Part B: Model Pricing, Hook Configuration and CLAUDE_HOME move from
Settings → Dashboard to the Claude Code tab's AgentHub sub-tab
(ClaudeAgentHubSettings.tsx); CODEX_HOME is editable in Codex → AgentHub.
A home that is or contains the OS home is refused (EBADHOME).
Settings persistence: dashboard-saved settings go to
~/.config/agenthub/settings.env (AGENTHUB_SETTINGS_FILE), outside the plugin
cache an update wipes. Load order: real env > the checkout's .env > that file.
The first CLAUDE_HOME change pins DASHBOARD_DATA_DIR so the DB stays put.
C1: codex-usage.js reads Codex rollout journals into sessions (new column
tool) and token_usage, tracked per file in codex_rollouts; forks deduped by
total vector, sub-agent threads folded into the root; pricing-sync also
prices OpenAI models; PricingSettings with a vendor filter.
C2: codex-hooks.js installs AgentHub's hooks into $CODEX_HOME/hooks.json on
a click, through a stable forwarder; Codex trust is never written. The hook
route creates tool='codex' sessions and folds sub-agent hooks into the root.
D: antigravity-usage.js reads agy's per-conversation SQLite files (tokens
from ModelUsageStats) into tool='antigravity' sessions; antigravity-hooks.js
installs the named hook `agenthub` (PreInvocation, PostToolUse, Stop — never
PreToolUse) into ~/.gemini/config/hooks.json; Settings → Antigravity tab;
Gemini pricing.
Whole-change review fixes: sameOriginGuard on both home PUTs, routers.missing
on a CLAUDE_HOME change, reset-pricing scope, GET /api/sessions?tool=,
clear-data covering the new tables, Codex child-thread events moving no
status, hooks owning status only while current, EFOREIGN for a forwarder
that is not ours, agy quiet window and turn rules, AGENTHUB_SETTINGS_FILE
honoured from .env.
Docs, OpenAPI and the CLAUDE.md marker updated. Server 2141, client 779.
Settings → Dashboard gains an AI Tools section (installed tools with
Enabled/Disabled and Set default, available ones with a homepage link,
Refresh). useAiTools shares one catalog copy across the app and replaces
LANE_TOOL_LABEL. The Add-lane modal picks Default (<tool>) or Other…; the
router row reads Native (<tool>) and disables AI Router profiles for tools
that cannot use one. The run form gives catalog tools resume and a red
per-run Bypass permissions switch (Codex too). Claude Code / Codex settings
tabs hide while that tool is disabled. The Workspace probes each lane tool's
CLI fresh.
Docs: CLAUDE.md tool/bypass rule, AGENTS.md, LANES.md, API.md, CLI.md,
server/README.md, OpenAPI fragment + regenerated openapi.yaml.
server/lib/ai-tools.js loads server/data/ai-tools/*.json (38 tools, data from
Orca, MIT) plus DASHBOARD_AI_TOOLS_DIR, detects each CLI by walking PATH, and
keeps enable/default in <data dir>/ai-tools.json. shell-path.js merges the
login shell's PATH at startup and on Refresh. /api/ai-tools lists, refreshes
and toggles; /api/run/binary resolves through the catalog.
Lanes accept any catalog id (EDISABLED for a disabled one on creation, ENOTOOL
at start once a tool left the catalog). pty-run keeps the Claude/Codex
builders and builds every other tool's argv from its catalog file, pasting an
after-start prompt through a tmux buffer; every pane gets the hydrated PATH.
A per-run bypass switch appends the catalog's bypass args (Codex included).
POST /api/run now forwards sandbox/approval/resumeLast, which it dropped.
agenthub lanes add gains --tool.
The plugin stops shipping its workflow layer:
- skills/ship-feature and skills/ship-feature-lane, the 18 stage/ship slash
commands (plugin.json keeps agenthub-doctor/update/open), and all four
agents/ (qc-local, senior-gate-reviewer, backend-reviewer, frontend-reviewer).
- CLI `agenthub skills install`, `agenthub lanes agents install`,
`agenthub ship-full|ship-quick`; routes POST /api/skills/install and
POST /api/lanes/:id/agents/install; the Settings button, the lane Tools
menu item and the Add-lane setup step that called them.
- The Superpowers auto-install: server/lib/skill-plugins.js, the bootstrap's
autoInstallSkillPlugins / skillPluginsSettled gate and the doctor row.
Pipelines are kept but dormant: server/data/pipelines/ is deleted, so no
template ships. `default` resolves to EMPTY_PIPELINE (no nodes, progress 0),
createLane still accepts it, POST /api/lanes/:id/stage answers 400
ENOPIPELINE, and the Workspace hides the stage timer, pickers, PipelineMap
and strip stage chip for a lane with no nodes. A template dropped into
DASHBOARD_PIPELINES_DIR turns it all back on. The server suite loads its
templates from server/__tests__/fixtures/pipelines/.
Docs: CLAUDE.md (marker, retired rules), README(.vi), INSTALL, ARCHITECTURE,
docs/LANES|PLUGINS|CLI|API, agenthub-introduction.html.
- antigravity: a third lane tool running Antigravity's `agy` CLI. `-i`
carries the first prompt, `--continue` resumes; `agy` has no model,
effort or permission flag, and `--dangerously-skip-permissions` is never
emitted. No pipeline, no agents install, and no AI Router (native only,
else EUNSUPPORTED). Add lane gains an Antigravity choice and the binary
probe maps it to `agy`.
- UI: the lane card no longer renders a frame holding only a "Codex" chip,
and the prompt box names the lane's CLI ("Ask Codex anything…"), via
LANE_TOOL_LABEL.
- runs: a pane that exits is released on a 10 s server beat
(healAllRunIds, DASHBOARD_RUN_HEAL_MS). Codex and Antigravity fire no
hook, so their lanes stayed "running" until something re-read them.
- codex: Start no longer sends resumeLast — only "Resume last session"
continues the previous thread.
- git (#50): stage-hunk/discard-hunk take an optional `block` — one run of
consecutive +/- lines inside an @@ hunk. `onlyBlock` (server/lib/git-diff.js)
cuts git's fresh hunk down to it, folding the other blocks into the side
being checked; LaneDiffView hangs one review bar per block (unified + split)
and its arrows walk blocks.
- git exclude: entries AgentHub writes for a lane go to the COMMON git dir with
the lane's path prefix (server/lib/git-exclude.js, used by router-apply and
lane-mcp) — a linked worktree's own info/exclude is never read by git, and a
subdir lane has no .git. MCP sync no longer fails with ENOTDIR on a worktree
lane or creates a stray .git in a non-repo lane.
- codex lanes carry no pipeline: codex-ingest mirrors only idle/running,
codex-nudge is removed, agents install writes nothing for codex, the
Workspace hides stage/pipeline/feature pickers/PipelineMap for them, and
every server start clears leftover codex stage data and the old AGENTS.md
stage block.
