Fifteen review findings on the dsh mode, the serious ones first:
- Multi-user: DEEPSEEK_BASE_URL joins the owner-clamped env keys.
_configureDeepSeek() forwards the SERVER's own DEEPSEEK_API_KEY into
every dsh pane and applyEnvOverrides() lands after it, so a non-granted
owner who could redirect the base URL would have the operator's key sent
as a bearer credential to a host of their choosing.
- Wait registry: until=stop/blocked is refused on docker and remote-SSH
dsh sessions (new deepSeekBridgeUnreachable fact in sessionHookOptions).
The HERDR triple is set via LOCAL tmux setenv, which crosses neither
docker exec nor ssh, so such a session can never post a hook event and
the wait burned its whole timeout on every turn.
- Approvals: a dsh item is an ALERT, not an answerable card. The answer
route refuses (the '1'/Esc keystrokes are Claude-dialog-shaped and the
option parser cannot read a third-party TUI's frames, so an answer was a
blind keystroke into a foreign composer), and the push notification
carries no Approve/Deny actions for dsh sessions.
- Status shim (v3): --seq is forwarded and the server drops stale retried
reports inside a 60s window (the TUI retries with backoff, so a retried
'working' could land after 'blocked' and resolve an approval whose
dialog was still on screen); 4xx responses exit 0 instead of retrying,
so one misconfigured session cannot feed the auth rate-limit bucket
until the hook endpoint 429s for the whole instance.
- Web-UI server: concurrent starts are serialized through a lock (two
racing POSTs used to pick the same port and orphan the winner), and the
readiness poll / timeout paths only clear or stop the singleton while it
is still theirs. First click actually opens the tab now
(refreshWebviews, not the nonexistent loadWebviews). DELETE
/api/deepseek/web requires the privileged grant in multi-user mode.
- Cron: deepseek jobs run the same two-part launch gate as the HTTP
create paths (impl moved into the resolver so all three share it) and no
longer stamp a Claude default model on the session.
- Parity sweeps: quick-start's docker branch rejects deepSeekConfig like
the remote branch; the Ralph auto-enable list gained deepseek;
HookEventType gained agent_working; the phone overview run menu filters
managed webview records like the desktop menu.
- install.sh: the dsh identity probe closes stdin (under curl|bash a
child that reads stdin eats the rest of the script), bounds the exec
with timeout where available, and is memoized to one scan per install.
- Welcome screen: .welcome-btn-deepseek styled in the #4d6bfe brand
identity (it rendered as an unstyled UA-grey button); stale markup
comment about the web shortcut rewritten; clamp docs updated.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Four review findings on the worker-transcript feature:
- Docker and remote-SSH dsh sessions now keep the pane segmenter: their
transcripts live in the container's / remote host's own ~/.dsh, which
the local reader can never see, so the transcript path returned
'nothing said yet' forever and an agent polling such a worker starved
on an answer that existed. Gated on !session.docker && !session.remote
(statically pinned) and documented in the integration guide.
- last-response reads are memoized on (path, mtime, size, blocks): the
skill's last_text polls once per second, and each poll decompressed and
reparsed the whole file on the event loop even when nothing had been
appended. An unchanged poll now costs one stat.
- The pairing ladder's comment claimed /new is served by step 2; in truth
the boot-window transcript wins for as long as it exists (deliberately:
preferring newest-eligible would hand a worker its busier sibling's
reply). The comment now states the real tradeoff instead of the
aspirational one. Same for decodeZstdFrames' 'skipped' wording — a
corrupt frame truncates the decode there, which is the safe behavior.
- stripReasoningPrefix no longer runs on user prompt text, so a prompt
containing a literal </think> renders whole in blocks view.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Three review findings on the detailed-rows feature, all in its edge cases:
- The App Settings width select consulted the handheld defaults blob
(tabRailWidth: 256) BEFORE the rich-aware default, which the renderer
never reads — so a tablet's unsized rich rail rendered 320 while the
dialog said 256, and a routine Save persisted the 256 (below the 288px
tight threshold, permanently). The chain now mirrors
applyTabRailWidth()'s actual resolution.
- _setTabRailWidth() re-rendered on a compact flip but never re-ran
applyTabWrapSettings(), the one owner of the folder line, whose railRich
input reads the compact class this function just toggled. A rich rail
dragged below 240px kept emitting folder rows — persistently, for a
stored width < 240, since the boot wrap pass runs before the class is
first applied. The wrap pass now re-runs on the flip, with exactly one
render either way.
- Both reset affordances (handle dblclick, Enter on the handle) reset to
the hardcoded 256 even on a rich rail, landing it below the tight
threshold; both now resolve the rich-aware default (320), via a new
optional defaultWidth input on resolveTabRailKeyboardWidth().
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The agent skill could spawn a worker in any mode, but it could only
DRIVE a claude one: every other CLI has neither a real end-of-turn
signal nor an answer to read, so the recipes route them through output
markers.
dsh has both halves now -- its harness reports idle/working/blocked to
Codeman, and the previous commit reads its transcript -- so it joins
claude as a mode the four verbs work on unchanged. `spawn_workers alpha
beta:deepseek` is a mixed fleet in one call, and `sendwait` / `last_text`
/ `delete_session` need no per-mode variant.
Preamble 1.20.0 (SKILL.md's §0 heredoc regenerated from it):
- `spawn_worker` grows a deepseek branch that gates on the harness
composer. ⚠️ Readiness there is NOT the stop signal: the harness
reports idle at BOOT ~300 ms before its composer paints (measured
2.26 s vs 2.56 s after spawn), so a send-and-wait fired straight after
quick-start resolves on the boot edge, reports a turn that never ran,
and strands the prompt in a pane not yet taking input. Waiting for the
composer also spends that edge, since signals are edge-triggered.
- `spawn_workers` takes `name[:mode]`, so a mixed fleet stays one
concurrent call. Case names still have to be unique -- the mode never
disambiguates two workers that would share a directory.
- `sendwait` asks for `wait:"stop,exit"` instead of the `wait:true`
default set. That set also carries `idle`, which for an external CLI is
inferred from output stabilization: on a dsh worker whose TUI repaints
rarely, the re-wait resolved in 0 ms with `signal:"idle"` on a turn
with three minutes left to run. It also makes a wrong mode loud -- the
modes that cannot deliver `stop` answer 400 before writing anything,
instead of resolving on a flap.
