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).
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.
`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.
`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.
The plugins tab could already update one plugin or one marketplace at a time. It gains an install-wide batch and, separately, the reload step that actually makes an update take effect.
Update all runs in a DETACHED worker rather than the request handler, for a structural reason: the batch ends with the plugin the dashboard ITSELF runs from, and that step's plugin-bootstrap.js --force SIGTERMs this very server. An in-process loop would die partway through with the remaining plugins unrun and nothing left to report. Order is marketplaces (one call, no name = all) -> every other installed plugin -> AgentHub last -> one reload. Each transition is journalled to cc-update-all.json in the runtime dir, outside the plugin cache the update replaces, so whichever server is alive serves progress from it; the POST's response is only an acknowledgement.
readUpdateAll DERIVES an abandoned status for a running journal older than an hour, because a detached worker killed by a reboot leaves nobody to write the terminal entry and the button would stay wedged forever. isRunning gates a concurrent batch: two claude plugin update chains over one config dir race.
Reload is its own capability (server/lib/cc-reload.js, POST /api/cc-config/reload). /reload-plugins and /reload-skills are scoped BY CLAUDE CODE to the session that runs them, so nothing can reload a session from outside - typing into a session already running is the only real mechanism, and that is what this does, over tmux send-keys into the dashboard's live lane panes. It spawns NO claude process on purpose: a one-shot claude -p "/reload-plugins" would reload only its own short-lived session, exit, and leave a junk session in the dashboard's history - a headless run gets a full sessions row and there is no filter anywhere to suppress it. Reusing live panes means no session is created at all. Panes at a shell prompt are skipped (the text would run as a shell command) and only claude lanes are addressed, since Codex has no such command.
Update-all reloads once at the end; updating a single plugin passes reload:true and reloads immediately, and never after a failed action.
Fixes a real pre-existing bug on the way: pluginUpdate omitted -y while runClaude closes stdin, and the CLI requires it whenever stdin is not a TTY - so the per-plugin Update button failed for any plugin whose marketplace declares an install command. Also completes CcConfigChangedPayload's action union, which claimed write|delete while nine other values were already being emitted.
Verified: server 1683/1683, client 486/486 (27 new tests), build, prettier, file-header audit, openapi.yaml regenerated (117 paths). Update all was NOT executed against this machine - it would update 7 real plugins and restart the running dashboard.
Update detection was a commit count against the default branch, which says nothing about whether a *version* shipped and is pure noise on the plugin cache (a depth-1 clone, where every distance is a graft artefact). It now reads the forge's own releases API and compares the published tag to this install's package.json version, falling back to the commit count only when no release can be read — with the reason in release_check, so "up to date" never silently means "we could not look".
Two states a commit count cannot express: dev_build (ahead of the newest release, missing nothing from it) now reads "Dev build - N ahead" instead of a false "Up to date"; and on a shallow clone every commits_*_release field is null and dev_build is never asserted, because the numbers would be meaningless.
POST /api/updates/apply gains a second mode, chosen by install kind and never by the request body. A checkout still fast-forwards onto the release tag, rebuilds and self-restarts. A ~/.claude/plugins/cache/ install is refused by applyUpdate (that directory is Claude Code's to replace) and handed instead to a detached chain: claude plugin marketplace update -> claude plugin update -y -> plugin-bootstrap.js --force from the NEW cache dir, which retires this server and starts its replacement. Both write the HTTP response before anything stops the process.
The chain records every outcome to plugin-update.json in the runtime dir, outside the cache the update replaces, so the next server reports it as last_plugin_update: a failed chain is otherwise indistinguishable from a slow one, and the tab would poll a server that never returns. /reload-plugins is deliberately not in the chain - it is REPL-scoped to the session that runs it, so no background process can perform it, and plugin-bootstrap derives PLUGIN_ROOT from its own __dirname, which is what makes it unnecessary for the dashboard.
Also rebrands the sidebar subtitle, splash and page title to SmartGift Solution, and corrects three ARCHITECTURE.md statements that claimed /apply did not exist.
Verified: server 1683/1683, client 486/486, build, prettier, file-header audit, openapi.yaml regenerated.
The repo shipped five install/publish surfaces that nothing in the project builds, tests or documents as supported any more. Each one carried its own stale docs, and the Docker/Podman path additionally described a Dockerfile whose COPY of statusline/ would have broken the build the moment that directory went away.
