 0 while the byte tail carries the real scrolled-off output. ⚠️ The refresh path (`_onSessionNeedsRefresh`) keeps its downgrade guard unchanged: a server-triggered refresh must not wipe local rows the user may be reading. Tests: `test/pane-history-memory.test.ts`, `test/fullscreen-tab-switch-capture.browser.test.ts` (browser suite).
 
-**A capture reports the geometry it was taken at** (#435): a visible frame repaints each row at an absolute position, counting up to the pane's height and out to the pane's width, so a terminal smaller than that pane damages it two ways at once. Too short and every address past the browser's own height clamps onto the last line, overwriting the rows underneath (measured: against a 50-row pane, a 30-row terminal rendered 28 of a 45-line command and drew the survivors twice). Too narrow and each row is painted out to the pane's width, so the browser wraps every painted row and the wrap on the last one scrolls the whole frame up by one. Nothing in the response used to say what geometry the frame was built for, so the client could not see either case. `PaneCaptureOptions.capturedGeometry` carries it out, and the terminal response publishes it as `captureCols`/`captureRows`. ⚠️ **Both fields are ABSENT unless a frame was really positioned**, and every consumer must test `Number.isFinite` rather than truthiness: `mux-visible` is necessary but not sufficient, because when the `display-message` cursor query fails `capturePaneBuffer` skips the snapshot repaint and returns the raw capture, and the route still labels that non-empty body `mux-visible`. A body that positioned nothing has no geometry to describe and nothing to repair, so a comparison that fires there buys a second capture, a reset plus chunked rewrite, a dropped and reopened WebSocket and a discarded xterm snapshot for no gain. ⚠️ **The comparison runs on `mux-visible` ONLY.** A full-history body is linear scrollback closed by a RELATIVE cursor move, which is relative precisely so the browser's row count need not match the pane's, and a byte-history body carries no row alignment at all, so a size mismatch damages neither and a replay repairs neither. That gate matters because the first select of every non-shell session per page takes the full-history path, where an ungated comparison would fire most often on the one response it cannot help, at the price of a second whole-scrollback capture. ⚠️ **The replay is capped at one attempt and latches per session when it cannot converge.** `resizeRetry` stops two competing fits trading replays forever; a pane already drawing at the size just requested is left alone, because a retry would capture the identical frame; and a pass that still does not converge joins `_geometryRetryUseless`, so the case `Session.resize` declines outright (a small viewport while a desktop viewport's size claim is live, where the retry re-sends the same declined resize and captures the same pane) costs one attempt per session per page load instead of one per select. ⚠️ **The clamp used to manufacture that equality, and no longer can** (#464). This paragraph previously explained it as the signature of a clamp: `getTerminalDimensions()` floors at 40x10 while `fitAddon.fit()` did not, so a terminal under 40 columns or 10 rows reported a pane permanently bigger than itself and would replay on every tab switch. That divergence is fixed at the source — `syncTerminalGeometry()` (terminal-ui.js) fits, floors and APPLIES in one step, so the browser terminal IS the size it reports. The only remaining reason the two can differ is a resize the server declined, which the server now reports back (`Session.ptyGeometry`, the `{"t":"zc"}` frame and the resize response) for the client to adopt by COLUMNS. The equality guard stays, for the plain case of a pane already at the requested size. ⚠️ **A retry pass must not re-arm `_fullHistoryLoaded`**: it did not consume the full-history pull, and re-arming it would spend a whole-scrollback capture on the next select. That branch is currently unreachable by construction, since reaching it needs `source === 'mux-visible'` while a `full=1` pass is answered `mux-full-history` or `history`; a static test over the source is the habit this repo uses for an invariant nothing can execute. Tests: `test/capture-geometry-retry.browser.test.ts` (eight cases, five of which fail against the merge base), `test/tmux-capture-full-history.test.ts`, `test/routes/session-routes.test.ts`.
+**A capture reports the geometry it was taken at** (#435): a visible frame repaints each row at an absolute position, counting up to the pane's height and out to the pane's width, so a terminal smaller than that pane damages it two ways at once. Too short and every address past the browser's own height clamps onto the last line, overwriting the rows underneath (measured: against a 50-row pane, a 30-row terminal rendered 28 of a 45-line command and drew the survivors twice). Too narrow and each row is painted out to the pane's width, so the browser wraps every painted row and the wrap on the last one scrolls the whole frame up by one. Nothing in the response used to say what geometry the frame was built for, so the client could not see either case. `PaneCaptureOptions.capturedGeometry` carries it out, and the terminal response publishes it as `captureCols`/`captureRows`. ⚠️ **Both fields are ABSENT unless a frame was really positioned**, and every consumer must test `Number.isFinite` rather than truthiness: `mux-visible` is necessary but not sufficient, because when the `display-message` cursor query fails `capturePaneBuffer` skips the snapshot repaint and returns the raw capture, and the route still labels that non-empty body `mux-visible`. A body that positioned nothing has no geometry to describe and nothing to repair, so a comparison that fires there buys a second capture, a reset plus chunked rewrite, a dropped and reopened WebSocket and a discarded xterm snapshot for no gain. ⚠️ **The comparison runs on `mux-visible` ONLY.** A full-history body is linear scrollback closed by a RELATIVE cursor move, which is relative precisely so the browser's row count need not match the pane's, and a byte-history body carries no row alignment at all, so a size mismatch damages neither and a replay repairs neither. That gate matters because the first select of every non-shell session per page takes the full-history path, where an ungated comparison would fire most often on the one response it cannot help, at the price of a second whole-scrollback capture. ⚠️ **The replay is capped at one attempt and latches per session when it cannot converge.** `resizeRetry` stops two competing fits trading replays forever; a pane already drawing at the size just requested is left alone, because a retry would capture the identical frame; and a pass that still does not converge joins `_geometryRetryUseless`, so the case `Session.resize` declines outright (a small viewport while a desktop viewport's size claim is live, where the retry re-sends the same declined resize and captures the same pane) costs one attempt per session per page load instead of one per select. ⚠️ **The clamp used to manufacture that equality, and no longer can** (#464). This paragraph previously explained it as the signature of a clamp: `getTerminalDimensions()` floors at 40x10 while `fitAddon.fit()` did not, so a terminal under 40 columns or 10 rows reported a pane permanently bigger than itself and would replay on every tab switch. That divergence is fixed at the source — `syncTerminalGeometry()` (terminal-ui.js) fits, floors and APPLIES in one step, so the browser terminal IS the size it reports. The only remaining reason the two can differ is a resize the server declined, which the server now reports back (`Session.ptyGeometry`, the `{"t":"zc"}` frame and the resize response) for the client to adopt by COLUMNS. The equality guard stays, for the plain case of a pane already at the requested size. ⚠️ **A retry pass must not re-arm `_fullHistoryLoaded`**: it did not consume the full-history pull, and re-arming it would spend a whole-scrollback capture on the next select. That branch is currently unreachable by construction, since reaching it needs `source === 'mux-visible'` while a `full=1` pass without `alt=recording` (every `useFullHistory` pass; only shell loads send it) is answered `mux-full-history` or `history`; a static test over the source is the habit this repo uses for an invariant nothing can execute. Tests: `test/capture-geometry-retry.browser.test.ts` (eight cases, five of which fail against the merge base), `test/tmux-capture-full-history.test.ts`, `test/routes/session-routes.test.ts`.
 
 ⚠️ **A frame dropped at the 128 KiB render cap MUST be recovered, and the recovery must verify itself** (`_scheduleDroppedOutputRecovery`, app.js): a hole in a TUI byte stream is a desynced cursor, which is muffled text (#464). It was a fire-and-forget 2s timer that nulled its own handle and then called `_onSessionNeedsRefresh()` — whose early returns (a buffer load in flight, a refresh already owning the session) are MOST likely to be true during exactly the burst that caused the drop, so the recovery was lost silently and the bytes were never replayed. `_onSessionNeedsRefresh` now returns whether it actually repainted, and the scheduler re-arms while it has not, bounded by `DROP_RECOVERY_MAX_ATTEMPTS` because the early returns it retries past are transient contention. A refresh that died at the capture fetch DEADLINE (it returns `'deadline'`) is a stalled link, not contention, and is NOT retried: each retry would be another `?full=1` capture waiting out a deadline of up to two minutes. The same debounce still collapses a burst into one attempt, and the `TERMINAL DROP` crash-trail line sits behind it, since one line per dropped frame evicted the whole 50-entry trail in under a second. Tests: `test/dropped-output-recovery.test.ts`, whose retry case and no-retry case only pin the fix as a pair.
 
@@ -229,7 +229,7 @@ Further detail: the `<prefix>: <title>` form (`w3-myapp: fix the login redirect`
 
 **Wheel/touch forwarding is NOT gated on viewport-at-bottom** (#205, `terminal-ui.js:_shouldForwardWheelToApp`): for sessions verified to scroll their own transcript on SGR wheel reports (claude ≥ 2.1.187 while `cliMouseTracking` is true, i.e. fullscreen; version via the local/docker/remote `--version` probes), the plain wheel AND touch drags forward as coalesced SGR reports (`_forwardScrollToApp` → `_sendSyntheticSgrWheel`, 40ms batches, 5-tick cap, 512-byte queue bound). It used to gate on the viewport being at the bottom so both scrollbacks stayed reachable, but a repaint-mode CLI keeps NO terminal scrollback of its own — xterm's buffer holds only replayed repaint frames, so local scrolling drags the CLI's pinned prompt box up the screen over stale frames; and `scrollToLastNonEmptyLine()` routinely parked the viewport off-bottom, silently pinning the wheel to local. Forwarding now snaps the viewport home first (SGR coordinates address the LIVE screen — a report computed from a scrolled-up viewport would hit-test the wrong row). Local scrollback remains on Shift+wheel and the `terminalWheelLocalScrollback` opt-out (both also cover touch via the shared gate; touch has no Shift, so the setting is its only local pin). `_wheelScrollLines()` normalizes `deltaMode` (Firefox fires LINE deltas ≈3/notch — read as pixels that rounded to 0 and fell to the ±1 fallback, ~4× too slow; PAGE deltas scale by `terminal.rows`) while keeping the #154 Shift-axis trap (macOS trackpads put Shift+scroll magnitude on deltaX). Tests: `test/terminal-touch-tap.test.ts`.
 
-**A false gate on a hollow pane must not mean a DEAD gesture** (#205 round 2, `_maybePageCliTranscript`): every way `_shouldForwardWheelToApp()` returns false leaves a repaint-mode pane scrolling a buffer that has nothing in it (`baseY === 0`) — the version probe came back empty, the CLI really is older than 2.1.187, the `cliMouseTracking` flag is unset (inline claude, or fullscreen right after a server restart), or the user turned on `terminalWheelLocalScrollback`. **opencode is the fifth case, and the gate is false there by design**: its TUI runs on the ALTERNATE SCREEN (1.18.31 measured: tmux `alternate_on=1`, `history_size=0`), so the buffer is hollow, and it IGNORES SGR wheel reports entirely (six `\x1b[<64;…M` reports against an idle pane left the capture byte-identical) while still paging its transcript on PageUp/PageDown (`messages_page_up/down`) — so paging is the only gesture that can reach it, and without it the wheel was silently dead in every opencode tab. The 1.12.0 retest reported exactly that shape for Claude: a wheel that did nothing at all while Fn+Up (PageUp) paged back through intact text, which is the proof that the CLI's own history and the PTY input path were both fine. So under the guard (`_localScrollbackIsHollow()` — `claude` or `opencode`, gate false, `baseY === 0`) wheel and touch travel is translated into coalesced `\x1b[5~` / `\x1b[6~` through the same 40ms queue as the SGR reports, at half a screen of travel per page key (the key jumps a whole screen; a 1:1 mapping was unusably slow with a discrete wheel). ⚠️ `shell`/`pi` own real terminal scrollback and are never paged, and codex/gemini/antigravity/grok/deepseek/omp page-key behaviour is unverified (`docs/scrollback-fix-plan.md`). ⚠️ Shift is excluded on purpose — it is the explicit "give me local scrollback" gesture and must keep that meaning. ⚠️ `terminalWheelLocalScrollback` is deliberately NOT scoped away from repaint-mode CLIs even though it is a footgun there: that would silently override an explicit user choice, so the fallback catches it instead. **Server-side counterpart**: `getClaudeCliVersion()` caches SUCCESS for the process lifetime but must never cache FAILURE — it used to, so one timed-out or PATH-starved probe at the first Claude session start disabled wheel-forwarding for every Claude session until the server restarted (a dead wheel on phone, tablet and laptop at once, the signature of a server-side cause). Failures now retry with a 1/2/4…15min backoff; the policy is the pure `resolveClaudeCliVersion()`. **Tiles** (a grid tile, the split's Pane B: `TerminalTile`) page the wheel through these same gates, called with the tile's own terminal and session (`_localScrollbackIsHollow(target)`, `_shouldForwardWheelToApp(ev, target)`) and the pure `CodemanTerminalInput.pageKeysForTravel`, so the mode list stays in terminal-ui.js alone. ⚠️ A tile's `baseY` is rarely 0 even when hollow: its first capture is taken at the PTY's previous, taller size and its row-shrinking fits push rows up, so it passes `localRows` with those rows discounted (`TerminalTile._localRows`). ⚠️ So a tile can be hollow with its viewport still up in those rows (Shift+PageUp, a scrollbar drag, a wheel during the first replay), and it pages only while `_terminalViewportAtBottom(tile.terminal)` holds: otherwise every wheel, wheel-down included, was paged and swallowed, the stale rows stayed on screen and the click report (which refuses an off-bottom viewport) went dead too. Off the bottom the wheel stays xterm's, so a wheel-down brings the viewport home. The primary pane needs no such gate: hollow there means `baseY === 0`, which is always at the bottom. Tiles have no touch path and do not forward SGR wheel (tile-grid-plan follow-up 4). Tests: `test/terminal-scroll-routing.test.ts`, `test/terminal-tile-scroll.test.ts`, `test/claude-cli-version-cache.test.ts`.
+**A false gate on a hollow pane must not mean a DEAD gesture** (#205 round 2, `_maybePageCliTranscript`): every way `_shouldForwardWheelToApp()` returns false leaves a repaint-mode pane scrolling a buffer that has nothing in it (`baseY === 0`) — the version probe came back empty, the CLI really is older than 2.1.187, the `cliMouseTracking` flag is unset (inline claude, or fullscreen right after a server restart), or the user turned on `terminalWheelLocalScrollback`. **opencode is the fifth case, and the gate is false there by design**: its TUI runs on the ALTERNATE SCREEN (1.18.31 measured: tmux `alternate_on=1`, `history_size=0`), so the buffer is hollow, and it IGNORES SGR wheel reports entirely (six `\x1b[<64;…M` reports against an idle pane left the capture byte-identical) while still paging its transcript on PageUp/PageDown (`messages_page_up/down`) — so paging is the only gesture that can reach it, and without it the wheel was silently dead in every opencode tab. The 1.12.0 retest reported exactly that shape for Claude: a wheel that did nothing at all while Fn+Up (PageUp) paged back through intact text, which is the proof that the CLI's own history and the PTY input path were both fine. So under the guard (`_localScrollbackIsHollow()`: the `transcriptPageKeys` capability, gate false, `baseY === 0`) wheel and touch travel is translated into coalesced `\x1b[5~` / `\x1b[6~` through the same 40ms queue as the SGR reports, at half a screen of travel per page key (the key jumps a whole screen; a 1:1 mapping was unusably slow with a discrete wheel). ⚠️ The first event of a trackpad-sized gesture (opening under 2 rows) pages AT ONCE and the travel it skipped is owed back by the rest of the gesture (one page per half screen still). Waiting for half a screen first meant an ordinary trackpad flick (well under 19 rows on a 38-row pane) sent nothing at all. A larger opening event is a mouse-wheel notch (100 px is 4 rows) and accumulates as before, or slow notches, each its own gesture past the gap, would page a full screen apiece. ⚠️ What starts a NEW gesture (and so may pre-pay) is narrow, because every boundary is another free page: (1) a TOUCH CONTACT is one gesture, never split by time: the touch paths tag every call, touchmove and momentum frames alike, with `touchId` (one per touchstart), and a different id (or a wheel event after a touch) is a boundary. touchmove only calls once per whole cell of finger travel, so a slow drag arrives as single lines further apart than any gap; on the gap alone 12 lines at 170 ms each sent 12 PageUps (measured through the real listeners, `test/terminal-touch-paging.browser.test.ts`). (2) Outside a contact, 150 ms of quiet (`PAGE_KEY_GESTURE_GAP_MS`), stretched to one second (`PAGE_KEY_DEBT_MS`) while a pre-paid page is still owed in the SAME direction, so slow small wheel ticks (40 px every 250 ms, a Firefox 1-line notch) pay the debt back instead of each pre-paying a page; after a second of quiet the debt lapses and the next small tick pages at once again. The debt also ends the moment travel repays it (reaches or crosses zero): the rest is plain travel and the 150 ms gap applies again. (3) A direction change inside a gesture (within the gap, or within one contact) drops a pre-paid debt (it belongs to the other direction) and accumulates WITHOUT pre-paying, so a +/-3 px trackpad wobble pages once for its opening event and then nothing; a reversal after the gap is a new gesture and pages at once. A tab switch also starts a new gesture. ⚠️ These gesture rules (first-event page, contact and debt lifetimes, reversal handling, notch accumulation, pinch and sideways-swipe consume) are ONE pure helper, `CodemanTerminalInput.pageKeysForGesture(state, ev, lines, rows, now)`, which both panes call with their own state; never re-derive them in a pane. ⚠️ Three gestures are consumed WITHOUT paging: a pinch (`ctrlKey`, Chrome's trackpad-pinch wheel events), a mostly horizontal swipe (both deltas present and `|deltaX| > |deltaY|`; the touch path locks its own axis first), and anything while the pane's OWN session has the alert `'action'` (a pending `permission_prompt` or `elicitation_dialog`, `updateTabAlertFromHooks` in app.js), so a page key never lands in an open dialog's selector; the primary pane logs that case as `page-keys-dialog-open` so a wheel dead under a stale alert leaves a `[scroll]` line. The dialog check is the one rule outside the helper, because it is per pane: the primary pane reads `tabAlerts` for `activeSessionId`, a tile for its own `sessionId` (usually not the active one). ⚠️ Which modes page is the registry's `capabilities.transcriptPageKeys`, published as `window.__codemanTranscriptPageKeys`, never an id check in terminal-ui.js. Today that is claude, codex and opencode. Codex is on it although it is never forwarded the wheel: newer codex draws on tmux's alternate screen, which the full strip hides from xterm, so its pane is hollow too, Line truncated
 
 **Why the wheel went where it went is LOGGED** (`_logScrollRouting`): one console line per session per distinct decision — `[scroll] <id> → forward-sgr|page-keys|local-scrollback|repull-refused-downgrade (mode=…, cliVersion=…, localScrollbackOptOut=…, mouseTracking=…, localScrollbackRows=…)`. #205 ran two rounds of remote guesswork over questions this line answers directly; keep it when touching the routing.
 
@@ -258,7 +258,7 @@ Further detail, closing: ⚠️ **Closing has the mirror-image race and one owne
 - **No start, attach or relaunch may be in flight** (`Session.paneLifecycleInFlight`, raised for the whole of `_setupOrAttachMuxSession()` and `restartCli()`). The dead-pane respawn revives an exited pane on purpose, and the pane reads as dead until `clearPaneExitForNewPane()` runs after its startup delay.
 - **The pane must have been up for `CLEAN_EXIT_MIN_PANE_LIFETIME_MS` (10 s)** since the last start, attach or relaunch finished (`Session.paneStartedAt`). A CLI that prints a startup error and exits 0 would otherwise lose its tab, and the error with it, seconds after launch; its row stays as `exited (0)` instead. An attach to an already running pane stamps it too, so an `/exit` within seconds of a server restart leaves a row to close by hand.
 
-Scoping needs no check of its own here: `setPaneExit()` already forces `paneExit` to UNKNOWN for direct-PTY, remote, docker and discovered sessions. There is no setting, by the maintainer's decision on #446. ⚠️ Do not flip local panes to `remain-on-exit failed` to get the same effect: a destroyed pane ends the tmux session, the PTY exit nulls the pid, and the browser's `selectSession()` then launches a fresh CLI. `cleanupSession()` keeps `{workingDir}/.claude-images` while another session still uses that directory (`pasteImageDirInUseByOtherSession()`, `paste-image-gc.ts`), since the sweep would otherwise routinely delete a live sibling's pasted images. Paths are compared by `realpath`, and a detached session counts through its persisted record, because `killMux=false` removes it from the map while its pane keeps running; only a session being KILLED is exempt. Each exit gets ONE close attempt (keyed by session id and `at`), and a session being closed refuses `startInteractive()`/`startShell()` (`Session.markClosing()`), so a start that races the close cannot orphan a tmux session. `planRebootRestore()` refuses a record whose persisted `paneExit` is clean (`agent-exited`), which covers an agent that exited just before the power went, before the sweep reached it. Tests: `test/pane-exit-sweep.test.ts`, `test/paste-image-dir-shared.test.ts`, `test/reboot-restore.test.ts`, `test/tmux-manager.test.ts`.
+Scoping needs no check of its own here: `setPaneExit()` already forces `paneExit` to UNKNOWN for direct-PTY, remote, docker and discovered sessions. There is no setting, by the maintainer's decision on #446. ⚠️ Do not flip local panes to `remain-on-exit failed` to get the same effect: a destroyed pane ends the tmux session, the PTY exit nulls the pid, and the browser's `selectSession()` then launches a fresh CLI. `cleanupSession()` keeps `{workingDir}/.codeman-uploads` (and the pre-move `.claude-images`) while another session still uses that directory (`pasteImageDirInUseByOtherSession()`, `paste-image-gc.ts`), since the sweep would otherwise routinely delete a live sibling's pasted images. Paths are compared by `realpath`, and a detached session counts through its persisted record, because `killMux=false` removes it from the map while its pane keeps running; only a session being KILLED is exempt. Each exit gets ONE close attempt (keyed by session id and `at`), and a session being closed refuses `startInteractive()`/`startShell()` (`Session.markClosing()`), so a start that races the close cannot orphan a tmux session. `planRebootRestore()` refuses a record whose persisted `paneExit` is clean (`agent-exited`), which covers an agent that exited just before the power went, before the sweep reached it. Tests: `test/pane-exit-sweep.test.ts`, `test/paste-image-dir-shared.test.ts`, `test/reboot-restore.test.ts`, `test/tmux-manager.test.ts`.
 
 ### Dead-pane respawn: the resume pin
 
@@ -270,6 +270,8 @@ Scoping needs no check of its own here: `setPaneExit()` already forces `paneExit
 
 So: `_confirmIdle()` (session.ts) requires the pane to go quiet, and then asks the SCREEN via `capturePaneText()` + `CLAUDE_WORKING_LINE_PATTERN` before believing it; a sustained run of repaints (`session-activity.ts`, pure + unit tested) is what marks a turn as started, with the same screen probe vetoing keystroke echo. Idle now lands ~3-5s after a turn ends instead of 2s into one.
 
+⚠️ **The workspace-trust dialog is not a composer.** Its `❯` sits on `No, exit`, so it arms the confirmation, and the dialog is quiet with no working line, so a fresh case used to be concluded idle about 4 s after launch while the dialog was still up (measured on Claude Code 2.1.294: `GET .../wait?until=idle` answered `idle` at 3.8 s in 6 of 12 fresh-dir launches). An agent that waited for idle and then prompted typed into the dialog, and the prompt's Enter confirmed `No, exit`: the pane died and the wait reported `exit`. `_probePaneWorking()` therefore also records whether its capture shows `isTrustDialogScreen()` (a failed capture keeps the last answer, since it proves nothing), and `_confirmIdle()` re-checks on `PANE_PROBE_RECHECK_MS` instead of concluding while it does, but only inside the auto-accept's own `TRUST_DIALOG_WINDOW_MS`: a later screen that merely quotes the dialog must not hold a session busy. The status is left alone rather than marked working, since nothing is running. Tests: `test/session-trust-dialog.test.ts`.
+
 ⚠️ **The composer glyph and the working line are per-CLI registry DATA** (`capabilities.workDetect`, #385), not Claude constants: claude declares `❯` plus the pattern above, codex declares `›` plus `[Ee]sc to interrupt`, pi declares its composer rule `─` plus the spinner it embeds there, opencode declares its composer bar `┃` plus its 8-cell footer spinner `[⬝■]{8}`, omp declares its input row `╰─` plus the spinner and elapsed time that replace the `π` in its status bar, gemini declares its composer's bottom bar `▀` plus its `(esc to cancel, <n>s)` spinner line, and a CLI that declares neither falls back to Claude's pair, which is what every session used before the registry carried one. Before that, this whole mechanism was gated Claude-mode-only on the reasoning that an external CLI has no `❯`, which was true and still left every Codex session reporting `idle` for its entire life.
 
 ⚠️ `workingLine` is config-supplied (a user `clis.json` can set it) and the compiled pattern runs on the PTY hot path, so it goes through `compileVersionRegex()` in BOTH the schema refine and `_workingLinePattern()`: a nested quantifier there is a ReDoS against the event loop, and the helper returns null rather than throwing so the fallback is structural.
@@ -278,6 +280,8 @@ So: `_confirmIdle()` (session.ts) requires the pane to go quiet, and then asks t
 
 ⚠️ A fourth, optional `workDetect` field, `awaitingLine`, makes a turn that ENDED waiting for workers the CLI started count as WORKING: with background agents or an ultracode workflow still running at turn end, Claude closes the turn with `✻ Waiting for 1 dynamic workflow to finish` instead of `✻ Brewed for 1m 18s`, then resumes by itself when they report back, so the pane is quiet with the composer up and every other signal read it as idle (reported 2026-09-28 on a live 2.1.283 session). `_probePaneWorking()` ORs it into the same capture. ⚠️ It must never be searched across the pane like `workingLine`: Claude renders the row from a `useState` snapshot taken at turn end and never redraws it, so the words stay on screen after the workers finish and a pane-wide match would pin the tab busy until they scroll off. `isAwaitingWorkers()` (pure, `session-activity.ts`) finds the composer (the LAST row carrying `promptGlyph`, bare or boxed), walks up at most `AWAITING_SEARCH_ROWS` past blank, box-drawn and indented rows, and tests only the first column-0 row. ⚠️ Keep the pattern anchored on column 0 (`^✻`): Claude's own rows start there and the agent's prose never does, which is what stops an agent from holding its own session busy. It is deliberately NOT tested on the PTY stream, where redraws of the stale row would re-mark working. Tests: `test/session-awaiting-workers.test.ts`.
 
+⚠️ A fifth, optional `workDetect` field, `blockedLine` (plus `blockedLines`), marks a pane showing a dialog the CLI is blocked on, and it WINS over the working line in the same capture (codex's question dialog carries `esc to interrupt`, its working line, in its own key hints). See [Screen-read blocked dialogs](#screen-read-blocked-dialogs-codex).
+
 ### The watching signal (a quiet pane that is not waiting for you)
 
 **A session that armed a monitor, backgrounded a shell or started a background terminal ends its turn and goes quiet, and a minute later Claude Code's idle notification arrives.** Before this existed, that prompt became an approval item like any other, so every surface filed the session under NEEDS YOU with nothing for a human to answer. The CLI says which kind of quiet it is on its own screen, and reading that row is the whole mechanism: `capabilities.workDetect.watchingLine` (optional, per CLI) plus `watchingLines` (how many non-blank rows at the foot of the screen may hold it, default `WATCHING_TAIL_LINES` = 1). `_confirmIdle()` already captures the pane at the moment a turn ends, so `_readWatching()` runs `watchingLabel()` (pure, `session-activity.ts`) over that same capture; the label lands on `Session.watching` and rides `toLightDetailedState()` out to every payload. ⚠️ It is cached BESIDE `_lastPaneProbeWorking` and goes stale with it, because the probe returns its cached boolean without re-capturing inside `PANE_PROBE_MIN_INTERVAL_MS` and a label from a capture nobody took is a guess. ⚠️ A capture that FAILS clears the label (and announces the change) rather than keeping the last one: a stale label opens the next idle prompt already acknowledged, so keeping it would turn a failed `capture-pane` into a missed alert, while clearing it costs at most an alert the next readable capture takes back. ⚠️ It then FREEZES once `_confirmIdle()` concludes — nothing looks at the pane again until it produces output — which is correct rather than tolerable, since work ending repaints the pane either way (a monitor firing wakes the agent; codex drops its background-terminal row by itself); a timer to keep it fresh would spend a `capture-pane` per idle session per tick to learn nothing. ⚠️ A server restart looks like a hole in that and is not one: the field is live state and starts empty, but reconciliation re-attaches the pane and the attach repaint arms the idle confirmation, which probes and re-reads the label with no input from anyone (measured 2026-09-23, back within ~20 s). A restored session showing no label has no chip on its screen.
@@ -286,7 +290,17 @@ So: `_confirmIdle()` (session.ts) requires the pane to go quiet, and then asks t
 
 ⚠️ **A chip that waits on a HUMAN must never count as watching**: Claude's Artifact comment monitor (an agent that published a page and hears nothing until somebody comments) is refused for the whole row by a `^`-anchored negative lookahead keyed on "Artifact" alone, so a footer cut off mid-chip is refused too, and the `^` is what stops the engine retrying past the chip and reporting a shell beside it.
 
-**⚠️ The label is pane-derived, so the window and the anchor are a trust boundary, not formatting.** An agent that gets its own text matched silences its own alert. Two things prevent it, and BOTH belong to whoever adds a pattern for a new CLI: the window must cover only rows that CLI draws, and the pattern must anchor on chrome only that CLI can produce. Claude satisfies both — its chip is the LAST row, so the default window of one row excludes even the status line directly above it, whose content comes from a `statusLine` command a bypassed session can write into its own `.claude/settings.json`. Codex does not: its row sits above the composer, and the slot it occupies holds the last row of the TRANSCRIPT whenever no terminal is running, so matching the complete row raises the bar without closing it. What contains that is `hooks: 'none'` — no hook event from a codex session reaches `notePrompt()`, so a forged label costs a wrong badge and nothing else, and a CLI that gains hook signals must not keep a pattern that soft. The label is also ANSI-stripped and capped (`MAX_WATCHING_LABEL_CHARS`) at the source, and every interpolation of it into markup goes through `escapeHtml()`, since a config-supplied capture group decides what it holds. Tests: `test/session-watching.test.ts` (the label and both CLI patterns), `test/watching-no-alert.test.ts` (the negative claim across all four surfaces).
+**⚠️ The label is pane-derived, so the window and the anchor are a trust boundary, not formatting.** An agent that gets its own text matched silences its own alert. Two things prevent it, and BOTH belong to whoever adds a pattern for a new CLI: the window must cover only rows that CLI draws, and the pattern must anchor on chrome only that CLI can produce. Claude satisfies both: its chip is the LAST row, so the default window of one row excludes even the status line directly above it, whose content comes from a `statusLine` command a bypassed session can write into its own `.claude/settings.json`. Codex does not: its row sits above the composer, and the slot it occupies holds the last row of the TRANSCRIPT whenever no terminal is running, so matching the complete row raises the bar without closing it. What contains that is `hooks: 'none'`: no IDLE prompt from a codex session reaches `notePrompt()` (its screen-read dialogs open PERMISSION items, which a watching label never acknowledges), so a forged label costs a wrong badge and nothing else, and a CLI that gains idle signals must not keep a pattern that soft. The label is also ANSI-stripped and capped (`MAX_WATCHING_LABEL_CHARS`) at the source, and every interpolation of it into markup goes through `escapeHtml()`, since a config-supplied capture group decides what it holds. Tests: `test/session-watching.test.ts` (the label and both CLI patterns), `test/watching-no-alert.test.ts` (the negative claim across all four surfaces).
+
+### Screen-read blocked dialogs (codex)
+
+**Codex is the one CLI Codeman launches with its approvals ON by default, and it has no hook to say it stopped for one** (`hooks: 'none'`), so a codex session parked on "Would you like to run the following command?" read as plainly idle, and one parked on a question as WORKING. `capabilities.workDetect.blockedLine` (optional, per CLI, plus `blockedLines`, 1 to 24, default 1) names the dialog's chrome; `_probePaneWorking()` reads it off the same capture the idle confirmation takes (`_paneShowsBlocked()` → `isBlockedScreen()`, pure, `session-activity.ts`: the last `blockedLines` non-blank rows joined with `\n`, tested once, so a pattern can demand several markers and end on `$`), and `Session` emits `blockedChanged`. `session-listener-wiring.ts` hands that to `web/screen-blocked.ts`, which raises EXACTLY what a Claude `permission_prompt` hook raises (`sessionWaits.notifySignal('blocked')`, an Approvals Inbox item, the `hook:permission_prompt` broadcast with `source: 'screen'`, push and webhook, a run-summary row), so the tab alert, phone overview, TUI and every other consumer needed no second code path; clearing resolves the item it opened, and `approval:resolved` clears every client.
+
+⚠️ **ALERT ONLY, never a keystroke.** The text is pane-derived and the agent writes most of the pane, so the pattern must end on chrome only the CLI draws at the FOOT of the screen, where a dialog replaces the composer (codex's: a `› N.` option row, then its key-hint row as the last row, allowing for that row to wrap; a transcript that quotes every word still has codex's column-0 composer row below it). The item is `answerInTerminal` (#610's shape), the answer route refuses it, push carries no `approvalId`, and Codeman never answers such a dialog itself, so a forged match costs a false alert and nothing more. The message every surface shows is built from the CLI's registry label, never from the screen (a question's text and a command preview are the agent's to choose).
+
+⚠️ **Raise on one reading, clear on two, and keep reading while it is up.** The raise needs no timer: a dialog is drawn at the END of a burst of output, so the settle probe (pane quiet, then capture) sees it. The clear does: answering resumes the turn, a busy pane never reaches the settle path again until the turn ends, and the red alert outlived the dialog by minutes, so `_blockedRecheckTimer` re-reads the screen every `BLOCKED_RECHECK_MS` (2 s, one `capture-pane`, only while a dialog is up; cleared in `_clearAllTimers()`). ⚠️ That read is a synchronous exec (about 8 ms of blocked event loop, measured 2026-10-14) and a dialog can sit unanswered overnight, so the timer SKIPS it while the pane is quiet: a dialog leaves the screen only through a repaint, which reaches the session as PTY output, so nothing drawn since the last reading that saw the dialog (`_lastActivityAt < _blockedSeenAt`) means it is still up; a safety-net read still runs every `BLOCKED_QUIET_RECHECK_MS` (30 s). Never once a reading has MISSED the dialog (`_blockedGoneReads > 0`): tmux updates the pane before it sends this client the repaint, and the confirming read must not wait for output. It clears after `BLOCKED_CLEAR_READS` (2) readable captures in a row without the dialog, because one capture can land inside a full-frame repaint, and a clear followed by a re-raise sends the alert, push and webhook twice for one dialog. An unreadable capture (`null`) changes nothing in either direction. A new pane (`_resetBuffers()`) and a PTY exit clear it, announced, while the listeners are still attached. The working probe decides blocked-over-working from THIS capture, never from `_screenBlocked`, which lags a vanished dialog by design.
+
+⚠️ **Hooks win.** `raiseScreenBlocked()` does nothing for a session whose CLI passes `hooksAvailableForMode()`, so a CLI that declared both could never alert twice for one dialog; and the inbox item gets no `capture` callback (only the frame, as context), because the inbox's option parser and staleness sweep are shaped for Claude's dialogs and the SESSION owns this item's lifetime. ⚠️ Codex's `Implement this plan?` after a plan turn is deliberately not matched: its `enter select · esc back` footer is the same as `/model`, `/permissions` and `/review`, menus the user opened at the keyboard, and the plan turn has ended there. Measured fixtures (codex-cli 0.162.0, 2026-10-14, widths 200 to 36): `test/fixtures/codex-dialogs/`. Tests: `test/codex-blocked-detection.test.ts` (the pattern and the session state), `test/screen-blocked.test.ts` (what the server raises), the codex blocks of `test/routes/session-wait-routes.test.ts` and `test/routes/approval-routes.test.ts`.
 
 ### Workspace-trust dialog auto-accept
 
@@ -300,7 +314,7 @@ So: `_confirmIdle()` (session.ts) requires the pane to go quiet, and then asks t
 
 ⚠️ It reads `capturePaneText()` and falls back to a deliberately SHORT tail of the terminal buffer only on a direct-PTY session, which has no pane: that buffer is append-only, so a longer tail would keep re-matching a dialog answered minutes ago.
 
-⚠️ **The scan must schedule its own next read** (`_trustDialogTimer`, cleared in `_clearAllTimers()`): it runs from the PTY `onData` handler, which was enough while one Enter answered the dialog, but the arrow that moves the cursor is the LAST output the pane produces, so a two-keystroke answer waiting on more output parks forever with the cursor sitting on the right option (measured on a live 2.1.252 spawn: cursor moved at 6 s, then nothing).
+⚠️ **The scan must schedule its own next read** (`_trustDialogTimer`, cleared in `_clearAllTimers()`): it runs from the PTY `onData` handler, which was enough while one Enter answered the dialog, but the arrow that moves the cursor is the LAST output the pane produces, so a two-keystroke answer waiting on more output parks forever with the cursor sitting on the right option (measured on a live 2.1.252 spawn: cursor moved at 6 s, then nothing). ⚠️ For the same reason a read the `TRUST_DIALOG_RETRY_MS` throttle refuses is DEFERRED onto that timer, never dropped: the dialog's own paint is often the last output a launching pane produces and lands inside the throttle of the read the blank screen before it triggered, so dropping it left 2 of 12 fresh launches on Claude Code 2.1.294 sitting on the dialog with no keystroke ever sent. And a stopped session never scans: killing the pane produces output, the buffer tail still holds the dialog, and a deleted session went on scanning it and arming the timer for several seconds.
 
 ### Owner tab layouts
 
@@ -310,12 +324,13 @@ So: `_confirmIdle()` (session.ts) requires the pane to go quiet, and then asks t
 
 - **Grouped iff vertical AND the owner has at least one group.** No layout, a failed `GET` (newest-wins via `createLoadCoordinator`, retried at 5/10/20/40 s and then left to the next SSE init or `tab:layoutChanged`) or zero groups renders the flat rail unchanged; the horizontal strip, phones and the sidebar never group. ⚠️ Adopting a layout rebuilds the strip ONLY when the structure key changed (`_applyTabLayout`): the server announces a change on every session create/close and order PUT, and the key deliberately leaves the layout version out, so those announcements cost a flat rail nothing.
 - **A render layer, never an order source.** `sessionOrder` (the server-projected global order), Alt+N, Ctrl+Tab and the palette are untouched; a grouped session row is the flat row's markup, so its badge still names its Alt+N slot. Web tabs keep their slot after every session wherever their group puts them (`renderWebviewTab`).
-- **Collapse is per-device** (`codeman:tab-groups-collapsed` in localStorage, ids of deleted groups garbage-collected on adoption). A store that throws means all-expanded; a malformed stored VALUE reads as empty and is rewritten, so it can never disable collapse on that device for good. A collapsed group still SHOWS the active row, and `_updateActiveTabImmediate` falls through to a full render whenever the structure key changes, since a class toggle cannot reveal a hidden row.
+- **Collapse is per-device** (`codeman:tab-groups-collapsed` in localStorage, ids of deleted groups garbage-collected on adoption). A store that throws means all-expanded; a malformed stored VALUE reads as empty and is rewritten, so it can never disable collapse on that device for good. A collapsed group still SHOWS the active row, and `_updateActiveTabImmediate` falls through to a full render whenever the structure key changes, since a class toggle cannot reveal a hidden row. The one exception is the rail's **Search sessions** box (#580): while it holds text `_projectTabGroups()` passes `collapsedGroupIds: []`, so every group is DRAWN open and a match inside a collapsed one can show, and `toggleTabGroupCollapsed()` refuses; the stored set and its localStorage key are never written, so clearing the search puts the collapse back exactly as it was (see Session list layout, the rail search).
 - **A collapsed header carries the most urgent alert it hides** (`hiddenGroupAlerts()`, applied by `_syncTabGroupHeaderAlerts` on BOTH render paths, since alerts change without a rebuild), in the tab alert language: `tab-alert-action` red, `tab-alert-idle` yellow. A permission prompt behind a collapse must never be invisible.
 - **Lineage arcs to a collapse-hidden session anchor to its group header** (`lineage-line--proxied`); two endpoints proxied to one header draw nothing.
 - **The upstream HTML5 drag stays off in the grouped rail**: a flat-order drop cannot express a group move, and the server re-ranks within the old group. The grouped rail has its OWN pointer drag instead (`_bindTabLayoutPointerDrag`, mouse/pen only, bound once on the container): rows move before/after a row or into a group, a header drag reorders groups, and the drop maps to ONE operation through the pure `dropOperation()`. Escape cancels a drag, and the click that ends one is swallowed. ⚠️ A press is only captured once it moves 6 px, so its release can land outside the rail: `pointerup`/`pointercancel` are heard on `window` while a press is pending, a move with the primary button up cancels it, and a new press cancels any previous one. Without that a stale press became a phantom drag on the next hover, and a replaced drag left its capture-phase Escape listener behind, swallowing every Escape before the terminal saw it. The flat rail and the header strip keep the HTML5 drag untouched. For the same reason Ctrl+Shift+{ / } only swaps with a neighbour in the active session's own section (`_canSwapActiveTabWith`, reading the projection's `sectionByRef`): a cross-group swap moves nothing on the server, gets no `session:orderChanged` back, and would leave this client's `sessionOrder` and Alt+N targets out of step with every other device.
 - **Edits are named operations, saved serially.** `createEditCoordinator` applies `createGroup` / `renameGroup` / `deleteGroup` / `reorderGroup` / `moveRef` to the rail at once, then sends ONE `PUT /api/tab-layout {baseVersion, layout}` at a time; edits made meanwhile wait and go out on the version that write returns. A 409 carries the server's layout: the in-flight operations are replayed onto it (an operation that no longer applies is dropped and reported) and re-sent, at most `maxAttempts` times; a 400 re-reads once (a second 400 is reported as a failed save, not as a race); anything else reports and re-reads. A `moveRef` moves the session together with the sessions that still follow it and marks a hand-moved child `placement: 'manual'`, mirroring the server's `moveRef`, and `normalizeLayout` keeps `placement` because whole layouts are written back. `_onTabLayoutChanged` / `_applyTabLayout` defer a read while a write is in flight and rebase unsaved edits onto a read otherwise. ⚠️ A FAILED read (`_applyTabLayout(null)`) while edits are pending keeps the held layout and the editor and re-reads after the write settles: disposing there would orphan the in-flight write, whose 409 then never gets its rebase. Any path that does drop unsaved work says so in a toast. On `pagehide`, unconfirmed operations go out in a `keepalive` PUT AND into sessionStorage as `{ owner, baseVersion, savedAt, operations }`; after reload they replay onto the fresh layout (a no-op when the keepalive landed), but only for the same owner, within `TAB_LAYOUT_PENDING_MAX_AGE_MS` (60 s), and never onto a layout older than the copy's base. A "Move to <group>" with no anchor carries no `index`, so a replay still puts the row last.
 - **Every way in has a keyboard path.** The session row menu (Shift+F10 in the tree, the rail's overflow button) gains Move up/down, Move to <group>, Move to Ungrouped and Move to new group in the vertical rail (only "new group" before the first group exists; nothing on the strip). A group header opens its menu with Shift+F10 / ContextMenu, right-click or its hover glyph (a non-focusable, `aria-hidden` span: a treeitem holds no interactive children), and F2 renames it inline (Enter commits and refocuses the header; a commit by BLUR leaves focus where it went, since refocusing from inside the blur handler overrides the user's click). The glyph stays visible under `@media (hover: none)`: a touch tablet has no hover and no long-press `contextmenu`. Group names in "Move to" labels are quoted, so a group named "New group" or "ungrouped" cannot read (or translate, case-insensitively) like the fixed entries. The menu closes on Escape (which it consumes before the global Escape handler), a pointer outside, Tab, focus leaving it, a resize, a second open and any full re-render. It borrows the `.tab-rail-action-menu` class for its look only: `closeTabRailActionMenu()` excludes `.tab-layout-group-action-menu`, so closing the row menu (every `session:deleted` does) cannot strand the group menu's listeners. The inline group editor shares `_activeRename` with the session rename, so only the CURRENT editor may release `_inlineRenameActive`.
+- **New, child and web tabs land in groups.** A group menu's "New session" (`createSessionInTabGroup`) runs ONE launch with `run({ count: 1, tabGroupId })`; every launch path puts `tabGroupId` on its create request (`POST /api/sessions` or `/api/quick-start`), and `TabLayoutService.sessionCreated(owner, id, { tabGroupId }, register)` places the session at the end of that group in the SAME versioned write that first records it. The session enters the live map inside the owner lock (`register`, rolled back if the write fails), so no queued layout edit can commit it first. The group is looked up in the creating owner's own layout, so an unknown, deleted or foreign id is ignored rather than failing the spawn, and an explicit group is a hand placement (a child placed this way is `manual`). The create response carries the committed `tabLayout`, adopted before the tab is drawn. A session spawned by another one (`parentSessionId`) and placed by nobody lands after its parent: the server's normalization puts it after the parent's rows in the parent's container, and `project({ parents })` draws a not-yet-stored child there at once. Whether it then FOLLOWS its parent is the synced `spawnedTabsFollowParent` setting (default OFF), which `sessionCreated()` reads fresh through the `childrenFollowParent` dep: off, the child is marked `manual` in that same creating write; on, it follows until moved by hand. Deciding at creation is enough, so flipping the setting leaves existing tabs alone. A hand move makes a following child `manual`; with the setting on, the row menu notes where a child sits (following, placed by hand, parent closed, parent loop; plain text, never a focusable menu item) and offers "Follow parent again" only to a `manual` child, which clears `placement` and is re-derived by the server on save. With the setting off the row menu says nothing about placement. A legacy `PUT /api/session-order` (strip or phone drag, a same-section Ctrl+Shift+{/}, a device without a layout, an admin's cross-owner order) marks a child `manual` ONLY when the requested order splits it from its parent's block (`applyLegacySessionRank`, `childrenSplitFromParent`); a PUT of the current order, or one that moves a family together, leaves every stored placement alone. The browser never echoes a create or delete through that PUT, with or without groups: the service already re-projects the order and broadcasts it (`session:orderChanged`), and `syncSessionOrder` adopts it. A child whose parent is gone stands on its own and is never re-adopted. Saved web tabs are layout refs like sessions, but OPENNESS stays per device: a grouped web tab renders (and takes an Alt+N slot) only while it is open on this device, and reopens where the layout keeps it.
 - **Only the grouped rail is a tree.** `#sessionTabs` ships as `role=tablist` with `role=tab` rows, and the header strip, sidebar and flat rail keep exactly that. While grouped, `_applyTabListRole` makes it `role=tree` (and restores `tablist` + its label when grouping ends), named-group headers are level-1 `treeitem`s that `aria-owns` their rows' `role=group` (rows sit beside the header, not inside it), and ungrouped rows plus a collapsed group's kept selection are level-1 items. A collapsed header owns nothing, a group with no open rows is a leaf (no `aria-expanded`, no owned group), and the "Ungrouped" heading is `aria-hidden`. Rows are re-roled in the DOM by `_applyTabTreeSemantics` after render, never by rewriting their markup, so a grouped row's content stays the flat row's.
 - **One tab stop in the tree.** Exactly one treeitem carries `tabindex=0` (the focused or selected item); every control inside a row drops to `-1`, which is why Shift+F10 / ContextMenu open a row's actions from the keyboard. Focus survives a full re-render by identity (`group:`/`session:`/`webview:`; a row a collapse just hid hands focus to its header), but only when focus was already inside the rail. The tree walk (`_tabTreeItems`) follows painted order WITHIN each group when the rail is sorted; the flat list keeps its own whole-list computed-order walk. `aria-posinset`/`aria-setsize` follow painted order too, so the incremental render path re-runs `_applyTabTreePositions` after it re-sorts rows in place. ⚠️ `_handleTabTreeKeydown` acts only when the key lands on the treeitem ITSELF: a key on a focused in-row control (close, overflow, the rename input) is that control's, or Enter on the overflow button re-selects the row instead of reopening its menu. ⚠️ The roving `tabindex=-1` also hides every item but the stop from the keyboard-dismiss selector's `[tabindex]` arm, which is why `MOBILE_KEYBOARD_DISMISS_EXEMPT_SELECTOR` lists `[role="treeitem"]` (see Dismissing the on-screen keyboard).
 
@@ -355,7 +370,7 @@ The escape hatch is the synced `workspaceHooksEnabled` setting (App Settings →
 
 ⚠️ Claude-mode only (others carry their conversation id in their own config object), and remote/docker sessions are never offered (`remote-or-docker`), because both need another host or container to be up.
 
-⚠️ A failed rebuild is undone with `discardPartiallyBuiltSession()`, deliberately NOT `cleanupSession()`: the delete path would count the session's tokens into the lifetime totals, demote a pinned record to `stopped` (which this pass reads as an intentional kill, making the session permanently unrestorable) and recursively remove the WORKSPACE's `.claude-images`.
+⚠️ A failed rebuild is undone with `discardPartiallyBuiltSession()`, deliberately NOT `cleanupSession()`: the delete path would count the session's tokens into the lifetime totals, demote a pinned record to `stopped` (which this pass reads as an intentional kill, making the session permanently unrestorable) and recursively remove the WORKSPACE's upload dirs (`.codeman-uploads`, and the pre-move `.claude-images`).
 
 ⚠️ `os.uptime()` reports the HOST's uptime, which a container shares, and that cuts both ways: after a genuine host reboot a containerized Codeman does see a short uptime and the banner works, but a container-only restart is invisible to it, which is the case where this would help most. Tests: `test/reboot-restore.test.ts`, `test/routes/reboot-restore-routes.test.ts`, `test/routes/reboot-restore-rebuild-failure.test.ts`, `test/discard-partially-built-session.test.ts`.
 
@@ -369,15 +384,19 @@ The escape hatch is the synced `workspaceHooksEnabled` setting (App Settings →
 
 A file path an agent prints is a link on both surfaces it can appear on, and clicking it opens the file-preview overlay. Three things make that work and each has bitten:
 
-**One pattern, two consumers.** `FILE_PATH_LINK_PATTERN` / `absoluteFilePathPattern()` live in `constants.js`; the xterm link provider (`registerFilePathLinkProvider`, terminal-ui.js) and the response viewer's `_linkifyFilePaths()` (app.js) both build a fresh instance from it. ⚠️ Fresh per call, never one shared object: `lastIndex` is per-object state on a `/g` regex. The pattern is anchored on a known absolute root and terminated by a known extension, so a fraction (`3/4`) or a date can't match and trailing punctuation stays out. Roots include `Users` and `mnt`, without which nothing was clickable on macOS or WSL. The linear-time guard and the "terminal-ui builds from the factory" structural check are in `test/link-provider-regex.test.ts`.
+**One pattern, two consumers.** `FILE_PATH_LINK_PATTERN` / `absoluteFilePathPattern()` live in `constants.js`; the xterm link provider (`registerFilePathLinkProvider`, terminal-ui.js) and the response viewer's `_linkifyFilePaths()` (app.js) both reach it through `findFilePathLinks()` (below), which builds a fresh instance per call. ⚠️ Fresh per call, never one shared object: `lastIndex` is per-object state on a `/g` regex. The pattern is anchored on a known absolute root and terminated by a known extension, so a fraction (`3/4`) or a date can't match and trailing punctuation stays out. Roots include `Users` and `mnt`, without which nothing was clickable on macOS or WSL. The linear-time guard and the "terminal-ui builds from the factory" structural check are in `test/link-provider-regex.test.ts`.
 
 **The chat linkifier walks text nodes.** `_linkifyFilePaths()` builds anchors with `createElement`/`textContent` on the rendered subtree, never by rebuilding sanitized markup as a string — the source is model output. Subtrees already inside an `<a>` are skipped (marked autolinks URLs; a nested anchor would swallow the click), and the anchor's text is the path verbatim so "copy code" still yields what the agent printed. `test/response-viewer-file-links.test.ts` pins both properties.
 
 **Out-of-workspace paths go through the attachment routes, not the file routes.** `file-content`/`file-raw` resolve against `workingDir` and 404 anything that escapes it, which is correct and unchanged — but the paths agents most often print (a `/tmp` capture, Claude's own scratchpad, another checkout) are exactly that, so clicking one used to report "File not found" for a file sitting on disk. `openFilePreview()` now detects the case (`_isExternalPreviewPath`, a string compare for ROUTING only; the real decision stays server-side) and registers the path via `POST /api/sessions/:id/attachments` first, rendering by id. ⚠️ That registration passes `notify: false`, which suppresses ONLY the `attachment:detected` broadcast — the guard, the registry entry and the by-id routes are identical either way. Without it every click also popped an attachment card announcing the file already filling the screen. ⚠️ The click is an explicit user action on the **explicit, Origin-guarded** registration route, which is why it may cross the workspace boundary at all; the passive magic-link scanner stays force-confined. A type outside `SUPPORTED_ATTACHMENT_EXTENSIONS` (`.svg`, `.bmp`) is refused with a message naming what IS previewable, rather than the registry's own policy term.
 
-⚠️ **The terminal routes an out-of-workspace path to the preview, not the log viewer.** The log viewer spawns `tail -f` and allows only the workspace, `/var/log` and `~/logs`, so an external `.log`/`.json`/code path answered `Path must be within working directory or allowed log directories` while the SAME path clicked in the response viewer previewed fine. `activate()` now checks `_isExternalPreviewPath` alongside `previewsInFileViewer`. In-workspace text keeps the tail viewer, which is the point of it (live follow); nothing widened `file-stream-manager`'s allowlist, so no `tail -f` is spawned on an arbitrary host path.
+⚠️ **The terminal routes an out-of-workspace path to the preview, not the log viewer.** The log viewer spawns `tail -f` and allows only the workspace, `/var/log` and `~/logs`, so an external `.log`/`.json`/code path answered `Path must be within working directory or allowed log directories` while the SAME path clicked in the response viewer previewed fine. `activate()` now checks `_isExternalPreviewPath` alongside `previewsInFileViewer`. In-workspace text keeps the tail viewer, which is the point of it (live follow), except markdown (`.md`/`.markdown`, in `FILE_PREVIEW_EXTENSIONS` since the owner's 2026-10-13 decision, rendered as a document) and `.html`/`.htm` (`rendersAsPage()`, rendered as a page, see below), which always go to the preview; nothing widened `file-stream-manager`'s allowlist, so no `tail -f` is spawned on an arbitrary host path.
 
-**Text reuses the edit-mode allowlist; markup stays download-only.** `TEXT_ATTACHMENT_EXTENSIONS` IS `EDITABLE_EXTENSIONS` (`config/file-editing.ts`) rather than a second curated list that would drift from it: if the viewer would open a file for editing inside the workspace, the same file outside it can be read. The justification for widening is that the agent in the session can already `cat` any of these and the picker already previews them, so the suffix was never the confidentiality gate; the path guard is (sensitive-file blocklist, `/root` and `/etc` trees, realpath first). ⚠️ Two consequences had to be handled at the same time: `~/.codeman*/state.json` joined `isSensitivePath` (it persists `SessionState.envOverrides`, and the env allowlist admits key-shaped names like `GEMINI_API_KEY`, so it can hold a live credential), and `html`/`htm` joined `svg` in `serveRawFile`'s **download-only** branch so that widening what can be READ never widens what can RUN on our own origin. Text with no dedicated MIME entry goes out as inert `text/plain; charset=utf-8` + `nosniff`, matching the picker. The by-id text preview is bounded like the workspace one: a `Range` request for the first 512KB (a real partial read, not a discarded 50MB download) plus a 500-line cap, with the footer saying so.
+**Path forms beyond the rooted one.** `findFilePathLinks(text)` (constants.js) is the single entry point both consumers call; it returns the rooted matches first, then `~/` and relative paths (`RELATIVE_FILE_PATH_LINK_PATTERN`, same extension list, pinned equal by `test/link-provider-regex.test.ts`), then folder paths, each only where it overlaps neither an earlier match nor an `http(s)` URL. A relative match must start at a boundary (line start, whitespace, quote, bracket, `=`, `,`, `>`), so it can never begin inside a word, an absolute path or a URL; brackets, `=` and `,` end it (`(shot.png)`, `OUT=shots/a.png`). `resolveLinkedFilePath()` resolves a relative path against the session's `workingDir` (`.`/`..` collapsed, so `../x.png` lands outside and routes like any external path) and passes absolute and `~/` paths through. ⚠️ `~` is the home of the SESSION'S host, which only the server knows: `~/` paths always take the attachment route, where `registerExternalAttachment` expands them (`homedir()` locally, `$HOME` inside the remote probe script), and the guard runs on the expanded, realpath-resolved path as before. ⚠️ Programs hard-wrap their own output short of the terminal edge (Claude Code breaks tool output at 59-118 columns), so xterm's `isWrapped` never fires. `terminalPathLine()` (link provider only; touch selection keeps `terminalLogicalLine`) joins the next row when the previous row ends in a path token with a `/` and no extension and the join reaches an extension; a row that starts a new path is never joined: rooted or `~/`, `./`/`../`, or a relative one that opens on the same top directory the broken token did (the next entry of `git ls-files`/`find`/`ls -R`: `src/components` + `src/index.ts` are two paths, not `src/componentssrc/index.ts`), while a wrapped path continues INSIDE a name, so `…/captures/2` + `026-10-10/…` still joins; a 3+ row chain is kept only if a later row completes the path, and the broken row must be at least as long as the continuing one (keeps `cd /mnt/foo` + a longer next line apart). Tests: `test/file-path-link-forms.test.ts`.
+
+**Folder grid: a folder is granted, then served by bare file name.** A folder path (rooted or `~/` whose last segment has no extension, relative only with a trailing `/`; `FOLDER_PATH_LINK_PATTERN` + `folderLinkPath()`) opens `openFolderGrid()` (panels-ui.js), as does the preview's "Show folder" button. `POST /api/sessions/:id/folders {path}` (`src/folder-listing.ts`) resolves the folder (locally a bounded `probePathKind(…, { pastCap: true })` first, then async `fs.realpath`; remotely one ssh round trip via `remoteListFolder`, through `runWithRemoteSshLimit`), runs the attachment guard on it (blocklist, `confineToWorkspace`, and in multi-user mode `folderAllowedForUser`: a non-admin only inside their own `userSpacePath` or the session's `workingDir`, as the filesystem path picker), locally BEFORE reading it, lists non-hidden subfolders and files of a supported attachment type (each file also blocklist-checked, newest first, capped at `MAX_FOLDER_ENTRIES`), and GRANTS it an id. Files are then fetched as `GET …/folders/:folderId/files/:name/(raw|preview|thumbnail)`: `:name` must be a bare file name of a supported type (no `/`, `\`, NUL, no leading `.`, since the listing hides those), the local file is probed bounded (`unknown` → 503, never 404), and the joined path goes through the SAME `resolveServableAttachmentPath` as an attachment by id (symlink re-resolution, remote re-probe, guard), then the same gate registration applies to what a symlink RESOLVES to (a supported extension on the resolved path and a regular file: locally `fs.stat`, remotely the probe kind; else 400, so `shot.png -> ../secret.bin` is not served and a FIFO or device behind a supported name never reaches `open`), then `folderAllowedForUser` once more on the resolved file, before `serveServableRaw` (the attachment raw route's body, shared) streams it. So browser requests still never carry a path. ⚠️ Grants live in their own registry (`folderGrants`, 32 per session, cleared with the session's attachments), NOT the attachment registry: that one is capped at 200 per session, and paging a folder of 300 renders would evict the session's real attachments. ⚠️ A local folder is never touched unbounded (#516 rule): a folder on a network mount that stopped answering answers 503 with `describeUnknownPath()`, not 404 and not a frozen server. ⚠️ The remote listing script tells "missing" (`E|n`, 404) from "is a file" (`E|f`, 400 `Not a folder`), because the client falls back to the file preview on `Not a folder`; a remote extensionless file otherwise answered "Folder not found". ⚠️ `.folder-grid-overlay` sits at `z-index: 5090`, one step BELOW the preview (5100) and above the response viewer (5000): a tile's preview opens on top and closing it lands back on the grid. ⚠️ ←/→/Esc are handled in a document CAPTURE listener with `stopPropagation`, because the terminal keeps keyboard focus under the overlays and an arrow that reached xterm would be typed into the session; inputs other than xterm's helper textarea keep their keys. ⚠️ A route PARAMETER is capped by the router (Fastify's default is 100 characters, answered as 404 "Route not found"), and a file name goes up to 255, percent-encoded: the server sets `routerOptions.maxParamLength` to `MAX_ROUTE_PARAM_LENGTH` (route-helpers.ts), the route test harness uses the same constant, and a wiring guard pins it. ⚠️ `~/` in a Docker case is refused with 400 (attachment POST and `POST …/folders`): the session is local for these routes, but `~` inside the container is the container's own home, and expanding it on the host would open a different file than the one the agent printed. Tests: `test/routes/folder-grid-routes.test.ts` (incl. the listing script on a real `sh`), `test/folder-grid-ui.test.ts`.
+
+**Text reuses the edit-mode allowlist; markup stays download-only.** `TEXT_ATTACHMENT_EXTENSIONS` IS `EDITABLE_EXTENSIONS` (`config/file-editing.ts`) rather than a second curated list that would drift from it: if the viewer would open a file for editing inside the workspace, the same file outside it can be read. The justification for widening is that the agent in the session can already `cat` any of these and the picker already previews them, so the suffix was never the confidentiality gate; the path guard is (sensitive-file blocklist, `/root` and `/etc` trees, realpath first). ⚠️ Two consequences had to be handled at the same time: `~/.codeman*/state.json` joined `isSensitivePath` (it persists `SessionState.envOverrides`, and the env allowlist admits key-shaped names like `GEMINI_API_KEY`, so it can hold a live credential), and `html`/`htm` joined `svg` in `serveRawFile`'s **download-only** branch so that widening what can be READ never widens what can RUN on our own origin. Text with no dedicated MIME entry goes out as inert `text/plain; charset=utf-8` + `nosniff`, matching the picker. ⚠️ **The one place HTML renders is `/html-view/<cap>/...`** (`html-view-routes.ts`, `html-view-capabilities.ts`), never these routes: `POST /api/sessions/:id/html-view` mints a capability for the HTML file's own directory, and every response carries `Content-Security-Policy: sandbox` WITHOUT `allow-same-origin`, so the page is opaque-origin even as a top-level tab and cannot read Codeman or drive its API. ⚠️ **Network access is open, by the owner's decision (2026-10-13, #547):** `HTML_VIEW_CSP` is the `sandbox` directive alone, which REPLACES Codeman's own CSP, with no fetch directive, so a page loads scripts, styles, fonts, images and media from anywhere and may fetch, XHR, open WebSockets, send beacons and post forms anywhere, like a file opened from disk. It can read its own folder through the capability and nothing else on this server; it cannot use Codeman's origin, cookies or API, or reach other folders. Never add a fetch directive or `allow-same-origin` without the owner; the exact header is pinned in `test/routes/html-view-routes.test.ts`. The directory is served recursively to the page's scripts (non-dot, asset-allowlisted, realpath-confined, blocklist applied, GET/HEAD only; the dot and extension rules run on the requested name AND on the symlink-resolved target, or a `cfg.json -> .git/config` link would serve the git config), which is why minting is refused for an HTML file in a hidden directory (for a file inside the workspace, a hidden directory below the workspace root, so a workspace that itself lives under one still renders) or directly in a broad root (`/`, home, `/tmp`, `/var/tmp` and `$TMPDIR` (named literally, since macOS sets `TMPDIR` in every terminal), the cases and user-space roots, a drive's mount point such as WSL's `/mnt/d`) or in any directory above one of them (`/home`, `/Users`, macOS's `/private`; `isBroadHtmlViewRoot`). A capability minted for a folder outside the workspace records that, and the serving route refuses it per request once `attachmentConfineToWorkspace` is on, the same re-check the attachment serve path makes. The preview renders it by default; the **Page** pill (`htmlRendered`, per-device like `mdRendered`) and any refused mint fall through to the source view with its Edit button. Tests: `test/routes/html-view-routes.test.ts`, the html-view block in `test/webview-auth-exemption.test.ts`, `test/webview-capability-revocation.test.ts`. The by-id text preview is bounded like the workspace one: a `Range` request for the first 512KB (a real partial read, not a discarded 50MB download) plus a 500-line cap, with the footer saying so.
 
 **Media is single-sourced across the two preview paths.** `VIDEO_ATTACHMENT_EXTENSIONS` / `AUDIO_ATTACHMENT_EXTENSIONS` live in `attachment-registry.ts` and are imported by `file-content`'s media classification, so a clip plays identically whether it is in the workspace or reached by id from outside it. They diverged first: the workspace path had its own inline sets and the registry allowlist had no media at all, so a video an agent wrote to `/tmp` was refused as an unsupported type while the same file inside the repo played. ⚠️ Three things have to line up for a player rather than a dead frame: the extension in the allowlist, a **real MIME entry** in `MIME_TYPES` (a `<video>` refuses to decode `application/octet-stream`, which presents as a player that renders and then does nothing), and the range-aware body (`serveRawFile` → `sendFileBody`) that makes the scrub bar work. `getAttachmentType()` returns the `video`/`audio` members of `AttachmentDetectedType` for them; the attachment card has no per-type CSS and its thumbnail falls back to the type label, since `generateFirstPageThumbnail` has no media branch and answers 204. ⚠️ The image-watcher keeps its OWN narrow detection list (`png/pdf/docx/pptx`), so this does not start popping cards for every video an agent writes.
 