- The self-heal resend carries `delivered:true` forward. The resend is a
tagged duplicate, so the server truthfully reports `delivered:false`
about a write it skipped, and §1's cleanup then read a completed turn
as an undelivered one and kept a finished worker forever.
- dsh workers spawn with the permission posture the Run button sends,
because the harness default still asks and a worker parked on an
approval row cannot finish a fan-out. The multi-user clamp still
applies.
Docs: a worked dsh flow in recipes.md, readiness and the signal rules in
verbs.md, and the corrections this makes necessary -- `stop`/`blocked`
are no longer claude-only, and `last-response` is no longer permanently
empty for deepseek. The integration guide gains a section on reading a
session back and driving one as a worker; its web-UI section was also
stale (that server moved out of a shell session).
The static guard that keeps those lists from naming some external CLIs but
not others is extended rather than exempted: it now knows the three real
classes inside that family (no transcript, no hook signals, and the
positive twin -- the modes whose answers can be read), with the hook class
derived from `hooksAvailableForMode()` so the predicate and the prose
cannot drift apart. Any other partial list still fails, and a new backend
belongs to none of the classes until someone says so.
`GET /api/sessions/:id/last-response` is how an agent (and the Response
Viewer) reads what a worker said. DeepSeek was falling through to the
pane segmenter with the other external CLIs, which for this mode is not
merely coarse but wrong: dsh-TUI paints a full-screen splash, so a
`last-response` call on a fresh dsh session answered with its ASCII-art
logo -- and anything polling for a worker's first reply reads that as a
reply.
dsh does not belong in that group. It writes a structured JSONL
transcript per session, so read it. Four things in that file shaped the
reader, all measured against real transcripts on disk:
1. dsh appends ONE ZSTD FRAME PER WRITE, and Node's zlib zstd decoder
(one-shot and streaming alike) stops at the first frame end: a real
56-line transcript decoded as 1 line / 158 bytes -- the session header
alone, i.e. a silent truncation that reads as "nothing said yet"
forever. `zstdFrameRanges()` walks frame and block headers to find
exact boundaries; splitting on the 4-byte magic would corrupt
everything after a magic sequence occurring inside compressed data.
zstd is resolved at RUNTIME because it landed in Node 22.15 while the
project floor is 22.0, so an older Node keeps the pane behaviour.
2. Every turn also records a plugin-sourced `user/message` (the runtime
context snapshot), which must not render as the user's own words.
3. A turn that ends in an error carries the provider's message; it is
surfaced as `Turn error: …` (and a non-error early stop as
`Turn ended: …`) rather than as an empty string, which an agent reads
as "still thinking" through fifteen polls.
4. Reply text is assembled per (turn, step): a finalized message wins and
the streamed deltas fill in only for a step that never finalized, so a
partial answer is readable mid-turn and never doubled. "Finalized" is
tracked as a set of steps rather than as non-empty text, because a
step whose whole reply was reasoning strips to '' at the `</think>`
boundary and would otherwise resurrect the raw deltas in its place.
Session-to-transcript pairing is by the transcript's own header `cwd`
plus a boot window against the session's createdAt, never by
reproducing dsh's directory mangling (already two forms on disk) and
never by newest-mtime alone -- mtime alone handed a freshly spawned
worker its predecessor's answer in the same case directory.
An empty result still wins over the pane; only a Node that cannot decode
zstd falls back to it.
Clicking "DeepSeek web UI..." opened two tabs: the web tab asked for, and a
shell tab running the server next to it. The shell was deliberate - the server
lived in an ordinary session so it was visible, scrollable, killable and died
with its tab, and nothing new had to supervise a long-lived HTTP server. That
reasoning was sound and the result was still wrong in use: opening a dashboard
should open one tab, and after the first launch the terminal is pure noise.
The server moves to a background child process owned by a new
`src/deepseek-web-server.ts`, behind `POST /api/deepseek/web`. What the session
gave away for free is now explicit, which is most of the module:
- Exactly one server. A second click reuses the running one instead of racing
it for a port; the session flow could not do this at all, because two clicks
were simply two sessions.
- Restarted when the requested authority changes. `--trusted-host` fences dsh's
own /api against the browser authority, and a Codeman reachable at both
loopback and a tailnet name has two. Reusing a server fenced for the other
origin renders a page whose every call 403s, which reads as a broken
dashboard rather than a misconfigured one, so a mismatch restarts instead.
- Killed on shutdown. The child is detached so its whole plugin tree can be
signalled at once, which also means it would outlive Codeman and hold its
port against the next start - the exact EADDRINUSE this feature already got
wrong once.
- Boot output captured and returned. With no shell tab there is nowhere else
for a stack trace to land, so a failed spawn reports its own tail.
The endpoint is fenced at the same bar as the profile installer and for the
same reason: booting a dsh profile executes the plugin code in it, so this is a
privileged action even though it reads as "open a page". `authority` comes from
the client (`location.host`) because only the browser knows which origin is in
play, and it is regex-confined at the schema boundary - defence in depth behind
the argv-array spawn, admitting host:port in the shapes a browser authority can
take and nothing readable as a second argument.
`GET /api/deepseek/web-port` is gone; port selection moved into the supervisor,
which is the thing that knows whether a server is already running. The two
client-side probe helpers went with it, since the server now owns the wait.
Verified over the tailnet authority end to end: no session is created (session
count unchanged, one tab), the server runs on 3081 beside the user's own dsh
web on 3080, status reports the tailnet authority, and the proxied dashboard
renders with zero 4xx. Full gate green (6148 passed, +6).
The `Run > DeepSeek web UI...` shortcut failed three ways at once against a real
install, and the three are independent.
1. It hardcoded `--port 3080`. That is dsh web's OWN default, which makes it
precisely the port a DeepSeek user is most likely to be serving on already,
so the launch died with EADDRINUSE against the user's own server. The port
now comes from `GET /api/deepseek/web-port`, which walks 3080..3119 for a
free loopback port by BINDING it (a connect probe cannot tell "free" from
"listening but not answering yet").