Removed: vscode-extension/ (16 files, marketplace extension for an ID we do not own); statusline/ (3 files, standalone CLI statusline); the repo-root landing/wiki PWA assets (sw.js, manifest.json, robots.txt, sitemap.xml, og-image.svg, favicon.svg, fonts/ - 17 files for pages that no longer exist); .superpowers/ (4 files of scratch brainstorm state, already covered by .gitignore:55); and the container publish path (Dockerfile, docker-compose.yml, .dockerignore, 5 Makefile targets, 2 npm scripts).
Kept deliberately - these are about the USER's containers, not ours: scripts/install-hooks.js container detection and AGENTHUB_ALLOW_CONTAINER_HOOKS (issue #193), server/lib/lane-detect.js docker-compose service detection, server/data/profile-templates/postgres-compose/, server/lib/stage-detect.js 'docker push' detection, the host.docker.internal Prometheus notes, the Config Explorer's reads of the user's own ~/.claude/statusline.py, and .devcontainer/ (its docker-in-docker feature is dropped since it existed only to build the removed Dockerfile).
Docs follow the code: ARCHITECTURE.md loses the VS Code, statusline and container sections plus two dead TOC anchors, and its PWA section is rewritten from three PWAs to one; INSTALL.md, SETUP.md, docs/DEPLOYMENT.md, server/README.md, client/README.md, docs/README.md, docs/HOOKS.md and docs/DATABASE.md drop the same. The Makefile target count in ARCHITECTURE.md is corrected 25 -> 17 (it was already stale).
The ten `agenthub-*` plugins were listed by the marketplace but could never be
installed — two schema mistakes present since the entries were first authored
(57dc915), both invisible until you actually run `claude plugin install`:
- marketplace entries used `"path": "plugins/<name>"`; the CLI only accepts
`"source"`, so every one failed with
`This plugin's marketplace entry is invalid: source: Invalid input`
- each `plugin.json` carried npm's `"repository": { "type", "url" }` object;
the schema wants a plain string, so the next attempt failed with
`Validation errors: repository: Invalid input`
Verified by installing all 11 entries into a throwaway CLAUDE_CONFIG_DIR:
1/11 before, 11/11 after. `plugins-marketplace.test.js` was asserting the
broken shape, which is why CI stayed green — it now asserts `source` and
rejects `path`.
Marketplace renamed `claude-code-agent-monitor-plugins` -> `agenthub`, so the
install line is `agenthub@agenthub`. ARCHITECTURE.md documented a marketplace
name (`smartgift-claude-code-monitor`) that never existed in any revision.
The documented `/plugin marketplace add Smartgift-AI/Claude-Code-Monitor` also
never worked: the `owner/repo` shorthand resolves to github.com only, and this
repo is self-hosted on Gitea. Confirmed the CLI rejects `host/owner/repo` and
ignores GH_HOST / GITHUB_HOST / CLAUDE_CODE_MARKETPLACE_URL, so the full
https:// clone URL is the only supported form. Same GitHub-habit class fixed
elsewhere: raw.githubusercontent.com icon URLs (404, and the repo ships
favicon.svg not favicon.ico — the client had the same dead /favicon.ico),
a smartgift.github.io docs link (404), and Gitea `/blob/<ref>/` paths that
must be `/src/branch/<ref>/`.
Repository renamed to Smartgift-AI/AgentHub throughout. Gitea 301-redirects
the old name, so these were stale rather than broken; `docs/superpowers/`
keeps the old name because those are dated decision records.
Adds `repo-links.test.js` to guard both classes (it runs under the existing
pre-commit gate) plus an openapi.yaml drift check — the committed spec had
silently fallen a minor version behind package.json. `generate-openapi-yaml.js`
now exports `render()` so the test compares against one definition of the dump
options instead of a second copy.
Lane Pipeline Pass Rates read `lanes.stages`, which is a snapshot of the
pass in flight: re-entering a node overwrites its record, and moving
backwards deletes every record ahead of the new stage. The report treated
one lane row as one run, so shipping ten features through a lane still
counted as one, and no node could ever exceed 1/1. History survived only
via `lane_features`, which is written when a feature is activated — a
path most users never take. The panel was not wrong, it was structurally
incapable of counting.
Add `lane_stage_events`, an append-only log written by every declaration
and every accepted detection, and aggregate from that. A run is one pass
through the template: the log starts a new one when the lane moves
backwards or leaves the template's last node. Existing snapshots are
seeded as one run each on first boot, so upgrading does not read as
"never run". Templates a lane is configured with are still listed at
zero — a pipeline going unused has to be visible as such.