- Changes panel: a whole file opens for reading and enters the editor on a
click on its text (caret at the clicked line), not automatically.
- macOS: LaunchAgent start-at-login (scripts/launchd-service.js, picked by
scripts/service-manager.js).
- Adds the AgentHub introduction slides and their image-embedding script.
Opening a conflicted file in the Changes panel now shows each conflict
block as its two sides, labelled from git's markers and tinted apart,
with Keep mine / Take theirs / Keep both per block (undoable, ^/v to
jump). Save and mark resolved writes the chosen text through the
whole-file editor's PUT /git/file (baseSha stale check) then runs
resolve done. The Merge conflicts group's buttons are text, and a
conflicted file is marked with a red ! in the Files tree.
Fix: ls-files --cached lists an unmerged path once per stage, so the
tree showed a conflicted file three times; lsFiles now dedupes.
The (cwd, tool) rebuild lists its columns by hand and omitted
router_profile, which its probe adds earlier in the same start — a legacy
DB lost the column (and every lane's router) until the next restart.
- tmux targets are exact (`=name`, `=name:` for panes/options): a bare
`-t agenthub-lane-1` resolves by prefix to agenthub-lane-12 once lane 1
is gone, so killing a finished lane could kill another.
- getRun/killRun refuse ids that are not `agenthub-lane-<n>`, so
DELETE /api/run/main can no longer kill a user's own tmux session.
- router/apply reports restart_required only for a running pane; getRun
returns a `gone` handle, so it was always true.
- needs_action is cleared only by the session that raised it (new nullable
lanes.needs_action_session); a second session in the same cwd used to
clear it on its second hook.
- validateRemoteUrl checks the literal scheme prefix git will receive and
refuses %/backslash in the authority (WHATWG normalises https:/host and
percent-escapes that git reads differently); route test for EBADURL.
- Parent-session scrub also drops AI_AGENT, and GIT_EDITOR when it is
Claude Code's `true`.
- canonicalCwd resolves a not-yet-existing path through its deepest
existing ancestor, so a symlinked parent still yields one lane.
- Note the clock-step ceiling on liveServers' boot-time filter.
- Docs/header wording: the Active runs count is lane-scoped.
What broke (code audit, untested on a Mac):
- Without systemd the server is always spawned from inside a Claude Code
session, so it inherited CLAUDECODE, CLAUDE_CODE_SESSION_ID,
CLAUDE_PLUGIN_ROOT, ... and tmux handed them to every lane pane (a codex
pane even resolved as `claude` in `agenthub`'s harness check).
- The ~/.local/bin/agenthub launcher pinned a versioned node path that
`brew upgrade`/nvm deletes, leaving the CLI dead until the next bootstrap.
- A discovery entry left over from before a reboot could name a recycled pid,
so the bootstrap's fast path believed a dashboard was live and
stopDashboard could signal an unrelated process.
- Lanes were stored under the typed cwd; macOS reports /private/tmp and the
on-disk case, so sessions never bound to a /tmp or differently-cased lane.
What changed:
- server/index.js deletes a DENYLIST of parent-session variables at startup;
user config (CLAUDE_CONFIG_DIR, CLAUDE_CODE_USE_BEDROCK, ANTHROPIC_*) stays.
- linkCli's launcher falls back to `node` on PATH when the pinned one is gone.
- liveServers ignores entries whose startedAt predates this boot (60 s margin).
- createLane and POST /api/lanes/ensure store/look up fs.realpathSync.native
of an existing cwd; the liveness reaper also tries the realpath. Existing
rows are untouched.
- session-liveness gains an exec seam so the darwin ps+lsof branch is tested.
- plugin-doctor names ~/.zshrc on darwin; INSTALL.md gains a macOS section
with the Mac verification checklist, mirrored in both READMEs and PLUGINS.md.
No launchd, no macOptionIsMeta, engines stays >=20.
What broke: a lane on a fresh `git init` repo had no working git surface.
`statusCounts` ran `rev-parse --short HEAD` unguarded, so `gitFacts` threw
and `GET /api/lanes/:id/git` answered `{available:false}` — the Changes
panel showed no header, and `commit`/`switch` 500'd through `isDirty`.
Add lane in Worktree mode accepted such a repo and left a `failed` row
behind, and a repo with commits but no remote had a permanently disabled
Push with no way to give it one short of a terminal.
What changed:
- worktree.js: an unborn HEAD is a normal state. `statusCounts` reports
`head: null`; `gitFacts` reads the branch via `symbolic-ref`, an empty
subject, and a new `unborn` field.
- git-ops.js: an 11th write, `publish` — `remote add origin -- <url>` then
`push -u origin HEAD`, only with a commit (ENOCOMMITS) and NO remote at
all (EREMOTEEXISTS, so it never re-points one), the url allowlisted by
`validateRemoteUrl` (https without userinfo, ssh://, scp-like; refuses
local paths, file://, ext::, fd::, git://, http://, leading `-`:
EBADURL), and `origin` removed again when the push fails.
- POST /api/lanes/worktree answers 400 EEMPTYREPO for a repo with no
commits instead of creating a row that can only fail.
- LaneGitBar: unborn repos show "no commits yet" and a commit-first hint;
no remote offers Publish… with an inline URL box; a remote but no
upstream relabels Push as "Publish branch". AddLaneModal disables Add
on an empty repo and offers "Use as a Repo lane".
- The two tests that used an unborn repo to force a provisioning failure
now use a `feat` ref clash, since that repo is refused up front.
- CLAUDE.md rule reworded to "the boring eleven" with publish's guards;
docs/API.md, docs/LANES.md, ARCHITECTURE.md, server/README.md, openapi.
Part A — a fetch/pull/push against a remote that needs a login no longer
dies on "terminal prompts disabled": the server answers 409 EGITAUTH with
the host, the git bar asks for a username + password/access token and
resends the same action. AgentHub keeps its OWN credential store
(<CLAUDE_HOME>/agenthub-git-credentials/, 0700/0600), one login per host,
shared by every user of the install, managed from Settings -> Git
credentials. A login is only sent to one of the repo's own remote hosts,
saved only after git accepted it, never returned by any route, and reaches
git through an exclusive credential helper (never argv).
Part B — merge a branch (or the diverged upstream) from the git bar; a
merge that stops on conflicts lists them in the Changes panel with Keep
mine / Take theirs / Mark resolved, then Complete merge or Abort merge
(confirm gate). Clean tree, repo-root lane and an existing ref only;
pull/switch/hunk actions are blocked mid-merge; staging a conflicted file
with markers left is refused. Rebase/reset/stash stay out.
- Changes panel at full screen: the whole Files view (finder, expand-all,
tree) becomes a left sidebar and the Files tab is dropped; a working-copy
file opens straight into the editor (Save/Cancel live once the text
differs, Cancel restores it, Whole file toggles back to the diff), and the
page is height-locked so the editor no longer collapses with no live run.