2. It opened the tab unconditionally. The crashed server left a saved dashboard
pointing at nothing, with the failure only visible in a shell tab nobody had
a reason to look at. The launch now polls the existing webview probe until
the URL answers, and on timeout reports the error naming the shell tab
instead of persisting a dead dashboard.
3. The saved tab was untrusted, so the frame was sandboxed without
`allow-same-origin` and the dashboard was broken twice over: the dsh
client-runtime reads `localStorage` while loading its plugins and died there
("the document is sandboxed and lacks the 'allow-same-origin' flag"), and an
opaque-origin frame sends `Origin: null`, so dsh's own trust fence 403'd
every `/api` call no matter which authority `--trusted-host` named. Passing
`location.host` only means anything once the frame actually carries that
origin, so `--trusted-host` had never once done its job. The managed tab is
now created `trusted: true`.
That trade is real and deliberate: a trusted proxied frame is same-origin
with Codeman and can reach Codeman's API. It is defensible only because this
dashboard is an agent harness Codeman just started itself, on loopback, which
can already run code as the user. It is not a precedent for trusting
third-party dashboards, which is why it is set at this one call site rather
than defaulted.
Separately, the shortcut listed its own dashboard twice: once as the menu entry
that starts it and once as the row that entry had written on the previous click.
Webviews now carry an optional `managed` marker, managed rows are filtered out
of the saved-dashboard list, and a relaunch repoints the existing row rather
than stacking one dead dashboard per restart (which the per-launch port would
otherwise guarantee). `managed` is declared in the schema because a plain
`z.object` strips undeclared keys, so an undeclared marker would never survive
the round trip.
`DEEPSEEK_WEB_PORT` is gone from constants.js; its doc comment asserted that a
hand-started `dsh web` and the shortcut "land on the same place and share one
saved tab", which is the bug stated as a feature.
Verified on a real install with the user's own `dsh web` holding 3080: the
shortcut takes 3081, the server answers, exactly one DeepSeek entry shows in the
run menu, and the proxied dashboard renders its workspaces and completes its own
API calls (the previously-403'd `api/settings.describe` now succeeds). Full gate
green (6142 passed), typecheck/lint/format/public-assets clean.
Two review nits on the vertical rail's detailed rows.
1. The tab-rail-tight rule (below 288px) hides `.tab-meta-created`, and its
comment claimed the value "survives in the row's title attribute either way".
It did not: the only title carrying it lived ON that element, and a
`display: none` element has no hover target, so the created stamp was not
shrunk but gone with no way to ask for it. Rather than just correcting the
comment, `_sidebarRichMetaHTML()` now puts BOTH absolute stamps on the
`.tab-meta` line itself, so the pill and the gaps around the stamps remain as
hover targets. An item's own title still wins where the item is visible.
2. applyTabOrientation() decided whether applyTabWrapSettings() had already
re-rendered by comparing `_tallTabsEnabled` before and after. That reads an
UNDEFINED previous value as "it rendered", but applyTabWrapSettings()
deliberately renders nothing on its first call ever (it only establishes the
baseline: `prevTallTabs !== undefined && prevTallTabs !== showFolder`). So on
a first call that also flips the folder row, neither function rendered and the
rows stayed stale. Reachable when the pre-paint script throws and leaves the
layout attributes on their catch-branch fallbacks for applyTabOrientation() to
correct. The guard now mirrors applyTabWrapSettings()'s own condition.
Both new tests were run against the unfixed code first and fail there, which is
the only thing that makes them regression tests. (The third, "does not render
twice", passes either way by design: it pins that fix 2 did not introduce a
double rebuild.)
Verified in a real browser against a live server with two sessions, driving the
narrowing through _setTabRailWidth() the way the resize drag does: at the 320
default the row reads "CREATED 2m ago · IDLE <1m" with the created element
displayed; at 256 the tight class is on, the created element computes to
display:none, the visible text drops to "IDLE <1m", and the meta line's title
still reads "First created: ...". At 220 the compact threshold drops rich rows
entirely. Screenshots confirm no truncation artifacts in either state.
Full gate green (6104 passed), typecheck, lint, format, frontend-syntax and
public-assets all clean.
The three smaller review nits, plus the first real test coverage for the status
shim (it had none: it is emitted as a STRING, so tsc never sees it).
1. The shim was written with a plain writeFileSync. The TUI can be exec'ing that
exact path while an upgraded Codeman refreshes it, and a reader catching a
half-written file gets a syntax error, exits non-zero, and is retried four
times per state change for a file that will never parse. Now temp + rename
(atomic within the directory), with the temp chmod'ed before the rename since
writeFileSync's mode only applies on create, and removed if the write throws.
SHIM_VERSION bumped to 2, because SHIM_SOURCE changed and an existing v1 shim
would otherwise keep matching the embedded marker and never be refreshed.
2. The pane-id comment claimed the ambient env "cannot be spoofed by an argument
the agent itself could influence". The agent runs IN that pane and can invoke
the shim with CODEMAN_SESSION_ID unset and any argv it likes. It buys nothing
it did not already have (the hook-secret file is readable from the same pane,
so it can POST /api/hook-event directly), but the comment read like a security
boundary. Rewritten to say what the preference actually buys: correct
attribution when a TUI mangles or re-uses the pane argument. Accidents, not
adversaries.
3. classifyProfile() folded the directory name into the same haystack as the
bundles, but only the TUI arm could match a bare name, so a stock profile
whose package.json has no dsh.profile.bundles (hand-edited, older layout,
mid-install) classified as `unknown` -> launchable -> eligible as the DEFAULT
pick, which is exactly the pane-dies-on-arrival failure the two-part
availability gate exists to prevent. The stock names are now a LAST-resort
fallback consulted after the bundle patterns, so real bundle evidence still
wins over a name the user chose. The loose `tui` arm gained word boundaries:
it decides which profile boots by default, and matching the middle of
`intuition` is not a rule anyone could predict.
New test/deepseek-status-shim.test.ts runs the generated script the way the
harness does -- real node process, real argv, real env, real listener -- and
covers the exit-code contract that makes the retry behaviour safe: mapped states
post and exit 0, an unknown verb or unmapped state exits 0 WITHOUT posting (a
non-zero there would be four HTTP requests per state change forever), a rejecting
server or an unreachable one exits non-zero so the caller retries, the hook secret
is read at execution time, and `node --check` parses the file (a template-literal
typo in SHIM_SOURCE is invisible to tsc).
