feat(orchestration): Control Panel — live coordinator DAG in Mission Control (O1 #6 + O2 #7)

[zaridan]

Part of #5. Implements O1 (#6) and O2 (#7): make the coordinator's live task DAG visible in Mission Control. Behind experimentalOrchestrators; additive; no regression to the existing aggregate-count path or the shipped-log section.

O1 — sync the per-run task DAG + dispatch detail to the renderer

Previously only three aggregate ints crossed to the renderer (OrchestrationActivity). Now the main→renderer graph sync also carries, per running coordinator, keyed by coordinator paneKey, run-scoped per F1 (#12):

• Tasks — { id, status, deps, title (display_name ?? task_title ?? first spec line), targetKey } from listTasksWithDispatch({ coordinatorRunId }).
• Dispatch per task — { assigneeHandle, assigneeAgent, status, lastHeartbeatAt, stale }. assigneeAgent is resolved in main from the worker PTY's launch agent; stale is computed in main (it owns the clock + hung threshold) so the renderer stays pure.
• Latest worker signal per dispatch — most recent heartbeat phase or worker_done summary from the messages table (addressed to the run's coordinator handle, so it stays run-scoped without a run column on messages).

listTasksWithDispatch now also surfaces the active dispatch's status + last_heartbeat_at (aliased to avoid colliding with the task's own status); the taskList RPC strips them so its shape is unchanged. A pure assembler (run-dag-snapshot.ts) maps DB rows → OrchestrationRunDag, caps the payload (200 tasks, reports the truncated count, console.warn on truncation), and indexes the latest signal per dispatch. A new orchestration-run-dag store slice reuses the equality-check pattern from orchestration-activity to avoid re-render churn on high-frequency title ticks.

O2 — rewire Mission Control to DAG-backed + render the live Control Panel

OrchestratorMissionControl now renders the coordinator's DAG using Orca's existing visual vocabulary — AgentStateDot for status and the shared agent glyphs (AgentIcon) for who's working. No new status visuals invented.

Field mapping:

UI	Source
Coordinator dot	deriveOrcastratorDotState (supervising/stalled)
Counts line	OrchestrationActivity (N tasks · M workers · K stalled)
recipe line	OrchestrationRunDag.recipe (absent for LLM directors)
Task row dot	task status → idle (pending/ready) · working (dispatched) · stalled (stale heartbeat) · blocked · done (completed) · interrupted (failed)
Task row icon	dispatch assigneeAgent → agent glyph
Task row message	latest heartbeat phase / worker_done summary / "waiting on <dep>" for unmet deps
Shipped section	existing .orcastrate/log.jsonl parse (unchanged)

Mode-agnostic + fallback

The same panel renders the current LLM director and a future recipe director — both emit the same CoordinatorRun + tasks + dispatches. When a director has no coordinator run yet (e.g. an LLM director that hasn't called orchestration.run), the existing lineage "Spawned work" + Shipped view renders unchanged, so nothing regresses.

Constraints

STYLEGUIDE tokens + shadcn + AgentStateDot; cross-platform; SSH/remote (sync already crosses that boundary). AGENTS.md: no max-lines disable (row/section/summary split into focused components), .ts over .d.ts, comments document the why, concrete file names. New UI strings synced across all locale catalogs.

Test evidence

• Sync shape + per-run scoping: run-dag-snapshot.test.ts (stale derivation, signal latest-wins for heartbeat phase / worker_done summary, deps parsing, task cap + truncation, recipe passthrough); db.test.ts (new dispatch columns surfaced; listTasksWithDispatch({ coordinatorRunId }) scoping — two runs don't bleed); orchestration-run-dag.test.ts (store equality avoids churn, replaces on change); orchestrator-mission-control-data.test.ts (selector picks the director's run by paneKey tab id; concurrent runs don't bleed).
• Render: orchestrator-task-row.test.ts (status→AgentStateDot mapping for each status incl. working vs stalled; message derivation); MissionControlTaskRow.test.tsx (correct dot/glyph/message per status); MissionControlTasksSection.test.tsx (row per task, empty state, truncation note); OrchestratorMissionControl.dag.test.tsx (renders DAG + counts when a run exists; falls back to lineage view when none; Shipped still renders).
• Full vitest green (the only failures seen were pre-existing flaky native-PTY/git-relay suites that pass in isolation, plus pre-existing TuiAgent typecheck errors in orca-runtime.test.ts — both unrelated to this change). typecheck (node/web/cli), electron-vite build, and oxlint all green; localization catalog parity verified.

🤖 Generated with Claude Code

[zaridan]

Review panel — not merge-ready (top findings verified against source)

3-lens panel; the load-bearing items re-verified against source by the coordinator. No blockers, but a real perf regression + a feature gap.