- Overview splits into Staged Changes / Changes groups in the work range.
- A changed file's whole tree row takes its status tint.
- Drag an image onto the lane terminal: POST /api/run/drop-image saves it
(png/jpeg/gif/webp, random name, 0600, pruned after a day) and the path
is pasted into the pane, which is how claude attaches it.
- Claude Opus 5.5 gets its own pricing row (4/20/0.2/5/8, fast 8/40);
claude-opus-5% had been pricing it as Opus 5.
- server/lib/pricing-sync.js: a model family used in token_usage with no
row of its own is inserted from LiteLLM's public price list, never
updated; off with AGENTHUB_PRICING_SYNC=0. repo-links allows exactly
that one third-party GitHub raw URL.
- Split view: every pane shows the selected lane, each in its own tmux
session (`slot` 0-3 on POST /api/lanes/:id/start; slot 0 stays the
lane's `run_id`, slots 1-3 are `agenthub-lane-<id>-<slot>`). A pane's
Active runs hides what the other panes show; stop/reset/remove kill
every slot. No per-pane lane picker any more.
- Whole-file editor: PUT /api/lanes/:id/git/file (lane-file-write.js)
with the read path's guards, existing text files only, and a baseSha
stale check (409 ESTALE). A file it cannot take says why.
- Per-hunk review bar under each hunk's last changed line: prev/next,
"n of N", Undo, Keep. Review diffs no longer nest their own scroll box
(unless windowed), so navigation moves the one real scroller.
- Files tree lists untracked-not-ignored files and colours entries by git
status (M/U/A/D/R), folders holding changes get a dot.
- Changes panel: Unstage all on the single commit/Stage-all row; sizes
Narrow (30%) / Medium / Full screen, no 70/30.
- Lane card: the needs-you banner is gone (the strip tile keeps it).
Several suites start real `agenthub-lane-<id>` tmux sessions and then kill
or reset them, with lane ids from their own throwaway database. On the
user's default tmux server those names collide with the dashboard's real
lane panes, so a test run — including the one the pre-commit hook runs on
every commit — killed whichever live session shared an id, the pane the
commit was made from included.
`npm run test:server` now goes through scripts/test-server.js, which drops
TMUX, points TMUX_TMPDIR at a fresh directory and kills that server after.
Also fixes the ERUNTIMEOUT test's tmux mock, which had the wrong (async,
cmd+args) signature and only passed by accident.
Review and the single-file view gain per-hunk Stage and Undo
(stage-hunk / discard-hunk: the hunk is named by its @@ line and
re-matched against a fresh diff, then git apply --cached / -R; undo
needs confirm: true and two clicks). Each review file header stays
pinned while its diff scrolls and carries the file's own
stage/unstage/discard.
- The Changes panel's width control gains a full-width step (narrow →
50/50 → 70/30 → full → closed); the console beside it is hidden with
CSS, never unmounted.
- "Expand all folders" is one GET /api/lanes/:id/git/tree-all (one
`git ls-files`, every level shaped like /git/tree) instead of one
request and one git process per folder.
- Overview: a Stage all button and the staged-only commit box sit above
the file list and stay pinned while it scrolls; the Files finder is
pinned too.
Server (git-diff.js, git-ops.js, worktree.js, routes/lanes.js):
- Every lane git call runs with GIT_LITERAL_PATHSPECS=1 and -z porcelain, so a
file name is a file name, never a glob, and odd names parse correctly.
- A lane that is a subdirectory of its repo sees and acts on ONLY its own
paths (reads, stage/unstage/discard/commit, and gitFacts.files).
- Any file named settings.local.json (the lane's router API key) is refused as
403 ESECRET on disk (symlinks resolved) and at every committed rev.
- commit refuses with 400 ESTAGEDOUTSIDE while anything outside the lane is
staged (checked with --no-renames, so a move into the lane is caught).
- push resolves its remote (upstream → only remote → origin), else 400
ENOREMOTE; unstaging a rename unstages both halves.
- Git writes serialize on the repo's git dir, not the lane id, so two lanes on
one index cannot collide on index.lock.
- git refusing an action (a numeric exit status) answers 409 EGIT with git's
stderr instead of a 500 carrying a numeric code; one shared parseNumstat.
Client:
- Changes panel: no stale answer from a previous lane or an older quick-open
query repaints the panel; discard disarms on lane switch; Enter on a file
path opens it instead of folding; bounded tree expand-all.
- Large diffs (> 500 rows) render only the visible rows, with stable width
and in-sync split columns; mirrored split scroll is clamped.
- Git bar: no "in sync" without counts, a clear no-remote message, branch list
re-read after a switch, and no previous-lane facts after a lane switch.
- Tooltips no longer flicker or close on unrelated scrolls.
- Settings → Appearance: mode Dark / Light / System (follows the OS live) and
a named theme remembered per mode — dark: Default, Darcula, Apple Dark;
light: Default, Apple Light, IntelliJ Light, Canvas. Each theme is a
CSS-variable override block keyed by `data-theme` on <html>; the sidebar
quick toggle gains System and swaps between the two remembered themes.
- An inline boot script in index.html applies the stored mode and theme
before first paint, so a reload no longer flashes the default colours.
- Diff +/- rows are easier to tell apart on every theme: a /20 status tint,
a 3px status bar and a bold coloured marker; the changed-span emphasis no
longer vanishes on light themes.
- #42: the Add-lane modal (Repo mode / "Current branch") asks
POST /api/lanes/ensure with `exact: true`. Without it a folder nested inside
an existing lane resolved to that parent lane with created:false, so the
modal closed and no new card appeared. When the folder really already has a
lane, the modal now says which one instead of closing.
Closes#42
An older Claude Code has no -y on `plugin install|update` and exits with
"error: unknown option '-y'", so the dashboard's Update now chain failed at
its plugin_update step (and so did Update all, the plugin buttons and the
superpowers auto-install). Those versions also predate the confirmation
prompt -y answers, so every wrapper now retries the same argv without it
when, and only when, the CLI rejects the flag. /agenthub-update, which the
failure message points users to, falls back the same way.
The Health Score's heap component divided heapUsed by heapTotal, which is
only what V8 has grown the heap to so far, so a healthy server read 70-90%
and dragged the score down. GET /api/settings/info now also reports
server.memory.heapLimit (v8 heap_size_limit, the ceiling the process dies
at); the V8 Heap bar and the Health Score divide by it, falling back to
heapTotal against an older server.
A doubled hook registration (plugin + leftover checkout install) POSTs every
event twice with a byte-identical payload, so the timeline showed each row
twice and token/alert/lane side effects ran twice. POST /api/hooks/event now
skips a payload identical to the session's LATEST event from the last 10s
(findRecentDuplicateEvent). Latest-only keeps two legitimately identical
Stop events of consecutive turns, which always have other events between.