Trap worth recording, hit while writing it: the tests must spawn the shim
ASYNCHRONOUSLY. The listener lives in the test process, so spawnSync blocks the
event loop that has to accept the connection, the shim waits out its own 1500ms
socket timeout and exits 1, and it reads exactly like a broken shim (measured:
Socket._onTimeout in its --trace-exit output, server logging nothing).
Verified: full gate green (6142 passed, +10), typecheck/lint/format clean.
Three review findings on the DeepSeek Harness mode, plus one the third exposed.
1. The multi-user clamp was bypassable by a sibling field on the same request.
clampExternalCliBypassForOwner() clamps deepSeekConfig.permissionMode, but
DSH_* is an allowlisted envOverrides prefix and applyEnvOverrides() runs AFTER
_configureDeepSeek(), so a non-granted owner sending
envOverrides.DSH_PERMISSION_MODE landed last and won. Measured on an isolated
instance: a session created with permissionMode "read-only" and that override
ran with DSH_PERMISSION_MODE=danger-full-access in its pane.
Every other CLI's bypass is a command-line flag reachable only through the
per-CLI config, which is why the config clamp alone is the whole gate for
them. clampEnvOverridesForOwner() adds the env-var half: for a non-granted
owner it DROPS DSH_PERMISSION_MODE and DSH_HOME (dropping falls through to
what _configureDeepSeek() exports, i.e. the clamped value). DSH_HOME is on
that list because it aims the launcher at a profile tree whose plugin code
runs at boot, before any approval row can apply. Verified end to end in real
multi-user mode: a non-granted user sending both now gets workspace-write and
no DSH_HOME, while an unrelated DSH_TELEMETRY_MODE passes through untouched.
2. POST /api/deepseek/install-profile could hang forever. spawn's own `timeout`
signals only the direct child, and a plugin install fans out into
package-manager children that keep the inherited stdio pipes open, so `close`
never fires and the held-open request leaks with no route-level deadline.
Reproduced: with a 1.5s built-in timeout the promise was still unsettled after
6s and both fan-out children were alive. Now detached: true plus negative-pid
SIGTERM/SIGKILL, the same escalation runGit() uses for the same reason, with a
last-resort reap for a grandchild that escaped the group. Same probe after the
change: close fires, direct child and both grandchildren dead.
3. hooksAvailableForMode() promised more than a dsh session can deliver.
deepSeekConfig.statusReporting: false disarms the HERDR_* export, and that
triple is the only reason a dsh session posts hook events, so `until=stop` was
accepted and then blocked for the caller's whole timeout: the exact
infinite-wait-dressed-as-a-timeout the predicate exists to prevent. It now
takes HookCapabilityOptions and every call site passes sessionHookOptions(),
with the deepseek arm reading `!== false` so a forgotten one degrades to the
old behaviour. The refusal names the setting rather than saying "no Claude
Code hooks", which would send the caller hunting a bug that is really a
setting they chose. Profile conformance stays unknowable at request time and
is documented as such. The stale "True for `claude` and nothing else" docblock
is corrected.
4. Exposed by (3): hooksAvailableForMode() was doing double duty as "is this a
claude session". Read My Mind (POST /api/sessions/:id/readmymind) and intent
capture read Claude's own transcript, and adding deepseek silently widened
both to a mode that has none. They compare mode === 'claude' directly now, and
a static check pins them there.
Verified: full CI gate green (6132 passed), typecheck/lint/format clean, and the
wait-signal gating exercised against a live server with a real dsh 0.1.1-rc.2 --
bridge off plus explicit until=stop is a 400 naming the setting, bridge off with
no `until` still 200s on idle/exit, bridge on accepts stop.
The vertical tab rail (tabOrientation 'vertical') listed names and nothing
else, while the rich sidebar and both home screens already answered the
question a docked column exists to answer: which of these sessions wants me
next, and how long has it been like that. The rail is a docked column too, so
it now draws the same row.
- New per-device setting tabRailDetail ('rich' | 'simple', default rich),
App Settings -> Appearance -> Tabs, in SettingsUpdateSchema + displayKeys and
stamped as data-tab-rail-detail by the pre-paint script, so a detailed rail
does not flash through simple rows on every load.
- ONE gate for both vertical surfaces: isRichTabRows() =
isSessionSidebarRich() || isTabRailRich(). The row model, the markup and the
20s in-place clock are the existing rich-sidebar ones, classified by
_mobileOverviewState/_mobileOverviewSince, so the rail, the sidebar, the
desktop home rail and the phone overview cannot disagree about what
"working" means or which stamp measures it.
- Detail rides on its OWN attribute, exactly as the sidebar's does, so every
existing [data-tab-orientation='vertical'] rule keeps matching both variants
untouched. A flip of detail ALONE still forces a full render (the stamps line
is emitted by the row template, not toggled by CSS) and re-runs
applyTabWrapSettings(), which owns the folder line and is now rail-aware.
- CSS: every rich paint rule gains a rail twin as a COMMA-GROUPED selector,
never :is() - an :is() list takes its most specific argument, which would
lift the sidebar arm from (0,3,1) to the rail's (0,5,1) and let these rules
outrank things they never used to.
- Width is why there are thresholds. At 256px the stamps line ellipsizes
mid-word, the same reason the rich sidebar is 300px, so a rail that has never
been sized defaults to 320 (RICH_DEFAULT_WIDTH, the existing Wide preset,
which also keeps the settings select on a named choice). A width the user has
chosen is never overridden: below 288px the created stamp is dropped rather
than truncated (tab-rail-tight, CSS only) and below 240px the rows go back to
simple (tab-rail-compact, which re-renders).
- The rich clock is armed and disarmed by applyTabOrientation() as well as
applySessionListLayout(); a leaked interval would rewrite stamps in a list
that no longer has any.
Also fixes a data-loss bug in the inline tab rename that predates the rail and
reproduces in every layout, header strip included: Escape set the input to ''
and blurred it, and the blur handler commits - so cancelling a rename PUT an
empty name, and the tab fell back to its folder label (measured against a live
server: ["rail-alpha","","rail-gamma"]). Escape now calls cancelRename(), which
invalidates the edit so the blur that follows the input's removal is a no-op.