@@ -385,6 +404,8 @@ A file path an agent prints is a link on both surfaces it can appear on, and cli
 
 **XLSX previews parse in the browser worker, and ExcelJS only ever sees what admission checked.** An `.xlsx` joins the raw routes like any other file (`?preview=true` caps it at 10 MB with a 413); the server never parses it. `spreadsheet-preview.js` hands the bytes to `spreadsheet-preview-worker.js`, the ONLY place the pinned `exceljs`/`fflate` vendor bundles load: fflate and `spreadsheet-xlsx-core.js` at worker start, ExcelJS only after `admitXlsx()` passes, and the page itself never loads either. `admitXlsx()` streams every entry through fflate and enforces the entry count, per-entry and total inflated bytes (64 MB), compression ratio, worksheet, cell, merge and style caps on the actual inflated output, refusing (never truncating) a workbook that trips one. ⚠️ Admission walks LOCAL headers while ExcelJS (JSZip) reads the CENTRAL directory, so overlapping entries (a stored entry hiding a whole `sheet1.xml`, a one-cell decoy later in the stream) once let a file be admitted as 1 cell and parsed as 300k. The worker therefore never hands ExcelJS the fetched bytes: it gets `buildAdmittedArchive()`, a STORE-only `fflate.zipSync(entries, { level: 0 })` of exactly the entries admission inflated, and a name streamed twice is refused. That costs one transient copy of the inflated entries (bounded by the same 64 MB cap) and is pinned by the overlapping-entry fixture in `test/spreadsheet-preview-worker.test.ts`. ⚠️ Admission also has to bound what ExcelJS EXPANDS, not just what it reads: ExcelJS 4.4.0 builds one object per covered cell of a `<mergeCell>`, per address of a `<dataValidation sqref>` and per index up to `<col max>`, so a 6.5 KB file admitted as one cell once cost a gigabyte. The counter therefore charges each merge its full AREA against the cell caps (an unparseable `ref` is refused), refuses a `<col>` whose `min`/`max` is past 16384, and the worker loads with `ignoreNodes: ['dataValidations']` (the preview never shows validations). Defined names expand the same way and live in `xl/workbook.xml`, where admission reads only the `<sheet>` ids: ExcelJS's `DefinedNames` model setter builds one object per cell of every range (a whole-sheet name exhausted a 4 GB heap), so right after `new ExcelJS.Workbook()` the worker replaces `_definedNames.model` with an `Object.defineProperty` stub whose getter returns `[]` and whose setter drops the value. Print areas and titles are split off in the workbook xform's reconcile, before that setter runs, so they are unaffected, and `defineProperty` throws if a future ExcelJS renames `_definedNames` rather than silently expanding again. ⚠️ Every name admission uses is the one ExcelJS will SEE, not the stored one: JSZip resolves each entry name on load (`utils.resolve`: `.` and empty middle segments dropped, `..` pops a segment) and ExcelJS then strips one leading `/` and matches worksheets with an UNANCHORED `xl/worksheets/sheet<N>.xml`. Checking the stored name once let `/xl/worksheets/sheet1.xml`, `xl/./...`, `xl//...`, `xl/xl/worksheets/sheet1.xml` and `xl/worksheets/sheet1.xml.x` skip every counter. `excelJsEntryName()` mirrors both steps; admission refuses two entries that resolve to one name, keys `inflatedEntries` (so the rebuilt archive) on the resolved name, picks the worksheet and styles counters from it, and counts any name matching ExcelJS's unanchored worksheet pattern as a worksheet. The local-versus-central consistency checks still compare the stored names. The counter reads `<mergeCell>`/`<col>` attributes IN ORDER from the tag name with a sticky regex that consumes each quoted value whole (a raw `>` or the other quote character is legal inside a value, so a first-match search could be fed a fake `ref`/`max`), and refuses a tag whose attributes do not parse up to `>` or repeat a name. It counts every `<row>` (ExcelJS keeps one Row object per element, cells or not) against per-sheet and total row caps, and reads each `<row>`'s attributes the same way. ⚠️ The INDEX a row or sheet claims is a cost too, not just how many there are: ExcelJS stores a row at `_rows[r - 1]` and walks `_rows` up to the largest index on every `eachRow` and `sheet.model` (five rows plus one `<row r="50000000">` took 5 s to load and 1.7 s per tile), and stores a sheet at `_worksheets[sheetId]`, which its `worksheets` getter slices and sorts (`sheetId="30000000"` on a one-cell workbook took 1.6 s and 557 MB). So a `<row r>` that is not plain digits in 1 to 1,048,576 is refused (an absent `r` is fine), and a counter for the resolved `xl/workbook.xml` reads every `<sheet>` tag in order and refuses one that does not parse or whose `sheetId` is not plain digits up to `LIMITS.maxSheetId` (65535; real ids are small, and the `(?=[\s/>])` lookahead keeps the `<sheets>` container out). The counter also counts every `<xf>` in styles.xml with no list-tracking state (a `</cellXfs>` inside a comment used to desync it). It carries everything from the last `<` into the next inflate chunk,Line truncated
 
+⚠️ **ExcelJS follows a worksheet's comment, comment-note (VML), table and drawing links UNCHECKED, so the worker repairs its parsed model before it links it.** ExcelJS 4.4.0's `reconcile` looks each relationship `Target` up verbatim in maps keyed by its own spelling (`../comments1.xml`, `../drawings/vmlDrawing1.vml`, `../tables/table1.xml`, `drawing1`) and dereferences the result. openpyxl, which most agents write spreadsheets with, uses absolute Targets, its own part names (`/xl/comments/comment1.xml`, which ExcelJS never even loads) and a drawing XML ExcelJS parses to nothing, so a workbook with ONE comment, table or chart threw (`reading 'comments'`, `'name'`, `'anchors'`) and the whole preview failed. Right after `new ExcelJS.Workbook()` the worker wraps `workbook.xlsx.reconcile` (which `load()` calls once on the parsed model) to run `repairExcelJsModel()` (core) first: it re-points a link to the key ExcelJS stored for the same part (openpyxl's `/xl/tables/table1.xml` becomes `../tables/table1.xml`, so the table is kept), and drops a link, `<tablePart>`, `<drawing>` or `<hyperlink>` reference, or parsed-to-nothing drawing/table that would dereference nothing. It never adds anything, looks keys up as own properties only (a `__proto__` Target resolves to nothing), and the preview shows none of these parts. The wrap throws if a future ExcelJS renames `reconcile`. Real openpyxl 3.1.5 fixtures (`test/fixtures/xlsx/openpyxl-{comment,table,chart}.xlsx`) fail the worker tests without it.
+
 ### Filesystem path picker
 
 **Filesystem path picker** (Link Existing "Browse" button + the extended mobile keyboard's `📁 Path` key): a lazy one-directory-at-a-time browser over `GET /api/filesystem/browse`, with `GET /api/filesystem/preview` serving the tapped file. It starts at the active session's working directory (falling back to the Codeman Cases root, then `/mnt/d`, then the first root), hides dot entries, and inserts the chosen path **without** Enter so the prompt is not submitted. The companion `⌫ All` key clears only the current unsent prompt buffer and must never emit the agent's `/clear` command.
@@ -420,7 +441,7 @@ Tests: `test/file-editing-policy.test.ts` (pure policy), `test/routes/file-write
 - ⚠️ **`data-i18n-skip` on the container.** The translator's MutationObserver translates inserted headings and paragraphs, and the `.file-preview-content` entry in its skip list matches nothing (no element has that class), so the attribute is what keeps a Chinese UI from rewriting a README.
 - ⚠️ **Toggles are per-device, in their own localStorage keys** (`codeman:filePreviewMdRendered` / `LineNumbers` / `Wrap`), for the same reason as the Files panel's show-hidden toggle: the app-settings object is rebuilt from the settings modal on every save, and they are display state, not synced settings (`SettingsUpdateSchema` is `.strict()`). MD re-renders from the kept source (`filePreviewContent`, which is also what Copy copies) without a refetch; Lines/Wrap are class flips on the one `<pre>`, whose rules are scoped `.file-preview-body > pre.file-preview-text` so they never leak into the document's own code blocks. Lines are one inline `<span class="fp-line">` per line joined by real newlines, the counter in `::before` with `user-select: none`, so select and copy return the exact text.
 - ⚠️ **Caps.** Markdown fetches `lines=10000` (the route's `MAX_LINES_LIMIT`), because a rendered document cut at 500 lines reads as the whole document; other text keeps 500, which is what stops a huge log locking the tab in one `<pre>`. The attachment (out-of-workspace) branch keeps its 512 KB Range read and skips the 500-line clip for markdown. Edit mode is unchanged and still re-fetches `edit=1`; the three toggles hide while editing and for images, media and PDFs.
-- ⚠️ **`md` is NOT in `FILE_PREVIEW_EXTENSIONS`.** Only an in-workspace `.md` path clicked in the terminal still opens the tail viewer (see File-path links above: in-workspace text keeps live follow, which is what Ralph's `fix_plan.md` needs). Everything else reaches `openFilePreview()` and renders it: the Files panel, a path clicked in the Response Viewer (its delegate calls `openFilePreview()` directly), an out-of-workspace terminal path, and attachment cards. `avif` and `ico` were added to the set (a printed `favicon.ico` used to tail binary noise), with `avif` also in file-content's image set and file-raw's MIME map; out-of-workspace avif/ico stay unregistrable, like svg/bmp.
+- ⚠️ **`md` and `markdown` ARE in `FILE_PREVIEW_EXTENSIONS`** (owner decision 2026-10-13). Until then an in-workspace `.md` path clicked in the terminal opened the tail viewer, which showed "streaming" over raw text instead of the document (the live follow it gave Ralph's `fix_plan.md` was not worth that). Now every surface reaches `openFilePreview()` and renders it: the terminal (the main pane, grid tiles and the split's Pane B share `registerFilePathLinkProvider`, and a touch tap activates the same provider), the Files panel, a path clicked in the Response Viewer (its delegate calls `openFilePreview()` directly), and attachment cards. `FILE_PATH_LINK_PATTERN` links `.markdown` (and `.htm`) as well, or such a path would not be clickable at all. Other in-workspace text (`.log`, `.txt`, `.json`, code) keeps the tail viewer. `avif` and `ico` were added to the set (a printed `favicon.ico` used to tail binary noise), with `avif` also in file-content's image set and file-raw's MIME map; out-of-workspace avif/ico stay unregistrable, like svg/bmp.
 
 Tests: `test/file-preview-markdown.test.ts` (jsdom-in-vm, pins every rule above), `test/routes/file-routes.test.ts` (avif).
 
@@ -441,6 +462,20 @@ Tests: `test/git-clone.test.ts` (pure half exhaustively, plus REAL git against a
 
 Further detail: the synchronous clone request is bounded by `GIT_CLONE_TIMEOUT_MS`.
 
+### Git worktrees of a case
+
+**Git worktrees** (case picker → **Git worktrees…**, the phone case sheet's **Worktrees**, Close Session's **Kill & remove worktree**): a LOCAL case whose folder is the top of a git repository keeps worktrees at `<repo>/.claude/worktrees/<name>` on a new branch (default `worktree-<name>`), so parallel agents each get their own checkout instead of sharing the case folder. That is Claude Code's own `--worktree` layout (measured in 2.1.296: same path, branch `worktree-<name>`, names of `[A-Za-z0-9._-]`, at most 64), and Codeman's names are a strict subset of Claude's. Core in `src/git-worktrees.ts` (pure name/branch gates, `worktree list --porcelain` parsing, exclude text; IO behind an injectable runner); routes in `routes/worktree-routes.ts` (`GET`/`POST /api/cases/:name/worktrees`, `DELETE /api/cases/:name/worktrees/:worktree`).
+
+- ⚠️ **Only `<repo>/.claude/worktrees/<valid-name>` is ever created or removed.** Every component is `lstat`ed (a symlinked `.claude`, `worktrees` or worktree folder is refused, never followed) and the real path must equal the lexical one, so a repository cannot redirect a create or remove outside itself; removal also requires git to list the folder as a worktree of THIS repository. A worktree anywhere else is listed read-only (`managed: false`). A managed worktree whose folder is GONE stays removable (`resolveManagedWorktree({ allowMissing })`, removal only, never a session start: git lists it at exactly the managed path, no component on the way is a symlink, and `git worktree remove` then only drops git's record), since otherwise its name could never be used again.
+- ⚠️ **git never runs for a case at or inside a Docker case workspace** (or whose common git dir is there), and Docker and remote cases are refused with a 400 before any git, the same rule as the git status indicator: `git worktree add` runs the repository's post-checkout hook and checkout filters ON THE HOST. That is acceptable for a local case, where the user's agent already runs code. A managed worktree that is, sits inside, or holds a Docker case workspace (a Docker case linked to the worktree folder) is never removed (400): `git worktree remove` runs a status inside the tree and then deletes it, ignored files included, under the container.
+- ⚠️ **Removal is `git worktree remove` WITHOUT `--force`**, refused (409) while any live session (any owner) runs at or inside the worktree, and git's own refusal (modified or untracked files, a lock) is the answer, shown through `textContent`. The optional branch delete is `git branch -d`, never `-D`: an unmerged branch survives and the response says why (`branchError`).
+- ⚠️ **The main checkout must not see the worktrees.** When `git check-ignore` says they are not ignored, `/.claude/worktrees/` is appended to the repository's LOCAL `<common git dir>/info/exclude` (never a tracked `.gitignore`), and so is `/.claude/settings.local.json`: Codeman writes its hooks there for every claude session, and an untracked copy in a worktree would make `git worktree remove` refuse forever (Claude Code covers that file with the global excludes file, which a machine may not have). The agent skill copy (`<skills dir>/codeman/`, injected while `agentSkillEnabled` is on) gets a line too, for every skills folder the registry names (`capabilities.agentSkillDir`, passed in by the route as data). A symlinked `info` dir or `exclude` file is reported, never written through.
+- **Names are gated before git**: charset and length (Zod and `worktreeNameError()`), then `git check-ref-format --branch`, whose output must EQUAL the input (`--branch` expands `@{-1}`, which the charset gate already refuses). The base is a resolved commit (`rev-parse --verify <base>^{commit}`), passed after `--` with `--no-track`; an existing branch is refused (`-b`, never `-B`, which would reset it). Create and remove are serialized per repository.
+- **Sessions start through the EXISTING create paths**, never a new spawn path: `POST /api/quick-start` with `worktree` (resolved with the same guards; hooks, the agent skill, the model override and the session all target the worktree, and a case that does not exist is a 404, never a scaffold), or `POST /api/sessions` with the worktree as `workingDir` (what the Run button uses for claude and shell). The frontend launches through `run({ worktree })` with a temporary `_runMode` swap, tabs named `w<n>-<worktree>`.
+- The unified session list names a live session's worktree from its folder layout (`worktreeFromPath`), so a codex or shell session gets the worktree pill before (or without) a transcript record.
+
+Tests: `test/git-worktrees.test.ts` (pure half, then REAL git: create, exclude, symlink guards, dirty/live/locked refusals, `branch -d`), `test/routes/worktree-routes.test.ts` (real repository through the routes and quick-start, Docker/remote/ownership refusals), `test/worktree-ui.test.ts`, `test/worktree-modal-i18n.test.ts`.
+
 ### Ultracode and workflow-run visualization
 
 **Ultracode / Workflow-run visualization** (opt-in `showUltracodeAgents`, default OFF; released 1.1.2): the Workflow tool ("ultracode") writes a COMPLETION artifact per run at `~/.claude/projects/<projHash>/<sessionUuid>/workflows/wf_*.json` (written only at run end); LIVE in-flight runs exist only as transcript dirs at `…/subagents/workflows/wf_<id>/` (journal.jsonl + `agent-*.jsonl`). `workflow-run-watcher.ts` (STANDALONE — deliberately never imports/touches `subagent-watcher.ts`; separate singleton, though it independently reads the same `subagents/workflows/` tree) scans BOTH sources via periodic poll + per-directory chokidar watchers with per-source mtime skip (LRU agentStatCache + journalCache), synthesizing ACTIVE runs (live per-agent tokens/tools/state from transcripts, title/phases from the workflow script) until the completion `wf_*.json` appears and supersedes, and broadcasts SSE `workflow:run_discovered`/`run_updated`/`run_removed`. The watcher is started when **either** `showUltracodeAgents` **or** `ultracodeFloatingWindows` is on (`server.ts` `isWorkflowAgentTrackingEnabled()` returns `(showUltracodeAgents ?? false) || (ultracodeFloatingWindows ?? false)`). Served via `GET /api/workflows` (optional `?minutes=` filter) and `GET /api/workflows/:runId`. Frontend `ultracode-panel.js` renders a docked master-detail view (LEFT: runs + phases; RIGHT: per-agent tokens + tool-calls; click an agent card → its live transcript via client-side `agentId` join). **Additionally**, `ultracode-windows.js` auto-pops a draggable **floating window per active run** (gated on a **DEDICATED** `ultracodeFloatingWindows` toggle, default OFF — independent of the dock panel's `showUltracodeAgents`; see `_ultracodeFloatingEnabled()`), connected by a glowing line to the originating session tab (resolved by `session.claudeSessionId === run.sessionUuid`) — same line idiom as subagent windows, drawn into the shared `#connectionLines` SVG from the tail of `_updateConnectionLinesImmediate`. The window auto-closes ~8s after its run finishes; explicit dismissals are remembered. Clicking an agent card opens an **in-page** connected transcript window (not a browser popup); both run and transcript windows minimize **into** the originating session tab as a merged `ULTRA` badge (🧬 runs / 📄 transcripts) with a restore/dismiss dropdown — minimized runs are skipped by auto-pop. Gesture beta: floating subagent/ultracode windows are pinch-draggable (a `window` grab kind in `entry.ts`). Types: `src/types/workflow-run.ts`. Config: `src/config/workflow-config.ts`.
@@ -608,7 +643,7 @@ Either way the actual launch (`runCustomModelEntry`) routes through `run()` itse
 
 ### Approvals Inbox
 
-**Approvals Inbox** (cross-session queue of prompts waiting on a human; `approvalsInboxEnabled`, SYNCED, default OFF: every surface is opt-in; only the store and answer endpoints run regardless, so flipping it ON shows anything already pending): `web/approval-inbox.ts` is a `sessionWaits`-style singleton fed by `/api/hook-event`, holding at most ONE item per session (a new prompt supersedes), claude-mode only, in-memory. Cards are answered via `POST /api/approvals/:id/answer`, which sends a digit / Esc / idle-prompt text through `writeViaMux` (menu answers never carry `\r`).
+**Approvals Inbox** (cross-session queue of prompts waiting on a human; `approvalsInboxEnabled`, SYNCED, default OFF: every surface is opt-in; only the store and answer endpoints run regardless, so flipping it ON shows anything already pending): `web/approval-inbox.ts` is a `sessionWaits`-style singleton fed by `/api/hook-event`, holding at most ONE item per session (a new prompt supersedes), in-memory; answerable items are claude-mode only, and the DeepSeek status bridge and codex's screen-read dialogs (`web/screen-blocked.ts`) open ALERT-ONLY items (`answerInTerminal`: the answer route refuses them BEFORE its hooks gate, push carries no `approvalId`, the card and overview strip say "answer in the terminal"). Cards are answered via `POST /api/approvals/:id/answer`, which sends a digit / Esc / idle-prompt text through `writeViaMux` (menu answers never carry `\r`).
 
 ⚠️ `option` digits are accepted ONLY when they match options parsed from the captured pane frame, and the answer path RE-CAPTURES the pane first (a dialog that no longer parses on screen means the keystroke would land in the composer, so refuse with 409).
 
@@ -662,7 +697,7 @@ Matching semantics: a query containing `/` matches the relative path, otherwise
 
 ### Raw file bodies: streamed and range-aware
 
-**Raw file bodies are streamed and range-aware**: `file-raw`, the attachments `/raw` route and `GET /api/download` always advertise `Accept-Ranges: bytes` and answer a `Range` header with `206` + `Content-Range` (single-range only; parser is pure + unit-tested in `src/web/http-range.ts`, a malformed spec is ignored → 200 while an out-of-bounds one is a 416).
+**Raw file bodies are streamed and range-aware**: `file-raw`, the attachments `/raw` route, `GET /api/download` and the rendered-HTML route `GET /html-view/<cap>/...` (so a rendered page's own `<video>`/`<audio>` can seek) always advertise `Accept-Ranges: bytes` and answer a `Range` header with `206` + `Content-Range` (single-range only; parser is pure + unit-tested in `src/web/http-range.ts`, a malformed spec is ignored → 200 while an out-of-bounds one is a 416).
 
 ⚠️ **The size cap on all three is a sanity bound, not memory protection** (`MAX_FILE_DOWNLOAD_BYTES` in `config/buffer-limits.ts`, default 2GB, env `CODEMAN_MAX_DOWNLOAD_BYTES`, `0` = unlimited): the bodies stream, so size costs a read stream and not RSS (measured: a 600MB download moved peak RSS by ~37MB). Its predecessor was a hardcoded 50MB whose comment still said "prevent memory exhaustion" long after the `readFile()` it described was replaced by `sendFileBody()`, so all it did was refuse legitimate downloads of build artifacts, videos and archives. `/api/download` was the last route that really did buffer the whole file, and now shares `sendFileBody()` with the other two.
 
@@ -676,7 +711,7 @@ Matching semantics: a query containing `/` matches the relative path, otherwise
 
 ### Command palette and shortcut registry
 
-**Command palette + shortcut registry** (COD-151/153/157/192, #146): `Ctrl/Cmd/Alt+K` opens the session palette (fuzzy search over live sessions; "Browse all sessions" → the Session Manager modal backed by `GET /api/sessions/unified`); the quick-start case `<select>` is fronted by a searchable picker (`buildCasePickerOptions`/`formatCasePickerLabel` — remote cases render `name @ hostId`). Shortcuts live in a rebindable registry (`DEFAULT_SHORTCUTS`/`getShortcutRegistry()`/`matchesShortcutEvent()` in app.js; overrides persist under `settings.shortcutOverrides` via `saveAppSettingsToStorage`); App Settings → Shortcuts renders capture/disable rows; `Ctrl+?` opens the registry-driven overlay (footer links to the full `#helpModal` reference). ⚠️ Palette-chord keys must ALSO be swallowed in `attachCustomKeyEventHandler` (terminal-ui.js) or xterm writes the control byte (0x0B) into the PTY. ⚠️ The global Escape handler (app.js) calls every overlay close method on every Escape, in the capture phase and so before the focused element's own Escape handler, which means a close method with side effects beyond hiding must return early when its overlay is not open: `closeCommandPalette()` and `closeSessionManager()` hand focus back on close (`_restoreOverlayFocus()`, panels-ui.js), and run unconditionally that restore stole focus from split Pane B and from any text field, and turned the inline tab rename's Escape into a commit. ⚠️ `saveAppSettings()` rebuilds settings from the DOM — keys edited elsewhere (`shortcutOverrides`, `showTokenCount`, `showCost`) need explicit `_prev` carry-over.
+**Command palette + shortcut registry** (COD-151/153/157/192, #146): `Ctrl/Cmd/Alt+K` opens the session palette (fuzzy search over live sessions; "Browse all sessions" → the Session Manager modal backed by `GET /api/sessions/unified`); the quick-start case `<select>` is fronted by a searchable picker (`buildCasePickerOptions`/`formatCasePickerLabel` — remote cases render `name @ hostId`). Shortcuts live in a rebindable registry (`DEFAULT_SHORTCUTS`/`getShortcutRegistry()`/`matchesShortcutEvent()` in app.js; overrides persist under `settings.shortcutOverrides` via `saveAppSettingsToStorage`); App Settings → Shortcuts renders capture/disable rows; `Ctrl+?` opens the registry-driven overlay (footer links to the full `#helpModal` reference). ⚠️ Palette-chord keys must ALSO be swallowed in `attachCustomKeyEventHandler` (terminal-ui.js) or xterm writes the control byte (0x0B) into the PTY. ⚠️ The global Escape handler (app.js) calls every overlay close method on every Escape, in the capture phase and so before the focused element's own Escape handler, which means a close method with side effects beyond hiding must return early when its overlay is not open: `closeCommandPalette()` and `closeSessionManager()` hand focus back on close (`_restoreOverlayFocus()`, panels-ui.js), and run unconditionally that restore stole focus from split Pane B and from any text field, and turned the inline tab rename's Escape into a commit. ⚠️ `saveAppSettings()` rebuilds settings from the DOM — keys edited elsewhere (`shortcutOverrides`, `showTokenCount`, `showCost`) need explicit `_prev` carry-over. ⚠️ Every xterm key handler (Pane A in terminal-ui.js, Pane B in terminal-tile.js) must call `isUserBoundShortcutEvent(ev)`, which is true only for a chord the user REBOUND a generically dispatched action to: the document capture handler's `preventDefault()` does not stop xterm, so a rebind to e.g. Alt+J would otherwise also type `ESC j` into the PTY. Defaults are deliberately not gated (Ctrl+L must reach the CLI for its redraw), so the gate also skips an entry re-captured onto its own default, `toggleSessionSidebar` (it has its own gate, `shouldToggleSessionSidebarFromShortcut`) and copy-selection. Capturing, re-enabling and Reset all go through `findShortcutConflict`/`findShortcutBindingConflict` (settings-ui.js); Alt+1-9, Ctrl/Shift+Enter and Ctrl/Cmd+V are reserved there.
 
 **A key gate on a chord that can carry Shift alone must swallow the KEYPRESS too** (#520). xterm calls the custom handler from `attachCustomKeyEventHandler` for `keydown`, `keypress` AND `keyup`, and its own `_keyPress` drops a keypress carrying Ctrl, Alt or Meta but not one whose only modifier is Shift. So a gate that returns `false` for `ev.type === 'keydown'` alone is safe for a Ctrl or Alt chord (the palette's Ctrl+K, Alt+B, Ctrl+V, Ctrl+Z and the keydown-only smart copy below) and leaks for a Shift-only one. That is how Shift+Enter submitted the prompt from April to #520: its keypress fell through to `onData` as a bare `\r`, which goes out synchronously over the WebSocket ahead of the async `send-key` fetch, so the CR reached the PTY before the newline did. Ctrl+Enter worked all along only because Chrome fires no keypress for it. ⚠️ The Shift/Ctrl+Enter gate in terminal-ui.js and its twin in Pane B (terminal-split.js) return `false` for EVERY event type and key only the `send-key` POST on `keydown`; never narrow either back to keydown, and give any new Shift-only chord the same shape. Swallowing the keyup as well is harmless (xterm only skips a focus and cursor-style refresh for that one event). Tests: `test/shift-enter-keypress-swallowed.test.ts` (CI gate, a source guard on both files) and `test/shift-enter-keypress.browser.test.ts` (browser suite, which the gate does not run: real keys against both real handlers in the real app). Settings → Terminal & Input → **Key tester** shows what a device reports for each of the three events.
 
@@ -732,7 +767,7 @@ The guard is a one-shot flag on each trap rather than a browser test or a readin
 
 **One visual language, three modals.** `#appSettingsModal`, `#sessionOptionsModal` and `#createCaseModal` share the `set-*` surface (left rail, sections of grouped row cards, label + description on the left, control pinned right) through a single `:is(#appSettingsModal, #sessionOptionsModal, #createCaseModal)` scope in `styles.css`. An `:is()` list takes the specificity of its **most specific argument**, and all three arguments are ids, so every rule kept exactly the weight it had when the block was `#appSettingsModal`-only: nothing downstream shifted in the cascade. That property is what let the surface absorb Session Options and then Add Case in two separate commits without a cascade audit each time.
 
-Anatomy: `.set-shell` → `.set-shell-head` (title + `.set-head-actions`) + `.set-body` (grid of `.set-rail` and `.set-doc`). The rail holds `.set-search`, `.set-rail-items` and `.set-rail-foot`; the document holds `.set-section`s, each with `.set-section-head` + `.set-section-blurb` + `.set-group`s of `.set-row`s (`.set-row-text` = `.set-row-label` + `.set-row-desc`; controls sit in `.set-row-actions`). ⚠️ Close is deliberately **first in the DOM** inside `.set-head-actions` so the focus trap lands on it rather than on Save; `row-reverse` is what puts Save to its left visually on phones.
+Anatomy: `.set-shell` → `.set-shell-head` (title + `.set-head-actions`) + `.set-body` (grid of `.set-rail` and `.set-doc`). The rail holds `.set-search`, `.set-rail-items` and `.set-rail-foot`; the document holds `.set-section`s, each with `.set-section-head` + `.set-section-blurb` + `.set-group`s of `.set-row`s (`.set-row-text` = `.set-row-label` + `.set-row-desc`; controls sit in `.set-row-actions`). ⚠️ Close is deliberately **first in the DOM** inside `.set-head-actions` so the focus trap lands on it rather than on Save; `row-reverse` is what puts Save to its left visually on phones. Apply (`applyAppSettings()`) sits beside Save in the same actions and saves without closing; it only sets a one-shot flag that `saveAppSettings()` consumes before its first await, so a Save clicked mid-Apply still closes the sheet.
 
 **The rail means two different things, and that is deliberate:**
 
@@ -797,6 +832,20 @@ Further detail: with many sessions the horizontal strip stops being scannable, w
 
 ⚠️ Leaving sidebar mode **clears `_sidebarFilter`**: the filter box only exists in the aside, so a stale filter would hide sessions from the header strip with no reachable control to clear it. ⚠️ On handhelds the closed drawer keeps `display: flex`, so without `inert` + `aria-hidden` (`_isSessionSidebarOverlay()`) its filter box and ~4 tab stops per session stay in the tab order; the DOCKED desktop rail must never be inerted, its rows are still clickable.
 
+⚠️ **The rail's Search sessions box and the sidebar's filter box are ONE row filter** (#580): `applySidebarFilter(query)` sets `_sidebarFilter`, the rail box (`setTabRailSearch`, `#tabRailSearch` at the top of `#tabRail`) sets `_tabRailSearch`, and both run `_applyTabListFilter()`, which reads the rows the last render drew and applies the pure `CodemanTabSearch` (constants.js) as classes. Only one needle is ever live, because the rail and the sidebar never host `#sessionTabs` at the same time (`applyTabOrientation()` forces horizontal while the sidebar owns the list); off the rail and out of a reachable sidebar the needle is empty and every class comes off. The rules that keep it a view filter:
+
+- **Classes only.** `tab-filtered-out` on a row, and on a `.tab-layout-group` or `.tab-cluster` the filter emptied; never a reorder, a re-render per keystroke, a layout write or a request, so `sessionOrder`, the Alt+N badges, the server order and the tab layout never move. Both render paths end in `applySidebarFilter()`, which re-applies it. The hide rules are layout-scoped like the sidebar's original one (`html[data-tab-orientation='vertical'] .tab-rail ...` for rows, groups and case boxes; `html[data-session-list="sidebar"] ...` for rows and case boxes), so a leaked class can never hide anything on the header strip, which has no box to clear it from.
+- **Matching.** The rail matches the NAME only (`.tab-name`'s `data-full-name`, a web tab's title, never a URL or folder); the sidebar keeps its name + working directory. Trimmed, case-insensitive substring, lower-cased with `toLowerCase()`, never `toLocaleLowerCase()` (a Turkish locale lowers "API" to "apı").
+- **A row with a tab alert is never hidden** (owner decision on #580, the collapsed-header promise again: a prompt waiting on you is never hidden by a view filter). The row goes into the core as `keep: true` when `tabAlerts` holds anything for it (red `action` or yellow `idle`); a kept row counts toward its section, so its group or case box stays on screen, but not toward `matchCount`, so "No sessions match" can show above a lone alerted row. Alerts come and go through `renderSessionTabs()`, whose tail re-runs the filter, so nothing else has to.
+- **Counts and the tree.** A section's count shows the rows left showing during a search and is restored from the `data-total` it stashed on first touch. In the grouped tree `_tabTreeItems()` skips a hidden group whole (header included), the filter tail re-runs `_applyTabTreePositions()` and moves the one roving stop off a hidden item, but only while a search hides something or right after one changed what shows (both render paths already set them over an unfiltered tree). Only the grouped rail is a tree; the flat rail stays a tablist.
+- **Collapsed groups.** A collapsed group's rows are not in the DOM, so a class cannot reveal them: the projection opens every group while the rail search runs (see Owner tab layouts), and `setTabRailSearch()` re-renders only when that changes the structure (`_isTabGroupStructureStale()`); every other keystroke only re-applies classes.
+- **Connectors.** Lineage lines and the subagent/ultracode connectors are anchored to row positions, and a keystroke re-renders nothing, so `_applyTabListFilter()` calls `updateConnectionLines()` whenever it actually showed or hid something (a row, a section, the state headings via `tabs-filtering`, the empty note). An unchanged re-apply at a render tail calls nothing, which keeps the incremental path's lineage gate meaningful. ⚠️ A hidden row still answers `getBoundingClientRect()` with an all-zero rect, which is truthy, so every floating window that anchors to its parent tab (the subagent and ultracode connectors' shared `tab:<id>` cache, both spawn positions, the ultracode genie) finds it through `_paintedSessionTab()` (app.js: the row only while `getClientRects()` is non-empty). An unpainted parent draws no connector, spawns where a window without a tab would, and tears down without the genie; otherwise all three started at the viewport's top-left corner. Lineage needs no guard, since `computeTree` drops zero-size rects.
+- **Escape.** The global key handler runs in the CAPTURE phase on `document`, before the box's inline `onkeydown`, so a `stopPropagation()` there comes too late: the global Escape branch claims an Escape whose target is `#tabRailSearch` while the box holds text (beside the group menu, the grouped-rail drag and the Tiles count menu) and routes it to `handleTabRailSearchKeydown()`. Without that, clearing the search also collapsed the Monitor and Subagents panels. An empty box leaves Escape to the global handler.
+- **Drag stays on while searching** (owner decision): searching for a tab and dragging it into a group is a useful flow. A drop uses the same anchor logic relative to the visible rows, so a row can land between visible neighbours with hidden rows on either side.
+- **In memory only, cleared off the rail.** The text is never persisted or sent. `applyTabOrientation()` calls `_resetTabRailSearch()` before its render when the list leaves the vertical rail, the same reason leaving sidebar mode clears `_sidebarFilter`.
+
+Tests: `test/tab-rail-search.test.ts` (gate) and `test/tab-rail-search.browser.test.ts` (browser suite, not in the gate; it installs the real `setupEventListeners()` so the capture-before-inline Escape order is the shipped one).
+
 ⚠️ **The detailed RAIL additionally wears the home rail's CARD, and the detailed SIDEBAR deliberately does not**: the rail is an occasional, resizable list you scan, the sidebar is a permanently-docked nav column where 20 stacked cards read as a wall — so the card rules are RAIL-SCOPED and were NOT added to the comma-grouped selectors above, which is the one-line change that would silently restyle the sidebar. Card state accents reuse `home-sessions-blink-red`/`-yellow` rather than declaring a second pair, and every state dot rule excludes `.tab-alert-action`/`.tab-alert-idle` by hand, because those alert rules are only (0,3,0) and the rail's are (0,5,1)+. There is no `.active` rule in that block on purpose: `.session-tab.active` already paints border/background/box-shadow `!important`.
 
 ⚠️ **Row ORDER is a third rail attribute** (`tabRailSort: 'activity'|'manual'`, `data-tab-rail-sort`, default **activity**): a sorted rail answers the home screens' question with the home screens' answer, `CodemanSessionOrder` over rows classified by `_mobileOverviewState` (`_tabRailSortOrder` in app.js). ⚠️ **It is applied as the flex `order` property, never by reordering the DOM**: `#sessionTabs` stays in `sessionOrder`, so the Alt+N badge (`_tabIdx`, which therefore does NOT run 1,2,3 down a sorted rail — it names a shortcut, not a position), drag-and-drop, the arrow-key walk, the sidebar filter and `_scrollActiveTabIntoView()` all keep reading the list they always read, and a session changing state moves ONE inline style instead of forcing the full rebuild that would restart every card's animation on every SSE tick. The incremental render path therefore has to re-apply it (a state flip adds no tab, so the full rebuild is never reached) and an empty string is what clears it when sorting stops. ⚠️ Web tabs are pinned past the cards by a CSS `order: 9999` rather than an inline one, since `renderWebviewTabs()` emits the same markup for every layout; the flex default of 0 would interleave them among the sorted sessions. ⚠️ `setupTabDragHandlers()` sets `draggable="false"` and returns while sorting is on: the drop rewrites `sessionOrder` correctly and the sort then puts the card straight back, so the affordance would be a lie — `'manual'` is the way back to drag-reordering.
@@ -809,17 +858,17 @@ Further detail: with many sessions the horizontal strip stops being scannable, w
 
 ### Tile grid
 
-**Tile grid** (`showTileGridButton`, header Tiles button + `Ctrl+Shift+G`, default ON on desktop and OFF on handhelds and touch-primary tablets (primary pointer coarse, resolved in `getDefaultSettings()`, never `isTouchDevice()`, so a touchscreen laptop keeps ON), per-device: in `displayKeys`, stripped from the settings PUT, not in `SettingsUpdateSchema`; desktop-only at 1180px by a JS width check with a live media listener plus a `@media (max-width: 1179px)` backstop, never in a solo window). 1 to 6 live sessions side by side in one window, each a `TerminalTile` (terminal-tile.js), orchestrated by tile-grid.js (load order 7.6, `CodemanApp.prototype` methods like the split's). Pure helpers live in constants.js as `window.CodemanTileGrid`: `computeTileLayout` (1x1, 2x1, 3x1 on a grid area at least 1800px wide else 2x2, 2x2, 3x2, and 3x3 up to `TILE_LAYOUT_MAX` (9), unreachable today; `fits` against a 480x240 minimum tile), `tileGridCapacity` (never more than `TILE_GRID_MAX`), `sanitizeTileGridState` (truncates a stored grid to the cap, keeps focus only if it survives), `buildTilePickerSessions`, `dragTrackFractions`, `tileNeighbor`, `tileInDirection`, `cycleTile`, `TILE_SCROLLBACK` (10,000), and `TILE_GRID_MAX` (6), the ONE cap (owner decision 7: six tested smooth on a real desktop, nine missed the headless frame bar). ⚠️ Every limit reads the cap through `_tileGridLimit()` (tile-grid.js: the window's capacity, at most the cap), never a literal, and its texts say which binds ("at most 6 tiles" vs "what this window fits"). The grid is a `<section class="tile-grid">` SIBLING of `.terminal-wrap`, swapped in by `.main.tiles-active` (no reparenting); `.main.webview-active .tile-grid` hides it like the single view.
+**Tile grid** (`showTileGridButton`, header Tiles button + `Ctrl+Shift+G`, default ON on desktop and OFF on handhelds and touch-primary tablets (primary pointer coarse, resolved in `getDefaultSettings()`, never `isTouchDevice()`, so a touchscreen laptop keeps ON), per-device: in `displayKeys`, stripped from the settings PUT, not in `SettingsUpdateSchema`; desktop-only at 1180px by a JS width check with a live media listener plus a `@media (max-width: 1179px)` backstop, never in a solo window). 1 to 6 live sessions side by side in one window, each a `TerminalTile` (terminal-tile.js), orchestrated by tile-grid.js (load order 7.6, `CodemanApp.prototype` methods like the split's). Pure helpers live in constants.js as `window.CodemanTileGrid`: `computeTileLayout` (1x1, 2x1, 3x1 on a grid area at least 1800px wide else 2x2, 2x2, 3x2, and 3x3 up to `TILE_LAYOUT_MAX` (9), unreachable today; `fits` against a 480x240 minimum tile), `tileGridCapacity` (never more than `TILE_GRID_MAX`), `sanitizeTileGridState` (truncates a stored grid to the cap, keeps focus only if it survives, reports the cells whose sessions went away as `freed`), `restoreTileGridCells` (a stored grid as it comes back), `rankTileSessions` (the order a fresh or filling grid takes sessions in), `buildTilePickerSessions`, `dragTrackFractions`, `tileNeighbor`, `tileInDirection`, `cycleTile`, `TILE_SCROLLBACK` (10,000), and `TILE_GRID_MAX` (6), the ONE cap (owner decision 7: six tested smooth on a real desktop, nine missed the headless frame bar). ⚠️ Every limit reads the cap through `_tileGridLimit()` (tile-grid.js: the window's capacity, at most the cap), never a literal, and its texts say which binds ("at most 6 tiles" vs "what this window fits"). The grid is a `<section class="tile-grid">` SIBLING of `.terminal-wrap`, swapped in by `.main.tiles-active` (no reparenting); `.main.webview-active .tile-grid` hides it like the single view.
 
 ⚠️ **The main terminal is parked while the grid is open.** Opening runs `_cleanupPreviousSession()` ONCE (its snapshot is right at that moment, and it closes the main socket), and `activeSessionId` is always the FOCUSED tile's session, so everything keyed on it (files panel, git status, respawn and Ralph panels, subagent windows, voice, image paste, the tab highlight) follows focus. With the main socket closed `_wsReady` is false, so every SSE terminal handler would write the focused tile's output into the hidden xterm: `_tilesOwnTerminal()` turns `_onSessionTerminal`, `_onSessionClearTerminal`, `_onSessionNeedsRefresh` (returns `false`, which the drop recovery reads), `_scheduleDroppedOutputRecovery`, the `writeln` in `_onSessionCompletion`/`_onSessionError`, `sendResize`, `throttledResize`'s fit and `_maybeRefetchFullHistory` into no-ops; `retryConnection` and `handleInit` re-arm the TILES' sockets instead of the main one (`handleInit` keeps live tiles, drops dead ones through `_reconcileTileGrid`, and re-selects only when the focused tile is gone, so an SSE blip never hides an open web tab). ⚠️ The page's SSE filter names `TILE_GRID_SSE_FILTER` (constants.js, an id no session takes) instead of the focused session while tiles own the terminal: the server's filter gates only `session:terminal` batches, which the parked terminal could only parse and drop. Both places that set the filter ask `_sseFilterSessionId()`: the live re-subscribe every tile focus runs (`_updateSseSubscription`) and the connect URL an SSE reconnect rebuilds (`connectSSE`); leaving the grid re-subscribes the shown session through `selectSession`. `test/sse-tile-grid-filter.test.ts` pins the server's side in multi-user mode (the id is accepted on connect and on re-subscribe, and session/hook events still reach their owner through it). ⚠️ The WebGL long-task observer watches the WHOLE page: it counts nothing while tiles own the terminal, or tile renders would write the sticky 7-day WebGL disable. The header connection dot reads the tile sockets (`_tileGridSocketState`; a tile stopped for good does not count). `_focusedPane()` answers with the focused tile even when DOM focus left every terminal, and `_forEachTile` reaches every grid tile (`{ grid: false }` skips them where tiles keep their own font size). Leaving the grid destroys every tile, resets `_lastResizeDims`, invalidates the main terminal's cached content for EVERY tiled id (`_xtermSnapshots`, `codeman-xs-<id>`, `terminalBufferCache`: written before the grid opened, and selectSession paints a snapshot as its first frame) and replays the focused session through `selectSession(id, { forceReload: true, auto: true })`.
 
-⚠️ **One load queue.** Every capture a tile fetches (initial load, refresh after a reconnect, server `{t:'r'}` refresh, server `{t:'c'}` refresh, dropped-output recovery refresh, shell history pull) goes through the grid's ONE `TileLoadQueue` (terminal-tile.js, the tile's `scheduleLoad` option): concurrency 1, a history pull first, then the focused tile, then reading order; a destroyed tile's waiting loads are dropped unrun and destroy() aborts its running fetch. A load holds the slot only while xterm parses its replay: `writeChunked` queues every 32 KB slice at once (one 1 MiB window at a time, since xterm's write queue throws past 50 MB) and settles on the callback of a write queued behind them, never one slice per animation frame (that pacing held the slot about a second per 1 MiB tile). ⚠️ destroy() settles a replay in progress (`_cancelReplay`): a disposed xterm never runs that callback, and an unsettled replay would hold the tile's flag and the grid's queue forever. Each `GET /api/sessions/:id/terminal` runs synchronous tmux calls, so N tiles loading at once (after a deploy restart all N reopen within a second) would stall every WebSocket and SSE stream on the server back to back. Grid tiles load the BOUNDED window (`full=1&tail=TERMINAL_TAIL_SIZE` for a TUI, `tail=` for a shell; the `boundedLoad` option), and every full capture of theirs (a TUI load, a shell history pull) also sends `lines=<TILE_SCROLLBACK + rows>`, so tmux reads no more history than the tile keeps instead of the whole history limit (the route's optional `lines` bound, docs/api-reference.md; the split's Pane B sends none), keep `TILE_SCROLLBACK` lines and their own per-device font (`codeman-tile-font-size`, Ctrl +/- while the grid is open). A refresh fetches at its turn, not when asked, and resets the screen with the queued in-stream `\x1bc` only once its capture is in hand, right before the replay (never xterm's `clear()` before the fetch), so a tile keeps its last frame through its wait and its own round trip; a failed, aborted or empty fetch writes nothing and resets nothing, and the tile keeps its last frame and every held live frame.
+⚠️ **One load queue.** Every capture a tile fetches (initial load, refresh after a reconnect, server `{t:'r'}` refresh, server `{t:'c'}` refresh, dropped-output recovery refresh, shell history pull) goes through the grid's ONE `TileLoadQueue` (terminal-tile.js, the tile's `scheduleLoad` option): concurrency 1, a history pull first, then the focused tile, then reading order; a destroyed tile's waiting loads are dropped unrun and destroy() aborts its running fetch. A load holds the slot only while xterm parses its replay: `writeChunked` queues every 32 KB slice at once (one 1 MiB window at a time, since xterm's write queue throws past 50 MB) and settles on the callback of a write queued behind them, never one slice per animation frame (that pacing held the slot about a second per 1 MiB tile). ⚠️ destroy() settles a replay in progress (`_cancelReplay`): a disposed xterm never runs that callback, and an unsettled replay would hold the tile's flag and the grid's queue forever. Each `GET /api/sessions/:id/terminal` runs synchronous tmux calls, so N tiles loading at once (after a deploy restart all N reopen within a second) would stall every WebSocket and SSE stream on the server back to back. Grid tiles load the BOUNDED window (`full=1&tail=TERMINAL_TAIL_SIZE` for a TUI, `SHELL_LOAD_BASE_QUERY` for a shell, the shell load without its fixed `lines=`; the `boundedLoad` option), and every full capture of theirs (a TUI load, a shell load, a shell history pull) also sends `lines=<TILE_SCROLLBACK + rows>`, so tmux reads no more history than the tile keeps instead of the whole history limit (the route's optional `lines` bound, docs/api-reference.md; the split's Pane B sends none for a TUI and `SHELL_LOAD_QUERY`'s `lines=10000` for a shell), keep `TILE_SCROLLBACK` lines and their own per-device font (`codeman-tile-font-size`, Ctrl +/- while the grid is open). A refresh fetches at its turn, not when asked, and resets the screen with the queued in-stream `\x1bc` only once its capture is in hand, right before the replay (never xterm's `clear()` before the fetch), so a tile keeps its last frame through its wait and its own round trip; a failed, aborted or empty fetch writes nothing and resets nothing, and the tile keeps its last frame and every held live frame.
 
 ⚠️ **Selections.** The tile branch of `selectSession` sits right after its "already active" early return: a tiled id is FOCUSED (`_selectTiledSession`: no cleanup, replay, resize or main socket; the shared `_refreshSessionPanels`; only a user-initiated pick acknowledges the idle alert). A USER-initiated pick of a session that is not tiled leaves the grid for the single view, remembered (decision 1), and so does `leaveTiles` (a followed `#session=` link); an `auto` pick never collapses it. Whoever calls `closeTileGrid({ reselect: false })` and then selects must null `activeSessionId` first, or `_cleanupPreviousSession` saves the parked terminal's stale content as a snapshot. App-driven fallbacks pick a tile: `closeSession` captures the neighbouring tile BEFORE it tears the tile down (like `wasActive`) and focuses it with `auto` before the delete request goes out; the `_onSessionDeleted` wrapper does the same for a delete from elsewhere, and leaves a close from this tab (`_closingSessions`) to `closeSession`. Ctrl+Tab and Alt+[ ] cycle the tiles. A popped-out (detached) session leaves the grid. Moving focus off a zoomed tile restores the grid (tmux `select-pane`); an automatic zoom (window too small for the minimum tile) follows focus instead.
 
 ⚠️ **Chords.** `toggle-tile-grid` (Ctrl+Shift+G), `focus-tile-*` (Alt+Shift+Arrows), `move-tile-*` (Ctrl+Shift+Arrows), `zoom-tile` (Alt+Shift+Enter) and `remove-tile` (unbound) are registry entries kept OUT of `SHORTCUT_ACTIONS`: `tileShortcutFor` decides whether one applies (the toggle while the grid is open, or where one could open AND `showTileGridButton` is on, an owner decision: with the setting off the chord is inert and passes through like any unbound key; the rest only while it is open), the capture handler dispatches it, and the main terminal's and every tile's xterm key handler return false for it, for every event type and BEFORE the Shift+Enter gate (Alt+Shift+Enter would otherwise send `S-Enter`). Outside the grid the focus chords reach the terminal untouched. The move chords also apply while a tile is zoomed (a no-op, so their keys never reach the CLI). The arrow chords, focus and move, never apply in a text field other than xterm's own textarea (`isTextFieldTarget`), where shifted arrows select.
 
-⚠️ **Opening.** The Tiles button's click and `Ctrl+Shift+G` are one function, `toggleTileGrid` (owner decision 8), so the two cannot drift. It opens the grid at once with the remembered count of tiles (`codeman:tile-count`, 2, 4 or 6, default 6, owner decision 10; at most `_tileGridLimit().capacity`), on `tileGridOpenSet` (constants.js, pure): (a) the stored grid if any of its sessions survive (an open split closes and is not seeded first: `openTileGrid(..., { mergeSplit: false })`), (b) else an open split's two sessions, (c) else the open sessions in tab order (no detached ones), the active session always included and focused; then `tileGridSetForCount` trims it from the end (the session to focus kept) or fills it from tab order to the count. A stored grid comes back with its tiles in their cells (`reformTileCells`: a count change is a shape change under `fitTileCells`, the added tiles fill the empty cells first): the count supersedes decision 8's "exactly the stored set", and only the page-load restore brings back exactly what was stored. Ctrl/Cmd+click with the grid closed opens the count in total, the clicked session among them and focused. Right-click (`oncontextmenu`, which also fires for Shift+F10 and the Menu key, and calls `preventDefault` for the browser menu) is the 2 / 4 / 6 count menu (`openTileCountMenu`); with the grid open a pick re-forms it (`_reformTileGrid`: the focused tile always stays, every joining tile is mounted and laid out before any connects, so each fits once). ⚠️ The menu owns its Escape: the global handler closes it alone and gives the keyboard back to the Tiles button, like the tab-group menu. ⚠️ It closes when the keyboard leaves it for another element (a `focusout` with a `relatedTarget` outside it; a focus going nowhere, a Safari button click, does not count): closing the grid starts a selection that focuses the single view's terminal when its replay lands, and a menu left open behind that sent its arrows, Enter and Escape into the PTY. ⚠️ Opening builds the tiles' terminals one per animation frame (`_connectTilesPaced`, the focused one first, a run token stops a stale run), so `openTileGrid` returns before they exist: a focus asked for meanwhile is handed over in `_connectTile` (`focusOnConnect`), never for `focus: false`, and `_connectTile` connects a tile once (`entry.connected`, reset by `_remountTile`). ⚠️ Motion (the grid's own, on by default): tiles enter staggered (`.tile--entering`, `--tile-enter-index`), a terminal stays transparent until its first capture lands (`.tile--revealing`, cleared by the load queue's idle, 15 s backstop), and the toggle's close leaves an inert cloned still copy (`_ghostTileGrid`, no xterm or socket) that holds until the single view's `selectSession` settles (at most 700 ms) and then fades; every keyframe animates opacity and transform only (FitAddon reads the untransformed box, so still one PTY resize per tile), nothing moves and no copy is made under `prefers-reduced-motion`, and `.main.webview-active` hides the copy.
+⚠️ **Opening.** The Tiles button's click and `Ctrl+Shift+G` are one function, `toggleTileGrid` (owner decision 8), so the two cannot drift. It opens the grid at once on `tileGridOpenSet` (constants.js, pure): (a) the stored grid if any of its sessions survive, EXACTLY as the user left it (decision 11: `restoreTileGridCells` keeps its cells and holes, its focus and a zoom the user chose; `_openStoredTileGrid` puts back its divider sizes and count; an open split closes and is not seeded first: `openTileGrid(..., { mergeSplit: false })`), never filled to the remembered count and never trimmed to the window (`_applyTileLayout`'s automatic zoom shows the focused tile until the window fits, the arrangement kept); (b) else an open split's two sessions; (c) else the open sessions as `rankTileSessions` orders them (`_tileGridRanking`, tile-grid.js: no detached ones; WORKING first, the most recently started turn first, keyed off `lastSubmitAt` ONLY because a working pane's `lastActivityAt` is always "now"; then the ones NEEDING INPUT, `needs`/`waiting`, the red and yellow tab alerts, most recent first; then the rest by `sessionActivityAnchor`, newest first; a 0 stamp last, tab order the final tiebreak; the states are `_mobileOverviewState`'s, and a page without it falls back to tab order), the active session always included and focused; then `tileGridSetForCount` trims (b) and (c) from the end (the session to focus kept) or fills them from the ranking to the remembered count (`codeman:tile-count`, 2, 4 or 6, default 6, owner decision 10; at most `_tileGridLimit().capacity`). ⚠️ A stored grid's freed cells (a session gone or popped out since) are filled from the ranking first, then other empty cells, only while it holds fewer tiles than its stored `count`; a hole the user made stays. A count picked in the menu with the grid closed (`_activateTileGrid({ count })`) re-forms a stored grid to it (`reformTileCells`, the added tiles in the empty cells first). Ctrl/Cmd+click with the grid closed opens what the toggle would with the clicked session focused, never past the count: it joins while the grid holds fewer, else takes the last tile's place. Right-click (`oncontextmenu`, which also fires for Shift+F10 and the Menu key, and calls `preventDefault` for the browser menu) is the 2 / 4 / 6 count menu (`openTileCountMenu`); with the grid open a pick re-forms it (`_reformTileGrid`: the focused tile always stays, every joining tile is mounted and laid out before any connects, so each fits once). ⚠️ The menu owns its Escape: the global handler closes it alone and gives the keyboard back to the Tiles button, like the tab-group menu. ⚠️ It closes when the keyboard leaves it for another element (a `focusout` with a `relatedTarget` outside it; a focus going nowhere, a Safari button click, does not count): closing the grid starts a selection that focuses the single view's terminal when its replay lands, and a menu left open behind that sent its arrows, Enter and Escape into the PTY. ⚠️ Opening builds the tiles' terminals one per animation frame (`_connectTilesPaced`, the focused one first, a run token stops a stale run), so `openTileGrid` returns before they exist: a focus asked for meanwhile is handed over in `_connectTile` (`focusOnConnect`), never for `focus: false`, and `_connectTile` connects a tile once (`entry.connected`, reset by `_remountTile`). ⚠️ Motion (the grid's own, on by default): tiles enter staggered (`.tile--entering`, `--tile-enter-index`), a terminal stays transparent until its first capture lands (`.tile--revealing`, cleared by the load queue's idle, 15 s backstop), and the toggle's close leaves an inert cloned still copy (`_ghostTileGrid`, no xterm or socket) that holds until the single view's `selectSession` settles (at most 700 ms) and then fades; every keyframe animates opacity and transform only (FitAddon reads the untransformed box, so still one PTY resize per tile), nothing moves and no copy is made under `prefers-reduced-motion`, and `.main.webview-active` hides the copy.
 
 ⚠️ **zh-CN.** Every string the grid puts on screen has its own entry in i18n.js's `ZH_CN` (or a `translateDynamic` pattern for counts, exit codes and the header tooltip's state plus duration, which requires the duration: bare state words stay out of the table, see mobile-overview.js), so nothing reaches the generic leading-verb fallback. `test/tile-grid-i18n.test.ts` drives the real tile code through every state that writes text, harvests each string and requires a full translation (and the same English back in `en`); a new tile string needs its entry or that test fails. Session and group names stay user text (`data-i18n-skip`). ⚠️ A refresh that skips unchanged text must compare with the last ENGLISH value it set (`entry.headerLabel`, `entry.overlayLabel`, `entry.zoomLabel`), never the DOM: in zh-CN the DOM holds the translation, so a DOM compare rewrites English on every `session:updated` for the observer to translate again.
 
@@ -829,7 +878,9 @@ Further detail: with many sessions the horizontal strip stops being scannable, w
 
 ⚠️ **Attach.** A tile whose session has no PTY (`pid === null`) or whose socket closed because it exited (4009) shows an Attach overlay (absolute, the body keeps its size): `POST /interactive` (or `/shell`) with NO body, one in flight per session, a tripped PTY-exit breaker only through the same confirm as the single view, then the tile is remounted (a stopped socket cannot reconnect). The routes report a refusal in the ENVELOPE of a 200, so the response body is read, not `res.ok`. An agent that exited in a live pane (`paneExit`) cannot be started again in place (both routes refuse while the pane's tmux client runs): its tile shows the exit and points at Close session.
 
-⚠️ **Persistence and joining.** `codeman:tile-grid` holds `{ v: 1, open, ids, focused, zoomed, colFr, rowFr }`, ids only, written as the grid changes; its `ids` are the CELLS, `null` for an empty one, and `sanitizeTileGridState` returns them as `cells` (a dropped id a hole, never a shift) beside the packed `ids` every list consumer wants; a restore puts the holes back only when the cell count matches the current shape, else the tiles stay packed, and `_closeStoredTileGrid` writes the cells back, never the packed list. The old packed format reads as cells with no hole; closing keeps it as `open: false` (the Tiles toggle, the picker and Ctrl/Cmd+click bring it back), the last tile leaving forgets it, a solo window never reads or writes it, an automatic zoom is not stored. Sessions created by THIS tab's Run join the open grid (`_joinTileGridFromRun`, called from session-ui.js `_ensureCreatedSessionVisible`; a wrapper from tile-grid.js would be overwritten by session-ui.js's later `Object.assign`); sessions created elsewhere arrive only by `session:created` and never join. ⚠️ A tile that joins that way connects BEFORE Run starts its pane: its resize reaches a session with no PTY, which `Session.resize()` only records (`_lastDesktopDims`) while the spawn uses a fixed 120x40, and Run's own resize step measures the parked main terminal (`display: none`, so `proposeDimensions()` is NaN and the step is skipped). Measured live: a 97x17 tile over a 120x40 pane. So `_renderTileChrome` remembers each tile's last-seen pid and calls `tile.paneStarted()` when it appears or changes (keyed on the sessions map, so a `handleInit` after an SSE drop counts too); `paneStarted()` forgets the sent size and sends it, and a hidden tile (a zoomed neighbour) sends nothing but keeps the size forgotten, so its next `fit()` sends it. A plain `fit({ force: true })` would lose that case. Tests: `test/tile-grid-*.test.ts` over the shared vm harness `test/mocks/tile-grid-vm.ts`.
+⚠️ **Persistence and joining.** `codeman:tile-grid` (localStorage, per browser, never sent to the server) holds `{ v: 1, open, ids, count, focused, zoomed, colFr, rowFr }`, session ids and layout only, never content, written on every change (move, divider pointer-up, add, remove, count pick, focus, zoom); its `ids` are the CELLS, `null` for an empty one, and `sanitizeTileGridState` returns them as `cells` (a dropped id a hole, never a shift) beside the packed `ids` every list consumer wants, plus `freed` (cells whose session went away since) and `count`. ⚠️ `count` is how many tiles the user's own last change left: `removeTile(..., { gone: true })` (the `_onSessionDeleted` wrapper, `_reconcileTileGrid`, `_onTileExit`, and app.js `_markDetached` for a session popped out while the grid is open) does NOT lower it, so the next activation refills that cell, while a removal by hand does; it is derived (the sessions the cells name) for a value written before it, and the format stays `v: 1` so an older build still reads a newer value. A restore keeps the cells when their shape matches the current one, else `reformTileCells` (positions kept when all fit, else packed). ⚠️ Every close keeps the grid as `open: false` (`closeTileGrid({ keepStored: true })` from every caller: the toggle, a non-tiled pick, `leaveTiles`, Home, the width gate, the last tile leaving, "Open group as tiles", kill-all); `_closeStoredTileGrid` (a `#session=` link on load) flips `open` on the RAW stored value, so a gone id still frees its cell; `_openStoredTileGrid` holds `_persistTileGrid` (`_tilePersistHold`) until the grid is back, so `openTileGrid`'s packed intermediate layout is never written over it. Only when none of its sessions survive does activation rank from scratch. ⚠️ A RELOAD's fill (`_restoreTileGrid` inside `handleInit`) ranks on status and stamps only: pending approvals arrive later through the async `seedApprovals`, so needs-input cannot rank for that one fill, and a late seed never re-forms the restored grid (holding the fill back would open fewer tiles, possibly another shape, and move the user's tiles a moment after the reload; pinned in `test/tile-grid-restore.test.ts`). A solo window never reads or writes it, an automatic zoom is not stored. Sessions created by THIS tab's Run join the open grid (`_joinTileGridFromRun`, called from session-ui.js `_ensureCreatedSessionVisible`; a wrapper from tile-grid.js would be overwritten by session-ui.js's later `Object.assign`); sessions created elsewhere arrive only by `session:created` and never join. ⚠️ A tile that joins that way connects BEFORE Run starts its pane: its resize reaches a session with no PTY, which `Session.resize()` only records (`_lastDesktopDims`) while the spawn uses a fixed 120x40, and Run's own resize step measures the parked main terminal (`display: none`, so `proposeDimensions()` is NaN and the step is skipped). Measured live: a 97x17 tile over a 120x40 pane. So `_renderTileChrome` remembers each tile's last-seen pid and calls `tile.paneStarted()` when it appears or changes (keyed on the sessions map, so a `handleInit` after an SSE drop counts too); `paneStarted()` forgets the sent size and sends it, and a hidden tile (a zoomed neighbour) sends nothing but keeps the size forgotten, so its next `fit()` sends it. A plain `fit({ force: true })` would lose that case. Tests: `test/tile-grid-*.test.ts` over the shared vm harness `test/mocks/tile-grid-vm.ts`.
+
+⚠️ **Detach Tiles** (`tileDetachEnabled`, per-device, default OFF: in `displayKeys`, stripped from the settings PUT, not in `SettingsUpdateSchema`; App Settings → Appearance, its row shown on desktop only). A tile leaves the grid for a window of its own, and comes back into any Codeman window's grid, by dragging, with nothing new on the server: a pop-out is another client of the same session (`detachSession`), and a tile is a `TerminalTile` in whichever window has it. (1) A header let go OUTSIDE the window opens it with `detachSession(id, placement)`, the tile's size (at least 480x320), its header near the drop point, after `TILE_TEAR_OFF_SETTLE_MS` (120 ms) that a claim from another window cancels. "Outside" is the dragend's own point off the viewport with `dropEffect: 'none'`; where a browser reports no point (0, 0) it is the page's record of a `dragleave` at the viewport's edge with nothing entered, cleared by any later `dragover`. A drop that any target took (`dropEffect: 'move'`, in this window or another) never opens one, a drop inside on nothing changes nothing, and with the setting on a lone or zoomed tile drags too (out only: no target here takes it). (2) A header dropped on ANOTHER Codeman window's tile or empty cell moves it there: that window's `_acceptTabDrops` takes a drag that is none of its own and carries `application/x-codeman-tile` (`_isForeignTileDrag`; the id is only readable on drop), docks it through the target's own drop (`_dockForeignTile` → `dropSessionOnTile` / `dropSessionOnSlot`, so `_reorderTiles` stays the one move path) and posts `tile-adopted` on the `codeman-windows` channel. Only the window the tile was dragged FROM acts on that (its drag still settling, or ended within `TILE_ADOPT_WINDOW_MS`): the tile goes with `focus: false`, and a last tile closes the grid onto the welcome screen, never onto that session, since two windows sizing one PTY garble it. (3) A pop-out's title (`#soloSessionTitle`, `_installSoloTileHandle`) carries the same type, so it drags back onto any grid: docking a popped-out session `_redock`s it first (a detached session never mounts), posts `close-request`, and ignores that pop-out's `detached` answers for `TILE_DOCK_GRACE_MS` (`_tileDockedRecently`), which would otherwise take the new tile straight back off. A pop-out that gets `close-request` is RELEASED (`_releaseSoloWindow`): it posts `redocked` at once, stops answering roll-calls, then closes; one a script may not close (opened by typing its URL) says where the session went instead of staying a pop-out every dashboard would mark detached again. A tile's ⋯ menu offers "Open in a new window" with the setting on (`detachTile`, placed over the tile): the keyboard's way to the drag. ⚠️ Chrome keeps a script-opened window on the opener's screen unless the page holds the Window Management permission, so a drop on another monitor opens at this screen's edge (drag the window across from there), and Wayland ignores the position. The gesture overlay reaches all of it through `tileAtPoint` / `canGrabTile` / `tileDropTargetAt` / `dropOnTileTarget` / `isOverTileGrid` / `detachTileAtPoint`, never move logic of its own; its pop-out has no click behind it, so a popup blocker refuses it unless the site's pop-ups are allowed (detachSession's toast says so, and the tile stays). ⚠️ Known limits: (a) every Codeman window in one browser shares ONE stored grid (`codeman:tile-grid`, localStorage), so reloading the window a tile left can restore the other window's grid (the window that loses a tile to another holds `_tilePersistHold` and never writes over it, but the next change in either window wins); (b) a drag between two browser PROFILES docks the tile, but neither its pop-out nor the source tile hears about it, since the `codeman-windows` BroadcastChannel is per profile; (c) a drag-out held longer than the browser's user-activation window (about 5 s in Chrome and Firefox, about 1 s in Safari) gets "Pop-out blocked" and the tile stays. Tests: `test/tile-grid-detach.test.ts`.
 
 ### Gesture control: the setting
 
@@ -837,7 +888,7 @@ Further detail: with many sessions the horizontal strip stops being scannable, w
 
 ### Gesture control: the source package
 
-**Gesture-control source lives in-repo** at `packages/gesture-control/` (workspace package `codeman-gesture-control`, was the standalone `Ark0N/codeman-gesture-control` repo). The transport-agnostic core is `src/gesture/*` (MediaPipe GestureRecognizer → One-Euro-filtered cursor → pinch state machine); `src/codeman/entry.ts` is the Codeman _consumer_ that maps grab/drag/drop onto real `.session-tab`/toolbar buttons and is the bundle entry. **Edit there, then run `npm run build:gesture`** (`scripts/build-gesture-bundle.mjs` → esbuild bundles `entry.ts`, MediaPipe JS included, into `src/web/public/gesture/gesture-codeman.js`) and **commit the regenerated bundle** — the committed bundle is what dev/`tsx` serves (no bundler at runtime), and `scripts/build.mjs` reruns the same step so prod always reflects current source. The MediaPipe **wasm + model** are NOT bundled — loaded at runtime from same-origin `/gesture/wasm` + `/gesture/gesture_recognizer.task`, fetched by `scripts/fetch-gesture-assets.mjs` (gitignored, see Gotchas). `entry.ts` mounts `window.__codemanGesture = new GestureBridge()` idempotently at module-eval. A standalone vite playground (`npm run dev` in the package — fake tabs, no Codeman) lets you iterate on gesture _feel_ in isolation. ⚠️ Keep `MP_VERSION` in `fetch-gesture-assets.mjs` in sync with `@mediapipe/tasks-vision` in `packages/gesture-control/package.json`.
+**Gesture-control source lives in-repo** at `packages/gesture-control/` (workspace package `codeman-gesture-control`, was the standalone `Ark0N/codeman-gesture-control` repo). The transport-agnostic core is `src/gesture/*` (MediaPipe GestureRecognizer → One-Euro-filtered cursor → pinch state machine); `src/codeman/entry.ts` is the Codeman _consumer_ that maps grab/drag/drop onto real `.session-tab`/toolbar buttons and is the bundle entry. **Edit there, then run `npm run build:gesture`** (`scripts/build-gesture-bundle.mjs` → esbuild bundles `entry.ts`, MediaPipe JS included, into `src/web/public/gesture/gesture-codeman.js`) and **commit the regenerated bundle** — the committed bundle is what dev/`tsx` serves (no bundler at runtime), and `scripts/build.mjs` reruns the same step so prod always reflects current source. The MediaPipe **wasm + model** are NOT bundled — loaded at runtime from same-origin `/gesture/wasm` + `/gesture/gesture_recognizer.task`, fetched by `scripts/fetch-gesture-assets.mjs` (gitignored, see Gotchas). `entry.ts` mounts `window.__codemanGesture = new GestureBridge()` idempotently at module-eval. A standalone vite playground (`npm run dev` in the package — fake tabs, no Codeman) lets you iterate on gesture _feel_ in isolation. ⚠️ Keep `MP_VERSION` in `fetch-gesture-assets.mjs` in sync with `@mediapipe/tasks-vision` in `packages/gesture-control/package.json`. The tile grid's verbs (a tile pinched anywhere on it and carried onto another tile or an empty cell; a tab carried onto the grid joins it there; with Detach Tiles a tile let go outside the grid pops out) ask tile-grid.js's gesture helpers (see Tile grid, Detach Tiles) and never move tiles themselves, so the hand and the mouse cannot disagree about a drop.
 
 ### Theme skins
 
@@ -881,7 +932,7 @@ Tests: `test/terminal-touch-tap.test.ts`.
 
 ### Entrance animations
 
-**Entrance animations** (`entrance-animations.js`, all OFF by default): opt-in animations for the four things that appear when work starts, chosen per surface via `data-tab-anim` / `data-term-anim` / `data-win-anim` / `data-line-anim` on `<html>`. Defaults are the `legacy` theme, so an untouched install behaves exactly as before and every hook short-circuits on its first line. Persisted to its own `codeman:*Anim` localStorage keys (per-device, deliberately NOT in the `.strict()` `SettingsUpdateSchema`); picker in App Settings → Appearance, full per-surface lab at `?animlab=1`.
+**Entrance animations** (`entrance-animations.js`, all OFF by default): opt-in animations for the five things that appear when work starts, chosen per surface via `data-tab-anim` / `data-term-anim` / `data-win-anim` / `data-line-anim` / `data-tile-anim` on `<html>`. Defaults are the `legacy` theme, so an untouched install behaves exactly as before and every hook short-circuits on its first line. Persisted to its own `codeman:*Anim` localStorage keys (per-device, deliberately NOT in the `.strict()` `SettingsUpdateSchema`); picker in its own App Settings → **Animations** section (`#settings-animations`: the Entrance Theme, Tile Animations, and an Open lab button), full per-surface lab at `?animlab=1`. New animation settings go in that section.
 