🟠 Must-fix

1. Hot-path perf regression. buildOrchestrationRunDagByPaneKey() runs at orca-runtime.ts:2746 inside syncWindowGraph — the 16ms-coalesced tick that fires on ANY terminal title/status/layout churn. Per running run it does listTasksWithDispatch (multi-table LEFT JOIN) + getAllMessagesForHandle(…,500,…), replacing 3 cheap COUNT(*)s. Fix: gate recompute on a change signal (max task/dispatch rowid + message-sequence high-water mark), cache otherwise — don't rebuild every coalesced tick.
2. Builder ignores experimentalOrchestrators. The flag gates only the renderer panel; the main-side builder runs whenever a run is active, so users with a running coordinator pay feat(orchestrators): Orcastrator — director sidebar, Send to Orcastrator, and Mission Control #1 un-opted-in and can't see the panel. Fix: gate the builder on the flag / skip when no Mission Control window is mounted.
3. worker_done summary unreachable on completed rows. listTasksWithDispatch's join is WHERE status IN ('pending','dispatched') (db.ts:691), so a completed task has dispatch_id = null; the signal is keyed on dispatch_id → deriveTaskMessage returns ''. The advertised summary never lands. Tests pass only by injecting signalsByDispatchId against a hand-built dispatch_id, bypassing the real join. Fix: source the worker signal by task_id (or a 2nd LEFT JOIN to the latest dispatch regardless of status), and fix the test to drive the real join.

🟠 Should-fix

4. Per-task stale omits the dispatched-at grace (run-dag-snapshot.ts:502) → freshly-dispatched workers flash amber stalled while the aggregate count shows 0. Add the same dispatched_at < threshold grace as getStaleDispatches.
5. Two different numbers both labeled "tasks" — summary shows pendingTasks (outstanding), section shows dag.tasks.length (all). Relabel summary to "active"/"outstanding".
6. Main buildOrchestrationRunDagByPaneKey untested — the "two runs don't bleed across panes" guarantee (pane resolution + signal scoping) has no test. Add one.
7. Hardcoded English row labels (status words + "waiting on") bypass i18n. Route through t().

Nits

Dedup listCoordinatorRuns/staleThresholdIso across the two builders; wrap the per-run loop so a throw doesn't forfeit the whole sync tick; satisfies never exhaustiveness on the dot/label mappers; empty-title → fall back to task id; omit waiting on <opaque-uuid> past the 200-task cap.

No-merge until #1–#3 land; #4–#7 same pass.

[zaridan]

Round 2 — must-fix #1–#3 + should-fix #4–#7 + nits landed

Pushed two commits (6bcdce6 main, 57af8a4 renderer).

Must-fix

1. Hot-path perf — buildOrchestrationRunDagByPaneKey no longer rebuilds every coalesced tick. A cheap per-run change token (OrchestrationDb.getRunDagChangeToken: MAX(rowid) of the run's tasks + dispatch_contexts, MAX(sequence) of the coordinator handle's messages) is computed each tick; the heavy listTasksWithDispatch + getAllMessagesForHandle(…,500,…) run only when it moves, otherwise the cached snapshot is served. Cache pruned to the running set. Test: heavy queries called once across two unchanged ticks, again when the token moves.
2. Flag gate — the builder is gated on experimentalOrchestrators (read from the runtime store); un-opted-in users with a running coordinator pay nothing. Test: flag-off → no DAG, listTasksWithDispatch never called.
3. worker_done on completed rows — signals are now keyed by task_id (on both heartbeat and worker_done payloads), not dispatch_id, so the summary reaches completed rows (whose join dispatch_id is null). New real-join test seeds an in-memory DB, completes a task with a worker_done, and asserts the summary lands on the completed node — it asserts dispatch_id === null as the precondition, so it fails against the dispatch-keyed code.

Should-fix

4. Stale grace — per-task stale now mirrors getStaleDispatches exactly (requires dispatched_at past the grace AND stale/null heartbeat); listTasksWithDispatch surfaces dispatched_at. Test: freshly-dispatched + null heartbeat within grace → stale === false.
5. Counts label — supervision summary relabeled N tasks → N outstanding (it's pendingTasks, distinct from the section header's total).
6. Main builder tested — added cross-pane no-bleed (two runs → distinct panes; task + per-handle signal scoping stay run-local) and the change-token cache test.
7. i18n — row status words + waiting on <dep> route through translate() with new keys synced across all locales; worker-authored content (phase/summary) still renders verbatim.

Nits

Deduped listCoordinatorRuns/threshold across the two builders (collectRunningOrchestration); per-run loop wrapped in try/catch so a malformed run can't forfeit the sync tick; blank title → task id; waiting on <dep> omitted (no raw uuid) when the dep is truncated past the 200-task cap. Skipped the satisfies never default — the mappers already get compile-time exhaustiveness from their return type, and an explicit default trips oxlint's switch-exhaustiveness-check.

Verification

typecheck (node/web/cli), electron-vite build, oxlint, and localization parity all green. Full vitest green for everything touched (orchestration DB/builder/runtime + renderer). The only full-suite failures were pre-existing flaky native-PTY / git-relay / subprocess / dev-startup suites that pass in isolation — unrelated to this change. Also fixed a pre-existing missing TuiAgent import in orca-runtime.test.ts so typecheck:node is now clean.

Not merging.