Count inferred reaches too, in their own column and their own colour.
Detection stays out of the pass rate ("inference is never evidence"), but
reporting a stage the hooks demonstrably saw as "never reached" is what
made the panel look broken to anyone who works without running the
pipeline commands.
Also reject a stage naming no node of the lane's pipeline (400 EBADSTAGE,
listing the ones that exist) instead of storing it with a warning nobody
reads. A real lane had accumulated five such keys, one of them `test` on
a ship-feature lane whose CI node is `gates` — which is why that row read
never-reached while the stage had in fact been declared. Aliases still
resolve, so shorthand is unaffected.
The Memory tab could only show this machine's user memory and the
dashboard's own project file. Setting up a project you had not run
through the dashboard meant opening a terminal — `PUT /file`'s
"project" scope always resolves against the server's cwd, on purpose.
Add one narrow route per tool that writes into the directory you name:
CLAUDE.md for Claude Code, AGENTS.md for Codex, both through the same
`project-memory.js`. Being the only write path that honours a
caller-supplied directory, it pays for that:
- the directory must be absolute and already exist — nothing is
created for you;
- the filename comes from the route, never from the body;
- an existing file, symlink included, is a 409 — this can only
create, never overwrite;
- it sits behind the loopback-Origin guard /api/run uses, so a page
in the user's browser cannot plant agent instructions in a project
directory. That guard moves to lib/same-origin.js to be shared.
The modal takes the dashboard's known project dirs as suggestions on a
free-text input rather than a closed list: a project it has never seen
is exactly the case this exists for.
Mutation-checked: dropping the origin guard, and swapping lstat for
stat, each fail a test.
`autoInstallSkillPlugins` takes `run` as a seam, so every existing test
stops short of the real `spawnSync` — the line that decides whether
CLAUDE_CONFIG_DIR is pinned was untested. Restoring the unconditional
pin here (the bug that sent `.claude.json` to `~/.claude/.claude.json`)
kept the whole suite green.
Export `runClaudePlugin` and drive it against a fake `claude` on PATH
that prints the variable back: one test for CLAUDE_HOME set, one for
unset. Mutation-checked — the unconditional pin fails the second.
skill-plugins.js carried its own copy of the unconditional pin, with the
same comment claiming it keeps writes where cc-discovery reads. Same rule
as runClaude now: pin only when the operator set CLAUDE_HOME.
Also corrects three comments that described the old behaviour — the
confidently-wrong pair at these two call sites is what let two readers
believe reads and writes were aligned when they were not.
The previous commit routed the read through getClaudeHome(), which appends
/.claude when CLAUDE_HOME is unset — pointing the dashboard at
~/.claude/.claude.json, a file the CLI never writes by default. Verified
against the real binary: with CLAUDE_CONFIG_DIR unset the CLI writes
$HOME/.claude.json; with it set, .claude.json sits flat inside that dir.
Branch on CLAUDE_HOME accordingly.
The test now drives a fake claude that writes where the real one does and
asserts getClaudeJsonPath() names that exact file, under both branches —
the previous test only restated the implementation.
getClaudeJsonPath() was the only one of cc-discovery's 20 read paths not
derived from getClaudeHome(), so a CLAUDE_HOME install read one file while
the CLI wrote another. The default case was fixed by dropping the
CLAUDE_CONFIG_DIR pin; this closes the mirrored half.
Also replaces the Remove button's display-string comparison with a
removable flag from the reader: rewording the source label silently
resurrected a button that always fails.
runClaude() unconditionally set CLAUDE_CONFIG_DIR to getClaudeHome()
(~/.claude), so the CLI wrote ~/.claude/.claude.json — a file
cc-discovery.getClaudeJsonPath() never reads (it reads ~/.claude.json,
beside ~/.claude/, by design). Every default-install MCP add/remove and
plugin/marketplace mutation was inert: add toasted success and vanished on
refetch, remove 422'd for every server the panel lists.
With CLAUDE_HOME unset (the default), leave CLAUDE_CONFIG_DIR unset too —
the CLI's own default config dir already lines up with what cc-discovery
reads, so alignment comes from doing nothing. Only pin it when the operator
explicitly set CLAUDE_HOME, which is the one case the pin is meant for.