 ⚠️ Tabs and connection lines are **destroyed mid-animation** on every re-render (`_fullRenderSessionTabs()` replaces the strip's innerHTML; `_updateConnectionLinesImmediate()` does `svg.innerHTML = ''`), so both are tracked by id and re-applied to the fresh element with a **negative `animation-delay`** to resume rather than restart.
 
@@ -893,6 +944,14 @@ Tests: `test/terminal-touch-tap.test.ts`.
 
 ⚠️ Window styles other than `beam` transform the window, which moves the rect its connection line is aimed at; `beam` deliberately animates opacity/filter only so its line can draw toward a stable target.
 
+**Tile grid entrances (the fifth surface, `data-tile-anim`).** A tile plays in two beats: its FRAME enters as it mounts in a `TILE_ANIM_STYLES` style (`settle`, `fly`, `deal`, `crt`, `beam`, `cascade`, `pop`, `soft`, `off`), and its SCREEN (`.tile-body`) plays the terminal pane's style when the load queue reports its first capture, through the same `html[data-term-anim]` rules (each names `.terminal-container.term-enter` AND `.tile-body.term-enter`). Content that lands while the frame is still entering waits for it (`--tile-screen-delay`), and the screen's `::before` wash fills `forwards` only, so a held wash never shows as a bright static block. Tiles are OFF by default: the style is its own setting, App Settings → Animations → **Tile Animations** (`#appSettingsTileAnim`, `codeman:tileAnim`), whose default `settle` is the grid's own fade and settle (`.tile--entering`, pinned by `test/tile-grid-motion.test.ts`) with no screen beat and the plain exit, exactly as before. No saved key means `settle`, so a theme saved before tiles existed gives them nothing new. Picking a theme PRESETS the tile style (every theme carries a `tile` key, `legacy` → `settle`), but the theme readout (`currentAnimTheme`) is matched on the four classic surfaces only, so changing the tile row afterwards never turns the theme into "Custom". The screen beat plays only for a tile style other than `settle`.
+
+⚠️ Tile FRAME keyframes animate **transform and opacity only**, colour on a `::before` wash: six tiles animate at once, so no `filter` (a tile's blur is its screen beat, one tile at a time as the queue serves captures). The keyframes are named `tile-enter*` / `tile-leave*` because that is what the mount's `animationend` and the still copy's removal listen for, ignoring any event with a `pseudoElement`. `test/entrance-animations.test.ts` pins both.
+
+⚠️ A themed tile is held one frame (`.tile--enter-hold`: invisible, `animation: none`), then `_runTileEntrances` orders the cascade from the final cells (`reading` / `wave` / `ripple`, a lab override in `codeman:tileAnimOrder`) and, for `fly`/`deal`/`beam`, measures its source (its session tab, else the Tiles button) on the final layout: `openTileGrid` packs tiles and a stored grid then moves them back, so measuring at mount would aim at the wrong cell. The offset rides `--tile-from-*` custom properties set BEFORE the hold comes off. The mount arms a backstop (4 s at once, the precise one from the module), so a frame that never comes cannot strand a tile invisible.
+
+⚠️ **A reload's restore always settles** (`_tileEnterQuiet` around `_restoreTileGrid`) and owes no screen beat: nothing animates on page load. A re-form (count change) keeps the plain fade on its still copy (`--now`), because that copy covers tiles that stay; only the Tiles toggle's close plays the style's own exit (`_stageTileExit`: back into the tabs, a CRT switch-off), and the copy's fallback timer waits as long as that exit says.
+
 ### Mobile tab strip scrolling
 
 **Mobile tab strip scrolling** (issue #257): under 768px the tab strip is a horizontal scroller (desktop wraps to a second row instead), so the active tab can sit off-screen. Three rules keep it reachable and they only work together: `_updateActiveTabImmediate()` scrolls the selected tab into view via `computeTabScrollLeft()` (pure, in constants.js) using **rect math on the strip's own `scrollLeft`**, never `scrollIntoView()`, which would also scroll the document under a fixed header; `_fullRenderSessionTabs()` **restores `scrollLeft`** across the `innerHTML` rebuild, since ambient rebuilds (a task badge appearing, a session created elsewhere) otherwise snap a mid-swipe strip back to 0; and it re-reveals the active tab **only when it changed** (`_lastRenderedActiveTabId`), so browsing the far end of the strip is not undone by background renders.
@@ -1041,7 +1100,7 @@ Tests: `test/mobile-prompt-composer.test.ts` (in the CI gate, deliberately not u
 
 ### Z-index layers
 
-**Z-index layers**: subagent windows (1000), split picker menu (1000, `.split-picker-menu`), Tiles count menu (1000, `.tile-count-menu`), the closing grid's still copy (`.tile-grid-ghosts`, z-index 3 inside `.main`, over the single view coming back; hidden under `.main.webview-active`), plan agents (1100), mobile/tablet fixed header (1200, `mobile.css`), modals on ≤768px (1300 — must beat the fixed header or the modal close button is buried), log viewers (2000), connection-loss overlay (2500, above the fixed header and modals), image popups (3000), response viewer (5000, backdrop 4999), file-preview overlay (5100 — must outrank the response viewer, which can launch it; at its old 2000 a path clicked in the chat opened BEHIND the chat), toasts/path picker (10000+, deliberately above the preview), the custom-model center-status banner (10001, `.center-status-banner` — `[hidden]` must re-assert `display: none` over its own `display: flex`, same trap as `.home-sessions[hidden]`, or `dismiss()` leaves an invisible click-blocker dead centre on screen), the swap-confirm and context-warning modals (10010, `#customModelSwapConfirmModal`/`#customModelContextWarningModal` — must clear both the plain `.modal` z-index of 1000 and the center-status banner it can appear over), terminal touch-selection bar (900 — above terminal content and the local-echo overlay, deliberately BELOW floating agent windows so it can never cover their controls), local echo overlay (7), iOS IME composition preview (EFFECTIVE layer depends on its home: with local echo on it is part of the local echo overlay at 7, drawn by `setComposition()` as an underlined tail after the pending text; with local echo off it is a span at z-index 6 inside `.xterm-helpers`, but `.xterm-helpers` is its own z-index 5 stacking context, so the span's effective layer is 5. That is below the overlay's 7, which is why the span cannot be used while the overlay holds text: typed text never reaches the PTY before Enter, the PTY cursor that places the span stays at the prompt start, and the overlay's opaque line div covers it).
+**Z-index layers**: subagent windows (1000), split picker menu (1000, `.split-picker-menu`), Tiles count menu (1000, `.tile-count-menu`), the closing grid's still copy (`.tile-grid-ghosts`, z-index 3 inside `.main`, over the single view coming back; hidden under `.main.webview-active`), plan agents (1100), mobile/tablet fixed header (1200, `mobile.css`), modals on ≤768px (1300 — must beat the fixed header or the modal close button is buried), log viewers (2000), connection-loss overlay (2500, above the fixed header and modals), image popups (3000), response viewer (5000, backdrop 4999), folder grid (5090, `#folderGridOverlay` — above the response viewer a folder link opens it from, below the file preview its tiles open, so a tile's preview and its ←/→ paging sit on top of the grid), file-preview overlay (5100 — must outrank the response viewer, which can launch it; at its old 2000 a path clicked in the chat opened BEHIND the chat), toasts/path picker (10000+, deliberately above the preview), the custom-model center-status banner (10001, `.center-status-banner` — `[hidden]` must re-assert `display: none` over its own `display: flex`, same trap as `.home-sessions[hidden]`, or `dismiss()` leaves an invisible click-blocker dead centre on screen), the swap-confirm and context-warning modals (10010, `#customModelSwapConfirmModal`/`#customModelContextWarningModal` — must clear both the plain `.modal` z-index of 1000 and the center-status banner it can appear over), terminal touch-selection bar (900 — above terminal content and the local-echo overlay, deliberately BELOW floating agent windows so it can never cover their controls), local echo overlay (7), iOS IME composition preview (EFFECTIVE layer depends on its home: with local echo on it is part of the local echo overlay at 7, drawn by `setComposition()` as an underlined tail after the pending text; with local echo off it is a span at z-index 6 inside `.xterm-helpers`, but `.xterm-helpers` is its own z-index 5 stacking context, so the span's effective layer is 5. That is below the overlay's 7, which is why the span cannot be used while the overlay holds text: typed text never reaches the PTY before Enter, the PTY cursor that places the span stays at the prompt start, and the overlay's opaque line div covers it).
 
 ## Security layers
 
@@ -1065,7 +1124,9 @@ Tests: `test/mobile-prompt-composer.test.ts` (in the CI gate, deliberately not u
 
 ### Buffers, uploads, and terminal history
 
-Target: 20 sessions, 50 agent windows at 60fps. Limits in `src/config/`: terminal 32MB (see below), text 1MB, messages 1000, max agents 500, max sessions 50, max SSE clients 100. **Terminal history** (`src/config/terminal-history.ts`, COD-80): tmux history-limit 100k lines, PTY buffer 32MB max / 24MB trim (env `CODEMAN_MAX_TERMINAL_BUFFER`/`CODEMAN_TRIM_TERMINAL_TO`; the env-derived trim is clamped ≤75% of max — trim ≥ max would disable `BufferAccumulator` trimming entirely = unbounded memory); browser xterm scrollback stays a separate hardcoded 50k (`DEFAULT_SCROLLBACK` in constants.js — 100k/tab is a mobile-memory hazard). tmux <3.7 allocates history at pane creation, so `createSession()` sets the global default in the same command queue immediately before `new-session`; tmux 3.7+ instead creates the session and targets only that pane, because changing the global option can resize and trim unrelated live panes. A settings change resizes tracked panes only on 3.7+ and otherwise affects future panes; no version can recover lines already evicted. Settings keys `terminalScrollbackLines`/`terminalBufferMaxBytes`/`terminalBufferTrimBytes` are schema-validated but inert (only `tmuxHistoryLimit` is wired); `buffer-limits.ts` re-exports the defaults. Text/message limits are env-overridable too (`CODEMAN_MAX_TEXT_OUTPUT`/`CODEMAN_TRIM_TEXT_TO`/`CODEMAN_MAX_MESSAGES`). **Image upload** (`image-input.js` / `config/buffer-limits.ts`): up to `_maxBatchImages` 20 images/batch (bounded concurrency 3), per-file `MAX_PASTE_IMAGE_BYTES` 50MB (env `CODEMAN_MAX_PASTE_IMAGE_BYTES`); the mobile camera-roll picker auto-downscales to fit before upload. **HEIC paste uploads** (#151): converted server-side to JPEG in a `worker_threads` worker (`web/heic-jpeg-worker.ts`, resourceLimits + 30s timeout) gated by `runWithConversionLimit()`; detection is magic-byte based (covers Android/MIUI HEIFs mislabeled as JPEG); headers declaring > 64MP are rejected 415 BEFORE decode (decompression-bomb guard). Deps: `heic-decode` + `jpeg-js`. Use `LRUMap` for bounded caches, `StaleExpirationMap` for TTL cleanup. Anti-flicker pipeline: `docs/terminal-anti-flicker.md`.
+Target: 20 sessions, 50 agent windows at 60fps. Limits in `src/config/`: terminal 32MB (see below), text 1MB, messages 1000, max agents 500, max sessions 50, max SSE clients 100. **Terminal history** (`src/config/terminal-history.ts`, COD-80): tmux history-limit 100k lines, PTY buffer 32MB max / 24MB trim (env `CODEMAN_MAX_TERMINAL_BUFFER`/`CODEMAN_TRIM_TERMINAL_TO`; the env-derived trim is clamped ≤75% of max — trim ≥ max would disable `BufferAccumulator` trimming entirely = unbounded memory); browser xterm scrollback stays a separate hardcoded 50k (`DEFAULT_SCROLLBACK` in constants.js — 100k/tab is a mobile-memory hazard). tmux <3.7 allocates history at pane creation, so `createSession()` sets the global default in the same command queue immediately before `new-session`; tmux 3.7+ instead creates the session and targets only that pane, because changing the global option can resize and trim unrelated live panes. A settings change resizes tracked panes only on 3.7+ and otherwise affects future panes; no version can recover lines already evicted. Settings keys `terminalScrollbackLines`/`terminalBufferMaxBytes`/`terminalBufferTrimBytes` are schema-validated but inert (only `tmuxHistoryLimit` is wired); `buffer-limits.ts` re-exports the defaults. Text/message limits are env-overridable too (`CODEMAN_MAX_TEXT_OUTPUT`/`CODEMAN_TRIM_TEXT_TO`/`CODEMAN_MAX_MESSAGES`). **Image upload** (`image-input.js` / `config/buffer-limits.ts`): up to `_maxBatchImages` 20 images/batch (bounded concurrency 3), per-file `MAX_PASTE_IMAGE_BYTES` 50MB (env `CODEMAN_MAX_PASTE_IMAGE_BYTES`); the mobile camera-roll picker auto-downscales to fit before upload. A camera-roll VIDEO (`?kind=video` on the same route, set by the client from the file's MIME type) is never buffered: it streams to disk under `MAX_PASTE_VIDEO_BYTES` 1GB (env `CODEMAN_MAX_PASTE_VIDEO_BYTES`), sniffed from the stream's first bytes (ISO BMFF / QuickTime atom, EBML for webm, Ogg for ogv) and unlinked on any failure. The client keys the flag on the MIME type, falling back to the name's extension because some Android file managers hand over a File with none; that classifier (`_promptAttachKind()`, image-input.js) is the ONE filter every pick and paste surface reads: Attach, drop, the Compose and paste-dialog pickers and the clipboard handlers. ⚠️ The flag is the ONLY way onto that path and admits video types alone (the registry's `VIDEO_ATTACHMENT_EXTENSIONS`, single-sourced so an uploaded video also previews: mp4/webm/mov/m4v/ogv; 3gp/mkv are out), so an image can never ride the larger cap through memory; without it the route is byte-identical to the image-only version. A DOCUMENT (the office formats pdf/docx/pptx/xlsx/zip plus the text family txt/md/csv/tsv/json/log/yaml/yml/xml, the curated `ALLOWED_DOCUMENT_EXTS`, never `SUPPORTED_ATTACHMENT_EXTENSIONS` or `EDITABLE_EXTENSIONS`) is admitted by its name's extension and buffered like an image, judged on its bytes by the same fail-closed `imageMagicMatchesExt` (a magic-number signature, or UTF-8 without a NUL byte for the text family) and saved under a sanitized `[A-Za-z0-9._-]` stem of its name; sh/py/js/html are a deliberate SCOPE exclusion by extension (the server declines to store what `file-raw` would serve or an agent could run), not a content filter, since the agent can already write any of them. `_promptAttachKind()` returns `image`, `video` or `document`, and the server's dispatch mirrors it: an image MIME wins whatever the name says, a document extension takes the document branch, and a foreign extension with no image MIME is refused `Unsupported file type` only after the image sniff fails, so nothing the image path accepted by its bytes is lost. **HEIC paste uploads** (#151): converted server-side to JPEG in a `worker_threads` worker (`web/heic-jpeg-worker.ts`, resourceLimits + 30s timeout) gated by `runWithConversionLimit()`; detection is magic-byte based (covers Android/MIUI HEIFs mislabeled as JPEG); headers declaring > 64MP are rejected 415 BEFORE decode (decompression-bomb guard). Deps: `heic-decode` + `jpeg-js`. Use `LRUMap` for bounded caches, `StaleExpirationMap` for TTL cleanup. Anti-flicker pipeline: `docs/terminal-anti-flicker.md`.
+
+**Prompt uploads live in `<workspace>/.codeman-uploads/`** (`POST /api/sessions/:id/paste-image`; the names live ONLY in `src/web/paste-image-gc.ts`): IN the workspace because that is the only path that resolves identically for a local agent and a container (only the workspace is bind-mounted, at the same absolute path; `~/.codeman` is not); hidden so it stays out of `git status` and the agent's view of the repository; FLAT because a nested `<workspace>/.codeman/` IS the data dir (`join(homedir(), '.codeman')`, `src/config/instance.ts`) when the workspace is the home directory, which nothing refuses; and self-ignoring through a `.gitignore` of `*` written once with `wx`, so a file already there is the user's and stays. The pre-move `.claude-images/` receives nothing but stays readable for one release: `UPLOAD_DIR_NAMES` lists both, the hourly sweep and the delete cleanup in `cleanupSession()` go through `uploadDirs()`, and the image watcher's ignore filter reads the names. ⚠️ `uploadDirs()` is async: the working directory is a user-chosen path, so it is probed BOUNDED first (`probePathKind()`, the #516 rule; `unknown` is skipped and never touched, the hourly sweep keeps the stall cap and `cleanupSession()` passes `pastCap`, since it acts on one path at the user's request), and the lstat and realpath after it are `fs.promises` calls, never sync, because the sweep runs 30 s after boot and hourly over every live session and a linked case on a dead mount would otherwise freeze the whole server. It returns only REAL directories (lstat at the leaf, so a planted `.codeman-uploads -> /other-case/.codeman-uploads` is never listed: `readdir` follows a link to a directory, and the sweep would otherwise age out the other case's uploads through it), none that is or contains `getDataDir()` (contrived, `CODEMAN_INSTANCE=uploads` makes `~/.codeman-uploads` the data dir of a home workspace and `CODEMAN_DATA_DIR` can point inside one, but it is one check next to a recursive delete; one strictly INSIDE the data dir is listed, since it only ever holds uploads and a workspace like the `~/.codeman/app` that `install.sh` clones would otherwise never have its uploads collected), and nothing for a remote (SSH) session, whose `workingDir` is the remote path and would otherwise name a same-named LOCAL directory for the sweep and the delete. The check is made when the directories are listed: a same-user process that swaps a listed directory for a link during the sweep's awaits is accepted (Node has no `openat()`, and that actor already writes anywhere this process can). The route refuses a remote (SSH) session with 400 before any disk touch: its `workingDir` is the remote path, so the file would land on THIS host where the agent cannot read it. ⚠️ And because a live session's workspace can go onto a network mount that later stops answering, the route probes `session.workingDir` BOUNDED (`probePathKind(..., { pastCap: true })`) before the first disk touch, refusing an `unknown` mount through `describeUnknownPath()` (OPERATION_FAILED) exactly as `POST /api/sessions` does, so a mount already unreachable cannot freeze a libuv worker on every upload (`ensureUploadDir()`'s own lstat/mkdir then start on a reachable path; a mount that dies between the probe and that write is the same residual race `POST /api/sessions` accepts, since a pre-check cannot bound the write that follows it). Tests: `test/paste-image-gc.test.ts`, `test/paste-image-dir-shared.test.ts`, the paste-image block of `test/routes/session-routes.test.ts`, `test/routes/session-create-unreachable-path.test.ts`.
 
 ### Process-tree walks are bounded
 

docs/browser-testing-guide.md

@@ -77,11 +77,13 @@ A test that starts a server binds an ephemeral port, never a fixed one:
 
 - `WebServer`: `new WebServer(0, false, true)`, then read the port the OS handed out from
   `server.boundPort` after `await server.start()`. `test/test-ports-guard.test.ts` fails
-  any `WebServer` built under `test/` on a non-zero port (a shrink-only legacy list
-  excepted).
-- A raw Fastify or `ws` server: `listen({ port: 0 })`, then `address().port`.
-- The mobile suite (`test/mobile/**`, via `createTestServer(PORT)`) keeps the fixed-port
-  convention in `test/mobile/README.md` for now.
+  any `WebServer` built under `test/` on a non-zero port.
+- A raw `http`, `net`, Fastify or `ws` server: `listen({ port: 0 })`, then
+  `address().port`. The guard also fails a raw listen on a number or a `…PORT` constant.
+- The mobile suite (`test/mobile/**`) gets its server from `createTestServer()`, which
+  binds an ephemeral port too; read it from `server.boundPort`.
+- The one exception is `test/codex-predictive-echo.test.ts`, which starts a separate lab
+  server process on port 3222 that the guard cannot see.
 - Never port 3000: that is the live instance.
 
 `scripts/browser-comparison.mjs` is a standalone script outside the guard and still uses

docs/cli-registry.md

@@ -25,7 +25,7 @@ Every run mode Codeman can launch — Claude Code, Terminal/Shell, OpenCode, Cod
 App Settings → Agents & CLIs → **CLI management** (`cliManagementEnabled`, default OFF; admin-only in multi-user mode) lists every entry with an installed/not-installed badge and:
 
 - toggles any entry on or off. A `kind: 'shell'` entry cannot be disabled, and the row shows no switch for it. A disabled CLI disappears from the Run menu, the welcome screen and the phone overview, and new session requests for it are rejected.
-- installs a missing **stock** CLI by running its shipped install command, after a confirm that names the exact command. Only one install per CLI runs at a time, and the command runs without any `CODEMAN_*` variable in its environment. A custom entry's install command is never executed.
+- installs a missing **stock** CLI by running its shipped install command, after a confirm that names the exact command. Only one install per CLI runs at a time, and the command runs without any `CODEMAN_*` variable in its environment. An `npm install -g` command is pointed at `~/.local` (which every resolver searches) when the npm global prefix is not writable by the server user, for example a system node under `/usr`; a writable prefix, a prefix that does not exist yet but could be created, and an explicit `NPM_CONFIG_PREFIX` are left alone. A custom entry's install command is never executed.
 - adds, edits and deletes **custom** entries (id, label, badge, binaries, launch argv). The server re-validates the whole assembled entry through `CliEntrySchema`, so the form cannot bypass the load-time rules.
 
 These are the only writes to `clis.json`. They are serialized, and a file that does not parse or has unsafe permissions is refused rather than overwritten; fix it (or `chmod 600` it) and retry. The HTTP routes are listed in `docs/api-reference.md` under *CLI management*.
@@ -36,7 +36,7 @@ These are the only writes to `clis.json`. They are serialized, and a file that d
 interface CliEntry {
   id: CliId; // 'codex'
   label: string; // 'Codex' — shown in menus
-  shortBadge: string; // short label ("Run CX", the Settings CLI list), e.g. 'CX'; tabs show the run-mode-dot logo instead
+  shortBadge: string; // short label ("Run CX", the Settings CLI list), e.g. 'CX'; tabs show the run-mode-dot logo instead (no mark at all with CLI Logos on Tabs off)
   accent: string; // single hex colour
   enabled: boolean;
   stock: boolean; // set by the loader; a custom entry can never claim it
@@ -46,9 +46,10 @@ interface CliEntry {
   launch: CliLaunch; // the structured argv template
   env: CliEnv; // exports, tmux setenv keys, the env-override allowlist
   capabilities: CliCapabilities; // what every call site reads instead of the id
-  //   .workDetect?: { promptGlyph, workingLine, watchingLine?, watchingLines?, awaitingLine? }
-  //   — how this CLI's pane shows work, work it started in the background, and a turn
-  //   that ended waiting for workers it will resume from
+  //   .workDetect?: { promptGlyph, workingLine, watchingLine?, watchingLines?, awaitingLine?,
+  //                   blockedLine?, blockedLines? }
+  //   (how this CLI's pane shows work, work it started in the background, a turn
+  //   that ended waiting for workers it will resume from, and a dialog it is blocked on)
   //   .modelDetect?: { screenLine, screenLines? }
   //   (where this CLI's own chrome names the model it runs: SessionState.displayModel)
   //   .launchDefaults?: { [launchParam]: settingsKey }
@@ -62,7 +63,7 @@ interface CliEntry {
 
 ### Regexes that come from config
 
-Five capability fields carry a regular expression an override file can set: `discovery.version.regex`, `capabilities.workDetect.workingLine`, `capabilities.workDetect.watchingLine`, `capabilities.workDetect.awaitingLine` and `capabilities.modelDetect.screenLine`. All five go through `compileVersionRegex()`, which caps the source at 200 characters, refuses the nested-quantifier shapes that cause catastrophic backtracking, and returns `null` rather than throwing so every caller degrades instead of crashing.
+Six capability fields carry a regular expression an override file can set: `discovery.version.regex`, `capabilities.workDetect.workingLine`, `capabilities.workDetect.watchingLine`, `capabilities.workDetect.awaitingLine`, `capabilities.workDetect.blockedLine` and `capabilities.modelDetect.screenLine`. All six go through `compileVersionRegex()`, which caps the source at 200 characters, refuses the nested-quantifier shapes that cause catastrophic backtracking, and returns `null` rather than throwing so every caller degrades instead of crashing.
 
 `workingLine` is the one that matters most, because it is compiled once per session and then run against every accumulated PTY chunk and every pane capture. A nested quantifier there is a ReDoS against the event loop for the whole server, not just that session. The guard therefore runs in two places, and neither is redundant: `schema.ts` rejects the entry at LOAD time so a bad pattern never reaches a session, and `_workingLinePattern()` in `session.ts` compiles through the same helper so the runtime cannot end up with a pattern the schema would have refused.
 
@@ -118,9 +119,65 @@ nothing the agent can write — not even the status line above it, whose command
 running with permissions bypassed can write into its own `.claude/settings.json`. Codex's row
 shares its slot with the last row of the transcript whenever no terminal is running, so a
 message ending in that exact line is matched. What keeps that harmless is `hooks: 'none'`: no
-hook event from a codex session reaches the approvals inbox, so a forged label costs a wrong
-badge and cannot silence an alert. Before giving a CLI both hook signals and a pattern, make
-sure its row is one the agent cannot write.
+IDLE prompt from a codex session ever reaches the approvals inbox (its screen-read dialogs,
+below, open PERMISSION items, which a watching label never acknowledges), so a forged label
+costs a wrong badge and cannot silence an alert. Before giving a CLI idle signals and a
+pattern, make sure its row is one the agent cannot write.
+
+### Dialogs a CLI blocks on (`blockedLine`)
+
+`blockedLine` is the screen-read source of the `blocked` state for a CLI with no hook to
+report it. Codex is the reason it exists: it is the one CLI Codeman launches with its
+approvals on by default, and with `hooks: 'none'` a codex session parked on "Would you like to
+run the following command?" used to read as plainly idle, and one parked on a question as
+WORKING (the question's own key hints end in `esc to interrupt`, codex's working line). A
+pane matching it is BLOCKED, which raises exactly what a Claude `permission_prompt` hook
+raises: the red "needs you" tab alert and desktop notification, web push and webhooks, the
+phone overview's NEEDS YOU row, an Approvals Inbox item, a run-summary row and the `blocked`
+wait signal (`web/screen-blocked.ts`). `stop` stays unavailable for such a CLI: nothing on a
+screen says a turn ended.
+
+The pattern runs over the last `blockedLines` (1 to 24, default 1) non-blank rows, joined
+with newlines and tested once (`isBlockedScreen()`, `session-activity.ts`), so it can demand
+several markers and anchor the dialog's footer on the END of the window (no `m` flag: `$` is
+the last row). Codex's needs a `› N.` option row and then its key-hint row (`Press enter to
+confirm or esc to cancel`, `tab to add notes | enter to submit answer | esc to interrupt`,
+`enter continue · esc quit` for the startup trust prompt, `esc back` when it follows the
+update prompt), allowing for that row to wrap onto rows of its own below about 60 columns.
+Measured on codex-cli 0.162.0 at widths 200 down to 36: the command, edit and
+network-escalation approvals, single and multi-question
+`request_user_input` dialogs (notes field open or closed), and the workspace trust prompt
+match; the idle and working views, popups the user opened (`/model`, `/permissions`,
+`/review`), the composer holding the dialogs' words and a transcript quoting them verbatim
+do not. `Implement this plan?` after a plan turn is deliberately not matched: its footer is
+the same as `/model`'s, and the turn has ended there. The fixtures are in
+`test/fixtures/codex-dialogs/`.
+
+Three rules hold for any CLI that declares one:
+
+- ⚠️ The text is pane-derived and the agent writes most of the pane, so the pattern must end
+  on chrome only the CLI draws, at the foot of the screen, where a dialog replaces the
+  composer. A match is an ALERT and nothing else: the inbox item is `answerInTerminal`, the
+  answer route refuses it, push offers no Approve/Deny, and Codeman never answers such a
+  dialog itself, so a forged match costs a false alert and never a keystroke.
+- The state raises on one reading (the session's settle probe, which reads the screen once
+  the pane goes quiet, which is when a dialog has just been drawn), and while it is up a timer
+  re-reads the screen every 2 s (`BLOCKED_RECHECK_MS`), because answering resumes the turn and
+  a busy pane would not reach the settle path again until the turn ended. The read is skipped
+  while the pane has drawn nothing since the last reading that saw the dialog (a dialog leaves
+  only through a repaint), with a safety-net read every 30 s (`BLOCKED_QUIET_RECHECK_MS`), so a
+  dialog left unanswered for hours does not cost a synchronous `tmux capture-pane` every 2 s.
+  It clears only after
+  two readable captures in a row without the dialog (`BLOCKED_CLEAR_READS`), so a capture
+  taken mid-repaint cannot clear and re-raise it (a second alert, push and webhook for one
+  dialog). An unreadable capture changes nothing in either direction. A blocked screen wins
+  over the working line in the same capture.
+- Hooks win: a session whose CLI reports blocked itself (`hooksAvailableForMode()`) gets
+  nothing from the screen, so declaring both can never alert twice for one dialog.
+
+A `GET /api/sessions/:id/wait?until=blocked` that arrives while such a dialog is already up is
+answered at once (`immediate: true`): unlike a hook's edge, this is a level the server is
+still reading. A caller that asked only for `idle` keeps getting `idle`.
 
 ### Three capabilities that must stay independent
 
@@ -128,6 +185,18 @@ sure its row is one the agent cannot write.
 
 `test/cli-capability-predicates.test.ts` asserts that no two of the three are equivalent across the catalog, so collapsing them fails the build rather than a user's session.
 
+### Where the agent skill lands
+
+`agentSkillInjection` says a CLI's sessions get the agent skill (`skills/codeman`) and its per-session preamble when the synced `agentSkillEnabled` setting is on; `agentSkillDir` says which workspace directory the CLI scans for skills, read through `agentSkillDir()` (hooks-config.ts), so no call site branches on a CLI id. Absent with the boolean on means `.claude/skills`, which is what the boolean meant before the field existed.
+
+| CLI | Skill directory |
+| --- | --- |
+| claude | `.claude/skills/codeman` |
+| codex, gemini, opencode, antigravity, grok, deepseek, omp | `.agents/skills/codeman` |
+| pi, shell | none |
+
+Measured 2026-10-11 against the installed builds or their source: every CLI in the middle row scans `<workspace>/.agents/skills/` (gemini only in its cwd, dsh only at the git root, else cwd), and Claude Code does not read it, so a case shared with a claude session lists the skill once there (opencode, grok and omp read both directories and see the same skill twice, identical, under one name). ⚠️ pi is left out on purpose: any project-level `.agents/skills` or `.pi/skills` makes pi ask "Trust project folder?" at startup, which would stall a spawned worker on a dialog; it reads the user-level `~/.agents/skills` without the prompt (`npx skills add Ark0N/Codeman --skill codeman -g`). A codex, gemini or other session's copy in a case that pi sessions also use triggers that prompt once for the human there. ⚠️ `agentSkillDir` is a closed enum (`AGENT_SKILL_DIRS`, schema.ts), never a free path: `clis.json` can set it and it decides where the server writes. ⚠️ Two codex traps the skill names rather than fixes: codex's default sandbox blocks network even to loopback (the skill needs **Bypass approvals and sandbox**), and codex 0.162 runs tool commands through a shared `app-server` daemon whose environment (`CODEMAN_SESSION_ID`, `CODEMAN_API_URL`) is that of the session that started it.
+
 ## The newline chord
 
 `capabilities.newline` (`'line-feed'` | `'esc-enter'`, absent = line feed) is the byte sequence the `send-key` route types into the pane for Shift+Enter. A line feed (`0x0a`, also Ctrl+Enter) is what Claude Code's Ink input reads as "insert a newline"; `esc-enter` (`ESC CR`, the Option/Alt+Enter chord) is there for a composer that ignores a bare line feed. No stock CLI declares it today: the bytes are typed by tmux on the server, so the browser's OS cannot change what a CLI reads, and Codex 0.147.0 was checked to take a line feed (a Shift+Enter that submits is the keypress leak fixed in #520, not a byte problem). A user `clis.json` can set it for a CLI that needs it. It is an enum rather than a byte string on purpose: config never carries bytes that get typed into a pane. Settings → Terminal & Input → **Key tester** prints what a browser reports for keydown/keypress/keyup, to see whether a device is sending what you think.
@@ -265,9 +334,11 @@ A module-level const freezes at first import, and the failure is asymmetric: a C
 
 ## MCP server sync
 
-`capabilities.mcpConfig` (`{ path, format, relocation? }`, `path` relative to the home directory) names the file a CLI keeps its user-level MCP server list in and the dialect it is written in. `src/mcp-sync.ts` reads that list from every ENABLED CLI that declares one, and that is installed or already has the file (a CLI that is neither is reported `absent`, never created), and adds any server a CLI is missing from the others. It writes other tools' own config, so it is **opt-in**: `mcpSyncEnabled` (synced, default OFF) gates `GET`/`POST /api/mcp-sync` (403 while off) and the Settings → Agents & CLIs → MCP servers controls. Declared today for claude, gemini, codex, opencode and antigravity; every format was checked against what the CLI's own `mcp add` writes, except opencode's (documented, not installed to check). A CLI with no entry (pi, grok, omp, deepseek) is not guessed at: it is listed as `unsupported` in the result when enabled. Adding one is a registry entry plus a small adapter in `mcp-sync.ts`, and a verified fixture in `test/mcp-sync.test.ts`.
+`capabilities.mcpConfig` (`{ path, format, relocation? }`, `path` relative to the home directory) names the file a CLI keeps its user-level MCP server list in and the dialect it is written in. `src/mcp-sync.ts` reads that list from every ENABLED CLI that declares one, and that is installed or already has the file (a CLI that is neither is reported `absent`, never created), and adds any server a CLI is missing from the others. It writes other tools' own config, so it is **opt-in**: `mcpSyncEnabled` (synced, default OFF) gates `GET`/`POST /api/mcp-sync` (403 while off) and the Settings → Agents & CLIs → MCP servers controls. Declared today for claude, gemini, codex, opencode and antigravity (plus GitHub Copilot CLI as a sync-only target, below); every format was checked against what the CLI's own `mcp add` writes, except opencode's (documented, not installed to check). A CLI with no entry (pi, grok, omp, deepseek) is not guessed at: it is listed as `unsupported` in the result when enabled. Adding one is a registry entry plus a small adapter in `mcp-sync.ts`, and a verified fixture in `test/mcp-sync.test.ts`.
 
-`relocation` (`{ envVar, path }`) names the env var the CLI itself reads to move that file: claude `CLAUDE_CONFIG_DIR` (`.claude.json` under it), codex `CODEX_HOME` (`config.toml`), opencode `XDG_CONFIG_HOME` (`opencode/opencode.json`) and gemini `GEMINI_CLI_HOME` (`.gemini/settings.json`); antigravity follows `$HOME` only, so it declares none. The var is read from the SERVER process env at call time, which is the env the CLIs Codeman spawns inherit. An absolute value moves the file to `<value>/<relocation.path>`, an empty one counts as unset (as it does for each CLI), and anything else reports the target `skipped` with the reason instead of writing a file the CLI never reads. A per-session relocation (a session's own `CLAUDE_CONFIG_DIR` in `envOverrides`) is not followed: the sync only knows the server's environment.
+**Tools that are not run modes.** GitHub Copilot CLI keeps an MCP list worth syncing but Codeman does not launch it, so it has no registry entry. `src/mcp-sync-targets.ts` declares such tools as plain data (`MCP_SYNC_ONLY_TOOLS`: id, label, config path, dialect, relocation var, the binary whose presence means "installed"). They join the registry CLIs as sync targets (listed after them, so a registry CLI's definition wins a same-name difference), under the same rules: installed or already configured, otherwise `absent`. Copilot's dialect is `copilot-json` (`~/.copilot/mcp-config.json`, relocated by `COPILOT_HOME`; checked against `copilot mcp add` 1.0.94): `mcpServers`, each entry with `tools` (`["*"]` = all), `type` `local` | `http` | `sse`, `command`/`args`/`env` or `url`/`headers`. `copilot mcp disable` does not mark the entry: it lists the name under `disabledMcpServers` in `settings.json` beside the config. Sync reads that list (never writes it) so a disabled server is not copied, and reports the target `unreadable` if `settings.json` is not valid JSON rather than guessing.
+
+`relocation` (`{ envVar, path }`) names the env var the CLI itself reads to move that file: claude `CLAUDE_CONFIG_DIR` (`.claude.json` under it), codex `CODEX_HOME` (`config.toml`), opencode `XDG_CONFIG_HOME` (`opencode/opencode.json`), gemini `GEMINI_CLI_HOME` (`.gemini/settings.json`) and, for the sync-only Copilot CLI, `COPILOT_HOME` (`mcp-config.json`); antigravity follows `$HOME` only, so it declares none. The var is read from the SERVER process env at call time, which is the env the CLIs Codeman spawns inherit. An absolute value moves the file to `<value>/<relocation.path>`, an empty one counts as unset (as it does for each CLI), and anything else reports the target `skipped` with the reason instead of writing a file the CLI never reads. A per-session relocation (a session's own `CLAUDE_CONFIG_DIR` in `envOverrides`) is not followed: the sync only knows the server's environment.
 
 The rules the module keeps and the tests pin: it only ADDS (a name already defined, in any shape, is never edited or removed; a same-name difference is reported as a conflict); a server switched off in its own CLI is not copied; it never writes a file it could not parse (opencode JSONC with comments, a TOML file with a duplicate table) and re-parses the new text before writing; codex TOML is read with a real parser (`smol-toml`), so CRLF files and inline tables are handled; names such as `__proto__` are ignored and every table keyed by an untrusted name has no prototype; a symlinked config is written through, not replaced; a file that receives `env`/`headers` is left `0600`; only one apply runs at a time; and its result carries server names only, never env values or headers, and never file text: a parse failure is reported by line and column, not by the parser's message (smol-toml prints a code frame of the offending lines and V8's JSON errors quote source, either of which can hold a secret). The schema restricts `path` and `relocation.path` to a relative path without `..`, since sync writes to it.
 

docs/cron-guide.md

@@ -5,7 +5,7 @@ spin up a Claude (or shell / OpenCode / Codex / Antigravity / Gemini / Pi) sessi
 feed it a prompt. Think "cron for agent sessions": _"every weekday at 3am, open a
 Claude session in `~/proj` and tell it to update dependencies and open a PR."_
 
-- **UI**: the **⏰ Cron** button in the header → the Cron Jobs modal (`#cronModal`).
+- **UI**: the **⏰ Cron** button in the bottom toolbar → the Cron Jobs modal (`#cronModal`).
 - **API**: `/api/cron/jobs*` and `/api/cron/runs`.
 - **Code**: `src/cron/cron-service.ts`, `src/cron/cron-time.ts`, `src/cron/cron-input.ts`,
   types in `src/types/cron.ts`, routes in `src/web/routes/cron-routes.ts`,
@@ -24,7 +24,7 @@ Claude session in `~/proj` and tell it to update dependencies and open a PR."_
 
 ### In the browser
 
-1. Click **⏰ Cron** in the header.
+1. Click **⏰ Cron** in the bottom toolbar.
 2. Click **+ New Job**.
 3. Fill in a **name**, pick an **agent type** and **working directory**, choose a
    **prompt** (inline text or a file path), pick a **schedule**, and leave
@@ -204,7 +204,7 @@ trailing carriage return:
 
 | Mode    | Mechanism                                                       | Use when                                         |
 | ------- | --------------------------------------------------------------- | ------------------------------------------------ |
-| `typed` | `session.writeViaMux()` — tmux `send-keys -l` (literal) + Enter | Default; behaves like a human typing the prompt. |
+| `typed` | `session.writeViaMux()` — the text as a tmux bracketed paste + Enter (a single character is typed, so a dialog answer selects) | Default; like a human pasting the prompt and pressing Enter. |
 | `paste` | `session.write()` — writes directly to the PTY/mux              | Bulk paste-style delivery.                       |
 
 > ⚠️ **Single-line only — enforced.** Like all programmatic input in Codeman,

docs/extending-codeman.md

@@ -163,9 +163,14 @@ respawn, Ralph, the orchestrator, search, and admin. Each route module carries a
 If the caller is an agent running _inside_ a Codeman session, install the packaged
 agent skill instead of teaching it these calls by hand: `skills/codeman` in the repo
 (`npx skills add Ark0N/Codeman --skill codeman -g`, or `codeman skill install
-[--case <name>]`, or the synced `agentSkillEnabled` App Setting for automatic
-per-case injection on Claude session create). The skill carries the guard, the
-safety rules, and verified wait/orchestration recipes.
+[--case <name>] [--mode <mode>]`, or the synced `agentSkillEnabled` App Setting for
+automatic per-case injection on session create, into the directory each CLI scans for
+skills). The skill carries the guard, the safety rules, and verified
+wait/orchestration recipes.
+
+If the caller speaks the Model Context Protocol (Claude Desktop, Cursor, Codex, any MCP
+client), point it at `codeman mcp` instead: seven tools over these same endpoints,
+configured with `CODEMAN_API_URL`. See the README section "MCP server".
 
 The common ones:
 
@@ -215,17 +220,18 @@ whether your integration works, and the last one is what actually bites:
   to die in the field.
 - **`wait.timeoutMs`** is the timeout after server-side clamping (600 s ceiling by
   default). Read it rather than assuming you got what you asked for.
-- **`stop` and `blocked` only exist for `claude` sessions**, and on a `shell` session
-  even `idle` fires only once at startup, so send-and-wait there can only time out.
+- **`stop` only exists for `claude` (and `deepseek`) sessions, `blocked` also for
+  `codex`** (read off its screen), and on a `shell` session even `idle` fires only once
+  at startup, so send-and-wait there can only time out.
   See the Gotchas below.
 
 ⚠️ **There is no readiness signal, and skipping readiness is the failure that looks
 like success.** A session reports `idle` before its CLI has spawned, and a `claude`
 worker in a brand-new case comes up on the CLI's **trust dialog**, which has a ❯
-prompt of its own. Prompt it at that moment and the text lands in the dialog, the
-`\r` does not get past it, and the session's startup `idle` lands inside the wait
-window: the wait resolves on `idle` in a couple of seconds with `timedOut: false`,
-indistinguishable from a finished turn. Wait for the pid, then wait for the
+prompt of its own. Prompt it at that moment and the text lands in the dialog,
+whose highlighted option is `No, exit`, so the prompt's `\r` quits the CLI. Codeman
+holds the session's `idle` while the dialog is on screen, which makes waiting for
+`idle` safe but still not a readiness signal. Wait for the pid, then wait for the
 composer, answering the dialog only as the bounded fallback.
 
 ⚠️ **Answering it is not "press Enter".** Claude Code 2.1.252 dropped the options'
@@ -433,9 +439,11 @@ Every one of these has cost somebody real time.
   a multi-word match is unreliable there. Match one short space-free token, ideally
   one you printed yourself, and keep it out of the typed line (your own keystrokes
   echo into the stream).
-- **`stop` and `blocked` never fire for `shell`, `opencode`, `codex`, `gemini`,
-  `antigravity` or `pi` sessions.** They come from Claude Code hooks, which no other mode
-  installs, so only `idle`, `working` and `exit` exist there. Asking for them
+- **`stop` and `blocked` never fire for `shell`, `opencode`, `gemini`, `antigravity` or
+  `pi` sessions, and `stop` never for `codex`.** They come from Claude Code hooks, which
+  no other mode installs, so only `idle`, `working` and `exit` exist there. The exception
+  is `blocked` on `codex`, which Codeman raises from the approval, question and trust
+  dialogs it reads off codex's screen. Asking for them
   explicitly is a `400`; omitting `until` is safe, since the server drops them from
   the default set and echoes what it actually waited on as `wait.until`. Even in
   `claude` mode, a Docker case needs `CODEMAN_DOCKER_BRIDGE_HOOKS=1` for hooks to

docs/media-assets.md

@@ -0,0 +1,70 @@
+# Media assets: videos and GIFs
+
+Codeman produces a lot of video: README GIFs, website loops, release-note clips, promo videos and their drafts, raw screen captures. This page says where each kind lives and which of them belong in git.
+
+## The rule
+
+**Only media that a public page embeds is pushed to GitHub. Everything else stays on the maintainer's machine, in the local media library.**
+
+| Kind | Where it lives | On GitHub? |
+| --- | --- | --- |
+| GIFs the README, the wiki or a published article embeds | `docs/images/` on `master` | yes |
+| Release-note and media-kit clips | the orphan `design-assets` branch | yes, but never on `master` |
+| getcodeman.com videos and GIFs | `pr/website/assets/` (local, deployed to Cloudflare Pages) | no |
+| Masters, promo videos, drafts, shorts, feature clips, raw captures | `pr/media/` (the local media library) | no |
+| Remotion render inputs for the promo videos | `scripts/remotion/public/` on the unpushed promo branch | no |
+
+`.gitignore` enforces it: `*.mp4`, `*.webm`, `*.mov`, `*.m4v`, `*.mkv`, `*.avi` and `*.gif` are ignored everywhere except `docs/images/*.gif`, plus `scripts/remotion/public/`, whose render inputs the promo branch tracks. A video that really must be committed needs `git add -f`, which makes the exception visible in review.
+
+A GIF that leaves `docs/images/` stays in git history, so removing it does not shrink a clone. The rule keeps new weight out and the tree honest about what is actually used.
+
+## Tracked GIFs (`docs/images/`)
+
+| File | Embedded in | Shows |
+| --- | --- | --- |
+| `tiles-crt-stats-1600-20261010.gif` | README hero (en + zh-CN) | six live agents in the tile grid, powering on and off with the CRT animation, header stats on |
+| `tile-grid-20261009.gif` | README, Tile Grid section | the Tiles button opening six live sessions side by side, and closing them |
+| `zerolag-demo-20260728.gif` | README, wiki `Input-And-Voice`, `packages/xterm-zerolag-input/README.md`, CHANGELOG, getcodeman.com docs | instant local echo next to slow server echo, on two phones |
+| `codeman-tab-states-20261010.gif` | README, Tab Alerts section | annotated tab states: working, blocked on a question, idle |
+| `codeman-skill-crt-20261010.gif` | README, agent skill section; wiki `Codeman-Skill` (getcodeman.com docs show the MP4 cut) | one prompt to Claude Code spawning and driving two worker sessions through the skill |
+| `codeman-skill-grid-20261015.gif` | README, agent skill section (Step 2); wiki `Codeman-Skill` (getcodeman.com docs show the MP4 cut) | one prompt to a Claude Code lead, five Claude Code workers powering on as tiles in the grid |
+| `codeman-skill-tabs-20261015.gif` | README, agent skill section (Step 3); wiki `Codeman-Skill` (getcodeman.com docs show the MP4 cut) | one prompt to a Claude Code lead, eight worker tabs powering on with lineage lines, then a click into a live worker |
+| `subagent-demo-20260724.gif` | the "mission control" article on dev.to, Hashnode and Medium (hotlinked from `master`) | subagent windows fanning out |
+
+Before removing one, search for hotlinks: `README*.md`, `docs/wiki/`, `packages/*/README.md`, `CHANGELOG.md`, the site (`pr/website/`), and published articles, which link `raw.githubusercontent.com/Ark0N/Codeman/master/docs/images/<file>` and break silently.
+
+On 2026-10-13, seven GIFs that no page embedded left `docs/images/` and moved into the library: `codeman-demo.gif`, `codeman-skill-20261010.gif`, `subagent-demo-new.gif`, `subagent-demo-opt.gif`, `tab-alerts-glow-20260815.gif`, `tiles-crt-20261010.gif`, `tiles-crt-stats-20261010.gif`.
+
+### Adding or replacing a README GIF
+
+1. Keep the master (the lossless MP4 or the full-size GIF) in `pr/media/clips/<feature>/`, or in the project folder it came from.
+2. Commit only the size the README shows, named `<feature>-<YYYYMMDD>.gif`, and keep it lean: the current ones are 0.8 to 10 MB.
+3. A replaced GIF leaves `docs/images/` in the same commit, unless something outside the repo hotlinks it (see above).
+
+## Website media (getcodeman.com)
+
+The site is built and deployed from `pr/website/` (gitignored) with `pr/deploy-site.sh`. Its `/docs` pages are generated from `docs/wiki/`, and a wiki GIF that has an MP4 cut of the same run on the site is shown there as that video (the README and the GitHub wiki keep the GIF, since GitHub cannot autoplay inline video). Its videos and GIFs sit in `pr/website/assets/` because that directory is what gets uploaded: they are web cuts (1280 to 1600 px wide, high CRF, posters next to them) derived from masters in the library. `pr/site-notes.md` records how each one was cut and from which master. A few superseded files stay deployed on purpose, so links that point at them keep resolving.
+
+## Release notes and the media kit (`design-assets` branch)
+
+GitHub releases and the Announcements discussions embed clips from the orphan `design-assets` branch (`release-1.40/`, `release-1.41/`, `media-kit-2026-10/`). They are referenced by a commit-pinned raw URL, so the branch can change later without breaking old notes, and none of it ever touches `master`.
+
+## The local media library (`pr/media/`)
+
+`pr/` is gitignored, so the library is never pushed. It holds every master, every finished promo cut and every superseded draft in one tree:
+
+```
+pr/media/
+  README.md          curated guide: what each collection is, which cut is current, where it is published
+  INVENTORY.md       generated list of every file (size, length, resolution, fps, date, published-or-not)
+  inventory.mjs      regenerates INVENTORY.md: node pr/media/inventory.mjs
+  videos/<production>/        finished promo videos; superseded renders in drafts/
+  shorts/<episode>/           per-feature cuts of a promo; older sets in drafts/
+  clips/<feature>/            feature loops and GIF masters (README / site / release clips start here)
+  crt-gifs/                   CRT tile-animation loops
+  files-one-click-20261011/   the Files panel loop project (own README)
+  screenshots-hq/             high-res stills and their capture masters
+  raw-captures/               unedited screen recordings, safe to prune
+```
+
+Two working directories keep files the library only archives, because tools read them from fixed paths: `pr/website/assets/` (the deploy upload) and the promo worktree's `scripts/remotion/public/` (Remotion's `staticFile()` inputs). Renders still land in `scripts/remotion/out/` and are filed into `pr/media/videos/<production>/` once checked. A finished render never goes to `tmp/`.

docs/multi-user-plan.md

@@ -161,8 +161,8 @@ Codeman now ships a global **Startup Mode** picker (App Settings, Claude CLI tab
 | Screenshots `/api/screenshots` (deprecated)                                | Deprecated: removed in a later MAJOR, so no per-user subdir is planned; the replacement `POST /api/sessions/:id/paste-image` is already session-scoped. Former plan: per-user subdir `~/.codeman/screenshots/<username>/` in multi-user mode. Note: `GET /:name` deliberately rejects `/` in names as traversal, so derive the subdir server-side from `req.authUser` and keep client-visible names flat                                                  |
 | Attachments                                                                | Already session-scoped; inherits the session owner check. `attachmentConfineToWorkspace` is a global, default-OFF setting today: in multi-user mode it is FORCED ON for non-admins regardless of the setting (their attachments must resolve inside their own space); the setting keeps meaning what it means for admins                                                                                                                                  |
 | File routes (browse/preview)                                               | Path allowlist adds: non-admin paths must resolve (realpath) inside their own space or their own sessions' workingDirs                                                                                                                                                                                                                                                                                                                                    |
-| Settings (`settings.json`)                                                 | Global, admin-only writes in multi-user mode; reads allowed (per-device display keys stay in localStorage as today). Per-user server settings: out of scope v1                                                                                                                                                                                                                                                                                            |
-| System ops (self-update, tunnel toggle, span-displays, docker image build) | Admin-only                                                                                                                                                                                                                                                                                                                                                                                                                                                |
+| Settings (`settings.json`)                                                 | Global, admin-only writes in multi-user mode; reads allowed (per-device display keys stay in localStorage as today). **Implemented (1.42.0):** `PUT /api/settings`, `PUT /api/execution/model-config`, `PUT /api/config` and `PUT /api/cases/order` answer 403 `FORBIDDEN` to a non-admin before parsing anything; quick-start and Ralph loop start skip a non-admin's `lastUsedCase`; the browser skips those writes for a non-admin (`_settingsReadOnly()`, settings-ui.js) and keeps their own run mode and respawn presets on the device; a non-admin's `GET /api/settings` omits `voiceSettings.apiKey`, `caseOrder` and `lastUsedCase` (`settingsForNonAdmin`). Per-user server settings: out of scope v1|
+| System ops (self-update, tunnel toggle, span-displays, docker image build) | Admin-only. **Implemented (1.42.0):** `POST /api/system/update`, `/api/system/span-displays`, `/api/tunnel/qr/regenerate` and `/api/cleanup-state` answer 403 `FORBIDDEN` to a non-admin; a non-admin's `POST /api/auth/revoke` revokes only their own login sessions and `GET /api/tunnel/info` lists only those |
 | `getLightState` init snapshot                                              | Filtered per connection. Actual contents to filter (verified): `sessions`, `scheduledRuns`, `respawnStatus`, `subagents`, `workflowRuns`, `planUsage` (host-plan telemetry: admin-only); `globalStats` stays coarse-global. Cron jobs are NOT in the snapshot (they have their own REST route; filter there). The snapshot is cached process-wide (`LIGHT_STATE_CACHE_TTL_MS`): either key the cache per role/user or filter AFTER the cache on each send |
 
 ## 7. SSE Event Filtering

docs/orchestrator-loop-architecture.md

@@ -244,7 +244,7 @@ RalphLoop.tick() → assignTasks()  // OR OrchestratorLoop does its own assignme
 session.sendInput(task.prompt)
   │
   ▼
-writeViaMux() → tmux send-keys -l "prompt..." + Enter
+writeViaMux() → tmux paste-buffer -p (prompt as one bracketed paste) + Enter
   │
   ▼
 Claude CLI receives prompt, executes, outputs results

docs/releasing.md

@@ -0,0 +1,33 @@
+# Releasing Codeman (the COM procedure)
+
+> Moved from the root `CLAUDE.md`, which keeps a pointer here. What SemVer covers is defined in `docs/versioning-policy.md`. The root file's Session Safety rules (branch check, explicit paths, never `git add -A`) apply to every step.
+
+When user says "COM":
+
+1. **Determine bump type**: `COM` = patch (default), `COM minor` = minor, `COM major` = major
+2. **Create a changeset file** (no interactive prompts). Write a `.md` file in `.changeset/` with a random filename:
+
+   ```bash
+   cat > .changeset/$(openssl rand -hex 4).md << 'CHANGESET'
+   ---
+   "aicodeman": patch
+   ---
+
+   Detailed description of ALL changes since last release (not just the most recent commit — review full git log since last version tag)
+   CHANGESET
+   ```
+
+   Replace `patch` with `minor` or `major` as needed. Include `"xterm-zerolag-input": patch` on a separate line if that package changed too.
+
+3. **Consume the changeset**: `npm run version-packages` (auto-bumps `package.json` files, updates `CHANGELOG.md`, runs `npm install --package-lock-only`, and verifies lockfile sync via `scripts/check-lockfile-sync.mjs` — all in one command; never hand-edit `CHANGELOG.md` or `package-lock.json` versions. It also writes `src/web/public/whats-new.json` (`scripts/generate-whats-new.mjs`): the bold lead-in of each change in the new `CHANGELOG.md` section, which the browser announces once per version in a toast. Commit it with the release; `test/whats-new.test.ts` fails when its version differs from `package.json`, so a hand bump needs `node scripts/generate-whats-new.mjs && npx prettier --write src/web/public/whats-new.json` too. A release whose notes have no bold lead-ins ships an empty list and shows no toast)
+4. **Sync CLAUDE.md version**: Update the `**Version**` line in the root `CLAUDE.md` (under COM Shorthand) to match the new version from `package.json`
+5. **Commit and deploy**: verify the branch first (`git branch --show-current`), then stage EXPLICIT paths — never `git add -A`, which has swept another session's WIP into a release. `git status --short` and account for every line before committing:
+   `git add <paths> && git commit -m "chore: version packages" && git push && npm run build && systemctl --user restart codeman-web`
+6. **Refresh the getcodeman.com version badge**: the landing page's status bar carries the release version (`v<x.y.z> · getcodeman.com · MIT`), so it goes stale on every release if nobody bumps it. The site source and its deploy script are maintained outside this repository, on the maintainer's machine only; follow the local site handbook there, which also covers the numbers strip and `sitemap.xml` refresh that belong in the same pass. Poll production (`curl -s https://getcodeman.com/ | grep v<x.y.z>`) before calling it done, since the edge lags a deploy by up to a minute. Not applicable to contributor clones — skip it and say so.
+7. **Wait for CI**: after `git push`, TWO workflows fire per master push — `CI` and `Release` (the npm publish + GitHub release). List both runs for the pushed commit with `gh run list --commit $(git rev-parse HEAD) --json databaseId,workflowName` and watch EACH with `gh run watch <id> --exit-status`. Confirm both pass before considering the release done (`gh run list -L 1` returns only one of the two).
+
+8. **Publish the new version to the MCP registry**: `codeman mcp` is listed in the official MCP registry as `io.github.Ark0N/codeman`, and the registry does not follow npm, so every release is published there by hand or the listing stays on the old version. `npm run version-packages` has already written the new version into `config/mcp-registry/server.json` (`scripts/sync-plugin.mjs`). Wait until npm serves it (`npm view aicodeman version`; it can lag the green `Release` run by several minutes), because the registry checks that the npm package exists and carries `mcpName: io.github.Ark0N/codeman`. Then run `cd config/mcp-registry && mcp-publisher publish`, preceded by `mcp-publisher login github` when the saved login has expired, signed in as the account that owns the `io.github.Ark0N/` namespace. Use mcp-publisher 1.8.1 or newer (`mcp-publisher --version`), the release binary from https://github.com/modelcontextprotocol/registry/releases: the Snap Store package is stuck at 1.1.0 and rejects the current `$schema` with a false "deprecated schema detected", so never edit `server.json` for that error. Older publishers also save their login as `.mcpregistry_github_token` and `.mcpregistry_registry_token` (live tokens) in the current directory; move those out of the repo, never commit them. Confirm with `curl -s 'https://registry.modelcontextprotocol.io/v0/servers?search=io.github.Ark0N/codeman' | jq -r '.servers[] | "\(.server.version) latest=\(._meta["io.modelcontextprotocol.registry/official"].isLatest)"'`: the new version must be the one with `latest=true`. Maintainer only (it needs that GitHub account); contributor clones skip it and say so.
+
+9. **Announce it in Discussions**: every release gets a post in the [Announcements](https://github.com/Ark0N/Codeman/discussions/categories/announcements) category, shaped like #418 and #302: the short version first (one bold lead-in per theme, features before fixes, what it does for the user rather than how it works), contributor @-mentions inline where their work is described (a mention notifies them, and a contributor reposting is the cheapest reach this project has), a link to the release, and the Thanks names at the end. Casual first-person voice, no em-dashes, humanizer pass when the skill is available. A same-day follow-on patch (1.28.1 after 1.28.0) is folded into the previous post as an edit, never a second thread. Post it with `gh api graphql -F body=@<file> -f title='Codeman <x.y.z>: <hook>' -f repo=R_kgDOQ-SMDg -f cat=DIC_kwDOQ-SMDs4DCHZE -f query='mutation($repo:ID!,$cat:ID!,$title:String!,$body:String!){createDiscussion(input:{repositoryId:$repo,categoryId:$cat,title:$title,body:$body}){discussion{number url}}}'` (the repo's node id and its Announcements category id; `pinDiscussion` does not exist in the API, so pinning stays a click in the UI). This step exists because announcements stopped at 1.18 (#302) while ten releases shipped with nobody notified; #418 is the backfill covering 1.19.0 to 1.28.1, and release notes only count as content once they reach a surface people are subscribed to.
+
+CI runs `npm run check:lockfile` on every push/PR, so lockfile drift fails the build even if the `version-packages` script is bypassed.

docs/remote-sessions.md

@@ -251,6 +251,71 @@ unreachable host answers "unknown", which also means do not revive. The answer
 is cached per session and cleared whenever the pane is next seen alive, so a
 stale `true` from one transport drop can never revive the NEXT clean exit.
 
+## Terminal history from the remote tmux
+
+The local pane runs the ssh attach, so it only sees the remote tmux client's
+full-screen repaints and keeps a fraction of the scrollback (none after a
+reattach). A full-history load (`GET /api/sessions/:id/terminal?full=1`) of a
+remote session therefore reads the scrollback rows from the REMOTE pane in one
+ssh round trip (`src/remote-pane-history.ts`) and splices them above the local
+visible frame and caret. Bounded (5 s for the round trip, at most 1.5 s queued
+for a slot in the shared ssh limiter) and fail-soft: any failure keeps the
+local capture, and a transport failure (ssh exit 255, a timeout, output that is
+not the script's) also skips that host for 30 s; a failure of the remote
+command itself (a remote tmux session that is gone) affects that session only.
+How much history exists is the remote tmux's own `history-limit`, which the
+launch command does not set (tmux's default is 2000 lines).
+
+The same round trip also reads the remote SCREEN, in the same tmux command list
+as the history, because two seams would otherwise drop rows at the join:
+
+- **Rows that scroll during the round trip.** The local frame is captured only
+  once the round trip is back, so rows that scrolled off the screen in between
+  are in neither the remote history nor the local frame (measured before this
+  read existed: 13 lines lost with a 150 ms round trip and fast output).
+- **A shared remote session with a taller client.** Under `window-size latest`
+  a remote session attached elsewhere is sized by its latest client; when that
+  client is taller than this pane, the rows above this pane's frame are only
+  on the remote screen (measured before: 20 rows).
+
+Both are rows of the remote screen above where the local frame begins, so the
+load keeps them between the history and the frame (`joinScrollbackAndFrame`,
+`findFrameInScreen` in `src/tmux-manager.ts`). The frame is placed by matching
+its top rows against the screen as text (escape sequences and trailing blanks
+ignored, since the two tmuxes need not encode attributes alike); the screen's
+trailing blank rows and its last content row, which output in flight can still
+change, are left out of the match, the overlap must be at least three rows
+with two of them non-blank, and the highest match wins. A placement too low
+would REPEAT the rows between it and the true one, so a shorter overlap (a lone
+`$` row that happens to match after a spinner row changed) is not trusted.
+When the frame cannot be placed (more output scrolled than the screen held, or
+nothing matches), the join is history over frame, as without the screen read.
+That includes an agent TUI mid-turn (Claude Code, codex, opencode): it keeps a
+live region (spinner, composer, footer) pinned to the bottom and inserts new
+rows above it, so the frame is not a shifted copy of the screen and is not
+placed. A plain shell places, unless the frame's top row is the second half
+of a wrapped line (the screen holds that line as one joined row, so nothing
+matches and the join is history over frame). What is left: when the screen
+repeats a whole block of rows, the frame can be placed at the higher copy,
+which leaves out the rows between the two copies.
+
+The history, the history size and the screen are ONE tmux command list
+(`capture-pane … \; display-message … \; capture-pane …`), not separate `tmux`
+runs. tmux runs a client's command list back to back without reading pane
+output in between, so the screen begins exactly where the history ends. As
+separate runs the shell kept writing between them and the rows that scrolled in
+those milliseconds were in neither capture (measured with output streaming: 1-2
+rows per load on a LAN host, every read on a local busy pane). This rests on how
+tmux's command queue behaves, not on a documented guarantee. A random marker row
+(`display-message`) separates the history from the size and the screen, since
+tmux output cannot carry the NUL markers around it. tmux before 3.2 prints no
+newline after `capture-pane -p` output, so when the history ends inside a
+wrapped line the marker is glued onto that row; the parser accepts the marker
+at the end of a row and re-terminates the history. The remote `tail -c` bound
+applies to that whole output, history first, and is the configured byte cap
+plus 1 MiB for the marker, size and screen, so a cut lands in the history; the
+route trims what it returns to the cap itself.
+
 ## File access over SSH
 
 A remote case's `workingDir` is an absolute path on the **remote** host
@@ -270,6 +335,8 @@ and it follows the same rule as the launch path: every ssh command line comes fr
 | `GET /api/sessions/:id/file-thumbnail` | `400` for remote files |
 | `POST /api/sessions/:id/attachments` | Registers an absolute path that lives on the **remote** host (a clicked link pointing outside the case directory) by probing it there |
 | `GET /api/sessions/:id/attachments/:attachmentId/raw` | Streams the registered remote file over ssh, same 200/206/416 contract; `preview` (office) and `thumbnail` answer `400` |
+| `POST /api/sessions/:id/folders` | Lists a folder (absolute or `~/`) on the **remote** host in one ssh round trip (`remoteListFolder`): resolved with `cd -P`, missing → `404`, a file → `400 Not a folder` |
+| `GET /api/sessions/:id/folders/:folderId/files/:name/raw` | Streams one file of a granted remote folder over ssh after the same re-probe + guard as an attachment; `preview` (office) and `thumbnail` answer `400` |
 | `GET /api/sessions/:id/attachments/:attachmentId`, `GET …/attachments` (history) | Size/mtime/existence resolved over ssh, so a remote entry is not reported `missing`; the history list resolves EVERY entry in one batched probe, never one connection per entry |
 
 ⚠️ The attachment route is the one a clicked path takes when it is **outside** the case

docs/security-architecture.md

@@ -374,6 +374,23 @@ Two registration paths, with **different trust**:
   file outside the session's own workspace. Cross‑workspace attach must go through
   the explicit POST path above.
 
