The owner's verdict on the chooser: nice, but not nice enough, too much just
blue. He also wants "Run a local model" on top of the start screen, above
Run Claude Code.
The memory bar is now a map: the model, its context and the brief peaks side
by side on the track of what this Mac can give a model, each with its size in
a legend, and the verdict on the fit beside the figure. recommend() sends what
the need is made of as `Candidate.memory`.
Colour now carries meaning: green for good and tested, amber for tight, orange
for slow, teal for the context. Violet is the feature's own colour, on its
start screen button and its main action. Every row has a size tile and a speed
gauge scaled to the fastest row in the list. The light skins get deeper values
of the same colours, and the window is 40 px taller to hold it without
scrolling.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The window is one frame for the whole flow: a fixed height on desktop, App
Settings' boxed close button, the buttons in a foot of their own outside
the part that scrolls, so a click never moves the next button away and
nothing is cut off; the short screens are centred in it on desktop and sit
above the buttons on a phone.
The chooser lists the pick first and then the other models by download
size, largest first, a column the reader can see. A pick that only just
fits says "Best fit, tight". The memory figure is an estimate and says so
once, in whole GB; its legend names the peak, and the installed line uses
the same words beside what the model takes now, so the two figures are not
read as a contradiction. The legend, the column headers and the foot's
promise are set in --text-dim, the bar's track in --text-muted so its free
part can be seen, and only the download figure keeps tabular figures.
The consent screen is the download as a table: the chosen row, the runtime,
the coding agent with its installer's command as the row's second line, the
total in the strong weight, then where it goes in one line. The download bar
means the model download alone and is gone once the model is in; the screen
that ends an install shows "Start a session" alone; an error on an installed
model is the headline with a grey dot; an error's first sentence is the
headline; Remove is quiet everywhere but red in its own question. The
no-recommendation reasons are one sentence each, and the untested note is
in Codeman's voice. A refusal the server names is printed from its sentence.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The connect step records the connection in the state file BEFORE it writes
the endpoint record and the settings flag (the flag's earlier value read by
inspectLocalEndpoint()), so a state write that fails or a server killed
between the two never leaves a record of ours that the state file does not
know; a record under the reserved id that says exactly what ours would say
is ours. changePort() writes the state first and puts the old port back when
the endpoint move is refused.
A job whose end could not be written (a full disk) is kept in memory: the
view shows it, Resume and a new install start from it, the next verb writes
it, and Cancel never answers a 500. A first install on a full disk is a 507
with the sentence. "installing" in the file while this process runs no job
always reads as paused or that failure.
An installed model whose runtime does not start says so (runtime-start-failed
with Start as the next step, and the launch check says it too); the stored
runtime version is started while its folder is unpacked. Stop, Remove and
the boot of anything that is not a running installation end a runtime an
earlier server left behind before any file is deleted. Stop and "Use another
port" answer 409 sessions-open while a live session uses the model. A pull
that failed because the runtime is gone is runtime-stopped, not a network
problem. After "does not fit" only a model that needs less memory is offered.
A replaced job's partial blobs are deleted. Every 400 carries its sentence as
data.message, and the six inline sentences are labels with zh-CN entries.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
POST /api/local-model/install, cancel, resume and runtime, and DELETE
/api/local-model. The GET gains an `install` object the window polls. All of
it is experimental and admin only in multi-user mode.
One job of five steps (runtime, model, verify, agent, connect) driven by a
pure reducer over explicit state, step and transition tables. One job at a
time (409). The consent is bound to the size shown: `acknowledgedBytes` must
be the server's own total (409 size-changed). A job a server restart
interrupted comes back paused and never continues without a click. Every
failure is a code with one next step.
The model is judged against the budget the runtime measured before it is
downloaded. A missing coding agent is installed with its stock command
through the shared runner, audited. Connect writes one endpoint record under
the reserved id through mutateCustomModelHosts() and switches
customModelEndpointsEnabled on through mutateSettingsFile(), remembering the
flag's value. The model's digest is stored and read again at boot and before
every session start on that endpoint; a changed model refuses the launch.
Remove answers 409 with the session names while a live session uses the
endpoint, undoes the two writes first, restores the flag only when this
feature switched it on and no other endpoint is saved, and deletes only
inside local-model/. The coding agent stays.
The server starts the runtime at boot when a model is installed and stops
it at shutdown beside stopDeepSeekWeb().
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The catalogue rows carry the figures measured on a Mac (the MLX build of
qwen3.6:35b-a3b is 23,105,798,956 bytes, about 37 KB held per context token,
about 3 GB of runner overhead, about 5 GB more for a moment while a long
prompt is processed). A model's need is now its peak. Rows that were not
measured use the same per-token and overhead figures and say `estimated`.
qwen3.6:35b-a3b gets the tested entry for an Apple M5 Pro with 48 GB and
OpenCode, with its measured 67.7 tokens per second, so that row may say
"Tested on a Mac like this".
The ranking changes with it: a pair tested on a Mac like this one comes
first, then speed (ok, unknown, slow), then model rank, harness rank and the
larger context. The fit grade decides whether a row is listed and what it
says, it no longer sorts, so a model is offered at the largest context it
fits at. recommend() takes the budget the runtime measured
(`budgetSource: measured`), and modelFitsBudget() judges one model against it.
