Owner decision 2026-10-13 (#547): a rendered .html page behaves like a
file opened from disk in a browser. It may load scripts, styles, fonts,
images and media from any host and make fetch, XHR, WebSocket, beacon
and form requests to any host.
The /html-view Content-Security-Policy is therefore the `sandbox`
directive alone (HTML_VIEW_CSP, now exported and documented): the same
sandbox flags as before, never `allow-same-origin`, and no fetch
directive. It replaces Codeman's own `default-src 'self'` policy on these
responses. Everything that confines a page stays: the opaque origin (no
Codeman cookies or document, API writes with `Origin: null` refused, API
reads not readable cross-origin), and the capability serving only the
page's own folder under every path, symlink, dot, extension and
confinement rule.
A new test builds the route behind registerSecurityHeaders, as in
production, and pins the exact header: one `sandbox` directive with the
current flags, no `allow-same-origin`, no `-src` directive. The security
architecture, architecture invariants, API reference, CLAUDE.md and the
Working With Files wiki page now say what a page can do (reach the
network, read its own folder) and cannot do (Codeman's origin, cookies,
API, other folders), and that this is the owner's decision.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
An in-workspace .md path clicked in the terminal went to the log viewer
(tail -f, a "streaming" header over raw text), because md was kept out
of FILE_PREVIEW_EXTENSIONS on purpose; every other surface already
rendered it. Owner decision 2026-10-13: markdown is a document to read,
so md and markdown join FILE_PREVIEW_EXTENSIONS and a click or tap opens
the file preview, rendered by default with the MD pill for source. The
main pane, every grid tile and the split's Pane B share the link
provider, so all of them route the same way. Other in-workspace text
(.log, .txt, .json, code) keeps the log viewer and its live follow.
FILE_PATH_LINK_PATTERN now also links `.markdown` and `.htm`, which
were not clickable at all.
Tests: file-preview-markdown.test.ts now asserts md and markdown ARE
routed to the preview (and plain text is not); terminal-touch-tap pins a
tapped .md, .markdown and .html opening the preview while .log, .txt and
.json open the log viewer, and a tile-registered provider routing the
same way against the tile's own session. Docs: CLAUDE.md, the
architecture invariants and the wiki (Working With Files, Mobile Guide).
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A rendered page's own <video> or <audio> could not seek, because the
route streamed a plain 200 without Accept-Ranges. sendFileBody() and
localFileSource() are now exported from file-routes.ts and the serving
route uses them, so it advertises Accept-Ranges and answers a Range with
206 + Content-Range (or 416), exactly like file-raw, the attachment raw
route and /api/download. The body sender hijacks the reply and copies
every header set before it (the sandbox CSP, CORS, no-store), which the
new test pins on a 206.
Media types stay in ASSET_TYPES: with Range support they work, and they
are served under the same sandbox CSP as everything else.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The mint refused an HTML file whose absolute path had any dot segment, so
a workspace that itself lives under a hidden folder (~/.local/src/app)
could never render one of its own pages. For a file inside the workspace
the rule now covers the path below the workspace root, which is all the
page's tree can reach; outside the workspace it still covers the whole
path. A hidden directory below the workspace root is still refused.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The serving route checked the blocklist per request but not
attachmentConfineToWorkspace, while the attachment serve path
(file-routes.ts) re-checks it every time. A capability minted for a
folder outside the workspace before an admin turned confinement on kept
serving that folder for the rest of its TTL.
The capability record now says whether its folder was admitted from
outside the workspace, and the serving route refuses such a record while
confinement is on. The flag is part of the reuse key, so an
out-of-workspace mint never picks up a capability recorded as
in-workspace for the same folder.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The serving route checked hidden segments and the asset allowlist on the
requested name only; after realpath it checked containment and the
blocklist. A symlink `cfg.json -> .git/config` inside the page's folder
was therefore served as 200 application/json. The dot rule now also runs
on the segments of relative(rootDir, resolvedPath), and the extension
allowlist on the resolved file name, which also types the response.
Both realpath calls on the request path (the mint of an out-of-workspace
file and the serving route) are now `await fs.realpath` instead of
realpathSync, so a slow mount stalls one request, not the event loop.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
os.tmpdir() returns $TMPDIR when it is set, and macOS sets it in every
terminal session, so with TMPDIR set /tmp, /private/tmp and /var/tmp were
never treated as broad and /tmp/report.html got a capability over all of
/tmp. The broad-root list now names /tmp and /var/tmp literally (through
canonical(), so macOS gets /private/tmp), refuses any directory that is an
ancestor of a broad root (/home, /Users, /private), and treats a drive's
mount point (/mnt/d on WSL, /media/x, /Volumes/x) as broad as well.
The new test sets TMPDIR elsewhere and still gets /tmp refused.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Conflicts resolved:
- CLAUDE.md: kept master's Prompt uploads paragraph and its longer
File-path links paragraph (XLSX admission limits), and inserted the
PR's html-view sentence after the download-only rule.