Left a residual, documented at both sites: with CLAUDE_HOME set, reads still
come from $HOME/.claude.json while the CLI writes
<CLAUDE_HOME>/.claude.json — a split predating this task, affecting a
configuration nobody here runs.
New tests pin the env contract directly (no real CLI needed): runClaude
must not set CLAUDE_CONFIG_DIR when CLAUDE_HOME is absent, and must set it
to CLAUDE_HOME's value when present.
readMcpServers only ever parses ~/.claude.json (top-level = user scope,
projects[<root>] = the panel's "projectScoped" bucket, which the CLI itself
calls `local`). The previous user/project scope choice was the exact
inverse: `--scope project` writes .mcp.json, which readMcpServers never
reads (a successful add would vanish), and every "projectScoped" row was
actually `local`, so Remove sent a scope the CLI could never match.
- cc-mcp-cli.js: MCP_SCOPES -> ["user", "local"]; drop the local
mutationQueue/serialize in favor of cc-plugin-cli.js's exported one, so a
plugin install and an MCP add can't spawn `claude` concurrently; drop the
String(v) coercion in env/header flags (a wrong-typed value is now a 400,
not a written "[object Object]"); assertEmbeddedValue allows empty strings
(a legitimate empty stdio arg was previously a 400).
- cc-plugin-cli.js: export `serialize` for reuse.
- CcConfig.tsx: CcMcpScope narrowed to user/local; the projectScoped bucket's
McpCard now sends scope="local"; rows sourced from
~/.claude/settings.json (no CLI scope addresses them) no longer render a
Remove button. Cleaned up i18n keys my own earlier change made dead
(explain.mcp.addCmd/remove/removeButton) and gave the scope <select> its
own short option labels instead of reusing full-sentence section headings.
Mutation goes through `claude mcp add|remove`, never a direct edit of
~/.claude.json — the running CLI rewrites that file concurrently, so a
dashboard write would race it. Reads stay on readMcpServers.
The plan specified a --url flag; the real CLI (v2.1.266) has none —
commandOrUrl is positional, and --header must trail it or the variadic
flag swallows the url. Built against the verified shape.
The Claude writer's containment was lexical only: a symlink at
~/.claude/<type>/<name>, or at any ancestor of it, is a string inside the
root while the write lands wherever it points. The backup destination had
the same exposure, and delete copies through the link before unlinking it.
PUT/DELETE /file also honoured ?cwd=, so a project-scope write could be
aimed at any directory on disk. PATCH /settings already refused that, with
the reasoning in a comment; the two other write routes never adopted it.
The guard moves into config-write.js so both writers run one copy.
--env's value is the RHS of one KEY=VALUE argv element, never its own
argv position, so a leading "-" can never be parsed as a flag — unlike
the key, which stays validated as an identifier. assertArg's leading-dash
guard was reused for the value too, refusing ordinary values like "-1".
Split it into assertEnvValue (no leading-dash check) for values only.
Plugins/marketplaces GET routes were silently empty because codex-cli.js
never unwrapped the CLI's {installed,available}/{marketplaces} objects
into arrays; backup writes bypassed the symlink-escape guard entirely;
readFileSafe was stricter than every other Codex reader, showing a
symlinked skill/rule in the list and then refusing to open it;
marketplaceAdd only accepted http(s) URLs, rejecting owner/repo sources
the CLI itself takes. Also: docs drift, a dead i18n key, and a scope-cast
hazard in the shared Editor/ConfirmDelete modals.
Fixes, one per finding: unwrap installed/marketplaces in codex-cli.js so
every list route keeps returning a bare list; guard the backup path in
codex-mutate.js with the same assertNotSymlinkEscape the write path
already used (Claude's cc-mutate.js has no such guard on its primary
write path either, so left untouched there); relax readFileSafe's
containment to lexical-only, matching the other four Codex readers,
while leaving the write side's realpath check untouched; reuse
cc-plugin-cli's assertSource for marketplaceAdd instead of a third
validator (ssh remains out of scope, no credential story for it yet).
Server: 1557/1557 (was 1548). Client: tsc -b clean, 469/469 tests.
Round 2's escape check rejected a quote whenever a single backslash
preceded it, so an escaped backslash before a real closing quote (e.g.
path = "path\\" # c) was misread as an escaped quote and the field was
dropped. Now counts the run of backslashes and only treats an ODD run as
escaping the quote. Folded the old two-branch quote handling into one
scan (it already finds the terminator correctly with or without a
trailing comment) and decode \\ -> \ in the extracted value, since the
scan already knows which backslash run is real.
readTomlMeta's trailing-comment stripper found the first quote character
regardless of escaping, so a description containing \"escaped quotes\"
before a real comment got silently truncated instead of dropped. The scan
now skips a quote preceded by a backslash.