Tests: rail-detail gate, the three ways it turns back off (simple, compact,
horizontal), sidebar-wins, render-on-detail-flip and the plumbing/CSS guards in
test/session-list-layout.test.ts; the rename cancel in test/inline-rename.test.ts
(browser suite), pinned by running it against the old code first. Verified live
against a real server on an isolated instance: detailed/simple/compact/header/
sidebar variants, click-select, the ... menu, inline rename, Alt+N, the in-place
stamp tick and a full settings-picker round-trip including reload.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Adds `mode: 'deepseek'` alongside claude/shell/opencode/codex/gemini/
antigravity/pi/grok, plus a shortcut that opens the harness's own browser UI
as a Codeman web tab.
DeepSeek is wired unlike its siblings in three ways, each of which is the
reason for a design decision rather than an accident:
1. The agent is a PROFILE, not the binary. `dsh` is a launcher over
$DSH_HOME/profiles/<name>, and DeepSeek ships only `web`, `headless` and
`base` -- the interactive terminal front door is always a third-party
plugin. So availability is two questions: `isDeepSeekAvailable()` (binary)
and `isDeepSeekRunnable()` (binary AND a pane-capable profile). The Run
button gates on the latter, because reporting only the binary would spawn a
pane that dies on arrival. When the binary is present but no profile is,
the run menu offers to install one (POST /api/deepseek/install-profile).
2. The permission switch is an env var, not a flag. The harness has no
command-line permission option; its sandbox/approval rows read
DSH_PERMISSION_MODE (read-only / workspace-write / danger-full-access).
Exported via `tmux setenv`, never on the spawn line. Absent = the harness's
own workspace-write, which still asks, so the multi-user clamp is the
only-if-sent branch and clamps to workspace-write, never read-only.
3. It is the only non-claude mode that passes hooksAvailableForMode(), and it
earned that. The terminal front door reports idle/working/blocked to a
supervising process over a generic env-gated contract; a generated shim
(deepseek-status-shim.ts) makes Codeman that supervisor and forwards each
report to /api/hook-event as stop / agent_working / permission_prompt. So a
dsh session gets definitive respawn triggers, real wait-endpoint signals and
real Approvals Inbox items instead of output-stabilization guesswork.
`agent_working` is new (157th SSE constant) and joins
APPROVAL_RESOLVING_EVENTS so a dialog answered in the terminal clears its
alert at once.
The resolver needs the strictest identity probe of the family: `dsh` is not
merely a squattable npm name, Debian ships an unrelated `dsh` (dancer's shell),
so `dsh --help` must print the harness's own banner before a candidate is
handed a spawn line.
Model is deliberately not a session field -- it is a composition entry in the
profile's config tree. Env allowlist gains DSH_* and DEEPSEEK_* only; provider
keys named by a settings-file `apiKeyEnv` stay out, which is pi's
34-provider-key problem in a new shape.
Verified live against dsh 0.1.1-rc.2 and @deepseek-harness-tui/dsh-tui: the
status endpoint's two-part answer, the no-profile refusal, the profile
bootstrap, a real session whose pane runs `dsh --profile dsh-tui` with the
permission mode injected via setenv, and the full status bridge -- a
send-and-wait returned signal "stop" from a real turn, and blocked/working
created and cleared an Approvals Inbox item.
Docs: docs/deepseek-integration.md (guide), docs/deepseek-integration-plan.md
(decisions + honest gaps). Tests: test/deepseek-mode.test.ts,
test/deepseek-cli-resolver.test.ts.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The agent-image grok step copied /root/.grok/bin/grok onto /usr/local/bin/grok
with cp -L. Newer versions of xAI's install.sh already create
/usr/local/bin/grok as a symlink to that same binary, so the copy failed with
'same file' and the --no-cache rebuild died at the grok layer (2026-08-24).
Stage the copy under a temp name, drop whatever the installer left at the
destination, then move into place - correct against both old and new
installers.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Post-merge follow-ups from the deep review of #334 and #335, so they ship in
the same release as the features.
Tab-layout foundation (#335):
- PUT /api/session-order drops unknown/foreign ids again instead of 400ing
the whole write, in both the owner and the admin path (single-user requests
are the synthetic admin, so that path is the one the browser hits). The
frontend debounces its reorder push and swallows errors, so a session
deleted inside the debounce window silently cost the user the entire
reorder - and the endpoint sits on the stable /api/v1 surface, where the
pre-layout server merged leniently.
- A failed mux restore no longer locks explicit deletions into 500s for the
process lifetime: runSessionDeletion and webviewDeleted degrade to
best-effort without layout coordination, while the automated stale sweep
(runStaleSessionCleanup) stays fail-closed.
- sse-events doc comment: no 'suppressed' hook event exists; hooks stay 8.
- registerSessionWithLayout resolves its owner through ownerLayoutKey()
instead of a hardcoded '@single'.
Vertical rail (#334) - all rail-awareness gaps in sidebar-only predicates,
unified behind the new _isVerticalTabList() (sidebar OR rail):
- Drag-reorder read the insertion side from clientX in the rail, so
before/after was effectively arbitrary on vertical rows; the drag-over
indicators now draw as top/bottom edges there like the sidebar's.
- The active tab is scrolled into view in the rail (Alt+N/palette selection
used to leave the row below the fold).
- Floating subagent/ultracode windows anchor to the RIGHT of rail tabs, and
the connector redraw gates (render tail + strip scroll) cover the rail.
- Server-seeded tabOrientation is applied when the async settings load
resolves, not only at boot, so a fresh device shows the rail immediately.
- The pre-paint script stamps data-tab-orientation and --tab-rail-width
(sidebar-wins and solo carve-outs included), removing the flash of the
header strip on every vertical-mode load.
- The session name font defaults to 12px, the sidebar's historical 0.75rem
size, so installs that never touch the new slider are not restyled.
Also documents the rail in CLAUDE.md (second #sessionTabs host, mover
ordering, the axis-predicate rule) and gives tab-rail-resize.js its
@dependency/@loadorder header. Full gate green (6093 tests); the excluded
browser suite was run by hand - only the known environmental failures
(opencode/codex binaries) remain, identical to pristine master.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The button next to the file preview's close icon was Copy Content, whose
overlapping-pages glyph reads as a pop-out control - and for a PDF or any
media/binary preview it was completely dead: those branches never fill
filePreviewContent, so the click hit an empty-content guard and did nothing,
with no feedback.
