Chat UI:
- Each Claude/Codex session keeps its own chat/terminal view; Settings only
sets how new sessions open.
- Composer + menu: upload, project files and folders (#54, also dragged from
the Files/Changes tree), context files, recent uploads, skills, plugins.
- Activity side panel (progress, outputs, files read, uploads, memory,
skills, connectors) with a file preview; edited-file counts per turn.
- Permission-mode chip (Shift+Tab steps, read back from the pane:
server/lib/pane-mode.js, GET|POST /api/run/:id/permission-mode).
- Stop (Esc) while the main agent works; the status is checked per agent
(with session_id /api/agents ignores its status filter) and re-read every
5 s while Stop shows.
- Dropped images shown in the conversation (GET /api/run/drop/:name, the
requester's own drop dir), full-size viewer.
- A plain-text paste over 5,000 characters is attached as pasted-text.txt,
so no prompt reaches tmux/argv limits (#68).
Changes panel: every /api/lanes/:id/git* route takes ?source=<lane_sources
id> (sourceFolder; ENOSOURCE otherwise) and a folder picker switches the
whole panel to a workspace's extra folder (#70).
Server mode S3 (one Linux user per account) — experimental, verified with
fake sudo plus a P6 managed-settings spike on a test account:
- member tmux listing runs in the background (cached, 15 s => unknown);
- transcript snapshot copy is async, chunked, temp file + rename;
- Express Layer patch only on Express 4;
- tmux capturePane/sendKey go to the member's own server.
Docs: the multi-user design moved in as docs/plans/server-mode/README.md
(the AgentHub-Server repo is gone); S3-PLAN P6 result; LANES/API/openapi.
Measured on a copy of a real 1.2 GB / 253K-event database, 10 simulated
browsers + 10 hooks/s for 90 s: event-loop delay p99 1239 ms → 32 ms (max
2.0 s → 94 ms), 266 failed requests → 0, 74 → 111 requests/s served, server
RSS 633 → 346 MB. Entry bundle 975 KB raw → 394 KB (106 KB brotli).
Server
- /api/analytics: the events-table scans run in the read worker
(scoped-stats eventAggregates, read-pool job analyticsEvents); runRead takes
maxAgeMs — analytics 10 s, /api/workflows 30 s — keyed by the args
- /api/workflows/skill-usage: +event_type and a tool_name range instead of
LIKE, so it uses idx_events_tool (784 ms → a few ms)
- idx_events_session_type_tool (session_id, event_type, tool_name,
created_at) replaces idx_events_session_type: covering reads for session
stats, the hook dedupes and the import resume point (1.2 s → 20 ms); 64 MB
cache_size and 256 MB mmap_size on both connections
- WAL: the server sets wal_autocheckpoint = 0 and checkpoints PASSIVE every
DASHBOARD_WAL_CHECKPOINT_MS (5 s) from the read worker's own writable
connection; a failed checkpoint restores the automatic one (hook COMMIT p95
109 ms → 15 ms at 100 hooks/s)
- tmux: one list-sessions answers a whole GET /api/lanes and the heal beat
(liveSessions), shared for 1 s and cleared by this process's new/kill —
GET /api/run's sync spawns had blocked 60 s of a 90 s run; the heal beat
skips tmux when nothing runs; attach options chained into 2 spawns
- hooks: TurnDuration/APIError deduped against one read per hook instead of a
query per transcript entry (23 ms/hook on a 763-turn session); one lane
scan per hook; hourly upserts skip rows that did not grow; frames are sent
after the transaction commits (and after the pane binding); new_event
carries its row id; a split pane's (slot 1–9) hooks no longer rebind the
lane's session_id
- watchdog liveness probe runs ps asynchronously; transcript newline scan via
indexOf; startup subagent backfill fingerprints the files it already checked
- pty attach: backpressure (pause past 1 MB unsent, resume under 256 KB)
- static: hashed assets served from their pre-compressed .br/.gz sibling
Client
- hidden tabs neither refetch on WS events nor poll (useLiveRefresh, panes,
lane cards, git header, changes panel, settings); no per-tab hourly
/updates/check (the server already checks)
- refetch storms throttled (workspace tree/session list 3 s, /api/run only on
a run-key change, notification pane lookup only when shown); WS reconnect
jitter; terminal retry backoff 1.5 → 30 s, stops on 1008
- heavier throttles for aggregates (15 s) and detail views; Monitor keepIfSame;
GET in-flight dedupe in request()
- lazy Dashboard/History, xterm (with the first terminal), lane modals, Sagi
(only while enabled); only the active locale loads; Workspace chunk
prefetched on a /workspace deep link
- Sagi: no work while disabled, pupil via CSS variables, memo avatar
- build: .br/.gz siblings, sourcemap hidden; service worker no longer caches
/assets/ (cache v3 drops the old one)
Tests: scripts/test-server.js drops the GIT_DIR-family variables a hook
exports, so a suite run from a linked worktree's pre-commit can no longer
commit into the real repository or set core.bare.
