Conflict resolution against the five PRs that landed while this was in review, plus
the items left for merge on the thread.
The real one was `session-ui.js`. #454 refactored all eight non-Claude `run*()`
functions to funnel through one `_launchQuickStartInstances()` helper that does the
POST itself, while this PR replaced that same POST in each of them with
`_quickStartWithCustomModelConfirm()`. Resolved in the helper rather than seven times
over: the helper now goes through the confirm path, and each body builder carries the
`customModel` spread. `runAntigravity` deliberately does NOT, since antigravity's
`customModelInjection` is `unsupported`; parity with this PR's own per-mode choices is
asserted rather than assumed.
That merge creates a question neither feature had alone: the confirm dialog now runs
inside a loop that can launch up to 20 instances. Both questions it can ask (context
window too small, and loading this will unload the model another session is using) are
decisions about the ENDPOINT, and every instance in a batch targets the same one, so
the answer is taken once and carried to the rest. Without that a 20-instance launch
asks the same question 20 times.
Also: `sse-events.ts` is 161 constants (master added two for remote wake, this adds
one, verified by counting rather than by arithmetic), `server.ts` keeps both new SSE
prefixes, the two comments pointing at code that no longer exists are corrected, and
CLAUDE.md's SSE and route counts move to 161 / ~236 / custom-model (6).
`pumpLlamaSwapLogTail`'s unparsed remainder is now capped at 64 KiB. It only shrank at
a `\n\n` frame boundary, so a backend that streams without one would grow it for the
life of a deliberately indefinite connection.
NOT changed, deliberately: the context warning and the swap-conflict warning still
share one `confirmed` flag with the context check first, so confirming "launch anyway"
on a too-small context also skips the "this unloads it for another session" ask. That
is the author's documented choice and the reviewer's own note calls it minor. Both
fixes are worse to make here than to defer: separate flags are new wire surface landed
unreviewed during a release, and reordering the checks adds a network round trip to a
path that currently short-circuits. Raised as a follow-up instead.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The MAC-count limit lived in two places that disagreed. RemoteHostSchema.wakeMac's
128-character cap admits seven comma-separated MACs while parseMacList takes at most
four, all-or-nothing, so a five-MAC value validated, was written to remote-hosts.json,
and then resolved to NO wake target: POST /api/sessions/:id/wake answered
"No wake-on-LAN target configured for this host" and the banner offered "Configure WoL"
for a host the user had just configured. MAX_WAKE_MACS now lives in
src/config/remote-wake-limits.ts and both sides refine against it. Its own module
because src/remote-wake.ts is import-fenced to session-routes.ts and server.ts (the
wiring guard that stops a watcher waking a host), and because schemas.ts must not drag
dgram/net/child_process into every request-validating module.
The documented 40 s request budget also omitted the wake's own cost. A `command` target
is bounded by REMOTE_WAKE_COMMAND_TIMEOUT_MS and runs BEFORE the readiness poll, so a
slow one pushed a wakeCommand host's worst case to ~68 s, past the 60 s
proxy_read_timeout the budget exists to stay under. _wakeAndWait now subtracts the
wake's measured elapsed time from the readiness budget, floored at one poll interval so
a wake that ate the whole budget still gets one probe. A magic packet is effectively
instant and is unaffected, which is why live testing never saw it.
Also: the two new endpoints are documented in docs/api-reference.md with the import
fence stated as the rule it is, CLAUDE.md's frontend module count moves to 34, and the
release changesets carry the Thanks section.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
#451 cleaned two things on copy. The trailing trim is right and every native
terminal does it. The shared leading-indent strip is this project's own rule,
and it is dropped here rather than shipped.
Measured against the shipped transform over 401,445 three-row windows across
1,010 tracked files in this repo, it fired on 73% of them: 92% inside a YAML
workflow, 76% over `git log` output, 48% in a TypeScript source. No width
threshold separates a margin from content because they are the same widths, a
live Claude Code pane's own margins measuring 2 and 5 columns while the most
common non-TUI shared run is 4. The failure modes are not symmetric either: a
wrong trailing trim costs nothing, while a wrong dedent silently deletes
information that was on the screen, with nothing in the clipboard to hint at
it, on git log bodies, on indented code read out of cat (semantic in Python),
on git diff context rows where the leading space is the marker, and on stack
traces.
It also could not be made self-consistent cheaply. Whether the first row joined
the measurement depended on the mousedown COLUMN, which the user never sees, so
one block of three rows produced three different clipboard results; and the
flag read getSelectionPosition().start, which is xterm's mousedown anchor and
is never normalised, so dragging UP through a block read it off the bottom row.
The PR's test stub hardcoded a downward drag, so its suite could not express
that case.
The transform, the wiring, the tests, the invariants, CLAUDE.md, the wiki page
and the changeset all move together. The test block now pins the ABSENCE as a
contract, with the git log, Python and git diff cases as its examples, so this
is not re-derived later. If it is ever revisited, the one qualification that
measured clean is painted trailing padding: zero false positives over all
401,445 windows.
Also from the review: the comments and invariant rule justifying the
padding-only clear described the pre-change code (the Ctrl+C gate reads the
CLEANED selection now, so such a selection falls through to the PTY on its own
and the clear is feedback rather than protection), the new 'Nothing to copy'
toast gained its zh-CN entry, and the invariants paragraph no longer repeats
its own opening sentence.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
#454: the behaviour the PR adds had no test, so a regression test drives
runGrok() at tabCount 3 and asserts three quick-start POSTs with sequential
w<n>-<case> names (verified to fail against master's session-ui.js). Each
caller now reads the count BEFORE its opening banner and announces it there,
the way runClaude() already did, so a launch no longer prints two headers and
a launch with another session already active still says how many are starting.
runClaude() calls the shared _readTabCount() instead of its own copy of the
1..20 clamp, and that helper optional-chains the element read, since hoisting
it above each caller's try block would otherwise let a missing #tabCount throw
where the launch-error path cannot report it.
#435: sizeMovedUnderLoad derived from data.source alone. `mux-visible` is not
sufficient: a failed display-message cursor query makes capturePaneBuffer skip
the snapshot repaint and return the raw capture, which the route still labels
mux-visible, so a size that moved during such a load bought a full forced
reload to repair a frame that was never positioned. It now tests
Number.isFinite(data.captureRows) like its two siblings.
Plus the invariants and CLAUDE.md lines promised on #435: a visible capture
reports its geometry and omits it when nothing was positioned, the comparison
runs on mux-visible only, and the replay is capped at one attempt and latches
per session when it cannot converge.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude Code 2.1.277 takes typed text the moment its composer paints but
ignores Enter for the first 30 to 50 seconds after it (measured 2026-09-19
through the input route: an Enter at 28 s stranded the prompt, one at 51 s
submitted it). The text+Enter pair `sendInput` sends 50 ms apart therefore
left every programmatic prompt sitting unsent, and every waiter burned its
timeout on a turn that never started.
Server: `SubmitVerifier` (session-submit-verifier.ts), armed from
`writeViaMux` for every mux write that carried a carriage return, reads the
pane on a 2 s to 60 s schedule and re-sends Enter only while the last
composer line (the CLI's own prompt glyph) still holds the head of what was
sent. An empty composer, other text, or no composer line at all ends it; a
newer write replaces the schedule.
Skill: `sendwait` gets the same loop (`_composer_text`, no-break space
stripped by its bytes for BSD sed) for servers that predate this, and the
preamble version moves to 1.30.1 so seeded agents pick up the fresh copy.