+**Folder grid grants** (`POST /api/sessions/:id/folders`, `src/folder-listing.ts`)
+follow the explicit-POST trust level: a deliberate, Origin‑guarded request, allowed
+cross‑workspace subject to the same guard (blocklist on the folder AND on every
+listed file, `confineToWorkspace` when set, hidden entries and unsupported types
+left out). The grant is an id for a resolved folder; files are served by
+`folderId` + a bare file name only (no `/`, `..` or NUL, supported type), and each
+request re‑runs `resolveServableAttachmentPath` on the joined path, so a file that
+was swapped for a symlink to a blocked target after listing is still refused.
+The file a symlink resolves to is held to the attachment gate as well (supported
+extension, regular file), so a supported-looking name cannot serve an arbitrary
+file, and a FIFO or device node is never opened.
+In multi‑user mode a regular user is held to the filesystem path picker's rule
+(`folderAllowedForUser`): the folder, and again each served file after symlink
+resolution, must sit inside their own `userSpacePath` or the session's
+`workingDir`, else 403. User spaces live inside `homedir()`, so without it one user
+could list and read another's space.
+
 ### SSE log‑tail route — intentional extra read roots
 
 The live file‑tail SSE route (`FileStreamManager`, used to stream a growing log
@@ -529,6 +546,7 @@ A saved dashboard URL renders as a tab, served through Codeman's own origin at `
 - **The proxy is exempt from cookie auth and the Origin/CSRF guard, and that is deliberate.** The iframe is sandboxed without `allow-same-origin`, so it is opaque‑origin: its requests are cross‑site, meaning the `SameSite=lax` session cookie is never attached and its writes and WS upgrades arrive with `Origin: null`. The credential is instead a 192‑bit capability in the path, minted only by an authenticated `POST /api/webviews/:id/open`, held in memory (a restart invalidates every one), rolling TTL, bound to the minting user, and granting nothing but "relay bytes to this one saved URL". ⚠️ **The Host allowlist is NOT bypassed**, so DNS‑rebinding protection is unaffected. A second `Referer`‑keyed form exists for root‑absolute assets and is the only exemption decided by a request‑supplied header, so it is fenced to safe methods on non‑`/api`, non‑`/ws`, non‑`/q` paths. Edges pinned by `test/webview-auth-exemption.test.ts`.
 - **The lost‑frame recovery page is the third unauthenticated 200, and the only one decided by request headers alone.** The proxy's runtime shim masks `/webview/<cap>/` off the page's own URL so a single‑page app routes on the path it expects; a navigation the page then starts itself (`location.reload()`, a root‑absolute `location.href`) lands on Codeman's root with no capability anywhere, no cookie (opaque origin) and a Referer naming the masked page. `serveLostWebviewFrame()` in `middleware/auth.ts` recognises it by shape (`GET`/`HEAD`, `Sec-Fetch-Dest: iframe` or `frame`, `Accept: text/html`, `Sec-Fetch-Mode: navigate` or absent) and answers, BEFORE the credential checks and without counting an auth failure, with a static page whose only content is a `postMessage` of the lost path to the parent tab (`default-src 'none'` plus the hash of that one script, `no-store`, `referrer: no-referrer`, no reflected input). It is fenced to paths that are NOT registered routes and never `/api/`, `/ws/` or `/q/`, with one carve‑out: `/` itself, because the landing page masks to exactly `/` and its reload otherwise rendered Codeman's app shell inside the web tab. `/` is admitted only when the request carries neither the `codeman_session` cookie nor an `Authorization` header: nothing in Codeman frames its own root and a sandboxed frame has neither, while a framed `/` that does carry credentials still gets the shell. On a passwordless install no auth hook runs, so the index route applies the same test itself (`isLostWebviewRootFrame`). ⚠️ Known property, accepted rather than mitigated: those headers are trivially set by a non‑browser client, so an unauthenticated caller can distinguish a registered route (401) from a non‑route (200) and enumerate the route table; the routes are public in `docs/api-reference.md`, so nothing is learned. Pinned by `test/webview-auth-exemption.test.ts` (password) and `test/webview-lost-root-frame.test.ts` (passwordless).
 - **Sandboxed by default; `allow-same-origin` is an explicit per‑dashboard opt‑in.** A proxied page is same‑origin with Codeman, so without the sandbox its JavaScript could read the Codeman document and call the agent‑spawning API. ⚠️ In BOTH modes the `Authorization` header and the `codeman_session` cookie are stripped before the upstream request, because a trusted (same‑origin) frame makes the browser attach Codeman's own Basic‑auth credentials to every proxied request; forwarding them would hand `CODEMAN_PASSWORD` to the dashboard.
+- **Rendered HTML files (`/html-view/<capability>/`) are the other capability-gated route that skips cookie auth.** A clicked `.html` file renders through it so an agent's report can be viewed from any device. The exemption is narrower than the proxy's: `GET`/`HEAD` only (plus the page's CORS preflight, which `registerSecurityHeaders` answers itself with `Access-Control-Allow-Methods: GET, HEAD` and never hands to a route), the capability in the path only (no `Referer` form), and it is a separate store, so a web-tab capability never opens a page and vice versa. The capability is minted by an authenticated `POST /api/sessions/:id/html-view` (session ownership via `findSessionOrFail`, the attachment guard for out-of-workspace files), held in memory with a rolling TTL and revoked with the web-tab ones on logout, admin logout and user deletion. Every response carries `Content-Security-Policy: sandbox` without `allow-same-origin`, so the page is opaque-origin even when opened as a top-level tab and cannot read Codeman's document, cookies or API. **Network access is open, by the owner's decision (2026-10-13):** that `sandbox` directive is the whole policy (it replaces Codeman's own `default-src 'self'` CSP and carries no fetch directive), so a rendered page behaves like a file opened from disk in a browser. What a page CAN do: load scripts, styles, fonts, images and media from any host (an agent's dashboard pulling Chart.js from a CDN), make fetch, XHR, WebSocket, beacon and form requests to any host, and read the files of its own folder through its capability. What it CANNOT do: act as Codeman's origin, read Codeman's cookies or document, use Codeman's API (a write arrives with `Origin: null` and the Origin guard refuses it; a read's response stays unreadable, because Codeman answers CORS only for its own localhost origins, and with a password set the request carries no session cookie either), or read any folder other than its own. The accepted consequence is that a page can send what it can read, its own folder, to any host, which is why the folder rules below are the boundary that matters. `Referrer-Policy: no-referrer` keeps the capability URL out of the `Referer` the page's third-party requests carry. Pinned by the exact-header test in `test/routes/html-view-routes.test.ts`. It may read its own directory tree (non-dot, asset-allowlisted, realpath-confined, blocklist applied; the dot and extension rules hold for the symlink-resolved target as well as the requested name), so minting is refused for an HTML file in a hidden directory or directly in a broad root (`/`, home, `/tmp`, `/var/tmp` and `$TMPDIR`, the cases and user-space roots, a drive's mount point) or in a directory above one (`/home`, `/Users`). Pinned by `test/routes/html-view-routes.test.ts` and the html-view block in `test/webview-auth-exemption.test.ts`.
 - **Not an open relay, and not a privilege boundary.** `resolveUpstreamUrl()` refuses anything leaving the saved origin, and cross‑origin redirects are handed back unchanged rather than followed. The proxy does reach whatever the SERVER can reach, which is not an escalation for someone who already commands `--dangerously-skip-permissions` agents, but in multi‑user mode it means a non‑admin's dashboard is fetched from the server's network position. Saved URLs are validated to plain http(s) with no embedded credentials, and there is deliberately **no magic‑link path**: terminal output can never create a webview (the mistake the attachment scanner had to be walled off from). The one refused destination class is link‑local and cloud‑metadata addresses (`169.254.0.0/16`, `fe80::/10`, `fd00:ec2::254`, `168.63.129.16`, `100.100.100.200`, `metadata.google.internal`): `webview-egress-policy.ts` refuses them at save time, and `webview-egress.ts` re‑judges the RESOLVED address at connect time through a `lookup` hook on the proxy's undici Agent and on its WebSocket client, so a DNS name pointing into those ranges is refused as well. Loopback and RFC1918 stay allowed on purpose. Capabilities are revoked on logout, admin logout and user deletion, and proxied responses carry `Referrer-Policy: same-origin` so a dashboard cannot hand the capability‑bearing URL to a third‑party host it links.
 
 ---

docs/tile-grid-plan.md

@@ -45,20 +45,28 @@ or settled a question the spec left open. The invariants as built are in
   `Right-click`) and `Arrows` translated. Every string has its own entry or pattern; refreshes
   compare with the last English value, not the translated DOM.
 - **The Tiles button opens the grid at once** (owner decision 8, with the count of
-  decision 10): a click (and `Ctrl+Shift+G`, the same `toggleTileGrid`) opens the remembered
-  count of tiles (default 6, at most what the window fits). `tileGridOpenSet`
-  (constants.js) picks the grid this tab last had, else an open split's two sessions, else
-  the open sessions in tab order, the active one always included and focused, and
-  `tileGridSetForCount` trims it (from the end, the session to focus kept) or fills it
-  (from tab order) to the count. A remembered grid comes back with its tiles first, in
-  their cells, then sessions in tab order, to the count in total: the count is a shape
-  change under the cell model's rule (`reformTileCells`: the tiles keep their row and
-  column when all fit, else they pack in reading order) and the added tiles fill the empty
-  cells first. This supersedes decision 8's "exactly the stored set" (owner answer). A
-  remembered grid still wins over an open split: the split closes and its sessions are not
-  seeded first. A page-load restore brings back exactly the stored grid, whatever the
-  count. Ctrl/Cmd+click on a tab with the grid closed opens the count in total, that
-  session among them and focused (owner answer: N, not N+1).
+  decision 10 and the layout memory of decision 11): a click (and `Ctrl+Shift+G`, the same
+  `toggleTileGrid`) brings back the grid this browser last had EXACTLY as the user left
+  it: which session sits in which cell, holes included, its tile count, divider sizes,
+  focus and a zoom the user chose (`restoreTileGridCells`, constants.js). It is never
+  filled to the remembered count and never trimmed to the window (a window too small
+  for it shows the focused tile alone until it fits, the arrangement kept). A session
+  that no longer exists frees its cell, which the ranking fills. Only with nothing
+  stored, or none of its sessions left, does `tileGridOpenSet` (constants.js) take an
+  open split's two sessions, else the open sessions as `rankTileSessions` orders them
+  (owner request: "prefer to load in tiles that are working and then the most recent,
+  so the oldest don't get opened"): WORKING first (the most recently started turn
+  first, keyed off `lastSubmitAt` only), then the ones NEEDING INPUT (the red and yellow
+  tab alerts), then the rest by most recent activity, tab order breaking ties; the
+  active one always included and focused, and `tileGridSetForCount` trims it (from the
+  end, the session to focus kept) or fills it (from the ranking) to the remembered count
+  (default 6, at most what the window fits). The states and stamps are the home
+  screens' own (`_mobileOverviewState`, `sessionActivityAnchor`). A remembered grid still
+  wins over an open split: the split closes and its sessions are not seeded first. A
+  page-load restore brings back the same grid as the toggle. Ctrl/Cmd+click on a tab
+  with the grid closed opens what the toggle would with that session among the tiles and
+  focused, never past the count (owner answer: N, not N+1): it joins the first empty cell
+  while the grid holds fewer than the count, else it takes the last tile's place.
 - **A hover card on the Tiles button says it** (owner feedback: "give me the hover info
   to right click over the tile button to adjust it"). It replaces the button's native title:
   "Tiles · N" (the remembered count, live), what a click does (open or close the grid),
@@ -80,9 +88,10 @@ or settled a question the spec left open. The invariants as built are in
   back on the Tiles button, Tab, a click elsewhere and the keyboard leaving it for another
   element close it (the single view a close starts focuses its terminal when its replay
   lands; a menu left open behind that would send its keys there). A pick is remembered per
-  device in `codeman:tile-count` (`codeman:tile-grid` stays ids only) and opens that many
-  tiles; with the grid open it re-forms it (`_reformTileGrid`): the focused tile always
-  stays, the others leave from the end or join from tab order, filling empty cells first,
+  device in `codeman:tile-count` and opens that many tiles (a stored grid re-formed to
+  it, its tiles first in their cells); with the grid open it re-forms it
+  (`_reformTileGrid`): the focused tile always stays, the others leave from the end or
+  join from the ranking, filling empty cells first,
   every joining tile mounted and laid out before any connects (one fit, one PTY resize
   each), and a zoom the user chose ends. The other ways in (Ctrl/Cmd+click, a dragged tab,
   "Open group as tiles", Run) still add up to the cap of 6.
@@ -97,7 +106,10 @@ or settled a question the spec left open. The invariants as built are in
   `test/header-icon-hover.test.ts`.
 - **The grid opens and closes with a short animation, on by default** (owner request:
   "when clicking on the tile button first make this animation nicer"). It is the grid's
-  own, not an `entrance-animations.js` theme (those are off by default). Opening, each tile
+  own `settle` style, the default of App Settings → Animations → Tile Animations, which
+  switches on other styles (`fly` out of the tabs, `deal` from the Tiles button, `crt`,
+  `beam`, ...; docs/architecture-invariants.md#entrance-animations).
+  Opening, each tile
   fades and settles in (opacity, translateY 6px and scale .97), 180 ms, 24 ms apart in
   reading order (`--tile-enter-index`): the last of six is done at 300 ms; a tile added
   later enters the same way. Its terminal stays transparent (`.tile--revealing`) until the
@@ -389,16 +401,34 @@ the harness/model request; see "As built")
 
 ### Persistence
 
-Decided: per device, restored on reload. Stored in localStorage key
-`codeman:tile-grid`:
+Decided: per device (per browser), restored on reload when it was open, and by
+the Tiles toggle however it was closed (decision 11). Stored in localStorage key `codeman:tile-grid`, never on the
+server:
 
 ```json
-{ "v": 1, "open": true, "ids": ["…", "…"], "focused": "…", "zoomed": null,
-  "colFr": [1, 1, 1], "rowFr": [1, 1] }
+{ "v": 1, "open": true, "ids": ["…", null, "…"], "count": 3, "focused": "…",
+  "zoomed": null, "colFr": [1, 1, 1], "rowFr": [1, 1] }
 ```
 
-Ids only, never content. A pure sanitizer drops unknown, deleted, detached and
-duplicate ids on load. Never restored in a solo window.
+Session ids and the layout, never content. `ids` are the CELLS in reading order,
+`null` for an empty one. `count` is how many tiles the user's own last change
+left (open, add, remove by hand, a count picked): a session that goes away by
+itself (deleted, popped out, its socket refused) does not lower it, so the next
+time the grid opens the ranking fills that place, while a hole the user made
+stays. It is written on every change (a move, a divider drag at pointer-up, a
+tile added or removed, a count picked, a focus, a zoom) and kept, as
+`open: false`, however the grid closes: the toggle, a non-tiled tab,
+`leaveTiles` or a `#session=` link (which flips `open` only, so a gone id still
+frees its cell), Home, the width gate, the last tile, "Open group as tiles",
+closing or killing sessions. Nothing is written while a stored grid is being
+put back, so a half-built grid never overwrites it.
+
+A pure sanitizer drops unknown, deleted, detached and duplicate ids on load,
+reports the cells their sessions freed (`freed`), and derives `count` for a
+value written before it existed (the number of sessions the cells name); the
+old packed `ids` (no nulls) read as cells with no hole, and anything that is not
+a v1 object is ignored. The format stays `v: 1`, so an older build still reads
+a newer value (it ignores `count`). Never read or written in a solo window.
 
 The restore runs INSIDE `handleInit`, in place of its initial
 `selectSession(restoreId, { auto: true })` (the non-`keepTerminal` branch), not
@@ -409,6 +439,17 @@ the main terminal never loads on that page load. A later `handleInit` (SSE
 reconnect after a server restart, the `keepTerminal` branch) reconciles ids
 against the live list without rebuilding tiles that are still alive.
 
+A cell freed since the grid was stored is filled during that restore, from a
+ranking that knows each session's status and stamps (the init payload) but not
+yet its pending approvals: `seedApprovals` asks the server for them
+asynchronously, and the restore has run by the time they land. So on a reload a
+session waiting on a permission dialog or an unseen finished turn ranks with the
+quiet ones for that one fill (working sessions still rank first). Accepted: a
+fill held back for the approvals would open fewer tiles, which can be another
+shape, and then reshape the grid and move the user's tiles a second after the
+reload; so approvals that land later never re-form a restored grid. The Tiles
+toggle, run once the page has loaded, ranks with them.
+
 ### Gating
 
 - Setting `showTileGridButton`, per device (in `displayKeys`, stripped from the
@@ -815,7 +856,8 @@ Separate follow-up PRs worth doing (see "Follow-ups").
   open, and an open menu owns the Escape (it closes alone and the keyboard goes
   back to the Tiles button, like the tab-group menu).
 - **User text** (names) via `textContent` / attributes, never `innerHTML`.
-- **No secrets in localStorage**: the stored grid holds ids only.
+- **No secrets in localStorage**: the stored grid holds session ids and its
+  layout only, never content.
 - **Memory**: everything a tile creates is released in `destroy()`.
 
 ## Delivery: two PRs
@@ -1008,13 +1050,17 @@ exits green. Use the browser runner for those files and read the file count.
    viewer keeps a stale width and renders garbled output (#464).
 3. WebSocket backpressure (`bufferedAmount` threshold, drop and send `{t:'r'}`
    on drain) for grids over slow links.
-4. Tile parity extras: mouse-wheel forwarding for Claude's fullscreen renderer,
-   a "Load full history" action inside a tile. (Done since: a tile pages a
-   hollow buffer's CLI transcript with PageUp/PageDown, the primary pane's
-   #555 route, and hand-reports a plain click while its session has
-   `cliMouseTracking` on, both through the primary pane's gates aimed at the
-   tile. The SGR wheel forwarding itself is still open: a fullscreen Claude
-   tile leaves the wheel to xterm.)
+4. Tile parity extras: a "Load full history" action inside a tile. (Done
+   since: a tile pages a hollow buffer's CLI transcript with PageUp/PageDown,
+   the primary pane's #555 route, hand-reports a plain click while its session
+   has `cliMouseTracking` on, and forwards the wheel to Claude's fullscreen
+   renderer as SGR wheel reports from its own cells
+   (`TerminalTile._maybeForwardWheelToCli`, encoding shared with the primary
+   pane via `CodemanTerminalInput.sgrWheelReports`), all through the primary
+   pane's gates aimed at the tile. Before that, a fullscreen Claude tile left
+   the wheel to xterm, which scrolled only stale replayed frames. Shift+wheel
+   scrolls the tile's local scrollback itself (`_maybeScrollLocalOnShift`),
+   since xterm turns it into a horizontal no-op off macOS.)
 5. WebGL in tiles, after measuring the DOM renderer with nine busy tiles.
 6. Named grid presets, possibly per owner on the server.
 7. The end state: the main terminal becomes a 1x1 grid of `TerminalTile`,
@@ -1053,7 +1099,8 @@ exits green. Use the browser runner for those files and read the file count.
    is on right-click of the button (its title says so, as do the wiki and the
    Help modal). Superseded in part by decision 10: right-click is now the count
    menu, and a remembered grid is filled to the count instead of opening
-   exactly as stored.
+   exactly as stored; decision 11 then restored "exactly as stored" and put a
+   ranking in place of the tab order.
 9. **No + in the tile header.** Decided by the owner ("remove the + button from
    these views"): the header is `● name ……… ⋯ ⤢ ×`. The + menu and its "New
    session in this case" went with it. Tiles are added from the Tiles button
@@ -1066,12 +1113,32 @@ exits green. Use the browser runner for those files and read the file count.
    right-click menu offers 2, 4 and 6, remembered per device; the session
    picker is gone, and decision 8's "picker on right-click" is superseded. The
    owner's answers on the details: the count wins over a remembered grid's
-   size (its tiles first, in their cells, holes filled first, then tab order);
+   size (its tiles first, in their cells, holes filled first, then tab order;
+   superseded by decision 11: a click brings the remembered grid back as it
+   was, and only a count picked in the menu re-forms it);
    Ctrl/Cmd+click with the grid closed opens the count in total, that session
    focused; shrinking keeps the focused tile; only the toggle animates the
    close; a remembered count larger than the window stays checked but greyed
    and a click opens what fits; the close keeps its dimmed still until the
    single view has painted (at most 700 ms); paced connect is in.
+11. **The grid keeps the layout the user arranged, and a fresh one ranks by
+   work.** Decided by the owner ("when I hit the tiles button, it should prefer
+   to load in tiles that are working and then the most recent working, so the
+   oldest dont get opened ... when I moved around and modified it, save it per
+   browser the layout, so when I turn tiles off and on, always keep what the
+   last setting was, if there was no setting before take the working ones, that
+   ones needs input and then the most recent ones in order"). The layout
+   (cells and holes, tile count, divider sizes, focus, a zoom the user chose)
+   is saved per browser on every change and comes back exactly from the toggle,
+   however the grid closed, and from a reload when the grid was open (a grid
+   closed before the reload stays remembered for the toggle; the page shows the
+   single view); it is never filled to the remembered count nor trimmed to the
+   window. A session gone since frees its cell for the
+   ranking; with none left, the grid opens from the ranking (`rankTileSessions`:
+   working, then needing input, then most recent), which also fills every place
+   the grid fills on its own (a count picked in the menu, a freed cell, an open
+   split's fill). Supersedes decision 10's "the count wins over a remembered
+   grid's size"; the count menu itself, its counts and its other answers stay.
 
 ## Code anchors
 

docs/usage-limits-display-plan.md

@@ -46,6 +46,8 @@ Claude Code (**v2.1.80+**; prod box runs **2.1.177**) pipes a JSON blob to a con
 
 The same stdin object also carries `model.display_name`, `context_window.{used_percentage, total_input_tokens, total_output_tokens, …}`, `cost.total_cost_usd`, `effort.level`, etc. The shipped feature uses **model + token totals + context %** to build the in-terminal footer (so the statusline stays useful even though we own it). The endpoint also broadcasts `contextUsedPercentage`/`costUsd`/`modelDisplayName` alongside the limits for future chip tooltips.
 
+Since Claude Code v2.1.251 the payload also carries a `prompt_cache` object for the main conversation: `warm`, `ttl` (`5m` or `1h`), `expires_at` (epoch seconds), `recache_tokens_if_cold`, `misses`, `expected_rebuilds`, `hit_ratio`, `last_miss_at` and `last_miss_cause`. `StatusTelemetrySchema` lists every documented scalar field null-tolerant (several are documented as possibly null, and the object gains fields across CLI versions, so a strict shape would 400 the post and blank the footer); `parsePromptCache()` normalizes only the subset Codeman stores. The footer appends `cache:until HH:MM` while warm or `cache:cold` otherwise, and Claude re-runs the statusline when `expires_at` passes, so the cold flip arrives without a turn. The Session Options readout adds the hit rate, the last miss's causes, and a `Context compacted at HH:MM` note after a compaction or a tool-result clearing (the first warm report whose `recache_tokens_if_cold` is null after one that carried it; `expected_rebuilds` rises only at the next request, so it is not read; never raised as an alert). The last report is restored on a Codeman restart and the readout is time-aware, so a value whose expiry has passed shows as cold rather than stale-warm.
+
 ### Alternatives considered & rejected
 
 | Source | Why not |

docs/wiki/Agent-CLIs.md

@@ -72,12 +72,12 @@ output. The other CLIs expose no equivalent.
 | Precise idle detection                            | Yes    | Codex, Pi, OpenCode, OMP and Gemini: same screen check, via their own prompt and working line. DeepSeek: reports its state itself. Others: output stabilization, coarser |
 | Auto-resume when a usage limit resets             | Yes    | No                                                   |
 | Plan usage chip                                   | Yes    | No                                                   |
-| Approvals Inbox                                   | Yes    | DeepSeek yes; others no                              |
+| Approvals Inbox                                   | Yes    | DeepSeek and Codex as alerts (answer in the terminal); others no |
 | Read My Mind                                      | Yes    | No                                                   |
 | Ralph loop and its task tracker                   | Yes    | No                                                   |
 | Subagent and team windows                         | Yes    | No                                                   |
 | Model, effort, advisor, and ultracode controls    | Yes    | No                                                   |
-| `stop` and `blocked` wait signals                 | Yes    | DeepSeek yes; elsewhere 400 if you ask for them explicitly |
+| `stop` and `blocked` wait signals                 | Yes    | DeepSeek yes; Codex `blocked` only; elsewhere 400 if you ask for them explicitly |
 | The bundled agent skill                           | Yes    | No                                                   |
 
 Everything that makes a session a session works everywhere. What is Claude-only is mostly
@@ -153,6 +153,17 @@ Two behaviours that are deliberate and worth knowing:
   does; before 1.26.1 every Codex session reported idle for its whole life. Codex
   conversations also appear in Past Sessions and can be resumed, and on phones the keyboard
   bar grows `⇧←` / `⇧→` for Codex's queued-message editing and prompt stack.
+- **A dialog waiting for you turns the tab red.** Codex is launched with its approvals on
+  unless you tick the bypass option, and it has no hooks to say it stopped, so Codeman reads
+  its dialogs off the screen: a command, edit or network approval, a question it asks you
+  (Plan mode), and the "Trust this folder?" prompt when it starts in a new folder. Each one
+  raises the same red "needs you" alert a Claude permission prompt does, with the desktop
+  notification, web push, webhook, the phone overview's NEEDS YOU row and an Approvals Inbox
+  card, and the alert clears a few seconds after you answer. Answer in the terminal: Codeman
+  never answers a Codex dialog for you, so the inbox card and the push carry no Approve or
+  Deny buttons. While such a dialog is up the session shows as idle (a question used to show
+  as working), and agents can wait on it with the `blocked` signal. "Implement this plan?"
+  at the end of a plan is not counted, since the turn has ended there.
 
 ### Gemini
 

docs/wiki/Codeman-Skill.md

@@ -0,0 +1,166 @@
+# The Codeman Skill
+
+The codeman skill lets the agent inside a Codeman session drive Codeman itself. You ask in
+plain English, and the agent starts worker sessions in any installed CLI, hands each one its
+task, waits for them, reads their answers and cleans up after itself. No curl, no endpoint
+names, no session ids in your prompt.
+
+It is one skill for every agent CLI, in the `SKILL.md` format that Claude Code, Codex,
+Gemini CLI, OpenCode, Grok, OMP and most other agent CLIs read. It lives in the repo at
+[`skills/codeman`](https://github.com/Ark0N/Codeman/tree/master/skills/codeman).
+
+## See it work
+
+Three real runs on a live Codeman: real sessions, real agents, one typed prompt each.
+Nothing on screen is mocked up; the long waits are sped up and labelled as time-lapse.
+
+### Five workers in the tile grid
+
+The lead Claude Code session gets `use codeman to start 5 claude workers, one per TODO item`,
+and five Claude Code workers power on, one per item, with all six agents working at once.
+
+![A real codeman skill run in the tile grid: the lead Claude Code session powers on as the only tile, the prompt use codeman to start 5 claude workers, one per TODO item is typed, then five Claude Code worker sessions power on as new tiles one after another with the CRT animation, and all six agents work at once](https://raw.githubusercontent.com/Ark0N/Codeman/master/docs/images/codeman-skill-grid-20261015.gif)
+
+Sessions an agent starts do not join the grid by themselves. In this recording each worker
+was put on the grid as it started, the way `Ctrl`+click or dragging its tab does it: see
+[Tile Grid](Tile-Grid).
+
+### Eight workers, eight tabs
+
+The same prompt with eight workers, in the normal view. Every worker gets its own tab, and a
+lineage line runs from the lead to each one, so you can see who spawned whom. Clicking a
+worker's tab shows it at work.
+
+![A real codeman skill run in the normal view: the prompt use codeman to start 8 claude workers, one per TODO item is typed, eight worker tabs power on left to right while lineage lines draw from the lead to each one, then clicking a worker tab shows a live Claude Code worker receiving its task](https://raw.githubusercontent.com/Ark0N/Codeman/master/docs/images/codeman-skill-tabs-20261015.gif)
+
+### Two CLIs from one prompt
+
+Workers do not have to run the same CLI as the lead. Here the prompt is
+`use codeman to have a dsh worker on qwen build search and a claude one build checkout`: a
+DeepSeek Harness worker on a local qwen model and a Claude Code worker power on side by side.
+
+![A real codeman skill run: one short prompt typed into Claude Code, the tile grid powering on, then a DeepSeek Harness worker on a local qwen model and a Claude Code worker powering on as new tiles, with lineage lines from the lead to both](https://raw.githubusercontent.com/Ark0N/Codeman/master/docs/images/codeman-skill-crt-20261010.gif)
+
+## Install it
+
+Pick one route. The skill only acts inside a Codeman session, so a global install is safe.
+
+| How                | Command                                                                         | Scope |
+| ------------------ | ------------------------------------------------------------------------------- | ----- |
+| Skills CLI         | `npx skills add Ark0N/Codeman --skill codeman -g`                               | Global, any skills-aware agent. The way to give it to pi. |
+| Claude Code plugin | `/plugin marketplace add Ark0N/Codeman`, then `/plugin install codeman@codeman` | Global, through Claude Code's plugin manager. `/plugin update codeman` follows releases. |
+| Bundled CLI        | `codeman skill install`                                                         | Global, at `~/.claude/skills/codeman`. |
+| Bundled CLI        | `codeman skill install --case <name> [--mode <mode>]`                           | One case. `--mode codex` (or any other run mode) writes it where that CLI looks for skills. |
+| Web UI             | **App Settings → Agents & CLIs → All agent CLIs → Codeman Agent Skill**         | Injects it into each case when a session is created, where that session's CLI looks: `.claude/skills/` for Claude Code, `.agents/skills/` for Codex, Gemini CLI, OpenCode, Antigravity, Grok, DeepSeek and OMP. Off by default. |
+
+Use the plugin **or** `codeman skill install`, not both: a Claude Code with both lists the
+skill twice (`codeman` and `codeman:codeman`).
+
+The web UI route skips pi, because a project skill makes pi ask to trust the folder at
+startup. A pi or plain shell session uses the [`codeman agent`
+commands](Driving-Codeman-From-An-Agent#the-codeman-agent-commands) instead, which are the
+same verbs as shell commands.
+
+### Remove it
+
+`codeman skill uninstall [--case <name>]` reverses the CLI installs (with `--case`, every
+Codeman-managed copy in that case, whichever CLI's directory it sits in). It never touches a
+`skills/codeman` you wrote yourself.
+
+Turning the web UI setting back off **does not remove copies it already injected**, because
+a sweep at session start would pull the skill out from under other live sessions sharing
+that directory. Remove them per case with `codeman skill uninstall --case <name>`.
+
+## Then just ask
+
+These prompts work as written:
+
+| You say                                                                                   | What happens |
+| ----------------------------------------------------------------------------------------- | ------------ |
+| "What sessions are running right now?"                                                     | Lists them with name, mode and status. Read-only. |
+| "Use codeman to start 5 claude workers, one per TODO item."                                | One worker per item, each handed its task, answers gathered as they finish. |
+| "Start a shell worker on the `myapp` case, run the test suite, tell me if it passes."      | Spawns, waits on a completion marker, reads the exit code, cleans up. |
+| "Spin up 3 workers for lint, typecheck and tests, run them in parallel, report failures."  | One session per task, all started first, then gathered as each finishes. |
+| "Have a claude worker summarize `src/session.ts`, then close it."                          | Spawns, waits until the worker is ready, sends and waits, reads the answer, deletes the session. |
+| "Have a DeepSeek Harness worker build search and a claude worker build checkout."          | A mixed fleet: one worker per CLI, started in one go. |
+| "Watch session w4 and tell me if it gets stuck on a permission prompt."                    | Blocks on the `blocked` signal and brings the question to **you**. |
+
+Sessions the agent creates are deleted when it is done, so you can watch the tabs appear and
+disappear in the dashboard while it works.
+
+### A run, start to finish
+
+> **You:** spin up 3 shell workers, run lint / typecheck / the frontend syntax check in
+> parallel, and tell me which failed.
+
+```text
+lint       -> 9f2d8e5f   dispatched
+typecheck  -> aff9c691   dispatched     3 tabs appear in the dashboard
+syntax     -> be9f1f15   dispatched
+
+lint         DONE_lint_17909      rc=0
+typecheck    DONE_typecheck_3409  rc=0   gathered as each one finishes
+syntax       DONE_syntax_18501    rc=0
+
+deleted 9f2d8e5f, aff9c691, be9f1f15    tabs disappear
+```
+
+The `DONE_<task>_<random>` strings are the
+[split-marker trick](Driving-Codeman-From-An-Agent#the-split-marker-trick): a shell session
+has no "turn finished" signal, so the skill waits for a marker in the output, typed in two
+halves so that only the command's real output ever contains the joined string.
+
+## Which workers it can drive
+
+Any run mode installed on the machine can be a worker. Claude Code and DeepSeek Harness are
+the two with real completion signals (Claude's from its hooks, DeepSeek's from the harness),
+so the skill can wait for "turn finished" and read the answer as clean text. Every other CLI
+and plain shell sessions are synchronized on an output marker instead, which works the same
+from your side. See [Agent CLIs](Agent-CLIs) for the run modes.
+
+With Claude Code workers, where the installed Claude Code supports it, the lead can also
+message a worker directly instead of typing into its terminal.
+
+## What it will and will not do
+
+- **It self-gates.** Outside a Codeman session (`CODEMAN_MUX` unset) it refuses to act and
+  never guesses an API URL, so a global install costs an unrelated agent session nothing.
+- **Unprompted, it may only** start sessions, prompt them, and delete the ones **it created
+  in that conversation, by exact id**, behind a guard that refuses to delete the agent's own
+  session.
+- **It will not** answer another session's permission prompt on your behalf. It brings the
+  question to you instead.
+- **Only when you ask, naming the target:** deleting a case (which erases a real directory
+  of your code), bulk kills, respawn, Ralph, cron, the orchestrator, and settings changes.
+
+## When the agent is Codex
+
+Two things to know:
+
+- Codex's default sandbox blocks network access, even to this machine, so the skill needs
+  **Bypass approvals and sandbox** (App Settings, Codex).
+- Codex 0.162 runs its commands in a shared background process, whose Codeman variables can
+  belong to the session that started it.
+
+The skill checks for both and says so instead of acting.
+
+## What is in the skill
+
+| File | Contents |
+| ---- | -------- |
+| [`SKILL.md`](https://github.com/Ark0N/Codeman/blob/master/skills/codeman/SKILL.md) | Safety rules, the fast path (start N workers, task them, collect the answers) and the verb index. Always loaded. |
+| [`reference/verbs.md`](https://github.com/Ark0N/Codeman/blob/master/skills/codeman/reference/verbs.md) | Every verb in detail: readiness, send-and-wait, markers, interrupts, cleanup. Loaded on demand. |
+| [`reference/recipes.md`](https://github.com/Ark0N/Codeman/blob/master/skills/codeman/reference/recipes.md) | Worked flows: claude, DeepSeek Harness and shell workers, fan-out, watching a blocked worker, messaging. Loaded on demand. |
+| [`reference/endpoints.md`](https://github.com/Ark0N/Codeman/blob/master/skills/codeman/reference/endpoints.md) | Endpoint tables, error codes, the signals each mode emits, capacity limits. Loaded on demand. |
+| [`reference/messaging.md`](https://github.com/Ark0N/Codeman/blob/master/skills/codeman/reference/messaging.md) | Talking to claude workers directly through Claude Code's cross-session messaging. Loaded on demand. |
+
+Only `SKILL.md` sits in the agent's context; the references load when a job needs them, so a
+quick run stays quick.
+
+## Read next
+
+- [Driving Codeman From An Agent](Driving-Codeman-From-An-Agent) - the `codeman agent`
+  commands, the MCP server, and the raw HTTP rules the skill is built on.
+- [Tile Grid](Tile-Grid) - watching several workers side by side.
+- [The Dashboard](The-Dashboard) - tabs, lineage lines and status dots.
+- [Watching Agents Work](Watching-Agents-Work) - the subagents inside each session.

docs/wiki/Contributing.md

@@ -63,9 +63,8 @@ provide what they need; that means "not runnable here", not a regression.
 Tests are tmux-safe by design: under vitest the tmux layer becomes an in-memory mock, so
 tests cannot touch real sessions. If you add a test that binds a port, bind port 0
 (`new WebServer(0, …)` + `server.boundPort`, or `listen({ port: 0 })` + `address().port`),
-or use `app.inject()` when no socket is needed. Never 3000. Mobile tests (`test/mobile/**`,
-via `createTestServer(PORT)`) keep the fixed ports in `test/mobile/README.md` for now,
-because that helper caches servers by port.
+or use `app.inject()` when no socket is needed. Mobile tests call `createTestServer()` and
+read `server.boundPort`. Never 3000.
 
 ## Finding your way around
 

docs/wiki/Core-Concepts.md

@@ -46,7 +46,8 @@ case. Each has a stable id, and that id is what the API, the wait primitives, an
 event use.
 
 Several sessions can share one case. That is the normal way to parallelize: three workers
-in the same repo, three tabs, one case.
+in the same repo, three tabs, one case. They then also share one folder; when the workers
+edit and commit in parallel, give each its own checkout with [Git Worktrees](Git-Worktrees).
 
 A session carries state the case does not:
 

docs/wiki/Cron-Jobs.md

@@ -4,8 +4,8 @@ Saved, named jobs that start a session and send it a prompt on a schedule. Cron
 sessions: *every weekday at 03:00, open a Claude session in `~/proj` and tell it to update
 dependencies and open a PR.*
 
-The ⏰ **Cron** header button is opt-in. Turn it on in
-**App Settings → Header & Panels**.
+The ⏰ **Cron** button in the bottom toolbar is opt-in. Turn it on under
+**App Settings → Header & Panels → Scheduling**.
 
 ## Creating a job
 
@@ -96,11 +96,19 @@ The skip policy has the details you would want it to have:
 
 Every fire is recorded per job, with a status:
 
-| Status    | Meaning                                                              |
-| --------- | -------------------------------------------------------------------- |
-| `created` | The run started and a session was created.                            |
-| `skipped` | The concurrency policy blocked it. Not counted as a run.              |
-| `failed`  | The prompt could not be resolved, or the working directory was gone.  |
+| Status            | Meaning                                                           |
+| ----------------- | ----------------------------------------------------------------- |
+| `created`         | The run record was created; the session has not started yet.       |
+| `session_started` | The session started; prompt delivery is still pending.             |
+| `prompt_sent`     | The prompt was sent to the session.                                |
+| `skipped`         | The concurrency policy blocked it. Not counted as a run.           |
+| `failed`          | Prompt resolution, session launch, or prompt delivery failed.      |
+
+History records whether the session started and the prompt was delivered, **not whether
+the agent's task succeeded**. `prompt_sent` is not a task-completion signal.
+If the session is closed during the readiness wait or the server restarts before
+delivery, a run can remain `session_started` indefinitely with `finishedAt: null`;
+clients must not poll forever waiting for a terminal status.
 
 The schedule is advanced **before** the session launches, so a slow start cannot cause the
 same job to re-trigger.

docs/wiki/Driving-Codeman-From-An-Agent.md

@@ -4,61 +4,107 @@ Everything the dashboard does is HTTP, so an agent can do it too. This page is f
 that makes Codeman interesting: **Claude Code running inside a Codeman session, spawning and
 supervising other sessions.**
 
-Two routes. Start with the skill.
+Three routes. Start with the skill, whichever CLI the agent is. Where the skill does not
+reach (a pi or shell session), use the `codeman agent` commands. Raw HTTP is there for
+everything else.
 
 ## The agent skill
 
-A Claude Code skill that teaches the agent the whole API, so you ask in plain English
-instead of pasting endpoint documentation into prompts.
+A skill in the `SKILL.md` format that Claude Code, Codex, Gemini CLI, OpenCode, Grok, OMP
+and most other agent CLIs read. It teaches the agent the whole API, so you ask in plain
+English instead of pasting endpoint documentation into prompts:
 
-### Install it
+```bash
+codeman skill install    # or: npx skills add Ark0N/Codeman --skill codeman -g
+```
 
-| How          | Command                                                     | Scope                                                       |
-| ------------ | ----------------------------------------------------------- | ----------------------------------------------------------- |
-| Skills CLI   | `npx skills add Ark0N/Codeman --skill codeman -g`            | Global, any skills-aware agent.                              |
-| Claude Code plugin | `/plugin marketplace add Ark0N/Codeman`, then `/plugin install codeman@codeman` | Global, through Claude Code's plugin manager. `/plugin update codeman` follows releases. Pick this or `codeman skill install`, not both, or the skill is listed twice (`codeman` and `codeman:codeman`). |
-| Bundled CLI  | `codeman skill install`                                      | Global, at `~/.claude/skills/codeman`.                       |
-| Bundled CLI  | `codeman skill install --case <name>`                        | One case.                                                    |
-| Web UI       | **App Settings → Agents & CLIs → Claude → Agent Skill**      | Injects into each case when a Claude session is created. Off by default. |
+Then ask the agent in a Codeman session something like *"use codeman to start 5 claude
+workers, one per TODO item"*. It starts the workers, hands each one its task, gathers the
+answers and deletes the sessions it created.
 
-`codeman skill uninstall [--case <name>]` reverses the CLI installs, and never touches a
-`skills/codeman` you wrote yourself.
-
-### Then just ask
-
-| You say                                                                                 | What happens                                                                             |
-| --------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------ |
-| "What sessions are running right now?"                                                   | Lists them with name, mode, and status. Read-only.                                          |
-| "Start a shell worker on the `myapp` case, run the test suite, tell me if it passes."     | Spawns, waits on a completion marker, reads the exit code, cleans up.                       |
-| "Spin up 3 workers for lint, typecheck and tests, run them in parallel, report failures." | One session per task, all started first, then gathered as each finishes.                    |
-| "Have a claude worker summarize `src/session.ts`, then close it."                         | Spawns, runs the readiness ladder, sends and waits, reads the answer, deletes the session.  |
-| "Watch session w4 and tell me if it gets stuck on a permission prompt."                   | Blocks on the `blocked` signal and surfaces the question to **you**.                        |
-
-Sessions the agent creates get deleted when it is done. You can watch the tabs appear and
-disappear in the dashboard while it works.
-
-### What it will and will not do
-
-- **It self-gates.** Outside a Codeman session it refuses to act and does not guess an API
-  URL, so a global install costs an unrelated Claude Code session nothing.
-- **Unprompted, it may only** spawn sessions, prompt them, and delete ones **it created in
-  that conversation, by exact id**, behind a guard that refuses to delete the agent's own
-  session.
-- **It will not** answer another session's permission prompt on your behalf. It surfaces the
-  question instead.
-- **Deleting a case** (which erases a real directory of your code), bulk kills, respawn,
-  Ralph, cron, orchestrator, and settings writes all require you to ask, naming the target.
-
-Turning the setting back off **does not remove already-injected copies**, because a
-create-time sweep would yank the skill out from under other live sessions sharing that
-directory. Remove them per case with `codeman skill uninstall --case <name>`.
-
-The skill ships with the verb index always loaded, plus on-demand references for the verbs,
-worked multi-worker recipes, endpoint tables, and cross-session messaging. It drives
+**[The Codeman Skill](Codeman-Skill)** has the rest: recordings of real runs, every install
+route (Claude Code plugin, skills CLI, per case, the web UI setting), prompts that work as
+written, what it will and will not do on its own, and the notes for Codex. It drives
 DeepSeek Harness workers the same way it drives Claude ones (`spawn_workers alpha
 beta:deepseek` is a mixed fleet in one call), since those are the two modes with real
 completion signals.
 
+## The `codeman agent` commands
+
+The skill drives the API with `curl` and `jq`. Where it does not reach (a pi or shell
+session, a machine without `jq`), the agent runs in the same environment, so `codeman agent`
+packages the same verbs as shell commands. It is a thin client over the endpoints in
+[the manual path](#the-manual-path), so auth and ownership apply unchanged, and it refuses to
+act outside a Codeman session. One line in a case's `AGENTS.md` is enough: *"other sessions:
+`codeman agent --help`"*.
+
+```bash
+codeman agent ls                                          # sessions; * marks this one
+SID=$(codeman agent spawn scratch-1 --mode claude)        # quick-start + wait for the composer where the mode has a ready mark
+codeman agent send "$SID" 'review src/, then say DONE' --until stop,exit --timeout 300000
+codeman agent read "$SID"                                 # last answer (as the server reads it for that mode)
+codeman agent read "$SID" --tail 3000                     # terminal tail, ANSI stripped (every mode)
+codeman agent send "$SID" 'run the tests, then print WORKDONE followed by _4711'   # hook-less modes: the marker in halves …
+codeman agent wait "$SID" --match WORKDONE_4711           # … and the wait on the joined form
+codeman agent interrupt "$SID"                            # a bare ESC, conversation intact
+codeman agent rm "$SID"                                   # any session except this one
+```
+
+- **Ids** may be the 8-character form `ls` prints. Anything shorter refuses, and so does an
+  ambiguous prefix.
+- **`send`** takes ONE quoted argument of printable text and presses Enter. A prompt that
+  starts with `-` goes after `--`: `codeman agent send "$SID" -- "- fix the bug"`.
+- **Markers** follow [the split-marker trick](#the-split-marker-trick): the echo of your own
+  prompt is output too, so ask for the marker in halves and wait on the joined form.
+- **Exit codes** are the same for every verb: `0` done, `1` error, `2` timeout, `3` the
+  worker exited, `4` refused. `--json` prints the response's `data`.
+
+## The MCP server
+
+`codeman mcp` is a [Model Context Protocol](https://modelcontextprotocol.io) server, so any MCP
+client (Claude Code, Claude Desktop, Codex, Gemini CLI, Cursor) can drive Codeman, from
+inside a Codeman session or from an assistant on your desktop. It is a client of the same
+HTTP API, with seven tools: `list_sessions`, `start_session`, `send_prompt`,
+`wait_for_session`, `read_output`, `interrupt_session` and `delete_session`.
+
+```bash
+# Claude Code
+claude mcp add codeman -e CODEMAN_API_URL=https://127.0.0.1:3000 -- codeman mcp
+```
+
+```json
+{
+  "mcpServers": {
+    "codeman": {
+      "command": "codeman",
+      "args": ["mcp"],
+      "env": { "CODEMAN_API_URL": "https://127.0.0.1:3000" }
+    }
+  }
+}
+```
+
+```toml
+# Codex (~/.codex/config.toml): raise the 60 s default tool timeout for long waits
+[mcp_servers.codeman]
+command = "codeman"
+args = ["mcp"]
+env = { CODEMAN_API_URL = "https://127.0.0.1:3000" }
+env_vars = ["CODEMAN_MUX", "CODEMAN_SESSION_ID"]  # Codex passes MCP servers a minimal env; these let it know the Codeman session it runs in
+tool_timeout_sec = 120
+```
+
+The JSON form is what Claude Desktop, Cursor and Gemini CLI take. Add `CODEMAN_PASSWORD` to
+`env` when the server has a password.
+
+- **It never guesses a server.** Without `CODEMAN_API_URL` (or `--url`) every call says how
+  to configure it.
+- **Only its own sessions.** Prompting, interrupting and deleting are limited to sessions it
+  started; listing, reading and waiting work on everything. `codeman mcp
+  --allow-any-session` lifts that, for an assistant that should drive your existing tabs.
+- **Waits are short by default** (45 s, at most 600 s), so they fit a client's tool
+  timeout. A `timed_out` answer means call again.
+
 ## The manual path
 
 The same operations as raw HTTP, for a CI bot, a shell script, or an agent without skill
@@ -95,17 +141,20 @@ Read these before writing any code. Each one has cost somebody an afternoon.
 5. **Wait instead of polling, and a timeout is not an error.** The wait endpoints answer
    `200` with `wait.timedOut: true`. Loop over short waits rather than one long call, because
    tunnels cut idle connections.
-6. **Only `claude` and `deepseek` sessions emit `stop` and `blocked`.** Claude's come from
-   Claude Code hooks, DeepSeek's from the harness reporting its state to Codeman. Shell and
-   the other external CLIs accept only `idle`, `working`, and `exit`; asking for `stop`
-   explicitly there is a `400`, while omitting `until` is always safe. On a shell session
+6. **Only `claude` and `deepseek` sessions emit `stop`; `blocked` also fires for `codex`.**
+   Claude's come from Claude Code hooks, DeepSeek's from the harness reporting its state to
+   Codeman, and Codex's `blocked` from Codeman reading its approval, question and trust
+   dialogs off the screen (a wait that asks for `blocked` while such a dialog is already up
+   answers at once). Shell and the other external CLIs accept only `idle`, `working`, and
+   `exit`; asking for `stop` explicitly there (codex included) is a `400`, while omitting
+   `until` is always safe. On a shell session
    `idle` fires **once at startup and never again**, so synchronize hook-less sessions with an
    output marker instead.
 7. **Nothing reports "ready", so wait for it explicitly.** A new session answers
    `{"signal":"exit","immediate":true}` until its PID exists, and that means *not started*,
-   not *crashed*. A Claude worker in a fresh case then sits on the CLI's trust dialog; prompt
-   it there and the wait resolves on idle in about two seconds looking exactly like a finished
-   turn, while your text sits stuck in the dialog.
+   not *crashed*. A Claude worker in a fresh case then sits on the CLI's trust dialog, whose
+   highlighted option is `No, exit`: prompt it there and your Enter quits the CLI. Codeman
+   holds `idle` while the dialog is on screen, but that is not a readiness signal either.
 
 ### Recipes
 
@@ -196,6 +245,7 @@ failing the spawn, because a cosmetic field must never break a worker.
 
 ## Read next
 
+- [The Codeman Skill](Codeman-Skill) - the skill in action, its install routes and its limits.
 - [HTTP API](HTTP-API) - the endpoint map and the envelope.
 - [Hooks And Integrations](Hooks-And-Integrations) - events flowing the other way.
 - [Watching Agents Work](Watching-Agents-Work) - seeing the fan-out in the UI.

docs/wiki/FAQ.md

@@ -207,10 +207,13 @@ cover them properly.
 ### Can agents start and supervise other agents?
 
 Yes. Codeman ships an agent skill that lets an agent inside a session drive the HTTP API:
-list sessions, spawn workers, send prompts, and block until a worker's turn finishes. It is
-off by default and enabled per case.
+list sessions, spawn workers in any installed CLI, send prompts, and block until a worker's
+turn finishes. Install it once with `codeman skill install` or the Claude Code plugin, then
+ask in plain English, for example *"use codeman to start 5 claude workers, one per TODO
+item"*. The web UI setting that injects it into every new session's case is off by default.
 
-See [Driving Codeman From An Agent](Driving-Codeman-From-An-Agent).
+See [The Codeman Skill](Codeman-Skill), with recordings of real runs, and
+[Driving Codeman From An Agent](Driving-Codeman-From-An-Agent) for the other routes.
 
 ### Can I run a case in a container?
 

docs/wiki/Git-Worktrees.md

@@ -0,0 +1,73 @@
+# Git Worktrees
+
+Sessions in one case share its folder. That is fine for one agent, but three agents in the
+same checkout edit the same files, overwrite each other's builds, and commit each other's
+half-finished work. A **git worktree** gives each of them a checkout of its own: same
+repository, same history, a separate folder on a separate branch.
+
+Codeman creates worktrees for a case whose folder is a git repository, starts sessions in
+them, and removes them again when the work is done.
+
+## Where they live
+
+`<case>/.claude/worktrees/<name>`, on a new branch `worktree-<name>`. That is exactly where
+Claude Code's own `claude --worktree <name>` puts them, so a worktree made in Codeman is one
+Claude Code recognizes, and the other way round. Codeman's Recent Sessions and Session Manager
+lists mark a session that runs in one with a `⑂ <name>` pill.
+
+The main checkout does not show the worktrees as untracked: Codeman adds `/.claude/worktrees/`
+(and `/.claude/settings.local.json`, the hooks file it writes for every Claude session, plus the
+`codeman` folder it writes when **Codeman Agent Skill** is on) to the repository's local
+`.git/info/exclude` when git does not ignore them already. Without that, those files would make
+every worktree look modified and git would refuse to remove it. Your tracked `.gitignore` is
+never touched.
+
+## Creating one and starting a session in it
+
+1. Pick the case in the toolbar's case picker.
+2. Open the picker again and choose **Git worktrees…** (on a phone: the case sheet's
+   **Worktrees** button).
+3. Under **New worktree**, type a name (letters, digits, `.`, `_`, `-`, up to 64 characters).
+4. Pick the CLI under **Start sessions with** (Claude Code, Codex, a shell, any installed run
+   mode).
+5. **Create & start session**.
+
+The worktree is created from the case's current commit, and the session opens in it as a tab
+named `w1-<name>`. Under **Branch and start point** you can name the branch yourself or start
+from another branch, tag or commit. An existing branch is never reused or reset: pick another
+name.
+
+**Open** on a worktree's row starts another session in it, with the CLI chosen above.
+
+## Removing one
+
+Two ways, and both use `git worktree remove` **without** `--force`:
+
+- **Remove…** on the worktree's row in the dialog.
+- Closing a session that runs in a worktree: the Close Session dialog offers **Kill & remove
+  worktree**, which kills the session and then removes its worktree.
+
+git refuses to remove a worktree that still holds uncommitted or untracked files, and Codeman
+shows that refusal as git worded it and leaves everything in place. Commit, stash or delete the
+work first, then remove again. A worktree a live session still runs in is refused too.
+
+**Also delete its branch** deletes the branch with `git branch -d`, which keeps a branch whose
+commits are not merged yet. The worktree is still removed, and the message says the branch was
+kept.
+
+If a worktree's folder was deleted by hand, its row says **missing** and offers only
+**Remove…**: that clears the record git keeps of it, so the name can be used for a new worktree.
+
+## What is not supported
+
+- **Docker cases and remote (SSH) cases.** A container can write to its workspace, and creating
+  a worktree runs the repository's hooks and checkout filters on the Codeman machine, so git is
+  never run there. A remote case's repository is on another machine. For the same reason a
+  worktree that a Docker case uses as its workspace (or that holds one) is never removed from
+  here: remove the Docker case first.
+- A case folder that is not the top of a repository (a subfolder of one), a bare repository, or
+  an empty repository with no commit yet.
+- Worktrees outside `.claude/worktrees/` (made by hand with `git worktree add`) are listed, but
+  Codeman neither starts sessions in them from the dialog nor removes them.
+
+The HTTP endpoints are in the [HTTP API](HTTP-API) reference (`/api/cases/:name/worktrees`).

docs/wiki/HTTP-API.md

@@ -72,7 +72,7 @@ Roughly 235 handlers across 26 route modules. By domain:
 | ------------------- | -------- | --------------------------------------------------- |
 | System              | 56       | Status, settings, digest, updates, tunnel.           |
 | Sessions            | 34       | Create, input, terminal, wait, last response, kill.  |
-| Cases               | 34       | Create, link, clone, remote and docker cases.        |
+| Cases               | 37       | Create, link, clone, worktrees, remote and docker cases. |
 | Files               | 17       | Preview, edit, raw, attachments, path picker.        |
 | Orchestrator        | 10       | Plans and phases.                                     |
 | Ralph               | 9        | Loop control and configuration.                       |
@@ -92,6 +92,32 @@ Roughly 235 handlers across 26 route modules. By domain:
 
 Each route module documents its own endpoints in its file header.
 
+## Cron jobs
+
+Saved scheduled jobs are distinct from the legacy `/api/scheduled` loops. Their
+versioned endpoints are:
+
+| Method | Path | Purpose |
+| --- | --- | --- |
+| GET | `/api/v1/cron/jobs` | List jobs |
+| POST | `/api/v1/cron/jobs` | Create a job |
+| GET | `/api/v1/cron/jobs/:id` | Read a job |
+| PUT | `/api/v1/cron/jobs/:id` | Update a job (partial body) |
+| DELETE | `/api/v1/cron/jobs/:id` | Delete a job |
+| PUT | `/api/v1/cron/jobs/:id/enabled` | Enable/disable with `{ enabled: boolean }` |
+| POST | `/api/v1/cron/jobs/:id/run` | Run now, without changing the schedule |
+| GET | `/api/v1/cron/jobs/:id/runs` | Read a job's run history |
+| GET | `/api/v1/cron/runs` | Read all accessible run history |
+
+Responses use the envelope above: job lists/history have arrays in `data`, a
+single-job GET has the job itself, create/update/enable have `{ job }`, delete has
+`{}`, and Run Now has `{ run, activeAgents }`. Prompt delivery is asynchronous;
+`session_started` and `prompt_sent` describe launch/delivery, not task success.
+
+The [cron API reference](https://github.com/Ark0N/Codeman/blob/master/docs/api-reference.md#cron-jobs)
+lists all request and response fields, validation, and ownership restrictions.
+[Creating a job](Cron-Jobs) covers the UI and schedule semantics.
+
 ## Long-polling instead of polling
 
 Three calls block until something happens, so an agent driving Codeman from a shell can wait
@@ -116,9 +142,10 @@ Three semantics that break callers who assume otherwise:
 `wait-output` matches a **literal substring, never a regex.** That is deliberate: no regex
 means no catastrophic backtracking on attacker-influenced output.
 
-Only `claude` and `deepseek` sessions emit `stop` and `blocked`: Claude's come from Claude
-Code hooks, DeepSeek's from the harness reporting its state to Codeman. Shell and the other
-external CLI sessions accept `idle`, `working`, and `exit`.
+Only `claude` and `deepseek` sessions emit `stop`: Claude's come from Claude Code hooks,
+DeepSeek's from the harness reporting its state to Codeman. `blocked` fires for those two
+and for `codex`, whose approval, question and trust dialogs Codeman reads off the screen.
+Shell and the other external CLI sessions accept `idle`, `working`, and `exit`.
 
 ## SSE
 

docs/wiki/Home.md

@@ -41,7 +41,8 @@ codeman web    # then open http://localhost:3000
 
 **Driving it from code**
 
-- [Driving Codeman From An Agent](Driving-Codeman-From-An-Agent) - the bundled skill, worker sessions, wait primitives.
+- [The Codeman Skill](Codeman-Skill) - one prompt, a whole crew: an agent starting and supervising worker sessions.
+- [Driving Codeman From An Agent](Driving-Codeman-From-An-Agent) - the `codeman agent` commands, the MCP server, wait primitives.
 - [HTTP API](HTTP-API) - the envelope, auth, the endpoint map, SSE events.
 - [Hooks And Integrations](Hooks-And-Integrations) - events flowing back into Codeman.
 
@@ -66,6 +67,7 @@ codeman web    # then open http://localhost:3000
 | [Agent CLIs](Agent-CLIs)                   | Which agent should this session run, and how do I set it up? |
 | [Custom Model Endpoints](Custom-Model-Endpoints) | How do I point a session at my own OpenAI-compatible endpoint? |
 | [Working With Files](Working-With-Files)   | How do I read, edit, and attach files?                     |
+| [Git Worktrees](Git-Worktrees)             | How do several agents work on one repository without sharing a folder? |
 | [Input And Voice](Input-And-Voice)         | How do I talk to an agent, including by voice?             |
 | [Mobile Guide](Mobile-Guide)               | How well does this work on a phone?                        |
 | [Keyboard Shortcuts](Keyboard-Shortcuts)   | What can I drive from the keyboard?                        |
@@ -101,6 +103,7 @@ codeman web    # then open http://localhost:3000
 
 | Page                                                              | What it answers                              |
 | ----------------------------------------------------------------- | -------------------------------------------- |
+| [The Codeman Skill](Codeman-Skill)                                | How do I get an agent to start and run workers for me? |
 | [Driving Codeman From An Agent](Driving-Codeman-From-An-Agent)    | How does an agent spawn and drive workers?   |
 | [HTTP API](HTTP-API)                                              | What can I call, and what comes back?        |
 | [Hooks And Integrations](Hooks-And-Integrations)                  | How do I wire Codeman into something else?   |

docs/wiki/Hooks-And-Integrations.md

@@ -22,8 +22,11 @@ terminal into something that can notify you.
 This is why several Codeman features are Claude-only. The one partial exception is DeepSeek
 Harness, whose terminal front door reports idle, working and blocked to Codeman over the
 harness's own supervisor contract, so it gets the hook-driven surfaces without any hook
-file. The other CLIs have no equivalent, so for them Codeman watches terminal output, which
-reveals that something happened but not what it was.
+file. Codex gets one of them from its screen: Codeman reads its approval, question and
+trust dialogs off the pane and raises the same red "needs you" alert a `permission_prompt`
+hook raises (see [Agent CLIs](Agent-CLIs#codex)). The other CLIs have no equivalent, so for
+them Codeman watches terminal output, which reveals that something happened but not what it
+was.
 
 ### How hooks get installed
 

docs/wiki/Keeping-Agents-Running.md

@@ -99,6 +99,14 @@ limit-paused. Without that, the next cycle would fire `/clear` and wipe the conv
 are waiting to resume. This is the single most useful setting for overnight runs on a
 subscription plan.
 
+## The prompt cache readout
+
+**Claude only.** On the Respawn tab, above the loop controls.
+
+Claude Code bills a turn against the cached conversation while that cache is warm, and lets the cache expire after a fixed idle time: one hour on a subscription within plan usage, five minutes on an API key, a cloud provider or extra-usage credits. The first turn after a longer break re-processes the whole context at the cache-write rate, which is why coming back to a big session after lunch costs more than the turns around it.
+
+Claude Code 2.1.251 and newer reports the cache state on the statusline after every response, and Codeman's statusline exporter keeps it per session. The row shows **Warm until HH:MM** with the lifetime and the tokens that would be re-cached if it went cold, or **Cold** with the size of the re-cache the next turn pays. When Claude reports them it also shows the session's cache hit rate and the cause of the last cache miss (for example `ttl_expired_1h` or `tools_changed`), so a surprise miss has a reason next to it. If the conversation is compacted while it stays warm (Claude Code now does this on its own on large idle sessions, a manual `/compact` does it too, and so does a clearing of old tool results), the row notes **Context compacted at HH:MM** until the next turn: Claude Code reports no re-cache size until then. None of this raises an alert. The row refreshes live and flips to cold on its own when the lifetime passes, and the last state is kept across a Codeman restart so it is not blank until the session's next turn. Sessions without a statusline of their own also get a short `cache:until HH:MM` or `cache:cold` group in the terminal footer. The readout rides the plan usage chip's telemetry collection, which is on by default on desktop.
+
 ## The plan usage chip
 
 **Claude only.** A header chip showing live subscription usage, on by default on desktop and

docs/wiki/Keyboard-Shortcuts.md

@@ -67,6 +67,13 @@ default. See [Tile Grid](Tile-Grid).
 
 **App Settings → Shortcuts.** Bindings live in a registry with per-user overrides, so a
 rebind is stored as an override on top of the default rather than replacing the table.
+Previous Session (`Alt+[`) and Next Session (alternate, `Alt+]`) are rows like any other,
+so they can be rebound too.
+
+A chord another action already uses is refused, and the message names that action. The same
+check runs when you re-enable a disabled shortcut and when you press Reset: if its chord was
+given to something else in the meantime, it stays as it was. `Alt+1` to `Alt+9`,
+`Shift+Enter`/`Ctrl+Enter` and `Ctrl+V`/`Cmd+V` are reserved.
 
 Two things are deliberately not rebindable:
 
@@ -79,8 +86,11 @@ Two things are deliberately not rebindable:
 
 The terminal sees keystrokes before the app does. Any chord the app claims has to also be
 swallowed at the terminal layer, or xterm writes the control byte into the session as well
-as triggering the action. If you rebind something to a chord the terminal cares about
-(`Ctrl+D`, say), expect the CLI to see it too.
+as triggering the action. A chord **you** rebind an action to is kept out of the terminal
+too, so rebinding Previous Session to `Alt+J` switches the session and sends nothing to the
+CLI. The built-in defaults are deliberately not held back this way: `Ctrl+L` clears the
+screen and also reaches the CLI, which redraws on it. Binding an action back onto its own
+default behaves like the default.
 
 `Alt+1` through `Alt+9` are matched on **physical key position** rather than the character
 produced, so macOS Option layouts that produce `¡™£` still switch tabs.

docs/wiki/Mobile-Guide.md

@@ -66,7 +66,8 @@ autocorrect: Enter adds a new line, and only Send delivers the text, as one past
 by Enter, so your line breaks reach the agent intact. Anything already typed on the terminal
 prompt moves into the editor when it opens. Drafts are kept per session and in memory only,
 so switching tabs keeps them and a page reload forgets them; a dot on the key shows a draft
-is parked. The editor's Image button attaches photos and puts their paths into the draft.
+is parked. The editor's Attach button attaches photos, videos or documents and puts their paths
+into the draft.
 Destructive commands need a double press, so you cannot fire `/clear` with a stray thumb. On
 Codex sessions the bar also shows `⇧←` and `⇧→`, the Shift-modified arrows Codex binds to
 editing the last queued message and walking the prompt stack.
@@ -100,13 +101,41 @@ looks exactly like a dead button.
 On phones this button replaces the desktop's **Run Shell** control; starting a shell moved
 into the Run dropdown.
 
+## The Attach button
+
+The toolbar's **Attach** button (the picture icon beside the mic) is the keyboard-down way to
+hand the agent a photo, a video or a document. One tap opens the system's sheet (camera,
+photo library, or a file from Files or a cloud drive), nothing else: no editor, no keyboard.
+The chosen files upload and their paths land on the prompt, so you then tap the terminal,
+write the rest of your message and press Enter. Videos (mp4, webm, mov, m4v, ogv)
+upload as recorded, up to 1 GB each by default (`CODEMAN_MAX_PASTE_VIDEO_BYTES` on the
+server changes it), so a long 4K clip takes a moment on the way up. Documents are PDFs,
+Office files (docx, pptx, xlsx), zips and plain-text formats such as txt, md, csv and json;
+a document keeps the plain-letters part of its name in the saved file, so the agent and you
+can tell the uploads apart. See [Working With Files](Working-With-Files) for where they land and how long they
+stay.
+
+If the keyboard was up when you tapped Attach, iOS closes it to show the photo sheet and does
+not bring it back, so the tap on the terminal is how you get it again. That is the system's
+behaviour around its own picker rather than a Codeman setting.
+
+The paths go onto the prompt as unsent text, exactly like something you typed, which means
+they travel with it. Tap **Compose** afterwards and the editor opens holding the paths *and*
+anything you had already typed, so you can write in front of a path, fix a word, and Send.
+**Use terminal keyboard** hands a single-line message back the other way, onto the prompt,
+so neither surface holds a half of your message. A multi-line message cannot make that trip
+(the prompt is a single line), so it stays parked as a draft and the dot on the Compose key
+shows it is waiting. The Compose editor's own Attach button remains for the times you are
+already writing in the editor.
+
 ## Tapping, links and copying
 
 - **Tap a link** in terminal output and it opens in a new tab. Same for a link in an agent's
   answer in the response viewer — it opens a tab rather than navigating the dashboard away,
   which on a phone would unload the whole session view.
-- **Tap a file path** an agent printed and the file-preview overlay opens; a log path opens the
-  log viewer. Works in scrolled-up transcript too.
+- **Tap a file path** an agent printed and the file-preview overlay opens (Markdown rendered,
+  HTML as a page); a log or other plain-text path opens the log viewer. Works in scrolled-up
+  transcript too.
 - A tap on the prose *beside* a link still places the cursor as usual, and a tap on a dialog's
   numbered choice still answers the dialog even when the row contains a path — the dialog wins,
   because on a phone it is the only interaction that matters.

docs/wiki/Multi-User-Mode.md

@@ -60,6 +60,36 @@ Admins see everything.
 Ownership threads through every list endpoint, the session lookup helper, the WebSocket
 layer, and file previews. A user cannot address another user's session even by id.
 
+## Server-wide settings are admin only
+
+There is one set of server settings (`~/.codeman/settings.json`) for everyone: the feature
+switches, the run mode, respawn presets, the case order, the Cloudflare tunnel, the default
+models. Every user can **read** them, but only an **admin** can change them. For a regular
+user the settings write routes answer `403 FORBIDDEN` (see the
+[HTTP API](HTTP-API) and [`docs/api-reference.md`](https://github.com/Ark0N/Codeman/blob/master/docs/api-reference.md)).
+A regular user reads them without what is the admin's: the Deepgram key an admin saved for
+voice dictation, and the admin's case names (case order, last used case).
+
+In the browser this means:
+
+- **App Settings → Save** still keeps a regular user's per-device settings (skin, fonts,
+  header buttons, keyboard bar and the rest of the per-device keys) on that device. If they
+  also changed a server-wide setting, one notice says it can only be changed by an admin,
+  and the setting keeps the server's value.
+- A regular user's run-mode pick and respawn presets are their own: kept in that browser
+  under their name, not sent to the server, and still theirs on the next visit (over what
+  an admin set). Their last used case is not recorded server-wide.
+- Reordering cases and switching the tunnel on or off are admin actions.
+- If the page could not tell yet who is signed in (a slow or failed sign-in check), the
+  server's refusal is read the same way: the same notice, and nothing more is sent.
+
+The operations that act on the whole server are admin actions too: updating Codeman,
+opening the multi-display window, regenerating the tunnel QR code (it clears every user's
+pending QR login) and cleaning up stale session state. **Revoke All** signs a regular user
+out of their own logins only, and the tunnel panel lists only their own.
+
+Per-user server settings are not part of this version.
+
 ## Safer defaults for regular users
 
 Non-admins get tighter defaults, and lifting them is an explicit per-user grant:

docs/wiki/Notifications-And-Approvals.md

@@ -34,8 +34,11 @@ reloading the dashboard while a permission dialog is blocking a session does not
 with a normal-looking tab.
 
 For Claude sessions, these come from Claude Code's hooks and are precise about *why* the
-session stopped; DeepSeek Harness sessions report the same states themselves. For the other
-CLIs there are no hooks, so you get the coarser output-based signal.
+session stopped; DeepSeek Harness sessions report the same states themselves. Codex has no
+hooks, but its dialogs are distinctive enough to read off the screen, so a Codex session
+waiting on an approval, a question or the "Trust this folder?" prompt turns red too, and
+clears a few seconds after you answer. For the other CLIs there are no hooks, so you get the
+coarser output-based signal.
 
