The open-tabs list on the welcome screen was a fixed 256px card floating
vertically centered in the left gutter, which read as debris rather than
chrome and left 12px type stranded on a wide display.
- Dock it: left/top/bottom 0, full height, hairline right border and a soft
background fade. The centered welcome content still does not move.
- Scale it off one knob: width clamp(250px, 19vw, 430px) plus a fluid
font-size on .home-sessions, every child sized in em. Measured 250px/12.2px
at the 1180px gate, 380px/15px at 2000px, 430px/17px at 2938px; the gap to
the centered content never goes negative.
- Show when each session was first created and last active, on a full-width
footer line so it does not fight the status pill, exact dates in the title.
Both stamps refresh in place on a 20s clock (disarmed when the home screen
goes away) rather than by re-rendering, which would restart every row's
blink animation and working ring twice a minute.
- Mute idle green: dot and pill mix toward --text-muted, so idle reads as
greyed-out next to the vivid green of a working session. Mixed rather than
hardcoded, so every skin keeps its own green.
Verified in a browser at 1180/2000/2938px and on a light skin, plus
test/home-sessions.test.ts, frontend-syntax, public-assets, prettier and a
PostCSS parse of styles.css.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Round 2 of the #251 review: settingsWriteBlocker covered only
writeHooksConfig and updateCaseModel, while applyStatusLineConfig,
stripCaseEnvKeys, updateCaseEnvVars, refreshStaleCodemanHooks and
ensureCodemanHooks still wrote the same repository-controlled path
unguarded (applyStatusLineConfig was demonstrated writing through a
symlinked settings.local.json).
All seven writers now go through withSafeSettingsWrite(), which runs
the blocker check INSIDE the per-path settings lock and then hands the
writer its claudeDir/settingsPath; none of them touch the settings path
directly anymore. Test pins all seven against a symlinked
settings.local.json at once (link target must stay byte-identical).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Addresses all four findings from the #251 review:
- Scaffolding no longer writes through repository-controlled symlinks.
The guard lives in hooks-config.ts (settingsWriteBlocker) so it also
covers quick-start/docker/ralph writers, not just the clone route:
refuses a symlinked .claude or settings.local.json, a .claude that is
a file, or one resolving outside the case. The clone route surfaces
the refusal as a user-visible warning, and the CLAUDE.md write checks
presence via lstat so a BROKEN repo-shipped symlink counts as present
(existsSync follows links and would have created the outside target).
- Failed-clone cleanup can no longer delete a concurrent winner's tree:
git clones into an attempt-owned temp sibling (.<name>.cloning-<rand>)
which is atomically renamed into place; the loser reports
DESTINATION_EXISTS and only ever removes its own temp dir.
- decodeURIComponent(url.pathname) is guarded: malformed percent-escapes
now come back as BAD_SYNTAX instead of an uncaught URIError 500.
- The git pool's waiter queue is bounded (CODEMAN_MAX_GIT_QUEUE, default
16): overflow answers BUSY immediately (HTTP 429 via RATE_LIMITED),
and queue time counts against the operation's own deadline.
Tests: hostile symlink fixture repo (route level), settingsWriteBlocker
units, concurrent same-destination race, temp-dir leak assertions,
percent-escape rejection, and a fake-git pool-bounds suite.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The feature as pitched in docs/readmymind-plan.md: pressing the header
brain button predicts the prompt you were about to type, from the case's
intent profile plus everything the session already knows.