There is now a real detach button that opens the previewed file in a browser
tab (raw route for PDFs/images/media/text, the server-converted PDF preview
for docx/pptx), severs window.opener by hand so a blocked pop-up stays
detectable, closes the overlay on success (which also stops any playing
media), and disarms on close so it can never open a stale file. The copy
button now toasts 'Nothing to copy in this preview' instead of staying
silent.
Verified live with Playwright against an isolated instance: button visible
and armed on a PDF preview, file-raw answers 200, clicking opens the URL and
tears the overlay down, text previews keep a working copy buffer.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The agent skill's endpoints.md is what other agents read as ground truth, and
three of its facts went stale when grok landed:
- `/api/v1/grok/status` was added to the probe list, but the sentence after it
still said only Pi's response carries `.data.version`. Grok's carries it for
the same reason (a squatted binary name), and an agent that trusts the old
wording has no way to tell a misresolved grok from an absent one.
- the `active-tools` bullet listed grok among the modes it stays empty for, then
claimed in the same breath that `isExternalCliMode` "lists only those five".
- its three source line refs had all drifted: `isExternalCliMode` is now
session.ts:174-183 (it was already wrong before this branch), the external-CLI
early return is session.ts:2261, and TEXT_COMMAND_PATTERN is
bash-tool-parser.ts:89.
CLAUDE.md and architecture-invariants.md counted modes in their Docker-cases and
Web-tabs paragraphs ("any of the five CLI backends", "never a sixth
SessionMode"). Both numbers were already stale before grok (antigravity and pi
had made it seven) and grok is now in the agent image, so the counts are gone
rather than incremented: the invariant those sentences carry is that Docker and
web tabs are not modes at all, which no number has ever helped state. The two
plan docs keep their original wording, being historical design records.
test/workflow-run-watcher.test.ts pinned its fixture's newest activity at
2026-06-14T20:06:40Z and then asked getRecentRunSummaries(100000) to return
it. That argument is MINUTES, so the window is 69.4 days: the assertion
expired at 2026-08-23T06:46:40Z and the file has failed on every branch
since, on a suite nobody had touched. The last green CI run finished at
06:47:43Z, about a minute inside the boundary, which is why it landed as a
surprise rather than a bisectable regression.
The fixture epochs now hang off a RUN_ANCHOR of Date.now() - 601s with every
offset preserved verbatim, so the parsed durations, the ordering and the
live-vs-done discriminators are all unchanged, and the recency filter is
still the thing under test. It just cannot rot again.
SessionMode gains 'grok', a first-class backend alongside Claude Code,
shell, OpenCode, Codex, Gemini, Antigravity and Pi: its own PTY, tmux
session, charcoal tab identity ('gk' badge), welcome button, run-mode
entry, cron agentType, Docker and remote-SSH command defaults, and
clone-repo Brain option. Flag surface verified live against grok 1.0.5.
Grok mixes two existing shapes and the wiring follows from that:
- Codex-shaped on permissions: the bypass switch is GrokConfig.alwaysApprove
(--always-approve, grok's bypassPermissions mode; config-level deny rules
still apply on top). The Run button sends it true, like runAntigravity(),
and clampExternalCliBypassForOwner() puts grok in the only-if-sent branch:
a bare grok spawn is grok's own ask-mode default, which is already safe,
so only a sent config needs the flag forced off. Cron needs nothing for
the same reason.
- OpenCode-shaped on rendering: grok is a fullscreen alternate-screen TUI
with mouse support, so it stays OUT of isAltScreenStripMode() and lands
on the narrow tmux-attach strip and the 'buffer' local-echo fallthrough
(unmeasured against an authenticated composer; documented fallback is the
'off' branch).
- Pi-shaped on resolution: 'grok' has npm squatters (@vibe-kit/grok-cli
also installs a grok bin), so grok-cli-resolver.ts version-probes every
candidate (grok --version, killSignal SIGKILL, VITEST-gated) and
GET /api/grok/status surfaces path AND version; GROK_VERSION_REGEX is
shared with the dependency registry so doctor and run mode cannot drift.
Env allowlist gains GROK_* plus the XAI_* vendor namespace (XAI_API_KEY is
grok's documented headless auth var), the same narrow-vendor reasoning as
GOOGLE_* for gemini. Resume is id-regexed on purpose: grok's own --resume
also matches session titles, which are arbitrary user strings that must
never reach the bash -c spawn line.
Docker: grok is not on npm, so the agent image installs it in its own step
(xAI's installer has no --dir override; the binary is copied to
/usr/local/bin and root's ~/.grok dropped in the same layer), and
credentials are seeded per-file (auth.json, config.toml, pager.toml; the
dir also holds sessions/, memory/ and the ~160MB binary). Remote SSH routes
through the login-shell wrapper like the other agent CLIs.
Verified end to end on an isolated CODEMAN_INSTANCE with grok 1.0.5
installed: /api/grok/status resolves and reports the probed version,
quick-start spawns a pane whose command line ends in 'grok
--always-approve', the real TUI renders (OAuth device screen on an
unauthenticated box), and grokConfig round-trips through state.json.
Docs: docs/grok-integration.md (user guide) + docs/grok-integration-plan.md
(decisions, verification record, follow-ups).
Tests: test/grok-mode.test.ts, test/grok-cli-resolver.test.ts, plus
extended clamp/system-routes/render-index-html/run-mode-ui/mobile-overview/
local-echo-gating coverage. npm test (the CI gate) green: 5910 tests.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
scripts/tmux-chooser.sh is deleted. codeman tui replaces it and does the
job better: sc numbered its entries globally but only accepted a single
[1-9] keypress, so sessions 10+ were listed and unselectable, and it
inferred nothing about what an agent was doing. The tui carries the
server's real states, answers permission dialogs, and leaves an attach
with one key.
install.sh no longer creates the tmux-chooser symlink or the sc alias.
It now sweeps both up instead, on update AND uninstall, so an update
cannot leave a symlink pointing at a script this version stopped
shipping. The alias removal is marker-owned: it matches the exact line
the installer wrote, so someone's own 'alias sc=' for another tool is
never touched, and it rewrites through 'cat >' so the profile keeps its
mode and ownership. Verified against three profile shapes.