- src/web/public/terminal-ui.js: kept master's tile-aware
getSessionId() in the shared file-path link provider and added the
PR's .html routing to the file preview on top of it.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The XLSX preview hands an admitted copy of the workbook to ExcelJS 4.4.0,
which links each worksheet to its comments, comment notes (VML), tables and
drawing by looking the relationship Target up verbatim in maps keyed by its
own spelling, then dereferences the result unchecked. openpyxl, the library
most agents write .xlsx files with, uses absolute Targets
(/xl/tables/table1.xml), its own comment part name (xl/comments/comment1.xml,
which ExcelJS never loads) and a drawing XML ExcelJS parses to nothing. One
comment, table or chart was enough for "Spreadsheet preview failed".
repairExcelJsModel() (spreadsheet-xlsx-core.js) runs on ExcelJS's parsed
model right before its reconcile: it re-points a link to the key ExcelJS
stored for the same part (so an openpyxl table is kept) and drops a link,
tablePart, drawing or hyperlink reference that would dereference nothing.
It adds nothing and resolves keys as own properties only. The worker wraps
workbook.xlsx.reconcile per instance to call it, and throws if a future
ExcelJS renames that method. SPREADSHEET_ASSET_VERSION is refreshed.
Tests: unit tests on model objects, and three real openpyxl 3.1.5
workbooks in test/fixtures/xlsx that fail the worker tests without the hook.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The agent skill only reached Claude Code: injection wrote
.claude/skills/codeman, which no other CLI reads, and the per-session
preamble was seeded for claude alone. A Codex, Gemini, OpenCode,
Antigravity, Grok, DeepSeek or OMP session inside Codeman never learned it
could start and drive workers.
- Registry field capabilities.agentSkillDir (a closed enum, since
clis.json can set it and it decides where the server writes) names the
skills directory each CLI scans. Claude keeps .claude/skills; those seven
CLIs get the same skill in <case>/.agents/skills/codeman (each measured to
scan it; Claude Code does not, so it never sees two copies). pi gets none:
a project skill makes pi ask "Trust project folder?" at startup.
- The preamble is seeded for every local session whose CLI takes the
skill; nothing in it is claude-specific.
- The skill text is CLI-neutral and names the two CLI-side failures to
report instead of retrying: a sandbox that blocks loopback (Codex's
default), and commands that do not carry this session's environment
(Codex 0.162 runs them in a shared app-server daemon).
- POST /api/sessions injects after the launch pre-flight, so a CLI that
cannot start leaves nothing in the workspace.
- codeman skill install gains --mode; uninstall --case removes the copy
from every directory the registry can write.
- The settings toggle moves to Agents & CLIs > All agent CLIs.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A Docker case is a local session for the file routes, but `~` inside the
container is the container's own seeded home. Expanding `~/x.png` on the host
opened a different file than the one the agent printed. The attachment POST
and POST …/folders now answer 400 for `~/` when the session has `docker` set
(absolute paths are unaffected: the workspace is a bind mount).
Docs: the resolved-target gate, the route-parameter limit, the listing join
rule and the Docker `~/` refusal in architecture-invariants.md,
security-architecture.md and CLAUDE.md.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
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.
- Multi-user: a non-admin's folder must resolve inside their own userSpacePath
or the session's workingDir (folderAllowedForUser, the path picker's rule),
checked on the resolved folder before it is read and again on each served
file after symlink resolution; 403 otherwise.
- listFolder probes the local folder with probePathKind({ pastCap: true })
first (unknown -> 503 via describeUnknownPath, never 404), then async
fs.realpath; resolveFolderFile uses probePathKind instead of fs.stat.
- The guard runs on the realpath before readdir.
- FOLDER_PATH_LINK_PATTERN ends a folder at ':' (quadratic `://` rescan on
colon runs; PATH=/a:/b linked as one folder).
- Load counters drop stale preview / folder-grid responses; closing the grid
invalidates a listing still in flight.
- validFolderFileName refuses a leading dot; the remote listing matches
extensions in any case.
- Docs: folder grid in the z-index lists; multi-user rule and bounded probe in
the invariants and security docs.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Review follow-ups for loading a shell from tmux's rendered history.
tmux adds no history while a pane is in the alternate screen, and a
nested tmux client (remote SSH and Docker cases, or tmux run by hand)
holds its pane there, so a full capture of it is one screen. The
display-message query that reads #{history_size} now also reads
#{alternate_on} (PaneCaptureOptions.capturedAlternateOn), and a request
carrying `alt=recording` (only SHELL_LOAD_QUERY sends it) is answered
the old way on such a pane: the visible frame over the tailed byte
recording. Agents, scroll pulls and Load full history are unchanged.
alternate_on cannot tell a nested tmux from vim or less, so a local
shell with one of those open takes the same fallback while it is open.