Backend:
- readmymind-context.ts: pure budgeted context assembler (9 ranked
sources: pending approval dialog, user goals, last assistant turn tail,
recent prompts, tool activity, git workspace signals, away context,
sibling sessions, rethink state; 30 KB budget, whole-section drop from
the bottom of the ranking, trust tiers stated in the prompt)
- readmymind-collectors.ts: transcript tail reader (the live watcher
keeps only a 500-char snippet) and git signal collection (execFile,
2s timeout, skipped for remote-SSH cases)
- readmymind-predictor.ts: one-shot claude -p in a throwaway tmux
session, opus by default (readMyMindModel setting), strict JSON
contract with 1-3 suggestions (continue / verify / redirect), newline
stripping, 90s timeout; mutable singleton so route tests can stub it
- POST /api/sessions/:id/readmymind: claude-mode only (400), one
prediction in flight per session (409 CONFLICT), rethink body
{ steer, rejected }; ownership via findSessionOrFail
Frontend:
- readmymind-ui.js (loadorder 11.3): header brain button, marker-hidden
until readMyMindEnabled is ON, desktop only (phone key is phase 3);
modal with editable suggestion + rationale and Send / Insert /
Rethink / Dismiss; suggestion text rendered via value/textContent only
and nothing ever auto-sends
- App Settings -> Panels checkbox for readMyMindEnabled; en + zh-CN
strings
Verified end to end against a live isolated instance: transcript
capture, a real opus prediction grounded in the stated goals, rethink
steering, the 409, and the browser modal incl. Insert leaving the text
unsubmitted on the composer. 41 new unit/route tests; full test:ci
sweep green (4680 tests).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adds an exact-key tier (ALLOWED_ENV_KEYS) beside ALLOWED_ENV_PREFIXES in
schemas.ts, admitting CLAUDE_CONFIG_DIR so a case can run on a separate
Claude subscription (client-billed accounts). Exact match only: other
CLAUDE_* keys and near-misses like CLAUDE_CONFIG_DIR_EXTRA stay rejected,
blocked keys stay blocked. The key also survives getEnvOverridesForPersist()
(a path, not a secret; dropping it would silently switch a rebuilt session
back to the default account after a reboot).
Docs cover the transcript caveat: a relocated config dir writes transcripts
outside ~/.claude/projects, so response viewer / subagent windows /
ultracode / Read My Mind go blind for that session unless projects is
symlinked back into the shared tree.
Design and spec contributed by @jordan8037310 in #255. Closes#255.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The brand "C" got a 44px-wide hit box in the previous commit but was capped at
36px tall by the bar it sits in. The phone header is now 44px, so the one
control that gets you back to the home screen is square at the platform
minimum, and every other header control gains the same 8px.
Redefined as --header-height inside the phone media query rather than as a
literal, so the panels positioned off that token (file browser, project
insights, plan overlays) follow the bar instead of drifting 8px underneath it;
.app's top offset is derived from it for the same reason. The header also stops
top-aligning its children on phones: that read as centred in a 36px bar whose
contents were ~31px, and leaves a visible gap under everything at 44px.
Costs 8px of terminal height on a phone.
Verified on a real isolated instance at 390px: header 44px, button 44x44
spanning the bar, a touch tap at (4,41) - inside the new area, outside the old
one - reaches the home screen, tabs centred, and content still clears the fixed
header. Tablet (48px) and desktop are untouched.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The welcome overlay centers ~560px of content in a ~1400px window, so both
gutters are dead space. The left one now carries the open tabs as a vertical
list (home-sessions.js): one row per live session plus saved web tabs, in TAB
order rather than by urgency, because the row badges are the Alt+1..9 indices.
Clicking a row enters that session.
Working state is deliberately the phone's, exactly: a pulsing green dot ringed
by the same tab-load-spin the tab strip uses while a tab loads, now with a green
halo added on both surfaces so "working" reads identically wherever you see it.
The column is position:absolute so the centered content never moves, which is
why it needs a width gate in two places (HOME_SESSIONS_MIN_WIDTH = 1180 in JS,
a max-width: 1179px media query as the backstop for a resize that outruns the
matchMedia listener). A test pins the two equal. State classification is reused
from mobile-overview.js rather than re-derived, so the two home screens cannot
disagree about what counts as needing you.
Phones keep the mobile overview, and their brand "C" was a 0.85rem inline span,
roughly a 12x13px target on the one control that gets you back to that screen.
It is now a 44px-wide button filling the full header height, with the glyph
scaled to match. 44 is horizontal only: the phone header is pinned to 36px and
clips overflow, so a true 44x44 would mean taking height off the terminal.
Verified end to end against a real isolated instance (own tmux socket + data
dir): 18 browser checks covering render, live update through the tab renderer,
the working dot's animation/glow/ring, row click, the narrow-window gate, the
phone fallback, and a real touch tap on the far corner of the new hit box.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
docs/readmymind.md covers phase 1 as a user guide: how to enable the synced
readMyMindEnabled setting via the API (no UI checkbox until phase 2), exactly
what is and is not captured, the hooks dependency (Docker bridge / remote-SSH
caveats), storage and wipe paths, curl examples for the three endpoints, the
agent-skill ground rules, and a troubleshooting table. Cross-linked from the
CLAUDE.md Key Patterns entry and the api-reference section.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Per-case profiles of user intent (docs/readmymind-plan.md): user-stated goals
plus the user's recently submitted prompts, captured from the Claude session
transcript behind the new synced readMyMindEnabled setting (default OFF).