Candidate rows now always carry `needSource`, `harnesses`, `installBytes`,
`installGb` and, where one applies, `tag`. `kvBytesPerToken` and the prompt
cache tiers are gone.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
- `LocalModelService.launchRefusal(endpointId)`: what a session launch asks
before it starts on the managed endpoint. It answers the error that
refuses the launch (the runtime child is not alive, another program holds
the port, the runtime keeps stopping, the model's digest changed) or
nothing. New error code `runtime-not-running` with its sentence, for a
runtime the person stopped: only a refused session start says it.
- The fake runtime's MacBook log line is rebuilt from the two values the
measurement notes quote, so it is `INFERENCE_COMPUTE_LINE_MACBOOK` now.
Only the Mac mini line is a verbatim capture and keeps `REAL_`.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The types, the module skeletons and the test fake that let the runtime, the
install job and the window be built at the same time.
Types:
- src/types/local-model-install.ts (new): the runtime pin, the install
states, steps and error codes with one sentence and one next step each,
the progress, the `install` object of GET /api/local-model, the bodies and
409 answers of the install verbs, the state file. `LocalModelStatus` moves
here; `install` is optional until the route emits it.
- src/types/local-model.ts: the catalogue row's Mac memory figures, the
measured speed of a tested entry, the `measured` budget source,
`RecommendOptions`, and the row tag, the coding agents and the install
total of a candidate. All optional until the catalogue and `recommend()`
carry them, so the read-only stage keeps compiling and its tests stay
green.
Modules with final signatures and bodies that throw:
runtime.ts (the compiled pin is final), ollama-client.ts, developer-tools.ts,
state-store.ts, install-job.ts (the state, step and transition tables are
final), connect.ts, custom-model-hosts-writer.ts, cli-install-runner.ts.
Nothing imports them yet, so the server is unchanged.
Implemented: paths.ts (every folder of the feature, from dataPath()), and
`modelFitsBudget()` in recommend.ts, the judgement the job makes with the
measured budget before it downloads a model. `recommend()` takes the options
parameter and does not read it yet.
Test fake: test/fixtures/fake-ollama.ts, a runtime on a port the system
picks, in the shapes measured on a Mac: version, tags, ps, the pull stream
(MLX and GGUF lines, an error line, a cut stream, a resumed pull), chat with
and without a tool call, the model list, a runtime archive with Range
support, and a fake child process that logs the device total.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Review fixes for the read-only recommendation and its card.
Recommendation (pure):
- `limited` is now a statement about the Mac: no model with 12 billion
parameters or more fits. It is judged on everything that fits, before
the disk filter and whatever the pick is. A 32 or 36 GB Mac whose pick
is a 9B model no longer reads "This Mac can run a small model" above a
list of 20B, 27B and 35B models, and knowing the chip no longer lowers
the verdict. A small pick under `ok` carries a note that says so.
- A model that fits the memory but has no room on the disk is no longer
dropped in silence: the best-ranked one is named in a note with the GB
that are missing. A full disk cannot turn a 48 GB Mac into a
small-model Mac.
- "Models fit, but not at a context a harness needs" has its own
sentence with the budget and the need, worded as an estimate. The
16 GB Mac mini misses by 1.3 points (10.3 of 10.4 GB) and was told
flatly that its memory is too little, the sentence of an 8 GB Mac.
- The installed memory is named by the size the Mac is sold as (48, not
51.5): the hardware line and the memory sentence. Every computed size
stays decimal GB from the one formatter. An unknown memory size is no
longer reported as "0.0 GB".
- Slow is graded on the unrounded speed: 9.95 to 9.9996 tokens per
second printed as 10 and read ok.
- The disk sentence names the 10 GiB that must stay free, so its
numbers add up for the reader.
Card and Settings group:
- The Settings group "Model on this Mac" is hidden on a server that is
not a Mac, and until the first answer says it is one.
- Only `403 FORBIDDEN` is a refusal. `403 PASSWORD_CHANGE_REQUIRED`
(an admin who must change the password) is a failed request and is
asked again, so that admin no longer loses the card until a reload.
- Each fit and speed label carries what it measures as its tooltip and
accessible name, an alternative with a speed estimate prints it, and
the count beside "Other models that fit" is gone (it was the number
of rows, capped at four, not the number of models that fit).
- The line saying nothing can be set up yet sits right under the pick,
in the colour of every other line (it was 3.9:1), and says whose step
it is.
- The dismissal key is named for this stage, so hiding the read-only
card does not hide the install offer later.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The label list was "frozen for the parallel build" and the card's test said
the route answers 501: both were true while the pieces were written side by
side and are not any more. Comments only.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The core of the read-only recommendation: what this Mac is, what the catalogue
knows, and which model fits. Nothing is downloaded, installed or written.
- hardware.ts: one bounded async spawn of /usr/sbin/sysctl with seven keys,
without -n, parsed as "key: value" lines by the pure parseSysctl(). Free disk
of the data dir through statfs behind the bounded path probe. Fallbacks from
node:os, a 60 s cache, never throws, never system_profiler. A wired limit of
0 means "not set".
