fix(orchestrator): don't synthesize the teardown stopped that follows a router-ceded terminal (#11689 residual)

[NubsCarson]

Problem

Since fd2e343 (#11689), every routed one-shot completion posts a spurious second message to the origin channel: the router posts the real completion, then seconds later swarm synthesis posts <label> stopped. (or a sanitized chunk of the session's raw lastOutput) to the same room, threaded to the user's message.

Mechanism, on current develop:

1. Every planner-spawned one-shot (TASKS op=create → runPromptAndClose / runPromptViaSmithers, actions/tasks.ts) emits task_complete on success, then in its finally always calls service.stopSession → AcpService.closeSession emits a stopped carrying response = lastOutput (raw agent output), plus a second explicit stopped. This sequence is deterministic — it happens on every routed success, not as a rare race.
2. fix(orchestrator): router owns origin-routed session completions — stop swarm-synthesis double-posts and suppressed-error leaks (#11634) #11689's router-owned task_complete skip in SwarmCoordinatorService.runSwarmComplete deliberately returns without claiming the synthesizedCompletionSessions dedupe slot (so a later genuine stop on a reused session isn't swallowed), and stopped is deliberately not router-owned (shouldInject in sub-agent-router.ts excludes it).
3. So the teardown stopped falls through every guard and synthesizes. server-helpers-swarm.ts posts it unconditionally (swarm_synthesis is EPHEMERAL, so the roomId+text delivery dedupe is bypassed) to originRoomId with replyToExternalMessageId — i.e. the user's Discord channel.

The same mechanism leaks a terminal stopped mid-respawn on error paths the router deliberately suppresses (account failover / state-lost): tasks.ts stops the session before the router stamps handedOffToSuccessorSessionId, so the #11720/#11721 handoff skips can't catch it (they help cache staleness, not a not-yet-stamped store).

Pre-#11689 the top-of-function dedupe claim made these teardown stops invisible; the commit converted the #11634 duplicate-completion into a spurious stop-post instead of eliminating the double post. Tests at swarm-coordinator-service.test.ts pinned the wrong behavior as expected (modeled as a rare user-stop race, but the one-shot runners produce the identical sequence deterministically on every routed success).

Fix

Track the cession in memory: when the router-owned skip cedes a task_complete/error, record the session in a new routerCededTerminalSessions set. A subsequent stopped on a ceded session is recognized as teardown plumbing of that same turn and skipped — without claiming the synthesis dedupe slot.

The marker is strictly turn-scoped, so nothing goes silent:

• Cleared when the session resumes (any non-terminal event): a genuine user stop on a later turn of a reused session still synthesizes.
• Cleared on legacy-task eviction and stop(): bounded memory; an idle-session stop after the grace window behaves exactly as before.
• A genuine user stop with no preceding ceded terminal (mid-turn cancel / no-output stop) hits no marker and still synthesizes — the fix(orchestrator): router owns origin-routed session completions — stop swarm-synthesis double-posts and suppressed-error leaks (#11634) #11689 invariant is preserved.
• Custom-validator (app-verification) completions are unaffected: they synthesize and claim the dedupe slot as before.
• The check runs before the swarm-synthesis posts one message per verify-retry lineage session teardown (stopped) — user sees 3 completions for 1 task #11711 store re-read, so it also covers the suppressed-failover leak where teardown lands before the router stamps the handoff marker. The store-backed handoff skips are kept — they cover a restarted coordinator / resumed-then-handed-off session where the in-memory marker is gone.

Per-session terminal serialization (maybeFireSwarmComplete) guarantees a same-tick teardown stopped observes the cession recorded by the terminal that preceded it.

Tests

Reproduced first (all three fail on unfixed develop, with the exact spurious payload completionSummary: "build-site stopped." targeted at the origin room):

• does NOT synthesize the teardown stopped that follows a router-owned task_complete (#11689 residual) — the deterministic one-shot success sequence (task_complete → closeSession stopped w/ raw lastOutput → explicit stopped) produces zero synthesis.
• does NOT leak a teardown stopped after a router-owned error the router suppresses (failover/state-lost) — no terminal <label> stopped. mid-respawn.
• suppresses the stopped racing a router-owned terminal on the same session (same-tick teardown) — rewrites the old race test that pinned the wrong behavior.
• a stopped AFTER the session resumes still synthesizes (cession is per-turn; sessions are reused) — preserves the reuse guarantee the old test at :842 guarded.
• The two swarm-synthesis posts one message per verify-retry lineage session teardown (stopped) — user sees 3 completions for 1 task #11711 tests now warm the enrichment cache via a non-terminal event so they keep exercising the store-backed handoff re-read layer (previously warmed via a router-owned task_complete, which now also sets the in-memory marker).

Unchanged and still green: genuine user stopped with no preceding terminal posts; router-inactive/no-origin sessions still synthesize; custom-validator completions still post.

Verification

plugins/plugin-agent-orchestrator: bun run test:unit  → 118 files, 1273 tests passed
plugins/plugin-agent-orchestrator: bun run typecheck  → clean
packages/agent server-helpers-swarm.test.ts           → 13 passed

Refs #11689 (fd2e343), #11634, #11711, #11720, #11721.

[greptile-apps]

Your trial has ended. Reactivate Greptile to resume code reviews.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: 4e85ca4a-24f6-4867-972f-2eed8276befb

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/rh-0

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[lalalune]

Validation pass after rebasing onto current origin/develop (includes #11811):

• Generated standard keyword artifacts with node packages/shared/scripts/generate-keywords.mjs --target ts for test imports.
• bunx vitest run plugins/plugin-agent-orchestrator/__tests__/unit/swarm-coordinator-service.test.ts --reporter=dot -> 49 tests passed.
• bun run --cwd plugins/plugin-agent-orchestrator typecheck -> passed.
• bunx biome check plugins/plugin-agent-orchestrator/src/services/swarm-coordinator-service.ts plugins/plugin-agent-orchestrator/__tests__/unit/swarm-coordinator-service.test.ts -> passed.
• bunx vitest run packages/agent/src/api/server-helpers-swarm.test.ts --reporter=dot -> 13 tests passed.
• bun run --cwd plugins/plugin-agent-orchestrator test:unit -> 118 files / 1281 tests passed.
• git diff --check origin/develop..HEAD -> passed.

Reviewed the session marker lifecycle: router-ceded task/error terminals mark the current turn, teardown stopped events are suppressed, non-terminal resume clears the marker, and genuine stops with no preceding ceded terminal still synthesize.

[greptile-apps]

Your trial has ended. Reactivate Greptile to resume code reviews.

[claude]

Claude encountered an error —— View job

---

I'll analyze this and get back to you.