- intent-store.ts: keyed by owner + realpath(workingDir), FIFO/size caps,
consecutive-dupe collapse, atomic 0600 writes to ~/.codeman/intents.json
- transcript-watcher.ts: new transcript:user_prompt event for typed user turns
(tool_result-only entries stay silent); capture wiring in server.ts is
claude-only and gated on the setting per event
- readmymind-routes.ts: GET/PUT/DELETE /api/sessions/:id/intent, ownership
via findSessionOrFail, strict Zod schema
- agent skill: SKILL.md recipe + endpoints.md rows so agents can read and
record intent (PUT replaces: read + merge; never delete unprompted)
- groundwork for the phase-2 predictor button; nothing is ever auto-sent
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
selectSession() ends with scrollToLastNonEmptyLine(), which parks the viewport
one row ABOVE the bottom for any session whose buffer is taller than the screen
and ends in blank rows, so that is the normal state after a tab switch. Nothing
pinned that a tap there still leaves the keyboard reachable.
The blocker reduced in #173 came back through exactly that gap in #244: a tap
classifier that treats "viewport is scrolled up" as a reason to blur, paired
with touchstart preventDefault cancelling the compatibility click, closes both
routes to focus on the same gesture and strands document.activeElement on
<body> with no way to type. The prompt row is no exception.
Measured on a 390x844 viewport, claude-mode session, dispatched touch gesture:
master leaves focus on textarea.xterm-helper-textarea, PR #244's terminal-ui.js
leaves it on body. Green here, red against that branch.
The test also pins the half that IS correct: SGR coordinates are meaningless
off-bottom, so the tap must send no mouse report.
It has to be a dispatched gesture. Calling the touchend handler directly
bypasses touchstart's preventDefault, which is half of what closes the focus
path, so a direct call reports the right intent and still misses the bug.
test/mobile/keyboard.test.ts: 4 failed | 32 passed (36), against 4 failed |
31 passed (35) without it. Same four pre-existing failures either way.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The working dot is the one glance-state a phone needs: busy tabs now get a
9px pulsing dot with a green glow (idle stays 4px). The glow needs !important
because the skin block's no-halo rule outranks mobile.css.
The simple keyboard accessory bar swaps /clear for Tab (/clear and /compact
stay in the extended bar with their double-tap confirm). The tab action now
flushes locally-buffered prompt text to the PTY before sending \t, so
completion applies to what was just typed instead of an empty composer.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Adds an Add Case -> "Clone Repo" tab plus two endpoints, implementing
@DodgyBadger's proposal in #236: clone a public repository straight into
codeman-cases/<name> and register it as a normal local case.
POST /api/cases/clone is synchronous by design (request held open, bounded
by GIT_CLONE_TIMEOUT_MS): no job store, no polling, no cancellation
surface. Success broadcasts the usual case:created event, so the case
still appears when a proxy idle-timeout kills the request mid-clone.
POST /api/cases/clone-preflight runs `git ls-remote --symref` so the UI can
say, while the user is still typing, whether the URL is cloneable without
credentials, what its default branch is, and which branches/tags exist.
Core lives in src/git-clone.ts, split into a pure half (URL parse, argv/env,
ls-remote parse, stderr classification) and a thin IO half, so every
security decision is unit-testable without spawning anything:
- `<name>::<payload>` transports are refused as a family, not by name:
ext:: is the famous one, but any of them dispatches to git-remote-<name>
and turns a clone into arbitrary command execution.
- A leading `-` is refused AND every spawn puts `--` before the operands.
Either alone is one edit away from being a hole.
- argv arrays, never a shell. URLs carrying user:password@ are refused.
- gitNonInteractiveEnv() closes all four ways git can block on a prompt
with no terminal attached (terminal prompt, askpass/GUI, ssh, GCM).
HOME/PATH stay inherited, so a user's own credential helper or ssh agent
keeps working; Codeman itself collects and stores nothing.
- The timeout signals the process GROUP, since clone fans out into
git-remote-https/index-pack children that outlive a signal to the parent.