The "never walks outside" containment test passed by construction (row.file
is always built off a fixed root) and would not fail if the isUnder guard
were removed. Replaced it with two tests that actually pin the regression
risk: a skill dir and an agent .toml symlinked in from outside CODEX_HOME,
asserting the row is discovered and its content read through the link.
isUnder stays as a defensive check (matches cc-discovery.js), now commented
as such.
readTomlMeta was dropping any field with a trailing `# comment` (common in
real agents/*.toml). readSkills/readAgents/readRules used bare
isDirectory()/isFile(), which miss symlinked entries per cc-discovery.js's
own documented Dirent quirk — reused its isDirLike/isFileLike instead
(exporting isDirLike from cc-discovery.js, which wasn't previously public).
Also swapped the "never walks outside" test for one that actually exercises
a symlink escape, and dropped the unnecessary isUnder re-export.
Two leftovers from the rebrand, done together because they touch the same
docs and manifests.
**Data directory.** `agent-dashboard` could not simply be renamed in the
source: it names a real directory holding the live SQLite database. Startup
now moves it once, via `migrateLegacyDataDir()` in `server/lib/claude-home.js`
— a same-filesystem rename, so it is instant even at multiple gigabytes and
cannot leave two divergent copies. `getDataDir()` stays PURE and resolves to
the legacy path until the move happens; the first version of this put the
rename inside the getter, which meant `npm run test:server` relocated the
developer's real database as a side effect. The call sits behind
`require.main === module` and ahead of the router requires, because those pull
in db.js, which opens SQLite at module load. A failed rename is silent and the
old directory keeps being used: an unfashionable directory name beats a blank
database beside an unreachable one. Same treatment for the port-discovery file
(`.agent-dashboard.json` → `.agenthub.json`, old one dropped on write).
Not renamed on purpose: the Sagi localStorage keys (renaming silently resets
everyone's prefs) and the PWA manifest `id` (changing it orphans an installed
app).
**MCP server.** 4 tool calls in 46 days across 577,132 events, against 4,556
lines of TypeScript, a 2 MB committed build artifact, and a pre-commit guard
that had already blocked commits twice. Deleted: `mcp/`, the maintainer skills
and `mcp-reviewer` agent, the build check, `plugins/agenthub/.mcp.json`,
`docs/MCP.md`, 12 npm scripts, 5 Makefile targets, and every doc section
describing it.
Kept, because they are unrelated things sharing the name: the Config
Explorer's MCP tab and its API, per-lane Playwright MCP, the agenthub-config
plugin's MCP audit, and the historical `mcp__plugin_agenthub_*` events that
analytics must keep attributing.
Verified: server 1432/1432, client 459/459, tsc clean, prettier clean. Both
directions of the migration proven — a sandbox-HOME test run leaves the legacy
directory untouched, and a real boot moves it with its contents intact.
Four defects, all from one design mistake: a profile's display name and its
filename were the same string.
- Saving the logged-in account made a SECOND profile of an account another
profile already held. Both then reported themselves active, so there was
nothing to switch between. `POST /profiles` now refuses `EDUPLICATE` and
names the profile that already holds the account. Re-saving into the
account's own profile stays allowed - that is the token-refresh path.
- Worse, an accidental duplicate could not be cleaned up: two profiles of one
account are both active, and the delete guard refused any active profile, so
neither was ever deletable. The guard now blocks only the LAST saved copy of
the live account, which is the actual credential-loss condition it exists
for. The trash button had to follow, or the server would allow what the UI
hid.
- A name could not contain spaces or diacritics, because it had to be
filename-safe. A profile now carries `name` (internal identifier, still the
filename, still what every route takes) and `label` (what the user reads and
edits: 1-60 free-form characters, control characters rejected, stored
trimmed, unique case-insensitively). A label never touches a path, which is
what makes "Tài khoản chính" a legal name.
- Naming happened up front in a "save current account as" field. That field is
gone. The server derives the identifier, the save button lives on the live
account's card and appears only while it is unsaved, and the label is edited
by clicking it (Enter saves, Escape cancels).