- local-model-catalog.ts: seven models, three harness rows (pi gated) and the
chips table as a compiled constant, validated by a Zod schema on first use.
No override file. Model ids are name:tag in lowercase without a slash; every
model carries paramsB and an (empty) tested list. Chip names match on the
end of the chip string, so 'M4' is not the chip of 'Apple M4 Pro'.
- recommend.ts: device budget (the wired limit when above 0, else 2/3 of
memory below 36 GB and 3/4 from 36 GB, less the reserve), need = weights +
KV cache + 1 GiB + prompt cache, fit good up to 0.85 and tight up to 0.98,
speed ok, unknown or slow (a chip in no chips row is unknown), and the
ranking. Verdicts ok, limited (best model under 12 billion parameters) and
none with its reason. "Recommended" only for a pick with a tested entry for
this Mac, which no row has today.
- types/local-model.ts: comments brought in line with the code (chip matching,
budgetShare rounding, a budget of 0 on an Intel Mac).
- cli-registry-no-id-branching.test.ts: the guard also catches a comparison of
a `cli` field with a CLI id (plus one stray blank line Prettier refused).
Tests: test/local-model-format.test.ts, test/local-model-hardware.test.ts
(the captured output of two real Macs), test/local-model-catalog.test.ts,
test/local-model-recommend.test.ts.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The contract three builders code against in parallel. Nothing is recommended
yet: the route answers 501 and the catalogue is empty.
- src/types/local-model.ts: HardwareFacts, the catalogue rows, HarnessFact,
Candidate, Recommendation, LocalModelStatus, and LOCAL_MODEL_LABELS, the one
list of sentences the server may put into the payload (the zh-CN dictionary
is keyed by them).
- src/local-model/format.ts: formatGb, the one place a byte count becomes a
size a person reads (decimal GB), and memClassGb.
- src/config/local-model-catalog.ts: the Zod schema (model ids are name:tag,
lowercase, no slash) and an empty stock catalogue. No override file.
- Skeletons with their final signatures: hardware.ts, recommend.ts,
harness-facts.ts, routes/local-model-routes.ts (injectable dependencies),
public/local-model-ui.js with its script tag.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
105 new em-dashes sat in comments, test names and one test fixture from the
foreign tmux adoption work. Replaced by hand with commas, colons or
parentheses, one reading at a time. The fixture now quotes the real refusal
text, which says "It may still be running, try again."
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The attach line now takes the rule #770 gave remote and docker launches: bashCArg() single-quotes it at both respawn-pane sites, so it reaches bash verbatim. quoteAdoptLaunch is retired, since its double quotes kept the same $ and backtick safety but lost a real newline. #510's no-cd condition stays.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
candidatesFrom had moved outside the try when the transports moved to
runForeignProbe, so a classifier failure would have reached the route as a 500.
The module promises it never throws, so the guard is back for local and docker.
Docs:
- The api reference said `blocked` waits are always refused, and listed Ralph as
refused while two of its file routes were not (they are now).
- The invariants doc and CLAUDE.md gain the rules this review turned up: target
syntax in names, the ping before list-panes, nothing typed into an adopted pane,
quoteAdoptLaunch.
- CLAUDE.md no longer says every claude session gets hooks or a stop wait.
- ForeignTmuxSession's mode and workingDir describe the first listed pane, not
the active one.
- The wiki had no page on adoption: The Dashboard gains an Opened outside Codeman
section, and the remote and docker pages point at it. docs/docker-cases.md now
separates an adopted container from an adopted tmux session.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The owner's answer to the PR's open question: the switch does not also pass
--sandbox workspace-write. Settings, wiki, skill and release note now state
the consequence: a folder Codex runs read-only keeps the network closed and
asks per command. A golden test pins the decision by name, and the stale
launch-defaults comments now name networkAccess.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Brings the 1.33.2 based adoption feature up to master. Conflicts were
resolved by keeping master's intent and re-applying the feature on top:
both sides kept for imports, the script tag and the stylesheet blocks; the
trust-dialog guard keeps master's stopped check plus the adopted refusal;
the Ralph and image watchers keep master's remote branch plus the adopted
skip; the close confirmation keeps master's worktree option plus the
adopted wording.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Conflict in src/session-cli-registry-bridge.ts: kept the PR's gateVersion() (gates measured against the gated CLI's own version) and added master's Discord startup hint engine value.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
getCustomModelForPersist() trusted homeEnvVar to name a path, but the registry
types it as any env name, so a clis.json override could write a session's own
credential to state.json. Only an absolute path is written now.
A re-apply resolved the CLI's real home from what the session's map held, which
under a selection is the injected directory, and recovery resolved none at all.
An entry the own home lacks (settings.yaml) was then linked from the server's
home on an endpoint switch and on recovery. The route now reads the new
Session.ownEnvValue(), and recovery reads the clamped record through the same
bounded guard, under the bulk cap since it walks every saved session. Also
corrects the comment that said only a removed directory gets its links again.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The lookup judges liveWorktree.path lexically and only when the pane moves,
while the route's probe and readdir both follow symlinks. A symlink put at
that path, or at a parent of it, without moving the pane re-rooted the listing
outside workingDir. Names and sizes leaked, while file-content and file-raw
still refused the paths.