A full capture without `lines=` was reported as a 'tail' cut whenever
the pane held more than the configured ceiling (possible on tmux before
3.7), offering a Load full history that could never return more. The
cut is now reported only for a client bound below the ceiling.
The response carries `historyLinesCut` when that bound left lines in
tmux, and the banner then says earlier output may be retained instead
of naming the bytes a later tail cut, which undercount what tmux holds.
A bounded grid tile's shell load sends its own `lines=` (via
SHELL_LOAD_BASE_QUERY). CLAUDE.md, docs/architecture-invariants.md and
docs/api-reference.md describe the shell load, the cut report and the
fallback.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Review follow-ups for the remote scrollback read.
The 1.5 s slot-wait bound was checked inside the task, so it only fired
once runWithRemoteSshLimit() had already handed over a slot, after
everything queued ahead of it. runWithRemoteSshLimit() now takes
`maxWaitMs`: a waiter that times out leaves the queue, rejects with
RemoteSshSlotTimeoutError, never runs and never touches `active`. The
history read uses it, and a slot timeout does not back the host off.
Only a transport failure backs the host off now (ssh exit 255, a
timeout, a spawn error, output that is not the script's), via
classifyRemoteAliveExit(). A remote command failing is about that
session only. A live run against a real host showed a missing session
comes back as exit 0 with empty markers, because the `| tail -c` pipe
takes tail's exit status, so markers that arrived without a size are
also per-session.
fetchRemoteScrollback() takes an optional runner as a test seam; tests
now cover the bounded wait, each back-off class, its expiry, and the
masked missing-session case. A source pin covers the `-S 0` capture
branch. The two seams that can drop rows at the join (rows that scroll
during the round trip, a taller shared client) are documented in
docs/remote-sessions.md, and the limiter's header and the remote-ssh
invariants now name the history read as an ssh consumer.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
pageKeysForGesture kept `prepaid` set after the travel had fully repaid a
pre-paid page, so the one-second debt lifetime still applied: on a 38-row
pane, 19 lines of -1 at 32 ms left { pending: 0, prepaid: true } and a -1
300 ms later sent nothing. `prepaid` now clears as soon as the pending
travel reaches or crosses zero, so the leftover is plain travel and the
150 ms gap applies again. Touch contacts are unaffected (never split by
time). The mixed parity step's reversal grows from 20 to 25 rows because
the -5 rows left after repayment are now plain travel and cancel against it.
Review of #527 (Ark0N, 2026-10-11):
1. A slow touch drag paged a full screen per line. touchmove calls the
fallback once per whole cell, so a slow drag's calls were further apart
than the 150 ms gesture gap and each pre-paid a page (12 lines at 170 ms
sent 12 PageUps). The touch paths now tag every call, touchmove and
momentum frames alike, with a per-contact touchId (incremented on
touchstart), and pageKeysForGesture never splits a contact by time; a
new contact is a new gesture.
2. A pre-paid debt now survives up to PAGE_KEY_DEBT_MS (1 s) of quiet when
the next event goes the same way, so slow small wheel ticks (40 px every
250 ms, a Firefox 1-line notch) pay it back instead of each pre-paying.
A reversal inside the gap (or inside one contact) drops the debt and
accumulates without pre-paying, so a +/-3 px wobble no longer pages
back and forth. A reversal after the gap still pages at once. Main pane
and tiles share the helper, so both get it.
3. A wheel consumed by the dialog check now logs page-keys-dialog-open.
4. The Dashboard wiki names newer Codex beside OpenCode; the invariants
paragraph describes the new gesture boundaries.
Tests: unit regressions through the real _maybePageCliTranscript, helper
cases, tile/main parity for ticks and wobble, and a Chromium test that
drives the real initTerminal touch listeners (12 PageUps before, 1 after).
One conflict, in i18n.js SKIP_SELECTOR: master added '.tab-name' itself
(the same fix this branch made for zh-CN tab names), so master's line is
kept and this branch's duplicate comment dropped.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Conflicts resolved by keeping both sides:
- terminal-ui.js: master's wheelDeltaWholeLines/sgrWheelReports next to
this branch's pageKeyRows/pageKeysForGesture.
- terminal-tile.js: master's tile SGR wheel forwarding and Shift+wheel
local scroll (_shiftScrollPending) kept; tile paging stays on the
shared pageKeysForGesture state (_pageKeyGesture) with its per-tile
dialog check.
- architecture-invariants.md: this branch's hollow-paging paragraph,
with master's tile SGR forwarding and Shift+wheel sentences replacing
the old "do not forward SGR wheel" note.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
CLAUDE.md key pattern and architecture-invariants (File-path links) get
the new path forms (`findFilePathLinks`, `~/` expanded server-side,
`terminalPathLine` join rules) and the folder grid (grant + bare-name
routes, own registry, z-order, captured keys); remote-sessions lists the
folder routes for a remote case; security-architecture states the trust
level of a folder grant.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The paste-image route wrote under session.workingDir via ensureUploadDir
with no bounded probe, so an upload onto a stalled network mount held one
of libuv's few threadpool workers until the mount answered. Probe the
workspace with probePathKind past the stall cap before the first disk
touch and refuse an unknown mount through describeUnknownPath, the way
POST /api/sessions already does.