 ## Window title and OS notifications
 
@@ -109,7 +112,9 @@ Behaviour worth knowing:
 ## The Approvals Inbox
 
 **Opt-in, off by default. Claude sessions, plus DeepSeek Harness sessions, whose terminal
-front door reports its prompts to Codeman.**
+front door reports its prompts to Codeman, and Codex sessions, whose dialogs Codeman reads
+off the screen.** DeepSeek and Codex cards are alerts: they say "Answer in the terminal" and
+carry no answer buttons, because Codeman cannot verify a keystroke into those dialogs.
 
 One queue of every prompt currently waiting on a human, across all your sessions, answerable
 in place. When you have eight workers running, this is the difference between checking eight
@@ -214,7 +219,8 @@ from the lock screen.
 - **iOS needs the home screen install.** A Safari tab will never receive push.
 - **The bell is invisible at zero.** That is deliberate, not a broken setting.
 - **Approvals need real signals.** They are built on hook events, which Claude emits and
-  DeepSeek Harness reports itself; the other CLIs do neither.
+  DeepSeek Harness reports itself, and on Codex's dialogs as its screen shows them; the
+  other CLIs give neither.
 - **A stale menu answer is refused, not sent.** If you answer a card for a dialog that has
   since gone away, Codeman declines rather than typing a digit into the composer.
 

docs/wiki/Settings-Reference.md

@@ -29,6 +29,10 @@ The practical rule: **appearance and input are per device, behaviour is shared.*
 did not follow you, it is in one of the first two rows, and you change it again on that
 device.
 
+In [Multi-User Mode](Multi-User-Mode) the server's settings are shared by every user, so
+only an **admin** can change them. A regular user's **Save** keeps their per-device settings
+as usual and tells them that server-wide settings can only be changed by an admin.
+
 ## App Settings
 
 ### Updates
@@ -90,7 +94,6 @@ every session or only the active tab.
 | Setting                | Notes                                                                                     |
 | ---------------------- | ----------------------------------------------------------------------------------------- |
 | Skin                   | Theme palettes, light ones included. Applied before first paint, so no flash of the wrong theme. |
-| Entrance Animations     | Per-surface animation styles for tabs, terminals, windows, and lineage lines. All default to the legacy no-animation behaviour. |
 | Display Name           | Your name in the UI. Cosmetic only; it never renames the package, CLI, API, or storage.    |
 | Interface Language     | English or Simplified Chinese. Per device.                                                 |
 | Session List Layout    | Header tab strip (default), a collapsible left sidebar, or the sidebar with detailed rows. See [The Dashboard](The-Dashboard#session-list-layout). |
@@ -99,11 +102,23 @@ every session or only the active tab.
 | State Order            | For *By state*: needs you on top (default) or at the bottom, right above the terminal. |
 | Vertical Rail Order    | *By activity* (default) sorts the rail the way the home screens are sorted; *Manual* keeps your tab order and drag-reordering. With *By state* or *By case* it orders the rows inside each section. |
 | Tall Tabs              | Taller tab strip.                                                                          |
+| CLI Logos on Tabs      | Each agent tab, and its row on the desktop home rail, shows the CLI's logo before the name. Off hides those logos on this device; the status dot and the shell's SH badge stay, and tiles, split headers and the Run menus keep their logos. On by default. |
 | Pop-out Button on Tabs | Adds the detach control to tabs, with a per-tab override.                                  |
 | Spawn Lineage Lines    | Lines from each tab to the sessions it spawned; the selected tab's family is drawn thicker. Desktop only, on by default. |
 | Auto-name Sessions     | Titles a new tab after its first prompt, keeping the case prefix (`w3-myapp: fix the login redirect`). Synced, off by default. See [The Dashboard](The-Dashboard#automatic-session-names). |
+| Spawned Tabs Follow Their Parent | A session started by another one (an agent's worker, a `parentSessionId` spawn) follows its parent's tab into its group and moves with it until you move it by hand; the row menu then says where it sits and offers *Follow parent again*. Off, it starts right after its parent and stays put. Decided when the tab is created, so flipping it leaves existing tabs alone. Synced, off by default. |
 | Overview Home Screen   | The phone home screen. On by default.                                                      |
 
+### Animations
+
+All per device, all off by default, applied as you pick them.
+
+| Setting                | Notes                                                                                     |
+| ---------------------- | ----------------------------------------------------------------------------------------- |
+| Entrance Theme         | One look for how new tabs, terminal panes, agent windows and their lines arrive (Terminal, Beam down, Launch, Soft focus, Quiet, Playful). Off by default. |
+| Tile Animations        | How tiles arrive when the tile grid opens and leave when it closes: fly out of their tabs, dealt from the Tiles button, CRT, beam down, cascade, pop or soft; each screen then plays the theme's terminal animation. Off by default (the grid's quick fade); picking a theme presets it. |
+| Animation Lab          | Opens the per-surface lab (the same as `?animlab=1`): every style side by side, with replay, stagger and speed. Closes settings first. |
+
 ### Models
 
 Claude model cards, the 1M context window switch, the thinking effort segment and the
@@ -135,14 +150,14 @@ instead of its native cloud backend. See [Custom Model Endpoints](Custom-Model-E
 | Allowed Tools                    | The list used by the explicit mode.                                                            |
 | Ralph / Todo Tracker             | Enables the Ralph loop surfaces.                                                               |
 | Agent Teams                      | Experimental teams. Also needs the CLI's own environment flag.                                 |
-| Codeman Agent Skill              | Injects the agent skill into new Claude sessions per case. Off by default. See [Driving Codeman From An Agent](Driving-Codeman-From-An-Agent). |
+| Codeman Agent Skill              | Under **All agent CLIs**. Injects the agent skill into each new session's case, where that session's CLI looks for skills (every CLI except Pi and a plain shell). Off by default. See [The Codeman Skill](Codeman-Skill). |
 | Remote auto-reconnect            | Reattaches dropped remote SSH sessions. On by default.                                         |
 | Nice priority / value            | Runs agent processes at a lower CPU priority.                                                  |
 | Default Codex model              | Model for new local Codex sessions; empty uses Codex's own config. Letters, digits, `.` `_` `-` `/` only. |
 | Default Codex reasoning effort   | Reasoning level for new local Codex sessions; empty uses Codex's own config.                   |
 | Bypass approvals and sandbox     | Starts new Codex sessions with `--dangerously-bypass-approvals-and-sandbox`. Read [Agent CLIs](Agent-CLIs) before enabling. |
 | Animated status effects          | Cosmetic.                                                                                      |
-| MCP server sync                  | Copies the MCP servers each installed, enabled CLI (Claude, Codex, Gemini, OpenCode, Antigravity) has into the others' own config files. Synced, off by default, admin only in multi-user mode. Turn it on and save, then **Preview** shows what would change and **Sync now** applies it. It only adds missing servers, keeps the previous file as `.codeman-bak`, and leaves a file that receives env values or headers readable by you only. A config dir moved by `CODEX_HOME`, `CLAUDE_CONFIG_DIR`, `XDG_CONFIG_HOME` or `GEMINI_CLI_HOME` in Codeman's own environment is followed. |
+| MCP server sync                  | Copies the MCP servers each installed, enabled CLI (Claude, Codex, Gemini, OpenCode, Antigravity) and GitHub Copilot CLI has into the others' own config files. Synced, off by default, admin only in multi-user mode. Turn it on and **Apply** or **Save** (Apply keeps Settings open), then **Preview** shows what would change and **Sync now** applies it. It only adds missing servers, keeps the previous file as `.codeman-bak`, and leaves a file that receives env values or headers readable by you only. A config dir moved by `CODEX_HOME`, `CLAUDE_CONFIG_DIR`, `XDG_CONFIG_HOME`, `GEMINI_CLI_HOME` or `COPILOT_HOME` in Codeman's own environment is followed. |
 
 ### Notifications
 
@@ -206,6 +221,8 @@ Some things are configured before the server starts, not in the UI:
 | `CODEMAN_ALLOW_UNAUTHENTICATED_NETWORK` | Acknowledges exposing the server with no password.                  |
 | `CODEMAN_BASE_URL`                  | Mounts Codeman under a sub-path behind a reverse proxy that forwards the prefix unchanged. See [Remote Access](Remote-Access). |
 | `CODEMAN_MAX_DOWNLOAD_BYTES`        | Cap on raw file bodies and downloads. 2 GB by default, `0` for none.    |
+| `CODEMAN_MAX_PASTE_IMAGE_BYTES`     | Cap on one pasted, dropped or attached image or document. 50 MB by default. |
+| `CODEMAN_MAX_PASTE_VIDEO_BYTES`     | Cap on one attached video (phone Attach, drop). 1 GB by default; videos stream to disk, so this bounds disk use, not memory. |
 | `CODEMAN_MAX_REMOTE_FILE_SSH`       | Concurrent ssh reads for files in remote cases. 4 by default.           |
 | `CODEMAN_PATH_PROBE_TIMEOUT_MS`     | How long a linked case's folder may take to answer before it is shown as unreachable. 1500 ms by default; raise it for a slow but healthy mount. |
 | `CODEMAN_PATH_PROBE_MAX_STALLED`    | Unanswered folder checks allowed to pile up before new ones are refused. 2 by default: one below the threadpool size minus one, so it follows `UV_THREADPOOL_SIZE` (4 unless set), and it is never allowed above that ceiling. A check you start by opening one case or session may use the one slot left above it. |

docs/wiki/The-Dashboard.md

@@ -29,7 +29,7 @@ Session List Layout** can move it into a vertical sidebar on the left instead, a
 | -------------------- | --------------------------------------------------------------------------------- |
 | **Header tab strip** | The default. One list in tab order unless you pick another [Tab layout](#tab-layouts); it scrolls sideways on a phone. |
 | **Left sidebar**     | A vertical list with a filter box and a live session count. `Alt+B` collapses it to a narrow rail that keeps the status dots and task badges visible. On a phone it is an off-canvas drawer rather than a docked rail. A detailed variant adds the home screen's per-session line (`created 3d ago · working 12m`) and a status pill. |
-| **Vertical rail**    | The strip turned vertical beside the terminal, resizable, with detailed rows by default. **Vertical Rail Order** sorts it by activity (blocked on you first, then longest running, then most recently quiet), the same order as the home screens; pick *Manual* to get your own order and drag-reordering back. **Tab groups:** pick *Move to new group* from a row's ⋯ menu (or Shift+F10 on it) to make the first one; a group header's menu (right-click, Shift+F10 or its ⋯ glyph) renames it (also F2), reorders or deletes it, rows move between groups from their own menu or by dragging with a mouse or pen, and a collapsed group stays collapsed on that device. Desktop and tablet only. |
+| **Vertical rail**    | The strip turned vertical beside the terminal, resizable, with detailed rows by default. **Vertical Rail Order** sorts it by activity (blocked on you first, then longest running, then most recently quiet), the same order as the home screens; pick *Manual* to get your own order and drag-reordering back. **Tab groups:** pick *Move to new group* from a row's ⋯ menu (or Shift+F10 on it) to make the first one; a group header's menu (right-click, Shift+F10 or its ⋯ glyph) renames it (also F2), reorders or deletes it, rows move between groups from their own menu or by dragging with a mouse or pen, and a collapsed group stays collapsed on that device. **Search sessions** at the top of the rail narrows it to the tabs whose name matches (a web tab by its title), across every group, collapsed ones included, without changing the groups or the order; a tab with an alert stays visible even when its name does not match; Escape or × clears it, and it is never saved. **Focus:** *Add to Focus* in a row's ⋯ menu (right-click on a web tab) pins a shortcut to that tab in a **Focus** section at the top of the rail, with or without groups; the tab itself stays where it is. Click a shortcut to switch to its tab, right-click it (or press the menu key) for *Remove from Focus*. Focus follows you across devices; whether it is collapsed is kept per device. Desktop and tablet only. |
 
 It is the same list either way, just re-hosted: tab order, drag-to-reorder, the `Alt+1`
 to `Alt+9` numbers and every status colour below behave identically in both. The setting is
@@ -73,6 +73,10 @@ precedence there.
 
 One tab per session, in your order, and that order syncs across your devices.
 
+An agent tab shows its CLI's logo before the name, and a shell tab an `SH` badge. **CLI Logos
+on Tabs** (App Settings → Appearance → Tabs) hides the logos on that device; the tile and split
+headers and the Run menus keep theirs.
+
 **Status is carried by the dot and the tab's own styling:**
 
 | Look                          | Meaning                                                                 |
@@ -224,8 +228,12 @@ Worth knowing:
   `~/.claude/settings.json`), so the wheel scrolls the conversation rather than the terminal.
   Claude's default inline view keeps its history in the terminal and scrolls locally. `Shift+Wheel` is
   always local scrollback. OpenCode's wheel and swipes page its own conversation
-  (PageUp/PageDown); in a grid tile or the split view's second pane the wheel does
-  too. Other CLIs scroll locally.
+  (PageUp/PageDown), and so do newer Codex versions, which draw on the alternate screen
+  and leave the terminal no history of its own (an older Codex that draws inline keeps
+  its history in the terminal and scrolls locally). The first small movement of a
+  gesture pages at once, then one page per half screen of travel; one finger on the
+  screen is one gesture however slowly it moves. In a grid tile or the split view's
+  second pane the wheel pages too. Other CLIs scroll locally.
 - **Selection copy.** `Ctrl+C` copies when text is selected and interrupts when it is not.
   `Ctrl+Shift+C` always copies.
 - **Selecting where the CLI owns the mouse.** `Shift+drag` starts a selection even in a pane
@@ -259,7 +267,7 @@ Two extras depending on the device:
 | Respawn          | Session Options                    | [Keeping Agents Running](Keeping-Agents-Running)                  |
 | Ralph            | Session Options                    | [Autonomous Loops](Autonomous-Loops)                              |
 | Orchestrator     | Toolbar                            | [Autonomous Loops](Autonomous-Loops)                              |
-| Cron             | Header ⏰ (opt-in)                 | [Cron Jobs](Cron-Jobs)                                            |
+| Cron             | Bottom toolbar ⏰ (opt-in)         | [Cron Jobs](Cron-Jobs)                                            |
 | Subagents        | Automatic while agents run         | [Watching Agents Work](Watching-Agents-Work)                      |
 | Ultracode        | Header (opt-in)                    | [Watching Agents Work](Watching-Agents-Work)                      |
 | File Viewer      | Header                             | [Working With Files](Working-With-Files)                          |

docs/wiki/Tile-Grid.md

@@ -16,23 +16,32 @@ It shows a **Tiles** button in the header, beside Split, and enables `Ctrl+Shift
 
 ## Opening a grid
 
-- **Tiles button**: one click shows the tiles straight away, as many as you last chose
-  (six until you choose; fewer if the window is too small or you have fewer sessions open).
-  You get the grid you last had, its tiles where they were, topped up with your open
-  sessions in tab order; if there is none, an open split's two first; otherwise your open
-  sessions in tab order, with the session you are on focused. With the grid open, the same
-  button closes it.
+- **Tiles button**: one click shows the tiles straight away. If you have used the grid in
+  this browser before, you get it back exactly as you left it: the same sessions in the
+  same places, an empty place where you left one, the same number of tiles, your column
+  widths and row heights, the tile you were in, and a zoomed tile still zoomed. A session
+  closed since frees its place, which is filled the way a new grid is filled (below).
+  Otherwise, or when none of those sessions is left, you get as many tiles as you last
+  chose (six until you choose; fewer if the window is too small or you have fewer sessions
+  open): an open split's two first; otherwise the sessions that are working (the most
+  recently started first), then the ones waiting on you (red and yellow tabs), then the
+  rest, the most recently used first, so the oldest are the ones left out. The session you
+  are on always comes along and is focused. With the grid open, the same button closes it.
 - **Rest the pointer on the Tiles button** (or tab to it) for a short card that shows the
-  count it opens and what a click and a right-click do.
+  count you chose and what a click and a right-click do.
 - **Right-click the Tiles button** (or press `Shift+F10` on it) to choose how many tiles:
-  **2**, **4** or **6**, each drawn as its layout. Your choice is remembered on this device
-  and is what the next click opens. With the grid open, picking a count re-forms it: the
-  tile you are in always stays, extra tiles leave from the end, new ones join from your tab
-  order. A count the window is too small for is greyed out, with the reason.
+  **2**, **4** or **6**, each drawn as its layout. Your choice is remembered in this
+  browser. Picking a count opens the grid with that many tiles (the grid you left, its
+  tiles in their places, new ones in the empty places first), and it is what a click opens
+  when there is no grid to bring back. With the grid open, picking a count re-forms it: the
+  tile you are in always stays, extra tiles leave from the end, new ones join working ones
+  first, then the ones waiting on you, then the most recent. A count the window is too
+  small for is greyed out, with the reason.
 - **`Ctrl+Shift+G`**: exactly what a click on the Tiles button does.
 - **`Ctrl`+click (or `Cmd`+click) a tab**: adds that session to the grid and focuses it. With
-  the grid closed it opens what the Tiles button would show, with that session among them
-  (still the count you chose in total). On macOS use
+  the grid closed it opens what the Tiles button would show with that session added: in the
+  empty place while the grid has fewer tiles than the count you chose, else in place of the
+  last tile (never more than the count). On macOS use
   `Cmd`: `Ctrl`+click there opens the tab's rename instead.
 - **Drag a tab onto a tile** to replace that tile with it (the replaced session keeps
   running), or onto an empty slot to add it. Dragging a session that is already tiled onto
@@ -89,7 +98,8 @@ keeps the focus.
 
 A moved tile takes the size of the place it lands in: column widths and row heights stay
 where you dragged the dividers. Tiles do not move while one is zoomed. Where everything is,
-the empty slot included, is saved with the grid and comes back on reload.
+the empty slot included, is saved with the grid and comes back when you turn the grid off
+and on, and on a page reload while the grid is open.
 
 Closing a tile leaves its place empty when the grid keeps its shape (six tiles to five), and a
 new tile takes the first empty place. When the number of tiles changes the grid's shape (four
@@ -122,8 +132,18 @@ session finder) shows that session on its own, the normal single view. The grid
 remembered: the Tiles button or `Ctrl+Shift+G` brings it straight back. Going Home does the
 same. Narrowing the window below the desktop width also returns to the single view.
 
-The grid is saved on this device and comes back when you reload the page, with its focus,
-zoom and column widths. A session that was closed in the meantime is simply left out.
+The grid is saved in this browser every time you change it (moving, resizing, adding or
+removing a tile, changing the count, focusing or zooming a tile), and never sent to the
+server. However you leave it (the Tiles button, another tab, Home, a link, closing its last
+tile or session), the Tiles button brings it back as it was. A page reload brings it back
+when the grid was open; after you left it, a reload shows the single view and the Tiles
+button still brings the grid back. A session that was closed or popped out into its own
+window in the meantime frees its place for another one, picked the way a new grid picks
+them; the place stays empty only when no other session is left. Right after a page reload
+the page does not know yet which sessions are waiting for your answer, so that pick goes by
+which sessions are working and which you used last. If the window has become too small for
+all the tiles, the tile you were in fills the grid until the window is wide enough again,
+and the rest of the layout is kept.
 
 Split shows the same logo, name and model above both of its panes.
 

docs/wiki/Working-With-Files.md

@@ -14,7 +14,7 @@ It renders what it can:
 | Kind                     | Behaviour                                                                 |
 | ------------------------ | ------------------------------------------------------------------------- |
 | Text and code            | Plain preview with Lines (line numbers) and Wrap toggles in the header. Long files are truncated in plain preview. |
-| Markdown                 | Rendered by default: headings, tables, code blocks with copy buttons, images and links relative to the file (root-relative ones resolve from the workspace root, as on GitHub). Opened from an attachment card, where the file's folder is unknown, relative images show their alt text and relative links show as plain text. The MD pill in the header flips to source. |
+| Markdown                 | Rendered by default: headings, tables, code blocks with copy buttons, images and links relative to the file (root-relative ones resolve from the workspace root, as on GitHub). Links to another heading of the same file (`[Install](#installation)`) scroll to it, with GitHub's heading names (lower-case, punctuation dropped, repeats numbered `-1`, `-2`), and never leave the page. Opened from an attachment card, where the file's folder is unknown, relative images show their alt text and relative links show as plain text. The MD pill in the header flips to source. |
 | Images                   | Inline.                                                                    |
 | Audio and video          | Inline with a working scrub bar, because range requests are supported.     |
 | Spreadsheets (`.xlsx`)   | Read-only grid, parsed in your browser (never on the server), up to 10 MB. `.xls` and `.ods` are download only. |
@@ -87,9 +87,10 @@ File paths in a session are links. That works in two places:
   render as underlined monospace links.
 
 Clicking one opens it in the preview: images and PDFs render, video and audio play with a
-working scrub bar, documents convert, text shows inline and Markdown renders. The exception is
-a text or Markdown file inside the workspace clicked in the terminal: that opens in the tail
-viewer instead, which follows a file that is still being written.
+working scrub bar, documents convert, text shows inline, Markdown renders and HTML opens as a
+page. The exception is a plain text file (a log, `.txt`, `.json`, code) inside the workspace
+clicked in the terminal: that opens in the tail viewer instead, which follows a file that is
+still being written. A Markdown file always opens rendered, wherever you click it.
 
 Paths **outside** the session's workspace work too, which matters because that is where most
 of an agent's output lands: a screenshot in `/tmp`, a capture in its own scratchpad, a file in
@@ -101,8 +102,24 @@ Outside the workspace the allowlist is images, video, audio, PDF, Office documen
 files, where "text" is the same list the viewer will let you edit: code, config, logs, csv,
 markdown. The reasoning is that a session can already `cat` any of those, so the file suffix
 was never what kept anything secret; the path guard is. Types outside the list (`.svg`,
-`.bmp`) say so rather than failing silently, and `.html` previews as source rather than being
-rendered, so nothing served this way can execute in the page.
+`.bmp`) say so rather than failing silently.
+
+An `.html` file opens as a rendered page by default, with its relative CSS, scripts, images and
+data files. The page is served from a separate `/html-view/...` address in a sandbox with no
+origin of its own, so its scripts can never read Codeman's page, cookies or API.
+
+By the project owner's decision, a rendered page has the same network access as a file you open
+from disk in your browser. It **can** load scripts, styles, fonts, images and media from
+anywhere on the internet (a dashboard that pulls Chart.js from a CDN works), send requests to
+any host, and read the files in its own folder. It **cannot** act as Codeman: it has no access
+to Codeman's page, cookies or API, and no access to any folder other than its own. Since a page
+can send what it can read to any host, open HTML only from sources you trust, as you would on
+your own desktop. The **Page**
+pill in the viewer header switches to the source view (where **Edit** works as for any text
+file), and the viewer falls back to source on its own when a page cannot be rendered: a remote
+(SSH) case, an HTML file inside a hidden folder, or one sitting directly in a broad folder such
+as your home folder or `/tmp` (the page may read the files in its own folder tree, so move it
+into a folder of its own). The raw file routes still serve HTML as a download only.
 
 Text previews are capped at the first 500 lines, fetched as a partial read, so clicking a
 one-gigabyte log does not try to paint one.
@@ -151,13 +168,39 @@ blocks sensitive trees. In multi-user mode a non-admin gets only their own user
 root, because per-user spaces live inside the home directory and a home-directory root would
 expose everyone.
 
-## Images into a session
+## Files into a session
 
-Paste from the clipboard or drag and drop straight onto the terminal. The image is written
+Paste from the clipboard or drag and drop straight onto the terminal. The file is written
 where the agent can read it and the reference is inserted into your prompt. On a phone, the
-image key in the keyboard bar opens the camera or photo library.
+toolbar's Attach button opens the camera, photo library or files, and puts the path on your
+prompt, where it behaves like text you typed; the Attach button inside the Compose editor
+does the same into the editor. Dropping a video or a document onto the terminal works the
+same way.
 
-HEIC images from an iPhone are converted to JPEG on the way in.
+Images (png, jpg, gif, webp, bmp, heic) are re-encoded on the device before they go up (a GIF
+goes as is, so it keeps its frames), and HEIC from an iPhone is converted to JPEG. Videos (mp4, webm, mov, m4v, ogv) are stored as recorded,
+up to 1 GB each by default (`CODEMAN_MAX_PASTE_VIDEO_BYTES`); they stream to disk rather than
+through server memory. Documents are the office formats (pdf, docx, pptx, xlsx, zip) and the
+text family (txt, md, csv, tsv, json, log, yaml, yml, xml). Each is checked on the way in,
+against its file signature for the binary formats and against its encoding for text (UTF-8,
+no NUL bytes), so a script renamed to `.pdf` is refused while a `.txt` can hold any text at
+all. Files with script or page extensions (sh, py, js, html) are not accepted, since a prompt
+upload is something for the agent to read, not something for Codeman to store and serve. A
+document keeps a trimmed copy of its original name, letters, digits, dot, dash and underscore
+with anything else turned into a dash, so `Q3 report.pdf` is findable as
+`paste-<time>-<id>-Q3-report.pdf`; a name made entirely of other characters saves without
+one. Uploads are refused for remote (SSH) sessions, whose working directory is on the other
+host.
+
+Uploads land in `.codeman-uploads/` inside the case, next to the code the agent is working
+on, because that is the one path a local agent and a Docker case (whose workspace is mounted
+at the same absolute path) both see the same way. The first upload writes a `.gitignore` of `*` into `.codeman-uploads/`, so the
+directory ignores itself and your repository never picks the files up, while your own ignore file is
+left alone; a `.gitignore` you already had there is never overwritten. Files older than seven
+days are swept hourly, and the directory goes when the last session of the case is closed (a
+session that is only detached keeps it). Earlier releases wrote to `.claude-images/`; that
+directory is still swept and cleaned up the same way for one release so a case in use before
+the move is not left with stray files.
 
 ## Generated artifacts
 

docs/wiki/_Sidebar.md

@@ -15,6 +15,7 @@
 - [Agent CLIs](Agent-CLIs)
 - [Custom Model Endpoints](Custom-Model-Endpoints)
 - [Working With Files](Working-With-Files)
+- [Git Worktrees](Git-Worktrees)
 - [Input And Voice](Input-And-Voice)
 - [Mobile Guide](Mobile-Guide)
 - [Keyboard Shortcuts](Keyboard-Shortcuts)
@@ -42,6 +43,7 @@
 
 **Automation**
 
+- [The Codeman Skill](Codeman-Skill)
 - [Driving It From An Agent](Driving-Codeman-From-An-Agent)
 - [HTTP API](HTTP-API)
 - [Hooks & Integrations](Hooks-And-Integrations)

package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "aicodeman",
-  "version": "1.40.0",
+  "version": "1.42.0",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "aicodeman",
-      "version": "1.40.0",
+      "version": "1.42.0",
       "hasInstallScript": true,
       "license": "MIT",
       "workspaces": [
@@ -13389,7 +13389,7 @@
       }
     },
     "packages/xterm-zerolag-input": {
-      "version": "0.4.0",
+      "version": "0.4.1",
       "license": "MIT",
       "devDependencies": {
         "@xterm/headless": "^6.0.0",

package.json

@@ -1,7 +1,8 @@
 {
   "name": "aicodeman",
-  "version": "1.40.0",
-  "description": "Mission control for AI coding agents - run 20 autonomous agents with real-time monitoring and session persistence",
+  "version": "1.42.0",
+  "description": "Self-hosted mission control for AI coding agents: run Claude Code, Codex, OpenCode, Gemini, DeepSeek, Grok and more 24/7 in tmux, from any device.",
+  "mcpName": "io.github.Ark0N/codeman",
   "type": "module",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",
@@ -38,11 +39,11 @@
     "check:public-assets": "node scripts/check-public-assets.mjs",
     "capture:subagents": "node scripts/capture-subagent-screenshots.mjs",
     "changeset": "changeset",
-    "version-packages": "changeset version && node scripts/sync-plugin.mjs && npm install --package-lock-only && node scripts/check-lockfile-sync.mjs",
+    "version-packages": "changeset version && node scripts/generate-whats-new.mjs && prettier --write src/web/public/whats-new.json && node scripts/sync-plugin.mjs && npm install --package-lock-only && node scripts/check-lockfile-sync.mjs",
     "check:lockfile": "node scripts/check-lockfile-sync.mjs",
     "check:plugin": "node scripts/sync-plugin.mjs --check && claude plugin validate --strict plugins/codeman && claude plugin validate --strict .claude-plugin/marketplace.json",
     "knip": "npx --yes knip@latest --config config/knip.json",
-    "release": "changeset publish"
+    "release": "node scripts/npm-release.mjs"
   },
   "prettier": {
     "singleQuote": true,
@@ -171,7 +172,7 @@
   "bugs": {
     "url": "https://github.com/Ark0N/Codeman/issues"
   },
-  "homepage": "https://github.com/Ark0N/Codeman#readme",
+  "homepage": "https://getcodeman.com",
   "files": [
     "dist",
     "scripts/postinstall.js",

packages/gesture-control/README.md

@@ -133,6 +133,13 @@ its CSP for WebAssembly). Deploy = copy the bundle into Codeman's
   drop point. Pinch the panel again to move it anywhere. It stays inside the
   camera-owning page, so the hand keeps control (the old `window.app.detachSession`
   `window.open` was a one-way trip). A small twitch-and-release cancels.
+- **Tiles** (tile grid open): pinch anywhere on a tile and carry it onto
+  another tile (the two trade places) or an empty cell (it moves there). A
+  session tab carried onto a tile or an empty cell joins the grid there. With
+  **Detach Tiles** on (App Settings, per device, off by default) a tile let go
+  outside the grid opens as a window of its own, which needs the site's pop-ups
+  allowed (a hand is no click). The moves are tile-grid.js's own
+  (`tileDropTargetAt` / `dropOnTileTarget` / `detachTileAtPoint`).
 - **Run / Run Shell** — pinch over the **Run** (`#runBtn`) or **Run Shell**
   (`.btn-shell`) toolbar button and release in place to fire it; drifting too
   far first cancels the tap. The button list is `CLICK_SELECTOR` in `entry.ts`.

packages/gesture-control/src/codeman/entry.ts

@@ -27,6 +27,14 @@
 //   • Button "tap" — pinch over a toolbar button (Run / Run Shell) and release
 //     in place → fires the button's real click handler. Drift too far first and
 //     it's treated as a stray move, not a tap.
+//   • Tile "grab-to-move": with the tile grid open, pinch anywhere on a tile
+//     and carry it onto another tile (the two trade places) or an empty cell
+//     (it moves there). With Detach Tiles on (App Settings, per device, off by
+//     default) a tile let go OUTSIDE the grid opens as a window of its own.
+//     A session TAB carried onto a tile or an empty cell joins the grid there,
+//     as a mouse-dragged tab does. Every move goes through app.js's own paths
+//     (tile-grid.js tileDropTargetAt / dropOnTileTarget / detachTileAtPoint), so
+//     the hand and the mouse can never disagree about what a drop does.
 //
 // Why in-page floats, not OS-window detach (decided 2026-06-08, see
 // docs/MULTIMONITOR_DESIGN.md): the hand only exists in the one page that owns
@@ -52,11 +60,22 @@ declare global {
       saveSubagentWindowStates?: () => void;
       subagentWindowZIndex?: number;
       ultracodeWindowZIndex?: number;
+      // The tile grid (tile-grid.js). All optional: an older dashboard without
+      // them simply has no tile verbs for the hand.
+      tileAtPoint?: (x: number, y: number) => string | null;
+      canGrabTile?: (id: string) => boolean;
+      tileDropTargetAt?: (x: number, y: number, draggedId: string) => TileTarget | null;
+      dropOnTileTarget?: (draggedId: string, target: TileTarget) => boolean;
+      isOverTileGrid?: (x: number, y: number) => boolean;
+      tileDetachEnabled?: () => boolean;
+      detachTileAtPoint?: (id: string, x: number, y: number, at?: { offsetX: number; offsetY: number }) => boolean;
     };
   }
 }
 
 const TAB_SELECTOR = '.session-tab';
+/** A tile of the open tile grid (tile-grid.js), keyed by `data-session-id`. */
+const TILE_SELECTOR = '#tileGrid .tile';
 /** An in-page floating session panel this layer spawned — re-grabbable to move. */
 const PANEL_SELECTOR = '.cg-float';
 /** The dashboard's own floating agent windows (subagent runs + ultracode run and
@@ -81,11 +100,20 @@ const FLOAT_H = 420;
 const DETACH_PULL_PX = 70;
 /** If a button-pinch drifts more than this, it's a stray move, not a tap. */
 const TAP_CANCEL_PX = 45;
+/** How far (px) a grabbed tile must travel before it targets anything: a pinch
+ *  that barely moves is a stray one, and the tile stays where it is. */
+const TILE_MOVE_PX = 40;
+/** The width (px) of the small copy of a grabbed tile that follows the hand. */
+const TILE_GHOST_W = 280;
 
 /** First letter colours: cyan left, violet right; green while pinching. */
 const handColor = (handedness: string, pinching: boolean): string =>
   pinching ? '#4ade80' : handedness === 'Right' ? '#a78bfa' : '#38bdf8';
 
+/** Where a session carried by the hand would land in the tile grid: another
+ *  tile, or an empty cell (tile-grid.js tileDropTargetAt). */
+type TileTarget = { kind: 'tile'; id: string; el: HTMLElement } | { kind: 'cell'; cell: number; el: HTMLElement };
+
 /** A floating in-page session panel (an iframe of `/session/:id`) the hand can
  *  place and re-grab. Stays in this page's DOM, so it never leaves hand reach. */
 interface FloatingPanel {
@@ -107,6 +135,29 @@ type Grab =
       oy: number;
       /** Pulled past the float-out threshold at least once. */
       armed: boolean;
+      /** With the tile grid open: the tile or empty cell it would join (highlighted). */
+      target: TileTarget | null;
+    }
+  | {
+      /** A tile of the open grid, carried to another cell or (Detach Tiles)
+       *  out of the grid. The tile itself never moves until the drop: it dims
+       *  (`tile--dragging`, the mouse drag's own class) and a small copy of
+       *  it follows the hand. */
+      kind: 'tile';
+      id: string;
+      el: HTMLElement;
+      ghost: HTMLElement;
+      /** Where in the tile the hand closed, so a pop-out window lands under it. */
+      offsetX: number;
+      offsetY: number;
+      ox: number;
+      oy: number;
+      /** Travelled past TILE_MOVE_PX at least once. */
+      armed: boolean;
+      /** The tile or empty cell it would drop onto (highlighted). */
+      target: TileTarget | null;
+      /** Outside the grid with Detach Tiles on: letting go opens a window. */
+      out: boolean;
     }
   | {
       kind: 'panel';
@@ -224,7 +275,7 @@ class GestureBridge {
       await this.gc.start();
       this.running = true;
       this.button.classList.add('on');
-      this.status.textContent = 'on — pinch a tab, window, or button';
+      this.status.textContent = 'on: pinch a tab, tile, window, or button';
     } catch (err) {
       // Surface the *real* cause: MediaPipe/Emscripten can throw a non-Error
       // (number/string), so `(err as Error).message` was logging "undefined".
@@ -295,6 +346,35 @@ class GestureBridge {
       return;
     }
 
+    // A tile of the open tile grid → carry it to another cell, or (Detach
+    // Tiles) out of the grid. Anywhere on the tile: the hand is choosing the
+    // whole tile, as with the agent windows. A tile that can do neither (alone
+    // or zoomed, Detach Tiles off) is left alone, so the pinch falls through.
+    const tileId = window.app?.tileAtPoint?.(x, y) ?? null;
+    const tileEl = tileId ? this.hitClosest(x, y, TILE_SELECTOR) : null;
+    if (tileId && tileEl && window.app?.canGrabTile?.(tileId)) {
+      const rect = tileEl.getBoundingClientRect();
+      const ghost = this.tileGhost(tileEl, rect);
+      document.body.append(ghost);
+      tileEl.classList.add('tile--dragging');
+      this.grabs.set(hand, {
+        kind: 'tile',
+        id: tileId,
+        el: tileEl,
+        ghost,
+        offsetX: Math.max(0, Math.min(x - rect.left, rect.width)),
+        offsetY: Math.max(0, Math.min(y - rect.top, rect.height)),
+        ox: x,
+        oy: y,
+        armed: false,
+        target: null,
+        out: false,
+      });
+      this.positionGhost(ghost, x, y);
+      this.status.textContent = 'moving tile';
+      return;
+    }
+
     // A session tab → grab-and-pull-out into a floating panel (ghost follows).
     const tab = this.hitClosest(x, y, TAB_SELECTOR);
     const id = tab?.dataset.id;
@@ -308,7 +388,7 @@ class GestureBridge {
       document.body.append(ghost);
 
       tab.classList.add('cg-grabbed');
-      this.grabs.set(hand, { kind: 'tab', id, tab, ghost, ox: x, oy: y, armed: false });
+      this.grabs.set(hand, { kind: 'tab', id, tab, ghost, ox: x, oy: y, armed: false, target: null });
       this.positionGhost(ghost, x, y);
       return;
     }
@@ -325,13 +405,42 @@ class GestureBridge {
 
   private onDrag(hand: string, x: number, y: number): void {
     const grab = this.grabs.get(hand);
+    if (grab?.kind === 'tile') {
+      this.positionGhost(grab.ghost, x, y);
+      if (!grab.armed && Math.hypot(x - grab.ox, y - grab.oy) >= TILE_MOVE_PX) grab.armed = true;
+      if (!grab.armed) return;
+      const app = window.app;
+      const target = app?.tileDropTargetAt?.(x, y, grab.id) ?? null;
+      this.setTileTarget(grab, target);
+      const out = !target && !!app?.tileDetachEnabled?.() && !app?.isOverTileGrid?.(x, y);
+      grab.out = out;
+      grab.ghost.classList.toggle('cg-armed', !!target);
+      grab.ghost.classList.toggle('cg-out', out);
+      this.status.textContent = target
+        ? target.kind === 'tile'
+          ? 'release to swap tiles'
+          : 'release to move the tile here'
+        : out
+          ? 'release to open in a new window'
+          : 'moving tile';
+      return;
+    }
     if (grab?.kind === 'tab') {
       this.positionGhost(grab.ghost, x, y);
       const pulled = Math.hypot(x - grab.ox, y - grab.oy) >= DETACH_PULL_PX;
-      if (pulled !== grab.armed) {
+      // The tile grid open: a tile or an empty cell under the hand takes the
+      // tab there, as a mouse-dragged tab does.
+      const target = pulled ? (window.app?.tileDropTargetAt?.(x, y, grab.id) ?? null) : null;
+      const targetChanged = (target?.el ?? null) !== (grab.target?.el ?? null);
+      this.setTileTarget(grab, target);
+      if (pulled !== grab.armed || targetChanged) {
         grab.armed = pulled;
         grab.ghost.classList.toggle('cg-armed', pulled);
-        this.status.textContent = pulled ? 'release to float out' : 'on — pinch a tab';
+        this.status.textContent = target
+          ? 'release to tile it here'
+          : pulled
+            ? 'release to float out'
+            : 'on — pinch a tab';
       }
       return;
     }
@@ -354,17 +463,44 @@ class GestureBridge {
     if (tap && Math.hypot(x - tap.ox, y - tap.oy) > TAP_CANCEL_PX) {
       tap.el.classList.remove('cg-tap-armed');
       this.taps.delete(hand);
-      this.status.textContent = 'on — pinch a tab, window, or button';
+      this.status.textContent = 'on: pinch a tab, tile, window, or button';
     }
   }
 
   private onDrop(hand: string, x: number, y: number): void {
     const grab = this.grabs.get(hand);
+    if (grab?.kind === 'tile') {
+      this.grabs.delete(hand);
+      grab.ghost.remove();
+      grab.el.classList.remove('tile--dragging');
+      this.setTileTarget(grab, null);
+      const app = window.app;
+      if (!grab.armed) {
+        this.flash('cancelled');
+        return;
+      }
+      // Asked again at the drop: the grid may have changed since the last frame.
+      const target = app?.tileDropTargetAt?.(x, y, grab.id) ?? null;
+      if (target) {
+        this.flash(app?.dropOnTileTarget?.(grab.id, target) ? 'moved tile' : 'cancelled');
+      } else if (grab.out && !app?.isOverTileGrid?.(x, y)) {
+        // A window opened with no click behind it is what popup blockers stop:
+        // detachSession then says so in a toast and the tile stays.
+        const opened = app?.detachTileAtPoint?.(grab.id, x, y, { offsetX: grab.offsetX, offsetY: grab.offsetY });
+        this.flash(opened ? 'opened in a new window' : 'pop-out blocked: allow popups for this site');
+      } else {
+        this.flash('cancelled');
+      }
+      return;
+    }
     if (grab?.kind === 'tab') {
       this.grabs.delete(hand);
       grab.ghost.remove();
       grab.tab.classList.remove('cg-grabbed');
-      if (grab.armed) this.floatPanel(grab.id, x, y);
+      this.setTileTarget(grab, null);
+      const target = grab.armed ? (window.app?.tileDropTargetAt?.(x, y, grab.id) ?? null) : null;
+      if (target) this.flash(window.app?.dropOnTileTarget?.(grab.id, target) ? 'tiled' : 'cancelled');
+      else if (grab.armed) this.floatPanel(grab.id, x, y);
       else this.flash('cancelled');
       return;
     }
@@ -505,6 +641,33 @@ class GestureBridge {
     ghost.style.top = `${y}px`;
   }
 
+  /** A small copy of a tile to follow the hand: its header (a copy: no
+   *  listeners come along) over an empty body, in the tile's proportions. */
+  private tileGhost(tileEl: HTMLElement, rect: DOMRect): HTMLElement {
+    const ghost = el('div', 'cg-ghost cg-tile-ghost');
+    const width = Math.min(TILE_GHOST_W, rect.width);
+    ghost.style.width = `${width}px`;
+    ghost.style.height = `${Math.max(48, Math.round((width * rect.height) / Math.max(1, rect.width)))}px`;
+    const header = tileEl.querySelector('.tile-header');
+    if (header) {
+      const copy = header.cloneNode(true) as HTMLElement;
+      copy.removeAttribute('draggable');
+      ghost.append(copy);
+    }
+    ghost.append(el('div', 'cg-tile-ghost-body'));
+    return ghost;
+  }
+
+  /** Highlights the tile or empty cell a carried session would drop onto
+   *  (the mouse drag's own `tile--drop-target`), clearing the last one. */
+  private setTileTarget(grab: { target: TileTarget | null }, target: TileTarget | null): void {
+    if (grab.target?.el !== target?.el) {
+      grab.target?.el.classList.remove('tile--drop-target');
+      target?.el.classList.add('tile--drop-target');
+    }
+    grab.target = target;
+  }
+
   private cancelAllGrabs(): void {
     // Floats themselves persist (they're placed windows) — only release any
     // in-progress grab cleanly, restoring a moved panel's interactivity.
@@ -512,6 +675,11 @@ class GestureBridge {
       if (grab.kind === 'tab') {
         grab.ghost.remove();
         grab.tab.classList.remove('cg-grabbed');
+        this.setTileTarget(grab, null);
+      } else if (grab.kind === 'tile') {
+        grab.ghost.remove();
+        grab.el.classList.remove('tile--dragging');
+        this.setTileTarget(grab, null);
       } else if (grab.kind === 'panel') {
         grab.panel.el.style.pointerEvents = '';
         grab.panel.el.classList.remove('cg-float-grabbed', 'cg-redock');
@@ -525,6 +693,7 @@ class GestureBridge {
     document
       .querySelectorAll(`${TAB_SELECTOR}.cg-grabbed, .cg-tap-armed, .cg-win-grabbed`)
       .forEach((t) => t.classList.remove('cg-grabbed', 'cg-tap-armed', 'cg-win-grabbed'));
+    document.querySelectorAll(`${TILE_SELECTOR}.tile--dragging`).forEach((t) => t.classList.remove('tile--dragging'));
   }
 
   private onStatus(fps: number, hands: HandState[]): void {
@@ -607,6 +776,13 @@ function injectStyles(): void {
     outline: 2px solid #38bdf8; outline-offset: -2px;
   }
   .cg-ghost.cg-armed { outline-color: #4ade80; box-shadow: 0 8px 28px rgba(74,222,128,.5); }
+  .cg-ghost.cg-out { outline-color: #fbbf24; box-shadow: 0 8px 28px rgba(251,191,36,.5); }
+  .cg-tile-ghost {
+    display: flex; flex-direction: column; overflow: hidden; transform: translate(-50%, -20%) scale(1);
+    background: var(--term-bg, #161b23); border: 1px solid var(--border-color, #333);
+  }
+  .cg-tile-ghost > .tile-header { flex: 0 0 auto; }
+  .cg-tile-ghost-body { flex: 1 1 auto; opacity: .5; }
   .cg-dock {
     position: fixed; right: 12px; bottom: 156px; z-index: ${Z + 3};
     display: flex; align-items: center; gap: 8px; font: 12px/1 system-ui, sans-serif;

packages/xterm-zerolag-input/CHANGELOG.md

@@ -1,5 +1,51 @@
 # xterm-zerolag-input
 
+## 0.4.1
+
+### Patch Changes
+
+- 87f1c9c: ### Thanks
+  - @Randalix for `codeman agent` (#557), the session verbs (`ls`, `spawn`, `send`, `wait`, `read`, `interrupt`, `rm`) for agents in every CLI mode, and for moving every test server onto an ephemeral port (#570), which finishes #440. Thanks also for reporting and fixing git's clone errors on non-English hosts (#568, shipped as #572 with your commit).
+  - @opticon454 for five PRs: MCP server sync for GitHub Copilot CLI (#581), an Apply button that saves Settings without closing them (#565), configurable toast and browser-notification display times (#564), in-document links in rendered markdown that scroll to their heading (#563), and npm-based CLI installs that work when the system npm prefix is root-owned (#562).
+  - @JDProfresh for three PRs: pasted and uploaded files moving into a hidden, self-ignoring `.codeman-uploads/` folder (#574, after your #553 proposal), local echo that keeps painting on phones when the view sits above the bottom (#576), and upload failures that say why they failed (#578).
+  - @aakhter for session search on the vertical tab rail (#580), built on the sidebar's existing filter instead of a second one, and for keeping alerted tabs visible during a search.
+
+  ![Codeman tile grid: six live agents powering on and off with the CRT tile animation](https://raw.githubusercontent.com/Ark0N/Codeman/08694b5862534b2b1224e9e272ad9858c787ebef/release-1.41/tiles-crt-stats-800.gif)
+
+  **Tile Animations (#571).** The tile grid from 1.40.0 can now open with a show. App Settings has a new **Animations** section (right after Appearance) that holds every animation setting: the Entrance Theme (moved out of Appearance), the new **Tile Animations** row, and a button that opens the animation lab. Tile styles: **CRT** (each tile switches on as a hot line in a diagonal wave, and switches off to a line and a dot), **Fly from tab** (each tile grows out of its session tab and flies back into it), **Deal** (dealt out of the Tiles button like cards), **Beam down**, **Cascade**, **Pop**, **Soft** and **None**. A styled tile plays in two beats: the frame enters in the tile style, then its screen powers on in the terminal style of your Entrance Theme. Picking a theme presets a matching tile style, and a new **Launch** theme flies the tiles out of their tabs. Off by default (the grid keeps its quick fade), per device, nothing moves under reduced motion, and the animations never cost an extra PTY resize. The terminal pane's Boot entrance style is gone; a saved Boot falls back to off.
+
+  **Tiles open the sessions you care about, and keep your layout.** With no grid arranged yet, the Tiles button (or `Ctrl+Shift+G`) now fills the grid with the sessions that are working first (most recently started first), then the ones waiting on you (a permission prompt or a finished turn you have not seen), then the most recently used, instead of the tabs in strip order, which opened the oldest ones. The session you are on is still always included. Once you arrange a grid (which session sits where, empty cells, the tile count, divider sizes, the focused and zoomed tile), that layout is remembered in this browser and the Tiles button brings it back exactly, however the grid was closed, and so does a reload while it was open. A session that has gone since frees its place, which the same ranking fills. A grid larger than the window opens in full with the focused tile zoomed, instead of being trimmed (and losing the rest of your layout).
+
+  **The wheel scrolls Claude inside a tile (#577).** In a tile or the split's second pane, a Claude session on its fullscreen renderer now scrolls its own conversation with the mouse wheel, exactly like the main terminal: the wheel goes to Claude as mouse reports, aimed at that tile's session and computed from the tile's own screen. Before, the wheel did nothing or scrolled the stale frames left over from loading the tile. Shift+wheel (scroll local history) works in every tile on Windows and Linux too; it was dead there.
+
+  **`codeman agent`: session verbs for every CLI (#557).** Agents in any mode (Codex, OpenCode, Gemini, Pi and the rest, not just Claude) can now drive other Codeman sessions from the command line: `codeman agent ls | spawn | send | wait | read | interrupt | rm`. It is a thin client over the existing session API: every call names the session that made it, `wait` blocks on a signal (`--until stop,exit`) or a literal output marker (`--match`), and exit codes say what happened (`0` ok, `1` error, `2` timeout, `3` exited, `4` refused). Ids shorter than 8 characters are refused, so a stray `rm 9` can never pick a session at random, and `rm` never deletes the session it runs in. See the README section "`codeman agent`" and the wiki page Driving Codeman From An Agent. This is phase 1 of #445.
+
+  **Settings: Apply (#565).** Next to Save, an Apply button saves the same way but keeps Settings open. Switching on MCP server sync makes its Preview and Sync usable straight away, and CLI management's add, enable and disable work without closing and reopening Settings.
+
+  **MCP server sync reaches GitHub Copilot CLI (#581).** With MCP server sync on (App Settings, off by default), Copilot CLI's `~/.copilot/mcp-config.json` now takes part like the agent CLIs' own files: its servers are copied to the others and theirs to it, additively, with the previous file kept as `.codeman-bak`. Copilot joins only when it is installed or already has that file, a server switched off in Copilot is never copied, and `COPILOT_HOME` is followed. Copilot is a sync target only, not a new run mode.
+
+  **Notifications stay up longer if you want (#564).** Settings → Notifications has a Toast display time and a Browser notification display time (1 second to 5 minutes, per device; the defaults stay 3 s and 8 s).
+
+  **CLI logos on tabs can be switched off (#569).** App Settings → Appearance → Tabs → **CLI Logos on Tabs** hides the agent logo on every tab surface (header strip, rails, sidebar, phone chips, the desktop home list) on this device. On by default. Tile headers, split pane headers and the Run menus keep their logos.
+
+  **Search sessions on the vertical rail (#580).** The vertical tab rail has a **Search sessions** box at the top: type part of a name and the rail narrows to the tabs that match (a web tab by its title), across every group, collapsed ones included, without touching your groups, their collapse or the tab order. A tab with an alert stays visible even when its name does not match, so a prompt waiting on you is never filtered away. Escape or × clears it, you can still drag a found tab into a group, and nothing is saved. The sidebar's filter box shares the same filter: a tab with an alert stays visible there too, and in the by-case tab layout a case with no match now hides.
+
+  **Closing a tab is instant.** Closing a session used to take half a second or more before the tab went away. The tab, tile or split pane now goes (and the next session is selected) the moment you click, while the server shuts the session down in the background, and the server side is faster too (about 450 ms down to 200-260 ms for a Claude session): it no longer sleeps fixed intervals, no longer freezes for about 70 ms per close on a synchronous tmux call, and scans for a session's subagents once instead of once per subagent. If the server refuses the close, the tab comes back where it was with the error. This also fixes a bug where a failed close still said "Session closed" while the session kept running.
+
+  **Fixes.**
+  - **Clone errors on non-English hosts (#572, from #568).** Cloning a repository as a case now classifies a failed clone correctly whatever the host's language: a missing branch or tag is "does not exist on the remote" (400) and a missing repository is a 404, instead of a generic 422 with git's German (or any other) error text. Git runs with `LC_ALL=C` for clones and repo status, so the repo status card's error text is English on every host as well.
+  - **Links within a markdown file (#563).** A link to another heading of the same document (`[Install](#installation)`) in the File Viewer or Response Viewer scrolls to that heading instead of doing nothing. Headings get GitHub-style slugs, repeated titles are numbered, and non-ASCII headings work.
+  - **npm CLI installs on a root-owned prefix (#562).** Installing an npm-based CLI from Settings (DeepSeek's `dsh`, pi, ...) no longer fails with EACCES when the system node keeps its global prefix under `/usr`: the install goes to `~/.local`, where Codeman already looks for CLIs. A prefix you set yourself, or one you can write to, is left alone, including when Codeman runs under `npm run`.
+  - **Uploads go to a hidden `.codeman-uploads/` folder (#574).** Images you paste or upload into a prompt are saved in `<workspace>/.codeman-uploads/` instead of `.claude-images/`. The folder ignores itself in git (it carries a `.gitignore` of `*`), stays hidden in the Files panel, and is cleaned up as before: files older than 7 days in an hourly sweep, and the folder when the last session in that workspace closes. The old `.claude-images/` folder gets nothing new and is still swept and removed during 1.41.x. An upload to a remote (SSH) session is now refused with a clear message, since the file would land on the Codeman host where the remote agent cannot read it.
+  - **Typing on a phone after a tab switch (#576, fixes #575).** With local echo on (the default on touch devices), text typed while the terminal sat above the bottom was buffered but never painted, so the keyboard looked dead. The view sits there after every tab switch and after the keyboard closes. The overlay now paints whenever the prompt row is on screen, and hides only when you scroll the prompt out of view.
+  - **Upload failures say why (#578).** When a prompt image upload fails, the toast shows the server's reason (for example a rate limit) instead of only "1 failed".
+  - **The npm page shows the English README.** npmjs.com had been rendering the Chinese README, because npm picks the package's readme from an unsorted file match at publish time. The publish now moves `README.zh-CN.md` aside while it runs (the file and every link to it stay as they are), and the package description and homepage say what Codeman is and point at getcodeman.com.
+  - **New cases ask for clickable file paths.** The CLAUDE.md generated into a new case asks the agent to report every file it created as a full absolute path, which Codeman turns into a link that opens the File Viewer, and mentions the codeman skill for starting and managing worker sessions.
+
+  **For contributors (#570).** Every in-process test server binds an ephemeral port, the mobile suite included, and the port guard now also refuses raw listeners on a fixed port, so two test runs on one machine never collide.
+
+  **Fixes applied while landing.** zh-CN translations for the two new notification display-time settings and for the new Animations section. A session whose name matches an interface word ("Lab", "New session") is no longer translated in the tab strip when the interface is in Chinese. Plus test and doc cleanups left over from review.
+
 ## 0.4.0
 
 ### Minor Changes

packages/xterm-zerolag-input/README.md

@@ -537,7 +537,7 @@ While flushed text exists the prompt column is locked, so a full-screen redraw c
 
 ### Scroll awareness
 
-The overlay hides when the viewport is scrolled up (`viewportY !== baseY`) and re-renders, debounced, when you scroll back to the bottom.
+The overlay hides while the cursor row is scrolled out of the viewport and re-renders, debounced, when it scrolls back into view. A viewport parked a few rows above the bottom keeps painting as long as the cursor row is on screen. A buffer that reports no `cursorY` keeps the bottom-only rule.
 
 ---
 

packages/xterm-zerolag-input/package.json

@@ -1,6 +1,6 @@
 {
   "name": "xterm-zerolag-input",
-  "version": "0.4.0",
+  "version": "0.4.1",
   "description": "Instant keystroke feedback overlay for xterm.js: Mosh-inspired local echo that removes perceived input latency over SSH, tunnels and other high-RTT connections",
   "type": "module",
   "main": "dist/index.cjs",

packages/xterm-zerolag-input/src/prompt-finder.ts

@@ -72,3 +72,20 @@ export function readTextAfterPrompt(terminal: XtermTerminal, prompt: PromptPosit
     return '';
   }
 }
+
+/**
+ * Whether the row the overlay draws on (the cursor row) is inside the viewport. A bare
+ * `viewportY === baseY` test is wrong for a host that parks the viewport a few rows above
+ * the bottom with the prompt still on screen; a buffer without `cursorY` keeps that rule.
+ */
+export function promptRowInViewport(terminal: XtermTerminal): boolean {
+  try {
+    const buf = terminal.buffer.active;
+    if (buf.viewportY === buf.baseY) return true;
+    if (typeof buf.cursorY !== 'number') return false;
+    const cursorRow = buf.baseY + buf.cursorY;
+    return cursorRow >= buf.viewportY && cursorRow < buf.viewportY + terminal.rows;
+  } catch {
+    return false;
+  }
+}

packages/xterm-zerolag-input/src/zerolag-input-addon.ts

@@ -8,7 +8,7 @@ import type {
   FontStyle,
 } from './types.js';
 import { getCellDimensions } from './cell-dimensions.js';
-import { findPrompt, readTextAfterPrompt } from './prompt-finder.js';
+import { findPrompt, readTextAfterPrompt, promptRowInViewport } from './prompt-finder.js';
 import { renderOverlay, charCellWidth } from './overlay-renderer.js';
 
 const DEFAULT_PROMPT: PromptFinder = { type: 'character', char: '>', offset: 2 };
@@ -122,11 +122,10 @@ export class ZerolagInputAddon implements XtermAddon {
     // Cache font properties
     this._cacheFont();
 
-    // Scroll detection: hide overlay when scrolled away from bottom
+    // Scroll detection: hide the overlay while the cursor row is scrolled out of view
     this._scrollHandler = () => {
       try {
-        const buf = this._terminal!.buffer.active;
-        if (buf.viewportY !== buf.baseY) {
+        if (!promptRowInViewport(this._terminal!)) {
           this._overlay!.style.display = 'none';
           if (this._scrollTimer) {
             clearTimeout(this._scrollTimer);
@@ -565,8 +564,8 @@ export class ZerolagInputAddon implements XtermAddon {
     try {
       const buf = this._terminal.buffer.active;
 
-      // Hide overlay when scrolled up — prompt is at bottom, not in viewport
-      if (buf.viewportY !== buf.baseY) {
+      // Hide the overlay while the cursor row is scrolled out of view
+      if (!promptRowInViewport(this._terminal)) {
         this._overlay.style.display = 'none';
         return;
       }

packages/xterm-zerolag-input/test/prompt-finder.test.ts

@@ -1,6 +1,6 @@
 import { describe, it, expect } from 'vitest';
 import { createMockTerminal } from './helpers.js';
-import { findPrompt, readTextAfterPrompt } from '../src/prompt-finder.js';
+import { promptRowInViewport, findPrompt, readTextAfterPrompt } from '../src/prompt-finder.js';
 import type { XtermTerminal, PromptFinder } from '../src/types.js';
 
 function term(lines: string[]) {
@@ -156,3 +156,28 @@ describe('readTextAfterPrompt', () => {
     cleanup();
   });
 });
+
+describe('promptRowInViewport', () => {
+  const term = (viewportY: number, baseY: number, cursorY: number | undefined, rows = 24) =>
+    ({ rows, buffer: { active: { viewportY, baseY, cursorY, getLine: () => undefined } } }) as never;
+
+  it('is true at the bottom regardless of the cursor', () => {
+    expect(promptRowInViewport(term(10, 10, undefined))).toBe(true);
+  });
+
+  it('is true for a viewport parked above the bottom while the cursor row is on screen', () => {
+    // scrollToLastNonEmptyLine() parks rows - 2 above the last non-empty row
+    expect(promptRowInViewport(term(0, 16, 5))).toBe(true);
+    // cursor exactly on the last visible row
+    expect(promptRowInViewport(term(0, 23, 0))).toBe(true);
+  });
+
+  it('is false once the cursor row is scrolled out of the viewport', () => {
+    expect(promptRowInViewport(term(0, 24, 0))).toBe(false); // one past the last row
+    expect(promptRowInViewport(term(0, 200, 3))).toBe(false); // deep in history
+  });
+
+  it('keeps the bottom-only rule when the buffer has no cursorY', () => {
+    expect(promptRowInViewport(term(0, 1, undefined))).toBe(false);
+  });
+});

packages/xterm-zerolag-input/test/zerolag-input-addon.test.ts

@@ -725,3 +725,34 @@ describe('ZerolagInputAddon', () => {
     });
   });
 });
+
+describe('viewport scrolled away from the bottom', () => {
+  function parked(viewportY: number, baseY: number, cursorY: number, rows = 24) {
+    const mock = createMockTerminal({ buffer: { lines: ['$ '], viewportY, baseY, cursorY }, rows });
+    const addon = new ZerolagInputAddon({ prompt: { type: 'character', char: '$', offset: 2 } });
+    mock.terminal.loadAddon(addon);
+    cleanups.push(() => {
+      addon.dispose();
+      mock.cleanup();
+    });
+    const overlay = mock.terminal.element.querySelector('.xterm-screen')!.lastElementChild as HTMLDivElement;
+    return { addon, mock, overlay };
+  }
+
+  it('still paints with the viewport parked above the bottom while the cursor row is on screen', () => {
+    // The host parks the viewport to keep trailing blank rows out of view; the
+    // prompt and cursor are still visible, so the user's text must be too.
+    const { addon, overlay } = parked(0, 1, 0);
+    addon.appendText('abc');
+    expect(addon.pendingText).toBe('abc');
+    expect(overlay.style.display).not.toBe('none');
+    expect(overlay.textContent).toContain('abc');
+  });
+
+  it('hides once the cursor row is scrolled out of the viewport, even over a stale prompt glyph', () => {
+    const { addon, overlay } = parked(0, 30, 0);
+    addon.appendText('abc');
+    expect(addon.pendingText).toBe('abc');
+    expect(overlay.style.display).toBe('none');
+  });
+});

plugins/codeman/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "codeman",
   "description": "Drive Codeman, the self-hosted session manager for AI coding agents, from inside a Claude Code session: spawn worker sessions, prompt them, wait for them, read their answers, clean up. Acts only inside a Codeman-managed session.",
-  "version": "1.40.0",
+  "version": "1.42.0",
   "author": {
     "name": "Ark0N",
     "url": "https://github.com/Ark0N"

plugins/codeman/skills/codeman/SKILL.md

@@ -1,11 +1,13 @@
 ---
 name: codeman
 description: >-
-  Drive Codeman, the session manager this agent is running inside, over its HTTP API:
-  list sessions, start worker sessions, send them prompts, block until they finish
-  (wait / wait-output / send-and-wait), read their output, and clean up; where
-  available, message claude workers directly (Claude Code cross-session messaging).
-  Use when asked to orchestrate or parallelize work across Codeman sessions, watch
+  Drive Codeman, the session manager this agent is running inside, over its HTTP API,
+  from any agent CLI (Claude Code, Codex, Gemini CLI, OpenCode, Antigravity, Grok
+  Build, DeepSeek Harness, OMP): list sessions, start worker sessions in any installed
+  CLI, send them prompts, block until they finish (wait / wait-output / send-and-wait),
+  read their output, and clean up; where available, message claude workers directly
+  (Claude Code cross-session messaging). Use when asked to orchestrate or parallelize
+  work across Codeman sessions, get a second opinion from another CLI or model, watch
   another session, or start and manage workers. Only usable inside a Codeman-managed
   session (CODEMAN_MUX=1); refuse to act otherwise.
 ---
@@ -29,11 +31,31 @@ worked multi-worker flows in [reference/recipes.md](reference/recipes.md), endpo
 tables and a symptom gallery in [reference/endpoints.md](reference/endpoints.md), and
 direct messaging to claude workers in [reference/messaging.md](reference/messaging.md).
 
+This skill works the same whichever CLI you are: Codeman puts it where each CLI looks
+for skills (`.claude/skills/` for Claude Code, `.agents/skills/` for Codex, Gemini CLI,
+OpenCode, Antigravity, Grok, DeepSeek and OMP). A **pi** or **shell** worker never gets
+it, but has the same environment: tell it *"other sessions: `codeman agent --help`"*.
+That bundled CLI (`ls`, `spawn`, `send`, `wait`, `read`, `interrupt`, `rm`) is the same
+verbs over the same endpoints, with the guards below (no control bytes, no self-delete,
+no guessed URL) enforced in code, and it is also your own way through when this
+machine has no `jq` for the preamble below.
+
 ## 0. Guard and bootstrap
 
 If `CODEMAN_MUX` is not `1`, **stop and say so**. Do not guess an API URL; a server
 you are not part of is not yours to drive.
 
+Two failures come from your CLI, not from Codeman: name them to the user instead of
+retrying.
+
+- **Every call fails at once with "couldn't connect" or "connection refused".** Your
+  CLI's sandbox blocks network access, even to this machine. Codex's default sandbox
+  does; it needs **Bypass approvals and sandbox** (Codeman App Settings, Codex).
+- **`"${CURL[@]}" -o /dev/null -w '%{http_code}' "$API/api/v1/sessions/$SELF"` prints
+  `404`.** The commands you run do not carry this session's environment: Codex can run them in a shared background
+  process another session started, so `$CODEMAN_SESSION_ID` (and with it `$SELF` and
+  the delete guard) is that session's. Delete nothing; tell the user.
+
 ⚠️ **Your shell state does not survive between tool calls.** Each Bash call starts a
 fresh shell, so `$API`, `$SELF`, the `CURL` array and `delete_session` are all gone by
 the next call, and `$$` is a different pid. **The filesystem does survive**, so write
@@ -41,13 +63,14 @@ the preamble to a file once and source it afterwards, rather than re-pasting a
 hundred-odd lines at the top of every call (a half-re-pasted preamble used to be the
 single most likely way to break a run).
 