The route now resolves both folders on every request and walks the worktree
only when it lies strictly inside the resolved workingDir, otherwise it walks
workingDir. Only the check uses the resolved paths: the walk keeps the session's
spelling, so a workspace reached through a symlink still gets paths relative
to workingDir.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
transcriptPageKeys applied to every pi, and a pi in `tuiMode: regular` binds
PageUp/PageDown to its editor. It emits no mouse enable, so the flag stays
false, and on a fresh pane the buffer is hollow: the wheel sent PageUp, the
caret of a multi-line draft moved, and the next keystroke landed on the wrong
line. Measured on real pi 1.1.0, 1.4.0 and 1.5.0. Regular mode is not exotic,
it only takes one line in pi's settings. Telling the two modes apart would
need a new capability, since the paging gate takes no id, so pi has no
fallback: a fullscreen pi with "Wheel scrolls local history" on stays where
master left it.
altScreen, the half of the fix that does the work, was pinned by nothing:
reverting it left 1328 tests green. The strip decision table reads its
expected value from the same registry entry. pi now has its own block using
the bytes it really emits, which also pins that focus reporting survives.
The fullscreen default dates from pi 1.0.0, not 1.4.0, so the comments say
that, and the sites still calling pi narrow-strip or main-screen by default
are corrected. Where a comment listed CLIs it now names the capability.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude Code's EnterWorktree moves the running process into
<repo>/.claude/worktrees/<name>, while Session.workingDir stays on the main
checkout. The Files panel kept walking workingDir, so it listed a tree the
agent had left.
paneCwd already follows the move, measured on Claude Code 2.1.318 in both
directions, so nothing new has to track the agent. It also follows every cd
in a shell, so a new SessionState.liveWorktree narrows it to a linked
worktree, read from the .git file without starting git.
Only a folder strictly below workingDir is ever an answer, and the listing
only starts there: returned paths stay relative to workingDir, so the
preview, raw and edit routes are unchanged and confinement does not widen.
A worktree outside the session folder stays unfollowed.
The panel asks with root=live, shows a strip naming the worktree with a
switch to the session folder, and follows the agent in and out while open.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Codeman itself always lived under $HOME; administrator rights were only
spent on what it needs first (system packages for Node.js, tmux and git,
Homebrew on a Mac). An account without those rights could not install it:
the installer died in ensure_sudo, or in Homebrew's own installer.
install.sh now has a no-admin route, taken by itself when the account has
no administrator rights (no sudo binary; outside sudo/wheel/admin and
`sudo -n true` fails; a refused sudo password; a Mac account outside the
admin group with no writable Homebrew) and forced with --no-admin or
CODEMAN_NO_ADMIN=1:
- Node.js: the official tarball from nodejs.org, checked against the
SHASUMS256.txt published beside it, into ~/.codeman/runtime/node.
- tmux: the tmux project's own build (tmux/tmux-builds 3.7c), checked
against a SHA-256 pinned in the installer, into ~/.codeman/runtime/bin.
- Both are linked from ~/.local/bin (never over a file that is not ours),
and the runtime is on PATH and on the LaunchAgent's PATH.
- git is never installed. Without a working git the source comes as an
archive, the install is marked with .install-source, and it updates by
re-running the installer: the new tree is built beside the old one and
swapped in only after the build succeeded. A later re-run on a machine
that has gained git turns it back into a clone.
- Nothing on this route runs sudo, Homebrew or `xcode-select --install`,
and the route is never decided by `sudo -v`.
What it cannot supply is checked before the first download: a C/C++
toolchain on Linux (node-pty 1.1.0 compiles there), musl, and macOS below
15 (the tmux build's deployment target).
Along the way:
- check_git no longer trusts `command -v git`. On a Mac without the Xcode
Command Line Tools /usr/bin/git is a stub that opens Apple's install
dialog; it is judged by `xcode-select -p` and never executed. Any other
git is asked for its version.
- scripts/git-hooks.mjs answers "not a checkout" without running git, for
the same reason: postinstall calls it on every npm install.
- The in-app updater knows the install kind `archive` and points such an
install at the installer instead of `npm i -g aicodeman@latest`.
- download() and download_to_stdout() resolve DOWNLOADER themselves. Only
main() used to set it, so `install.sh cloudflared` died on `set -u` with
the message swallowed by the step log.
- The AI CLI menu leaves out a Homebrew-only install command where there
is no Homebrew.
Verified with the real installer: a bare Ubuntu 24.04 container as an
unprivileged user (install, a PTY round trip through the API, update by
re-run, cloudflared, uninstall), the same without a compiler (stops before
any download), with a broken then working git, with sudo present but not
usable, and on macOS 26 (arm64) in an isolated HOME with sudo, brew, git,
the compilers, launchctl and tailscale replaced by tripwires. Not verified
on a real standard macOS account: that account was an administrator.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Applying a custom model writes the endpoint's values over the keys it
injects, and clearing the model unset those keys. A session that had set
one of them itself lost that value: its own DSH_HOME, CODEX_HOME,
GROK_HOME or GEMINI_API_KEY was gone once the model was cleared, and so
was CLAUDE_CONFIG_DIR, which silently moved a Claude session onto the
default account. Switching endpoints did the same.
setCustomModel() now records what the session held under each injected
key and restores it on clear and on switch. Quick-start passes the
caller's values to the Session constructor so they are seen before the
selection lands.