A shell's automatic loads (tab select, dropped-output recovery, clear
terminal, a split/grid tile's load) replayed `session.terminalBuffer`,
the raw PTY recording, ahead of the visible frame. That recording holds
full-screen repaints (an ssh'd tmux, less, any TUI: cursor home, then the
rows joined by newlines, prompt on the last row) drawn for the pane
height of the moment. Replayed into a terminal with fewer rows, each
repaint's last newline scrolls and pushes its top row into scrollback,
so one line tmux holds once shows up dozens of times (26 copies in the
case that found this, after the pane went from 24 to 23 rows). Remote
SSH shell sessions hit it constantly: their inner tmux repaints often.
Shells now load a bounded window of tmux's rendered history instead
(`SHELL_LOAD_QUERY`: full=1&tail=1MiB&lines=10000), the same source the
scroll-to-top pull already used. `lines=10000` bounds the synchronous
capture-pane, measured at ~8 ms per 1000 lines (~0.1 s here vs ~0.8 s
for a 100k-line history). Because that bound is shallower than xterm's
scrollback, the server now reports a cut: the capture's existing
display-message query also reads `#{history_size}`
(`PaneCaptureOptions.capturedHistorySize`), and the route marks a full
capture `truncated` / 'tail' when tmux holds more than it read, so the
banner keeps offering Load full history and scroll-to-top keeps pulling
older lines. The byte cap still reports 'capped'. The banner no longer
names an amount it cannot know ("less than 1 KB more") for such a cut.
Agent sessions are unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A phone-only toolbar Attach button hands the agent a photo or a video with the
keyboard down. It owns its own file input, positioned over the button so iOS
anchors the picker popover to it; the picked paths land on the prompt as unsent
text, so Compose adopts them together with anything already typed and Use
terminal keyboard hands a single-line draft back the other way.
Videos take a second path on the same route: ?kind=video streams the file to
disk under MAX_PASTE_VIDEO_BYTES (1GB, env CODEMAN_MAX_PASTE_VIDEO_BYTES)
instead of buffering it, sniffs the container from the first bytes as they
arrive, drains a mislabeled body unwritten and unlinks on any failure. It draws
on its own 3/min per-(IP, session) rate bucket beside the image path's 30/min:
at 1GB each, 30/min would be 30GB of disk a minute. Without the flag the route
is byte-identical to the image-only version, same 50MB cap and same bucket.
One client classifier, _promptAttachKind (MIME first, then the name's
extension), drives every pick and paste surface: the Attach button, terminal
drop, the tile grid's drop, the Compose and paste-dialog pickers and the
clipboard handlers. A pick that lands while the tile grid owns the terminal
goes to the PTY rather than the parked main-terminal overlay.
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.
Gesture control now moves tiles: with the tile grid open, pinch anywhere
on a tile and carry it onto another tile (they swap) or an empty cell (it
moves there); a session tab carried onto the grid joins it there. The
overlay asks tile-grid.js what is under the hand (tileAtPoint,
tileDropTargetAt) and drops through the mouse drag's own paths
(dropSessionOnTile / dropSessionOnSlot, so _reorderTiles stays the one
move path).
Detach Tiles (App Settings > Appearance, per device, OFF by default,
desktop only):
- a tile's header let go outside the browser window opens the session as
a pop-out (detachSession, which now takes an optional size and screen
position), the tile's size, near the drop point, after a 120 ms settle;
- dropped on another Codeman window's tiles it moves there: that window
docks it through the target's own drop and posts 'tile-adopted' on the
window channel, and only the window it was dragged from lets go of its
copy (a last tile goes to the welcome screen, never onto that session);
- a pop-out's title drags back onto any grid: the session is taken back,
the pop-out is asked to close and is released (stops answering
roll-calls; one a script cannot close says where the session went), and
its stale 'detached' answers are ignored for 2 s;
- a tile's menu offers "Open in a new window", and a hand that lets a tile
go outside the grid pops it out too.
With the setting off nothing changes. No server change.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A page reload puts a stored open grid back inside handleInit and fills a
cell freed since (its session deleted while the grid was closed) from the
ranking. Pending approvals, the source of the needs-input group, reach the
page only through seedApprovals, an async GET that lands after the restore,
so for that one fill a session waiting on a permission dialog or an unseen
finished turn ranks with the quiet ones (working sessions still rank first).