SKILL.md's heredoc and the plugin mirror are regenerated.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Three follow-ups to the source gate, each one measured rather than reasoned.
A pane already drawing at the size the client just requested is left alone. The
replay runs at `dimsAfterLoad`, so it can only change what is on screen if the
pane was drawing at some other size; when the reported geometry already IS that
size, the second pass captures the identical frame and pays a full reload to do
it, including a visible re-flash, a dropped and reopened WebSocket and a deleted
xterm snapshot. That equality is the signature of a clamp rather than a race:
`getTerminalDimensions()` floors at 40x10 while `fitAddon.fit()` does not, so a
terminal narrower than 40 columns or shorter than 10 rows reports a pane
permanently bigger than itself and replayed on every tab switch without ever
converging. A race never produces the equality, since its premise is that the
pane was still at the size it was asked to leave. The declined-resize case does
not produce it either, so that one still costs the single capped attempt and
needs the pane-ownership question this does not touch.
The full-history re-arm is unreachable and now says so. A pass that consumed the
flag sent `full=1`, and the route answers `full=1` with `mux-full-history` or
`history`, never `mux-visible`, so the source gate already rules out every such
pass. The line stays for the invariant, but its comment no longer reads as if a
page load retries, and the suite pins that it does not.
The response no longer reports geometry for a body that carries no capture. The
full-history path writes `capturedGeometry` from the cursor query and then
returns '' for a pane holding nothing visible, which drops the source to
`history` with the geometry already recorded: a `full=1` request whose capture
reported 100x50 and returned nothing answered `source: "history"` with both
fields set. Nothing acted on it, because the client ignores geometry on any
other source, but the field said a frame had been drawn at a size when none had.
The browser stub now derives `source` from the request the way the route does,
rather than answering `full=1` with `mux-visible`, which the route cannot
produce. Each case reaches a visible-frame response the way production does, by
not being the first select of the page. Three cases pin the new behaviour and
each fails without its guard: the clamp case sees two fetches instead of one,
the scope case and the full-history case both see a replay the gate forbids, and
the width case sees one fetch instead of two.
The changeset now describes the change from 1.29.x rather than the difference
between the two commits on this branch.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Only a visible-frame capture positions its rows absolutely, so only that frame
can be damaged by a terminal of the wrong size. A `full=1` body is linear
scrollback closed by a relative cursor move, which is relative precisely so the
browser's row count need not match the pane's, and a `history` body is the byte
stream, which carries no row alignment to protect. The geometry comparison ran
on all three, so it fired most often on the one response it cannot help:
`_fullHistoryLoaded` is empty on the first select of every non-shell session per
page, and a session whose pane a desktop tab holds too tall to ever fit then
paid a second whole-scrollback capture, reset and replay on every page load and
every first tab switch.
`framePositionsRowsAbsolutely` gates both the captured-geometry comparison and
`sizeMovedUnderLoad`. A size that moved under a byte-stream or scrollback replay
is healed by xterm's own reflow plus the SIGWINCH the trailing `sendResize`
already sends.
A pane WIDER than the terminal damages the same frame a second way, so
`captureCols` is now compared rather than only logged. `formatPaneSnapshot`
paints each row out to the pane's own width, so a narrower browser wraps every
painted row, and the wrap on the last one scrolls the whole frame up by a row.
The terminal response no longer falls back to `session.ptyCols`/`ptyRows` when
the capture reported no geometry. The cursor query is what produces the absolute
addressing in the first place, so a capture that lost it returned a raw frame
that was never positioned, and a byte-history response was never positioned
either. Naming the session's own PTY size there described a frame that does not
exist and invited a repair for damage that is not present. `_ptyCols` is also
written only by `resize()` while the PTY is spawned at the size queried from
tmux, so it can be wrong on its own terms. Both fields are now absent instead,
and the `Session` getters added for that fallback go with it.
Two browser cases cover the new behaviour and each fails without its fix: a
`mux-full-history` response with both dimensions mismatched asserts one fetch
(two without the gate), and a `mux-visible` response wider than the terminal
but short enough to fit asserts two (one without the width comparison).
Corrects a claim in the comment above `capturedGeometry` in tmux-manager.ts.
Both replay paths do not address rows absolutely; the full-history one ends in a
relative move, which is the whole reason the gate is right.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A visible-frame capture repaints each row at an absolute position, counting up
to the pane's height. A terminal shorter than that clamps every address past its
own height onto its last line. The overflow rows then overwrite one another, and
the rows underneath are lost. Replaying a real 50-row capture into a 30-row
terminal rendered 28 lines of a 45-line command and drew the frame twice.
Nothing in the response said what height the frame was built for, so the client
could not detect this. A capture now reports the geometry it was really taken at
through `capturedGeometry` on `PaneCaptureOptions`, and the terminal response
carries it as `captureCols` and `captureRows`. When the captured pane is taller
than the terminal, or the size that produced the capture did not survive the
load, `selectSession` replays once at the size that stuck. `resizeRetry` caps
that at one attempt, so two competing fits cannot trade replays forever.
The retry re-arms the full-history flag only when the pass that ran had consumed
it. A tab switch takes the bounded tail, so its retry takes the tail too:
clearing the flag unconditionally would upgrade that switch into a fresh
scrollback capture the user never asked for, which the route's own comments put
at tens of megabytes.
What this repairs is a capture that won a race against the resize meant to
precede it. It does not repair a capture whose pane was too tall because
`Session.resize` declined the resize outright, which it does for a small
viewport while a desktop viewport's size claim is live. The retry re-sends the
same declined resize and captures the same pane, and `resizeRetry` then stops
it. Repairing that means changing who owns the pane size, which is a policy
question this does not touch. The reported geometry still helps there, because
the client can see the mismatch at all rather than being blind to it.
Follows #395, #396 and #397, which fixed the other ways the replayed frame and
the terminal could disagree.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The four edits the review said would be folded in at merge, none of them
code: the changeset becomes one user-facing paragraph, since it is what
CHANGELOG.md and the release notes print; the `_bufferLoadFinishOpts` comment
now names the second contributor to the duplicate window (`captureActivePaneBuffer`
is `execSync`, so anything painted into the pane before the server read it is
in the capture and is broadcast after the reply) and says why a `history`
payload keeps the pre-existing discard when its exposure is the same; the
`_finishBufferLoad` doc block moves from above `_beginBufferLoad` onto the
function it documents; and the test file's header describes both rules the
file now pins instead of only COD-144.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
xterm hands back whole screen rows and trims only the cells that were
never written to, so the real spaces a full-screen TUI paints across the
unused part of a row count as content and reach the clipboard. Measured
against Claude Code in a 282-column pane, single lines arrived carrying
138 trailing spaces, and every line carried the two-space transcript
indent as well. Windows Terminal, iTerm2 and GNOME Terminal all trim that
for you, decideAutoCopy already calls a wall of spaces "never what the
gesture meant", and _selectTouchSelectionLine already treats those cells
as padding — the mouse and keyboard paths never had the same rule.
CodemanCopySelection.clean lives in constants.js beside decideAutoCopy,
its pure sibling. It drops the trailing run from each line, and removes
the leading run only where every selected row shares one. A selection of
a single row keeps its run, because one row shares nothing with anything
and stripping it would silently reindent one line of `git log` body text
or one line out of `less`. A drag that began inside a row keeps its
partial first line untouched and out of the measurement, which otherwise
pins the shared run to zero and leaves every following row indented.