- Changes card beside lane detail with five modes: Files tree, overview,
Review (whole diff, per-file collapse), single file, commit history.
- LaneDiffView: unified / side-by-side with independent scroll on both axes
and mirrored scrollLeft, expand/collapse context, intra-line emphasis.
- Staged-only commit box; git actions (fetch/pull/push/stage/unstage/
discard/commit/switch/branch-create/branch-delete) guarded in git-ops.js.
- Server: git-diff.js + git-ops.js behind GET /api/lanes/:id/git/* and
POST /api/lanes/:id/git/:action; gitFacts gains upstream/ahead/behind.
- Add lane: new/existing branch, current branch, worktree list, prefill
from the git bar's "New lane here"; /api/run/cwds gains a repo kind.
- GlobalTooltip upgrades every native title into a themed anchored bubble;
chart/info popovers use theme tokens.
Terminal (issue #33):
- `convertEol` is gone from xterm. It rewrote every bare LF as CR+LF, so a
tmux pane that scrolled (tmux emits a bare LF and keeps drawing at the
column it was on) had the rest of that draw land at column 0, stamping a
line's first characters into the left margin of the rows below — the torn
left edge that only showed up once content scrolled.
- The attach carries the client's real size (`?cols=&rows=`, `sizeFromQuery`
in `server/websocket.js`, bounded because it reaches a TIOCSWINSZ ioctl).
`attach()` spawned `tmux attach-session` at a hardcoded 80x24 on every
attach, and tmux sizes the window to the newest client, so `claude` redrew
its whole UI at 80 columns until the first `resize` message landed.
- The terminal takes the keyboard on mount, so typing reaches tmux without
clicking the pane first.
- A pane that ends (double Ctrl+C, `exit`, a crash) reports itself through a
new `onExit` prop and `LaneConsolePane` runs `newRun()`, so the console
returns to the setup form instead of leaving a dead pane behind a
"New run" button.
Pane environment:
- `server/lib/desktop-env.js` forwards the graphical session's DISPLAY /
WAYLAND_DISPLAY / XAUTHORITY / XDG_RUNTIME_DIR (from
`systemctl --user show-environment`, this process's own values winning)
into the pane. The server normally runs as a systemd user service with none
of them, so `claude` there could not read the clipboard and answered "no
image found in clipboard" on every image paste.
Run setup:
- `auto` is a permission mode and is now the default everywhere — allowlists,
the `--permission-mode` fallback, the select, and the pane's initial state.
- The model picker sends the CLI's aliases (`fable`/`opus`/`sonnet`/`haiku`)
instead of pinned ids. `claude --model` resolves an alias to the latest
model itself, so a release like Fable 5.1 needs no edit here; pricing is
unaffected because it keys off the resolved id the transcript reports.
- **Create/update readme file** moved from the Add-lane modal into the lane's
Tools menu: creating a lane and writing its docs are separate decisions,
and the second is worth re-running long after the first.
- The Workspace's no-lane state is a real banner; the lane strip (add tile
included) renders only once a lane exists.
Refs #32, #33.
Tests: npm run test:server (1828 pass), client vitest (558 pass).
Run console:
- Picking a pipeline no longer pre-writes anything into the prompt box
(`promptPrefill` in `LaneConsolePane.tsx` is gone — it wrote the driving
slash command on a claude lane and a stage checklist on a codex one).
Whatever sits in that box is sent as a real prompt the moment the run
starts, so the box is the user's alone now.
- Resuming a session sends NO prompt at all (`handleStart` in `RunSetup.tsx`):
the resumed session carries its own transcript, and leftover text in the box
fired an extra turn into it.
Notifications:
- `client/src/lib/notify.ts` is now the ONE sender, used by both
`useNotifications` and Settings' test button, and it reports where the
notification landed: `POST /api/push/send` answering `{native:false,pushed:0}`
means it went NOWHERE, so the client falls back to a local notification
instead of counting the 200 as delivered.
- A notification carries the lane that raised it (`/workspace?lane=<id>`), and
Workspace selects that lane on arrival.
Profiles:
- `server/lib/cc-profiles.js` reads and writes the macOS login Keychain item
`Claude Code-credentials`, where the `claude` CLI actually keeps
`claudeAiOauth` on darwin — a switch used to rewrite a file nothing reads.
The file stays the fallback, is kept in step only when it already exists, and
`AGENTHUB_CC_KEYCHAIN=0` pins it (tests set it).
Tests: npm run test:server (1820 pass), client vitest (553 pass).
Three defects that all showed up as "the dashboard's numbers are wrong", plus
the notification and brand work asked for alongside them.
Tokens: Claude Code writes ONE JSONL line per content block of an assistant
message, and every one of those lines repeats the same message-level `usage`.
Both parsers summed them, so a tool-heavy session's tokens — and its cost — ran
1.5-2x high (one 4h session: $45.82 stored against $27.72 real). `lastUsageMsgId`
counts a message once, carried across the transcript cache's incremental chunk
boundaries. `replaceTokenUsage`'s high-water baselines mean the fixed parser
cannot correct what is already stored, so `setTokenUsage` (exact, baselines
cleared) is added for the deliberate re-parse path and `reconcileTokens` now
covers every session with a transcript on disk rather than imported ones only.
Turns: every `TurnDuration` row was stamped NOW while its dedup guard matched on
the transcript's timestamp, so each hook re-inserted the session's whole turn
history — 587k rows, 75% of a real events table, one session holding 68,747
copies of the same turns. The row now carries the transcript timestamp, which
fixes the guard and the feed's ordering together. Same class of bug in the
session metadata: `thinking_blocks` / `turn_count` / `total_turn_duration_ms`
are assigned from the full parse instead of `+=`, since `extract()` returns
whole-transcript totals on every call.
Events: leftover checkout hook entries in `~/.claude/settings.json` double every
event, every subagent row and every token figure. `stripLegacyHooks` has always
removed them — but only inside `bootstrap()`, which `fastPathOk` skips on
exactly the healthy installs where nothing else forces a full pass, so they
survived indefinitely. `hasLegacyHooks` (pure fs) now blocks the fast path.
A one-time repair at the first start after upgrading (`user_version` 3) deletes
the duplicated `TurnDuration` rows and re-parses every transcript with exact
writes. Marker 3 and not 2: an install that ran the earlier pass had its
corrected rows re-inflated by whatever old build was still ingesting hooks.
Router: switching a lane's AI Router while a pane is live only half-applies —
Claude Code re-applies `env` to a running session but keeps the model it
resolved at startup, so the old router's model name reaches the new endpoint and
the next prompt fails with "model not found". The apply route reports
`restart_required` / `run_id`, and the picker offers a Restart that respawns the
pane from the lane id alone (replaying the old `--model` would re-create the
mismatch).