BREAKING CHANGE: the 'sc' command and the 'tmux-chooser' symlink are
gone. Use 'codeman tui' (and 'codeman tui --list' / 'codeman tui <n>').
codeman tui supersedes the sc bash chooser: it reaches sessions 10+,
carries the server's real states instead of a static list, and leaves an
attach with one key. Every place that told a user to run sc now names the
tui equivalent, including the two wiki pages and install.sh's next-steps
banner. Both wiki pages also carried the wrong detach chord (Ctrl+A D;
the socket's prefix is C-b), which the tui makes moot.
Source comments that explained themselves as "the sc -l replacement" now
just say what they do. docs/tui-plan.md and CHANGELOG.md are historical
records and keep their references.
CLAUDE.md gained "any new `tmux -L` caller through `resolveTmuxSocketName()`"
but architecture-invariants.md, which that bullet points at for the
mechanism, still described only the dataPath() half. Say why the rule
exists there too: the TUI is the first non-server process to shell out
to tmux.
The sc chooser runs a plain `tmux attach-session` and binds nothing, so
F1 does not detach from it; only `codeman tui`'s attach claims that key,
and only for its own duration. The line it replaced was wrong too (the
socket's prefix is C-b, not C-a), so name the real chord and say which
command gives you the single key instead.
The guide and the README both said to detach with `Ctrl+B D`. Beta testing
proved that wrong twice over: tmux binds lowercase `d` to `detach-client` and
capital `D` to `choose-client`, and even the correct letter fails for anyone who
keeps Ctrl held, because that sends `Ctrl+D`, which tmux leaves unbound. A
tester followed the documented instruction, stayed attached, and exited the
agent to escape.
Both now say `F1`, and the attach section describes what actually happens: the
session strip across the top of the pane, `Alt+1`..`Alt+9` switching without
returning to the dashboard, and `r` to resume a session whose pane has died.
Also corrected: `1-9` switches rather than jump-attaches, `x` confirms with `y`
rather than a typed name, and a new session opens straight into its pane.
`docs/tui-plan.md` is deliberately untouched — it is the design record of what
was planned, not a description of what shipped.
The loose end from 777f974, now explained. Sessions came back from a detach on
`window-size latest` instead of `manual`, and the restore primitive round-tripped
correctly in isolation, so the corruption had to be upstream of it. It was: the
snapshot was taken from state this code had already broken.
`bindSwitchKey` passed a bare `;` between the two commands it wanted in one
binding. That is a command separator to tmux's OWN parser, not an argument: it
ended the `bind-key` and executed what followed immediately. So the binding kept
only `switch-client`, and `set-window-option ... window-size latest` RAN against
every switchable session at attach time — before the sizing snapshot was taken.
Every session was therefore snapshotted as `latest` and faithfully restored to
`latest`.
Proven against real tmux both ways before fixing: a bare `;` leaves the session
on `latest` and stores a one-command binding, while `\;` leaves it `manual` and
stores both commands.
Verified end to end: 7 sessions manual before, 1 latest + 6 manual during the
attach (the attached one follows the terminal, the rest are pre-sized), no dot
padding on a switch, and all 7 back to 120x40 manual after the detach.
This also means the "follow the terminal after switching" half of 777f974 never
actually worked — it was never in the binding.
The overview showed the same session twice, one frame above another, after
switching sessions (reported from the beta with a screenshot).
Claude repaints by ABSOLUTE CURSOR POSITIONING, not by clearing: a 198KB pane
tail carries 1142 `CSI r;c H` and exactly one `CSI 2J`. The replay honoured the
COLUMN of those sequences and ignored the ROW, so a repaint could never
overwrite what came before and was appended instead. That same tail replayed as
FIFTY stacked copies of one frame. The preview shows the last N lines, so on a
short terminal you saw the newest frame by luck and on a tall one you saw the
end of the previous frame above it.
A cursor HOME now starts the buffer over. That is not a heuristic but the
line-based equivalent of what a home means: a full-screen app announcing it is
repainting from the top, with everything on screen about to be overwritten in
place. Only row 1 column 1 counts — any other address is a write position
inside the frame being painted, and resetting on those would erase live
content.
Measured on the real tail that produced the screenshot: 198599 bytes and 50
copies of the welcome frame collapse to 40 lines carrying exactly one.
The old test pinned the append behaviour, including a spurious leading empty
line that the initial CUP produced; both are gone.
The bar now reads "alt+1-9 switch · F1 back to the codeman dashboard", so the
switch keys are discoverable instead of secret. Shown only when those keys were
actually claimed, the same rule the way-out key follows: a bar naming a key
that does nothing is the bug this series started with.
THE DOT GRID. Switching landed in a pane occupying part of the terminal with
tmux's dot fill everywhere else. It was never a size mismatch — the window was
already the right size. `window-size latest` only resizes a window while a
client is ON it, and the sessions behind the tab strip have none until you
switch, so the resize happened AT the switch: tmux painted the newly-available
area with dots and an idle claude had no reason to redraw into it. Every
switchable session is now pre-sized to the attaching terminal, which moves that
repaint to attach time while the user is still looking at the first session,
and the switch binding restores `window-size latest` on arrival so a mid-attach
terminal resize still follows. Measured: 14 consecutive switches across 7
sessions, zero dot-padded rows, against 1-in-6 before.
⚠️ Known loose end, deliberately not papered over: after a detach the window
SIZE is restored exactly but the window-size MODE can come back as `latest`
rather than `manual`. The restore primitive round-trips correctly in isolation
(manual -> presize -> latest -> restore = manual) and no call site in the TUI
or the server sets `latest` afterwards, so the cause is not yet identified. The
practical effect is nil: the remaining client keeps the window at its own size
and Codeman re-pins `manual` on the browser's next resize.
Switching with Alt+N landed in a pane that filled part of the terminal with
tmux padding the rest as a dot grid — reported from the beta with a screenshot
showing the pane in the left half and dots everywhere else.
Codeman pins every window `window-size manual` at the BROWSER's size
(tmux-manager.ts), so no attaching client can resize it. The attach already
lifted that for the session it opened, which is why a plain attach looked
right; `switch-client` then moved the user into a session that had never been
lifted, and the old pin reasserted itself. `window-size latest` now goes on
every session the strip can reach, alongside the bar those sessions already
get, and each one's original sizing is snapshotted and restored on detach.