Every pass over a line is a scan rather than a regex. `/[ \t]+(\r?)$/` is
quadratic on a line whose spaces are followed by a non-space character,
which is what right-aligned or centred TUI content looks like: measured
over 50 000 rows with a 280-column run it took 2.9s, against 1.3ms for
the scan, and a 2 000-column run took 16s. The scan is also the faster of
the two on an ordinary padded row.
cleanedTerminalSelection in terminal-ui.js is the half that needs the
live terminal. It returns a COLUMN selection untouched: Alt+drag makes
one, and a rectangle's rows lining up is the point of the gesture, so
both halves of the clean would destroy it. xterm exposes the mode nowhere
public, so the check reads terminal._core._selectionService, the way this
file already reads terminal._core for cell dimensions, and cleans
normally if a future xterm renames the field. A test pins that assumption
against the library rather than against a stub repeating the literal.
The Ctrl+C chord decides on the cleaned selection, not the raw one. A
drag across the blank part of a row selects real padding spaces, so the
raw text is truthy, and testing it would spend that press on a copy of
nothing and make the user press again to interrupt. A padding-only
selection is now dropped and the press falls through to the PTY, while
Ctrl+Shift+C still never falls through. copyTerminalSelection gates on
trim() for the same reason, since a multi-row drag across padding cleans
to line breaks alone and a bare newline pasted into a chat composer
submits it.
All four of the main terminal's copy paths go through it: the Ctrl+C
chord, right-click, the phone selection button and Auto Copy. The
browser's own Edit menu copy, a disabled copy shortcut and the subagent
windows still copy raw rows, as they did before, and the invariants doc
now says so rather than claiming every copy is cleaned. Auto Copy
resolves its own toggle before it reads the selection, since it is off by
default and a selection can run to the 50 000-row scrollback ceiling.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A capture load now replays its queued tail, and that replay runs through
`batchTerminalWrite`, which samples `_wasAtBottomBeforeWrite` before it queues.
It runs inside `chunkedTerminalWrite`, before that promise resolves, with the
terminal freshly reset and rewritten — so the sample is always true. The caller
then restored the reader's position and the next `flushPendingWrites` scrolled
straight back to the bottom off the latched flag, undoing it. The only thing in
the way was `_hasRecentUserScrollUp()`, a 1500ms window a server-triggered
refresh is usually past.
`_syncStickyScrollBaseline()` re-takes the flag from wherever the viewport now
sits, and the two paths that restore a position call it right after doing so:
`_onSessionNeedsRefresh` and `_maybeRefetchFullHistory`. Those are the paths
#259 and #205 exist for, and they are also where a non-empty queue is most
likely, since a needsRefresh fires when output is flooding. Re-taking rather
than suppressing the sampling: suppressing leaves whatever stale value the flag
held from before the load, which on the full-history re-pull has no reason to
be false. `selectSession` and `_onSessionClearTerminal` deliberately end at the
bottom, so the sampled true is already the truth there and they do not call it.
`_bufferLoadFinishOpts` gains the coverage the CI gate can see: both mux
sources flush, `history` does not, and a payload naming no source does not.
Its only coverage was the browser suite, which CI does not run.
The JSDoc and the changeset now record the one duplicate window this cutoff
cannot close. The server appends output to the byte buffer in the same tick it
emits, but broadcasts on a batch timer — 8ms over WebSocket, 16 to 50ms over
SSE — so a batch pending when `capture-pane` ran leaves the server after the
reply and is replayed although the capture holds it. It is one batch interval
wide against a recovery window spanning the whole chunked write, and closing it
means flushing that batch server side before the capture.
The second browser test asserts its session was created, so a failed create
fails it instead of passing with zero hits.
docs/architecture-invariants.md no longer claims the replay leaves the
queued-event discard window alone. That clause now describes what decides how a
load ends, the baseline rule, the batch window, and the three covering tests.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Blocker: .center-status-banner never actually disappears.
- Add `.center-status-banner[hidden] { display: none; }`, same trap as
`.home-sessions[hidden]`: the author-level `display: flex` beat the
UA `[hidden]` rule, so `dismiss()` set `el.hidden = true` and the
card stayed laid out at `opacity: 0` with its text/cancel/close
children still `pointer-events: auto` -- an invisible 442x67 click
blocker dead centre over the terminal until the page reloaded.
- Added a regression test pinning the CSS rule, and documented the
banner (10001) and the swap-confirm/context-warning modals (10010)
in CLAUDE.md's Z-index layers list.
Stale wording pointed at the reverted sticky-toast default:
- .changeset/run-menu-custom-model-picker.md, CLAUDE.md, and the
`.toast-message` comment in styles.css all still said "toasts
default to sticky" after 1f32128c put the flat 3s default back.
Reworded all three to describe the actual behaviour: one call site
passes an explicit `duration: 0`.
Smaller items from the same review:
- docs/api-reference.md said discovery failures answer
`502 OPERATION_FAILED`; OPERATION_FAILED is 422 per src/types/api.ts
and the error-code table earlier in the same file.
- The periodic re-discovery sweep (server.ts) never read
customModelEndpointsEnabled, so turning the feature off left
Codeman polling every saved endpoint forever. Added
readCustomModelEndpointsEnabled() (custom-model-routes.ts, same
shape as readPlanUsageTelemetryEnabled) and gated the interval
callback on it.
- Reverted the formatting-only Prettier pass docs/api-reference.md
picked up (table padding, *x* to _x_, JSON re-indent) by re-merging
the new Custom Model Endpoints section onto the pre-PR file, so the
diff is reviewable. No prose content was lost -- verified by diffing
the result against the pre-revert file (formatting-only) and against
the merge-base file (only the new section added).
- docs/custom-model-endpoints.md now states that a custom-model Claude
session's isolated CLAUDE_CONFIG_DIR loses the user's global
settings.json, user-level skills/agents/commands, and MCP servers
from ~/.claude.json -- only `projects` is symlinked back.
Design question left open in the review (does `confirmed: true` need
to be two flags so "launch anyway" on the context warning doesn't also
skip the llama-swap displacement warning): keeping the single flag, as
offered. The 20s displacement sweep still catches a resulting swap
after the fact, so it's a surprise rather than a silent failure, and
splitting it is real behavioural surface I have no way to verify live
in this environment.
`npm run test:browser` could not be run in this environment (no tmux,
no downloaded Playwright browser binary) -- none of its suite's files
touch code this fix changes, but it still needs a real pass before
merge, same as any frontend change.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ea59JhUmHBm1gRCsiYF33R
* chore: version packages
* chore: sync the CLAUDE.md version line to 1.30.0
The changesets bot does not touch this line, and pushing it to master
after merging the version PR starts a second Release run that has raced
the first before. Riding the bot's own branch keeps it to one push.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Codeman maintainer <noreply@anthropic.com>
Changeset text becomes user-facing CHANGELOG, so the #429 entry is cut
from five bullets of internal bash-array detail down to what the change
does for someone running the installer, as promised on the PR. The #441
entry loses its em-dashes, which are not house style. Adds an entry for
the maintainer fixes applied while landing #442, and the Thanks section.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Replaces the size-scaled expected-time estimate + matching auto-timeout
with a generic hardware/model-size disclaimer and a user-driven Cancel
button, per explicit request. Real load time depends on hardware this
feature has no way to know (VRAM, storage speed, GPU contention), so
the old estimate/timeout was a guess dressed up as a fact — worse, one
that could kill a genuinely slow load partway through on slower
hardware.