Notifications (#29): the two moments Claude is BLOCKED on the user now have
their own switches and default to on — `onAwaitingInput` (a permission prompt or
a question) and `onTurnComplete` (end of turn). Before, the first was unswitchable
and the second hung off a preference that defaults to off, so the one thing a
user walks away from went unannounced. Prefs shape/defaults/storage move to
`client/src/lib/notifPrefs.ts`; Settings and the hook had private copies.
Brand: the sidebar shows the SmartGift wordmark with the AgentHub caption under
it, the collapsed rail shows the mark, and the favicon is vendored to match.
Closes#24.
**The review node ran nothing.** Both templates declared
`/superpowers:requesting-code-review` — the skill about how to HAND OVER a
review, not one that reviews anything — so `agenthub lanes node review`
performed no review at all. They now declare the bare `/code-review`, Claude
Code's own. It carries no `plugin:` prefix, so `requiredSkillPlugins` skips it
and the doctor's required-plugin row is unchanged (verified against the live
function: still exactly the five superpowers skills the templates name).
**The rest is the brainstorm → implement → review span, made cheaper to run and
harder to lose track of.** `ship-feature` gains four sections it never had:
- **Right-size the run.** The pipeline is not free — every dispatching phase
costs a model call and minutes, and the fix-loop pays it again each lap. A
typo does not enter the pipeline at all; a small single-layer change skips the
plan-debate, independent-test and gate agents and lets the user's acceptance
be the gate; medium/large runs everything. Sizing is a claim made out loud,
and a task that grows re-sizes upward.
- **Dispatching a judgement.** The fresh-context/FOREGROUND/never-`fork`/
hand-only-verifiable-artifacts/one-parseable-line contract, written once
instead of re-argued in three phases — plus the rule that was missing
entirely: never dispatch what a command answers. An agent is for judgement,
not for typing.
- **Resuming after a compaction.** Long runs outlive the context window; the
lane skill had this and this one did not. Re-derive position from the declared
stage and the diff, and treat a phase whose evidence you cannot find as one
that did not happen.
- **A `Leave when:` line on every phase.** Objective exit criteria are what stop
both the silent skip and the endless polish: intake leaves when the answers
are in hand, review leaves when the latest pass over the CURRENT diff is
clean, the gate leaves when verification really ran AND the user said yes.
And, narrower: intake asks only what the repo cannot answer (a question the code
could have answered spends the user's round and earns a worse answer);
implement works the approved plan task by task, re-checks each acceptance
criterion, heartbeats on long work, and treats a plan that turns out wrong as a
return to `plan` rather than a quiet substitution; tests runs the suite the diff
needs rather than both by reflex.
The `default` 8-node pipeline walked past its own intake: the node declared
`agenthub stage intake`, a stage report with no work attached, so "intake" was
a label the run passed through without ever asking a question. It now declares
`/superpowers:brainstorming` (and detects it — that rule moves off `plan`),
and both driving skills make the brainstorm mandatory, asked with
AskUserQuestion and waited on rather than answered on the user's behalf.
Three stops replace three silent slides:
- After the plan is written and adversarially challenged, it goes to the USER
for approval. Implementation does not start without it, and work that turns
out to need a different approach comes back for a fresh one.
- Review picks its reviewers from `git diff --name-only`: Claude Code's own
`code-review` always, a domain reviewer only for a layer the diff actually
touched, and — new — only once the change is big enough to have seams. Three
agents re-reading twenty lines buys one opinion at three times the tokens, on
every fix-loop lap. Review is also a loop now: it ends on a clean pass over
the current diff, not on one pass having happened.
- The gate asks the user to accept before anything is claimed done (and in the
lane pipeline, before the PR is published — publishing is outward-facing).
A NO-GO re-enters by SIZE: a defect goes back to implement, a wrong approach
back to plan. Nothing carries a NO-GO forward with a note about it.
Also here, from a real stall seen on a lane: `ship-feature` had no waiting rule
at all, and a session at the gate invented one with `ScheduleWakeup` — a
/loop-only primitive that paces nothing here and is refused outright when
`auto` mode cannot reach its safety classifier. The rule the lane skill already
had now exists in both.
Cost, measured in what each agent is actually for: the suite run leaves its
agent (nothing is judged, so nothing is delegated — an agent there buys a
second model's tokens to read output the session can read for free), and
`backend-reviewer` drops from opus to sonnet like its frontend twin; both read
a scoped diff against a fixed checklist, and the senior gate above them is
still opus.
Skill bugs found in the same sweep: doc-map told agents not to recreate the
removed translations and then, three lines later, to mirror every README change
into them; the docs mapping still carried `+VN+CN` and the punctuation left by
their removal; Procedure jumped 2 → 6; Verify required `scripts/doc-coverage.sh`,
which does not exist in this repo.
A lane's `remove` tore down what the lane pointed at: `git worktree remove
--force`, a branch delete, and a drop of the lane's own database. One
mis-click on a dashboard row cost a whole branch's unpushed code — and the
row is the only thing AgentHub actually owns. `remove` now stops the run and
the stack, clears AgentHub's own runtime bookkeeping under
`LANES_ROOT/.state/lane<slot>/`, and deletes the row. The working tree, its
branch, git's worktree records and the database stay exactly where they are,
for the user to delete by hand when they mean to.
That also fixes#28. Removal used to route through `assertDestroyable`, whose
third check demands the cwd appear in `git worktree list`; a managed lane
whose directory git no longer listed answered `ENOTWORKTREE` ("lane is not
listed as a worktree in …") and could never be dropped from the dashboard at
all. Removal no longer asks git anything, so no preflight blocker stops it,
and the unpushed-commits force-gate now applies to `reset` alone — nothing
`remove` does can discard a commit.
The UI follows: one label for both lane kinds ("Delete lane record"), copy
that says what survives, and no Force checkbox on removal.
`removeWorktree`/`removeLaneData` remain in place as guarded primitives; no
route calls them any more.
The quota check failed with `invalid_grant` for the account the user was
logged into - the one profile they would try first. An OAuth refresh token is
single use, so every refresh Claude Code performs on its own account retires
the copy sitting in that account's saved profile: the stored credentials of
the LIVE account are the one set guaranteed to be rejected, however recently
the profile was saved. Confirmed on a real install - the saved profile held a
refresh token three rotations behind `.credentials.json`.
`cc-profiles.liveCredentialsFor` now matches a profile's accountUuid against
the live credentials and, on a hit, the usage call uses what the CLI is
holding. That token is never refreshed and that file is never written:
rotating it would log the user out of their own terminal. An expired live
token answers ELIVEEXPIRED - using Claude Code once makes the CLI refresh it -
and a genuinely spent refresh token is now reported as ESTALE with what to do
about it, instead of echoing Anthropic's raw 400.
Opening the Profiles modal also asks for every saved account at once, rather
than one button per row: comparing accounts is the only reason the panel
exists. The guard against rewriting credentials over and over is no longer
"one row at a time" but the rules above - the live account costs a plain read,
and any other profile refreshes only once its token has actually expired,
which then leaves it valid for hours of further opens. Each row resolves on
its own, so a superseded account still lets its neighbours report.