Renaming rewrites the label in place and deliberately does NOT move `name`, so
it cannot half-fail and leave a profile unreachable. Profiles written before
labels existed report `label` equal to their `name`, so nothing needs migrating.
Adds `PATCH /api/cc-config/profiles`; `POST` now takes an optional `name`.
Verified: server 1431/1431 (11 new), client 459/459, tsc clean. Also exercised
end to end against a COPY of the real profile store on this machine: the
duplicate deleted, the last copy still refused, a Vietnamese label accepted,
and the stored access/refresh tokens and 0600 mode intact after a rename.
Adds a "Switch profile" button left of Backups on /settings?tool=claude-code.
A *profile* is one Anthropic account - one quota - saved under a name, so the
local `claude` CLI can be pointed at a different account without disturbing
anything else.
A switch swaps exactly two identity artifacts and nothing more: `claudeAiOauth`
inside `<CLAUDE_HOME>/.credentials.json`, and `oauthAccount` + `userID` inside
the `.claude.json` config file(s). Those files are key-merged in place, and only
where they already exist, so `projects`/`numStartups`/everything else survives.
Skills, subagents, commands, memory, plugins, marketplaces, MCP servers and
settings are never read or written here - that is what makes a switch keep the
whole workspace and change only the account. `mcpOAuth` stays put too: those are
the machine's MCP server tokens, not account identity.
Profiles are one JSON file each under `<CLAUDE_HOME>/agenthub-profiles/` (dir
0700, files 0600, atomic temp+rename) and hold live OAuth tokens, so no route
ever returns token material - `GET /profiles` answers identity metadata only.
The dashboard cannot log an account in; only `claude auth login` mints tokens.
Adding a profile is therefore always "log in from a terminal, then snapshot what
the CLI wrote". `activateProfile` re-snapshots the outgoing account first so
tokens Claude Code refreshed since its last save are not lost, and refuses
(`EUNSAVED`) to switch away from an account that was never saved - that would
discard the only copy of its credentials. A switch applies to new sessions only,
since a running session rewrites the config on exit.
Routes: GET/POST/DELETE /api/cc-config/profiles, POST /profiles/activate.
Verified: server 1420/1420 (14 new, incl. token-free responses and 0600/0700
permissions), client 459/459.
Clean break, no compatibility alias. Renames the product name, the CLI
binary (`ccam` -> `agenthub`), the env var prefix (`CCAM_` ->
`AGENTHUB_`), the Prometheus metric namespace (`ccam_*` ->
`agenthub_*`), the tmux session prefix (`ccam-lane-<id>` ->
`agenthub-lane-<id>`), the per-repo runtime profile directory
(`.ccam/profile` -> `.agenthub/profile`), the lanes state root
(`~/.claude/ccam-lanes` -> `~/.claude/agenthub-lanes`), the bundled
plugins (`plugins/ccam*` -> `plugins/agenthub*`) and their slash
commands (`/ccam-doctor` -> `/agenthub-doctor`, `/ccam:*` ->
`/agenthub:*`).
Because there is no alias, an existing install has to be updated by
hand: rename its `.ccam/profile` directory and any `CCAM_*` variables,
and reinstall the plugin under its new name.
Two other name families are deliberately left alone here, since they are
not just strings: the `Claude-Code-Monitor` repository/marketplace name,
and the `agent-dashboard` npm package plus its `~/.claude/agent-dashboard`
runtime directory, which holds the live database and needs a migration
rather than a rename.
Verified: server 1406/1406, client 459/459, `grep -i ccam` clean across
tracked files, `openapi.yaml` regenerated (95 paths, unchanged count).
`plugins` sat in WATCH_SUBDIRS, so it got fs.watch({recursive:true}).
On Linux that's a userland tree walk costing ONE inotify watch per
directory, and ~/.claude/plugins is the plugin cache: full git checkouts
with node_modules, 44k+ directories. The dashboard process held 65,464
watches against the default fs.inotify.max_user_watches of 65536,
leaving none for systemd/IBus — IBus exited on boot and Unikey died
with it.
Config Explorer only reads two top-level files out of that tree
(installed_plugins.json, known_marketplaces.json), so watch plugins/
flat: 1 watch instead of 44k. The small config subdirs stay recursive.
Measured on the running service: 65,464 -> 96 watches, and touching
installed_plugins.json still broadcasts cc_config_changed.
The ship-feature pipelines drive `/superpowers:*` commands as their per-node
drivers, so a fresh install had a dependency it was only told about — the
`/ccam-doctor` WARN row and a line in the README.