- Bounded output (redacted stderr tail, capped ls-remote stdout, 500 refs
each) and a global 2-op pool, so N large clones cannot exhaust the host.
Repository contents beat scaffolding: an existing CLAUDE.md is kept, hooks
are merged into whatever .claude/settings.local.json the repo shipped, and
a repo that ships its own Claude settings is reported back as a warning
(those hooks run locally as soon as a session starts there). A failed clone
removes only the directory the attempt created, and refuses a pre-existing
destination outright, so it can never squat on a case name.
Not admin-gated in multi-user mode, unlike /api/cases/link: it writes only
inside the caller's own case space. Local-path/file:// sources are the
exception and stay admin-only there.
UI: live verdict under the URL field, case name filled from the parsed repo
until the user types their own, branch/tag as a datalist of the remote's
real refs, optional shallow clone, and a Brain picker (installed CLIs only)
that points the Run button at the chosen agent. Starting a session stays
opt-in. The tab hides itself when the server reports no git.
Tests: the pure half exhaustively (every refusal has a case), plus real git
against a real local bare repo for clone/ref/timeout/cleanup, and a
route-level suite with unmocked fs that clones through the endpoint.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The picker behind Link Existing's "Browse" and the mobile keyboard's Path key
refused every path with a dot-prefixed segment, so `.github/workflows/ci.yml`
could not be selected and a hidden folder could not be opened at all. It gains
the same `.*` toggle as the File Viewer: default OFF, per-device, and applied to
both the listing and the preview endpoint, which re-resolves the path
independently.
That dotfile filter was quietly doing security work. The picker's roots include
Home, so with every hidden path unreachable the shared blocklist never had to
name the credentials that live in dot-directories. Lifting the filter removes
that accident, so `isSensitivePath` now covers them explicitly: SSH keys at any
depth rather than only under $HOME, GPG keyrings, AWS/GCloud/Azure/Docker/
Kubernetes credentials, npm, Yarn, git, gh, netrc, PyPI, RubyGems, Cargo and
Terraform tokens, .pgpass and .my.cnf, and the Claude and Codeman agent
credentials. `~/.codeman/` and `~/.claude/` stay attachable as trees, since the
publish skill and the review-card loop read from them; only their secret-bearing
members are named.
Everything else still applies with the toggle on: blocked trees, sensitive
files, root confinement, ownership scoping and symlink-escape checks. A hidden
entry whose realpath is a secret is dropped from the listing, and opening it is
refused.
Follows #221
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
One switch now governs the whole feature: with approvalsInboxEnabled off
(the default), sendPushNotifications strips the actions and approvalId
from permission push payloads, so the buttons no longer render at all
(pre-inbox they rendered and did nothing). The page-side action relay is
gated the same way for stale notifications sent before the toggle
flipped. Only the store and answer endpoints keep running, so enabling
the toggle surfaces anything already pending immediately.
sendPushNotifications is async now (cached settings read); all call
sites were already fire-and-forget. Covered by three new payload tests
alongside the existing hostTitle suite.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
A session on a fresh directory sat on Claude's "Quick safety check: Is
this a project you created or one you trust?" dialog until a human
pressed Enter. Reproduced on a new case, then read off the wire:
1.\x1b[C Yes,\x1b[C I\x1b[C trust\x1b[C this\x1b[C folder
tmux repaints a row by writing each word followed by a cursor-forward
escape instead of a space, and Ink colours each word separately, so
`data.includes('trust this folder')` could never match a chunk. The
spaces are not there to strip: they were never sent. The auto-accept has
been dead for every session that hit the dialog.
Match on whitespace-free, ANSI-free, lowercased text instead
(`compactScreenText`), which survives both that repaint style and the
spaced full-screen redraw.
Answering means pressing Enter into a session, so three guards bound it:
- Read the RENDERED SCREEN (capturePaneText), not the chunk. The terminal
buffer is append-only and keeps the dialog in its tail long after it
has been answered, so a retry driven off the buffer would type into a
live session. Direct-PTY sessions, which have no pane, fall back to a
short buffer tail.
- Require a trust phrase AND the dialog's own confirm affordance. One
phrase is not enough, since an agent's transcript can quote it.
- Only look during the first 90s of the pane's life, and cap it at three
attempts. Ink can drop a keystroke while it is still mounting the
widget, which is the other half of why sessions got stuck, but a
dialog that will not clear must not become an Enter loop.