-**Codeman seeds the preamble file for you** when it spawns a claude session (server
-1.18.3+), so the bootstrap is usually nothing at all: these are the two lines every
+**Codeman seeds the preamble file for you** when it spawns your session (claude since
+server 1.18.3, the other CLIs since this skill reached them), so the bootstrap is
+usually nothing at all: these are the two lines every
 later call opens with, and your first REAL call performs them anyway:
 
 ```bash
 . "${XDG_CACHE_HOME:-$HOME/.cache}/codeman-agent-$CODEMAN_SESSION_ID.sh" 2>/dev/null
-[ "${CODEMAN_PREAMBLE:-}" = 1.33.4 ] || { echo "preamble missing or stale; run the full §0 block"; exit 1; }
+[ "${CODEMAN_PREAMBLE:-}" = 1.42.0 ] || { echo "preamble missing or stale; run the full §0 block"; exit 1; }
 ```
 
 ⚠️ **Never spend a Bash call on this check alone.** §1's block opens with this same
@@ -75,8 +98,11 @@ PRE="${XDG_CACHE_HOME:-$HOME/.cache}/codeman-agent-$CODEMAN_SESSION_ID.sh"
 mkdir -p "$(dirname "$PRE")"
 # Rewrite unless the file already ends with THIS version's stamp, so a stale or a
 # half-written file self-heals here instead of costing you a round trip to rm it.
-grep -qs '^CODEMAN_PREAMBLE=1.33.4$' "$PRE" || (umask 077; cat > "$PRE" <<'PREAMBLE'
-# ---- Codeman agent preamble 1.33.4 (seeded by Codeman at session spawn; the SKILL.md §0 bootstrap rewrites it when missing or stale) ----
+grep -qs '^CODEMAN_PREAMBLE=1.42.0$' "$PRE" || (umask 077; cat > "$PRE" <<'PREAMBLE'
+# ---- Codeman agent preamble 1.42.0 (seeded by Codeman at session spawn; the SKILL.md §0 bootstrap rewrites it when missing or stale) ----
+# Positional parameters are always written with braces, ${1} and never a bare dollar-digit: Claude Code
+# replaces a bare one in a skill body with the matching word of the skill's arguments, and §0 is pasted
+# from that body (test/agent-skill.test.ts pins it).
 API="${CODEMAN_API_URL:?CODEMAN_API_URL not set; refusing to guess}"
 SELF="${CODEMAN_SESSION_ID:?CODEMAN_SESSION_ID not set}"
 # Credentials, cheapest first. Your session has usually INHERITED the server's
@@ -85,7 +111,7 @@ SELF="${CODEMAN_SESSION_ID:?CODEMAN_SESSION_ID not set}"
 # reads. The data dir is wherever the hook-secret file lives. Values may be
 # quoted or `export`-prefixed.
 ENV_FILE="${CODEMAN_HOOK_SECRET_FILE:+${CODEMAN_HOOK_SECRET_FILE%hook-secret}.env}"
-envval() { sed -n "s/^\(export \)\{0,1\}$1=//p" "$ENV_FILE" | tail -1 | sed 's/^"\(.*\)"$/\1/; s/^'\''\(.*\)'\''$/\1/'; }
+envval() { sed -n "s/^\(export \)\{0,1\}${1}=//p" "$ENV_FILE" | tail -1 | sed 's/^"\(.*\)"$/\1/; s/^'\''\(.*\)'\''$/\1/'; }
 if [ -z "${CODEMAN_PASSWORD:-}" ] && [ -n "$ENV_FILE" ] && [ -f "$ENV_FILE" ]; then
   CODEMAN_USERNAME=$(envval CODEMAN_USERNAME)
   CODEMAN_PASSWORD=$(envval CODEMAN_PASSWORD)
@@ -120,15 +146,15 @@ delete_session() {
 
 # ---- fast path: the four verbs, already written. §1 composes them. ----
 _composer_up() {   # <sid> <timeoutMs> -> "true"/"false". `shift+tab` is the one token
-  "${CURL[@]}" -G "$API/api/v1/sessions/$1/wait-output" \
+  "${CURL[@]}" -G "$API/api/v1/sessions/${1}/wait-output" \
     --data-urlencode 'match=shift+tab' --data-urlencode 'from=buffer' \
-    --data-urlencode "timeout=$2" | jq -r '.data.wait.matched // false'
+    --data-urlencode "timeout=${2}" | jq -r '.data.wait.matched // false'
 }
 _dsh_up() {        # <sid> <timeoutMs> -> "true"/"false". The DeepSeek Harness TUI's
   # composer glyph. Override with DSH_READY_MARK for a profile that draws another one.
-  "${CURL[@]}" -G "$API/api/v1/sessions/$1/wait-output" \
+  "${CURL[@]}" -G "$API/api/v1/sessions/${1}/wait-output" \
     --data-urlencode "match=${DSH_READY_MARK:-❯}" --data-urlencode 'from=buffer' \
-    --data-urlencode "timeout=$2" | jq -r '.data.wait.matched // false'
+    --data-urlencode "timeout=${2}" | jq -r '.data.wait.matched // false'
 }
 # ---- the workspace-trust dialog: READ the screen, never press Enter blind ----
 # Claude Code 2.1.252 dropped the option numbers, REVERSED them, and highlights
@@ -144,7 +170,7 @@ _trust_key() {     # <sid> -> "confirm" | "move" | "" (nothing safe to press)
   # full=1 returns the RENDERED pane; a claude pane keeps no tmux history, so that
   # is the current frame rather than every repaint since launch. tail -1 anyway,
   # because the freshest marked row is the only one still true.
-  "${CURL[@]}" -G "$API/api/v1/sessions/$1/terminal" --data-urlencode 'full=1' \
+  "${CURL[@]}" -G "$API/api/v1/sessions/${1}/terminal" --data-urlencode 'full=1' \
     | jq -r '.data.terminalBuffer // empty' \
     | sed -e "s/$(printf '\033')\[[0-9;?]*[a-zA-Z]//g" -e "s/$(printf '\033')[()][AB0]//g" \
     | tr -d ' \t' | grep -i '❯[0-9.]*\(yes,itrustthisfolder\|no,exit\)' | tail -1 \
@@ -162,7 +188,7 @@ _composer_text() { # <sid> -> the composer's text with ALL whitespace removed: "
   # the LAST `❯` line: Claude Code echoes a submitted prompt with the same glyph higher
   # up in the transcript, so only the last one says whether the text was taken.
   local t
-  t=$("${CURL[@]}" -G "$API/api/v1/sessions/$1/terminal" --data-urlencode 'full=1' \
+  t=$("${CURL[@]}" -G "$API/api/v1/sessions/${1}/terminal" --data-urlencode 'full=1' \
     | jq -r '.data.terminalBuffer // empty' \
     | sed -e "s/$(printf '\033')\[[0-9;?]*[a-zA-Z]//g" -e "s/$(printf '\033')[()][AB0]//g" \
     | tr -d '\r' | grep -a '^[[:space:]]*❯' | tail -1)
@@ -172,7 +198,7 @@ _composer_text() { # <sid> -> the composer's text with ALL whitespace removed: "
   printf '%s' "$t" | sed 's/^[[:space:]]*❯//' | tr -d '[:space:]' | sed "s/$(printf '\302\240')//g"
 }
 _accept_trust() {  # <sid> -> 0 once it has answered the dialog, 1 if it could not
-  local sid="$1" k i=1
+  local sid="${1}" k i=1
   while [ "$i" -le 6 ]; do
     k=$(_trust_key "$sid")
     [ -n "$k" ] || return 1   # no dialog on screen, or a layout this cannot read
@@ -373,7 +399,7 @@ sendwait() {
 last_text() {
   local t="" prev="${2:-}"
   for _ in $(seq 1 15); do
-    t=$("${CURL[@]}" "$API/api/v1/sessions/$1/last-response" | jq -r '.data.text // empty')
+    t=$("${CURL[@]}" "$API/api/v1/sessions/${1}/last-response" | jq -r '.data.text // empty')
     [ -n "$t" ] && [ "$t" != "$prev" ] && { printf '%s\n' "$t"; return 0; }
     sleep 1
   done
@@ -384,10 +410,10 @@ last_text() {
 # The stamp is the LAST line on purpose (a truncated write leaves it unset) and is kept
 # bare on purpose: the write condition above anchors on it with $, so an inline comment
 # here would fail that match and rewrite this file on every single bootstrap.
-CODEMAN_PREAMBLE=1.33.4
+CODEMAN_PREAMBLE=1.42.0
 PREAMBLE
 )
-. "$PRE"; [ "${CODEMAN_PREAMBLE:-}" = 1.33.4 ] || { echo "preamble at $PRE is stale or truncated: rm it and re-run this block"; exit 1; }
+. "$PRE"; [ "${CODEMAN_PREAMBLE:-}" = 1.42.0 ] || { echo "preamble at $PRE is stale or truncated: rm it and re-run this block"; exit 1; }
 ```
 
 Every later Bash call that touches the API starts with the same two loader lines from
@@ -438,7 +464,7 @@ and no per-call body to hand-build.
 
 ```bash
 . "${XDG_CACHE_HOME:-$HOME/.cache}/codeman-agent-$CODEMAN_SESSION_ID.sh" 2>/dev/null   # §0 loader
-[ "${CODEMAN_PREAMBLE:-}" = 1.33.4 ] || { echo "preamble missing or stale; run the full §0 block"; exit 1; }
+[ "${CODEMAN_PREAMBLE:-}" = 1.42.0 ] || { echo "preamble missing or stale; run the full §0 block"; exit 1; }
 N=(alpha beta)                    # INVENT one fresh case name per worker; never list cases first
                                   # (a name may carry a mode: `beta:deepseek`, see below)
 T=('reply with one line: the absolute path of your working directory'

plugins/codeman/skills/codeman/preamble.sh

@@ -1,4 +1,7 @@
-# ---- Codeman agent preamble 1.33.4 (seeded by Codeman at session spawn; the SKILL.md §0 bootstrap rewrites it when missing or stale) ----
+# ---- Codeman agent preamble 1.42.0 (seeded by Codeman at session spawn; the SKILL.md §0 bootstrap rewrites it when missing or stale) ----
+# Positional parameters are always written with braces, ${1} and never a bare dollar-digit: Claude Code
+# replaces a bare one in a skill body with the matching word of the skill's arguments, and §0 is pasted
+# from that body (test/agent-skill.test.ts pins it).
 API="${CODEMAN_API_URL:?CODEMAN_API_URL not set; refusing to guess}"
 SELF="${CODEMAN_SESSION_ID:?CODEMAN_SESSION_ID not set}"
 # Credentials, cheapest first. Your session has usually INHERITED the server's
@@ -7,7 +10,7 @@ SELF="${CODEMAN_SESSION_ID:?CODEMAN_SESSION_ID not set}"
 # reads. The data dir is wherever the hook-secret file lives. Values may be
 # quoted or `export`-prefixed.
 ENV_FILE="${CODEMAN_HOOK_SECRET_FILE:+${CODEMAN_HOOK_SECRET_FILE%hook-secret}.env}"
-envval() { sed -n "s/^\(export \)\{0,1\}$1=//p" "$ENV_FILE" | tail -1 | sed 's/^"\(.*\)"$/\1/; s/^'\''\(.*\)'\''$/\1/'; }
+envval() { sed -n "s/^\(export \)\{0,1\}${1}=//p" "$ENV_FILE" | tail -1 | sed 's/^"\(.*\)"$/\1/; s/^'\''\(.*\)'\''$/\1/'; }
 if [ -z "${CODEMAN_PASSWORD:-}" ] && [ -n "$ENV_FILE" ] && [ -f "$ENV_FILE" ]; then
   CODEMAN_USERNAME=$(envval CODEMAN_USERNAME)
   CODEMAN_PASSWORD=$(envval CODEMAN_PASSWORD)
@@ -42,15 +45,15 @@ delete_session() {
 
 # ---- fast path: the four verbs, already written. §1 composes them. ----
 _composer_up() {   # <sid> <timeoutMs> -> "true"/"false". `shift+tab` is the one token
-  "${CURL[@]}" -G "$API/api/v1/sessions/$1/wait-output" \
+  "${CURL[@]}" -G "$API/api/v1/sessions/${1}/wait-output" \
     --data-urlencode 'match=shift+tab' --data-urlencode 'from=buffer' \
-    --data-urlencode "timeout=$2" | jq -r '.data.wait.matched // false'
+    --data-urlencode "timeout=${2}" | jq -r '.data.wait.matched // false'
 }
 _dsh_up() {        # <sid> <timeoutMs> -> "true"/"false". The DeepSeek Harness TUI's
   # composer glyph. Override with DSH_READY_MARK for a profile that draws another one.
-  "${CURL[@]}" -G "$API/api/v1/sessions/$1/wait-output" \
+  "${CURL[@]}" -G "$API/api/v1/sessions/${1}/wait-output" \
     --data-urlencode "match=${DSH_READY_MARK:-❯}" --data-urlencode 'from=buffer' \
-    --data-urlencode "timeout=$2" | jq -r '.data.wait.matched // false'
+    --data-urlencode "timeout=${2}" | jq -r '.data.wait.matched // false'
 }
 # ---- the workspace-trust dialog: READ the screen, never press Enter blind ----
 # Claude Code 2.1.252 dropped the option numbers, REVERSED them, and highlights
@@ -66,7 +69,7 @@ _trust_key() {     # <sid> -> "confirm" | "move" | "" (nothing safe to press)
   # full=1 returns the RENDERED pane; a claude pane keeps no tmux history, so that
   # is the current frame rather than every repaint since launch. tail -1 anyway,
   # because the freshest marked row is the only one still true.
-  "${CURL[@]}" -G "$API/api/v1/sessions/$1/terminal" --data-urlencode 'full=1' \
+  "${CURL[@]}" -G "$API/api/v1/sessions/${1}/terminal" --data-urlencode 'full=1' \
     | jq -r '.data.terminalBuffer // empty' \
     | sed -e "s/$(printf '\033')\[[0-9;?]*[a-zA-Z]//g" -e "s/$(printf '\033')[()][AB0]//g" \
     | tr -d ' \t' | grep -i '❯[0-9.]*\(yes,itrustthisfolder\|no,exit\)' | tail -1 \
@@ -84,7 +87,7 @@ _composer_text() { # <sid> -> the composer's text with ALL whitespace removed: "
   # the LAST `❯` line: Claude Code echoes a submitted prompt with the same glyph higher
   # up in the transcript, so only the last one says whether the text was taken.
   local t
-  t=$("${CURL[@]}" -G "$API/api/v1/sessions/$1/terminal" --data-urlencode 'full=1' \
+  t=$("${CURL[@]}" -G "$API/api/v1/sessions/${1}/terminal" --data-urlencode 'full=1' \
     | jq -r '.data.terminalBuffer // empty' \
     | sed -e "s/$(printf '\033')\[[0-9;?]*[a-zA-Z]//g" -e "s/$(printf '\033')[()][AB0]//g" \
     | tr -d '\r' | grep -a '^[[:space:]]*❯' | tail -1)
@@ -94,7 +97,7 @@ _composer_text() { # <sid> -> the composer's text with ALL whitespace removed: "
   printf '%s' "$t" | sed 's/^[[:space:]]*❯//' | tr -d '[:space:]' | sed "s/$(printf '\302\240')//g"
 }
 _accept_trust() {  # <sid> -> 0 once it has answered the dialog, 1 if it could not
-  local sid="$1" k i=1
+  local sid="${1}" k i=1
   while [ "$i" -le 6 ]; do
     k=$(_trust_key "$sid")
     [ -n "$k" ] || return 1   # no dialog on screen, or a layout this cannot read
@@ -295,7 +298,7 @@ sendwait() {
 last_text() {
   local t="" prev="${2:-}"
   for _ in $(seq 1 15); do
-    t=$("${CURL[@]}" "$API/api/v1/sessions/$1/last-response" | jq -r '.data.text // empty')
+    t=$("${CURL[@]}" "$API/api/v1/sessions/${1}/last-response" | jq -r '.data.text // empty')
     [ -n "$t" ] && [ "$t" != "$prev" ] && { printf '%s\n' "$t"; return 0; }
     sleep 1
   done
@@ -306,4 +309,4 @@ last_text() {
 # The stamp is the LAST line on purpose (a truncated write leaves it unset) and is kept
 # bare on purpose: the write condition above anchors on it with $, so an inline comment
 # here would fail that match and rewrite this file on every single bootstrap.
-CODEMAN_PREAMBLE=1.33.4
+CODEMAN_PREAMBLE=1.42.0

plugins/codeman/skills/codeman/reference/endpoints.md

@@ -635,7 +635,7 @@ Three bounded long-polls. Shared semantics:
 | `idle` | output stabilized + prompt detected, heuristic, can flap mid-turn | every mode |
 | `working` | session started producing output | every mode |
 | `stop` | Claude Code `stop` hook, the definitive end-of-turn | `claude` only |
-| `blocked` | `permission_prompt` / `elicitation_dialog` hook, the worker needs an answer | `claude` only |
+| `blocked` | `permission_prompt` / `elicitation_dialog` hook, the worker needs an answer | `claude`; also `codex`, whose approval, question and trust dialogs Codeman reads off its screen (answer them in the terminal, never through the inbox) |
 | `exit` | PTY exited or session deleted | every mode |
 
 ⚠️ **`claude` mode is necessary for `stop`/`blocked`, not sufficient. The real

plugins/codeman/skills/codeman/reference/recipes.md

@@ -21,7 +21,7 @@ by sourcing the preamble file the §0 bootstrap wrote, and checking its version
 
 ```bash
 . "${XDG_CACHE_HOME:-$HOME/.cache}/codeman-agent-$CODEMAN_SESSION_ID.sh" 2>/dev/null
-[ "${CODEMAN_PREAMBLE:-}" = 1.33.4 ] || { echo "preamble missing or stale; re-run the §0 bootstrap"; exit 1; }
+[ "${CODEMAN_PREAMBLE:-}" = 1.42.0 ] || { echo "preamble missing or stale; re-run the §0 bootstrap"; exit 1; }
 ```
 
 Do **not** re-paste the preamble body into each call. Sourcing it is what retires the
@@ -391,10 +391,10 @@ the worker remembering to print a token.
 
 ## Flow 5: watch for a worker stuck on a prompt
 
-Claude workers can block on a permission dialog. `blocked` is a wait signal
-(claude-mode only, and it needs Codeman's hooks in the worker's directory: see Flow 7
-step 4), so watch for it and surface the question to the user instead of guessing an
-answer. Expect it routinely on a server whose `claudeMode` is not the default bypass
+Claude workers can block on a permission dialog. `blocked` is a wait signal (a claude
+worker needs Codeman's hooks in its directory for it: see Flow 7 step 4; a codex worker
+gets it from its screen, see below), so watch for it and surface the question to the user
+instead of guessing an answer. Expect it routinely on a server whose `claudeMode` is not the default bypass
 one (the same setting that decides whether the readiness marker in Flow 1 ever
 appears):
 
@@ -409,10 +409,13 @@ if [ "$(jq -r '.data.wait.signal' <<<"$R")" = blocked ]; then
 fi
 ```
 
-Where the worker has no hooks, `blocked` never fires and a stuck worker looks exactly
-like a slow one: your marker wait burns its whole bound. The fallback is the same
-terminal tail, taken when a bound runs out, and the same rule about not answering it
-yourself.
+Codex workers are the one exception without hooks: Codeman reads their approval,
+question and trust dialogs off the screen, so the same watch works with
+`until=blocked,exit` (codex has no `stop`, a `400`), and the dialog is answered in the
+terminal, never through the Approvals Inbox. Anywhere else without hooks, `blocked` never
+fires and a stuck worker looks exactly like a slow one: your marker wait burns its whole
+bound. The fallback is the same terminal tail, taken when a bound runs out, and the same
+rule about not answering it yourself.
 
 ## Flow 6: claude fan-out over messaging
 

plugins/codeman/skills/codeman/reference/verbs.md

@@ -377,7 +377,9 @@ only when the workspace actually has them, see [§5.1](#51-where-to-spawn)) **an
 `deepseek`** — the one external CLI that reports its own lifecycle, so its `stop` is a
 real end-of-turn signal rather than a guess. On
 `shell`/`opencode`/`codex`/`gemini`/`antigravity`/`pi`/`grok`/`omp`, requesting them explicitly is a
-400, and lifecycle transitions there are coarse (a short shell command may emit **no**
+400 (except `blocked` on `codex`, which Codeman reads off its screen when an approval,
+question or trust dialog is up; a wait for it answers at once if one already is), and
+lifecycle transitions there are coarse (a short shell command may emit **no**
 `idle` transition at all, verified live), so synchronize those with markers.
 
 ⚠️ A dsh session can still refuse them for a per-SESSION reason: `statusReporting:

scripts/generate-whats-new.mjs

@@ -0,0 +1,18 @@
+#!/usr/bin/env node
+// Writes src/web/public/whats-new.json for the version in package.json from CHANGELOG.md.
+// Runs as part of `npm run version-packages`, so every release carries its own notes and the
+// UI can announce them once per version without anyone remembering to.
+import { readFileSync, writeFileSync } from 'node:fs';
+import { dirname, resolve } from 'node:path';
+import { fileURLToPath } from 'node:url';
+import { buildWhatsNew } from './lib/whats-new.mjs';
+
+const root = resolve(dirname(fileURLToPath(import.meta.url)), '..');
+const { version } = JSON.parse(readFileSync(resolve(root, 'package.json'), 'utf8'));
+const changelog = readFileSync(resolve(root, 'CHANGELOG.md'), 'utf8');
+const notes = buildWhatsNew(changelog, version);
+if (notes.highlights.length === 0) {
+  console.warn(`whats-new: no bold highlights found for ${version}; the in-app announcement will be skipped`);
+}
+writeFileSync(resolve(root, 'src/web/public/whats-new.json'), JSON.stringify(notes, null, 2) + '\n');
+console.log(`whats-new: ${version} -> ${notes.highlights.length} highlight(s)`);

scripts/lib/whats-new.mjs

@@ -0,0 +1,42 @@
+/**
+ * Pulls the headline of each change out of one version's section of CHANGELOG.md.
+ *
+ * Release notes here lead each change with a bold phrase ("**Tab layouts (#538, optional).** ..."),
+ * so a highlight is that phrase. The Thanks block is credit, not a feature.
+ */
+
+const MAX_HIGHLIGHTS = 8;
+
+/** The text between `## <version>` and the next `## ` heading, or '' when absent. */
+export function versionSection(changelog, version) {
+  const lines = changelog.split(/\r?\n/);
+  const start = lines.findIndex((l) => l.trim() === `## ${version}`);
+  if (start === -1) return '';
+  let end = lines.findIndex((l, i) => i > start && /^## /.test(l));
+  if (end === -1) end = lines.length;
+  return lines.slice(start + 1, end).join('\n');
+}
+
+// Bold phrases the release notes use as group labels rather than as the headline of a change:
+// "**Fixes.**" heads the list of fixes, "**Fixes applied while landing.**" the review fixes.
+const SECTION_LABELS =
+  /^(?:thanks\b.*|fixes|fixes applied while landing|docs|docs and tests|new features|also in this release)$/i;
+
+/** Bold lead-ins at the start of a paragraph or bullet, cleaned for display. */
+export function extractHighlights(section) {
+  const out = [];
+  for (const line of section.split('\n')) {
+    // `changeset version` prefixes the first line of a committed changeset with its commit
+    // ("- d18deed: **Lead.** text"), which is the shape every changesets-bot release has.
+    const m = /^\s*(?:[-*]\s+)?(?:[0-9a-f]{7,40}:\s+)?\*\*([^*]{3,120})\*\*/.exec(line);
+    if (!m) continue;
+    const text = m[1].replace(/[\s.:]+$/, '').trim();
+    if (!text || SECTION_LABELS.test(text)) continue;
+    if (!out.includes(text)) out.push(text);
+  }
+  return out.slice(0, MAX_HIGHLIGHTS);
+}
+
+export function buildWhatsNew(changelog, version) {
+  return { version, highlights: extractHighlights(versionSection(changelog, version)) };
+}

scripts/npm-release.mjs

@@ -0,0 +1,71 @@
+#!/usr/bin/env node
+/**
+ * @fileoverview `npm run release`: `changeset publish` with `README.zh-CN.md` moved aside.
+ *
+ * npmjs.com renders the package's top-level `readme`, and npm picks it at publish time:
+ * @npmcli/package-json's normalize globs `{README,README.*}` in the package root UNSORTED and
+ * keeps the first `.md` it sees. With `README.zh-CN.md` next to `README.md` that was the
+ * Chinese one on this machine and on CI, so npmjs.com showed the Chinese README for months.
+ * `files` cannot help: npm-packlist always includes every root `README.*`.
+ *
+ * Renaming the file would break every link to it, so for the length of the publish only it
+ * moves to a name npm does not treat as a readme (a leading dot), and is put back afterwards,
+ * whatever the publish did. The move happens HERE, inside the publish command, never as a
+ * step before `changesets/action` in release.yml: that action also runs the version path and
+ * commits the working tree into its version PR, which would commit the deletion.
+ *
+ * A previous run killed between the move and the restore leaves the aside copy behind; the
+ * next run puts it back first. `xterm-zerolag-input` (packages/) has only a README.md.
+ *
+ *   node scripts/npm-release.mjs    what the Release workflow runs (via `npm run release`)
+ */
+import { existsSync, renameSync } from 'node:fs';
+import { spawnSync } from 'node:child_process';
+import { join } from 'node:path';
+import { fileURLToPath } from 'node:url';
+
+export const HIDDEN_README = 'README.zh-CN.md';
+export const ASIDE_NAME = '.README.zh-CN.md.release-aside';
+
+/**
+ * Runs `publish` with `HIDDEN_README` moved aside in `root`, restoring it afterwards, also
+ * when `publish` fails or throws. Returns the exit code `publish` returned.
+ *
+ * @param {{ root: string, publish: () => number, log?: (msg: string) => void }} opts
+ * @returns {number}
+ */
+export function publishWithReadmeAside({ root, publish, log = (msg) => console.log(msg) }) {
+  const original = join(root, HIDDEN_README);
+  const aside = join(root, ASIDE_NAME);
+  if (existsSync(aside) && !existsSync(original)) {
+    renameSync(aside, original);
+    log(`npm-release: restored ${HIDDEN_README} left aside by an earlier run`);
+  }
+  const moved = existsSync(original);
+  if (moved) {
+    renameSync(original, aside);
+    log(`npm-release: ${HIDDEN_README} moved aside so npm picks README.md as the readme`);
+  }
+  try {
+    return publish();
+  } finally {
+    if (moved) {
+      renameSync(aside, original);
+      log(`npm-release: ${HIDDEN_README} restored`);
+    }
+  }
+}
+
+function runChangesetPublish() {
+  const result = spawnSync('changeset', ['publish'], { stdio: 'inherit', shell: process.platform === 'win32' });
+  if (result.error) {
+    console.error(`npm-release: could not run changeset publish: ${result.error.message}`);
+    return 1;
+  }
+  return result.status ?? 1;
+}
+
+if (process.argv[1] && fileURLToPath(import.meta.url) === process.argv[1]) {
+  const root = join(fileURLToPath(new URL('.', import.meta.url)), '..');
+  process.exitCode = publishWithReadmeAside({ root, publish: runChangesetPublish });
+}

scripts/sync-plugin.mjs

@@ -29,6 +29,9 @@ const SOURCE = 'skills/codeman';
 const PLUGIN_DIR = `plugins/${PLUGIN_NAME}`;
 const MIRROR = `${PLUGIN_DIR}/skills/codeman`;
 const MANIFESTS = [`${PLUGIN_DIR}/.claude-plugin/plugin.json`, '.claude-plugin/marketplace.json'];
+// The MCP registry manifest for `codeman mcp` (published with `mcp-publisher` from its
+// directory): the server and its npm package both carry the release version.
+const MCP_REGISTRY_MANIFEST = 'config/mcp-registry/server.json';
 
 const check = process.argv.includes('--check');
 const { version } = JSON.parse(readFileSync('package.json', 'utf8'));
@@ -78,6 +81,21 @@ for (const file of MANIFESTS) {
   if (changed && !check) writeFileSync(file, JSON.stringify(json, null, 2) + '\n');
 }
 
+// 3. The MCP registry manifest.
+{
+  const json = JSON.parse(readFileSync(MCP_REGISTRY_MANIFEST, 'utf8'));
+  const targets = [json, ...(json.packages ?? []).filter((p) => p.registryType === 'npm')];
+  let changed = false;
+  for (const target of targets) {
+    if (target.version !== version) {
+      drift.push(`${MCP_REGISTRY_MANIFEST}: ${target.version} -> ${version}`);
+      target.version = version;
+      changed = true;
+    }
+  }
+  if (changed && !check) writeFileSync(MCP_REGISTRY_MANIFEST, JSON.stringify(json, null, 2) + '\n');
+}
+
 if (drift.length === 0) {
   console.log(`plugin in step: mirror identical, manifests at ${version}`);
 } else if (check) {

skills/codeman/SKILL.md

@@ -1,11 +1,13 @@
 ---
 name: codeman
 description: >-
-  Drive Codeman, the session manager this agent is running inside, over its HTTP API:
-  list sessions, start worker sessions, send them prompts, block until they finish
-  (wait / wait-output / send-and-wait), read their output, and clean up; where
-  available, message claude workers directly (Claude Code cross-session messaging).
-  Use when asked to orchestrate or parallelize work across Codeman sessions, watch
+  Drive Codeman, the session manager this agent is running inside, over its HTTP API,
+  from any agent CLI (Claude Code, Codex, Gemini CLI, OpenCode, Antigravity, Grok
+  Build, DeepSeek Harness, OMP): list sessions, start worker sessions in any installed
+  CLI, send them prompts, block until they finish (wait / wait-output / send-and-wait),
+  read their output, and clean up; where available, message claude workers directly
+  (Claude Code cross-session messaging). Use when asked to orchestrate or parallelize
+  work across Codeman sessions, get a second opinion from another CLI or model, watch
   another session, or start and manage workers. Only usable inside a Codeman-managed
   session (CODEMAN_MUX=1); refuse to act otherwise.
 ---
@@ -29,11 +31,31 @@ worked multi-worker flows in [reference/recipes.md](reference/recipes.md), endpo
 tables and a symptom gallery in [reference/endpoints.md](reference/endpoints.md), and
 direct messaging to claude workers in [reference/messaging.md](reference/messaging.md).
 
+This skill works the same whichever CLI you are: Codeman puts it where each CLI looks
+for skills (`.claude/skills/` for Claude Code, `.agents/skills/` for Codex, Gemini CLI,
+OpenCode, Antigravity, Grok, DeepSeek and OMP). A **pi** or **shell** worker never gets
+it, but has the same environment: tell it *"other sessions: `codeman agent --help`"*.
+That bundled CLI (`ls`, `spawn`, `send`, `wait`, `read`, `interrupt`, `rm`) is the same
+verbs over the same endpoints, with the guards below (no control bytes, no self-delete,
+no guessed URL) enforced in code, and it is also your own way through when this
+machine has no `jq` for the preamble below.
+
 ## 0. Guard and bootstrap
 
 If `CODEMAN_MUX` is not `1`, **stop and say so**. Do not guess an API URL; a server
 you are not part of is not yours to drive.
 
+Two failures come from your CLI, not from Codeman: name them to the user instead of
+retrying.
+
+- **Every call fails at once with "couldn't connect" or "connection refused".** Your
+  CLI's sandbox blocks network access, even to this machine. Codex's default sandbox
+  does; it needs **Bypass approvals and sandbox** (Codeman App Settings, Codex).
+- **`"${CURL[@]}" -o /dev/null -w '%{http_code}' "$API/api/v1/sessions/$SELF"` prints
+  `404`.** The commands you run do not carry this session's environment: Codex can run them in a shared background
+  process another session started, so `$CODEMAN_SESSION_ID` (and with it `$SELF` and
+  the delete guard) is that session's. Delete nothing; tell the user.
+
 ⚠️ **Your shell state does not survive between tool calls.** Each Bash call starts a
 fresh shell, so `$API`, `$SELF`, the `CURL` array and `delete_session` are all gone by
 the next call, and `$$` is a different pid. **The filesystem does survive**, so write
@@ -41,13 +63,14 @@ the preamble to a file once and source it afterwards, rather than re-pasting a
 hundred-odd lines at the top of every call (a half-re-pasted preamble used to be the
 single most likely way to break a run).
 
-**Codeman seeds the preamble file for you** when it spawns a claude session (server
-1.18.3+), so the bootstrap is usually nothing at all: these are the two lines every
+**Codeman seeds the preamble file for you** when it spawns your session (claude since
+server 1.18.3, the other CLIs since this skill reached them), so the bootstrap is
+usually nothing at all: these are the two lines every
 later call opens with, and your first REAL call performs them anyway:
 
 ```bash
 . "${XDG_CACHE_HOME:-$HOME/.cache}/codeman-agent-$CODEMAN_SESSION_ID.sh" 2>/dev/null
-[ "${CODEMAN_PREAMBLE:-}" = 1.33.4 ] || { echo "preamble missing or stale; run the full §0 block"; exit 1; }
+[ "${CODEMAN_PREAMBLE:-}" = 1.42.0 ] || { echo "preamble missing or stale; run the full §0 block"; exit 1; }
 ```
 
 ⚠️ **Never spend a Bash call on this check alone.** §1's block opens with this same
@@ -75,8 +98,11 @@ PRE="${XDG_CACHE_HOME:-$HOME/.cache}/codeman-agent-$CODEMAN_SESSION_ID.sh"
 mkdir -p "$(dirname "$PRE")"
 # Rewrite unless the file already ends with THIS version's stamp, so a stale or a
 # half-written file self-heals here instead of costing you a round trip to rm it.
-grep -qs '^CODEMAN_PREAMBLE=1.33.4$' "$PRE" || (umask 077; cat > "$PRE" <<'PREAMBLE'
-# ---- Codeman agent preamble 1.33.4 (seeded by Codeman at session spawn; the SKILL.md §0 bootstrap rewrites it when missing or stale) ----
+grep -qs '^CODEMAN_PREAMBLE=1.42.0$' "$PRE" || (umask 077; cat > "$PRE" <<'PREAMBLE'
+# ---- Codeman agent preamble 1.42.0 (seeded by Codeman at session spawn; the SKILL.md §0 bootstrap rewrites it when missing or stale) ----
+# Positional parameters are always written with braces, ${1} and never a bare dollar-digit: Claude Code
+# replaces a bare one in a skill body with the matching word of the skill's arguments, and §0 is pasted
+# from that body (test/agent-skill.test.ts pins it).
 API="${CODEMAN_API_URL:?CODEMAN_API_URL not set; refusing to guess}"
 SELF="${CODEMAN_SESSION_ID:?CODEMAN_SESSION_ID not set}"
 # Credentials, cheapest first. Your session has usually INHERITED the server's
@@ -85,7 +111,7 @@ SELF="${CODEMAN_SESSION_ID:?CODEMAN_SESSION_ID not set}"
 # reads. The data dir is wherever the hook-secret file lives. Values may be
 # quoted or `export`-prefixed.
 ENV_FILE="${CODEMAN_HOOK_SECRET_FILE:+${CODEMAN_HOOK_SECRET_FILE%hook-secret}.env}"
-envval() { sed -n "s/^\(export \)\{0,1\}$1=//p" "$ENV_FILE" | tail -1 | sed 's/^"\(.*\)"$/\1/; s/^'\''\(.*\)'\''$/\1/'; }
+envval() { sed -n "s/^\(export \)\{0,1\}${1}=//p" "$ENV_FILE" | tail -1 | sed 's/^"\(.*\)"$/\1/; s/^'\''\(.*\)'\''$/\1/'; }
 if [ -z "${CODEMAN_PASSWORD:-}" ] && [ -n "$ENV_FILE" ] && [ -f "$ENV_FILE" ]; then
   CODEMAN_USERNAME=$(envval CODEMAN_USERNAME)
   CODEMAN_PASSWORD=$(envval CODEMAN_PASSWORD)
@@ -120,15 +146,15 @@ delete_session() {
 
 # ---- fast path: the four verbs, already written. §1 composes them. ----
 _composer_up() {   # <sid> <timeoutMs> -> "true"/"false". `shift+tab` is the one token
-  "${CURL[@]}" -G "$API/api/v1/sessions/$1/wait-output" \
+  "${CURL[@]}" -G "$API/api/v1/sessions/${1}/wait-output" \
     --data-urlencode 'match=shift+tab' --data-urlencode 'from=buffer' \
-    --data-urlencode "timeout=$2" | jq -r '.data.wait.matched // false'
+    --data-urlencode "timeout=${2}" | jq -r '.data.wait.matched // false'
 }
 _dsh_up() {        # <sid> <timeoutMs> -> "true"/"false". The DeepSeek Harness TUI's
   # composer glyph. Override with DSH_READY_MARK for a profile that draws another one.
-  "${CURL[@]}" -G "$API/api/v1/sessions/$1/wait-output" \
+  "${CURL[@]}" -G "$API/api/v1/sessions/${1}/wait-output" \
     --data-urlencode "match=${DSH_READY_MARK:-❯}" --data-urlencode 'from=buffer' \
-    --data-urlencode "timeout=$2" | jq -r '.data.wait.matched // false'
+    --data-urlencode "timeout=${2}" | jq -r '.data.wait.matched // false'
 }
 # ---- the workspace-trust dialog: READ the screen, never press Enter blind ----
 # Claude Code 2.1.252 dropped the option numbers, REVERSED them, and highlights
@@ -144,7 +170,7 @@ _trust_key() {     # <sid> -> "confirm" | "move" | "" (nothing safe to press)
   # full=1 returns the RENDERED pane; a claude pane keeps no tmux history, so that
   # is the current frame rather than every repaint since launch. tail -1 anyway,
   # because the freshest marked row is the only one still true.
-  "${CURL[@]}" -G "$API/api/v1/sessions/$1/terminal" --data-urlencode 'full=1' \
+  "${CURL[@]}" -G "$API/api/v1/sessions/${1}/terminal" --data-urlencode 'full=1' \
     | jq -r '.data.terminalBuffer // empty' \
     | sed -e "s/$(printf '\033')\[[0-9;?]*[a-zA-Z]//g" -e "s/$(printf '\033')[()][AB0]//g" \
     | tr -d ' \t' | grep -i '❯[0-9.]*\(yes,itrustthisfolder\|no,exit\)' | tail -1 \
@@ -162,7 +188,7 @@ _composer_text() { # <sid> -> the composer's text with ALL whitespace removed: "
   # the LAST `❯` line: Claude Code echoes a submitted prompt with the same glyph higher
   # up in the transcript, so only the last one says whether the text was taken.
   local t
-  t=$("${CURL[@]}" -G "$API/api/v1/sessions/$1/terminal" --data-urlencode 'full=1' \
+  t=$("${CURL[@]}" -G "$API/api/v1/sessions/${1}/terminal" --data-urlencode 'full=1' \
     | jq -r '.data.terminalBuffer // empty' \
     | sed -e "s/$(printf '\033')\[[0-9;?]*[a-zA-Z]//g" -e "s/$(printf '\033')[()][AB0]//g" \
     | tr -d '\r' | grep -a '^[[:space:]]*❯' | tail -1)
@@ -172,7 +198,7 @@ _composer_text() { # <sid> -> the composer's text with ALL whitespace removed: "
   printf '%s' "$t" | sed 's/^[[:space:]]*❯//' | tr -d '[:space:]' | sed "s/$(printf '\302\240')//g"
 }
 _accept_trust() {  # <sid> -> 0 once it has answered the dialog, 1 if it could not
-  local sid="$1" k i=1
+  local sid="${1}" k i=1
   while [ "$i" -le 6 ]; do
     k=$(_trust_key "$sid")
     [ -n "$k" ] || return 1   # no dialog on screen, or a layout this cannot read
@@ -373,7 +399,7 @@ sendwait() {
 last_text() {
   local t="" prev="${2:-}"
   for _ in $(seq 1 15); do
-    t=$("${CURL[@]}" "$API/api/v1/sessions/$1/last-response" | jq -r '.data.text // empty')
+    t=$("${CURL[@]}" "$API/api/v1/sessions/${1}/last-response" | jq -r '.data.text // empty')
     [ -n "$t" ] && [ "$t" != "$prev" ] && { printf '%s\n' "$t"; return 0; }
     sleep 1
   done
@@ -384,10 +410,10 @@ last_text() {
 # The stamp is the LAST line on purpose (a truncated write leaves it unset) and is kept
 # bare on purpose: the write condition above anchors on it with $, so an inline comment
 # here would fail that match and rewrite this file on every single bootstrap.
-CODEMAN_PREAMBLE=1.33.4
+CODEMAN_PREAMBLE=1.42.0
 PREAMBLE
 )
-. "$PRE"; [ "${CODEMAN_PREAMBLE:-}" = 1.33.4 ] || { echo "preamble at $PRE is stale or truncated: rm it and re-run this block"; exit 1; }
+. "$PRE"; [ "${CODEMAN_PREAMBLE:-}" = 1.42.0 ] || { echo "preamble at $PRE is stale or truncated: rm it and re-run this block"; exit 1; }
 ```
 
 Every later Bash call that touches the API starts with the same two loader lines from
@@ -438,7 +464,7 @@ and no per-call body to hand-build.
 
 ```bash
 . "${XDG_CACHE_HOME:-$HOME/.cache}/codeman-agent-$CODEMAN_SESSION_ID.sh" 2>/dev/null   # §0 loader
-[ "${CODEMAN_PREAMBLE:-}" = 1.33.4 ] || { echo "preamble missing or stale; run the full §0 block"; exit 1; }
+[ "${CODEMAN_PREAMBLE:-}" = 1.42.0 ] || { echo "preamble missing or stale; run the full §0 block"; exit 1; }
 N=(alpha beta)                    # INVENT one fresh case name per worker; never list cases first
                                   # (a name may carry a mode: `beta:deepseek`, see below)
 T=('reply with one line: the absolute path of your working directory'

skills/codeman/preamble.sh

@@ -1,4 +1,7 @@
-# ---- Codeman agent preamble 1.33.4 (seeded by Codeman at session spawn; the SKILL.md §0 bootstrap rewrites it when missing or stale) ----
+# ---- Codeman agent preamble 1.42.0 (seeded by Codeman at session spawn; the SKILL.md §0 bootstrap rewrites it when missing or stale) ----
+# Positional parameters are always written with braces, ${1} and never a bare dollar-digit: Claude Code
+# replaces a bare one in a skill body with the matching word of the skill's arguments, and §0 is pasted
+# from that body (test/agent-skill.test.ts pins it).
 API="${CODEMAN_API_URL:?CODEMAN_API_URL not set; refusing to guess}"
 SELF="${CODEMAN_SESSION_ID:?CODEMAN_SESSION_ID not set}"
 # Credentials, cheapest first. Your session has usually INHERITED the server's
@@ -7,7 +10,7 @@ SELF="${CODEMAN_SESSION_ID:?CODEMAN_SESSION_ID not set}"
 # reads. The data dir is wherever the hook-secret file lives. Values may be
 # quoted or `export`-prefixed.
 ENV_FILE="${CODEMAN_HOOK_SECRET_FILE:+${CODEMAN_HOOK_SECRET_FILE%hook-secret}.env}"
-envval() { sed -n "s/^\(export \)\{0,1\}$1=//p" "$ENV_FILE" | tail -1 | sed 's/^"\(.*\)"$/\1/; s/^'\''\(.*\)'\''$/\1/'; }
+envval() { sed -n "s/^\(export \)\{0,1\}${1}=//p" "$ENV_FILE" | tail -1 | sed 's/^"\(.*\)"$/\1/; s/^'\''\(.*\)'\''$/\1/'; }
 if [ -z "${CODEMAN_PASSWORD:-}" ] && [ -n "$ENV_FILE" ] && [ -f "$ENV_FILE" ]; then
   CODEMAN_USERNAME=$(envval CODEMAN_USERNAME)
   CODEMAN_PASSWORD=$(envval CODEMAN_PASSWORD)
@@ -42,15 +45,15 @@ delete_session() {
 
 # ---- fast path: the four verbs, already written. §1 composes them. ----
 _composer_up() {   # <sid> <timeoutMs> -> "true"/"false". `shift+tab` is the one token
-  "${CURL[@]}" -G "$API/api/v1/sessions/$1/wait-output" \
+  "${CURL[@]}" -G "$API/api/v1/sessions/${1}/wait-output" \
     --data-urlencode 'match=shift+tab' --data-urlencode 'from=buffer' \
-    --data-urlencode "timeout=$2" | jq -r '.data.wait.matched // false'
+    --data-urlencode "timeout=${2}" | jq -r '.data.wait.matched // false'
 }
 _dsh_up() {        # <sid> <timeoutMs> -> "true"/"false". The DeepSeek Harness TUI's
   # composer glyph. Override with DSH_READY_MARK for a profile that draws another one.
-  "${CURL[@]}" -G "$API/api/v1/sessions/$1/wait-output" \
+  "${CURL[@]}" -G "$API/api/v1/sessions/${1}/wait-output" \
     --data-urlencode "match=${DSH_READY_MARK:-❯}" --data-urlencode 'from=buffer' \
-    --data-urlencode "timeout=$2" | jq -r '.data.wait.matched // false'
+    --data-urlencode "timeout=${2}" | jq -r '.data.wait.matched // false'
 }
 # ---- the workspace-trust dialog: READ the screen, never press Enter blind ----
 # Claude Code 2.1.252 dropped the option numbers, REVERSED them, and highlights
@@ -66,7 +69,7 @@ _trust_key() {     # <sid> -> "confirm" | "move" | "" (nothing safe to press)
   # full=1 returns the RENDERED pane; a claude pane keeps no tmux history, so that
   # is the current frame rather than every repaint since launch. tail -1 anyway,
   # because the freshest marked row is the only one still true.
-  "${CURL[@]}" -G "$API/api/v1/sessions/$1/terminal" --data-urlencode 'full=1' \
+  "${CURL[@]}" -G "$API/api/v1/sessions/${1}/terminal" --data-urlencode 'full=1' \
     | jq -r '.data.terminalBuffer // empty' \
     | sed -e "s/$(printf '\033')\[[0-9;?]*[a-zA-Z]//g" -e "s/$(printf '\033')[()][AB0]//g" \
     | tr -d ' \t' | grep -i '❯[0-9.]*\(yes,itrustthisfolder\|no,exit\)' | tail -1 \
@@ -84,7 +87,7 @@ _composer_text() { # <sid> -> the composer's text with ALL whitespace removed: "
   # the LAST `❯` line: Claude Code echoes a submitted prompt with the same glyph higher
   # up in the transcript, so only the last one says whether the text was taken.
   local t
-  t=$("${CURL[@]}" -G "$API/api/v1/sessions/$1/terminal" --data-urlencode 'full=1' \
+  t=$("${CURL[@]}" -G "$API/api/v1/sessions/${1}/terminal" --data-urlencode 'full=1' \
     | jq -r '.data.terminalBuffer // empty' \
     | sed -e "s/$(printf '\033')\[[0-9;?]*[a-zA-Z]//g" -e "s/$(printf '\033')[()][AB0]//g" \
     | tr -d '\r' | grep -a '^[[:space:]]*❯' | tail -1)
@@ -94,7 +97,7 @@ _composer_text() { # <sid> -> the composer's text with ALL whitespace removed: "
   printf '%s' "$t" | sed 's/^[[:space:]]*❯//' | tr -d '[:space:]' | sed "s/$(printf '\302\240')//g"
 }
 _accept_trust() {  # <sid> -> 0 once it has answered the dialog, 1 if it could not
-  local sid="$1" k i=1
+  local sid="${1}" k i=1
   while [ "$i" -le 6 ]; do
     k=$(_trust_key "$sid")
     [ -n "$k" ] || return 1   # no dialog on screen, or a layout this cannot read
@@ -295,7 +298,7 @@ sendwait() {
 last_text() {
   local t="" prev="${2:-}"
   for _ in $(seq 1 15); do
-    t=$("${CURL[@]}" "$API/api/v1/sessions/$1/last-response" | jq -r '.data.text // empty')
+    t=$("${CURL[@]}" "$API/api/v1/sessions/${1}/last-response" | jq -r '.data.text // empty')
     [ -n "$t" ] && [ "$t" != "$prev" ] && { printf '%s\n' "$t"; return 0; }
     sleep 1
   done
@@ -306,4 +309,4 @@ last_text() {
 # The stamp is the LAST line on purpose (a truncated write leaves it unset) and is kept
 # bare on purpose: the write condition above anchors on it with $, so an inline comment
 # here would fail that match and rewrite this file on every single bootstrap.
-CODEMAN_PREAMBLE=1.33.4
+CODEMAN_PREAMBLE=1.42.0

skills/codeman/reference/endpoints.md

@@ -635,7 +635,7 @@ Three bounded long-polls. Shared semantics:
 | `idle` | output stabilized + prompt detected, heuristic, can flap mid-turn | every mode |
 | `working` | session started producing output | every mode |
 | `stop` | Claude Code `stop` hook, the definitive end-of-turn | `claude` only |
-| `blocked` | `permission_prompt` / `elicitation_dialog` hook, the worker needs an answer | `claude` only |
+| `blocked` | `permission_prompt` / `elicitation_dialog` hook, the worker needs an answer | `claude`; also `codex`, whose approval, question and trust dialogs Codeman reads off its screen (answer them in the terminal, never through the inbox) |
 | `exit` | PTY exited or session deleted | every mode |
 
 ⚠️ **`claude` mode is necessary for `stop`/`blocked`, not sufficient. The real

skills/codeman/reference/recipes.md

@@ -21,7 +21,7 @@ by sourcing the preamble file the §0 bootstrap wrote, and checking its version
 
 ```bash
 . "${XDG_CACHE_HOME:-$HOME/.cache}/codeman-agent-$CODEMAN_SESSION_ID.sh" 2>/dev/null
-[ "${CODEMAN_PREAMBLE:-}" = 1.33.4 ] || { echo "preamble missing or stale; re-run the §0 bootstrap"; exit 1; }
+[ "${CODEMAN_PREAMBLE:-}" = 1.42.0 ] || { echo "preamble missing or stale; re-run the §0 bootstrap"; exit 1; }
 ```
 
 Do **not** re-paste the preamble body into each call. Sourcing it is what retires the
@@ -391,10 +391,10 @@ the worker remembering to print a token.
 
 ## Flow 5: watch for a worker stuck on a prompt
 
-Claude workers can block on a permission dialog. `blocked` is a wait signal
-(claude-mode only, and it needs Codeman's hooks in the worker's directory: see Flow 7
-step 4), so watch for it and surface the question to the user instead of guessing an
-answer. Expect it routinely on a server whose `claudeMode` is not the default bypass
+Claude workers can block on a permission dialog. `blocked` is a wait signal (a claude
+worker needs Codeman's hooks in its directory for it: see Flow 7 step 4; a codex worker
+gets it from its screen, see below), so watch for it and surface the question to the user
+instead of guessing an answer. Expect it routinely on a server whose `claudeMode` is not the default bypass
 one (the same setting that decides whether the readiness marker in Flow 1 ever
 appears):
 
@@ -409,10 +409,13 @@ if [ "$(jq -r '.data.wait.signal' <<<"$R")" = blocked ]; then
 fi
 ```
 
-Where the worker has no hooks, `blocked` never fires and a stuck worker looks exactly
-like a slow one: your marker wait burns its whole bound. The fallback is the same
-terminal tail, taken when a bound runs out, and the same rule about not answering it
-yourself.
+Codex workers are the one exception without hooks: Codeman reads their approval,
+question and trust dialogs off the screen, so the same watch works with
+`until=blocked,exit` (codex has no `stop`, a `400`), and the dialog is answered in the
+terminal, never through the Approvals Inbox. Anywhere else without hooks, `blocked` never
+fires and a stuck worker looks exactly like a slow one: your marker wait burns its whole
+bound. The fallback is the same terminal tail, taken when a bound runs out, and the same
+rule about not answering it yourself.
 
 ## Flow 6: claude fan-out over messaging
 

skills/codeman/reference/verbs.md

@@ -377,7 +377,9 @@ only when the workspace actually has them, see [§5.1](#51-where-to-spawn)) **an
 `deepseek`** — the one external CLI that reports its own lifecycle, so its `stop` is a
 real end-of-turn signal rather than a guess. On
 `shell`/`opencode`/`codex`/`gemini`/`antigravity`/`pi`/`grok`/`omp`, requesting them explicitly is a
-400, and lifecycle transitions there are coarse (a short shell command may emit **no**
+400 (except `blocked` on `codex`, which Codeman reads off its screen when an approval,
+question or trust dialog is up; a wait for it answers at once if one already is), and
+lifecycle transitions there are coarse (a short shell command may emit **no**
 `idle` transition at all, verified live), so synchronize those with markers.
 
 ⚠️ A dsh session can still refuse them for a per-SESSION reason: `statusReporting:

src/attachment-registry.ts

@@ -9,7 +9,8 @@
 import { randomUUID } from 'node:crypto';
 import { realpathSync } from 'node:fs';
 import fs from 'node:fs/promises';
-import { basename, extname, isAbsolute } from 'node:path';
+import { homedir } from 'node:os';
+import { basename, extname, isAbsolute, join } from 'node:path';
 import { isBlockedAttachmentPath, isUnderTree, loadAttachmentGuardConfig } from './config/attachment-guard.js';
 import { EDITABLE_EXTENSIONS } from './config/file-editing.js';
 import { validateSessionFilePath } from './web/route-helpers.js';
@@ -148,6 +149,11 @@ class AttachmentRegistry {
 
 export const attachmentRegistry = new AttachmentRegistry();
 
+/** Every extension {@link isSupportedAttachmentExtension} accepts (lowercase, no dot). */
+export function supportedAttachmentExtensions(): string[] {
+  return [...SUPPORTED_ATTACHMENT_EXTENSIONS];
+}
+
 export function isSupportedAttachmentExtension(extension: string): boolean {
   return SUPPORTED_ATTACHMENT_EXTENSIONS.has(extension.toLowerCase().replace(/^\./, ''));
 }
@@ -330,13 +336,19 @@ export async function registerExternalAttachment(
   requestedPath: string,
   options: RegisterExternalAttachmentOptions = {}
 ): Promise<AttachmentRegistrationResult> {
-  if (!requestedPath || !isAbsolute(requestedPath)) {
+  // `~/…` is the home of the session's HOST: expanded here for a local session,
+  // and by the probe on the remote host for a remote one (`buildRemoteProbeCommand`),
+  // whose home this server does not know. A clicked terminal link sends it as is
+  // (Claude Code echoes attachments as `~/…`). Every guard below runs on the
+  // resolved absolute path either way.
+  const homeRelative = typeof requestedPath === 'string' && requestedPath.startsWith('~/');
+  if (!requestedPath || (!isAbsolute(requestedPath) && !homeRelative)) {
     throw new AttachmentRegistrationError('Attachment path must be an absolute local path');
   }
 
   const resolved = await (options.remote
     ? resolveRemoteAttachment(requestedPath, options.remote, options.sessionWorkingDir, options.remoteProbes)
-    : resolveLocalAttachment(requestedPath));
+    : resolveLocalAttachment(homeRelative ? join(homedir(), requestedPath.slice(2)) : requestedPath));
 
   // COD-53: enforce the active attachment-guard policy on the symlink-resolved
   // path before doing anything else.