Docs: ARCHITECTURE, DATABASE, API, DEPLOYMENT, client README,
docs/plans/perf/README.md (audit, checklist, results).
- A session picks its own AI tool: `tool` on POST /api/lanes/:id/start
(a catalog id, known + enabled, else 400 EBADTOOL/EDISABLED); a resume
keeps the tool its session ran on (`sessions.tool`). The workspace's
`lanes.tool` is now only the default.
- The workspace home's prompt box is the one place a session starts. Its
fields run in order: AI → AI Router (Native or a profile) → Model (the
tool's own list, or the router's models) → Permission (Claude mode /
Codex sandbox+approval / catalog bypass) → Effort.
- The pane's run form (RunSetup) is gone: New run, a single pane's exit and
an empty split pane's New session all go to the home. RunSetup.tsx keeps
PromptEditor, CwdAutocomplete and ModelPicker (new `choices` prop).
- The tool icon moved from the sidebar's workspace row to each session row
(right edge, `session.tool`); the split-pane header shows the run's tool.
- An ended session opens in its workspace (`&session=`, SessionReplay.tsx):
its conversation read-only plus a prompt box; only a SENT message resumes
it (resumeSessionId + prompt, first free slot). The console is only
hidden behind it.
- Analyst button beside Files/Changes in the right panel opens
/sessions/:id for the opened session.
- Gemini CLI ships its own icon (client/public/ai-tools/gemini.svg, from
zonalogo.com).
- RouterPicker always shows (Native disabled for tools without router
support); useAiRouters() exported.
- Tests: server start-with-tool cases; SessionReplay, home field order/tool/
router models, empty pane, Analyst; dropped the cwd-field suites
(Run.defaultCwd, Workspace.laneCwd). Docs: CLAUDE.md rules, API.md,
LANES.md, README, ARCHITECTURE, INSTALL.
AI Router is a session's choice, not a workspace's:
- an AI Router select beside the model picker in the workspace home and a
pane's run form (RouterPicker.tsx; hidden off Claude/Codex or with only
Native; last pick remembered in agenthub.session.router)
- `router` on POST /api/lanes/:id/start and POST /api/run; Claude gets
--settings=<data dir>/router-settings/<name>.json (0600, outranks project
and user settings, key never in argv), Codex -p <name> + key via tmux -e
- dashboard_runs.router; a resume naming none goes back to it
- profile PATCH rewrites the session files, DELETE removes them
- gone: POST /api/lanes/:id/router/apply, LaneRouterRow, the Add-workspace
router radio, the claude-/codex-home `routers` answers; clearLaneRouters
takes the old keys out of each workspace's settings.local.json at start
Sidebar: Settings pinned to the foot, the collapse toggle beside it (stacked
on the rail) instead of in the brand's corner.
Icons: shipped SVG marks for claude, codex, openclaw and hermes (a
black/white pair swapped by theme) in client/public/ai-tools/, used by
AiToolIcon before the server's favicon.
Two optional named themes from Orca's neutral palette join the Appearance
picker (Orca Dark keeps a blue accent so white-on-accent buttons stay
legible; Orca Light's status colours are darkened to clear 4.5:1 on their
chips). A shown notification can play a sound — None (default), Chime, Pop
or Alert — chosen globally and per notification type in Settings →
Notifications; the WAVs are generated by scripts/gen-notification-sounds.js.
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.
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.
Internal SmartGift build of a Claude Code monitoring dashboard.
Lanes: a durable unit of parallel agent work, one per working directory,
tracked across session restarts. Managed lanes are git worktrees the
dashboard provisions and can reset or remove behind a three-check destroy
guard and a counted preflight; adopted lanes are directories you already
own and are never destroyable.
Pipelines: a lane moves through pipeline stages. A stage the agent declares
with evidence renders green; a stage inferred from the tool-event stream
renders dashed amber and never counts as done. Detection is forward-only
within a 30-minute window, and never writes the declared stage.
Workspace: one page at /run with a lane grid, the selected lane's pipeline,
and a full Claude console behind a disclosure.
nntrivi2001