Verified end to end on a fresh case: dialog answered on attempt 1, one
Enter sent in total, session went straight to the composer and answered a
prompt. Before the fix the same flow parked on the dialog indefinitely.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Opening Codeman with nothing reachable (phone off the tailnet, VPN down,
server stopped) rendered a normal-looking UI: the service worker serves the
cached app shell, every /api call fails, and the only tell was an 8px red dot
in the header corner. On a phone that reads as "there are no sessions".
Two surfaces, chosen by whether there is anything worth looking at:
- Full-screen overlay while no server state has loaded this page load. It
names the host, lists the three things to check (network, VPN/Tailscale,
server), counts down to the next retry, and offers "Retry now" plus
"Show cached view" to demote itself to the banner.
- Non-blocking banner once state HAS loaded, so a mid-session drop leaves the
terminal scrollback readable.
A 2.5s grace keeps a COM deploy (SSE is back in ~200ms) from flashing the
banner every release; navigator.onLine === false skips the grace, since the
device saying "no network" is never a blip. Retry re-arms the terminal
WebSocket as well as SSE: planWsReconnect can give up outright, and the SSE
backoff caps at 30s, so waiting it out is not always an option.
The decision is pure (computeConnectionLossUi in constants.js, unit-tested in
a node VM like the WS reconnect policy); app.js only writes the DOM.
Owner decision: every Approvals Inbox UI surface (header bell, drawer,
phone overview answer strips, reload seeding) now requires enabling
approvalsInboxEnabled in App Settings -> Panels; only an explicit true
turns it on. The store, endpoints, and push Approve/Deny actions keep
running regardless (the push buttons are already opt-in per subscription).
Also replaces em-dashes with plain punctuation across the newly authored
comments, docs, and strings.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The tree endpoint has accepted `showHidden=true` since it was written; the
panel hardcoded `showHidden=false`, so dot-prefixed entries were unreachable
from the File Viewer and opening one meant guessing its path.
Adds a `.*` toggle to the panel header. It re-fetches instead of re-rendering
the cached tree (the filtering is server-side), preserves the expanded
directories so toggling does not collapse the tree, and persists per-device to
its own `codeman:fileBrowserShowHidden` key. That key is deliberately not part
of the app-settings object, which `saveAppSettings()` rebuilds from the
settings-modal DOM and would drop it on the next save.
Default is OFF, so an untouched install behaves exactly as before.
Closes#221
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The overview already had a `working` state; nothing ever reached it,
because the status it reads was wrong (see previous commit). Now that a
row can actually be in it, the state needed to look like something.
- The row gets a slow green breathing edge (2.2s). Deliberately calmer
and slower than the red/yellow alert blinks, since working is not an
alert and must not compete with the two states that do want you.
- The dot keeps its `pulse` and picks up a spinning ring: the same 2px
ring with a bright leading edge that a tab shows while it loads,
reusing the `tab-load-spin` keyframes from styles.css rather than
re-declaring them, so the two cannot drift. Green rather than the tab's
blue because here it means "running", not "loading": the motion is the
shared part, the color still belongs to the state.
- The pill animates "working ...".
Reduced motion drops all three to static: a green edge, a full ring, a
static ellipsis.
Verified in headless Chromium at 390px against a live working session:
row breathe-green, dot pulse plus tab-load-spin ring, pill dots.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Every working Claude session reported `status: "idle"` about two seconds
into its turn. Measured on live workers: two sessions mid-tool-call at 13
and 17 minutes both read `idle` while their panes showed
`✻ Actualizing… (13m 23s · ↓ 47.5k tokens)`.
Two things had drifted apart:
1. The working indicator changed. Claude animates the glyph through
`· ✢ ✳ ∗ ✻ ✽` and randomizes the gerund per turn, so neither
SPINNER_PATTERN (braille, no longer drawn) nor the keyword list
(Thinking/Writing/Reading/Running) matches a turn anymore.
2. A `❯` sighting is not the end of a turn. Claude redraws the composer
roughly once a second all the way through one, and that redraw armed
the "2s later, call it idle" timer.
Matching the new status line in the STREAM does not fix it either: tmux
ships partial repaints, so the complete line reached the PTY about once
every 20 seconds while the `❯` arrived every second.