`server/lib/skill-plugins.js` now carries the whole thing, shared by the
doctor (reports) and the bootstrap (installs) so the two cannot disagree
about what is missing. The bootstrap installs at the very end of its DETACHED
worker — the install clones from the network and nothing in the 10 s
SessionStart hook may wait on it — retrying once behind a `marketplace add`,
logging a failure with the manual command, and never throwing: the dashboard,
hooks, CLI and MCP all work without these plugins. The outcome is recorded as
`state.skillPluginsSettled`, which `fastPathOk()` now also requires, so the
check runs exactly once per install and `/ccam-update` is the retry.
`CCAM_AUTO_INSTALL_PLUGINS=0|false|off` opts out.
Reporting is derived, installing is allowlisted. Requirements come from the
loaded templates, which is right for a report; installing from that same
source would not be, because a template is a file anyone can drop into
`DASHBOARD_PIPELINES_DIR` — it would turn "edit a JSON file" into "run this
vendor's install command on the next session". Auto-install consults
`AUTO_INSTALLABLE`, a constant in code (`superpowers` today), and reports
every other requirement as skipped.
Tests inject a `run` seam so no case reaches a real `claude plugin install`,
and cover the refusal explicitly.
A pipeline template named ONE command for the whole run (`skill`), so the
only way to know what actually moves a given node was to read that node's
`detect` regexes. Each node now carries its own `command`: a slash command
(`/superpowers:writing-plans`) or a shell one (`ccam lanes hook ci-gate`),
whichever its detect rules already watch for.
The values are derived from each node's own detect rules, not invented —
every `ccam …` string was checked against bin/ccam.js and every
`/superpowers:*` against the installed plugin. Nodes reaching which is a
declaration rather than an action (`done`, `merged`, `reported`,
`e2e-feature-passed`, `qc-plan`) are deliberately left empty.
`pipelines.js` normalizes a missing or non-string value to "", so a
custom template written before this field keeps loading rather than
rendering "undefined". `nodeStates()` and GET /api/lanes/pipeline-stats
pass it through. Never executed by the dashboard — display text, the same
contract `skill` already had.
Surfaced where each view has room for it: the Workflows pass-rate panel
shows it under the node label and copies it on click; PipelineMap folds
it into the node tooltip, because a 16-node pipeline has to fit across
one card and there is no width for a second line.
`command?: string` is optional in the client type only to spare four
unrelated test fixtures — the server always sends the key, and the UI has
to branch on the empty string regardless.
Verified: npm run test:server (1373, +3), npm run test:client (459, +2),
tsc --noEmit, prettier and check-headers clean. docs/LANES.md § Custom
template format documents the field. openapi.yaml needs no change: the
lane payload and pipeline-stats were never schematized there.
A plugin card showed "7 skills, 5 agents" and nothing more — the counts
were the whole story, so finding out WHICH skills a plugin ships meant
opening its install directory by hand. The contribution chips are now
buttons: clicking one expands a list naming each artifact with its
description, one surface open at a time. Commands render with their
slash. Hooks stay a plain chip — a plugin's hooks are matcher/command
pairs declared inline in plugin.json, not named artifacts.
`contributes.items` carries the four surfaces server-side; the existing
counts keep their type and are now derived from those lists, so a chip
reading "3 commands" can never sit beside a list holding two.
That also fixes a real blind spot. Each surface was read only from the
conventional install subdirectory, but plugin.json may declare its own
paths — this repo's own `ccam` lists three files under
plugins/ccam/commands/ and keeps commands/ empty, so it reported zero
commands. Both sources are now read and de-duplicated. A declared path
resolving outside the plugin's install directory is dropped: a manifest
is third-party data and must not turn the dashboard into an arbitrary
file reader.
Also pins the Refresh button's width: both labels share one grid cell so
the button is always as wide as "Refreshing…" / "Đang làm mới…" and no
longer resizes the toolbar when the state flips. Measuring the longest
real string beats a min-width that would be wrong in the next language.
Verified: npm run test:server (1370 tests, +6 for the new discovery),
npm run test:client (457, +3 for the expand behaviour), tsc --noEmit,
prettier and check-headers clean. Docs synced: docs/API.md, openapi.yaml
(new CcConfigPluginItem schema), ARCHITECTURE.md, docs/PLUGINS.md.