- _watchLlamaSwapLoading (session-ui.js): dropped maxWaitMs/deadline
entirely — polls indefinitely until ready or cancelled, no automatic
give-up. Message is now "Loading <model> (<size>) on <endpoint> —
this can take a while depending on your hardware and the model
size.", with the real llama.cpp log line still on its own second
line. Removed _MODEL_LOAD_TIME_MATRIX/_estimateModelLoad/
_formatRemaining (dead code once the countdown is gone) —
_lookupModelSizeGB is kept, the GB figure still shows.
- _showCenterStatus (panels-ui.js) gains opts.onCancel: renders a real
"Cancel" button (distinct from the error-type "×" close button,
since Cancel has a real consequence) that calls it on click. Caller
owns what cancelling actually means, same split as the swap-confirm
modal's promise-resolving buttons.
- Cancelling dismisses the banner, shows an info toast (not an error —
this was deliberate), and closes the session, mirroring what the old
timeout used to do automatically but now on the user's own call.
- New .center-status-cancel CSS (bordered pill button, distinct from
the plain "×" close glyph).
Test changes: removed the now-invalid timeout-auto-close/estimate
tests, added cancel-flow tests (dismiss/toast-type/session-close,
never-closes-with-no-sessionId, unbounded-polling), and real-DOM tests
for the new Cancel button (bootAppWithRealCenterStatus, evaluating
panels-ui.js instead of stubbing _showCenterStatus, since this button
is worth verifying for real rather than just through the stub every
other test in the file uses). Typecheck/lint/frontend-syntax clean;
full suite shows no new regressions.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RqZeHrRS6DYcGcGX2p9EwG
Answers the underlying request behind investigating llama.cpp log
access: surface what the backend is actually doing, live, on top of
the existing countdown timer during a model load.
- getLatestLlamaSwapLogLine()/pruneIdleLlamaSwapLogTails()
(custom-model-routes.ts): one persistent GET /api/events (SSE)
connection held open per endpoint, parsing logData frames and
keeping the latest source:"upstream" (backend llama-server) line —
filtering out llama-swap's own source:"proxy" request-access lines.
Idle-closed after 30s of no polling, same 20s sweep as the existing
swap-displacement check.
- running-status route now returns logLine alongside the existing
isLlamaSwap/running fields.
- Frontend: _watchLlamaSwapLoading's banner gains a second line
("llama.cpp: <line>", bootlog timestamp/level/component prefix
stripped for display) that stays on the last real thing llama.cpp
said rather than clearing to blank between polls.
⚠️ Caught and fixed before merge, not after: the first cut targeted
GET /logs (the endpoint the name suggests), shipped a working-looking
implementation with passing tests, and only failed a live check against
the real Nemesis llama-swap deployment — /logs turns out to carry ONLY
llama-swap's own proxy request-access log and never once showed a
single backend line, even seconds after a real, confirmed model swap
triggered via a direct API call. GET /api/events's logData frames
(with an explicit source field distinguishing upstream from proxy) are
the only source that actually has backend output; corrected and
re-verified live end-to-end through an actual forced swap before
writing this commit, confirmed live to hold its connection open
indefinitely (unlike /logs, which closes after a fixed ~100KB).
12 tests for the corrected /api/events parsing (SSE frame buffering
across chunk boundaries, source filtering, malformed/wrong-type frames,
connection reuse, idle pruning) plus 2 for the frontend banner
rendering. Typecheck/lint/frontend-syntax clean; full suite shows no
new regressions (14 more passing than baseline, matching the new
tests; same pre-existing Windows-environment failures).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RqZeHrRS6DYcGcGX2p9EwG
The llama-swap conflict check on the apply/create routes only ever runs
at THAT session's own launch/apply moment, and cannot see a swap caused
by a DIFFERENT session's later, ordinary use. Confirmed live: a second
Codex session picking a different model launched with no warning at
all — nothing conflicted at that exact instant — yet it silently
evicted the first session's model regardless (llama.cpp runs one model
at a time). Reproduced and root-caused via direct API calls against a
live test-picker instance rather than guessing.
- detectCustomModelSwapDisplacements() (custom-model-routes.ts): groups
live sessions with a customModel by endpointId, checks each group's
endpoint via GET /running once, and flags a session whose own modelId
is no longer in the running list. Read-only, best-effort per endpoint
like refreshAllCustomModelHosts's sibling sweep.
- Notifies once per displacement via a caller-owned de-dupe Set: a
session id is added when displaced, removed once its own model is
loaded/ready again, so a later genuinely-new displacement can notify
again.
- New periodic sweep in server.ts (CUSTOM_MODEL_SWAP_CHECK_INTERVAL_MS,
20s — much shorter than the 5-minute model-list refresh, since this
is time-sensitive) broadcasts a new custom-model:swapped-out SSE
event per displacement. De-dupe Set cleared per-session on session
cleanup to avoid an unbounded leak.
- Frontend: global toast (not tied to the displaced session's tab,
since the point is warning before the user types into it) naming the
session, its previous model, and what's currently loaded.
Chose the "detect after the fact" scope (vs. checking before every
message send, which would add a round-trip to every turn on every
custom-model session) per explicit user decision after being presented
the trade-off.
9 new tests for the detection logic (flag/clear/re-flag cycle,
unreachable/deleted endpoints, non-llama-swap servers, multiple
sessions on one endpoint). SSE registry bumped 158->159, parity test
passing. Typecheck/lint/frontend-syntax clean; full suite shows no new
regressions (9 more passing than baseline, matching the new tests;
same pre-existing Windows-environment failures).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RqZeHrRS6DYcGcGX2p9EwG
A fresh, isolated CLAUDE_CONFIG_DIR (used to keep an injected API key
away from a stored claude.ai OAuth login) looks like a brand-new Claude
Code profile to the CLI, so it replays its ENTIRE first-run sequence on
every single launch: the theme picker, the security-notes screen, the
per-project "trust this folder?" dialog, and (running with
--dangerously-skip-permissions) a one-time bypass-permissions warning —
confirmed live, none of which a real, already-onboarded profile shows
again.
- New registry-declared env-kind field `skipFirstRunPrompts` (alongside
apiKeyTrustFile, which it reuses) — claude's entry only, carried
through buildCustomModelInjection (pure) into
applyCustomModelInjection (IO).
- seedFirstRunOnboardingState(): merges hasCompletedOnboarding: true and
this session's own projects[workingDir].hasTrustDialogAccepted: true
into the same <configDir>/.claude.json the API-key trust file already
writes to — other projects and other fields on this session's own
entry are left untouched.
- seedSkipBypassPermissionsPrompt(): merges
skipDangerousModePermissionPrompt: true into <configDir>/settings.json,
a separate file, same corrupt-tolerant merge behavior.
- applyCustomModelInjection() gains an optional workingDir parameter,
threaded from session.workingDir (dedicated apply route) /
resolvedCasePath (quick-start route) — boot recovery omits it
(a dialog already answered once needs no re-seed on the same,
persisted isolated directory).
Tests added at the pure-builder, IO-wrapper (including merge-preserves-
other-fields and corrupt-file-tolerance cases), and existing directory-
listing assertions updated for the new settings.json file. Typecheck/
lint/format clean; full suite shows no new regressions (baseline
pre-existing Windows-environment failures unchanged, 8 more passing
tests than before — the ones added here).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RqZeHrRS6DYcGcGX2p9EwG
Claude Code's own fixed per-turn overhead (system prompt + tool schemas,
~36.4K tokens measured live) can exceed a small local model's entire real
context before any conversation history exists to compact — confirmed
live twice as an in:0 out:0 failure on the very first message sent.