The selection overwrites the only other copy, tmux's setenv, so the
session's own home directory has to survive a Codeman restart on its
own. __customModel therefore stores that one path, the single exception
to persisting keys only, and recovery runs it through
clampEnvOverridesForOwner() so a revoked grant cannot bring it back. Any
other own value, a credential included, stays in memory only.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Owner decision. Typing the note into a brand-new non-Claude tab spent a model
turn per tab and, for a spawned worker or a scheduled run, landed with the first
real prompt. Claude keeps its hint through the system prompt, and every other CLI
is told nothing. The setting goes with the path it gated: sessionIdle, the fresh
set, isHintableSession and the session:idle wiring. sendInput loses `fromUser`,
which was only ever false for the hint, and the deps test pins that a reply is a
person's own input.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
`cwd` beside the fixed `workingDir` reads ambiguously on a public API surface, and
`paneCwd` says where the value comes from, like `paneExit`. Also pin the field on
the payload the browser really receives (getSessionStateWithRespawn), and give the
pane-directory text its own architecture-invariants paragraph.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Measured on 1.47.0 with the same "spawn a shell worker, echo, read, delete"
task: DeepSeek Harness and OpenCode already pass end to end; Codex 0.162
failed three ways, fixed here.
- Start Codex 0.162+ with --no-daemon (new noDaemonFlag gate, fail-closed).
The shared app-server daemon stopped every launch on an "incompatible
feature settings" dialog when another version's daemon was running, and
ran commands with another session's CODEMAN_* environment.
- Evaluate registry version gates against the gated CLI's own version
(resolveSessionCliVersion) instead of claude's for every entry.
- Add the opt-in synced setting App Settings > Codex > Allow network access
(codexNetworkAccess, default off): --config
sandbox_workspace_write.network_access=true, filled as a launch default
and clamped off for non-granted owners like the bypass switch.
- Skill: carry on with the curl path when `codeman` is not on PATH, re-run a
sandbox-blocked call once through the CLI's approval path, name the new
setting. Wiki and registry docs updated.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A relative path printed in the terminal was resolved against the session's
workingDir, which is fixed at create time and never follows a cd. In a shell
that moved into a subdirectory or a sibling repo, the link opened the wrong
file or reported "File not found".
The pane-exit watcher's batched list-panes read already carries
#{pane_current_path}. derivePaneCurrentPaths() turns those rows into one
directory per single-pane session, TmuxManager.getPaneCurrentPath() serves it
from memory, and the server applies it on the existing paneExitsUpdated tick
as SessionState.cwd, broadcasting only on a change. No new tmux call, no new
SSE event. A failed read keeps the previous answers.
Session.setCwd() uses the paneExit scoping rule: remote and Docker sessions
never get a cwd (their local pane runs the ssh or docker exec client), nor do
direct-PTY or rediscovered sessions. The value is runtime-only.
In the browser, relativeResolutionBase() picks the live cwd when known and
workingDir otherwise. It only changes the base a relative path joins: the
result is routed as before, so a target outside the workspace still goes
through the attachment routes, and file-content/file-raw are unchanged.
The scheduler is the part that runs with nobody watching, so it is decided per folder
rather than per session: sessions sharing a folder are the same clone and share one run
that both see, `inFlight`/`lastRunAt` are keyed by the resolved folder, the schedule is
stamped when a folder is picked instead of when its run starts, and a tick is skipped
while the previous one is still going. The click route joins a run already going for the
same folder rather than starting a second fetch in one clone.
The tick keeps ONE bound for the git processes of the whole run and hands it to every
folder, and bounds its folders with a bound of its own: the same semaphore cannot serve
both levels, because a folder holding a slot would then wait for a slot its own fetch
needs. Both bounds are `FETCH_CONCURRENCY`, so the ceiling is unchanged. The bound claim
in CLAUDE.md, the `Limiter` jsdoc and docs/api-reference.md now describes it as it is.
Before anything is fetched the folder is probed with the bounded path probe, so a case on
a mount that stopped answering is skipped and never touched.
The rest of the review: the hooks are named as they really are (`reference-transaction`,
`pre-auto-gc` through the maintenance a fetch triggers, config commands such as
`credential.helper`) and `FETCH_CONFIG_ARGS` turns maintenance and gc off for an
unattended run; the claim that an agent in the same repository is not disturbed is
replaced by the ref-lock collision it actually is; the click route publishes its own run
as `trigger: 'click'`; `autoSaveGitFetch` uses `_apiJson` so a 400 shows the error toast
and applies the returned settings; and the two wording fixes.