Holding the fill back until the seed lands would open fewer tiles, which
can be another shape, and then reshape the grid and move the user's tiles a
moment after the reload, so a late seed never re-forms a restored grid. The
plan, architecture-invariants#tile-grid and the wiki now say so, and a test
pins it: the seed arrives after the restore, the ranking then puts the
needs-input session first, and the restored cells and the stored layout
stay as they came back.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A session popped out while the grid was open (its pop-out button, or
another dashboard tab announcing the pop-out over the window channel) left
the grid through _markDetached, which removed its tile as if the user had
closed it: the stored count dropped to five and its cell stayed a hole the
ranking never filled, so after the window re-docked Tiles off and on showed
five tiles and an empty cell. Popping out with the grid closed, or a reload
reconcile, already kept the count and refilled the cell.
_markDetached now removes the tile as `gone` (it left by itself), the same
as a deleted session or a refused socket, so the count stays and the next
activation fills the cell from the ranking.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
After a session's first load per page, a tab switch fetched a 1 MiB tail of
the server's byte stream. For a fullscreen claude pane (alternate screen,
tmux `history_size` 0) that tail is old repaints of one frame, and an idle
tab parsed it twice (the cached copy, then the fresh one).
Every terminal response now records `paneHistoryLines` per session, and
`selectSession` takes `full=1` whenever the last report was 0. For such a
pane that capture is the visible frame, a few KB, which is also what a page
load already shows. A response without the field forgets the entry, so
unknown never counts as empty; a pane that starts keeping history goes back
to the tail on the next switch. Shells and the refresh path are unchanged.
Measured with real prod payloads through the real client in chromium:
median switch 1.26-1.66 s -> 0.10-0.19 s, parse ~1.1-1.5 s -> 2-3 ms.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The tile-grid plan records owner decision 11 (the layout is kept per
browser and comes back exactly; a fresh or filling grid ranks working,
then needing input, then most recent) and marks where it supersedes
decision 10's "the count wins over a remembered grid's size". The plan,
architecture-invariants#tile-grid and the CLAUDE.md Tile grid paragraph
now describe the stored format (`count`, freed cells, still v: 1), every
close path keeping it, and the open order; the wiki tells users what the
Tiles button brings back and how a new grid is chosen.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The "Showing the most recent 1.0 MB of this session. 4.8 MB more may still
be retained." bar covered the top of the terminal on nearly every tab switch,
and for fullscreen claude panes it could never be acted on: `truncated`
measures the server's byte stream, while Load full history asks tmux, which
holds no scrollback for a pane that lives in the alternate screen.
- The capture reads `#{history_size}` in the cursor query it already makes
and the terminal route returns it as `paneHistoryLines`.
- A reported 0 hides the notice; a positive count replaces the byte gap in
the message.
- The notice is lazy: it appears once a wheel/touch gesture reaches the top
of the browser buffer (after the pull that gesture starts), leaves when
the user scrolls back to live output or switches tabs, and its dismissal
sticks per session until reload.
- formatHistoryBytes no longer prints "1024 KB" just under 1 MiB.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Owner decision: searching for a tab and dragging it into a group is a useful
flow, so the drag-off guard is reverted and the docs and changeset say drag
works during a search.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
183efacc turned drag off while the rail search narrows the list, as the
owner asked, by refusing the flat manual rail's dragstart. That also
refused a drag that never reorders anything: with the tile grid open, a
found tab dragged onto a tile or an empty cell set no draggedTabId, so
_acceptTabDrops() (tile-grid.js) saw nothing and the drop was lost. On
the PR head that drop replaced or swapped the tile. The owner's reason
covers the reorder only: a reorder drop is saved for every device,
relative to rows the search hides, while the tile grid is per-device and
lands beside nothing hidden.
dragstart is unconditional again. The rows refuse the reorder instead:
their dragover returns before preventDefault while _tabRailSearchActive()
(the browser shows no-drop and fires no drop there), and their drop
returns before touching sessionOrder, for anything above the row that
might let a drop through. `draggable` stays on, so a cleared search
reorders with the same rows. The grouped rail's pointer drag keeps its
refusal in _onTabLayoutPointerDown(): it has no drop target besides the
rail.
Test: during a search that leaves both rows showing, dragstart is not
cancelled and sets draggedTabId, dragover on another row is not
accepted, a drop there leaves sessionOrder alone and saves nothing, the
real _acceptTabDrops() binder accepts the drag on a tile and hands it
the id, and after clearing the same rows reorder. It fails on the
previous commit, without the dragover guard, and without the drop guard.
CLAUDE.md, the invariants, the Dashboard wiki row and the release
changeset now say reordering by drag is off and a tile drop still works.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A floating subagent or ultracode window whose parent session's row the
rail search hid drew its connector from the viewport's top-left corner.
The hidden row is display:none, and getBoundingClientRect() still
answers it with an all-zero DOMRect, which is truthy, so the
`if (!tabRect) continue` paths never skipped it and _tabAnchor() put the
line at (0, 0) ("M 0 0 C 350 0, 350 400, 700 400" in Chromium). The
sidebar filter already had the same defect; #580 brought it to the rail,
and the keystroke redraw from 5f373a93 now ran it on every keystroke that
hides the parent.