CLAUDE_CODE_MAX_CONTEXT_TOKENS cannot fix this: it only governs when
history gets compacted, and there is none on message one.
- exceedsSafeContextFloor() (custom-model-routes.ts): true when a CLI's
registry entry declares contextLengthVar (currently only claude) and
the model's discovered context is below CLAUDE_MIN_SAFE_CONTEXT_TOKENS
(40000). A no-op for every other CLI by construction.
- Both apply routes (POST /api/sessions/:id/custom-model and the
quick-start customModel path) check this before the swap-conflict
check and before launching/restarting anything, returning
{requiresContextWarning, modelId, contextLength, minSafeContextTokens}
— skipped when confirmed:true.
- Frontend: #customModelContextWarningModal + _confirmContextWarning/
_resolveContextWarningConfirm (session-ui.js), wired into both
_quickStartWithCustomModelConfirm and _runCustomModelEntryViaRestart
(the path Claude actually uses) ahead of the swap-confirmation check.
Explains the fix in-modal: give the model an explicit larger -c/
--ctx-size in llama-swap instead of relying on --fit-ctx, which
optimizes for the biggest model that fits rather than the biggest
context.
Tests added for the route-level warning/confirm/skip cases and the
frontend modal + launch-flow wiring. Docs updated (custom-model-
endpoints.md, wiki/Custom-Model-Endpoints.md) and the PR's running
changeset extended.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RqZeHrRS6DYcGcGX2p9EwG
Fixes the visible double-launch reported on Codex: picking a custom-model
Run-menu entry launched natively first, waited for it to settle, then
restarted it in place with the endpoint applied. Necessary for the design at
the time, but visibly a native boot immediately followed by a second one -
worst on a CLI whose TUI fully reinitializes on a restart, confirmed live on
Codex.
POST /api/quick-start gains an optional customModel field
({endpointId, modelId, confirmed?}). When present, the route mints the
session's id itself (crypto.randomUUID()) before constructing it, computes
the same injection the existing POST /api/sessions/:id/custom-model route
computes (including the llama-swap conflict check from the last commit -
same {requiresConfirmation, currentlyLoadedModel, affectedSessions} shape,
no session created until confirmed), and launches the session already
pointed at the endpoint: env vars via the constructor, and the launchModel
override merged onto piConfig/grokConfig/ompConfig using the registry's own
launch.legacyConfigField the same way session.ts's restart path already
does. No restart at all - setCustomModel() afterward is bookkeeping only.
Wired into 7 of 8 launch functions (session-ui.js): openCode, codex, gemini,
pi, grok, deepseek, omp. Claude stays on the original launch-then-restart
path for now: its own --resume-based restart is far less jarring than the
other seven's, and runClaude()'s multi-tab launch plus docker-config-drift
confirm/retry loop make folding it into the one-shot path separate,
higher-risk work than the other seven's each-a-single-simple-launch shape.
Also fixes a pre-existing 'mode === omp' branch flagged by the CLI-id
static guard (test/cli-registry-no-id-branching.test.ts) - the ompConfig
launchModel merge is the same 'legacy <Mode>Config plumbing' category as
the six sibling branches already allowlisted there, just newly literal
where it was previously only inside resolveOmpConfigForCreate's own check.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RqZeHrRS6DYcGcGX2p9EwG
Covers #430's full scope so far: the picker itself, the model-selection
dialog, periodic re-discovery, and the session-busy/toast/CLAUDE_CONFIG_DIR/
context-length/llama-swap-conflict fixes found through live validation
against a real llama-swap server.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RqZeHrRS6DYcGcGX2p9EwG
A host reboot takes the tmux server down with it, so every pane dies,
reconciliation finds nothing to attach to, and the board comes up empty.
Picking yesterday's work back up meant finding each conversation in history
and resuming it by hand, one at a time.
The boot pass now works out what the reboot killed and leaves it on offer.
It runs inside restoreMuxSessions(), in the window where reconciliation has
reported the dead sessions and cleanupStaleSessions() has not pruned their
records yet, which is the only place the records can still be read. The
board shows a banner, and nothing is created until the user clicks it.
A click rather than an automatic restore is what makes the reboot heuristic
acceptable. The heuristic cannot tell a reboot from a crash that took tmux
down inside the same window, so it decides whether to ASK, never whether to
act: a wrong yes costs a line of text the user dismisses instead of N CLI
processes nobody asked for.
Four things are re-checked when the click arrives rather than trusted from
boot, because hours can pass and the board moves on. The owner's privilege
grant re-resolves through the env clamp. The workspace must still be on
disk. A conversation the user already resumed by hand from the Resume list
is skipped, since two panes running --resume on one conversation would
fight over the same transcript. Entries leave the plan synchronously before
the first await, and the route is single-flighted, so a double-click or two
devices cannot both reach the same entry.
A restored session comes back attached, idle and disarmed. Respawn
controllers and Ralph loops are deliberately not re-armed: a machine that
just came up is the worst moment to turn an autonomous run loose. Its
workspace hooks are installed by the restore route itself, because the
boot-time sweep sits behind a gate that is false after a reboot and has
finished long before the click; without them a session goes silently blind,
with no stop or idle events for respawn, no Approvals Inbox item and no red
tab on a blocking dialog. Stats collection starts the same way.
The pane is new, so the conversation continues and the terminal scrollback
does not. The banner says so rather than letting an empty pane read as a
broken restore.
The plan lives in memory only. A server restart drops it, which costs the
convenience this adds and never the conversation: the conversation is the
transcript under ~/.claude/projects, which the Welcome screen's Resume list
and the Session Manager already read, so a dropped plan returns the user to
resuming by hand.
clampEnvOverridesForOwner moves to src/session-env-clamp.ts, since the
question it answers is about session privilege rather than about HTTP and
it now has a caller outside the route layer. Its test hook stays re-exported
from session-routes.ts.
Claude sessions only for this pass. The other CLIs name their thread in
their own config object, which this does not thread through yet. Remote and
docker sessions are skipped on purpose, because both need another host or a
container to be up and a freshly booted machine cannot promise either.
Refs #411
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Every message typed on an Android phone lost its last character.
An Android soft keyboard commits the last typed character and sends the
Enter key in ONE InputConnection transaction, so the committed-text
`input` event and the Enter keydown are both processed before any
zero-delay timer runs. The orphaned-input recovery from #388 resolved
its candidate only on such a timer, and that lost the character twice
over:
* ORDER — xterm emits `\r` synchronously from the Enter keydown, and
the local-echo composer submits `pendingText` right there. The
recovered character arrived one macrotask too late to be part of the
prompt.
* LOSS — that same `\r` bumps the canonical counter, so by the time
the candidate resolved, `canonicalCount > snapshot` read as "xterm
spoke for this keystroke" and stood the recovery down. The character
was not merely late, it was dropped.
Drain pending candidates synchronously at the next keydown instead, from
xterm's custom key handler, which runs before xterm processes that key.
The counter then still holds the value it had while the candidate's own
keystroke was current, so the stand-down decision is made against the
right keystroke, and the recovered byte reaches the composer ahead of
whatever the new key emits. The timer stays as the fallback for a
keystroke with no key after it.
Physical keyboards are unaffected: there the timer has already resolved
the candidate long before the next key arrives.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* chore: version packages
* chore: sync the CLAUDE.md version line to 1.29.1
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
---------
Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
Co-authored-by: Codeman maintainer <noreply@anthropic.com>
Live terminal events are queued while a buffer load runs, and the load discards
that queue when it ends. That is right when the loaded buffer is the server's
accumulated byte history. The route appends to that history right up to the
moment it serializes the response, so a queued event already appears in it and
replaying it would duplicate output, most visibly Ink's cursor-up redraws.