Closes#19, #20, #21, #27.
#19 — a saved auth profile now reports how much of its account's quota is
already spent, which is the only number that decides which one is worth
switching to. Nothing under CLAUDE_HOME records it, so it comes from the same
endpoint the CLI's own /usage reads, answered for whichever account's OAuth
token is presented. It is a POST and a button, never a poll: every profile that
is not the active one is normally expired, so answering means refreshing it
first, and a refresh ROTATES the stored refresh token. The rotated pair is
written straight back — dropping it would leave the profile holding a
credential Anthropic has already retired, i.e. an account needing a fresh
login. The token is read, put in an Authorization header, and never returned,
logged or echoed; what leaves the server is two percentages and two reset
times.
#20 — a settings patch that changes nothing wrote nothing but still took a
timestamped backup, so opening the config panel quietly grew a backup
directory inside every lane. patchSettings now compares the rendered output
with what is already on disk and returns early, fixed once where all callers
route through. A lane's git exclude also picks up .claude/.cc-config-backups/
beside settings.local.json.
#21 — the router apply route was the only mutating lane route that never
broadcast its lane. The write reached disk and the column, but the client's
cached row kept the old router_profile forever: the picker looked right until
the user switched lanes and back, at which point the effect re-read the stale
row and flipped the selection to the previous profile. One broadcastLane, plus
a lane-switch guard so an apply that resolves after the user has moved on no
longer paints its result onto whichever lane is now on screen.
#27 — Add-lane in worktree mode left the dialog up after a clean setup with a
dead confirm button and no obvious way out. It closes itself when nothing
failed, and when something did, the summary stays and the confirm becomes a
real Close.
Closes#26, #25.
`runs` is the denominator every pass rate on the Workflows page is read
against, and any backwards move opened a new one. Stage detection bounces by
design — one feature reads `implement → tests → implement` as the session
edits, runs the suite and edits again — so a single feature was recorded as
five or six runs, and a lane in constant use reported a handful of passes
spread across dozens of phantom runs. A backwards move now opens a run only
when it is DECLARED; a backwards detection continues the run in flight.
Reaching the final node still closes the run for either, since work after
`done` is new work however it was observed. Replayed over this repo's own lane
the rule takes 26 recorded runs down to the 5 features that actually shipped
through it. The panel itself is unchanged: passes are still declarations only.
`/reload-plugins` and `/reload-skills` were typed back to back. Claude Code
opens its slash-command menu the moment `/` is typed and the Enter that follows
lands on that menu, so the second command was appended to the line still in the
composer and the pane submitted `/reload-plugins/reload-skills` — which Claude
Code runs as `/reload-plugins` with an argument, never reloading skills, and
leaving a stray prompt turn in the scrollback. Reproduced and fixed against a
live pane: SETTLE_MS (1s) between the two sends, which makes
`reloadOpenSessions` async rather than block the event loop, and the same pause
in the detached update-all worker.
The lane strip's tiles are now drag-reorderable (Alt+←/→ is the keyboard
path), and a lane's name can be set in place from the detail header. The
"Lane 23 ·" prefix is gone from both — the bare `#23` stays on the tile,
because that number is how the CLI addresses a lane.
Card order is a per-viewer preference, not a lane fact: it lives in
localStorage (`agenthub.workspace.laneOrder`) beside the split-view layout
and writes nothing server-side, so `GET /api/lanes` and `agenthub lanes`
keep answering in id order. Renaming is the opposite — it is PATCH
/api/lanes/:id with {title}, the one lane column a user now types into
directly.
That made a patch value reachable by hand for the first time, so
`bindable` now rejects any non-scalar for every PATCHABLE column with
EBADFIELD, and the route answers 400 for that and for the
SQLITE_CONSTRAINT family, rather than handing better-sqlite3 a datatype
throw and reporting 500 on plain bad input.
The uncommitted-files tooltip became a real panel. `GET /api/lanes/:id/git`
now returns `files` as {path, status, added, deleted} records — deltas from
one `git diff --numstat HEAD`, null where git has none to report (untracked,
binary, renamed) — and the card renders them portalled to the body, so the
Workspace's own scroll box cannot clip the list.
GET /api/lanes/:id/git now returns `files`, the raw `git status --porcelain`
lines the dirty/untracked counts were already parsed from, capped at 20 so a
lane mid-`npm install` cannot ship thousands of paths through a polled
endpoint. The lane card hovers them as the tooltip on its
"N modified · N untracked" text, appending "… +N" when the cap bit.
No new subprocess and no new endpoint — statusCounts was counting those lines
and throwing them away.
Two gaps left by the AI Router feature: a profile could be saved without ever
being reachable, and a lane's router could only be chosen at creation.
**POST /api/ai-routers/test** answers the question saving cannot. It sends one
Anthropic-style message capped at a single output token to `<baseUrl>/v1/messages`
with a 15s timeout — the wire shape a claude lane's eight env keys point at, so
what is tested is what will actually run. A router that answers is a 200 here
whatever it replies: `{ ok, status, message }` carries the router's own words,
and `status` is null when nothing answered at all. Only bad input is a 4xx.
The stored key is only ever sent to the endpoint it was stored against. A blank
`apiKey` falls back to the key of the profile named by `name`, but only while
the submitted `baseUrl` still equals the saved one — so testing a retyped
endpoint needs a retyped key (`EBADKEY`). Without that rule this route would be
a way to make the server POST a saved API key to any URL a browser names.
Network failures report `error.cause`, not `error.message`: node's fetch reports
every transport failure as the same useless "fetch failed" and hides the part
the user can act on (ECONNREFUSED, ENOTFOUND, a TLS error) one level down.
**The lane card carries an AI Router select.** The Add-lane modal was the only
place a router was chosen, and a lane outlives its creation. Switching goes
through `POST /api/lanes/:id/router/apply` — never a lane PATCH — because the
DB column and the lane's own config on disk may only move together. A failed
apply puts the select back rather than leaving the card claiming a router the
lane did not get. Hidden when `native` is the only profile: a select with one
option is a control that cannot do anything.
Adds `Lane.router_profile` to the client type (the payload already carried it).
Claude Opus 5 had no pricing row at all, so every token it burned was costed
at $0 — the pattern list stopped at the 4.x Opus family and nothing matched
`claude-opus-5`. Add it (5/25, fast mode 10/50, same shape as 4.8).
Sonnet 5's launch discount expired 2026-08-31 and the discounted rate became
the published standard, so the row's 3/15 standard rate started overcharging
the day after. Fold 2/10 into DEFAULT_PRICING and empty DEFAULT_INTRO_PRICING;
no model ships a promo today.
Fable/Mythos 5.1 fell through to the `claude-fable-5%` row, which caches at
$1/MTok against 5.1's $0.25 — give them their own rows. That overlap is only
resolvable if the lookup prefers the more specific pattern, and `matchPricing`
took `LIMIT 1` with no ORDER BY, i.e. whichever row was inserted first. Order
by pattern length descending, which is what docs/DATABASE.md already claimed.