Tests: the scheduler pins one run per folder, the two bounds, the probe before anything
runs, and that closing the session holding the override keeps the folder while another
session is still open on it; the route test pins the click publish and the join, asserting
the join before the held run is released — releasing first lets the run finish and clears
the lane, so the joiner would be entitled to its own fetch and the test would pass for the
wrong reason.
Fetch is the only git command Codeman runs that writes, and the only one that runs a
repository's own fetch-post hook, so it stays behind a click and off by default. The Git
status window's header row gains Fetch; App Settings -> Bottom bar gains the switch and
the interval, synced because the scheduler acts on them, and Session Options -> Git
overrides both per case. The scheduler reads settings fresh each tick, times the interval
from the start of a run, never overlaps runs, and re-checks the Docker and remote rules
every tick before any git runs. Nothing but remote-tracking refs moves: HEAD, the index
and the working tree are proven unchanged against real git while origin/main advances.
Tests: src/git-fetch.ts and the scheduler, the per-case settings route, the synced/display
setting split, the two new i18n rows, and the Fetch block of the Git status browser suite.
Conflict in CLAUDE.md's State Files line resolved by keeping #661's new
config-backup entries and this PR's "case default groups" in state.json.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Case settings gains a "Default session group" selector (Ungrouped plus every
named tab group of the owner). A new session of the case lands in that group
in the same versioned layout write that registers it, so it is drawn in place
with no Ungrouped-to-group jump.
- Stored per owner and case in state.json (caseDefaultGroups), as the group's
stable id or null for Ungrouped; read and written only through
TabLayoutService under the owner lock.
- Precedence in sessionCreated: explicit tabGroupId, then the parent
placement, then the case default, else Ungrouped. A deleted or foreign id
never fails a spawn and is forgotten only after the creating write lands.
- POST /api/quick-start (local, linked, remote, Docker) and
/api/ralph-loop/start pass their case; POST /api/sessions infers it
lexically from workingDir. The create response carries tabLayout when a
case default placed the session.
- GET/PUT /api/cases/:name/default-group; deleting a case (or an admin
deleting a user's case folder) forgets the value.
- The selector mirrors desktop and phone, reverts on a refused save, and is
redrawn from every adopted tab layout; zh-CN strings added.
Claude Code lets a subscription's one-hour prompt cache expire after an hour idle, and the first turn back re-writes the whole conversation at the cache-write rate; nearly every gap over an hour goes cold, while the saving is concentrated in gaps of one to four hours. Codeman now sends one short keep-alive turn five minutes before the expiry Claude Code reports on its statusline, so the next real prompt still reads the conversation from cache, and stops two hours (configurable) after the last real prompt.
Codeman owns the idle policy for these sessions, as agreed in Discussion #554: a launch with keep-warm on also carries idleCompaction: false in its one --settings JSON, so Claude Code's own idle compaction cannot replace the conversation with a summary a minute before the nudge. It is decided once per launch (keep-warm on, the statusLine exporter injected, Claude Code 2.1.290 or newer, an unknown version fails closed like --name), recorded on the mux record, and picked up by any relaunch; a launch with keep-warm off is byte-identical. A session switched on after launch is kept warm best-effort and the readout says so.
Every guard lives in one pure rulebook: the opt-in (an agent-spawned session does not inherit the app default), a launch that reports cache state, a warm 1h cache with enough tokens at stake and no compaction this gap, a live pane, no usage-limit pause, no respawn, Ralph, task or waiting agent, idle, no open dialog, an empty composer and no recent keystroke. The nudge is a submit and moves lastSubmitAt, which the response viewer's conversation correlation needs; the cap counts from its own persisted lastRealPromptAt instead. Read My Mind drops the exact nudge text, and the turn's stop hook and idle prompt open pre-acknowledged so nothing alerts. The last cache report and the clock survive a restart, so the timer re-arms.
Review of 8659ba25..be987fbb:
- Usage: a Free account answers premium_interactions {entitlement: 0, remaining: 0,
percent_remaining: 0, unlimited: false, has_quota: false}, which read as 100% used. The parser now
returns null when has_quota is false or the entitlement is not positive, with that payload as a
fixture. Free meters chat (200) and completions (2000) too, so the comment no longer says they
are unlimited on every plan.
- agent read: rows were skipped until a `❯ <prompt>` row, so an answer taller than the pane
(prompt scrolled off the top) read as "". With no prompt row AND no banner / trust dialog on
screen, the rows above the composer are now the tail of the last answer. Text typed in the
composer (a prompt-shaped row between the two rules) is not a prompt row.
- Page keys: measured on 1.0.95 in tmux that Copilot pages on PageUp/PageDown (three PageUp moved
lines 109-140 to 010-042, PageDown returned) as well as on SGR wheel reports, so copilot joins
claude and opencode in _localScrollbackIsHollow(); the stale-tracking-flag fallback the comment
and docs describe is now true, for the main pane and a TerminalTile (same function), with tests.
- X-GitHub-Api-Version is 2022-11-28, a version GitHub supports (verified against the live
endpoint); readPlanUsageTelemetryEnabled() is checked before isCopilotAvailable() so a switched-off
chip skips the resolver probe too.
- Palette label for Copilot in panels-ui.js; stale comments in types/session.ts and stock.ts.