A tmux pane capture is a photograph, current only as of the instant
`capture-pane` ran. Output printed afterwards was queued and then dropped, and
nothing scheduled a re-fetch to recover it: `_onSessionNeedsRefresh` is wired
only to the 128KB overflow path. The CLI's next partial redraw then landed on a
frame the terminal never received.
How much went missing depended on which capture the route served. A `?full=1`
load returns the capture alone, with no history in front of it, so it lost
everything from the capture to the end of the chunked write. A `?tail=` load
returns history, a clear, and then the capture, and the route reads that history
after the capture, so it lost everything from the response to the end of that
write. The chunked write dominates either way. An agent CLI hides the loss on
its next full redraw; a shell session does not, because its output is linear and
nothing repaints it.
Queue entries now carry their arrival time, and `_finishBufferLoad` takes a
`since` cutoff, so a capture load replays exactly the tail that arrived after
the response headers. The earlier events stay dropped, because a payload that
carries history does hold those.
All four paths that fetch a terminal buffer and write it now decide this the
same way, through one `_bufferLoadFinishOpts` helper, so they cannot drift
apart: `selectSession`, `_onSessionNeedsRefresh`, `_onSessionClearTerminal` and
`_maybeRefetchFullHistory`. The second of those is the one that stings. It
exists to restore output the client already dropped once under backpressure, and
it was dropping more output while performing that recovery. The cache-hit write
inside `selectSession` stays on discard deliberately: it runs before the fetch,
so its queue holds only events the capture that follows already contains.
Two further things had to change for that tail to still exist when the load
ends, and a browser test is what found both. `chunkedTerminalWrite` is what ends
the load for every non-empty buffer, so the flush policy travels to its own
finish calls; the call in `selectSession` runs only when the write was skipped.
`_beginBufferLoad` no longer empties the queue when one load re-enters it, which
it does on every write, because that reset discarded the whole fetch window
before anything could replay it.
The response already distinguishes the sources. `source` reads `mux-visible` or
`mux-full-history` for a capture and `history` for the byte stream.
Follows #395, #396 and #397, which fixed the ways the replayed frame itself
could disagree with the terminal.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Finishes #376. The contributed keystroke tracker sat on the raw byte stream
and named tabs wrong five ways (every prompt, every write path, a bare Esc
eating the next prompt's first character, pasted newlines as Enter, any CSI
clearing the draft) and replaced the whole name, which dropped the case from
the tab and reset the w<n> counter. This lands the feature with each of those
closed:
- First prompt means the first: applyAutoName() flips a placeholder to
`auto` whether or not the string changed. nameSource is now the tri-state
placeholder | auto | manual; the name setter is the only manual path.
- Only user-originated input counts: write()/writeViaMux() take
SessionWriteOptions.fromUser, set by the browser WS path and POST /input
only, so Ralph, respawn, cron, approvals and the trust-dialog keys can
never name a tab. A startMode 'shell' CLI never feeds the tracker (a
capability, not an id check); the send-key route feeds trackUserInput()
because its line feed bypasses the session.
- Prefix form `w3-case: title`: parseSessionPrefix() already renders it as
the title with the prefix in the tooltip and the next-session counter
still matches it. Composed within MAX_SESSION_NAME_LENGTH.
- Tracker rules per key: bare Esc resolves at chunk end; mouse/focus
reports, Tab, cursor keys, Shift+Tab are no-ops; Up/Down and Ctrl+P/N/R
taint the draft so Enter submits nothing rather than a fragment;
bracketed-paste newlines and Ctrl+J / Shift+Enter join with one space;
the draft keeps its head past 8192 code points; an escape past 64 bytes
is abandoned.
- Title: slash commands by shape (a path is a prompt), `!` escapes
refused, first sentence only past 8 code points ("e.g." is not a title),
72 code points on a word boundary.
- Synced `autoNameSessions` setting, default OFF (the prompt reaches
mux-sessions.json, session:updated and /api/search), App Settings ->
Appearance -> Tabs, read fresh per prompt after the eligibility check.
Tests: test/session-auto-name.test.ts (tracker, title, composition,
ownership, emit gating), the wiring test (once, prefix, setting off,
manual protected), test/routes/session-name-routes.test.ts (PUT /name
flips to manual and persists). Verified live on an isolated instance: API
and browser-typed prompts name the tab, a second prompt does not, shells
and renamed tabs are untouched, nameSource survives a restart.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Addresses the "left as they are"/"worth knowing" items Ark0N named when
merging #380 (the CLI-catalogue-driven install.sh + Docker agent image
PR), none of which were correctness-blocking but all of which were real:
- Removed install.sh's dead _cli_index/check_cli/get_cli_path helpers:
the catalogue-driven menu and hints stopped calling them and nothing
else ever did.
- The generator no longer emits CLI_KIND/CLI_NPM, two bash arrays
install.sh never read (the .mjs/docker-hosts.ts producers already
read the JSON catalogue's kind/npmPackage fields directly, so only
the bash copies were dead).
- detect_all_clis now skips a disabled entry's probe entirely instead
of running it and filtering the result downstream. No stock entry
ships disabled today, so this closes a latent inefficiency before it
is a latent bug rather than fixing an observed one.
- The install hint for a launcherProfile entry (DeepSeek today) now
explains in one line why it's a docs link and not a command: its own
docs page documents `npm install -g @deepseek-ai/dsh`, which installs
the launcher only and can't drive a pane, the exact trap the menu
already avoids by withholding the command. Driven by a new generated
CLI_LAUNCHER_ONLY array (from discovery.launcherProfile), not an id
check, so any future launcherProfile entry gets the same caveat free.
- Corrected the non-interactive-default comment: on a wget-only host,
Claude's curl one-liner is filtered out of the offered list first, so
the default becomes whichever npm-based entry sorts earliest instead
(Codex today), not always Claude. Behaviour is unchanged — it was
already printed, never silent — only the comment overclaimed.
Tests: extended test/install-sh-invariants.test.ts with a positive
guard for the new array and the trimmed array list, a negative guard
that CLI_KIND/CLI_NPM/the three dead helpers cannot come back, and two
real-bash tests (driven the same way the existing skip-menu tests are)
proving a disabled entry is genuinely never probed rather than merely
filtered after the fact.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011WzDjJnbK7zug8iQWnCc9z
flushPendingWrites() captured the viewport of a user who was reading
scrollback, called terminal.write(), and restored the anchor on the next
line. xterm parses on its own schedule, so at that point the buffer has not
moved: the guard `viewportY !== preserveViewportY` was false, scrollToLine
was never called at all, and the Codex redraw landed a tick later and took
the viewport to the live bottom with nothing left to pull it back. Scrolling
up during a stream still got dragged down, which is what #358 reports, and a
refresh was the only way back to a coherent view.
The restore moves inside xterm's write callback, the first moment the
redraw's effect exists, and runs before _scheduleTerminalWriteFlush() so a
deferred remainder re-captures the restored anchor rather than the bottom.
Two things follow from it running later:
- A live anchor now wins over the sticky scroll-to-bottom. The two are
captured at different moments (_wasAtBottomBeforeWrite at the frame's
first batchTerminalWrite, the anchor at flush time), so a scroll-up in
between leaves both set, and running both would jump to the bottom and
come back a frame later instead of staying put.
- The anchor is dropped if the active session changed or a buffer load
started while the write was in flight. It indexes the buffer it was
captured from, and selectSession() resets the terminal and chunk-loads a
different scrollback.