Verified by round-tripping a session pinned at 120x40 manual: latest 190x49
while attached, back to 120x40 manual after, with no dot rows at either step
and the bar intact at full width after a switch.
The renderer's own FOOTER_KEYS table still said 'jump'. It is only reached when
the app layer supplies no footerKeys, so nothing visible was wrong, but a
fallback that contradicts the live footer is exactly the kind of drift that
turns into a bug report later.
Four faults, all reported at once, and three of them were mine from the last
two commits.
THE HINT VANISHED. Two independent causes. First, a leaked F1 binding: an
attach whose TUI was killed leaves `F1 -> detach-client` in tmux's root table,
and the claim treated "already bound" as someone else's key, so every later
attach fell back to advertising the tmux chord — the bar stopped saying F1
while F1 still worked. A key already bound to `detach-client` now counts as
ours. Second, width: tmux truncates a status line that overflows and drops the
RIGHT-aligned segment, which is the hint. The strip now gets a budget measured
from the terminal's width minus the hint, and it drops tabs from the far end
until it fits. ⚠️ Measured on VISIBLE columns, not format bytes: `#[reverse]`
costs zero columns, and counting it made a strip that "fitted" still truncate
the hint at 80, 100, 120 and 176 columns on a real terminal.
ALT+N DID NOT SWITCH. On the dashboard, a bare digit meant jump AND ATTACH, and
a terminal sends Alt+N as ESC then N: when those land in separate reads —
routine over SSH — the chord decodes as Escape plus a bare digit, so "switch to
tab 2" threw the user into tab 2's pane. A digit now SELECTS, matching what
Alt+N means in the web UI; Enter is how you go in. Inside a pane the keys never
reached the TUI at all, since tmux owns the terminal, so the attach now binds
Alt+1..9 in tmux's root table to `switch-client` — the strip is usable rather
than decorative. ⚠️ The bar is applied to every session the strip can reach,
each highlighting its own tab: with it on the attached session only, switching
landed the user in a pane with no strip and no way out on screen.
⚠️ The leaked-state sweep was missing `status-position`, so it removed the
marker and left the position behind — and with no marker the leftover no longer
matched, making it permanently unsweepable. Found by diffing every session's
options after a detach.
Attaching made every other session disappear: the dashboard is gone, tmux owns
the terminal, and there is nothing left saying what else is running. The attach
bar now carries the session strip, numbered exactly as the dashboard numbers
them, with the session you are in inverted, and it sits at the TOP of the pane
where the web UI keeps its tabs.
The strip is a WINDOW around the active tab, not the whole list, with ellipses
marking each end that is actually cut. The bar is one line shared with the way
out, and that hint is the only instruction a user gets while tmux has the
terminal, so it must never be crowded off; a test drives 20 long-named sessions
through the bar and asserts it survives.
⚠️ The strip is a snapshot taken at attach time and never refreshed. The TUI is
blocked in `spawnSync` for the whole attach so there is no loop to update from,
and tmux's own format language cannot map a `codeman-<hex>` session name back
to a label a human recognises. Slightly stale beats absent.
The way out moves from F12 to F1, which sits beside Esc where a hand backing
out already goes. Verified against BOTH encodings a terminal sends for it:
xterm's SS3 (ESC O P) and PuTTY's default (ESC [ 1 1 ~).
`status-position` joins the snapshot, so a session that had its bar at the
bottom gets it back there on detach along with everything else.
Three separate "why are there boxes" reports, and I fixed them one glyph at a
time instead of as a class, so the next one was always waiting. Grouping the
tester's terminal by unicode block made the rule obvious:
RENDERS Latin-1 (·), Box Drawing (─ │), Block Elements (█ ▛ ▐),
Geometric Shapes (○ ▶), General Punctuation (…), Arrows
TOFU Miscellaneous Technical (⏎ U+23CE, ⏵ U+23F5), the sparse end
of Dingbats (❯ U+276F)
That is an ordinary font, not a broken one, so it is the profile to design
against. The working spinner moves off Dingbats and Math Operators onto
quadrant blocks (▖▘▝▗) — the same block as the `▛█▐` art claude itself draws,
which that font renders fine — and the blocked marker moves off `⚠`
(Misc Symbols, emoji presentation on many terminals) onto `▲`, the block that
already gives us `▶` and `○`.
The preview fold gains claude's own spinner dingbats (✢ ✳ ∗ ✻ ✽ ✴ → `*`) and
`⚠` → `!`. Its animated status line is exactly where a reader looks, so tofu
there is the most visible kind there is.
A test now enforces this as a CLASS: no glyph in the unicode set may come from
Misc Technical, Misc Symbols or Dingbats, with U+2714 the single documented
exception because it was observed rendering on the very font that failed the
others. Verified by scanning a live frame driven with the tester's exact
environment: zero glyphs from any of the three blocks.
Killing demanded the session's NAME typed out in full. That is the right
ceremony for dropping a production database and the wrong one for closing a
pane you are looking at; the beta tester's verdict was "thats stupid, just make
me type Y to confirm". `x` then `y` is already two deliberate keystrokes on a
row the user selected, and the conversation lives in its transcript, which a
kill does not touch.
Everything that is not `y` CANCELS rather than being ignored, so a stray key
closes the dialog instead of leaving a destructive prompt armed and waiting for
whatever gets typed next. Enter cancels too: it is the key most likely to be
hit by reflex, and this is the one dialog that destroys something.
⚠️ Found while verifying the new dialog: it did not name the session. The label
was computed as `row.session.name ?? id.slice(0, 8)`, and `??` falls back only
on null or undefined, so every session the server left with an EMPTY name — all
of them, until the TUI started naming its own — sailed through and the box read
"Kill ?". A destructive prompt that cannot say what it will destroy is worse
than no prompt, and it is now a single keystroke. The caller passes the same
label the LIST shows, so the dialog names the row in front of the user.
The typed-name machinery goes with it: TuiConfirmState.typed, setConfirmInput(),
confirmAccepts() and the 'typing'/'reject' steps are all removed rather than
left as unreachable branches.
Codeman maintainer