(The nested REPLACE(model_pattern, '%', '%') it dropped was a no-op.)
Retire Claude 3.x, Opus 4/4.1 and Sonnet 4 from the defaults. The startup
top-up only ever ADDS missing rows, so removing them from DEFAULT_PRICING
would leave every existing install untouched — prune them once, guarded by
PRAGMA user_version (0 -> 1) so a model deliberately re-added in Settings
survives later restarts. The same one-shot migration corrects the existing
Sonnet 5 row. Only rates are deleted; token_usage is never touched, and
Reset Defaults still restores the full current set.
`claude plugin update` unpacks the new version into a new cache directory and
says "Restart to apply changes". A Claude Code process that was already running
does not: it keeps loading the OLD directory, hooks and all.
Every session such a process opened then ran the OLD plugin-bootstrap, which saw
a state.json naming a different root, concluded the server had to be replaced,
repointed agenthub.service at itself and restarted the dashboard onto the
version the user had just upgraded away from. The dashboard reported the old
version and offered the update again — once per session, forever, while
installed_plugins.json and the cache both said the new version was installed.
A /agenthub-update from that session "succeeded" and changed nothing.
- `supersededBy()` reads Claude Code's installed_plugins.json. When a SIBLING
cache directory is the recorded install, the bootstrap logs one line and
returns "superseded" — no deps, no unit, no server. `--force` does not
override it: the self-update chain forces from the NEW directory, which is
never superseded, so the only --force refused is a deliberate downgrade.
It cannot hand over either — the process it would replace is the one loading
it — so the log line says to restart Claude Code.
- Matched by sibling directory, not by the <plugin>@<marketplace> key: correct
without parsing a cache path or trusting a name. A phantom entry (installPath
that no longer exists) is ignored rather than wedging the bootstrap shut.
Also adds `pruneRetiredPluginRoots()`, since the same updates leave a full copy
of the repo — node_modules and all — per release; twelve of them measured 1.7 GB
on one machine. It runs after writeState, so a bad prune cannot cost the record
of a good bootstrap, and it protects four independent readings of "in use":
this build, every installPath Claude Code records, the directory
agenthub.service names (`unitRoot`, foreign unit included), and any directory
still named in the user's settings.json hooks — that last one is the un-restarted
process above, seen from the other side. Plus the 2 newest survivors as a
rollback cushion, and nothing at all outside ~/.claude/plugins/cache/.
A new **AI Router** tab in Settings, next to Codex, stores named router
endpoints (9Router, OpenRouter, a self-hosted gateway) as profiles: a base
URL, an API key, a default model and an optional per-tier override. The tab
only stores them; a lane CHOOSES one in the Add-lane modal, via a Native /
AI Router radio, and that choice is written into the lane's own config.
The two CLIs are configured differently, and the difference is theirs:
- a claude lane gets eight `env` keys merged into its own
`.claude/settings.local.json` (`.local` because the value holds a key);
- a codex lane gets `$CODEX_HOME/<name>.config.toml` — the layer
`codex -p <name>` loads — because Codex has NO project-level config.
Its base `config.toml` is never touched.
Native is a built-in profile rather than a code branch, so "go back to
native" is the same write path. It blanks the two connection variables
instead of removing them (an absent key at project scope lets a USER-level
router apply) and removes the model names instead of writing them (writing
them would override the user's own global model choice).
Key material never leaves the server: profiles live one file each under
`<CLAUDE_HOME>/agenthub-routers/` (0700/0600), `GET /api/ai-routers`
reports only `hasKey`, and there is deliberately no route returning one
profile in full.
Three things this had to get right, each found by review and verified
against the real tools rather than assumed:
- `wire_api = "chat"` is rejected outright by codex-cli 0.149 ("no longer
supported"), so every codex lane on a router would have failed to start.
The layer writes "responses", and writes `env_key` only when the profile
actually has a key — naming an unset variable is itself a startup error.
- The key must never be TYPED into a pane. `tmux.sendCommand` used
`send-keys` with a `KEY='…' cmd` prefix, which the dashboard streams to
every attached browser and bash appends to history; it now takes over the
idle pane with `respawn-pane -k -e`, passing env through argv the way
`newSession` always has. Verified on tmux 3.4: multi-element argv arrives
intact and the pane renders no trace of the value.
- Attaching COPIES a profile, so editing or deleting one in Settings has to
reach the lanes already using it. An edit re-applies it (a rotated key
would otherwise never reach the lane still holding the revoked one); a
delete resets those lanes to native and removes the Codex layer file.
Both report per-lane failures, which the tab surfaces.
`.claude/settings.local.json` is added to the lane's git exclude, for both
native and router lanes: `agenthub lanes reset` runs `git clean -fd` without
`-x`, so an un-excluded settings file is deleted on reset — and untracked
meant the key was committable.
DB: `lanes.router_profile`, nullable, additive. Written only by
`setRouterProfile` beside the disk write, never through `PATCHABLE`.
This project's marketplace entry is `"source": "./"` — the plugin IS the
marketplace repo — but pluginMarketplaceUrl only accepted a URL string or
`{url}`, so it returned null and every plugin-cache install answered
`no_remote`: no install of this plugin could ever see a published release.
A relative source now falls back to the manifest's own `repository`. An
`owner/repo` shorthand still returns null — it names a different repo, so
borrowing the marketplace's URL would be a confidently wrong answer.
The residue an interrupted update leaves behind is invisible: the server
runs, every doctor row reads OK, and the unit quietly still names a cache
directory Claude Code deleted during the update. Nothing surfaces it until
a reboot, when the dashboard does not come back and nothing recorded why.
- fastPathOk now rejects a stale unit, so the next SessionStart repairs it
instead of short-circuiting. `unitIsStale` is pure fs on purpose: that
path runs on every session and may not spawn a process.
- Repairing the unit is not restarting the server. A current, healthy
server keeps running and the corrected unit takes effect at the next
start (`pendingSystemdRestart = systemdOwned && needsServer`), and the
detached-spawn fallback is guarded by needsServer so a stale unit alone
cannot start a second server onto a bound port.
- systemdOwned now also requires the unit to be ACTIVE. Installed-but-
stopped owns nothing: skipping the stop there and asking systemd to start
would bind an occupied port and crash-loop under Restart=always.
- /agenthub-doctor gains a systemd row naming the consequence ("will NOT
come back after a reboot") and the fix. WARN, never FAIL — the dashboard
in front of you works, and having no unit at all is a normal setup.
The fast-path test pins the regression against an otherwise fully healthy
install, live server included; without a live server it would have passed
against the unfixed code either way.