The existing regression passed throughout, because its write mock moved the
viewport synchronously, which real xterm never does. The harness now models
an asynchronous parse (redraw lands, then the callback fires), and all five
of the anchor tests fail against the old code.
Fixes#358
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Every per-skin xterm palette declared its selection layer as `selection`,
the key xterm.js renamed to `selectionBackground` in v5. An ITheme is a
plain object handed straight to the terminal, so an unknown key is not an
error, it is dropped: all seven skins have been drawing xterm's built-in
default, rgba(255,255,255,0.3), rather than the colour sitting next to it
in the palette.
Nobody saw it on the dark skins, where white at 30% is close to what those
palettes asked for. On the four light skins it is white over a near-white
background: blended, Paper Gray's selection differs from its own background
by 3/255. That is not a subtle highlight, it is no highlight, and it looks
exactly like a selection gesture that failed, which is part of what #360
reports on Android Chrome.
test/skin-themes.test.ts pins both halves: the key name, and that the
blended selection stays at least 16/255 from the background on every skin,
plus the light-skin fallback landing under that floor, which is what makes
this a fix rather than a rename.
Refs #360
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The two keys #408 adds to the mobile keyboard accessory bar send
Shift+Left and Shift+Right, which are Codex bindings (edit the last
queued message, step back through the prompt stack). They shipped on
both agent layouts, so a claude, pi, grok, omp, deepseek or gemini
session got two keys that do nothing. That was not only cosmetic: a tap
goes through sendNavKey(), which adds the session to
_echoPassthroughSessions and hands editing to plain PTY echo until Enter
or Ctrl+C, so on a phone a dead key also switched off the local echo
that makes typing feel instant there.
The reveal now follows the shape the 🧠 key already uses. The buttons
stay in both templates, carry an accessory-btn-codex marker class, and
are display:none in styles.css until the bar element carries
codex-enabled. The class has to live on the bar rather than on the keys
because setMode() rebuilds the buttons' innerHTML on every layout
switch. syncCodexKeys() toggles it from the active session's mode
(the same lookup _isShellSession() uses) and is called at init and from
refreshForActiveSession(), which selectSession() already invokes on
every switch. A session's mode is readonly on the server and fixed at
create, so no other event can change the answer; the welcome screen
(no active session) reads as not codex and hides the keys.
The frontend id-branching guard (test/cli-registry-no-id-branching.test.ts)
scans only src/**/*.ts, so the mode comparison in a public JS file is
in bounds, the same as the existing shell check beside it.
Tests: the new describe block in test/mobile-shell-keyboard.test.ts pins
the marker class in both templates, the CSS pair, the class for a codex
session in both layouts, its absence for claude/shell/pi/omp/deepseek,
the re-sync in both directions on a session switch, the no-session case,
and the init + refresh wiring. All six positive assertions fail without
the source change. README and the changeset now say the keys are
Codex-only.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
A clicked path that points OUTSIDE the case directory goes through the attachment
routes (the frontend's `_isExternalPreviewPath` sends every absolute path not under
`workingDir` to `POST /attachments`), and those had the same local-`fs` assumption
as file-raw: `realpathSync`/`fs.stat` on a path that only exists on the remote host,
so the file never opened — the case the #415 report was actually about.
- `registerExternalAttachment()` accepts `remote` and resolves through
`remoteProbePaths` (canonical path, size/mtime, kind, plus the workspace root for
the confinement check). Everything around it — blocklist, extension allowlist,
workspace confinement, registry/dedupe — is now shared by both branches, so the
remote path cannot drift from the local one.
- The by-id routes (`raw`, `preview`, `thumbnail`), the metadata poll and the
attachment history list resolve over ssh too. `raw` streams with the same
Range contract as file-raw; `preview` (office) and `thumbnail` answer 400 for a
remote record; an unreachable host answers 502, a vanished file 404.
- Which host a record is read from follows the SESSION, never the path string: the
same absolute path is a different file on each host, and a remote session never
falls back to a local file with that name.
- Codex generated artifacts keep force-workspace confinement for a remote case: the
well-known artifact directories are anchored at THIS host's home, so only a file
inside the remote workspace is trusted.
Still local-only by design: writes, office conversion, thumbnails, the file
tree/picker and tail-file.
A remote case's workingDir is an absolute path on the remote host, but the
file read routes resolved it with local `fs`: `validateSessionFilePath`'s
realpathSync fails for a path that does not exist on the Codeman host, so
every preview of an agent-written file answered "File not found" (#415).
Add src/remote-files.ts as the single remote-read layer, built on the same
buildSshConnectionArgs() the launch uses:
- remoteProbePaths(): ONE round trip returning realpath + stat for the
requested path AND the workspace root, so containment is checked against a
remotely canonicalized root (a symlinked remotePath is ordinary).
- remoteCreateReadStream(): streams the body (cat, or tail -c +N | head -c L
for a Range) with nothing buffered in memory, and reaps the ssh child when
the response ends so an aborted download cannot orphan it.
- remoteReadFile(): bounded read for file-content.
file-raw, file-content, file-preview and file-thumbnail now share one local/
remote target resolution. Guards keep their local strength: lexical pre-check,
remote realpath, workspace containment, sensitive-path blocklist, and the size
cap applied to the remote size before any bytes are read. An unreachable host
answers 502 with the remote reason instead of a misleading 404. Nothing is ever
copied to the Codeman host and there is NO local fallback (an sshfs mount of
the same tree must not shadow the remote bytes).
Deliberately unchanged: writes (edit=1 / PUT now answer 400 explicitly while
the viewer hides its Edit affordance), office previews, thumbnails, file tree,
picker, external attachment registration and tail-file stay local-only.
Rebased onto current master (the one real conflict was the import line
in docker-hosts.ts Ark0N flagged; kept both), then addressed every
point from the review:
**1. Rebase.** Done — this branch now sits on current upstream/master.
**2. Agent-image special cases are data now, not an id-keyed table
outside stock.ts.** `AGENT_IMAGE_SPECIAL_CASE_IDS`/`AGENT_IMAGE_SPECIAL_CASES`
are gone. `CliDiscovery.install.agentImageLayer?: { kind: 'dedicated';
reason: string }` is a field on the registry entry itself (pi,
deepseek), `reason` is required by schema.ts, both producers
(docker-hosts.ts and cli-catalog.mjs) filter on its presence instead
of an id, and the coverage test reads it from the generated catalogue.
Also added the npm-package-name validation to the TS producer, which
only the .mjs one had — same SAFE_PACKAGE regex, duplicated
(necessarily, one side can't import the other) and now pinned
byte-identical by a new parity test.
**3. Changeset said five, it's eight.** (Not nine — see the DeepSeek
point below, which changes the true count.) Reworded to state it
structurally rather than pin a number that will go stale again.
Then the four behavior-changing findings:
- **DeepSeek was offered as a normal install option but can't actually
drive a pane.** `npm install -g @deepseek-ai/dsh` installs the
launcher only; DeepSeek ships no profile that can run standalone.
The generator now emits an empty install command for any
`launcherProfile` entry, so install.sh's menu (which requires a
non-empty command) skips it and falls through to its docs URL hint
instead — matching what the old hand-written code did before this
PR replaced it.
- **wget-only hosts lost every automatic install, including the npm
ones that never needed curl.** The menu-building loop now filters
PER ENTRY (only a command starting with `curl ` is held back) rather
than wiping the whole menu when DOWNLOADER != curl.