So the decision moves off the stream:
- An unbroken run of repaints marks a turn as started. Sampled once a
second for 12s over six live sessions, the two working ones produced
output in 12/12 windows and the four idle ones in 0/12. Pure helpers in
session-activity.ts carry the thresholds.
- Idle now needs the pane to go quiet AND the screen to agree.
`_confirmIdle()` asks tmux what is rendered (new `capturePaneText()`,
one plain `capture-pane`, floored at 1.5s per session and only ever at
a transition) and re-checks every 5s while the screen still shows work.
A turn can sit silent for tens of seconds inside one tool call, so
silence alone proves nothing.
- The same screen check vetoes keystroke echo, which is a steady stream
of repaints too but is not work.
CLAUDE_WORKING_LINE_PATTERN matches the `… (elapsed)` shape rather than
the glyph, because the FINISHED line (`✻ Cooked for 2m 49s`) carries the
same glyph and would otherwise pin a session at working forever.
Claude mode only. An external CLI has no `❯`, so nothing would arm the
confirmation and such a session would latch busy.
respawn-patterns.hasWorkingPattern() had the same blind spot (its gerund
list cannot see "Actualizing"), so it takes the pattern as an extra
signal. That can only make respawn less eager, never more.
Idle now lands about 3 to 5 seconds after a turn ends instead of 2
seconds into one. Verified end to end against a live worker, sampled
against the CLI's own "esc to interrupt" footer as independent ground
truth: busy for all 25s of a turn, idle 3s after it ended.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Permission dialogs, AskUserQuestion questions and idle prompts from every
session now land in a server-side inbox (web/approval-inbox.ts, one item per
session, claude-mode only) and are answerable in place: a header bell + drawer
on desktop, inline answer strips on the phone overview's NEEDS YOU rows, and
working push Approve/Deny buttons (previously dead ends, now answered straight
from sw.js with no tab open). Pending alerts survive reloads because the
frontend seeds from GET /api/approvals on init.
Answering sends the digit / Esc / prompt text through the existing tmux input
path; option digits are accepted only when they match options parsed from the
captured pane frame, and the answer path re-captures the pane first so a
dialog that already left the screen refuses with 409 instead of typing into
the composer. New elicitation_complete / elicitation_response hook matchers
resolve question items the moment they are answered in the terminal;
refreshStaleCodemanHooks heals existing cases.
Verified end-to-end against a live claude session: a real AskUserQuestion
dialog parsed into 5 option buttons and was answered from the drawer.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Version-gated fail-closed at 2.1.224 (the cross-session-messaging release,
flag presence verified against that binary): an unknown or older CLI yields
a spawn command byte-identical to before, because claude aborts startup on
an unknown option and that would kill every session spawn. The value is
allowlist-sanitized ahead of the double-quoted interpolation, and only the
local command carries the flag; docker/remote builders never see it since
their CLI is not the probed binary. Verified E2E on an isolated instance:
cmdline shows --name, ListAgents lists the session name, replies arrive
tagged from-name.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude Code v2.1.224+ gives sessions ListAgents/SendMessage and a per-session
inbox socket. Codeman's claude workers are ordinary local Claude Code sessions,
so the agent skill now teaches task delivery and result collection over
messaging where available (multi-line exactly-once messages, mid-turn steering,
latched replies), with the HTTP primitives keeping spawn, readiness,
synchronization, liveness and delete, and a bounded fallback to the HTTP
recipes whenever the feature is absent. All mechanics verified live against
claude-cli 2.1.226.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Measured on a live claude worker: `GET /api/v1/sessions/:id` reported
`status: "idle"` while the worker was mid-turn and actively producing output, with
`lastActivityAt` equal to the moment of the call. The skill already warned that a
worker which dies inside its pane also reads `idle`, so the field is unreliable in
both directions and nothing an agent does should depend on it.
Synchronize on `stop` via send-and-wait or on an output marker. To judge from
outside, sample `terminal?tail=` twice a few seconds apart: a changing buffer is the
only cheap positive proof a worker is still working. `wait?until=exit` stays the
death check.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Second review round on the #241 follow-ups. Three defects in my own previous commit,
each reproduced before and after.
1. The temp path was shared between runs (`${dest}.tmp`), so two concurrent runs
fought over it: 4 of 4 concurrent pairs had one run die. Worse than a crash, a
sibling's cleanup landing between the esbuild and the alias append makes
`appendFileSync` CREATE the file, so the rename publishes a bundle-less file
containing only the alias tail, which still satisfies the content check and
would be blessed by the cache forever. The name now carries the owning pid.