A plugin self-update took the dashboard down and left it down. The chain
ran correctly up to the point where the bootstrap SIGTERMs the server it is
replacing — and that server was systemd-managed, with the bootstrap running
as its descendant, i.e. inside its cgroup. Two failures fired at once:
* the server exits 0, and `Restart=on-failure` reads a clean exit as
"meant to stop", so systemd did not bring it back;
* `KillMode=control-group` reaped the bootstrap itself, mid-run, before it
could repoint the unit at the new plugin root or write state.json.
Evidence from the incident: bootstrap.log ends at "stopped the server
started from the previous plugin version" with no systemd line after it,
.bootstrap.lock survived with a dead pid, state.json still named the
previous version, and the unit's ExecStart still pointed at the old cache
dir. The same truncation is visible on three earlier updates — those only
recovered because a later SessionStart bootstrapped again.
systemd-service.js now splits install into writeUserService (repoints the
unit, starts nothing) and restartUserService (--no-block). The bootstrap
skips the manual stop whenever a unit of ours owns the server, and defers
the restart past writeState and releaseLock, so being killed by its own
request costs nothing. The unit is Restart=always: a graceful SIGTERM from
anything must no longer read as "leave it stopped".
installUserService keeps its old semantics for `agenthub service install`,
which runs from a user's shell and is not in the cgroup.
`agenthub update-check` prints `message` as its headline, so every
plugin-cache install saw the literal "null" where the verdict belongs — and
that is the one install kind with no commit count to fall back on, so the
line carried nothing at all.
The plugin branch of getUpdatesStatus compares versions exactly like the
checkout branch does; it now builds the same four sentences from what it
already computed (newer release / newest / no local version to compare /
release lookup failed).
The nogit regression test asserted `message === null` as a proxy for "did
not fall through to the not-a-git-clone payload" — it now asserts that on
the message's content instead, which is what it always meant.
Starting a run on a machine without tmux surfaced `spawnSync tmux ENOENT`
as a 400 — blaming the request for a missing host package, naming no fix,
and reaching the user as a raw errno in the console pane.
The guard goes in spawnRun, which both spawn paths (POST /api/run and
POST /api/lanes/:id/start) already route through, so one check covers
them; the message carries the install command. Both routes map the code to
503: the request was fine, the dependency is not installed. The client
already renders `error.message` from the body verbatim, so no UI change.
Also: README's node advice said "stay on 20/22/24"; 1712 server + 488
client tests are green on v26.8.1, so only 25 is actually the bad one.
The READMEs never said tmux was needed, so a fresh install only found out
when Start Run failed with nothing explaining why. Both doctors were silent
about it too.
- README/README.vi: a prerequisites table (node, npm, git, tmux, claude,
codex, gh, systemd) and per-OS tmux install + verify steps; the feature
list is regrouped into observe/drive/configure and now covers analytics,
alerts, import, remote sources, the config explorer and Update all.
- INSTALL: Step 0 for the system packages, plus a troubleshooting entry
separating "tmux missing" from "the server has a different PATH".
- SETUP/docs/README: the same troubleshooting entry and a pointer to Step 0.
- plugin-doctor: a tmux row, WARN and never FAIL - the run console and
`lanes shell` break without it, nothing else does. Reuses tmux.js's own
isTmuxAvailable().
- agenthub doctor: asks GET /api/run/tmux instead, because the PATH that
decides whether a run starts is the server's, not the caller's shell's.
After the 1.5.2 -> 1.5.3 self-update the server ran 1.5.3 while the dashboard
still served the 1.5.2 UI. The client build had failed, and `runtime/client-dist`
still held the previous version's bundle, so nothing looked broken — the only
signal was the Settings page's "UI build v1.5.2 (rebuild client to match)".
Cause: the whole client build toolchain (vite, typescript,
@vitejs/plugin-react, tailwindcss, @types/*) is a devDependency, and npm reads
NODE_ENV=production as --omit=dev (`NODE_ENV=production npm config get omit` ->
`dev`). The dashboard server runs with NODE_ENV=production, and every bootstrap
descended from it inherits that: the plugin-update chain, and any SessionStart
fired by a `claude` running in a tmux pane the server started. So `npm ci`
dropped vite, `tsc -b` died on missing type declarations, and buildClient
returned "failed" with the stale bundle left in place.
`--include=dev` on both install paths — a CLI flag outranks the env, which is
the point. Verified by running buildClient with NODE_ENV=production against the
installed plugin root: previously "failed", now "built", and the served bundle
carries 1.5.3.
Also: plugin-update-helper writes its verdict BEFORE the bootstrap handover as
well as after. That step retires the process tree the helper hangs off, and the
observed run reached the new server with no verdict file at all — the dashboard
reported `last_plugin_update: null` for an update that had actually succeeded.
`dashboard_runs.mode` is NOT NULL and `recordRun` swallows insert errors by
design, so pty-run.js passing `mode: null` meant NO tmux-backed run was ever
persisted — the newest row in a live database dated from before pty-run.js
replaced run-spawner.js. `publicRun` reads cwd, model, prompt preview,
started_at and session_id off that row and the handle itself is recomputed
from tmux, so the runs modal rendered a live row with no cwd, no model, no
prompt and a start date of Jan 1 1970. The same missing row is why a resumed
session kept its separate `external` badge: the client dedupes an external
session against the session ids carried by run rows, and no run row ever
carried one.
- pty-run.js records `mode: "interactive"` — every run here is one pane.
- dashboard-runs.js still never fails a live run on a db error, but warns
instead of swallowing silently; that silence hid this for six weeks.
- RunHistory renders `startedAt: 0` as "-" rather than dating a row with no
known start time to the epoch.
Also closes the other half of the same lie: adopting a pane that is already
running an agent now throws ERUNLIVE (409 from both the lane start action and
POST /api/run) when the request carried a prompt or a resumeSessionId. That
argv cannot be typed over a live agent, so answering 200 reported a resume
that never spawned. A start asking for nothing but a pane still adopts.
An archive-sourced marketplace entry ("source": "url" — this project's own)
makes Claude Code unpack the plugin rather than clone it, so the cache dir
has no .git. getUpdatesStatus probed for .git first and returned "not a git
clone; check for updates manually", skipping the entire release layer:
update_available stayed false however many releases were published, and
plugin_marketplace/plugin_name were never set, so POST /api/updates/apply's
plugin branch could never fire (it answered 409, claiming nothing newer was
published). The bug reads as "up to date", so it stayed invisible for as
long as the install happened to be current.
The plugin branch now runs BEFORE the .git probe and answers without git at
all: the repository URL comes from the marketplace manifest on disk
(known_marketplaces.json -> .claude-plugin/marketplace.json -> source.url,
null rather than a guessed host for a github owner/repo shorthand), and
package.json is the only local comparison signal. A failed release lookup
is reported as release_check, never as "nothing newer".
UI/CLI: the "not a git clone" note is gated on install_kind !== "plugin", so
the Update now button is offered where the dashboard can actually check and
apply.
nntrivi2001