A new _paintedSessionTab() in app.js returns the parent's row only while
getClientRects() is non-empty, and every place a floating window
measures its parent tab goes through it:
- the shared `tab:<id>` rect cache, filled by the subagent connectors
and by both ultracode connector layers (all three fill it, so guarding
only the first would let the next one cache the zero rect itself);
- the ultracode window's spawn position, which now cascades;
- the subagent window's fly-from-tab spawn, the same defect, not in the
report, now positioned as a window without a tab;
- the ultracode genie on minimize, which now tears down at once.
Lineage lines need nothing: they measure on a cache miss and computeTree
already drops zero-size rects.
Tests: in the gate, with the page laid out by hand (jsdom has no
layout), a subagent window, an ultracode run window and an ultracode
agent window from a hidden row draw no line and draw it again from the
row once the search is cleared, an ultracode window spawns from a
painted row and cascades from a hidden one, and the genie is skipped for
a hidden row. In the browser suite, real Chromium shows the hidden row's
rect is all zero and the connector is gone, then back after Clear. All
four fail on the previous commit. CLAUDE.md and the invariants'
Connectors bullet record the rule.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Closing a tab took ~0.55-0.7s on an idle machine, more with child
processes or recently active subagents. Most of it was waiting:
- The web UI kept the tab until DELETE returned, then removed it on the
100ms tab-render debounce, which every session update restarts.
closeSession() is now optimistic: the tab, tile and split go and the
next session is selected before the request is sent, rendered at
once. A refused delete (checked with a GET, since a delete can land
and lose its reply) puts the row back at its old index with the
error toast. This also fixes a latent bug: _apiDelete never throws,
so an HTTP error used to report "Session closed" while the session
kept running. SSE upserts skip ids that are being closed.
- The kill path slept fixed intervals (100ms PTY grace, 200ms for the
pane's children, 100ms for the process group) and verified in 100ms
steps. waitForProcessesExit() (utils/process-exit-wait.ts) keeps
every deadline but returns once the processes are gone, counting a
zombie as exited. Signal decisions keep kill(pid, 0).
- tmux kill-session and the pane-pid lookup ran via execSync, freezing
the server for ~70ms per close. Now async.
- killSubagentsForSession() ran a full `pgrep -f claude` scan per
active/idle subagent (~85ms each with ~100 matching processes). It
now scans once for all of them.
Measured on an isolated instance: click to tab gone 540-690ms -> 58-95ms;
DELETE of a claude session ~450ms -> ~200-260ms.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
CLAUDE.md gained a rule for the rail's Search sessions box and linked it
to architecture-invariants#session-list-layout-header-strip-vs-left-sidebar,
which said nothing about it, and the "Collapse is per-device" bullet under
Owner tab layouts had an exception it did not record.
- Session list layout: one paragraph on the shared _applyTabListFilter()
over the pure CodemanTabSearch (classes only, layout-scoped hide rules,
rail matches the name and the sidebar name + folder, locale-independent
lower-casing), the alert-row keep (owner decision), the data-total count
restore, the tree walk and roving-stop fix-up, the projection opening
every group, the connector redraw, the global Escape claim, no drag
while searching, and the reset off the rail.
- Owner tab layouts: the collapse bullet notes that a search draws every
group open and refuses toggles without writing the stored set.
- CLAUDE.md: the Escape claim and the connector redraw as one clause on
the existing rail search rule.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Conflicts resolved against the release branch: docs/api-reference.md keeps
both new sections (codeman agent CLI, then Prompt uploads);
test/test-ports-guard.test.ts takes the release side (#570 already removed
the legacy list); test/paste-image-dir-shared.test.ts keeps #570's ephemeral
port and this PR's bounded-path-probe mock.
Also at merge: the Docker sentence in the route comment and the API reference
is narrowed to owned cases, since an adopted container mounts nothing.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Review fixes for "place new, child and web tabs into groups".
- Spawned tabs follow their parent only with the new synced setting
spawnedTabsFollowParent (App Settings > Appearance > Tabs, off by default).
TabLayoutService.sessionCreated() reads it fresh through a
childrenFollowParent dep: off, a spawned session lands after its parent and
is marked manual in the same write; on, it follows until moved by hand. The
row menu shows the placement note and "Follow parent again" only when on.
"New session" in a group menu is unaffected. zh-CN label and a
Settings-Reference row.
- applyLegacySessionRank marks a child manual only when the requested order
splits it from its parent's block. 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 PUT /api/session-order,
with or without groups: the layout service already re-projects the order and
broadcasts it, and syncSessionOrder adopts it.
- The placement note in the row menu is plain text described by the menu, not
a focusable menu item.
- A tab that first appears inside a collapsed group briefly marks that group's
header.
- i18n skips .tab-name, so tab names are never translated.