src/cli-agent.ts

@@ -0,0 +1,985 @@
+/**
+ * @fileoverview `codeman agent …` — session-to-session verbs for the agent running
+ * inside a Codeman session, in every CLI mode.
+ *
+ * A thin HTTP client over endpoints that already exist (`quick-start`, `input`,
+ * `wait`, `wait-output`, `last-response`, `terminal`, `DELETE sessions/:id`). It
+ * invents no route and no transport: everything goes through `CODEMAN_API_URL`, so
+ * auth, ownership and the per-session waiter cap apply unchanged. The behaviour is
+ * the packaged agent skill's (`skills/codeman`), ported from shell prose into code
+ * with tests, so a `codex`/`opencode`/`pi` agent — which never gets the claude-only
+ * preamble — has the same verbs from one line in its AGENTS.md.
+ *
+ * Invariants (each asserted in `test/cli-agent.test.ts`):
+ * 1. Refuses outside a Codeman session (`CODEMAN_MUX=1` + `CODEMAN_API_URL`); it
+ *    never guesses a URL — a server you are not part of is not yours to drive.
+ * 2. `send` transmits printable text plus `\r` only. ESC exists solely as
+ *    `interrupt`, which never appends `\r`. A stray control byte is a dead session
+ *    in the fullscreen TUIs (opencode's `Ctrl+C` is `app_exit`).
+ * 3. `rm` fails closed: empty id, a short self id, or a prefix match in EITHER
+ *    direction refuses. Ids appear in full and 8-char form, so equality alone
+ *    misses a real combination — and the miss deletes the caller.
+ * 4. An id shorter than 8 characters refuses (exit 4) before any request, on every
+ *    verb. `9` would resolve to whichever session is alone with that first
+ *    character, the user's own interactive tab included.
+ *
+ * Commands live here as functions returning an exit code, not calling
+ * `process.exit`, so the whole surface is unit-testable against a fake server.
+ *
+ * @module cli-agent
+ */
+import http from 'node:http';
+import https from 'node:https';
+import type { Command } from 'commander';
+import {
+  basicAuthHeader,
+  credentialsFrom,
+  readCodemanEnvFile,
+  type CodemanCredentials,
+} from './codeman-credentials.js';
+import { getCli } from './config/cli-registry/registry.js';
+import { GLYPH, palette, table } from './cli-style.js';
+import { getErrorMessage } from './types.js';
+import { stripAnsi as stripAnsiSequences } from './utils/regex-patterns.js';
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Context and guard
+// ─────────────────────────────────────────────────────────────────────────────
+
+export interface AgentContext {
+  /** Base URL of the Codeman server, from `CODEMAN_API_URL`. */
+  apiUrl: string;
+  /** This session's id, from `CODEMAN_SESSION_ID`. */
+  selfId: string;
+  /** Basic-auth credentials, when the server has a password. */
+  auth?: CodemanCredentials;
+}
+
+/** Thrown when the process is not inside a Codeman-managed session. */
+export class AgentGuardError extends Error {}
+
+/** Exit codes shared by every verb; a shell agent can branch on them. */
+export const EXIT = {
+  ok: 0,
+  error: 1,
+  timeout: 2,
+  dead: 3,
+  refused: 4,
+} as const;
+
+/**
+ * Resolve the context from the environment, or throw `AgentGuardError`.
+ *
+ * Credentials in the order every client of the API uses (`credentialsFrom`, shared
+ * with `codeman attach` and the TUI): each field from the environment (a session
+ * inherits the server's), then the data dir's `.env`. No password means the server
+ * is open (single-user) — or it is not, and the 401 says so.
+ */
+export function resolveAgentContext(
+  env: NodeJS.ProcessEnv = process.env,
+  envFile: () => Record<string, string> = readCodemanEnvFile
+): AgentContext {
+  if (env.CODEMAN_MUX !== '1') {
+    throw new AgentGuardError('Not inside a Codeman-managed session (CODEMAN_MUX is not 1); refusing to act.');
+  }
+  const apiUrl = env.CODEMAN_API_URL?.trim();
+  if (!apiUrl) {
+    throw new AgentGuardError('CODEMAN_API_URL is not set; refusing to guess a server.');
+  }
+  const selfId = env.CODEMAN_SESSION_ID?.trim();
+  if (!selfId) {
+    throw new AgentGuardError('CODEMAN_SESSION_ID is not set; cannot tell which session is me.');
+  }
+  const credentials = credentialsFrom(env, envFile());
+  return { apiUrl, selfId, auth: credentials.password ? credentials : undefined };
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Pure helpers (the invariants)
+// ─────────────────────────────────────────────────────────────────────────────
+
+/**
+ * Is `id` this session? Prefix in BOTH directions, because ids appear in full and
+ * in 8-char form (mux names, UI surfaces, Docker's truncated `$SELF`). A self id
+ * shorter than 8 characters cannot prove anything and is treated as "maybe me".
+ */
+export function isSelfSession(selfId: string, id: string): boolean {
+  if (!id || selfId.length < 8) return true;
+  return id.startsWith(selfId) || selfId.startsWith(id);
+}
+
+/** Why `rm` refuses, or `undefined` when the delete may go ahead. */
+export function deleteRefusal(selfId: string, id: string): string | undefined {
+  if (!id) return 'refusing: empty session id';
+  if (selfId.length < 8) return 'refusing: own session id unset or too short to prove this is not me';
+  if (isSelfSession(selfId, id)) return `refusing: ${id} is me`;
+  return undefined;
+}
+
+/**
+ * The prompt `send` was given, which must be ONE argument. Joining several with spaces
+ * would turn an unquoted `$(cat notes.txt)`, which the shell splits on every newline,
+ * back into a single line, so the multi-line refusal below would never see it.
+ */
+export function sendPromptFromArgs(words: readonly string[]): { text: string } | { error: string } {
+  if (words.length === 1) return { text: words[0] };
+  return {
+    error: `refusing: the prompt must be ONE argument, got ${words.length} — quote it (\`send <id> "…"\`; a prompt that starts with "-" goes after --: \`send <id> -- "- fix the bug"\`)`,
+  };
+}
+
+/**
+ * Why `send` refuses this text, or `undefined` when it is printable. The composer
+ * takes one line; the server strips `\r`/`\n` but everything else below 0x20 (and
+ * DEL) reaches the pane as a keypress. None of that is a prompt.
+ */
+export function inputRefusal(text: string): string | undefined {
+  if (text.length === 0) return 'refusing: empty input (use `interrupt` for ESC, `send <id> ""` is never a prompt)';
+  // The composer is one line: the server strips newlines, which silently joins the
+  // lines into one prompt, and a tab reaches the pane as a keypress (claude: mode toggle).
+  if (/[\n\r\t]/.test(text)) {
+    return 'refusing: input must be a single line (the composer strips newlines and would join your lines) — join them yourself, or write a file into the workspace and send its path';
+  }
+  // C0, DEL and C1 (U+0080–U+009F: an 8-bit CSI is still a CSI to a terminal).
+  // eslint-disable-next-line no-control-regex
+  const control = text.match(/[\x00-\x1f\x7f-\x9f]/);
+  if (control) {
+    const code = control[0].charCodeAt(0).toString(16).padStart(2, '0');
+    return `refusing: input contains control byte 0x${code}; send transmits printable text only (ESC is \`interrupt\`)`;
+  }
+  return undefined;
+}
+
+/**
+ * Body for `POST /sessions/:id/input` on the send path: text plus `\r` unless the
+ * caller asked to type without submitting. Never anything else.
+ */
+export function buildSendBody(
+  text: string,
+  options: { enter: boolean; clientId: string; seq: number; wait?: string | true; waitTimeout?: number }
+): Record<string, unknown> {
+  const body: Record<string, unknown> = {
+    input: options.enter ? `${text}\r` : text,
+    useMux: true,
+    clientId: options.clientId,
+    seq: options.seq,
+  };
+  if (options.wait !== undefined) body.wait = options.wait;
+  if (options.waitTimeout !== undefined) body.waitTimeout = options.waitTimeout;
+  return body;
+}
+
+/** Body for the interrupt path: a bare ESC, and nothing appended — ever. */
+export function buildInterruptBody(clientId: string, seq: number): Record<string, unknown> {
+  return { input: '\u001b', useMux: true, clientId, seq };
+}
+
+/** `clientId` for this caller: fixed per sending session, so `seq` stays monotonic. */
+export function defaultClientId(selfId: string, suffix = ''): string {
+  return `codeman-agent-cli-${selfId.slice(0, 8)}${suffix ? `-${suffix}` : ''}`;
+}
+
+/**
+ * Strip a terminal buffer for humans: the shared ANSI strip (CSI, OSC such as window
+ * titles, keypad modes) plus the charset designators (`ESC ( B`) it leaves in.
+ */
+export function stripAnsi(text: string): string {
+  // eslint-disable-next-line no-control-regex
+  return stripAnsiSequences(text).replace(/\x1b[()][AB0]/g, '');
+}
+
+/** Parse a positive-integer option (`--timeout` ms, `--tail` bytes); the server rejects anything else. */
+export function parsePositiveInt(raw: string | undefined, fallback: number, flag = '--timeout'): number {
+  if (raw === undefined) return fallback;
+  const n = Number(raw);
+  if (!Number.isInteger(n) || n <= 0) {
+    throw new Error(`${flag} must be a positive integer, got "${raw}"`);
+  }
+  return n;
+}
+
+/**
+ * Exit code for a wait result: matched or a signal → ok, `exit` or `ended` → dead,
+ * timeout → timeout. `ended` is checked BEFORE the happy paths: a worker that dies
+ * during `--until stop` comes back as `ended:true, signal:null` (the registry only
+ * satisfies waiters that listed `exit`, then cancels the rest), and a `--match` on
+ * a dead worker as `ended:true, matched:false` — both are "dead", never "done".
+ */
+export function waitExitCode(wait: WaitResult | undefined): number {
+  if (!wait) return EXIT.error;
+  if (wait.signal === 'exit' || wait.ended) return EXIT.dead;
+  if (wait.timedOut) return EXIT.timeout;
+  if (wait.matched === false) return EXIT.timeout;
+  return EXIT.ok;
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// HTTP
+// ─────────────────────────────────────────────────────────────────────────────
+
+export interface ApiEnvelope<T = unknown> {
+  success: boolean;
+  data?: T;
+  error?: string;
+  errorCode?: string;
+}
+
+export interface ApiResponse<T = unknown> {
+  status: number;
+  /** Parsed envelope, or `undefined` when the body was not JSON (auth guards answer in plain text). */
+  json?: ApiEnvelope<T>;
+  text: string;
+}
+
+export interface WaitResult {
+  /** The signal that fired, or null when the wait ended without one. */
+  signal?: string | null;
+  timedOut?: boolean;
+  /** The session went away (deleted / torn down / the write failed) before the wait resolved. */
+  ended?: boolean;
+  timeoutMs?: number;
+  until?: string[];
+  matched?: boolean;
+  match?: string;
+  snippet?: string;
+  immediate?: boolean;
+}
+
+export interface RequestOptions {
+  method: 'GET' | 'POST' | 'DELETE';
+  path: string;
+  query?: Record<string, string | number | boolean | undefined>;
+  body?: Record<string, unknown>;
+  headers?: Record<string, string>;
+  /** Socket timeout; long-polls pass their own timeout plus headroom. */
+  timeoutMs?: number;
+  /** Aborts the request (an MCP client cancelling a long wait); the promise rejects. */
+  signal?: AbortSignal;
+}
+
+export type ApiRequest = (ctx: AgentContext, options: RequestOptions) => Promise<ApiResponse>;
+
+/** Every request carries these; they are ignored on endpoints that do not read them. */
+export function baseHeaders(ctx: AgentContext): Record<string, string> {
+  const headers: Record<string, string> = { Accept: 'application/json' };
+  // Tags sessions this caller spawns as its children (lineage in the web UI).
+  // Cosmetic, never fails a call. NOT X-Codeman-Agent-Origin: that one marks a case
+  // directory as deletable agent scratch, so it rides only the spawn request that
+  // may create one (see agentSpawn), never anything else. Absent for a caller outside
+  // any session (`codeman mcp` launched by a desktop MCP client), which has no self.
+  if (ctx.selfId) headers['X-Codeman-Parent-Session'] = ctx.selfId;
+  const authorization = ctx.auth ? basicAuthHeader(ctx.auth) : undefined;
+  if (authorization) headers.Authorization = authorization;
+  return headers;
+}
+
+/** The real transport. `rejectUnauthorized:false` because the HTTPS install uses a self-signed cert. */
+export const httpRequest: ApiRequest = (ctx, options) => {
+  const url = new URL(options.path, ctx.apiUrl);
+  for (const [key, value] of Object.entries(options.query ?? {})) {
+    if (value !== undefined) url.searchParams.set(key, String(value));
+  }
+  const bodyText = options.body === undefined ? undefined : JSON.stringify(options.body);
+  const headers: Record<string, string | number> = { ...baseHeaders(ctx), ...(options.headers ?? {}) };
+  if (bodyText !== undefined) {
+    headers['Content-Type'] = 'application/json';
+    headers['Content-Length'] = Buffer.byteLength(bodyText);
+  }
+  const transport = url.protocol === 'https:' ? https : http;
+  return new Promise((resolve, reject) => {
+    const req = transport.request(
+      {
+        protocol: url.protocol,
+        hostname: url.hostname,
+        port: url.port,
+        method: options.method,
+        path: `${url.pathname}${url.search}`,
+        rejectUnauthorized: false,
+        headers,
+        timeout: options.timeoutMs ?? 30_000,
+        signal: options.signal,
+      },
+      (res) => {
+        const chunks: Buffer[] = [];
+        res.on('data', (chunk: Buffer) => chunks.push(chunk));
+        // A connection dropped mid-body emits 'error' and never 'end'; without this the
+        // promise never settles (and a long-lived `codeman mcp` never answers the call).
+        res.on('error', reject);
+        res.on('end', () => {
+          const text = Buffer.concat(chunks).toString('utf-8');
+          let json: ApiEnvelope | undefined;
+          try {
+            json = JSON.parse(text) as ApiEnvelope;
+          } catch {
+            json = undefined;
+          }
+          resolve({ status: res.statusCode ?? 0, json, text });
+        });
+      }
+    );
+    req.on('timeout', () => req.destroy(new Error(`request timed out after ${options.timeoutMs ?? 30_000} ms`)));
+    req.on('error', reject);
+    if (bodyText !== undefined) req.write(bodyText);
+    req.end();
+  });
+};
+
+/** One line describing a failed response, for humans. Plain-text guards (401/403/429) have no envelope. */
+export function describeFailure(res: ApiResponse): string {
+  if (res.json && !res.json.success) {
+    return `${res.json.errorCode ?? 'ERROR'}: ${res.json.error ?? 'request failed'} (HTTP ${res.status})`;
+  }
+  const text = res.text.trim().split('\n')[0] ?? '';
+  if (res.status === 401)
+    return `HTTP 401 ${text}: the server wants a password (CODEMAN_PASSWORD, or the data dir's .env)`;
+  return `HTTP ${res.status}${text ? ` ${text}` : ''}`;
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Commands
+// ─────────────────────────────────────────────────────────────────────────────
+
+export interface AgentIo {
+  out: (line: string) => void;
+  err: (line: string) => void;
+}
+
+export interface AgentDeps {
+  ctx: AgentContext;
+  request: ApiRequest;
+  io: AgentIo;
+  json: boolean;
+  /** Clock for `seq`; injectable so tests are deterministic. */
+  now?: () => number;
+}
+
+/** Print `data` as JSON (the `--json` path) — always the envelope's `data`, never a reshaped copy. */
+function emitJson(deps: AgentDeps, data: unknown): void {
+  deps.io.out(JSON.stringify(data, null, 2));
+}
+
+function fail(deps: AgentDeps, message: string, code: number = EXIT.error): number {
+  if (deps.json) {
+    deps.io.out(JSON.stringify({ success: false, error: message }));
+  } else {
+    deps.io.err(palette.err(`${GLYPH.fail} ${message}`));
+  }
+  return code;
+}
+
+interface SessionRow {
+  id: string;
+  name?: string;
+  mode?: string;
+  status?: string;
+  workingDir?: string;
+  pid?: number | null;
+  parentSessionId?: string | null;
+}
+
+/** A full session id (the only form the routes accept); `ls` prints the 8-char prefix. */
+const FULL_ID_LENGTH = 36;
+
+/**
+ * Shortest prefix that may name a session: the 8-char form `ls` prints, and the floor
+ * the server's own resolver uses (`PARENT_SESSION_ID_MIN_PREFIX`, route-helpers.ts).
+ */
+export const MIN_ID_PREFIX_LENGTH = 8;
+
+/**
+ * Turn the id a human typed into the one the routes accept. `ls` prints 8-char
+ * prefixes and the routes answer 404 to those (measured live), so anything shorter
+ * than a full id resolves through the session list; an ambiguous prefix refuses
+ * rather than picking one. Below 8 characters it refuses before the list: "unique"
+ * means nothing for `9` — it names whatever session happens to be alone with that
+ * first character, and `rm`/`send` would act on it.
+ */
+export async function resolveSessionId(
+  deps: AgentDeps,
+  id: string
+): Promise<{ id: string } | { error: string; code: number }> {
+  if (!id) return { error: 'refusing: empty session id', code: EXIT.refused };
+  if (id.length < MIN_ID_PREFIX_LENGTH) {
+    return {
+      error: `refusing: "${id}" is shorter than ${MIN_ID_PREFIX_LENGTH} characters — use the 8-character id \`agent ls\` prints, or the full id`,
+      code: EXIT.refused,
+    };
+  }
+  if (id.length >= FULL_ID_LENGTH) return { id };
+  const res = await deps.request(deps.ctx, { method: 'GET', path: '/api/v1/sessions' });
+  if (!res.json?.success) return { error: describeFailure(res), code: EXIT.error };
+  const matches = ((res.json.data as SessionRow[] | undefined) ?? []).filter((s) => s.id.startsWith(id));
+  if (matches.length === 1) return { id: matches[0].id };
+  if (matches.length === 0) return { error: `no session starts with "${id}" (see \`agent ls\`)`, code: EXIT.error };
+  return { error: `"${id}" is ambiguous: ${matches.map((s) => s.id.slice(0, 13)).join(', ')}`, code: EXIT.error };
+}
+
+/** `agent ls` — every session the caller can see, self marked. */
+export async function agentLs(deps: AgentDeps): Promise<number> {
+  const res = await deps.request(deps.ctx, { method: 'GET', path: '/api/v1/sessions' });
+  if (!res.json?.success) return fail(deps, describeFailure(res));
+  const sessions = (res.json.data as SessionRow[] | undefined) ?? [];
+  if (deps.json) {
+    emitJson(
+      deps,
+      sessions.map((s) => ({ ...s, self: isSelfSession(deps.ctx.selfId, s.id) }))
+    );
+    return EXIT.ok;
+  }
+  if (sessions.length === 0) {
+    deps.io.out(palette.muted('(no sessions)'));
+    return EXIT.ok;
+  }
+  const rows = sessions.map((s) => [
+    isSelfSession(deps.ctx.selfId, s.id) ? '*' : ' ',
+    s.id.slice(0, 8),
+    s.mode ?? '?',
+    s.status ?? '?',
+    s.name || s.workingDir || '',
+  ]);
+  deps.io.out(table([[' ', 'ID', 'MODE', 'STATUS', 'NAME'], ...rows], { gap: 2 }));
+  deps.io.out(
+    palette.muted(`* = this session (${deps.ctx.selfId.slice(0, 8)}). status is a UI hint, never a sync signal.`)
+  );
+  return EXIT.ok;
+}
+
+export interface SpawnOptions {
+  caseName: string;
+  mode: string;
+  name?: string;
+  /** Wait for the composer before returning, where the registry gives the mode a ready mark. */
+  ready: boolean;
+  timeoutMs: number;
+}
+
+/**
+ * Agent-scratch label for a case directory a spawn CREATES (the server applies it only
+ * when quick-start makes the directory). The Add Case UI offers a recursive delete for
+ * such directories, so this header must never ride any other request: mislabelling a
+ * real repo there is the one failure in this area that costs actual work.
+ */
+export const AGENT_ORIGIN_HEADER = { 'X-Codeman-Agent-Origin': 'codeman-agent-cli' } as const;
+
+/**
+ * What the mode's TUI draws once its composer can take a prompt, from the CLI registry
+ * (`capabilities.composerReadyMark`); undefined means the mode has no readiness wait.
+ */
+export function composerReadyMark(mode: string): string | undefined {
+  return getCli(mode)?.capabilities.composerReadyMark;
+}
+/** `agent spawn` — quick-start with lineage, then the readiness ladder where the mode has one. */
+export async function agentSpawn(deps: AgentDeps, options: SpawnOptions): Promise<number> {
+  const body: Record<string, unknown> = {
+    caseName: options.caseName,
+    mode: options.mode,
+    parentSessionId: deps.ctx.selfId,
+  };
+  if (options.name) body.sessionName = options.name;
+  const res = await deps.request(deps.ctx, {
+    method: 'POST',
+    path: '/api/v1/quick-start',
+    body,
+    headers: { ...AGENT_ORIGIN_HEADER },
+  });
+  const data = res.json?.data as { sessionId?: string; caseName?: string; casePath?: string } | undefined;
+  if (!res.json?.success || !data?.sessionId) return fail(deps, describeFailure(res));
+  const sid = data.sessionId;
+
+  let ready: boolean | undefined;
+  let readinessError: string | undefined;
+  let dead = false;
+  const mark = composerReadyMark(options.mode);
+  if (options.ready && mark) {
+    const wait = await deps.request(deps.ctx, {
+      method: 'GET',
+      path: `/api/v1/sessions/${encodeURIComponent(sid)}/wait-output`,
+      query: { match: mark, from: 'buffer', timeout: options.timeoutMs },
+      timeoutMs: options.timeoutMs + 10_000,
+    });
+    // A failed readiness call (waiter cap, 400, network) is its own error, not "the
+    // composer never showed up": report the real reason instead of the trust-dialog hint.
+    if (!wait.json?.success) readinessError = describeFailure(wait);
+    else {
+      const result = (wait.json.data as { wait?: WaitResult } | undefined)?.wait;
+      // A worker that died while we waited is exit 3 like every other wait, not a
+      // "composer not seen" timeout that sends the caller looking for a dialog.
+      dead = waitExitCode(result) === EXIT.dead;
+      ready = !dead && Boolean(result?.matched);
+    }
+  }
+
+  if (deps.json) {
+    emitJson(deps, { ...data, ready, readinessError });
+  } else {
+    // Human lines go to stderr so `SID=$(codeman agent spawn …)` captures the id alone.
+    const say = (line: string) => deps.io.err(line);
+    say(palette.ok(`${GLYPH.ok} spawned ${sid} (${options.mode}, case ${data.caseName ?? options.caseName})`));
+    if (ready === true) say(palette.muted('  composer up: the worker can take a prompt'));
+    if (dead) say(palette.err(`${GLYPH.fail} the worker exited during the readiness wait`));
+    if (ready === false && !dead) {
+      say(
+        palette.warn(
+          `${GLYPH.warn} composer not seen within ${options.timeoutMs} ms — read \`agent read ${sid.slice(0, 8)} --tail 2000\` before sending (a startup dialog?)`
+        )
+      );
+    }
+    if (readinessError) say(palette.err(`${GLYPH.fail} readiness check failed: ${readinessError}`));
+    if (ready === undefined && !readinessError && options.ready) {
+      say(
+        palette.muted(
+          `  ${options.mode} has no readiness mark; give it a moment, then use --match markers to synchronize`
+        )
+      );
+    }
+    deps.io.out(sid);
+  }
+  if (readinessError) return EXIT.error;
+  if (dead) return EXIT.dead;
+  return ready === false ? EXIT.timeout : EXIT.ok;
+}
+
+export interface SendOptions {
+  id: string;
+  text: string;
+  enter: boolean;
+  /** `undefined` = fire-and-forget; `true` = default signal set; string = comma list. */
+  wait?: string | true;
+  timeoutMs?: number;
+  clientId?: string;
+  seq?: number;
+}
+
+/** `agent send` — printable text plus `\r`, exactly-once, optionally blocking on end of turn. */
+export async function agentSend(deps: AgentDeps, options: SendOptions): Promise<number> {
+  if (isSelfSession(deps.ctx.selfId, options.id)) {
+    return fail(deps, `refusing: ${options.id} is me — typing into my own composer is not a message`, EXIT.refused);
+  }
+  const refusal = inputRefusal(options.text);
+  if (refusal) return fail(deps, refusal, EXIT.refused);
+  const target = await resolveSessionId(deps, options.id);
+  if ('error' in target) return fail(deps, target.error, target.code);
+  const body = buildSendBody(options.text, {
+    enter: options.enter,
+    clientId: options.clientId ?? defaultClientId(deps.ctx.selfId),
+    seq: options.seq ?? (deps.now ?? Date.now)(),
+    wait: options.wait,
+    waitTimeout: options.wait !== undefined ? options.timeoutMs : undefined,
+  });
+  const res = await deps.request(deps.ctx, {
+    method: 'POST',
+    path: `/api/v1/sessions/${encodeURIComponent(target.id)}/input`,
+    body,
+    timeoutMs: (options.timeoutMs ?? 60_000) + 10_000,
+  });
+  if (!res.json?.success) return fail(deps, describeFailure(res));
+  const data = res.json.data as
+    | { delivered?: boolean; duplicate?: boolean; buffered?: boolean; dropped?: boolean; wait?: WaitResult }
+    | undefined;
+  if (deps.json) emitJson(deps, data ?? {});
+  // Fire-and-forget to a remote session whose host is asleep (wake-on-LAN): the server
+  // holds the chunk and types it once the pane is back (`buffered`), or the chunk was
+  // over the wake buffer's cap and is gone (`dropped`). The seq is spent either way,
+  // so a retry needs a new one (the default, the clock, gives it that).
+  if (data?.dropped) {
+    if (!deps.json) {
+      deps.io.err(
+        palette.err(
+          `${GLYPH.fail} dropped: ${target.id}'s host is waking and its input buffer is full — nothing will be typed; send again once it is back`
+        )
+      );
+    }
+    return EXIT.error;
+  }
+  // `delivered:false` without `duplicate` is the route's "the bytes went nowhere":
+  // the PTY exited or send-keys hit a dead pane. The field exists so a client does not
+  // say "wait longer" when the truth is "restart the worker" — so it is a failure here.
+  if (data?.delivered === false && !data.duplicate) {
+    if (!deps.json) {
+      deps.io.err(
+        palette.err(`${GLYPH.fail} not delivered: ${target.id} has no live worker (pane exited) — restart it`)
+      );
+    }
+    return EXIT.dead;
+  }
+  if (!deps.json) {
+    const noEnter = options.enter ? '' : ' (no Enter)';
+    if (data?.duplicate) {
+      deps.io.out(palette.warn(`${GLYPH.warn} duplicate (clientId/seq already applied): nothing typed`));
+    } else if (data?.buffered) {
+      deps.io.out(
+        palette.ok(
+          `${GLYPH.ok} buffered for ${target.id}${noEnter}: its host is asleep; Codeman is waking it and types this once the pane is back`
+        )
+      );
+    } else if (data?.delivered === true) {
+      deps.io.out(palette.ok(`${GLYPH.ok} delivered to ${target.id}${noEnter}`));
+    } else {
+      // Fire-and-forget answers before the write, so there is no delivery report here.
+      deps.io.out(palette.ok(`${GLYPH.ok} accepted for ${target.id}${noEnter} (no delivery report without --wait)`));
+    }
+    if (data?.wait) deps.io.out(describeWait(data.wait));
+  }
+  if (options.wait === undefined) return EXIT.ok;
+  return waitExitCode(data?.wait);
+}
+
+function describeWait(wait: WaitResult): string {
+  if (wait.signal === 'exit') return palette.err(`${GLYPH.fail} the session exited`);
+  if (wait.ended) {
+    return palette.err(
+      `${GLYPH.fail} the wait ended without an answer: the session went away (dead worker, deleted, or nothing was written)`
+    );
+  }
+  // A timeout is a 200 with `timedOut`, an answer rather than a failure: exit 2 says it,
+  // so the line stays neutral instead of looking like an error to whoever reads the log.
+  if (wait.timedOut) return palette.muted(`timed out after ${wait.timeoutMs ?? '?'} ms (exit 2)`);
+  if (wait.matched !== undefined) {
+    return wait.matched
+      ? palette.ok(`${GLYPH.ok} matched "${wait.match}"${wait.snippet ? `: ${wait.snippet}` : ''}`)
+      : palette.muted('not matched (exit 2)');
+  }
+  return palette.ok(
+    `${GLYPH.ok} signal: ${wait.signal}${wait.immediate ? ' (immediate: current state, not a transition)' : ''}`
+  );
+}
+
+export interface WaitOptions {
+  id: string;
+  until?: string;
+  match?: string;
+  from?: 'buffer' | 'now';
+  fresh?: boolean;
+  nocase?: boolean;
+  timeoutMs: number;
+}
+
+/** `agent wait` — a signal (`--until`) or a literal output marker (`--match`). */
+export async function agentWait(deps: AgentDeps, options: WaitOptions): Promise<number> {
+  if (options.until && options.match)
+    return fail(deps, 'use either --until <signals> or --match <marker>, not both', EXIT.refused);
+  const target = await resolveSessionId(deps, options.id);
+  if ('error' in target) return fail(deps, target.error, target.code);
+  const sid = encodeURIComponent(target.id);
+  const res = options.match
+    ? await deps.request(deps.ctx, {
+        method: 'GET',
+        path: `/api/v1/sessions/${sid}/wait-output`,
+        query: {
+          match: options.match,
+          from: options.from ?? 'buffer',
+          nocase: options.nocase ? 1 : undefined,
+          timeout: options.timeoutMs,
+        },
+        timeoutMs: options.timeoutMs + 10_000,
+      })
+    : await deps.request(deps.ctx, {
+        method: 'GET',
+        path: `/api/v1/sessions/${sid}/wait`,
+        query: { until: options.until, fresh: options.fresh ? 1 : undefined, timeout: options.timeoutMs },
+        timeoutMs: options.timeoutMs + 10_000,
+      });
+  // A 400 here is the server saying "this mode has no such signal" (until=stop on an
+  // external CLI). Passed through, never papered over: the marker path is the answer.
+  if (!res.json?.success) return fail(deps, describeFailure(res));
+  const data = res.json.data as { wait?: WaitResult; status?: string; limitPaused?: boolean } | undefined;
+  if (deps.json) {
+    emitJson(deps, data ?? {});
+  } else if (data?.wait) {
+    deps.io.out(describeWait(data.wait));
+    if (data.limitPaused)
+      deps.io.out(palette.warn(`${GLYPH.warn} session is paused on a usage limit; a timeout is expected`));
+  }
+  return waitExitCode(data?.wait);
+}
+
+export interface ReadOptions {
+  id: string;
+  /** Bytes of raw terminal to fetch; ANSI is stripped for humans. */
+  tail?: number;
+  /** Whole conversation (`context=full`) instead of the last assistant message. */
+  full?: boolean;
+}
+
+/** `agent read` — the last answer (the route picks the transcript reader or the pane segmenter) or a terminal tail. */
+export async function agentRead(deps: AgentDeps, options: ReadOptions): Promise<number> {
+  const target = await resolveSessionId(deps, options.id);
+  if ('error' in target) return fail(deps, target.error, target.code);
+  const sid = encodeURIComponent(target.id);
+  if (options.tail !== undefined) {
+    const res = await deps.request(deps.ctx, {
+      method: 'GET',
+      path: `/api/v1/sessions/${sid}/terminal`,
+      query: { tail: options.tail },
+    });
+    if (!res.json?.success) return fail(deps, describeFailure(res));
+    const buffer = (res.json.data as { terminalBuffer?: string } | undefined)?.terminalBuffer ?? '';
+    if (deps.json) emitJson(deps, res.json.data);
+    else deps.io.out(stripAnsi(buffer));
+    return EXIT.ok;
+  }
+  const res = await deps.request(deps.ctx, {
+    method: 'GET',
+    path: `/api/v1/sessions/${sid}/last-response`,
+    query: { context: options.full ? 'full' : undefined },
+  });
+  if (!res.json?.success) return fail(deps, describeFailure(res));
+  const data = res.json.data as
+    | { text?: string; timestamp?: string; messages?: Array<{ role: string; text: string }> }
+    | undefined;
+  if (deps.json) {
+    emitJson(deps, data ?? {});
+    return EXIT.ok;
+  }
+  if (options.full && data?.messages) {
+    for (const m of data.messages) deps.io.out(`${palette.emph(m.role)}: ${m.text}`);
+    return EXIT.ok;
+  }
+  const text = data?.text ?? '';
+  if (!text) {
+    deps.io.err(
+      palette.muted('(empty: nothing answered yet, or nothing the server could segment as an answer; try --tail 3000)')
+    );
+    return EXIT.ok;
+  }
+  deps.io.out(text);
+  return EXIT.ok;
+}
+
+/** `agent interrupt` — a bare ESC keypress, no Enter, conversation intact. */
+export async function agentInterrupt(deps: AgentDeps, options: { id: string }): Promise<number> {
+  if (isSelfSession(deps.ctx.selfId, options.id)) return fail(deps, `refusing: ${options.id} is me`, EXIT.refused);
+  const target = await resolveSessionId(deps, options.id);
+  if ('error' in target) return fail(deps, target.error, target.code);
+  const body = buildInterruptBody(defaultClientId(deps.ctx.selfId, 'interrupt'), (deps.now ?? Date.now)());
+  const res = await deps.request(deps.ctx, {
+    method: 'POST',
+    path: `/api/v1/sessions/${encodeURIComponent(target.id)}/input`,
+    body,
+  });
+  if (!res.json?.success) return fail(deps, describeFailure(res));
+  if (deps.json) emitJson(deps, res.json.data ?? {});
+  else
+    deps.io.out(
+      palette.ok(
+        `${GLYPH.ok} ESC sent to ${target.id} — one Esc does not always land; read the tail before the next prompt`
+      )
+    );
+  return EXIT.ok;
+}
+
+/** `agent rm` — delete a session that is provably not this one. */
+export async function agentRm(deps: AgentDeps, options: { id: string }): Promise<number> {
+  const refusal = deleteRefusal(deps.ctx.selfId, options.id);
+  if (refusal) return fail(deps, refusal, EXIT.refused);
+  const target = await resolveSessionId(deps, options.id);
+  if ('error' in target) return fail(deps, target.error, target.code);
+  // The guard again on the RESOLVED id: a prefix that is not me can still resolve
+  // to me only if the list is lying, but a delete is the one call worth the paranoia.
+  const resolvedRefusal = deleteRefusal(deps.ctx.selfId, target.id);
+  if (resolvedRefusal) return fail(deps, resolvedRefusal, EXIT.refused);
+  const res = await deps.request(deps.ctx, {
+    method: 'DELETE',
+    path: `/api/v1/sessions/${encodeURIComponent(target.id)}`,
+  });
+  if (!res.json?.success) return fail(deps, describeFailure(res));
+  if (deps.json) emitJson(deps, res.json.data ?? {});
+  else deps.io.out(palette.ok(`${GLYPH.ok} deleted ${target.id} (its case directory stays on disk)`));
+  return EXIT.ok;
+}
+
+// ─────────────────────────────────────────────────────────────────────────────
+// Commander wiring
+// ─────────────────────────────────────────────────────────────────────────────
+
+const DEFAULT_WAIT_MS = 60_000;
+
+/** Build deps from the live environment; the guard's message is the only thing a non-session caller sees. */
+function liveDeps(json: boolean): AgentDeps | undefined {
+  try {
+    return {
+      ctx: resolveAgentContext(),
+      request: httpRequest,
+      json,
+      io: { out: (line) => console.log(line), err: (line) => console.error(line) },
+    };
+  } catch (err) {
+    if (err instanceof AgentGuardError) {
+      console.error(palette.err(`${GLYPH.fail} ${err.message}`));
+      return undefined;
+    }
+    throw err;
+  }
+}
+
+/** Run a verb with the live transport and turn its exit code into the process exit. */
+async function run(json: boolean, verb: (deps: AgentDeps) => Promise<number>): Promise<void> {
+  const deps = liveDeps(json);
+  if (!deps) {
+    process.exitCode = EXIT.refused;
+    return;
+  }
+  try {
+    process.exitCode = await verb(deps);
+  } catch (err) {
+    console.error(palette.err(`${GLYPH.fail} ${getErrorMessage(err)}`));
+    process.exitCode = EXIT.error;
+  }
+}
+
+/** Register `codeman agent …` on the program. */
+export function registerAgentCommands(program: Command): Command {
+  const agent = program
+    .command('agent')
+    .description('Talk to other sessions from inside one (any CLI mode): list, spawn, send, wait, read, interrupt, rm');
+
+  agent
+    .command('ls')
+    .alias('list')
+    .description('List sessions; * marks this one')
+    .option('--json', 'Machine-readable output')
+    .action((options: { json?: boolean }) => run(Boolean(options.json), agentLs));
+
+  agent
+    .command('spawn <case>')
+    .description(
+      'Start a worker session in a case (created if missing) and wait for its composer where the mode draws one'
+    )
+    .option('-m, --mode <mode>', 'Run mode id, as the Run menu names it', 'claude')
+    .option('-n, --name <name>', 'Session name shown in the UI')
+    .option('--no-ready', 'Return as soon as the session exists, without the readiness wait')
+    .option('-t, --timeout <ms>', 'Readiness budget in ms', String(DEFAULT_WAIT_MS))
+    .option('--json', 'Machine-readable output')
+    .action(
+      (caseName: string, options: { mode: string; name?: string; ready: boolean; timeout?: string; json?: boolean }) =>
+        run(Boolean(options.json), (deps) =>
+          agentSpawn(deps, {
+            caseName,
+            mode: options.mode,
+            name: options.name,
+            ready: options.ready,
+            timeoutMs: parsePositiveInt(options.timeout, DEFAULT_WAIT_MS),
+          })
+        )
+    );
+
+  agent
+    .command('send <id> <text...>')
+    .description(
+      'Type a prompt into another session and press Enter (ONE quoted argument, printable text only; a prompt that starts with "-" goes after --: send <id> -- "- fix the bug")'
+    )
+    .option('-w, --wait', 'Block until end of turn (the default signal set; see --until)')
+    .option('-u, --until <signals>', 'Signals to wait for, comma list such as stop,exit (implies --wait)')
+    .option('-t, --timeout <ms>', 'Wait budget in ms (with --wait)', String(DEFAULT_WAIT_MS))
+    .option('--no-enter', 'Type the text without submitting it')
+    .option('--client-id <id>', 'Exactly-once tag (default: one per calling session)')
+    .option(
+      '--seq <n>',
+      'Sequence number for the tag (default: the current epoch ms). Must stay monotonic per client id: a reused or lower value is a silent duplicate, nothing is typed'
+    )
+    .option('--json', 'Machine-readable output')
+    .action(
+      (
+        id: string,
+        words: string[],
+        options: {
+          wait?: boolean;
+          until?: string;
+          timeout?: string;
+          enter: boolean;
+          clientId?: string;
+          seq?: string;
+          json?: boolean;
+        }
+      ) =>
+        run(Boolean(options.json), (deps) => {
+          const prompt = sendPromptFromArgs(words);
+          if ('error' in prompt) return Promise.resolve(fail(deps, prompt.error, EXIT.refused));
+          return agentSend(deps, {
+            id,
+            text: prompt.text,
+            enter: options.enter,
+            wait: options.until ?? (options.wait ? true : undefined),
+            timeoutMs: parsePositiveInt(options.timeout, DEFAULT_WAIT_MS),
+            clientId: options.clientId,
+            seq: options.seq === undefined ? undefined : parsePositiveInt(options.seq, 1, '--seq'),
+          });
+        })
+    );
+
+  agent
+    .command('wait <id>')
+    .description(
+      'Block until a signal (--until) or an output marker (--match) — timeout exits 2, a dead worker exits 3'
+    )
+    .option(
+      '-u, --until <signals>',
+      'Comma list: stop,idle,exit,working,blocked (stop needs hook signals for the session, blocked hooks or a dialog Codeman reads off the screen, as for codex; where there are none the server answers 400, passed through)'
+    )
+    .option(
+      '-m, --match <marker>',
+      'Literal substring to wait for in the output (ANSI-stripped, no regex). The echo of your own prompt is output too, so never put the marker verbatim in the prompt: ask for it in halves ("print WORKDONE followed by _4711") and wait on the joined form (WORKDONE_4711)'
+    )
+    .option('--from <where>', 'buffer (scan existing output first, the default) or now', 'buffer')
+    .option('--nocase', 'Case-insensitive --match')
+    .option('--fresh', 'Require an actual transition (--until only)')
+    .option('-t, --timeout <ms>', 'Budget in ms', String(DEFAULT_WAIT_MS))
+    .option('--json', 'Machine-readable output')
+    .action(
+      (
+        id: string,
+        options: {
+          until?: string;
+          match?: string;
+          from: string;
+          nocase?: boolean;
+          fresh?: boolean;
+          timeout?: string;
+          json?: boolean;
+        }
+      ) =>
+        run(Boolean(options.json), (deps) =>
+          agentWait(deps, {
+            id,
+            until: options.until,
+            match: options.match,
+            from: options.from === 'now' ? 'now' : 'buffer',
+            nocase: options.nocase,
+            fresh: options.fresh,
+            timeoutMs: parsePositiveInt(options.timeout, DEFAULT_WAIT_MS),
+          })
+        )
+    );
+
+  agent
+    .command('read <id>')
+    .description("Print a session's last answer (as the server reads it for that mode) or, with --tail, its terminal")
+    .option('--tail <bytes>', 'Raw terminal tail in bytes, ANSI stripped (works in every mode)')
+    .option('--full', 'The whole conversation instead of the last assistant message')
+    .option('--json', 'Machine-readable output')
+    .action((id: string, options: { tail?: string; full?: boolean; json?: boolean }) =>
+      run(Boolean(options.json), (deps) =>
+        agentRead(deps, {
+          id,
+          tail: options.tail === undefined ? undefined : parsePositiveInt(options.tail, 3000, '--tail'),
+          full: options.full,
+        })
+      )
+    );
+
+  agent
+    .command('interrupt <id>')
+    .description('Send a bare ESC to stop the current turn (the conversation survives; deleting would not)')
+    .option('--json', 'Machine-readable output')
+    .action((id: string, options: { json?: boolean }) =>
+      run(Boolean(options.json), (deps) => agentInterrupt(deps, { id }))
+    );
+
+  agent
+    .command('rm <id>')
+    .description('Delete any session except this one (refuses your own id)')
+    .option('--json', 'Machine-readable output')
+    .action((id: string, options: { json?: boolean }) => run(Boolean(options.json), (deps) => agentRm(deps, { id })));
+
+  return agent;
+}

src/cli.ts

@@ -15,9 +15,21 @@ import { existsSync, readFileSync } from 'node:fs';
 import { isAbsolute, join } from 'node:path';
 import { homedir } from 'node:os';
 import { dataPath } from './config/instance.js';
+import { readCodemanCredentials } from './codeman-credentials.js';
 import { casePath } from './config/cases-dir.js';
 import { assertValidBasePath } from './config/base-path.js';
-import { installAgentSkillInto, removeAgentSkillFrom, type AgentSkillApplyResult } from './hooks-config.js';
+import {
+  agentSkillDir,
+  agentSkillDirFor,
+  CLAUDE_AGENT_SKILL_DIR,
+  installAgentSkillInto,
+  removeAgentSkillFrom,
+  type AgentSkillApplyResult,
+} from './hooks-config.js';
+import type { AgentSkillDir } from './config/cli-registry/types.js';
+import { getCli, listClis } from './config/cli-registry/registry.js';
+import { registerAgentCommands } from './cli-agent.js';
+import { resolveMcpConfig, runMcpStdio } from './mcp-server.js';
 import { getSessionManager } from './session-manager.js';
 import { getTaskQueue } from './task-queue.js';
 import { getRalphLoop } from './ralph-loop.js';
@@ -42,32 +54,8 @@ function makeAttachmentMagicLink(filePath: string): string {
   return `codeman://attach?path=${encodeURIComponent(filePath)}`;
 }
 
-function readCodemanEnv(): Record<string, string> {
-  const envPath = dataPath('.env');
-  try {
-    const text = readFileSync(envPath, 'utf-8');
-    const result: Record<string, string> = {};
-    for (const rawLine of text.split(/\r?\n/)) {
-      const line = rawLine.trim();
-      if (!line || line.startsWith('#')) continue;
-      const match = line.match(/^([A-Za-z_][A-Za-z0-9_]*)=(.*)$/);
-      if (!match) continue;
-      let value = match[2].trim();
-      if ((value.startsWith('"') && value.endsWith('"')) || (value.startsWith("'") && value.endsWith("'"))) {
-        value = value.slice(1, -1);
-      }
-      result[match[1]] = value;
-    }
-    return result;
-  } catch {
-    return {};
-  }
-}
-
 async function postAttachment(apiUrl: string, sessionId: string, filePath: string): Promise<boolean> {
-  const envFile = readCodemanEnv();
-  const username = process.env.CODEMAN_USERNAME || envFile.CODEMAN_USERNAME || 'admin';
-  const password = process.env.CODEMAN_PASSWORD || envFile.CODEMAN_PASSWORD;
+  const { username, password } = readCodemanCredentials();
   const url = new URL(`/api/sessions/${encodeURIComponent(sessionId)}/attachments`, apiUrl);
   const body = JSON.stringify({ path: filePath });
   const transport = url.protocol === 'https:' ? https : http;
@@ -162,7 +150,8 @@ export function resolveCliCasePath(name: string): string {
 /**
  * Resolve where `skill install` / `skill uninstall` operate. Global is
  * `~/.claude/skills/codeman` (Claude Code's user-scope skill dir, read by every new
- * session); `--case <name>` targets `<case>/.claude/skills/codeman`, resolved through
+ * session); `--case <name>` targets `<case>/<skillDir>/codeman` for the given skills
+ * directory (default: Claude Code's, `<case>/.claude/skills/codeman`), resolved through
  * `resolveCliCasePath()` above. The web server's automatic per-case injection
  * (`agentSkillEnabled`) covers multi-user spaces; this CLI is a local operator tool
  * and stays single-user.
@@ -173,17 +162,18 @@ export function resolveCliCasePath(name: string): string {
  */
 export function resolveSkillTargetPath(options: {
   case?: string;
+  skillDir?: AgentSkillDir;
 }): { target: string; missingCase?: undefined } | { target?: undefined; missingCase: string } {
   if (options.case) {
     const casePath = resolveCliCasePath(options.case);
     if (!existsSync(casePath)) return { missingCase: casePath };
-    return { target: join(casePath, '.claude', 'skills', 'codeman') };
+    return { target: agentSkillDirFor(casePath, options.skillDir ?? CLAUDE_AGENT_SKILL_DIR) };
   }
   return { target: join(homedir(), '.claude', 'skills', 'codeman') };
 }
 
 /** Exit-owning wrapper around `resolveSkillTargetPath()` for the two commands below. */
-function resolveSkillTarget(options: { case?: string }): string {
+function resolveSkillTarget(options: { case?: string; skillDir?: AgentSkillDir }): string {
   const resolved = resolveSkillTargetPath(options);
   if (resolved.missingCase !== undefined) {
     console.error(palette.err(`✗ Case not found: ${resolved.missingCase}`));
@@ -192,8 +182,44 @@ function resolveSkillTarget(options: { case?: string }): string {
   return resolved.target;
 }
 
-/** Print an AgentSkillApplyResult for humans; exit non-zero when nothing was done. */
-function reportSkillResult(result: AgentSkillApplyResult, target: string): void {
+/**
+ * The skills directory `skill install --mode <mode>` writes into, or why it cannot: a
+ * mode the registry gives no skill, or a user-level install for a CLI other than Claude
+ * Code (each CLI keeps its user-level skills somewhere else, and `npx skills add`
+ * already knows every one of those places).
+ */
+export function resolveSkillInstallDir(options: {
+  mode?: string;
+  case?: string;
+}): { skillDir: AgentSkillDir; error?: undefined } | { skillDir?: undefined; error: string } {
+  const mode = options.mode ?? 'claude';
+  const cli = getCli(mode);
+  if (!cli) return { error: `Unknown mode: ${mode}` };
+  const skillDir = agentSkillDir(cli.capabilities);
+  if (!skillDir) return { error: `Mode ${mode} takes no Codeman agent skill` };
+  if (!options.case && skillDir !== CLAUDE_AGENT_SKILL_DIR) {
+    return {
+      error: `A user-level install for ${mode} goes through the skills installer: npx skills add Ark0N/Codeman --skill codeman -g (or pass --case <name>)`,
+    };
+  }
+  return { skillDir };
+}
+
+/**
+ * Every distinct skills directory the registry can inject into, so
+ * `skill uninstall --case` removes whatever any CLI's session left there.
+ */
+export function registryAgentSkillDirs(): AgentSkillDir[] {
+  const dirs = new Set<AgentSkillDir>();
+  for (const cli of listClis()) {
+    const dir = agentSkillDir(cli.capabilities);
+    if (dir) dirs.add(dir);
+  }
+  return [...dirs];
+}
+
+/** An AgentSkillApplyResult in words, and whether it counts as success. */
+function describeSkillResult(result: AgentSkillApplyResult, target: string): { ok: boolean; text: string } {
   const messages: Record<AgentSkillApplyResult, { ok: boolean; text: string }> = {
     installed: { ok: true, text: `Agent skill installed: ${target}` },
     refreshed: { ok: true, text: `Agent skill refreshed (was stale): ${target}` },
@@ -209,7 +235,12 @@ function reportSkillResult(result: AgentSkillApplyResult, target: string): void
       text: `${target} (or its parent) is a symlink, refusing to write through it.`,
     },
   };
-  const message = messages[result];
+  return messages[result];
+}
+
+/** Print an AgentSkillApplyResult for humans; exit non-zero when nothing was done. */
+function reportSkillResult(result: AgentSkillApplyResult, target: string): void {
+  const message = describeSkillResult(result, target);
   if (message.ok) {
     console.log(palette.ok(`✓ ${message.text}`));
   } else {
@@ -227,9 +258,19 @@ skillCmd
   .description('Install the agent skill globally (~/.claude/skills/codeman) or into one case')
   .option('-g, --global', 'Install into ~/.claude/skills/codeman, picked up by every new session (the default)')
   .option('-c, --case <name>', 'Install into <case>/.claude/skills/codeman instead (linked cases resolve too)')
-  .action(async (options: { global?: boolean; case?: string }) => {
+  .option(
+    '-m, --mode <mode>',
+    'With --case: install where that run mode looks for skills (codex: <case>/.agents/skills/codeman)',
+    'claude'
+  )
+  .action(async (options: { global?: boolean; case?: string; mode?: string }) => {
     try {
-      const target = resolveSkillTarget(options);
+      const resolved = resolveSkillInstallDir(options);
+      if (resolved.error !== undefined) {
+        console.error(palette.err(`✗ ${resolved.error}`));
+        process.exit(1);
+      }
+      const target = resolveSkillTarget({ case: options.case, skillDir: resolved.skillDir });
       reportSkillResult(await installAgentSkillInto(target), target);
     } catch (err) {
       console.error(palette.err(`✗ Failed to install agent skill: ${getErrorMessage(err)}`));
@@ -239,19 +280,70 @@ skillCmd
 
 skillCmd
   .command('uninstall')
-  .description('Remove a Codeman-managed agent skill copy (never touches a user-authored one)')
+  .description('Remove Codeman-managed agent skill copies (never touches a user-authored one)')
   .option('-g, --global', 'Remove from ~/.claude/skills/codeman (the default)')
-  .option('-c, --case <name>', 'Remove from <case>/.claude/skills/codeman instead (linked cases resolve too)')
+  .option(
+    '-c, --case <name>',
+    "Remove every Codeman-managed copy from that case instead, whichever CLI's directory it is in (linked cases resolve too)"
+  )
   .action(async (options: { global?: boolean; case?: string }) => {
     try {
-      const target = resolveSkillTarget(options);
-      reportSkillResult(await removeAgentSkillFrom(target), target);
+      if (!options.case) {
+        const target = resolveSkillTarget(options);
+        reportSkillResult(await removeAgentSkillFrom(target), target);
+        return;
+      }
+      let failed = false;
+      let removedAny = false;
+      for (const skillDir of registryAgentSkillDirs()) {
+        const target = resolveSkillTarget({ case: options.case, skillDir });
+        const result = await removeAgentSkillFrom(target);
+        if (result === 'absent') continue;
+        if (result === 'symlink') {
+          // A shared-skills layout (`.agents/skills -> ../.claude/skills`) is deliberate:
+          // skip it rather than fail the whole sweep.
+          console.log(palette.warn(`- Skipped ${target}: it (or its parent) is a symlink, so it was not touched`));
+          continue;
+        }
+        removedAny = true;
+        const message = describeSkillResult(result, target);
+        if (message.ok) console.log(palette.ok(`✓ ${message.text}`));
+        else {
+          console.error(palette.err(`✗ ${message.text}`));
+          failed = true;
+        }
+      }
+      if (!removedAny) console.log(palette.ok(`✓ No Codeman-managed agent skill in case ${options.case}`));
+      if (failed) process.exit(1);
     } catch (err) {
       console.error(palette.err(`✗ Failed to remove agent skill: ${getErrorMessage(err)}`));
       process.exit(1);
     }
   });
 
+// ============ Agent Commands (session-to-session, any CLI mode) ============
+
+registerAgentCommands(program);
+
+// ============ MCP server (any MCP client drives Codeman) ============
+
+program
+  .command('mcp')
+  .description(
+    'Run a Model Context Protocol server on stdio, so MCP clients (Claude Code, Claude Desktop, Codex, Gemini CLI, Cursor, ...) can start, prompt, wait on, read and delete Codeman sessions'
+  )
+  .option('--url <url>', 'Codeman server URL (default: $CODEMAN_API_URL; never guessed)')
+  .option(
+    '--allow-any-session',
+    'Let the tools prompt, interrupt and delete sessions this server did not start (default: only its own)'
+  )
+  .action(async (options: { url?: string; allowAnySession?: boolean }) => {
+    await runMcpStdio(
+      resolveMcpConfig({ url: options.url, allowAnySession: options.allowAnySession, version: pkg.version })
+    );
+    process.exit(0);
+  });
+
 // ============ Session Commands ============
 
 const sessionCmd = program.command('session').alias('s').description('Manage Claude sessions');
@@ -641,9 +733,7 @@ function probeWebServerAt(base: string): Promise<WebServerProbe | null> {
   } catch {
     return Promise.resolve(null);
   }
-  const envFile = readCodemanEnv();
-  const username = process.env.CODEMAN_USERNAME || envFile.CODEMAN_USERNAME || 'admin';
-  const password = process.env.CODEMAN_PASSWORD || envFile.CODEMAN_PASSWORD;
+  const { username, password } = readCodemanCredentials();
   const transport = url.protocol === 'https:' ? https : http;
   const headers: Record<string, string> = { Accept: 'application/json' };
   if (password) {

src/codeman-credentials.ts

@@ -0,0 +1,75 @@
+/**
+ * @fileoverview Credentials for a client of this Codeman instance's own API.
+ *
+ * Env first, the data dir's `.env` as the fallback — the hand-authored file
+ * `codeman attach`, `codeman tui` and `codeman agent` all read. One reader, so the
+ * three clients cannot drift on quoting, comments or the default username.
+ *
+ * @module codeman-credentials
+ */
+import { readFileSync } from 'node:fs';
+import { dataPath } from './config/instance.js';
+
+export interface CodemanCredentials {
+  username: string;
+  /** Absent when no password is configured (or only the server's environment has it). */
+  password?: string;
+}
+
+/**
+ * Parse a `KEY=value` env file: blank lines and `#` comments skipped, an `export `
+ * prefix tolerated (the file is hand-authored, often sourced by a shell too), one
+ * layer of matching quotes stripped, anything that is not an assignment ignored.
+ */
+export function parseEnvFile(text: string): Record<string, string> {
+  const result: Record<string, string> = {};
+  for (const rawLine of text.split(/\r?\n/)) {
+    const line = rawLine.trim();
+    if (!line || line.startsWith('#')) continue;
+    const match = line.match(/^(?:export\s+)?([A-Za-z_][A-Za-z0-9_]*)=(.*)$/);
+    if (!match) continue;
+    let value = match[2].trim();
+    if ((value.startsWith('"') && value.endsWith('"')) || (value.startsWith("'") && value.endsWith("'"))) {
+      value = value.slice(1, -1);
+    }
+    result[match[1]] = value;
+  }
+  return result;
+}
+
+/** The data dir's `.env`, parsed. Absent or unreadable means `{}`. */
+export function readCodemanEnvFile(envFilePath: string = dataPath('.env')): Record<string, string> {
+  try {
+    return parseEnvFile(readFileSync(envFilePath, 'utf-8'));
+  } catch {
+    return {};
+  }
+}
+
+/**
+ * The lookup order every client uses, per field: the environment, then the `.env`
+ * file, then (username only) `admin`. Pure, so a caller with its own environment
+ * object (`codeman agent`'s guard takes one for testability) gets the same answer.
+ */
+export function credentialsFrom(env: NodeJS.ProcessEnv, fileEnv: Record<string, string>): CodemanCredentials {
+  const username = env.CODEMAN_USERNAME || fileEnv.CODEMAN_USERNAME || 'admin';
+  const password = env.CODEMAN_PASSWORD || fileEnv.CODEMAN_PASSWORD;
+  return password ? { username, password } : { username };
+}
+
+/**
+ * Credentials for the API. No password means no auth is configured, or the user has
+ * it only in the server's environment, in which case the API answers 401.
+ */
+export function readCodemanCredentials(
+  envFilePath: string = dataPath('.env'),
+  env: NodeJS.ProcessEnv = process.env
+): CodemanCredentials {
+  return credentialsFrom(env, readCodemanEnvFile(envFilePath));
+}
+
+/** `Authorization` header value, or undefined when there is no password to send. */
+export function basicAuthHeader(credentials: CodemanCredentials): string | undefined {
+  if (!credentials.password) return undefined;
+  return `Basic ${Buffer.from(`${credentials.username}:${credentials.password}`).toString('base64')}`;
+}

src/config/buffer-limits.ts

@@ -113,6 +113,16 @@ export const FILE_PEEK_BYTES = 8 * 1024 - 1; // 8KB (inclusive end offset)
  */
 export const MAX_PASTE_IMAGE_BYTES = parseInt(process.env.CODEMAN_MAX_PASTE_IMAGE_BYTES || '') || 50 * 1024 * 1024; // 50MB
 
+/**
+ * Maximum size (bytes) of a single video uploaded via POST
+ * /api/sessions/:id/paste-image?kind=video (the phone toolbar's Attach button
+ * offers the camera roll's videos beside its photos). Videos stream to disk
+ * rather than through memory, so this is a disk-fill sanity bound, not a memory
+ * bound: a minute of 4K60 from an iPhone is about 400MB, so the image cap would
+ * refuse most clips. Override: CODEMAN_MAX_PASTE_VIDEO_BYTES (bytes)
+ */
+export const MAX_PASTE_VIDEO_BYTES = parseInt(process.env.CODEMAN_MAX_PASTE_VIDEO_BYTES || '') || 1024 * 1024 * 1024; // 1GB
+
 // ============================================================================
 // File Download Limits
 // ============================================================================

src/config/cli-registry/schema.ts

@@ -15,7 +15,10 @@
 import { z } from 'zod';
 import { compileVersionRegex, countCaptureGroups, TOKEN_PATTERNS } from './patterns.js';
 import { isKnownLauncherProfile, isKnownSetenvProfile } from './profiles.js';
-import type { LaunchDefaultSettingKey, McpConfigFormat, ModelConfigResolverName } from './types.js';
+import type { AgentSkillDir, LaunchDefaultSettingKey, McpConfigFormat, ModelConfigResolverName } from './types.js';
+
+/** `capabilities.agentSkillDir`: every value must stay a fixed, relative, `..`-free path. */
+export const AGENT_SKILL_DIRS = ['.claude/skills', '.agents/skills'] as const satisfies readonly AgentSkillDir[];
 
 /** A bare CLI id: lowercase, starts with a letter, at most 24 chars. Also used as a CSS/URL token. */
 const cliId = z
@@ -294,6 +297,10 @@ const capabilitiesSchema = z
     wheelForward: z
       .object({ mode: z.enum(['never', 'version-gated']), minVersion: z.string().max(20).optional() })
       .strict(),
+    // Whether a hollow pane's wheel/touch scroll may be turned into PageUp/PageDown for
+    // the CLI to page its own transcript. Distinct from wheelForward (SGR reports).
+    // Optional: absent means no paging, the safe default for an unmeasured CLI.
+    transcriptPageKeys: z.boolean().optional(),
     keyboardAccessory: z.enum(['agent', 'shell']),
     privilegedCommandGate: z.boolean(),
     startMode: z.enum(['interactive', 'shell']),
@@ -302,6 +309,9 @@ const capabilitiesSchema = z
     respawn: z.boolean(),
     effort: z.boolean(),
     agentSkillInjection: z.boolean(),
+    // A closed list on purpose: it decides where the server WRITES, and this file is
+    // user-editable, so it can never be a free path (`..`, absolute).
+    agentSkillDir: z.enum(AGENT_SKILL_DIRS).optional(),
     statusLineTelemetry: z.boolean(),
     // How many columns this CLI indents its transcript body by, so a copy can take
     // that much off the clipboard. Bounded, because it is the whole strip: a copy
@@ -320,6 +330,8 @@ const capabilitiesSchema = z
     // A declared width cannot do either. Absent means no strip, so a CLI whose
     // transcript layout nobody has measured is never touched.
     transcriptGutter: z.number().int().min(1).max(8).optional(),
+    // Literal text matched by a `wait-output` long-poll, never compiled as a regex.
+    composerReadyMark: z.string().min(1).max(64).optional(),
     workDetect: z
       .object({
         promptGlyph: z.string().min(1).max(8),
@@ -356,6 +368,22 @@ const capabilitiesSchema = z
             'awaitingLine must be a regex compileVersionRegex() accepts: at most 200 characters, no nested quantifiers'
           )
           .optional(),
+        // Same guard again: run over the foot of a pane capture every time a session
+        // settles and on a timer while a dialog is up, and ~/.codeman/clis.json can set it.
+        blockedLine: z
+          .string()
+          .min(1)
+          .refine(
+            (src) => compileVersionRegex(src) !== null,
+            'blockedLine must be a regex compileVersionRegex() accepts: at most 200 characters, no nested quantifiers'
+          )
+          .optional(),
+        // Wider than `watchingLines` because a dialog's options wrap at narrow widths (codex
+        // needs 20 at 36 columns). That is safe only for a pattern that ends on the dialog's
+        // own footer: the rows above it, which the agent may write, can then supply a
+        // marker but never the footer. Still bounded, since it is also how much text the
+        // pattern runs over.
+        blockedLines: z.number().int().min(1).max(24).optional(),
       })
       .strict()
       // A window with nothing to search is a typo, not a configuration. Refused at LOAD
@@ -366,6 +394,10 @@ const capabilitiesSchema = z
         (v) => v.watchingLines === undefined || v.watchingLine !== undefined,
         'watchingLines has nothing to bound without a watchingLine'
       )
+      .refine(
+        (v) => v.blockedLines === undefined || v.blockedLine !== undefined,
+        'blockedLines has nothing to bound without a blockedLine'
+      )
       .optional(),
     model: z
       .object({ source: z.enum(['flag', 'claude-settings-file', 'none']), param: z.string().optional() })
@@ -441,6 +473,9 @@ const capabilitiesSchema = z
       .object({
         // Home-relative, no traversal: sync writes to this path.
         path: mcpRelativePath,
+        // Home-relative fallback used when `path` is absent. Same no-traversal rule as `path`:
+        // sync reads it and, when it parses as plain JSON, writes to it.
+        altPath: mcpRelativePath.optional(),
         // Every value must be a known McpConfigFormat (types.ts); mcp-sync.ts's dialect table is
         // keyed by the same type, so an adapter-less format fails to compile there.
         format: z.enum([
@@ -449,6 +484,7 @@ const capabilitiesSchema = z
           'codex-toml',
           'opencode-json',
           'antigravity-json',
+          'copilot-json',
         ] as const satisfies readonly McpConfigFormat[]),
         // The env var the CLI reads to move the file, and the path under it (same no-traversal
         // rule: sync writes there too). Resolved from the server env at call time, never here.

src/config/cli-registry/stock.ts

@@ -76,6 +76,21 @@ function agentDefaults(): Pick<
   };
 }
 
+/**
+ * The agent skill in `<workspace>/.agents/skills/codeman`, the one project skills
+ * directory every CLI that spreads this scans. Measured per CLI on
+ * 2026-10-11 against the installed builds or their source: codex 0.162 (every dir from
+ * the git root down to cwd), gemini (cwd only, behind folder trust, which `--skip-trust`
+ * already answers), opencode 1.3 (cwd up to the worktree), agy 1.3.2 (its `.agents/`
+ * customization root), grok 1.0.5 (each tier from cwd up to the repo root), dsh 0.1.1
+ * (the git root, else cwd) and omp (cwd up to the git root). Claude Code does NOT read
+ * `.agents/skills`, so a case shared with a claude session lists each skill once there.
+ */
+const AGENTS_DIR_SKILL: Pick<CliEntry['capabilities'], 'agentSkillInjection' | 'agentSkillDir'> = {
+  agentSkillInjection: true,
+  agentSkillDir: '.agents/skills',
+};
+
 // `accent` on every entry below (except SHELL, which the frontend renders no
 // distinct color for) is measured from the actual `.btn-toolbar.btn-run.mode-<id>`
 // CSS rule's `border-color` on the OG skin (styles.css) — the single cleanest
@@ -218,6 +233,9 @@ const CLAUDE: CliEntry = {
     // in them, so a copy can drop two and paste flush. Claude and codex are the only
     // entries that declare this, because theirs are the only gutters that have been measured.
     transcriptGutter: 2,
+    // The composer's own hint text (`⏵⏵ … (shift+tab to cycle)`), not `❯`, which the
+    // trust dialog's selected row also carries. Measured by the agent skill's spawn_worker.
+    composerReadyMark: 'shift+tab',
     // The historical hard-coded pair, now stated as data. `workingLine` matches both the
     // `✻ Actualizing… (39s · ↓ 2.0k tokens)` status line and the bare `esc to interrupt`
     // footer, because tmux repaints partially and only one of the two may land in a chunk.
@@ -267,6 +285,10 @@ const CLAUDE: CliEntry = {
     // Declared-for-later: the live rule (`_shouldForwardWheelToApp`, terminal-ui.js) is this version
     // AND the server-published `cliMouseTracking` flag (#498), so wiring this field up needs both.
     wheelForward: { mode: 'version-gated', minVersion: '2.1.187' },
+    // A repaint-mode Claude pane keeps no local history, so when the wheel is not
+    // forwarded (older CLI, inline renderer, the local-scrollback opt-out) a scroll is
+    // turned into PageUp/PageDown, which Claude pages its transcript on (#205).
+    transcriptPageKeys: true,
     keyboardAccessory: 'agent',
     privilegedCommandGate: false,
     startMode: 'interactive',
@@ -275,6 +297,7 @@ const CLAUDE: CliEntry = {
     respawn: true,
     effort: true,
     agentSkillInjection: true,
+    agentSkillDir: '.claude/skills',
     statusLineTelemetry: true,
     model: { source: 'claude-settings-file' },
     privilegedParams: [],
@@ -491,8 +514,14 @@ const OPENCODE: CliEntry = {
   },
   capabilities: {
     ...agentDefaults(),
+    ...AGENTS_DIR_SKILL,
     altScreen: 'strip-mux-and-mouse',
     echo: { policy: 'buffer', anchor: { kind: 'cursor' }, predictProfile: undefined },
+    // Pages its own transcript on PageUp/PageDown (`messages_page_up/down`) and ignores
+    // SGR wheel reports, and its TUI runs on the alternate screen (1.18.31: tmux
+    // `alternate_on=1`, `history_size=0`), so a pane keeps no local history and the
+    // wheel is turned into page keys or it does nothing at all (#555).
+    transcriptPageKeys: true,
     // Measured on a live opencode 1.3.0 pane (capture-pane every 250-300 ms through real
     // turns at 40, 60, 120 and 200 columns, plus the raw PTY stream, 2026-10-09). Every
     // composer row starts with a `┃` bar, and the submitted prompt lands in the transcript
@@ -533,6 +562,7 @@ const OPENCODE: CliEntry = {
     // opencode's global config dir is xdg-basedir's `$XDG_CONFIG_HOME/opencode`.
     mcpConfig: {
       path: '.config/opencode/opencode.json',
+      altPath: '.config/opencode/opencode.jsonc',
       format: 'opencode-json',
       relocation: { envVar: 'XDG_CONFIG_HOME', path: 'opencode/opencode.json' },
     },
@@ -624,6 +654,7 @@ const CODEX: CliEntry = {
   },
   capabilities: {
     ...agentDefaults(),
+    ...AGENTS_DIR_SKILL,
     // Codex draws `› Ask Codex to do anything` on its composer row and
     // `Working (2m 49s • esc to interrupt)` above it while a turn runs. It animates no
     // braille spinner, and it never prints `esc to interrupt` at rest, so that phrase
@@ -646,10 +677,11 @@ const CODEX: CliEntry = {
     // which the agent writes (and while a prompt is typed, the last two). Matching the complete row raises the bar — an assistant
     // message has to end with this exact line, to the character — but nothing here makes
     // forging it impossible, so do not read the Claude comment above as applying here.
-    // What contains it is that codex declares `hooks: 'none'`: no hook event from a codex
-    // session ever reaches `notePrompt()`, so there is no idle item to pre-acknowledge
-    // and a forged label costs a wrong badge and nothing else. A CLI that gains hook
-    // signals must not keep a pattern this soft.
+    // What contains it is that codex declares `hooks: 'none'`: no IDLE prompt from a codex
+    // session ever reaches `notePrompt()` (its screen-read dialogs below open PERMISSION
+    // items, which a watching label never acknowledges), so there is no idle item to
+    // pre-acknowledge and a forged label costs a wrong badge and nothing else. A CLI that
+    // gains idle signals must not keep a pattern this soft.
     // ⚠️ Background TERMINALS are the only background work codex advertises on screen.
     // A sub-agent started without waiting outlives the turn just as a terminal does —
     // measured 2026-09-22, the sandboxed process was still running — and the pane shows
@@ -658,11 +690,45 @@ const CODEX: CliEntry = {
     // composer. So a codex session waiting on a sub-agent reads as plainly idle here.
     // Nothing is misfiled by that (codex raises no idle prompts), and there is no row to
     // match until codex pins one.
+    // Codex is the one CLI Codeman launches with its approvals on by default, and it has no
+    // hook to say it stopped for one, so its dialogs are read off the screen. Measured on a
+    // live codex-cli 0.162.0 pane on 2026-10-14 (`-a on-request -s read-only`, widths 200,
+    // 80, 60, 44 and 36): every dialog it blocks on REPLACES the composer and its footer
+    // with a numbered option list, the selected one marked `›`, over a key-hint row:
+    //   command approval (a network         `› 1. Yes, proceed (y)` ...
+    //   escalation too), edit approval
+    //                                       `  Press enter to confirm or esc to cancel`
+    //   question (request_user_input)       `  › 1. Blue  Write the color blue.` ...
+    //                                       `  tab to add notes | enter to submit answer | esc to interrupt`
+    //   (its notes field open)              `  tab or esc to clear notes | enter to submit answer`
+    //   workspace trust, at startup         `› 1. Trust and continue`
+    //                                       `  enter continue · esc quit`
+    //   (the same prompt reached after the  `  2. Back to Agent Command Center`
+    //   update prompt, 0.162.0, 2026-10-16) `  enter continue · esc back`
+    // The pattern needs both: a `› N.` option row, then the key-hint row, which must END
+    // the window. Below about 60 columns that hint wraps onto rows of its own
+    // (`  Press enter to confirm or esc to` / `  cancel`, or `tab to add notes` /
+    // `enter to submit answer` / `esc to interrupt` one per row), so the hint may run on
+    // over rows that start with two spaces and a lowercase word or arrow, and must end on
+    // one of its own closing words. The transcript can quote every one of those strings
+    // (measured: an answer that printed them verbatim) and still not match, because while
+    // no dialog is up codex keeps its composer row (`› ...`, column 0, which ends the
+    // run) and its footer below anything the agent prints. The longest option list
+    // measured, a question at 36 columns, put the `›` row 16 non-blank rows up; 20 leaves
+    // room for longer option text, and a dialog whose marker scrolls out of the window is
+    // simply not seen (the behaviour before this field). Not matched on purpose:
+    // `Implement this plan?` after a plan turn, whose `enter select · esc back` footer is
+    // the same as `/model`, `/permissions` and `/review`, menus the user opened at the
+    // keyboard; the plan turn has ended there, and the session reads idle as before.
+    // ⚠️ The question dialog's hint contains `esc to interrupt`, the working line, so
+    // such a pane read as WORKING before this field: a blocked screen now wins over it.
     workDetect: {
       promptGlyph: '›',
       workingLine: '[Ee]sc to interrupt',
       watchingLine: String.raw`^\s{0,4}(\d+ background terminals?) running · /ps to view · /stop to close$`,
       watchingLines: 4,
+      blockedLine: String.raw`(?:^|\n) {0,4}› \d\. [\s\S]*\n {2}(?:Press enter to confirm|tab (?:to add|or esc to clear) notes|enter continue)(?:[^\n]|\n {2}(?=[a-z←])){0,120}(?:cancel|interrupt|quit|back|answer|all)$`,
+      blockedLines: 20,
     },
     // The footer under the composer, measured on a live 0.147.0 pane:
     // `  gpt-5.6-terra default · ~/codeman-cases/th-scratch` (model, reasoning effort,
@@ -703,6 +769,11 @@ const CODEX: CliEntry = {
     altScreen: 'strip-full',
     echo: { policy: 'predict', anchor: { kind: 'cursor' }, predictProfile: 'codex' },
     wheelForward: { mode: 'never' }, // #227: codex ignores SGR wheel reports, never forward
+    // ...but it DOES page its transcript on PageUp/PageDown. Measured on 0.157.1 (local,
+    // alternate screen inside tmux, so xterm holds no history) and 0.160 (remote): PageUp
+    // moved the visible transcript while SGR wheel reports and local scrolling did nothing.
+    // This is the page-key fallback for a hollow pane, not wheel forwarding.
+    transcriptPageKeys: true,
     maxFrameBytes: 32 * 1024,
     // codex's own bare-spawn default (no config sent) is already safe (no bypass flag), so
     // the multi-user clamp only needs to force an EXPLICITLY-SENT bypass back off.
@@ -813,6 +884,7 @@ const GEMINI: CliEntry = {
   },
   capabilities: {
     ...agentDefaults(),
+    ...AGENTS_DIR_SKILL,
     altScreen: 'strip-full',
     echo: { policy: 'buffer', anchor: { kind: 'cursor' } },
     // Measured on a live Gemini CLI 0.63.0 pane (capture-pane every 300 ms through real
@@ -921,6 +993,7 @@ const ANTIGRAVITY: CliEntry = {
   },
   capabilities: {
     ...agentDefaults(),
+    ...AGENTS_DIR_SKILL,
     altScreen: 'strip-mux-only',
     echo: { policy: 'buffer', anchor: { kind: 'cursor' } },
     // Like codex: an ABSENT config already defaults safe (no bypass flag), so only a
@@ -1016,6 +1089,11 @@ const PI: CliEntry = {
   },
   capabilities: {
     ...agentDefaults(),
+    // No agent skill (agentSkillInjection stays false): any project-level `.agents/skills`
+    // or `.pi/skills` makes pi ask "Trust project folder?" at startup
+    // (hasTrustRequiringProjectResources, pi 1.1.0), which would stall a spawned worker on
+    // a dialog. pi reads the user-level `~/.agents/skills` without that prompt, so a pi
+    // user installs it there (`npx skills add Ark0N/Codeman --skill codeman -g`).
     altScreen: 'preserve', // pi's TUI renders into the main screen with terminal-owned scrollback
     echo: { policy: 'buffer', anchor: { kind: 'cursor' } },
     // Measured on a live pi 1.1.0 pane (capture-pane every 250 ms through a turn,
@@ -1166,6 +1244,7 @@ const GROK: CliEntry = {
   },
   capabilities: {
     ...agentDefaults(),
+    ...AGENTS_DIR_SKILL,
     // Fullscreen alt-screen TUI with mouse support, same strip as antigravity until measured
     // (opencode's mouse strip is #443): only the tmux-attach-time smcup strip, not Ink's full
     // erase-scrollback+DECSET strip.
@@ -1256,8 +1335,14 @@ const DEEPSEEK: CliEntry = {
     launcherTargetParam: 'profile',
     install: {
       command: {
-        linux: 'npm install -g @deepseek-ai/dsh',
-        darwin: 'npm install -g @deepseek-ai/dsh',
+        // pnpm rides along: `dsh plugin` spawns a literal `pnpm` with no npm fallback, so the
+        // terminal profile this CLI needs before it can run cannot be installed without it (#352).
+        // ⚠️ Only when none is on PATH. `corepack enable` puts a `pnpm` shim in node's own bin
+        // directory, which is npm's global bin for nvm and Homebrew, and `npm install -g pnpm`
+        // refuses to overwrite it (EEXIST). Asked for in the same command as dsh, that failure
+        // took dsh down with it, on exactly the machines that already had a working pnpm.
+        linux: 'npm install -g @deepseek-ai/dsh && (command -v pnpm >/dev/null || npm install -g pnpm)',
+        darwin: 'npm install -g @deepseek-ai/dsh && (command -v pnpm >/dev/null || npm install -g pnpm)',
       },
       npmPackage: '@deepseek-ai/dsh',
       docsUrl: 'https://github.com/deepseek-ai/deepseek-harness',
@@ -1338,10 +1423,13 @@ const DEEPSEEK: CliEntry = {
   },
   capabilities: {
     ...agentDefaults(),
+    ...AGENTS_DIR_SKILL,
     // Definitive rather than inferred: the harness TUI reports idle/working/blocked to a
     // supervisor and Codeman is that supervisor. 'supervised' rather than 'always' because
     // the session can disarm the bridge, and docker/remote cannot reach it at all.
     hooks: 'supervised',
+    // dsh's composer glyph, drawn once the harness TUI can take a prompt.
+    composerReadyMark: '❯',
     transcript: 'deepseek-zstd',
     altScreen: 'strip-mux-only',
     echo: { policy: 'buffer', anchor: { kind: 'cursor' } },
@@ -1501,6 +1589,7 @@ const OMP: CliEntry = {
   },
   capabilities: {
     ...agentDefaults(),
+    ...AGENTS_DIR_SKILL,
     // Fullscreen alt-screen TUI, same shape as antigravity/grok: only the
     // tmux-attach-time smcup strip, not Ink's full erase-scrollback+DECSET strip.
     altScreen: 'strip-mux-only',

src/config/cli-registry/types.ts

@@ -69,6 +69,14 @@ export type Cond =
  */
 export type QuoteStyle = 'auto' | 'bare' | 'double' | 'single';
 
+/**
+ * Workspace-relative directories an agent CLI scans for project skills. A closed list
+ * rather than a free path: the value decides where the server WRITES, and
+ * `~/.codeman/clis.json` can set it, so it must never be able to name `..` or an
+ * absolute path.
+ */
+export type AgentSkillDir = '.claude/skills' | '.agents/skills';
+
 /** One argv element. */
 export type ArgSpec =
   /** A bare literal word, e.g. the base binary or codex's `resume` subcommand. */
@@ -105,7 +113,13 @@ export type ModelConfigResolverName = 'deepseek-route';
 export type LaunchDefaultSettingKey = 'codexModel' | 'codexReasoningEffort';
 