8 concurrent pairs afterwards: no failures, no strays, aliases intact.
2. The content check only covered the bundle, so a truncated xterm.min.js with a
fresh mtime stayed truncated. This script can no longer produce one, but
postinstall.js writes the same directory in place, so a Ctrl+C during
`npm install` does, and a 200-byte xterm.min.js means `Terminal` is undefined
and every mobile test dies on a null. A copy must now match its source byte for
byte, and a derived output must clear a floor far below the real ratios
(measured 0.97-1.00 minified, 0.51 for the bundle) while a truncation misses by
orders of magnitude. Verified: 200-byte and 50-byte poisonings both repaired.
3. The try block ended before the append and rename, so a rename failure leaked its
temp behind a raw stack. It now covers both and reports which asset failed.
Per-pid names mean a killed run's temp is never reclaimed by a later rebuild, so
startup sweeps temps whose owning process is gone, and only those: `kill(pid, 0)`
throwing ESRCH. Deleting a live run's temp would recreate the collision fix 1
removes. Verified both directions, plus SIGKILL mid-build leaving no litter. The
sweep swallows its own errors, because reclaiming litter must never fail the run:
a directory named like a dead temp otherwise crashed the whole prepare step.
Security-reviewed: no shell (execFileSync with an array, `shell` unset), every
argument from the static asset table plus a numeric pid, all writes confined to the
vendor dir under strace, `process.kill` only ever with signal 0 (and pid 0 skipped,
since to kill(2) it means this process group), no new dependencies, no network, no
eval, nothing published. The emitted browser bundle is byte-identical to the one
scripts/build.mjs ships, tail included.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Follow-ups to #241 (thanks @Lint111), from an independent review of that PR. The
script is a real fix for a real gap; these are the four defects the review found,
each reproduced before and after.
1. A wrong-but-fresh output was never repaired. The zerolag bundle is finished by a
SECOND step (the alias append), so anything landing between esbuild and the
append is permanent: the file looks complete, carries a current mtime, and the
mtime-only cache reports "up to date" forever while the suite dies on
`LocalEchoOverlay is not defined`. Reproduced by replaying #241's own two
commits: running the first and then pulling the second kept the broken bundle.
Fixed twice over, because the two halves address different cases. Builds now go
to a temp file and `renameSync` into place, so this script can never publish a
half-written output (that also covers an interrupted esbuild or copy, and two
concurrent runs). And `isFresh` verifies the bundle actually contains its alias
tail, which is what repairs a file an EARLIER version already poisoned; a rename
alone cannot fix what is already on disk.
2. Freshness compared against the entry file only, but esbuild bundles its four
siblings too, so editing overlay-renderer.ts left the suite testing a stale
overlay while reporting "up to date". Editing those siblings is exactly the
single-source workflow CLAUDE.md mandates. It now stats every `.ts` in the
package source dir. A full rebuild is ~2s, so the cache was not buying much.
3. `execFileSync('npx', ...)` passed no cwd, unlike scripts/build.mjs, so a run from
another directory missed the repo's pinned esbuild and would fetch an unpinned
one from the registry. Both calls now pass `cwd: ROOT`.
4. Every invocation in test/mobile/README.md was a bare `npx vitest`, which skips
the `pretest:mobile` hook npm only fires for `npm run test:mobile`, so the
documented commands all bypassed the fix. Rewritten, with a note on why.
Also: an esbuild failure printed a raw stack; it now names the asset and its input,
matching the missing-input message. And the header comment no longer implies the
vendor dir is always empty: scripts/postinstall.js already writes these same seven
outputs, so what this script adds is freshness and independence from install time.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Three gaps found while auditing the agent skill.
`codeman skill install` / `uninstall` had no tests at all, including the linked-case
resolution that shipped in 1.14.2 with nothing guarding it. Covered now: global target
resolution, `--case` resolving through linked-cases.json, `--case` falling back to the
cases dir for an unlinked name, a missing or malformed registry degrading to the
fallback instead of throwing, and a nonexistent case being rejected. `resolveSkillTarget`
called `process.exit(1)` for a missing case, which would have killed the test runner, so
the pure resolution is split out and exported; CLI behavior is unchanged.