- Tests: legacy-order cases for an unchanged order, a family reorder and a
split; service cases for the setting and for order PUTs; the scheduled
registration test now exercises register/undo; the grouped-create test binds
port 0; one route case runs through the real layout service.
Review follow-up after the rebase onto 1.40.0, where tiles began paging a
hollow buffer through the shared CodemanTerminalInput helpers and opencode
joined the hollow check (#555).
- opencode declares transcriptPageKeys in stock.ts, so the paging list read
by both panes is claude, codex and opencode. Without it, reading the
capability instead of the old id check would have turned opencode paging
off again. The stock-set and render tests name it.
- The gesture rules (first trackpad-sized event pages at once and is owed
back, wheel notches accumulate, a pre-paid debt dies with its gesture,
pinch and mostly horizontal swipes are consumed without paging) move out
of the primary pane into the pure CodemanTerminalInput.pageKeysForGesture.
The primary pane and TerminalTile both call it with their own state, so a
Codex or opencode tile pages exactly like the main pane.
- The dialog check stays per pane: the primary pane reads tabAlerts for the
active session, a tile for its own sessionId, so a wheel over a tile never
pages into that session's permission prompt while the focused tab is clear,
and a tile is not blocked by a dialog in another tab.
Tests: pure helper cases in terminal-scroll-routing, tile cases for Codex
paging, first-event paging, notch accumulation, pinch and swipe, byte parity
with the primary pane over a mixed sequence, and the per-session dialog
check in both directions.
A grid tile or the split's Pane B (TerminalTile) left the mouse wheel to
xterm for a session running Claude's fullscreen renderer (claude 2.1.187+
with mouse tracking on, cliMouseTracking). That renderer scrolls its own
transcript on SGR wheel reports, which the primary pane sends it, while the
tile's xterm holds only Codeman's replayed repaint frames (tmux keeps no
history for such a pane). So in a grid of fullscreen Claude sessions the
wheel either scrolled nothing or dragged stale frames, Claude's pinned input
box with them, up the tile, and Claude's transcript never moved. This was
tile-grid-plan follow-up 4.
The tile now forwards the wheel the way the primary pane does
(TerminalTile._maybeForwardWheelToCli): the primary pane's own gate,
_shouldForwardWheelToApp(ev, target), asked for the tile's terminal and
session; the cell from _clientPointToCell(x, y, tile.terminal); a scrolled-up
viewport snapped to the live screen first; and the reports flushed through
the tile's own 40 ms coalescer to the tile's session (the primary queue
flushes to the active session). The encoding moved into pure helpers in
terminal-ui.js, CodemanTerminalInput.wheelDeltaWholeLines and
sgrWheelReports, which the primary pane's _wheelScrollLines and
_sendSyntheticSgrWheel now call too, so the two panes send identical bytes.
Shift+wheel, the explicit local-scrollback gesture, was dead in every tile
off macOS: Chrome on Windows delivers it as a horizontal wheel (deltaX), and
xterm's own scroller turns a Shift+vertical wheel into a horizontal one. The
tile now scrolls it itself (_maybeScrollLocalOnShift: scrollLines() on the
dominant axis, sub-line travel carried over, a shell tile's history pull
still asked on the way up), as the primary pane's capture-phase handler does.
Unchanged: inline Claude, opencode and older Claude still take the
PageUp/PageDown route (#555), shells and other modes keep xterm's own plain
wheel, and a tracking xterm or the alternate buffer stays xterm's.
Tests: test/terminal-tile-scroll.test.ts covers forwarding (geometry, tick
cap, coalescing, viewport snap, the tile's session rather than the active
one, Shift/tracking/alternate exclusions, byte equality with the primary
pane) and Shift+wheel (Windows deltaX shape, sub-line carry, shell history
pull). test/terminal-tile-scroll.browser.test.ts adds a real-Chromium case
with trusted page.mouse.wheel() events.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Review follow-up for the hollow-pane PageUp/PageDown fallback
(_maybePageCliTranscript):
- A mostly horizontal wheel swipe (both deltas present, |deltaX| > |deltaY|)
is consumed without paging. The touch path already locks its axis.
- A pinch (wheel events with ctrlKey, as Chrome reports one) is consumed
without paging.
- Only a trackpad-sized opening event (under 2 rows) pages at once. A wheel
notch (100 px, 4 rows) accumulates toward half a screen as before, so five
slow notches send one PageUp again instead of five. A debt from a pre-paid
page still dies with its gesture; plain wheel travel carries across notches.
- Nothing pages while the active tab's alert is 'action' (a pending
permission_prompt or elicitation_dialog), so a page key never reaches an
open dialog's selector.
Docs: the invariants paragraph now says which Codex versions were seen
drawing on the alt screen (0.157.1, 0.160) and that 0.147.0 still draws
inline, without claiming a cutoff.