A codex session emits no hooks and rarely calls `ccam stage` on its own, so
its card sat at `idle` forever while the amber "detected" badge was the only
thing that ever moved. `codex-ingest.js` now advances the declared `stage`
from its own written detection — the one place in CCAM where an inference
moves the blue `current` ring, and only for `codex` lanes.
It stays honest: gated on `recordDetection` having written (so forward-only,
declared-wins and the detection TTL are inherited unchanged), passes NO
evidence (the node renders `passed-no-evidence` amber, never `done`), and no
template's last node carries detect rules, so a lane can never be
auto-declared `done`. `setStage` grows a `keepDetection` option — its sole
caller — so the detection columns survive the advance and the card keeps
showing which command moved it instead of looking like an agent's own claim.
Two levers then get the evidence auto-advance cannot supply:
- `codex-nudge.js` now nudges on a stage held 2+ min with no evidence,
instead of on a detection sitting ahead of the declared stage — after
auto-advance the old condition can never fire, since the two now match.
- `ccam lanes agents install` on a codex lane writes the marker-delimited
`<!-- ccam:stage:start -->` block into the lane's own AGENTS.md — the one
instruction channel Codex reads — instead of copying agent templates into
`.claude/agents/`, which it never loads. Appended, user prose preserved,
reinstall replaces exactly that block.
Verified with a real `codex exec` run: stage moved to `watching-pr` with
`evidence: null` and the detection signal intact. 1364 server tests pass.
Three defects let a lane's pipeline map claim progress that was not this
pass's work, all of them on the "going back" path a new request takes:
- `setStage` cleared the detection columns on a real transition but never
pruned the `stages` records AHEAD of the new stage. A lane that ran through
`done` and was then declared back at `plan` rendered `plan` as `current`
while `tests`/`review`/`done` still read `done`, carrying the previous
pass's evidence, and contradicting `progress` (computed from the declared
index alone). A backwards declaration now drops every record it can prove
is ahead; a key matching no node is unplaceable and left alone.
- `activateFeature` rewrote the live lane row for the incoming feature but
carried the outgoing one's `detected_*` across. A never-seen slug starts at
`stage='idle'`/`stages={}`, exactly the shape `inferredDisplayStage` renders
from detection, so the new feature opened painted with the old one's
progress — and forward-only rejected every real detection for it until the
carried-over one aged past DETECTION_TTL_MS. Now cleared, like every other
real transition; deliberately not archived, since detection is live-only.
- `codex-ingest` classified a rollout journal by raw `now - mtime`, which goes
negative for a future mtime (clock stepped back, NTP correction, a restore
that preserved timestamps) and read as "written milliseconds ago". A
finished thread's journal could replay as live work; a file claiming to be
from the future is now idle.
Also fixes `check-headers.sh`, which audited via `find` and so reported local
gitignored runtime state (a lane's own `.ccam/profile/hooks/*.sh`) that is not
this repo's source and can never be made compliant here. It now lists files
through git (tracked + untracked-but-not-ignored), so a file an agent just
wrote is still caught before it is staged while ignored state is not.
Tests: 1359 pass (4 new). Header audit exits 0.
A lane whose sessions never call `ccam stage` had no declaration at all, so
its pipeline map stayed 100% pending even though detection was tracking real
progress. inferredDisplayStage renders such a lane AT its last detection
instead (display only — the DB's declared `stage` and detection-as-evidence
rules are untouched). Also stops tmux's own "no session" stderr noise from
spamming the journal on every poll.
A codex lane's declared stage has always depended entirely on the session
reading AGENTS.md and calling `ccam stage` on its own — no hook, no wrapper,
no enforcement. Two additions make that more reliable without touching the
declared-wins / inferred-never-done invariants:
- LaneConsolePane pre-fills a codex pane's opening prompt with a plain-text
stage-declaration checklist built from the lane's own pipeline_nodes —
the same up-front priming a Claude lane gets from its pre-filled
/ccam:ship-feature-lane skill command, just as prose since Codex has no
slash commands.
- server/lib/codex-nudge.js sweeps every 60s and types a one-line reminder
into a lane's live tmux pane (tmux.sendText, a new literal-text primitive
alongside sendCommand) when its fresh detected_stage has sat undeclared
for over two minutes. It never writes stage/detected_stage itself — any
resulting declaration is still the codex session's own real ccam stage
call, just server-prompted instead of relying on memory alone.
server 1354/1354, client 449/449, tsc clean.
nntrivi2001