 /** The MCP config dialects `src/mcp-sync.ts` has an adapter for. */
-export type McpConfigFormat = 'claude-json' | 'gemini-json' | 'codex-toml' | 'opencode-json' | 'antigravity-json';
+export type McpConfigFormat =
+  | 'claude-json'
+  | 'gemini-json'
+  | 'codex-toml'
+  | 'opencode-json'
+  | 'antigravity-json'
+  | 'copilot-json';
 
 export interface CliLaunch {
   params: Record<string, ParamSpec>;
@@ -392,6 +406,32 @@ export interface CliCapabilities {
      * tested. See `isAwaitingWorkers()` in `session-activity.ts`.
      */
     awaitingLine?: string;
+    /**
+     * Source of a regex matching the dialog this CLI draws when it stops and waits for the
+     * user to answer something (an approval, a question, a trust prompt): Codeman reads
+     * such a pane as BLOCKED, with the same red "needs you" alert, push, webhook and
+     * `blocked` wait signal a Claude `permission_prompt` hook raises. It is the source of
+     * that state for a CLI with no hook of its own (`hooks: 'none'`); a CLI whose hooks
+     * report it never consults the screen for it.
+     *
+     * Tested against the last `blockedLines` non-blank rows of a pane capture, joined
+     * with `\n`, so the pattern can require several markers at once and anchor on the
+     * end of the window (no `m` flag: `$` is the last row). It must anchor on chrome
+     * only that CLI draws, at the FOOT of the screen, where a dialog replaces the
+     * composer: the text above is the agent's own, and it can print anything. A forged
+     * match may cost a false alert and nothing else, because Codeman never answers such a
+     * dialog. See `isBlockedScreen()` in `session-activity.ts`.
+     *
+     * A CLI that omits it is never blocked from the screen, which is what every CLI was
+     * before the field existed.
+     */
+    blockedLine?: string;
+    /**
+     * How many non-blank rows at the foot of the screen `blockedLine` sees. A dialog's
+     * options wrap at narrow widths, which is what pushes its markers up the screen, so
+     * this is sized for the narrowest pane the CLI was measured at. Default 1.
+     */
+    blockedLines?: number;
   };
   /**
    * How many columns this CLI indents its transcript body by, so a copy taken from its
@@ -416,6 +456,17 @@ export interface CliCapabilities {
    * Absent means no strip at all, the same fail-safe direction `workDetect` takes.
    */
   transcriptGutter?: number;
+  /**
+   * Literal text the TUI draws once its composer can take a prompt — what `codeman agent
+   * spawn` waits for (a `wait-output` match) before it calls a worker ready.
+   *
+   * Deliberately NOT `workDetect.promptGlyph`: claude's `❯` also marks the selected row
+   * of its workspace-trust dialog, which is exactly the screen a readiness wait must not
+   * mistake for a composer, so claude declares its composer's own hint text instead.
+   * Absent means no readiness wait: a spawn returns as soon as the session exists, and
+   * the caller synchronizes on `wait-output` markers.
+   */
+  composerReadyMark?: string;
   /** No direct-PTY fallback: the CLI must run inside tmux (secrets ride tmux setenv). */
   requiresMux: boolean;
   /**
@@ -490,6 +541,20 @@ export interface CliCapabilities {
   };
   /** Forwarding the wheel to the CLI's own transcript. 'never' keeps local scrollback. */
   wheelForward: { mode: 'never' | 'version-gated'; minVersion?: string };
+  /**
+   * The CLI pages its OWN transcript on PageUp/PageDown, so a wheel or touch scroll over
+   * a pane with no local history can be turned into those keys instead of doing nothing
+   * (`_maybePageCliTranscript` in terminal-ui.js). Read by the frontend through
+   * `window.__codemanTranscriptPageKeys`, which the server builds from this field.
+   *
+   * ⚠ NOT wheel forwarding. `wheelForward` above sends SGR mouse wheel REPORTS; this sends
+   * plain keys, and only when the local buffer is hollow (normal buffer, `baseY === 0`).
+   * A CLI can ignore one and honour the other: codex ignores SGR wheel reports (so its
+   * `wheelForward` stays 'never') yet pages on PageUp/PageDown.
+   *
+   * Absent means no paging, so a CLI nobody has measured keeps plain local scrolling.
+   */
+  transcriptPageKeys?: boolean;
   keyboardAccessory: 'agent' | 'shell';
   /** Multi-user: this CLI is a raw shell, so its commands need the privileged gate. */
   privilegedCommandGate: boolean;
@@ -498,7 +563,20 @@ export interface CliCapabilities {
   ralph: boolean;
   respawn: boolean;
   effort: boolean;
+  /**
+   * Codeman may write its agent skill (`skills/codeman`) into this CLI's workspace and
+   * seed its per-session preamble (behind the synced `agentSkillEnabled` setting,
+   * add-only, marker-owned). Where it lands comes from `agentSkillDir` below.
+   */
   agentSkillInjection: boolean;
+  /**
+   * The workspace-relative directory this CLI scans for project skills, so the skill
+   * lands at `<workspace>/<agentSkillDir>/codeman/SKILL.md` where the CLI discovers it.
+   * Measured per CLI. Absent with `agentSkillInjection: true` means Claude Code's
+   * `.claude/skills`, which is what the boolean meant before this field existed, so a
+   * `~/.codeman/clis.json` written against the old schema keeps its behaviour.
+   */
+  agentSkillDir?: AgentSkillDir;
   statusLineTelemetry: boolean;
   /** Where a model override is delivered. Claude uniquely writes settings.local.json. */
   model: { source: 'flag' | 'claude-settings-file' | 'none'; param?: string };
@@ -627,7 +705,12 @@ export interface CliCapabilities {
    * reported `skipped` rather than written somewhere the CLI never reads. Absent = the file
    * only follows `$HOME`.
    */
-  mcpConfig?: { path: string; format: McpConfigFormat; relocation?: { envVar: string; path: string } };
+  mcpConfig?: {
+    path: string;
+    altPath?: string;
+    format: McpConfigFormat;
+    relocation?: { envVar: string; path: string };
+  };
   /**
    * How this CLI is pointed at a user-supplied custom OpenAI-compatible
    * endpoint (local, e.g. llama.cpp, or cloud, e.g. Azure AI Foundry) — the

src/folder-listing.ts

@@ -0,0 +1,284 @@
+/**
+ * @fileoverview Folder listings for the file-preview folder grid.
+ *
+ * A folder the user opens (a clicked folder path, or "Folder" in the file
+ * preview) is resolved, checked against the same attachment guard a single file
+ * goes through, and GRANTED an id. The grid then asks for files by
+ * `folderId + file name` only, so browser requests never carry arbitrary paths —
+ * the rule the attachment routes follow for single files.
+ *
+ * Grants live in their own small registry, not in the attachment registry: that one
+ * is capped at 200 records per session, and a folder of 300 renders would evict the
+ * session's real attachments.
+ *
+ * Listings contain non-hidden subfolders and files whose type the preview supports
+ * (`isSupportedAttachmentExtension`), newest first, capped at
+ * {@link MAX_FOLDER_ENTRIES}. Local folders are read with `fs`; a remote (SSH) case's
+ * folder is listed on its host in one ssh round trip (`remoteListFolder`).
+ */
+
+import { randomUUID } from 'node:crypto';
+import fs from 'node:fs/promises';
+import { homedir } from 'node:os';
+import { extname, isAbsolute, join, posix } from 'node:path';
+import {
+  getAttachmentType,
+  isSupportedAttachmentExtension,
+  supportedAttachmentExtensions,
+} from './attachment-registry.js';
+import { isBlockedAttachmentPath, isUnderTree, loadAttachmentGuardConfig } from './config/attachment-guard.js';
+import { isMultiUserMode, userSpacePath } from './config/multiuser.js';
+import { remoteListFolder, RemoteFileAccessError, type RemoteFolderEntry } from './remote-files.js';
+import type { AttachmentDetectedType, AuthUser } from './types.js';
+import type { SessionRemote } from './types/session.js';
+import { describeUnknownPath, probePathKind } from './utils/bounded-path-probe.js';
+import { validateSessionFilePath } from './web/route-helpers.js';
+
+/** Most entries (files + subfolders) one listing returns. */
+export const MAX_FOLDER_ENTRIES = 1000;
+
+/** Folder grants kept per session; the oldest is dropped beyond this. */
+const MAX_FOLDER_GRANTS_PER_SESSION = 32;
+
+export interface FolderFileEntry {
+  name: string;
+  extension: string;
+  attachmentType: AttachmentDetectedType;
+  size: number;
+  mtimeMs: number;
+}
+
+export interface FolderListing {
+  folderId: string;
+  /** The folder with symlinks resolved (on the session's host). */
+  path: string;
+  files: FolderFileEntry[];
+  /** Non-hidden subfolder names, sorted. */
+  folders: string[];
+  truncated: boolean;
+}
+
+export class FolderListingError extends Error {
+  constructor(
+    message: string,
+    readonly statusCode: number = 400
+  ) {
+    super(message);
+  }
+}
+
+interface FolderGrant {
+  folderId: string;
+  dir: string;
+}
+
+class FolderGrants {
+  private bySession = new Map<string, Map<string, FolderGrant>>();
+
+  grant(sessionId: string, dir: string): string {
+    let grants = this.bySession.get(sessionId);
+    if (!grants) {
+      grants = new Map();
+      this.bySession.set(sessionId, grants);
+    }
+    for (const existing of grants.values()) {
+      if (existing.dir === dir) {
+        // Refresh its position so an often-used folder is the last to be dropped.
+        grants.delete(existing.folderId);
+        grants.set(existing.folderId, existing);
+        return existing.folderId;
+      }
+    }
+    const folderId = `fld_${randomUUID()}`;
+    grants.set(folderId, { folderId, dir });
+    while (grants.size > MAX_FOLDER_GRANTS_PER_SESSION) {
+      const oldest = grants.keys().next().value;
+      if (oldest === undefined) break;
+      grants.delete(oldest);
+    }
+    return folderId;
+  }
+
+  get(sessionId: string, folderId: string): FolderGrant | undefined {
+    return this.bySession.get(sessionId)?.get(folderId);
+  }
+
+  clearSession(sessionId: string): void {
+    this.bySession.delete(sessionId);
+  }
+}
+
+export const folderGrants = new FolderGrants();
+
+/**
+ * A file name inside a granted folder, or null when it is not one: no path
+ * separator, not hidden (which also rules out `.`/`..`; the listing leaves hidden
+ * entries out, so the routes do too), no NUL, and a type the preview supports.
+ */
+export function validFolderFileName(name: unknown): string | null {
+  if (typeof name !== 'string' || !name || name.length > 255) return null;
+  if (name.startsWith('.') || name.includes('/') || name.includes('\\') || name.includes('\0')) return null;
+  const extension = extname(name).toLowerCase().replace(/^\./, '');
+  return isSupportedAttachmentExtension(extension) ? name : null;
+}
+
+/** Where a granted folder's file lives (on the session's host). */
+export function folderFilePath(dir: string, name: string): string {
+  return posix.join(dir, name);
+}
+
+interface ListFolderOptions {
+  sessionWorkingDir?: string;
+  remote?: SessionRemote;
+  /** The requesting identity; in multi-user mode a regular user is held to {@link folderAllowedForUser}. */
+  user?: AuthUser;
+}
+
+/**
+ * Multi-user rule for folder grants, the same one the filesystem path picker follows:
+ * a regular user reaches only their own space (`userSpacePath`) or the session's
+ * `workingDir`. User spaces live inside `homedir()`, so without it a user could list
+ * and read another user's space. Always true in single-user mode and for admins; a
+ * remote session's folders are on its host, where only the `workingDir` applies.
+ *
+ * @param path Already resolved (realpath on the session's host).
+ */
+export function folderAllowedForUser(
+  path: string,
+  user: AuthUser | undefined,
+  scope: { workingDir?: string; remote?: SessionRemote }
+): boolean {
+  if (!isMultiUserMode()) return true;
+  if (!user) return false;
+  if (user.role === 'admin') return true;
+  if (scope.remote) return !!scope.workingDir && isUnderTree(path, scope.workingDir);
+  return [userSpacePath(user.username), scope.workingDir].some(
+    (root) => !!root && !!validateSessionFilePath(root, path)
+  );
+}
+
+/**
+ * Resolve, guard, list and grant a folder for `sessionId`.
+ *
+ * A local folder is probed bounded first (`probePathKind`): one on a network mount
+ * that stopped answering must not freeze the server, and "did not answer" is never
+ * reported as "not found". The guard runs on the resolved folder before it is read.
+ *
+ * @param requestedPath Absolute, or `~/…` (the home of the session's host).
+ * @throws FolderListingError 400 not absolute / not a folder, 403 blocked or outside the
+ *   user's space, 404 missing, 502 remote unreachable, 503 local folder not answering.
+ */
+export async function listFolder(
+  sessionId: string,
+  requestedPath: string,
+  options: ListFolderOptions = {}
+): Promise<FolderListing> {
+  const homeRelative = typeof requestedPath === 'string' && (requestedPath === '~' || requestedPath.startsWith('~/'));
+  if (!requestedPath || typeof requestedPath !== 'string' || (!isAbsolute(requestedPath) && !homeRelative)) {
+    throw new FolderListingError('Folder path must be absolute');
+  }
+
+  const guard = await loadAttachmentGuardConfig();
+  const workingDir = options.sessionWorkingDir;
+  const assertAllowed = (dir: string): void => {
+    const confined = !guard.confineToWorkspace
+      ? true
+      : options.remote
+        ? !!workingDir && isUnderTree(dir, workingDir)
+        : !!workingDir && !!validateSessionFilePath(workingDir, dir);
+    if (!confined || isBlockedAttachmentPath(dir, guard.blockedTrees)) {
+      throw new FolderListingError('Access to this folder is blocked', 403);
+    }
+    if (!folderAllowedForUser(dir, options.user, { workingDir, remote: options.remote })) {
+      throw new FolderListingError('Folder is outside your space', 403);
+    }
+  };
+
+  let dir: string;
+  let entries: RemoteFolderEntry[];
+  let truncated = false;
+  if (options.remote) {
+    const remotePath = requestedPath === '~' ? '~/' : requestedPath;
+    let listing;
+    try {
+      listing = await remoteListFolder(options.remote, remotePath, supportedAttachmentExtensions(), MAX_FOLDER_ENTRIES);
+    } catch (err) {
+      throw new FolderListingError(err instanceof RemoteFileAccessError ? err.message : 'remote host unreachable', 502);
+    }
+    if (!listing) throw new FolderListingError('Folder not found', 404);
+    if (listing === 'not-a-folder') throw new FolderListingError('Not a folder');
+    ({ dir, entries, truncated } = listing);
+    // One ssh call lists and resolves together, so the guard can only run after it;
+    // nothing of a refused listing leaves this function.
+    assertAllowed(dir);
+  } else {
+    const expanded = homeRelative ? join(homedir(), requestedPath.slice(2)) : requestedPath;
+    const kind = await probePathKind(expanded, { pastCap: true });
+    if (kind === 'unknown') {
+      throw new FolderListingError(describeUnknownPath('Folder', expanded, { pastCap: true }), 503);
+    }
+    if (kind === 'absent') throw new FolderListingError('Folder not found', 404);
+    if (kind !== 'directory') throw new FolderListingError('Not a folder');
+    try {
+      dir = await fs.realpath(expanded);
+    } catch {
+      throw new FolderListingError('Folder not found', 404);
+    }
+    assertAllowed(dir);
+    ({ entries, truncated } = await listLocalFolder(dir));
+  }
+
+  const files: FolderFileEntry[] = [];
+  const folders: string[] = [];
+  for (const entry of entries) {
+    if (entry.name.startsWith('.')) continue;
+    // A secret that happens to sit in an open folder stays out of the grid too.
+    if (isBlockedAttachmentPath(folderFilePath(dir, entry.name), guard.blockedTrees)) continue;
+    if (entry.kind === 'directory') {
+      folders.push(entry.name);
+      continue;
+    }
+    const extension = extname(entry.name).toLowerCase().replace(/^\./, '');
+    if (!isSupportedAttachmentExtension(extension)) continue;
+    files.push({
+      name: entry.name,
+      extension,
+      attachmentType: getAttachmentType(extension),
+      size: entry.size,
+      mtimeMs: entry.mtimeMs,
+    });
+  }
+  files.sort((a, b) => b.mtimeMs - a.mtimeMs || a.name.localeCompare(b.name));
+  folders.sort((a, b) => a.localeCompare(b));
+
+  return { folderId: folderGrants.grant(sessionId, dir), path: dir, files, folders, truncated };
+}
+
+/** Local counterpart of the remote listing script: subfolders + supported files, stat'ed through symlinks. */
+async function listLocalFolder(dir: string): Promise<{ entries: RemoteFolderEntry[]; truncated: boolean }> {
+  const dirents = await fs.readdir(dir, { withFileTypes: true });
+  const entries: RemoteFolderEntry[] = [];
+  let truncated = false;
+  for (const dirent of dirents) {
+    if (dirent.name.startsWith('.')) continue;
+    if (entries.length >= MAX_FOLDER_ENTRIES) {
+      truncated = true;
+      break;
+    }
+    const extension = extname(dirent.name).toLowerCase().replace(/^\./, '');
+    const maybeFile = isSupportedAttachmentExtension(extension);
+    if (!dirent.isDirectory() && !maybeFile && !dirent.isSymbolicLink()) continue;
+    let stat;
+    try {
+      stat = await fs.stat(join(dir, dirent.name));
+    } catch {
+      continue; // dangling symlink, or gone since readdir
+    }
+    if (stat.isDirectory()) {
+      entries.push({ kind: 'directory', name: dirent.name, size: 0, mtimeMs: 0 });
+    } else if (stat.isFile() && maybeFile) {
+      entries.push({ kind: 'file', name: dirent.name, size: stat.size, mtimeMs: stat.mtimeMs });
+    }
+  }
+  return { entries, truncated };
+}

src/git-clone.ts

@@ -495,6 +495,10 @@ export function gitNonInteractiveEnv(base: NodeJS.ProcessEnv = process.env): Nod
     SSH_ASKPASS_REQUIRE: 'never',
     DISPLAY: '',
     GCM_INTERACTIVE: 'never',
+    // classifyGitFailure() matches git's ENGLISH stderr; a German or French locale would
+    // turn a missing ref into a generic FAILED (422 instead of 400).
+    LC_ALL: 'C',
+    LANG: 'C',
     GIT_SSH_COMMAND:
       base.GIT_SSH_COMMAND || 'ssh -oBatchMode=yes -oStrictHostKeyChecking=accept-new -oConnectTimeout=10',
   };

src/git-worktrees.ts

@@ -0,0 +1,953 @@
+/**
+ * @fileoverview Git worktrees for a LOCAL case: list, create and remove the worktrees a
+ * repository-backed case keeps under `<repo>/.claude/worktrees/<name>`, so several agents working
+ * on one repository each get their own working tree instead of sharing the case folder (sessions
+ * in one case share its folder, and two agents editing, building and committing in one checkout
+ * step on each other). Routes: `src/web/routes/worktree-routes.ts`; a session starts in one
+ * through the existing create paths (`POST /api/quick-start` with `worktree`, or
+ * `POST /api/sessions` with the worktree as `workingDir`).
+ *
+ * WHERE. `<repo>/.claude/worktrees/<name>` on a new branch `worktree-<name>` is Claude Code's own
+ * `--worktree` layout (measured in Claude Code 2.1.296: path `<repo>/.claude/worktrees/<name>`,
+ * branch `worktree-<name>`, names of letters, digits, `.`, `_`, `-`, at most 64 characters), and
+ * Codeman already recognizes it (the worktree pill, `CLAUDE_WORKTREE_PATH` in session-routes.ts).
+ * Names here are a strict subset of Claude's (one segment, starting with a letter or digit), so a
+ * worktree made here is one `claude --worktree <name>` would also find. The main checkout must not
+ * show the worktrees as untracked, so `/.claude/worktrees/` is appended to the repository's LOCAL
+ * `info/exclude` (in the real common git dir, never a tracked `.gitignore`) when git does not
+ * already ignore it. `/.claude/settings.local.json` gets the same treatment: Codeman installs its
+ * hooks there for every claude session, and an untracked copy in a worktree makes
+ * `git worktree remove` refuse forever (Claude Code itself ignores that file through the global
+ * excludes file, which is not guaranteed to be set up on a given machine). The copy of Codeman's
+ * agent skill (`<skills dir>/codeman/`, injected while `agentSkillEnabled` is on) is excluded for
+ * the same reason, for every skills folder the registry names (`WorktreeOptions.agentSkillDirs`).
+ *
+ * Split like the other git modules: a PURE half (name and branch validation, `worktree list`
+ * parsing, exclude text, the managed-path rule) unit-tested without git, and an IO half that
+ * spawns git through an injectable runner.
+ *
+ * Rules the code keeps and the tests pin:
+ *   - git is spawned with an argv array, never a shell, with `--` before every path operand, the
+ *     non-interactive environment of `gitNonInteractiveEnv()`, a timeout that kills the whole
+ *     process group, and capped output. Read-only calls add `--no-optional-locks`.
+ *   - Names and branches are validated BEFORE they reach git (charset, length, then
+ *     `git check-ref-format --branch`, whose answer must equal the input: `@{-1}` style shorthands
+ *     are refused up front, since `--branch` would expand them).
+ *   - Only `<repo>/.claude/worktrees/<valid-name>` is ever created or removed. Every component of
+ *     that path is `lstat`ed (a symlinked `.claude` or `worktrees` directory, or a symlink at the
+ *     worktree itself, is refused) and its real path must be exactly the expected one, so a
+ *     repository cannot redirect a create or a remove outside itself. What is left is a same-user
+ *     race (an agent swapping a directory between the check and git), which is accepted: that
+ *     agent already runs code as this user.
+ *   - Removal is `git worktree remove` WITHOUT `--force` (git refuses a tree with modified or
+ *     untracked files, and that refusal is the answer), refused while a live session runs inside
+ *     the worktree, and the optional branch delete is `git branch -d` (never `-D`, so an unmerged
+ *     branch survives).
+ *   - git never runs for a case at or inside a Docker case workspace (or whose common git dir is
+ *     there): a container could plant a checkout filter or a hook, and `git worktree add` runs the
+ *     repository's post-checkout hook and filters on the HOST. On a local case that is acceptable,
+ *     since the user's agent runs code there anyway. A managed worktree that is, sits inside, or
+ *     holds a Docker case workspace is never removed either: `git worktree remove` runs a status
+ *     inside the tree, then deletes it under the container.
+ *   - A repository that merely encloses the case and is the home folder or above (a dotfiles repo
+ *     in `$HOME`) is not the case's repository, and a case folder below a repository's top level
+ *     is refused: the worktrees would not be where the case is.
+ *
+ * @module git-worktrees
+ */
+
+import { spawn } from 'node:child_process';
+import { promises as fs } from 'node:fs';
+import { homedir } from 'node:os';
+import { join, resolve, sep } from 'node:path';
+import { gitNonInteractiveEnv, isSafeGitRef, sanitizeGitOutput } from './git-clone.js';
+import { isInsideAny, isUnrelatedAncestor } from './git-workspace-status.js';
+
+// ─── Limits ──────────────────────────────────────────────────────────────────
+
+/** Claude Code's own limit for a `--worktree` name (2.1.296). */
+export const MAX_WORKTREE_NAME_LENGTH = 64;
+/** Longest accepted branch name. git allows more; 200 covers every real one. */
+export const MAX_WORKTREE_BRANCH_LENGTH = 200;
+/** A read-only git call (rev-parse, worktree list, check-ignore, check-ref-format). */
+export const WORKTREE_READ_TIMEOUT_MS = 15_000;
+/**
+ * `git worktree add` checks out a whole tree and runs the repository's post-checkout hook and
+ * smudge filters (Git LFS downloads objects there), so it gets the longest budget.
+ */
+export const WORKTREE_ADD_TIMEOUT_MS = 180_000;
+/** `git worktree remove` runs a status of the tree first, then deletes it. */
+export const WORKTREE_REMOVE_TIMEOUT_MS = 60_000;
+/** Captured stdout. `worktree list` is one short block per worktree. */
+const MAX_STDOUT_BYTES = 1_000_000;
+/** Captured stderr tail. */
+const MAX_STDERR_BYTES = 8_192;
+/** Worktrees listed. A repository with more is unusual; the count says when it was cut. */
+export const MAX_WORKTREES_LISTED = 200;
+
+/** Path segments of the managed worktrees folder, relative to the repository root. */
+export const WORKTREES_DIR_SEGMENTS = ['.claude', 'worktrees'] as const;
+/** The `info/exclude` line that hides the managed worktrees from the main checkout. */
+export const WORKTREES_EXCLUDE_LINE = '/.claude/worktrees/';
+/** The `info/exclude` line for the hooks file Codeman writes into every claude workspace. */
+export const SETTINGS_LOCAL_EXCLUDE_LINE = '/.claude/settings.local.json';
+
+/** A workspace-relative skills folder (`.claude/skills`, `.agents/skills`): plain segments only. */
+const SKILLS_DIR_RE = /^[A-Za-z0-9._-]+(?:\/[A-Za-z0-9._-]+)*$/;
+
+/**
+ * The `info/exclude` line for the copy of Codeman's agent skill in `skillsDir` of a workspace
+ * (`<skillsDir>/codeman/`, written into the folder a session runs in while `agentSkillEnabled`
+ * is on), or null for a folder that is not a plain relative path.
+ */
+export function agentSkillExcludeLine(skillsDir: string): string | null {
+  if (!SKILLS_DIR_RE.test(skillsDir) || skillsDir.split('/').some((s) => s === '.' || s === '..')) return null;
+  return `/${skillsDir}/codeman/`;
+}
+
+// ─── Pure: names ─────────────────────────────────────────────────────────────
+
+const WORKTREE_NAME_RE = /^[A-Za-z0-9][A-Za-z0-9._-]*$/;
+const BRANCH_NAME_RE = /^[A-Za-z0-9][A-Za-z0-9._/-]*$/;
+
+/**
+ * Why `name` cannot be a worktree name, or null when it can. One path segment of letters, digits,
+ * `.`, `_` and `-`, starting with a letter or digit, at most 64 characters, never `..` (the
+ * session path schema refuses any path containing it), and not ending in `.` or `.lock` (git's
+ * own ref rules, since the default branch is derived from the name).
+ */
+export function worktreeNameError(name: unknown): string | null {
+  if (typeof name !== 'string' || name.length === 0) return 'A worktree name is required.';
+  if (name.length > MAX_WORKTREE_NAME_LENGTH)
+    return `A worktree name is at most ${MAX_WORKTREE_NAME_LENGTH} characters.`;
+  if (!WORKTREE_NAME_RE.test(name))
+    return 'A worktree name may contain only letters, digits, ".", "_" and "-", and must start with a letter or digit.';
+  if (name.includes('..')) return 'A worktree name may not contain "..".';
+  if (name.endsWith('.') || name.toLowerCase().endsWith('.lock'))
+    return 'A worktree name may not end with "." or ".lock".';
+  return null;
+}
+
+/** The branch a new worktree gets when the caller names none: Claude Code's `worktree-<name>`. */
+export function defaultWorktreeBranch(name: string): string {
+  return `worktree-${name}`;
+}
+
+/**
+ * Why `branch` cannot be a new branch name, or null when it passes the charset gate. The final
+ * word is `git check-ref-format --branch` (see `createCaseWorktree`); this gate exists so nothing
+ * option-shaped, `@{...}`-shaped or otherwise surprising ever reaches git at all.
+ */
+export function worktreeBranchError(branch: unknown): string | null {
+  if (typeof branch !== 'string' || branch.length === 0) return 'A branch name is required.';
+  if (branch.length > MAX_WORKTREE_BRANCH_LENGTH)
+    return `A branch name is at most ${MAX_WORKTREE_BRANCH_LENGTH} characters.`;
+  if (!BRANCH_NAME_RE.test(branch))
+    return 'A branch name may contain only letters, digits, ".", "_", "-" and "/", and must start with a letter or digit.';
+  if (branch.includes('..') || branch.includes('//')) return 'A branch name may not contain ".." or "//".';
+  if (branch.endsWith('/') || branch.endsWith('.') || branch.toLowerCase().endsWith('.lock'))
+    return 'A branch name may not end with "/", "." or ".lock".';
+  if (branch.split('/').some((part) => part.startsWith('.'))) return 'No part of a branch name may start with ".".';
+  return null;
+}
+
+/** Why `base` cannot be a start point, or null. The same gate Clone Repo uses for a ref. */
+export function worktreeBaseError(base: unknown): string | null {
+  if (typeof base !== 'string' || !isSafeGitRef(base))
+    return 'The base must be a branch, tag or commit (letters, digits, ".", "_", "-", "/" and "+").';
+  return null;
+}
+
+// ─── Pure: paths ─────────────────────────────────────────────────────────────
+
+/** `<repo>/.claude/worktrees`. */
+export function managedWorktreesDir(repoRoot: string): string {
+  return join(repoRoot, ...WORKTREES_DIR_SEGMENTS);
+}
+
+/** `<repo>/.claude/worktrees/<name>`. The caller validates `name` first. */
+export function managedWorktreePath(repoRoot: string, name: string): string {
+  return join(managedWorktreesDir(repoRoot), name);
+}
+
+/**
+ * The managed worktree name of `path`, or null when it is not exactly
+ * `<repoRoot>/.claude/worktrees/<valid-name>`. Lexical: both sides must already be real paths.
+ */
+export function managedWorktreeName(repoRoot: string, path: string): string | null {
+  const dir = managedWorktreesDir(repoRoot) + sep;
+  const clean = path.endsWith(sep) && path.length > 1 ? path.slice(0, -1) : path;
+  if (!clean.startsWith(dir)) return null;
+  const name = clean.slice(dir.length);
+  if (!name || name.includes(sep)) return null;
+  return worktreeNameError(name) === null ? name : null;
+}
+
+/** True when `child` is `root` or inside it (lexical, on resolved paths). */
+export function isPathWithin(child: string, root: string): boolean {
+  const c = resolve(child);
+  const r = resolve(root);
+  return c === r || c.startsWith(r === sep ? r : r + sep);
+}
+
+/**
+ * `path` as the repository spells it: a path under `alias` (the case folder as its case resolves
+ * it, which can run through a symlink) moved under `root` (the repository's real top level), else
+ * `path` unchanged. Lexical, so it costs no filesystem call per session. Sessions keep the
+ * `workingDir` they were created with, while the worktree paths git lists are real, so without
+ * this a session spelled through the case's link counted for no worktree and did not block its
+ * removal.
+ */
+export function respellUnder(path: string, alias: string, root: string): string {
+  const p = resolve(path);
+  const a = resolve(alias);
+  if (a === resolve(root) || !isPathWithin(p, a)) return p;
+  return join(root, p.slice(a.length));
+}
+
+// ─── Pure: `git worktree list --porcelain` ───────────────────────────────────
+
+export interface ParsedWorktree {
+  path: string;
+  /** Commit checked out (absent for a bare entry). */
+  head: string | null;
+  /** Short branch name (`refs/heads/` stripped), null when detached or bare. */
+  branch: string | null;
+  detached: boolean;
+  bare: boolean;
+  locked: boolean;
+  lockReason?: string;
+  prunable: boolean;
+  prunableReason?: string;
+}
+
+/**
+ * Undo git's C-style path quoting (`"a\\"b\\303\\251"` → `a"bé`). Paths are printed quoted when
+ * they hold a `"`, a `\` or a control character even with `core.quotePath=false`; octal escapes are
+ * raw bytes, decoded as UTF-8. Text that is not quoted comes back unchanged.
+ */
+export function unquoteGitPath(text: string): string {
+  if (text.length < 2 || !text.startsWith('"') || !text.endsWith('"')) return text;
+  const body = text.slice(1, -1);
+  const bytes: number[] = [];
+  const simple: Record<string, number> = { a: 7, b: 8, f: 12, n: 10, r: 13, t: 9, v: 11, '"': 34, '\\': 92 };
+  for (let i = 0; i < body.length; i++) {
+    const ch = body[i];
+    if (ch !== '\\') {
+      for (const b of Buffer.from(ch, 'utf-8')) bytes.push(b);
+      continue;
+    }
+    const next = body[i + 1];
+    if (next !== undefined && /[0-7]/.test(next)) {
+      const oct = /^[0-7]{1,3}/.exec(body.slice(i + 1))![0];
+      bytes.push(parseInt(oct, 8) & 0xff);
+      i += oct.length;
+    } else if (next !== undefined && next in simple) {
+      bytes.push(simple[next]);
+      i += 1;
+    } else {
+      bytes.push(92);
+    }
+  }
+  return Buffer.from(bytes).toString('utf-8');
+}
+
+/**
+ * Parse `git worktree list --porcelain` (newline-terminated form: `-z` needs git 2.36, and Ubuntu
+ * 22.04 ships 2.34). Records are separated by blank lines; the first is the main worktree. Paths
+ * are read through `unquoteGitPath`, which together with `-c core.quotePath=false` keeps non-ASCII
+ * repository paths readable.
+ */
+export function parseWorktreeList(text: string): ParsedWorktree[] {
+  const out: ParsedWorktree[] = [];
+  let cur: ParsedWorktree | null = null;
+  const flush = () => {
+    if (cur) out.push(cur);
+    cur = null;
+  };
+  for (const raw of text.split('\n')) {
+    const line = raw.replace(/\r$/, '');
+    if (!line) {
+      flush();
+      continue;
+    }
+    const space = line.indexOf(' ');
+    const key = space === -1 ? line : line.slice(0, space);
+    const value = space === -1 ? '' : line.slice(space + 1);
+    if (key === 'worktree') {
+      flush();
+      cur = {
+        path: unquoteGitPath(value),
+        head: null,
+        branch: null,
+        detached: false,
+        bare: false,
+        locked: false,
+        prunable: false,
+      };
+      continue;
+    }
+    if (!cur) continue;
+    const entry: ParsedWorktree = cur;
+    if (key === 'HEAD') entry.head = value || null;
+    else if (key === 'branch') entry.branch = value.startsWith('refs/heads/') ? value.slice(11) : value || null;
+    else if (key === 'detached') entry.detached = true;
+    else if (key === 'bare') entry.bare = true;
+    else if (key === 'locked') {
+      entry.locked = true;
+      if (value) entry.lockReason = value;
+    } else if (key === 'prunable') {
+      entry.prunable = true;
+      if (value) entry.prunableReason = value;
+    }
+  }
+  flush();
+  return out;
+}
+
+// ─── Pure: exclude file ──────────────────────────────────────────────────────
+
+/**
+ * What to append to an `info/exclude` holding `existing` so that every line of `lines` is in it:
+ * the text ('' when nothing is missing) and the lines it adds. A file that does not end in a
+ * newline gets one first, so an appended pattern never fuses with the user's last line.
+ */
+export function excludeAppendText(existing: string, lines: readonly string[]): { text: string; added: string[] } {
+  const present = new Set(existing.split(/\r?\n/).map((l) => l.trim()));
+  const added = lines.filter((l) => !present.has(l));
+  if (added.length === 0) return { text: '', added };
+  const lead = existing.length > 0 && !existing.endsWith('\n') ? '\n' : '';
+  return {
+    text: `${lead}# Added by Codeman: git worktrees and the files Codeman writes into each workspace stay local\n${added.join('\n')}\n`,
+    added,
+  };
+}
+
+// ─── Pure: git's answers ─────────────────────────────────────────────────────
+
+/**
+ * The part of git's stderr a user should see: its `fatal:`/`error:` lines (with any `hint:` lines
+ * dropped, they repeat what the UI says), sanitized and capped. Shown through `textContent` only.
+ */
+export function gitRefusalText(stderr: string): string {
+  const clean = sanitizeGitOutput(stderr, 2000);
+  const lines = clean
+    .split('\n')
+    .map((l) => l.trim())
+    .filter((l) => l && !/^hint:/i.test(l));
+  const diagnostic = lines.filter((l) => /^(fatal|error|warning):/i.test(l));
+  const text = (diagnostic.length ? diagnostic : lines).join('\n');
+  return text.length > 1000 ? `${text.slice(0, 1000)}…` : text;
+}
+
+// ─── IO: git runner ──────────────────────────────────────────────────────────
+
+export interface WorktreeGitResult {
+  code: number | null;
+  stdout: string;
+  stderr: string;
+  timedOut: boolean;
+  spawnError?: string;
+}
+
+/** Runs `git <args>` in `cwd`; never throws. Injected so tests can refuse to run git at all. */
+export type WorktreeGitRunner = (
+  cwd: string,
+  args: string[],
+  opts?: { timeoutMs?: number }
+) => Promise<WorktreeGitResult>;
+
+/**
+ * Global options for a call that only reads: no optional index refresh, no filesystem monitor,
+ * no path quoting of non-ASCII names. Must precede the subcommand.
+ */
+const READ_ONLY_GIT_OPTIONS = ['--no-optional-locks', '-c', 'core.fsmonitor=false', '-c', 'core.quotePath=false'];
+
+/**
+ * The production runner: argv only (no shell), the non-interactive environment, captured output
+ * caps, and a timeout that signals the whole process group (`detached: true`), because
+ * `git worktree add` fans out into checkout children, hooks and filters that a signal to the
+ * parent alone would leave running.
+ */
+export const runWorktreeGit: WorktreeGitRunner = (cwd, args, opts) =>
+  new Promise<WorktreeGitResult>((resolveRun) => {
+    let child: ReturnType<typeof spawn>;
+    try {
+      child = spawn('git', args, {
+        cwd,
+        env: { ...gitNonInteractiveEnv(), GIT_OPTIONAL_LOCKS: '0' },
+        stdio: ['ignore', 'pipe', 'pipe'],
+        detached: true,
+      });
+    } catch (err) {
+      resolveRun({ code: null, stdout: '', stderr: '', timedOut: false, spawnError: String(err) });
+      return;
+    }
+    let stdout = '';
+    let stdoutBytes = 0;
+    let stderr = '';
+    let timedOut = false;
+    let settled = false;
+    let killTimer: NodeJS.Timeout | undefined;
+    const killTree = (signal: NodeJS.Signals) => {
+      try {
+        if (child.pid) process.kill(-child.pid, signal);
+      } catch {
+        try {
+          child.kill(signal);
+        } catch {
+          /* already gone */
+        }
+      }
+    };
+    const timer = setTimeout(() => {
+      timedOut = true;
+      killTree('SIGTERM');
+      killTimer = setTimeout(() => killTree('SIGKILL'), 3_000);
+    }, opts?.timeoutMs ?? WORKTREE_READ_TIMEOUT_MS);
+    child.stdout?.on('data', (chunk: Buffer) => {
+      stdoutBytes += chunk.length;
+      if (stdoutBytes <= MAX_STDOUT_BYTES) stdout += chunk.toString('utf-8');
+    });
+    child.stderr?.on('data', (chunk: Buffer) => {
+      stderr += chunk.toString('utf-8');
+      if (stderr.length > MAX_STDERR_BYTES * 2) stderr = stderr.slice(-MAX_STDERR_BYTES);
+    });
+    const finish = (result: WorktreeGitResult) => {
+      if (settled) return;
+      settled = true;
+      clearTimeout(timer);
+      if (killTimer) clearTimeout(killTimer);
+      resolveRun(result);
+    };
+    child.on('error', (err) => finish({ code: null, stdout, stderr, timedOut, spawnError: String(err) }));
+    child.on('close', (code) => finish({ code, stdout, stderr, timedOut }));
+  });
+
+// ─── IO: results ─────────────────────────────────────────────────────────────
+
+/** Failure codes, named after the API error codes the routes map them to. */
+export type WorktreeFailureCode = 'INVALID_INPUT' | 'NOT_FOUND' | 'ALREADY_EXISTS' | 'CONFLICT' | 'OPERATION_FAILED';
+
+export interface WorktreeFailure {
+  ok: false;
+  code: WorktreeFailureCode;
+  message: string;
+  /** git's own words, when git refused (sanitized; render as text). */
+  gitMessage?: string;
+}
+
+const fail = (code: WorktreeFailureCode, message: string, gitMessage?: string): WorktreeFailure => ({
+  ok: false,
+  code,
+  message,
+  ...(gitMessage ? { gitMessage } : {}),
+});
+
+function gitFailure(run: WorktreeGitResult, what: string): WorktreeFailure {
+  if (run.spawnError && /enoent/i.test(run.spawnError))
+    return fail('OPERATION_FAILED', 'git is not installed on this machine (or not on the server’s PATH).');
+  if (run.timedOut) return fail('OPERATION_FAILED', `git timed out while ${what}.`);
+  const text = gitRefusalText(run.stderr);
+  return fail('OPERATION_FAILED', text ? `git failed while ${what}: ${text}` : `git failed while ${what}.`, text);
+}
+
+// ─── IO: the case's repository ───────────────────────────────────────────────
+
+export interface CaseRepository {
+  /** The repository's top level, which is the case folder (real path). */
+  repoRoot: string;
+  /** The real common git dir (`.git` of the main checkout), where `info/exclude` lives. */
+  commonDir: string;
+}
+
+export interface WorktreeOptions {
+  git?: WorktreeGitRunner;
+  /** Host paths of every Docker case workspace: git never runs at or inside one. */
+  dockerWorkspaces?: string[];
+  /** Overrides `os.homedir()` for the unrelated-ancestor rule (tests). */
+  home?: string;
+  /**
+   * Every workspace skills folder a session can get Codeman's agent skill in (the registry's
+   * `capabilities.agentSkillDir` values, passed in as data by the caller). A create excludes
+   * `<dir>/codeman/` for each, like the hooks file: a session in a worktree gets an untracked
+   * copy there, and `git worktree remove` would refuse that worktree forever.
+   */
+  agentSkillDirs?: readonly string[];
+}
+
+const realOr = async (p: string): Promise<string> => {
+  try {
+    return await fs.realpath(p);
+  } catch {
+    return resolve(p);
+  }
+};
+
+const DOCKER_REFUSAL =
+  'Worktrees are not available for a folder inside a Docker case workspace: a container can write there, and git would run its hooks and filters on this machine.';
+
+const DOCKER_WORKTREE_REFUSAL =
+  'This worktree is, or holds, a Docker case workspace: git would run inside a folder a container can write, and removing it would delete that workspace. Remove the Docker case first.';
+
+/**
+ * The repository a case folder IS, or why it is not one Codeman makes worktrees for. Never runs
+ * git when the folder (or the repository's common git dir) is at or inside a Docker case workspace.
+ */
+export async function resolveCaseRepository(
+  casePath: string,
+  opts: WorktreeOptions = {}
+): Promise<{ ok: true; repo: CaseRepository } | WorktreeFailure> {
+  const git = opts.git ?? runWorktreeGit;
+  const dockerRoots = await Promise.all((opts.dockerWorkspaces ?? []).filter(Boolean).map(realOr));
+  if (await isInsideAny(casePath, dockerRoots)) return fail('INVALID_INPUT', DOCKER_REFUSAL);
+
+  const run = await git(casePath, [...READ_ONLY_GIT_OPTIONS, 'rev-parse', '--show-toplevel', '--git-common-dir'], {
+    timeoutMs: WORKTREE_READ_TIMEOUT_MS,
+  });
+  if (run.code !== 0 || run.spawnError || run.timedOut) {
+    if (/must be run in a work tree/i.test(run.stderr)) {
+      return fail('INVALID_INPUT', 'This case folder is a bare git repository, which has no work tree to branch from.');
+    }
+    if (/not a git repository/i.test(run.stderr)) {
+      return fail('INVALID_INPUT', 'This case folder is not a git repository.');
+    }
+    return gitFailure(run, 'reading the repository');
+  }
+  const [top, common] = run.stdout.split('\n').map((l) => l.trim());
+  if (!top || !common) return fail('OPERATION_FAILED', 'git did not report the repository root.');
+  const [repoRoot, caseReal] = await Promise.all([realOr(top), realOr(casePath)]);
+  const commonDir = await realOr(resolve(casePath, common));
+  if (await isInsideAny(commonDir, dockerRoots)) return fail('INVALID_INPUT', DOCKER_REFUSAL);
+  if (repoRoot !== caseReal) {
+    if (await isUnrelatedAncestor(repoRoot, caseReal, opts.home ?? homedir())) {
+      return fail('INVALID_INPUT', 'This case folder is not a git repository.');
+    }
+    return fail(
+      'INVALID_INPUT',
+      `Worktrees need the case folder to be the top of a git repository; this folder is inside ${repoRoot}.`
+    );
+  }
+  return { ok: true, repo: { repoRoot, commonDir } };
+}
+
+// ─── IO: listing ─────────────────────────────────────────────────────────────
+
+export interface CaseWorktree extends ParsedWorktree {
+  /** The repository's main checkout (git lists it first). */
+  main: boolean;
+  /** The case folder itself (the main checkout, or a linked worktree linked as the case). */
+  isCase: boolean;
+  /** Under `<repo>/.claude/worktrees/<valid-name>`: Codeman may start sessions in it and remove it. */
+  managed: boolean;
+  /** The managed name, when `managed`. */
+  name?: string;
+}
+
+export async function listCaseWorktrees(
+  repo: CaseRepository,
+  opts: WorktreeOptions = {}
+): Promise<{ ok: true; worktrees: CaseWorktree[]; truncated: boolean } | WorktreeFailure> {
+  const git = opts.git ?? runWorktreeGit;
+  const run = await git(repo.repoRoot, [...READ_ONLY_GIT_OPTIONS, 'worktree', 'list', '--porcelain'], {
+    timeoutMs: WORKTREE_READ_TIMEOUT_MS,
+  });
+  if (run.code !== 0 || run.spawnError || run.timedOut) return gitFailure(run, 'listing worktrees');
+  const parsed = parseWorktreeList(run.stdout);
+  const truncated = parsed.length > MAX_WORKTREES_LISTED;
+  const worktrees = parsed.slice(0, MAX_WORKTREES_LISTED).map((wt, index): CaseWorktree => {
+    const name = index === 0 ? null : managedWorktreeName(repo.repoRoot, wt.path);
+    return {
+      ...wt,
+      main: index === 0,
+      isCase: wt.path === repo.repoRoot,
+      managed: name !== null,
+      ...(name ? { name } : {}),
+    };
+  });
+  return { ok: true, worktrees, truncated };
+}
+
+// ─── IO: path guards ─────────────────────────────────────────────────────────
+
+type DirState = 'dir' | 'absent' | 'symlink' | 'other';
+
+async function lstatState(path: string): Promise<DirState> {
+  try {
+    const st = await fs.lstat(path);
+    if (st.isSymbolicLink()) return 'symlink';
+    return st.isDirectory() ? 'dir' : 'other';
+  } catch (err) {
+    const code = (err as NodeJS.ErrnoException).code;
+    if (code === 'ENOENT') return 'absent';
+    throw err;
+  }
+}
+
+/**
+ * Make sure `<repo>/.claude` and `<repo>/.claude/worktrees` are real directories (created when
+ * absent, refused when a symlink or a file) whose real path is exactly the lexical one.
+ */
+async function prepareWorktreesDir(repoRoot: string): Promise<{ ok: true; dir: string } | WorktreeFailure> {
+  let cur = repoRoot;
+  for (const segment of WORKTREES_DIR_SEGMENTS) {
+    cur = join(cur, segment);
+    const state = await lstatState(cur);
+    if (state === 'symlink') {
+      return fail(
+        'INVALID_INPUT',
+        `${cur} is a symlink. Codeman will not create worktrees through a link that could point outside the repository.`
+      );
+    }
+    if (state === 'other') return fail('INVALID_INPUT', `${cur} exists and is not a directory.`);
+    if (state === 'absent') {
+      try {
+        await fs.mkdir(cur);
+      } catch (err) {
+        if ((err as NodeJS.ErrnoException).code !== 'EEXIST') {
+          return fail('OPERATION_FAILED', `Could not create ${cur}: ${(err as Error).message}`);
+        }
+        if ((await lstatState(cur)) !== 'dir') return fail('INVALID_INPUT', `${cur} is not a plain directory.`);
+      }
+    }
+  }
+  if ((await realOr(cur)) !== cur) {
+    return fail('INVALID_INPUT', `${cur} does not resolve to itself; refusing to create a worktree there.`);
+  }
+  return { ok: true, dir: cur };
+}
+
+/**
+ * The managed worktree `name` of `repo` as it is on disk now: every component a real directory
+ * (no symlinks), the real path the expected one, and git listing it as a worktree of this
+ * repository. Used by removal and by starting a session in it.
+ *
+ * `allowMissing` (removal only, never a session start) also accepts a worktree whose folder is
+ * GONE (deleted by hand, or by an agent) while git still lists it, as prunable, at exactly the
+ * managed path: no component on the way may be a symlink or a file, and `missing` says so.
+ * Without it such a worktree was a dead end: it could not be removed, and its name could not be
+ * used again (the branch exists, then git refuses "a missing but already registered worktree").
+ */
+export async function resolveManagedWorktree(
+  repo: CaseRepository,
+  name: string,
+  opts: WorktreeOptions & { allowMissing?: boolean } = {}
+): Promise<{ ok: true; path: string; worktree: CaseWorktree; missing: boolean } | WorktreeFailure> {
+  const nameError = worktreeNameError(name);
+  if (nameError) return fail('INVALID_INPUT', nameError);
+  const target = managedWorktreePath(repo.repoRoot, name);
+  let cur = repo.repoRoot;
+  let missing = false;
+  for (const segment of [...WORKTREES_DIR_SEGMENTS, name]) {
+    cur = join(cur, segment);
+    const state = await lstatState(cur);
+    if (state === 'absent') {
+      if (!opts.allowMissing) return fail('NOT_FOUND', `No worktree named "${name}" in this case.`);
+      // Nothing below an absent folder exists, so nothing below it can redirect anything.
+      missing = true;
+      break;
+    }
+    if (state === 'symlink') {
+      return fail('INVALID_INPUT', `${cur} is a symlink. Codeman only manages worktrees that are real directories.`);
+    }
+    if (state === 'other') return fail('INVALID_INPUT', `${cur} is not a directory.`);
+  }
+  if (!missing && (await realOr(target)) !== target) {
+    return fail('INVALID_INPUT', `${target} does not resolve to itself; refusing to use it.`);
+  }
+  const listed = await listCaseWorktrees(repo, opts);
+  if (!listed.ok) return listed;
+  for (const wt of listed.worktrees) {
+    if (wt.main) continue;
+    if (missing) {
+      // Lexical only: the folder is gone, so there is no real path to compare. git never calls a
+      // LOCKED worktree prunable (prune skips it), so a locked one is taken too, and the caller's
+      // lock check then refuses it with git's lock reason.
+      if (wt.path === target && (wt.prunable || wt.locked)) {
+        return { ok: true, path: target, worktree: { ...wt, managed: true, name }, missing: true };
+      }
+      continue;
+    }
+    if (wt.path === target || (await realOr(wt.path)) === target) {
+      return { ok: true, path: target, worktree: { ...wt, managed: true, name }, missing: false };
+    }
+  }
+  if (missing) return fail('NOT_FOUND', `No worktree named "${name}" in this case.`);
+  return fail(
+    'NOT_FOUND',
+    `${target} is not a registered git worktree of this repository (a leftover folder, or one git no longer knows about).`
+  );
+}
+
+// ─── IO: excludes ────────────────────────────────────────────────────────────
+
+/**
+ * Append each of `lines` that git does not already ignore to `<commonDir>/info/exclude`. Returns
+ * the lines written. A symlinked `info` dir or `exclude` file is left alone (reported, never
+ * written through).
+ */
+async function ensureExcluded(
+  repo: CaseRepository,
+  probes: Array<{ line: string; probe: string }>,
+  git: WorktreeGitRunner
+): Promise<{ added: string[]; warning?: string }> {
+  const needed: string[] = [];
+  for (const { line, probe } of probes) {
+    const run = await git(repo.repoRoot, [...READ_ONLY_GIT_OPTIONS, 'check-ignore', '-q', '--', probe], {
+      timeoutMs: WORKTREE_READ_TIMEOUT_MS,
+    });
+    // 0: ignored. 1: not ignored. Anything else (128) is a git error: assume not ignored, since an
+    // extra exclude line is harmless and a missing one leaves the worktrees untracked.
+    if (run.code !== 0) needed.push(line);
+  }
+  if (needed.length === 0) return { added: [] };
+  const infoDir = join(repo.commonDir, 'info');
+  const excludePath = join(infoDir, 'exclude');
+  try {
+    const infoState = await lstatState(infoDir);
+    if (infoState === 'absent') await fs.mkdir(infoDir);
+    else if (infoState !== 'dir')
+      return { added: [], warning: `${infoDir} is not a plain directory; nothing was excluded.` };
+    const fileState = await lstatState(excludePath);
+    if (fileState === 'symlink' || fileState === 'dir') {
+      return { added: [], warning: `${excludePath} is not a plain file; nothing was excluded.` };
+    }
+    const existing = fileState === 'absent' ? '' : await fs.readFile(excludePath, 'utf-8');
+    const { text, added } = excludeAppendText(existing, needed);
+    if (text) await fs.appendFile(excludePath, text, { mode: 0o644 });
+    return { added };
+  } catch (err) {
+    return { added: [], warning: `Could not update ${excludePath}: ${(err as Error).message}` };
+  }
+}
+
+// ─── IO: one operation per repository at a time ──────────────────────────────
+
+const repoLocks = new Map<string, Promise<unknown>>();
+
+/** Serialize create/remove per repository: two creates must not both pass the "absent" check. */
+async function withRepoLock<T>(repoRoot: string, fn: () => Promise<T>): Promise<T> {
+  const prev = repoLocks.get(repoRoot) ?? Promise.resolve();
+  const next = prev.then(fn, fn);
+  const tail = next.catch(() => undefined);
+  repoLocks.set(repoRoot, tail);
+  try {
+    return await next;
+  } finally {
+    if (repoLocks.get(repoRoot) === tail) repoLocks.delete(repoRoot);
+  }
+}
+
+// ─── IO: create ──────────────────────────────────────────────────────────────
+
+export interface CreateWorktreeRequest {
+  name: string;
+  /** New branch; default `worktree-<name>`. */
+  branch?: string;
+  /** Start point; default `HEAD` (the case's current commit). */
+  base?: string;
+}
+
+export interface CreatedWorktree {
+  name: string;
+  path: string;
+  branch: string;
+  base: string;
+  /** The commit the new branch starts at. */
+  head: string;
+  /** Lines appended to the repository's `info/exclude` (empty when git already ignored them). */
+  excludeAdded: string[];
+  warnings: string[];
+}
+
+export async function createCaseWorktree(
+  repo: CaseRepository,
+  req: CreateWorktreeRequest,
+  opts: WorktreeOptions = {}
+): Promise<{ ok: true; worktree: CreatedWorktree } | WorktreeFailure> {
+  const git = opts.git ?? runWorktreeGit;
+  const nameError = worktreeNameError(req.name);
+  if (nameError) return fail('INVALID_INPUT', nameError);
+  const branch = req.branch || defaultWorktreeBranch(req.name);
+  const branchError = worktreeBranchError(branch);
+  if (branchError) return fail('INVALID_INPUT', branchError);
+  const base = req.base || 'HEAD';
+  const baseError = worktreeBaseError(base);
+  if (baseError) return fail('INVALID_INPUT', baseError);
+
+  return withRepoLock(repo.repoRoot, async () => {
+    // A name already taken says so in those words, before the branch check would report the
+    // derived branch instead. Rechecked after the folders are prepared, below.
+    try {
+      if ((await lstatState(managedWorktreePath(repo.repoRoot, req.name))) !== 'absent') {
+        return fail('ALREADY_EXISTS', `A worktree named "${req.name}" already exists in this case.`);
+      }
+    } catch {
+      /* unreadable: the check after the folders are prepared reports it */
+    }
+    // git's own word on the branch name. The output must be the input: `--branch` expands
+    // shorthands, and anything it rewrote is not the name the user typed.
+    const refFormat = await git(repo.repoRoot, [...READ_ONLY_GIT_OPTIONS, 'check-ref-format', '--branch', branch], {
+      timeoutMs: WORKTREE_READ_TIMEOUT_MS,
+    });
+    if (refFormat.code !== 0 || refFormat.stdout.trim() !== branch) {
+      return fail('INVALID_INPUT', `"${branch}" is not a valid branch name.`, gitRefusalText(refFormat.stderr));
+    }
+    const existingBranch = await git(
+      repo.repoRoot,
+      [...READ_ONLY_GIT_OPTIONS, 'show-ref', '--verify', '--quiet', `refs/heads/${branch}`],
+      { timeoutMs: WORKTREE_READ_TIMEOUT_MS }
+    );
+    if (existingBranch.code === 0) {
+      return fail('ALREADY_EXISTS', `A branch named "${branch}" already exists. Pick another worktree name or branch.`);
+    }
+    const resolved = await git(
+      repo.repoRoot,
+      // No `--end-of-options` (git 2.30+): the base already passed the ref gate, which refuses a
+      // leading `-`, so it cannot be read as an option.
+      [...READ_ONLY_GIT_OPTIONS, 'rev-parse', '--verify', '--quiet', `${base}^{commit}`],
+      { timeoutMs: WORKTREE_READ_TIMEOUT_MS }
+    );
+    const head = resolved.stdout.trim();
+    if (resolved.code !== 0 || !/^[0-9a-f]{40,64}$/.test(head)) {
+      return fail('INVALID_INPUT', `"${base}" is not a commit in this repository (an empty repository has none yet).`);
+    }
+
+    const prepared = await prepareWorktreesDir(repo.repoRoot);
+    if (!prepared.ok) return prepared;
+    const target = join(prepared.dir, req.name);
+    try {
+      if ((await lstatState(target)) !== 'absent') {
+        return fail('ALREADY_EXISTS', `${target} already exists. Pick another worktree name.`);
+      }
+    } catch (err) {
+      return fail('OPERATION_FAILED', `Could not inspect ${target}: ${(err as Error).message}`);
+    }
+
+    const skillProbes: Array<{ line: string; probe: string }> = [];
+    for (const dir of new Set(opts.agentSkillDirs ?? [])) {
+      const line = agentSkillExcludeLine(dir);
+      if (line) skillProbes.push({ line, probe: `${dir}/codeman` });
+    }
+    const excluded = await ensureExcluded(
+      repo,
+      [
+        { line: WORKTREES_EXCLUDE_LINE, probe: `${WORKTREES_DIR_SEGMENTS.join('/')}/${req.name}` },
+        { line: SETTINGS_LOCAL_EXCLUDE_LINE, probe: '.claude/settings.local.json' },
+        ...skillProbes,
+      ],
+      git
+    );
+
+    // The resolved commit, not the ref text, after `--`: nothing here can be read as an option,
+    // and `--no-track` keeps a worktree started from `origin/main` from tracking (and pulling) main.
+    const add = await git(repo.repoRoot, ['worktree', 'add', '--no-track', '-b', branch, '--', target, head], {
+      timeoutMs: WORKTREE_ADD_TIMEOUT_MS,
+    });
+    if (add.code !== 0 || add.spawnError || add.timedOut) {
+      const text = gitRefusalText(add.stderr);
+      if (/already exists/i.test(text)) return fail('ALREADY_EXISTS', `git refused: ${text}`, text);
+      return gitFailure(add, 'creating the worktree');
+    }
+    return {
+      ok: true,
+      worktree: {
+        name: req.name,
+        path: target,
+        branch,
+        base,
+        head,
+        excludeAdded: excluded.added,
+        warnings: excluded.warning ? [excluded.warning] : [],
+      },
+    };
+  });
+}
+
+// ─── IO: remove ──────────────────────────────────────────────────────────────
+
+export interface RemoveWorktreeRequest {
+  name: string;
+  /** Also `git branch -d` the worktree's branch (never `-D`). */
+  deleteBranch?: boolean;
+  /** True when a live session runs at or inside `path`: the removal is refused. */
+  isInUse: (path: string) => boolean | Promise<boolean>;
+}
+
+export interface RemovedWorktree {
+  name: string;
+  path: string;
+  branch: string | null;
+  branchDeleted: boolean;
+  /** git's refusal to delete the branch (typically "not fully merged"); the worktree is gone regardless. */
+  branchError?: string;
+}
+
+export async function removeCaseWorktree(
+  repo: CaseRepository,
+  req: RemoveWorktreeRequest,
+  opts: WorktreeOptions = {}
+): Promise<{ ok: true; removed: RemovedWorktree } | WorktreeFailure> {
+  const git = opts.git ?? runWorktreeGit;
+  return withRepoLock(repo.repoRoot, async () => {
+    // A worktree whose folder is gone is removable too: `git worktree remove` then only drops
+    // git's record of it (measured on git 2.43: exit 0, nothing on disk touched).
+    const found = await resolveManagedWorktree(repo, req.name, { ...opts, allowMissing: true });
+    if (!found.ok) return found;
+    // The case passed the Docker rule, but the worktree folder itself can be a Docker case's
+    // workspace (a Docker case linked to it), or hold one. `git worktree remove` runs a status
+    // INSIDE the tree first (container-written files, checkout filters on this machine), then
+    // deletes the whole tree, ignored files included, under a container that still mounts it.
+    // A folder that is gone has neither problem: git only drops its record.
+    if (!found.missing) {
+      const dockerRoots = await Promise.all((opts.dockerWorkspaces ?? []).filter(Boolean).map(realOr));
+      if (dockerRoots.some((root) => isPathWithin(found.path, root) || isPathWithin(root, found.path))) {
+        return fail('INVALID_INPUT', DOCKER_WORKTREE_REFUSAL);
+      }
+    }
+    if (await req.isInUse(found.path)) {
+      return fail(
+        'CONFLICT',
+        `A live session is still running in worktree "${req.name}". Close it first, then remove the worktree.`
+      );
+    }
+    if (found.worktree.locked) {
+      return fail(
+        'CONFLICT',
+        `Worktree "${req.name}" is locked${found.worktree.lockReason ? ` (${found.worktree.lockReason})` : ''}. Unlock it with git first.`
+      );
+    }
+    const removed = await git(repo.repoRoot, ['worktree', 'remove', '--', found.path], {
+      timeoutMs: WORKTREE_REMOVE_TIMEOUT_MS,
+    });
+    if (removed.code !== 0 || removed.spawnError || removed.timedOut) {
+      if (removed.timedOut || removed.spawnError) return gitFailure(removed, 'removing the worktree');
+      const text = gitRefusalText(removed.stderr);
+      // git's own refusal (modified or untracked files, a lock): Codeman never forces, so this is
+      // the answer, in git's words.
+      return fail(
+        'CONFLICT',
+        text ? `git refused to remove the worktree: ${text}` : 'git refused to remove the worktree.',
+        text
+      );
+    }
+    const branch = found.worktree.branch;
+    let branchDeleted = false;
+    let branchError: string | undefined;
+    if (req.deleteBranch && branch) {
+      const del = await git(repo.repoRoot, ['branch', '-d', '--', branch], { timeoutMs: WORKTREE_READ_TIMEOUT_MS });
+      if (del.code === 0) branchDeleted = true;
+      else branchError = gitRefusalText(del.stderr) || 'git refused to delete the branch.';
+    }
+    return {
+      ok: true,
+      removed: {
+        name: req.name,
+        path: found.path,
+        branch,
+        branchDeleted,
+        ...(branchError ? { branchError } : {}),
+      },
+    };
+  });
+}

src/hooks-config.ts

@@ -36,6 +36,7 @@ import { join, dirname } from 'node:path';
 import { fileURLToPath } from 'node:url';
 
 import type { HookEventType } from './types.js';
+import type { AgentSkillDir, CliCapabilities } from './config/cli-registry/types.js';
 import { HOOK_TIMEOUT_SECONDS } from './config/auth-config.js';
 import { dataPath } from './config/instance.js';
 import { readJsonConfig, SETTINGS_PATH } from './web/route-helpers.js';
@@ -1228,8 +1229,8 @@ interface AgentSkillFile {
  */
 async function readAgentSkillSource(): Promise<AgentSkillFile[]> {
   const src = agentSkillSourceDir();
-  const skill = await readFile(join(src, 'SKILL.md'), 'utf-8');
-  const files: AgentSkillFile[] = [{ relPath: 'SKILL.md', content: `${skill.trimEnd()}\n\n${AGENT_SKILL_MARKER}\n` }];
+  const skillMd = await readFile(join(src, 'SKILL.md'), 'utf-8');
+  const files: AgentSkillFile[] = [{ relPath: 'SKILL.md', content: `${skillMd.trimEnd()}\n\n${AGENT_SKILL_MARKER}\n` }];
   let referenceNames: string[] = [];
   try {
     referenceNames = (await readdir(join(src, 'reference'))).filter((name) => name.endsWith('.md')).sort();
@@ -1285,7 +1286,8 @@ export type AgentSkillApplyResult =
 
 /**
  * Install or refresh the Codeman agent skill into `skillDir` (a `.../codeman`
- * directory, e.g. `<case>/.claude/skills/codeman` or `~/.claude/skills/codeman`).
+ * directory, e.g. `<case>/.claude/skills/codeman`, `~/.claude/skills/codeman` or
+ * `<case>/.agents/skills/codeman`).
  *
  * Refuses two shapes rather than writing through them:
  * - a SYMLINK at the skill dir or its `skills/` parent: this repo's own dogfooding
@@ -1336,16 +1338,20 @@ export async function installAgentSkillInto(skillDir: string): Promise<AgentSkil
 }
 
 /**
- * Seed a claude session's agent preamble file (`$XDG_CACHE_HOME/codeman-agent-<id>.sh`,
+ * Seed a session's agent preamble file (`$XDG_CACHE_HOME/codeman-agent-<id>.sh`,
  * default `~/.cache/`) from the packaged `skills/codeman/preamble.sh`, so the agent
  * skill's §0 bootstrap collapses to a two-line loader instead of a ~150-line block the
  * model has to type out (measured live: that paste alone cost a spawn run ~47 s of
  * generation time). The path formula must match the skill's
  * `${XDG_CACHE_HOME:-$HOME/.cache}` exactly; sessions inherit the server's env, so
  * reading the server's own XDG_CACHE_HOME keeps the two in agreement (`||` mirrors the
- * shell's `:-`, treating empty as unset). Callers gate to LOCAL claude sessions (a
- * remote or in-container HOME is not this filesystem) and treat it as best-effort: the
- * skill's §0 fallback block self-heals a missing or stale file.
+ * shell's `:-`, treating empty as unset). Callers gate to LOCAL sessions whose CLI gets
+ * the skill (`agentSkillDir()` non-null, any CLI: the preamble is bash + curl + jq, nothing
+ * claude-specific; a remote or in-container HOME is not this filesystem) and treat it as
+ * best-effort: the skill's §0 fallback block self-heals a missing or stale file. Codex
+ * caveat: codex 0.162 runs tool commands in a shared app-server daemon whose
+ * `CODEMAN_SESSION_ID` is the starting session's, so its loader looks for that session's
+ * file and its self id is wrong; the skill tells the agent how to notice.
  */
 export async function seedAgentSessionPreamble(sessionId: string): Promise<void> {
   const content = await readFile(join(agentSkillSourceDir(), 'preamble.sh'), 'utf-8');
@@ -1478,31 +1484,59 @@ export async function removeAgentSkillFrom(skillDir: string): Promise<AgentSkill
     }
     await rmdir(join(skillDir, 'reference')).catch(() => {}); // fails when non-empty, fine
     await rmdir(skillDir).catch(() => {});
-    await rmdir(dirname(skillDir)).catch(() => {}); // prune `.claude/skills` if now empty
+    await rmdir(dirname(skillDir)).catch(() => {}); // prune `.claude/skills` (or the CLI's own dir) if now empty
     return 'removed';
   });
 }
 
 /**
- * Add or remove the Codeman agent skill in `<case>/.claude/skills/codeman`,
- * mirroring `applyStatusLineConfig`'s shape. Gated by the synced `agentSkillEnabled`
- * app setting (default OFF); callers gate on Claude mode, since the skill is discovered
- * via `.claude/skills/`, which only Claude Code reads.
+ * Claude Code's project skills directory, and the meaning `agentSkillInjection: true`
+ * had before `capabilities.agentSkillDir` existed.
+ */
+export const CLAUDE_AGENT_SKILL_DIR: AgentSkillDir = '.claude/skills';
+
+/**
+ * The workspace-relative skills directory this CLI scans, i.e. where its sessions get
+ * the skill, or `null` when they get none. Pure registry data, so no call site
+ * branches on a CLI id.
+ */
+export function agentSkillDir(
+  capabilities: Pick<CliCapabilities, 'agentSkillInjection' | 'agentSkillDir'> | undefined
+): AgentSkillDir | null {
+  if (!capabilities?.agentSkillInjection) return null;
+  return capabilities.agentSkillDir ?? CLAUDE_AGENT_SKILL_DIR;
+}
+
+/** `<casePath>/<dir>/codeman`, the directory the skill's SKILL.md lives in. */
+export function agentSkillDirFor(casePath: string, dir: AgentSkillDir): string {
+  return join(casePath, dir, 'codeman');
+}
+
+/**
+ * Add or remove the Codeman agent skill in `<case>/<dir>/codeman`, where `dir` is the
+ * skills directory the session's CLI scans (default: Claude Code's `.claude/skills`;
+ * callers resolve it with `agentSkillDir()`), mirroring `applyStatusLineConfig`'s
+ * shape. Gated by the synced `agentSkillEnabled` app setting (default OFF).
  *
  * Call-site policy is ADD-ONLY on session create (callers pass `enabled: true` or
- * skip the call), for the statusLine reason: sessions in a repo share one `.claude/`
- * dir, so a single create while the setting is off must not yank the skill out from
+ * skip the call), for the statusLine reason: sessions in a repo share one workspace,
+ * so a single create while the setting is off must not yank the skill out from
  * under other live sessions.
  *
  * ⚠️ Consequence: turning `agentSkillEnabled` OFF sweeps nothing. There is deliberately
  * no server-side toggle-off sweep (it would have to walk every case, including ones
- * with live sessions, and would hit exactly the shared-`.claude/` hazard above), so
+ * with live sessions, and would hit exactly the shared-workspace hazard above), so
  * already-injected copies stay on disk until removed per case with
- * `codeman skill uninstall --case <name>`. The `enabled: false` branch here backs that
- * CLI and the tests; it has no server call site. Keep the README's Agent Skill note in
- * sync if this ever changes.
+ * `codeman skill uninstall --case <name>`, which removes the Codeman-managed copy from
+ * every skills directory the registry could have written. The `enabled: false` branch
+ * here backs that CLI and the tests; it has no server call site. Keep the README's
+ * Agent Skill note in sync if this ever changes.
  */
-export async function applyAgentSkill(casePath: string, enabled: boolean): Promise<AgentSkillApplyResult> {
-  const skillDir = join(casePath, '.claude', 'skills', 'codeman');
+export async function applyAgentSkill(
+  casePath: string,
+  enabled: boolean,
+  dir: AgentSkillDir = CLAUDE_AGENT_SKILL_DIR
+): Promise<AgentSkillApplyResult> {
+  const skillDir = agentSkillDirFor(casePath, dir);
   return enabled ? installAgentSkillInto(skillDir) : removeAgentSkillFrom(skillDir);
 }