Two changes to the last-resort page-key scroll (_maybePageCliTranscript), the
path a wheel or touch gesture takes when it is not forwarded to the CLI and
the local xterm buffer has no history (normal buffer, baseY 0).
Codex now takes that path too. Newer codex draws on tmux's alternate screen,
which the full strip hides from xterm, so a codex pane is as hollow as a
repaint-mode Claude pane and the wheel did nothing. Codex still ignores SGR
wheel reports and is still never forwarded the wheel; this sends plain
PageUp/PageDown keys, which codex does page its transcript on (measured on
0.157.1 locally and 0.160 over a remote session). Which modes may page is a
new registry capability, capabilities.transcriptPageKeys (claude and codex in
stock.ts), published to the page as window.__codemanTranscriptPageKeys, so
terminal-ui.js holds no CLI id for it.
The first event of a gesture now pages at once. The fallback used to send a
key only after half a screen of accumulated travel (19 rows, about 475px on a
38-row pane), so an ordinary trackpad flick sent nothing. The skipped travel is
owed back by the rest of the gesture, keeping one page per half screen; a pause
of more than 150 ms, a direction change or a tab switch starts a new gesture.
New sessions, spawned child sessions and saved web tabs now land in tab
groups instead of always starting in Ungrouped.
New session in a group:
- The group menu gains "New session" (first item). It expands a collapsed
group and runs ONE launch with the current case and run mode; a launch
already in flight is reported, not queued.
- run({ count, tabGroupId }) threads the group through every launch path
(local POST /api/sessions, remote/docker and external-CLI
/api/quick-start). The count overrides the steppers for that launch only.
- Server: POST /api/sessions and /api/quick-start accept an optional
tabGroupId (Zod, 1-100 chars; null/empty is a 400). The routes pass it to
addSession(), and TabLayoutService.sessionCreated() places the session at
the end of that group in the SAME versioned write that first records it.
The session is added to the live map inside the owner lock (with a
rollback when 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 instead of failing the spawn. An
explicit group is a hand placement: a child placed this way is manual.
- The create response carries the committed tabLayout (only when tabGroupId
was sent), adopted before the new tab is drawn. Responses without the
field are unchanged.
Child sessions (parentSessionId / X-Codeman-Parent-Session):
- A child nobody placed follows its parent: the server's existing
normalization already puts it after the parent's rows; the browser now
draws a not-yet-stored child there too (project({ parents })) and, while
groups exist, adopts the server's order instead of echoing its own append
back through PUT /api/session-order (which pinned every child manual).
- The row menu says where a child sits (follows, placed by hand, parent
closed, parent loop) and offers "Follow parent again" only to a hand-placed
child. A child whose parent is gone stands on its own and is never
re-adopted.
Saved web tabs:
- Web tabs are layout refs like sessions; openness stays per device. A
grouped web tab shows (and takes an Alt+N slot) only while open here and
reopens where the layout keeps it. Pinned by tests; no behaviour change.
Unchanged: the flat rail and header strip (markup, HTML5 drag, append and
save on create), the tree semantics and single tab stop, sessionOrder and
Alt+N, and every request that does not send tabGroupId.
The hourly sweep called lstatSync and realpathSync on every live session's working directory, so a linked case on a mount that stopped answering blocked the event loop 30 s after boot and then every hour; the old sweep was fully async. uploadDirs() now probes the workspace with probePathKind() first and skips an unknown path without touching it, then uses fs.promises for the lstat and realpath. The sweep keeps the probe's stall cap; cleanupSession(), acting on one path at the user's request, passes pastCap and removes the directories with fs.rm.
Refusing an upload dir that sits strictly inside the data dir protected nothing (it only ever holds uploads, and a link is already excluded by lstat) while the route kept writing there, so uploads under a workspace like the ~/.codeman/app that install.sh clones were never swept and never removed. Only the two cases that matter stay refused: the upload dir being the data dir, or containing it.
Also: the ignore file's take-back after a failed write no longer replaces the error that caused it with its own, and the shared-dir test's header no longer names the fixed port it stopped using.
A pasted image landed in <workspace>/.claude-images/, a name that belongs to another tool, in a folder nothing ignored, so it showed up in git status of every case that ever received a paste. Uploads now go to a flat <workspace>/.codeman-uploads/ that carries a .gitignore of `*` (written once, never over a file already there): in the workspace because that is the only path identical for a local agent and a container, flat because a nested .codeman/ is the data dir itself when the workspace is the home directory.
The directory names live in paste-image-gc.ts alone. The old folder receives nothing but stays readable for one release: the hourly sweep and the delete cleanup go through uploadDirs(), the image watcher's ignore filter reads the names. uploadDirs() lists only real directories, none that is, contains or sits inside the data dir, and nothing for a remote session, since both consumers delete. The route refuses a remote (SSH) session before touching disk: the file would land on this host under the remote path, where the agent cannot read it.
test/paste-image-dir-shared.test.ts binds port 0 and leaves the port guard's legacy list.
Decided in #553.