- Docs: Past Sessions continues with --continue (resumeSessionId is for API callers), nothing sets
continueSession on a respawn, and only a tab name the user set is passed as --name; the usage
wording for Free; the wheel and mouse-strip paragraphs in CLAUDE.md and
architecture-invariants; the changeset describes the whole feature.
Copilot (`copilot`, npm @github/copilot) as a stock CLI entry with the same launch options as the other agent CLIs: model (--model), resume (--resume), continue (--continue) and allow-all (--yolo, clamped off for non-granted multi-user owners). Adds a resolver and GET /api/copilot/status probe, the agent-image package, Run-menu/mobile/logo entries, skill docs and tests.
Claude Code's tracker nulls recache_tokens_if_cold at a compaction or a tool-result clearing and only counts the expected rebuild at the next request, which also closes the window, so no report carried both and the old rule never fired. Stamp on the first warm null report after one that carried a value, carry it through idle re-reports and drop it at the next request; the readout now says "Context compacted at HH:MM" since a clearing opens the same window.
Also persist a changed report so a crash restores the latest one, rename the session field to statusLineExporter (statusLineTelemetry is the registry capability) and document its values, word the stake as re-cached, move the readout block above the SSE staleness heading, and drop keep-warm mentions until it lands.
Claude Code's statusline has carried a prompt_cache object since v2.1.251, and Codeman's statusline exporter already received it, but StatusTelemetrySchema stripped the whole object, so Codeman never knew whether a session's conversation was still cached. Keep it and show it.
StatusTelemetrySchema now 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), while parsePromptCache normalizes only the subset Codeman stores, so a hostile localhost payload cannot bloat state. Session.promptCache rides toState with a change-only session:promptCache event, and the route broadcasts the stored object.
The Session Options (Respawn tab) readout shows warm-until or cold with the re-read stake, the cache hit rate, the last miss's causes, and, for Claude Code's own idle compaction (recache_tokens_if_cold null while expected_rebuilds rose), a "compacted at HH:MM" note that is never raised as an alert. setPromptCache stamps that on a new object rather than mutating the parsed report, so the stamp reaches the broadcast. Sessions without a statusline of their own also get a cache:until HH:MM / cache:cold group in the terminal footer. The last report is restored on a Codeman restart (like displayModel) so the readout is not blank until the pane's next render; the readout is time-aware, so a value whose expiry has passed reads as cold, not stale-warm. A launch without the exporter records why (collection off, workspace statusline, remote) so the row says the cause instead of "not reported yet" forever.
This is the sensor a later opt-in keep-warm would read, and the TTL it reports is the source of truth rather than the plan tier.
Comment and doc corrections that #555 made stale, no behaviour change.
- stock.ts: the grok and omp altScreen comments compared their strip to
opencode's, which is now strip-mux-and-mouse rather than the narrow
strip. Grok now says it shares antigravity's strip until measured, and
omp drops opencode from its comparison.
- terminal-ui.js: the touch-tap comment named Claude/Codex/Gemini as the
stripped modes, but the gate is now the cliMouseTracking flag alone and
covers opencode too, so it names the two stripping flavours instead.
- src/types/session.ts: the cliMouseTracking JSDoc (the flag the browser
now gates on exclusively) listed only claude/codex/gemini; it now names
the strip-full and strip-mux-and-mouse modes, including opencode under
tmux.
- src/session.ts: the usesMux getter doc now names isMuxMouseStripMode,
since the replay strip passes usesMux to it as well.
- docs/architecture-invariants.md: the narrow-strip list gains
grok/deepseek/omp (matching the PR's own CLAUDE.md line), the
"must REMEMBER" heading covers both DECSET-stripping flavours, and the
cliMouseTracking writer is described as the full-or-mouse branch it
really is.
- docs/wiki/The-Dashboard.md: the user manual said every non-Claude CLI
scrolls locally; opencode's wheel and swipes now page its conversation.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
displayModel gains a `config` source, ranked below any report from the running
CLI and above the launch model: custom endpoint, then statusline or screen,
then config, then launch, then nothing. The screen still wins whenever it
names a model, since that is what the running TUI uses.
- Registry data: capabilities.modelDetect gains `configResolver`, a NAMED
reader (src/model-config-resolvers.ts), like a launcher profile; dsh names
'deepseek-route' (the reader from the previous commit). `screenLine` becomes
optional; the schema refuses a modelDetect naming nothing, an unknown
reader, or screenLines without a screenLine.
- Session: the reader runs from _withPaneLifecycle's finally, so at every pane
start, attach and relaunch, with the session's own launch config
(legacyConfigForMode) and env (its clamped overrides, then the server's), so
a per-session DSH_HOME is the home read. Async; a read that lands after a
newer one or after the session stopped is dropped; a remote or docker
session reads nothing locally. A change emits displayModelChanged
(broadcast and persist). Not restored after a restart: the next attach
reads it again, and a restored screen value outranks it.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A session header can only name the model a session runs if the server knows
it, so SessionState gains `displayModel: { model, source }`, resolved in a pure
module (src/session-display-model.ts), strongest first:
- custom-endpoint: a Custom Model Endpoint Profile's modelId answers the
session, whatever alias the CLI prints;
- statusline / screen: the newest report from the running CLI itself.