The `POST /api/sessions` injection call site was never exercised, because the shared
route mock hardcoded the gate off. The mock's gate is overridable per test now (default
still off, since other tests rely on that), and there is coverage that the path injects
when the setting is on, does not when it is off, and is claude-mode gated.
Nothing guarded skills/codeman/reference/endpoints.md against drifting from the routes
it documents, which is how it drifted in the first place. A static guard parses the
endpoints out of the markdown and asserts each is really registered, tolerating the
/api/v1 alias and path params.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
README.zh-CN.md taught a recipe that cannot work: its programmatic-input example had no
trailing `\r`, so Enter was never sent and the prompt sat unsubmitted forever, and its
read step used `/output`, whose `textOutput` is always empty for interactive tmux-backed
sessions. A reader following the Chinese README walked into both of the silent failures
the English one warns about. Its agent/automation section is now brought in line with
README.md: the `\r` rule and every example that needs it, and the correct read path.
CLAUDE.md's "Single-line prompts only" gotcha described the newline restriction but
never mentioned that input must end with `\r` or Enter is never sent, which is the most
common silent failure when driving the API.
docs/agent-control-plan.md asserted as still-open several things that shipped in 1.14.1
and 1.14.2 (the wait endpoints, the packaged skill, the install CLI, agentSkillEnabled).
The status header and the stale bullets now match reality; the historical design content
is untouched, since the document is a record.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Two ways the injection could go wrong quietly.
`installAgentSkillInto()` wrote each file with a bare `writeFile`, no lock and no
temp+rename, while every sibling mutator in hooks-config.ts goes through
`withSettingsLock`. Two Claude sessions created concurrently in one repo both wrote the
same ~16KB SKILL.md, and any reader loading it mid-write could observe a truncated
file. Writes now go through a temp+rename helper under the same lock the neighbours
use, so a reader sees either the old file or the new one.
Both server call sites discarded the outcome with `.catch(() => {})`, so the two
refusal results were invisible: `foreign` (a user-authored skills/codeman is present,
so we declined to touch it) and `symlink` (the skill dir or its parent is a symlink, so
we declined to write through it). Turning `agentSkillEnabled` on, seeing nothing appear
and having no way to find out why was the reportable-as-a-bug outcome. Refusals are now
logged with the path and what to do about it. The boring outcomes stay silent, since
they happen on every session create. Injection remains best-effort: a refusal or a
thrown error still cannot fail session creation.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The readiness gate matched `bypass`, which is the status bar of ONE permission mode.
`buildPermissionArgs()` also spawns `--permission-mode auto`, `--allowedTools` and
plain `normal`, and the mode is not exposed on `GET /api/v1/sessions/:id`, so an agent
cannot know which token to expect. A non-default worker was therefore reported broken
after burning the whole ladder.
Measured one pane per mode against claude-cli 2.1.226:
--dangerously-skip-permissions -> "bypass permissions on"
--permission-mode auto -> "auto mode on"
--allowedTools Read,Grep -> "don't ask on"
(none, normal) -> "don't ask on"
--permission-mode plan -> "plan mode on"
Every one ends `(shift+tab to cycle)`, so `shift+tab` is the single space-free token
that means "the composer is up" in every mode, and it is what the ladder matches now.
Verified live end to end on a virgin case: stage 1 misses while the trust dialog is up,
stage 2 accepts it, stage 3 matches in 623ms.
⚠️ `shift+tab` contains a `+`, so it only works through `--data-urlencode`. In a
hand-built query the `+` decodes to a space and the server searches for `shift tab`,
which never appears; the response echoes `match: "shift tab"`, which is how to spot it.
Measured both ways. The stage-4 fallback (make the worker echo a split token, proving
readiness by answering rather than by chrome) stays as the last resort, and is now also
verified live: it matched in 2.5s, with the token surviving the space-less TUI intact.
Also portable ANSI stripping: the read pipelines used `sed 's/\x1b...'`, and BSD sed
(the macOS default) has no `\xHH` escape, so on macOS the strip silently removed
nothing and handed the agent raw ANSI. They now build a real ESC with `printf`.
And endpoints.md gaps: the `FORBIDDEN` 403 row and which auth responses are plain text
rather than the JSON envelope, the input size cap, the undocumented `killMux` parameter
on DELETE, and the fact that zero/negative/non-integer timeouts are rejected with a 400
rather than clamped.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Codeman maintainer