- **The DISPLAY/TRUSTED split and the catalogue refresh didn't hold up
under review** (refresh's only real write was the label; it ran
before the Node existence check; its own eval-detection test was
tripped by the word "eval'd" in a comment). Dropped entirely per
your own recommendation — embedded catalogue only, no network
fetch, no second array. install-sh-invariants.test.ts now asserts
the refresh/DISPLAY machinery does not exist rather than testing its
internals.
The three take-or-leave items, applied:
- `dsh_banner_probe`'s bash 3.2 empty-array bug: `${runner[@]}` →
`${runner[@]+"${runner[@]}"}`. Verified live in a real `bash:3.2.57`
container with `timeout` removed from PATH — crashed before, clean
now, full `detect_all_clis` path exercised end to end.
- `docker-agent-image-coverage.test.ts` now anchors on each layer's
`<binary> --version` proof line instead of `Dockerfile.includes(binary)`,
which stayed true if a layer were deleted but its comment survived.
- Doc drift: docs/docker-cases.md (four → five, and now describes the
data field), docker/agent.Dockerfile's "other four CLIs" comment (no
longer a magic number — CLI_NPM_PACKAGES is generated and can grow),
CLAUDE.md's install.sh size (104KB → ~112KB) and its stale mention of
the now-dropped refresh.
Verified: tsc clean, prettier clean, the full targeted suite (142
tests across the 8 affected files) green, and the full `npm test` gate
diffed BY TEST NAME against a clean upstream/master baseline run on
this same machine — identical 201-name failure set both sides (168
tests / 67 files, all pre-existing Windows-environment noise: symlinks,
PTY spawning, POSIX permission bits — none of it touching anything
this PR changes), zero new failures either side of the diff.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R9ZSTEenc8soSu9bTi8Xru
Adds a "Consumers outside the server" section covering the two generated
artifacts, why each exists (neither install.sh nor a .mjs can import
TypeScript), what is deliberately NOT exported and why, the three-rule install
command trust boundary, and the bash 3.2 constraint with the offset/length
window shape it forces.
The adding-a-CLI checklist gains the regenerate step, since forgetting it is how
the installer would keep detecting the old set while the server offers the new
one — the drift this change removes, one level out.
docs/docker-cases.md gains how CLI_NPM_PACKAGES is derived, why it reads the
stock catalogue and not the merged registry, and a table of the four documented
Dockerfile special cases with their reasons. CLAUDE.md gains a command row and
names the generated block, the bash 3.2 rule and the trust boundary in its
install.sh paragraph.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015EMxQreQUZX5ZyybxAGh12
A Fable 5.1 reviewer read the whole release diff against 1.26.2 and returned
SHIP WITH FIXES. These are its findings, verified before acting on each.
**The changelog advertised a feature the code refuses (major).** The #401
changeset and docs/web-tabs.md both listed `*.localhost` in the loopback set.
The follow-up in 02b0e278 moved it out of the auto-route set on security
grounds and updated CLAUDE.md but neither of those, and that changeset becomes
the 1.27.0 CHANGELOG entry: a user would have read the release notes, tapped
`http://app.localhost:3000/` on a phone and got a connection error from a
documented feature. Both corrected, and the user guide now says why it is
excluded and that adding such a dashboard by hand still works.
**Dictation delivered its text twice (minor, #388).** `keydownSnapshot` started
`null`, so `keydownSnapshot ?? canonicalCount` at the input event read a counter
xterm had ALREADY bumped: on a fresh page load with no keydown yet, xterm's own
capture listener forwards the `insertText` itself (it is not gated behind a
keydown), then the snapshot equals the bumped count, `count > snapshot` is
false, and the controller emits the same text again. Reproduced directly
against the module: it emitted `hello` for input xterm had already delivered.
A `0` baseline restores that file's own invariant, that a missed recovery is
acceptable and a duplicated keystroke is not. Two regression tests, covering
both the xterm-already-delivered and genuinely-dropped halves.
**The sorted rail's arrow-key walk followed the DOM (minor).** `_tabKeydownHandler`
steps `querySelectorAll` order, which is `sessionOrder`, while a sorted rail
paints its rows with the flex `order` property, so ArrowDown from the top card
landed wherever that session happened to sit in the tab order. It now sorts its
node list by the COMPUTED order first: computed rather than inline, because web
tabs take their `order: 9999` from CSS and would otherwise read as 0 and lead
the walk. This is the one place that follows the paint; the Alt+N badge, the
drag model and the filter all still deliberately read the DOM.
**A trusted dashboard was auto-reused by a tapped link (minor, #401).** The
reuse loop skipped `managed` and direct-mode records but not `trusted`. A
trusted frame is mounted with `allow-same-origin`, i.e. on Codeman's origin
with the user's cookie, and these links come from agent output, which is the
threat model the loopback allowlist was just narrowed for. An agent that can
write into the dev server's tree could print a path that one tap opens inside
that privileged frame. Excluded from auto-reuse, with a test; opening it from
the Run dropdown is still an explicit action and unchanged.
**Two documentation claims that were no longer true.** CLAUDE.md said
test/location-overlay-commands.test.ts pins every remote pane command, but
remote claude and remote omp now have their own arm in `buildRemoteLaunchCommand`
and never reach `defaultRemoteCommandForMode`, which is what that test asserts,
so it pins nothing for them and changing either arm will not fail it. Named the
real pins instead. Also documented the arrow-key-walk exception in the rail
paragraph.
Left as follow-ups, deliberately: `POST /api/webviews` does not dedupe by URL
server-side, so two devices tapping one link concurrently can still save two
dashboards for one origin (pre-existing endpoint behaviour that #401 makes
reachable by a tap), and the location-overlay golden should assert the real
remote claude/omp commands rather than a branch neither reaches.
Full gate green: 359 files, 6869 tests.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A dashboard served through a web tab saw `/webview/<cap>/` as its
`location.pathname`, and no app has a route for that: a React Router, Vue
Router or Vite dev-server page painted its HTML and CSS and then replaced
them with its own "page not found" the moment its script ran (reproduced
with a minimal history-routed page).
The proxy's runtime shim now rewrites the history entry to the path the
page would see on its own origin, before any page script runs. The base
element still resolves relative URLs inside the prefix and every root-
absolute sink is rewritten back into it, so only what the page READS
changes. With the document URL masked the Referer-keyed 404 rescue can no
longer help a request the shim misses, so the remaining URL-taking entry
points (`Worker`, `SharedWorker`, `navigator.sendBeacon`, `window.open`)
are covered by the shim as well.
A navigation the page starts itself afterwards — `location.reload()`
(a dev server's full-reload HMR), a root-absolute `location.href` — lands
on Codeman's root with no capability anywhere: no prefix in the path, no
cookie in an opaque-origin frame, a Referer naming the masked page. It is
recognised by shape (a top-level iframe navigation asking for HTML, for a
path Codeman does not serve) and answered with a static page whose only
script posts `{type:'codeman:webview-lost', path}` to the parent; the tab
that owns the frame (matched by `event.source`, never by the payload)
remounts it inside the prefix at that path, bounded per frame. The
unauthenticated form is answered in the auth middleware before the
credential checks, so a dev server that reloads on every save cannot
rate-limit its own user out of Codeman; the authenticated form (Basic
auth, trusted mode) is answered by the 404 handler.
Verified end to end against a history-routed page: boots on `/`, its
API call succeeds, a reload inside the frame comes back routed on the
path it had pushed, `location.href = '/about'` comes back on `/about`,
and a deep link opens on its path.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Codeman maintainer