Claude's statusLine exporter already posts model.display_name on every
render; the status-telemetry route now records it (only for a CLI with
capabilities.statusLineTelemetry). A CLI whose registry entry declares the
new capabilities.modelDetect has its footer read off the pane capture the
idle/working probe already takes (no extra tmux call), so an in-session
/model switch is followed at the next transition;
- launch: the model the session was launched with (claude's --model or the
app-wide default, another CLI's <cli>Config.model), read where the registry
says the model param lives;
- nothing known: no field, never a placeholder.
modelDetect is registry data, measured on live panes: dsh-TUI's status line
on the row under its composer (qwen3.8-27b on the owner's route) and codex's
`<model> <effort> ·` footer on its last row. Both anchor on chrome only that
CLI draws, over the last rows of the screen only; a transcript line shaped like
the footer is never taken (fixture tests). The pattern goes through
compileVersionRegex() with exactly one capture group, checked at load time.
An unreadable or covered footer keeps the last model (unlike the watching
label: a model does not stop running when something covers its row). Model
text is untrusted: escape sequences and control characters are stripped and it
is capped at 64 characters. A change emits displayModelChanged, broadcast
(session:updated) and persisted; a restart restores a CLI-reported model until
the next report.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
- hooks-config: a probe the bulk cap refused gets ONE bounded re-probe past the
cap (probeBeforeTouching), and whatever is still unknown is skipped. The
per-spawn hook and statusLine helpers used to fall back to an unbounded
lstat/readFile there, which on a dead workspace never settled and could take
the last threadpool workers (and hang the boot hook sweep). New test: cap
engaged, stat/lstat/readFile hanging on two more paths; both helpers return.
- describeUnknownPath()/unknownPathReason(): POST /api/sessions, quick-start and
GET /api/cases/:name now say a folder was not checked (other mounts are still
not answering) instead of blaming a healthy folder at the stall ceiling.
errorCodes unchanged.
- #535 x #516: Create in a custom folder probes the parent through the bounded
probe before realpath/stat/lstat/readdir touch it; an unknown parent is 422
OPERATION_FAILED (UNREACHABLE) within the probe timeout. New test.
- Docs: MAX_STALLED default is 2 (follows UV_THREADPOOL_SIZE), CaseInfo
.unreachable covers a refused probe, the boot sweep skips an unanswering
workspace, a CLAUDE.md gotcha for bounded probes, verbs.md documents the 422
(plugin mirror synced), api-reference documents the custom-folder 422.
- Tests: the launcher case-lookup describe is no longer nested in the Grok
block, and the cap-below-ceiling test no longer depends on an inherited
UV_THREADPOOL_SIZE / CODEMAN_PATH_PROBE_MAX_STALLED.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The bounded path probe answered "absent" both when a path did not exist and
when it simply did not answer, so a stalled linked case 404'd and the Run
button scaffolded a stray local case over it, and two stalled paths anywhere
made every unrelated path read as absent (hooks skipped, statusLine
overridden, the clone warning lost).
- probePath()/probePathKind() are tri-state: present (or directory/file),
absent (ENOENT/ENOTDIR only) and unknown (timeout, other errors, refusal).
boundedPathExists() stays as the display-only boolean.
- A stalled path takes only its own mount out of probing (deepest mount
point from /proc/self/mounts, never /; just the path itself when there is
no mount table). Unrelated paths keep probing. The process-wide cap is a
backstop that answers unknown, and a single-path user request can probe
past it ({ pastCap: true }), still bounded and still recorded as stalled.
One console.warn when a path first stalls and one when the cap engages.
- GET /api/cases/:name keeps NOT_FOUND for definite absence only. An
unreachable linked case answers with its registered path and
unreachable: true; a local one answers OPERATION_FAILED. runClaude and
runShell create a case only on errorCode NOT_FOUND. The case list keeps an
unreachable linked case, marked unreachable, instead of dropping it, and
fix-plan reports an unreadable plan as an error, not "no plan".
- applyWorkspaceHooks and the statusLine helpers skip only a workspace that
is absent or on the stalled mount; a capacity refusal no longer stops
hooks being installed elsewhere, and an unreadable settings file never
lets the exporter override a user's own statusLine.
- The clone flow's repo-settings warning is back on its synchronous check,
and stripCaseEnvKeys uses pathExistsForWrite.
- POST /api/sessions (workingDir) and POST /api/quick-start (case folder)
probe with the bounded probe instead of statSync/existsSync. Missing and
non-directory keep INVALID_INPUT; unknown is OPERATION_FAILED, and
quick-start never scaffolds over a folder that did not answer.
- PATH_PROBE_TIMEOUT_MS and MAX_STALLED_PATH_PROBES move to
src/config/path-probe.ts, overridable via CODEMAN_PATH_PROBE_TIMEOUT_MS
(default 1500) and CODEMAN_PATH_PROBE_MAX_STALLED (default 3), and are
documented in the Settings Reference.
- The probe is exported from the utils barrel and imported from there.
Codeman maintainer