fix(grok): Harden ACP resume with replay-idle load readiness

[mwolson]

Summary

Closes #3560.

• Race session/prompt against xAI _x.ai/session/prompt_complete so Grok turns settle when the standard ACP RPC remains stranded.
• Make Stop interrupt the wedged prompt, settle the active turn exactly once, and return the session to ready.
• Resume Grok sessions without replay flooding by dropping replay updates and using replay-idle load readiness when session/load never returns.
• Keep one strict root-turn boundary: an AcpSessionRuntime only accepts session/update events for its started root session, and Grok drains those accepted events before emitting turn.completed.
• Keep the actual running turn unfolded in the web timeline when latestTurn temporarily regresses.

Deliberate scope

This is a narrow V1 Grok/ACP bridge. Child-session chunks are ignored instead of being flattened into the parent transcript. After merge, the V2 follow-up can route those chunks through its existing subagent lineage and progress model.

The earlier projection, reducer, shared-message, markdown-throttling, sidebar, and generic orchestration workarounds have been removed. turn.completed remains the canonical terminal barrier, and ProviderRuntimeIngestion retains its strict provider-independent finalization behavior.

Validation

• vp check passes with 7 unrelated existing mobile warnings.
• vp run typecheck passes.
• vp test packages/effect-acp/src/client.test.ts apps/server/src/provider/Layers/CursorAdapter.test.ts apps/server/src/provider/Layers/GrokAdapter.test.ts apps/server/src/provider/acp/AcpJsonRpcConnection.test.ts apps/server/src/provider/acp/AcpRuntimeModel.test.ts apps/web/src/components/ChatView.logic.test.ts apps/web/src/components/chat/MessagesTimeline.logic.test.ts apps/web/src/components/chat/MessagesTimeline.test.tsx: 8 files, 130 tests passed.

Note

Harden Grok ACP turn resumption with replay-idle load readiness and atomic prompt settlement

• Session load on resume now races between the session/load RPC and a replay-idle timer (waitForSessionLoadReplayIdle), completing as soon as replayed notifications go quiet for a configurable gap; replayed and child-session updates are dropped during load.
• Prompt settlement is now atomic and context-validated via settlePromptInFlight: late or cross-session completions are ignored, interrupted turns cannot be revived by late RPC results, and multiple in-flight prompts for the same turn merge with only the last settling the turn.
• interruptTurn now accepts a specific turnId, pre-marks it in interruptedTurnIds, cancels outstanding prompt slots, and moves the session to ready without emitting spurious terminal events.
• xAI prompt completion fallback (makeXAiPromptCompletionRuntime) races the native ACP prompt RPC against _x.ai/session/prompt_complete notifications, tags requests with promptId/requestId, dedupes completed IDs, and aborts pending fallbacks on cancel.
• AcpSessionRuntime serializes concurrent prompts, adds drainEvents/EventStreamBarrier for queue synchronization, and resolves an interrupted prompt to { stopReason: 'cancelled' } instead of hanging.
• MessagesTimeline and ChatView now propagate runningTurnId so fold behavior tracks the active turn rather than a regressed latestTurn, and interrupt actions include the active turn ID.
• Risk: turn.completed events may now carry a null stopReason when the xAI provider omits one from its notification.

^{Macroscope summarized bf4b5b5.}

---

Note

High Risk
Changes core Grok turn lifecycle, cancel/settlement ordering, and ACP event routing—areas where race bugs directly affect user-visible transcripts and session readiness.

Overview
Hardens the Grok ACP bridge so turns can finish when the standard session/prompt RPC hangs, resumes do not flood the transcript, and Stop/cancel races do not double-settle or leak late updates.

The ACP session runtime now ignores replay and foreign-session session/update traffic, can complete session/load when replay goes idle (not only when the RPC returns), serializes prompts, interrupts in-flight prompt fibers on cancel, and adds drainEvents / EventStreamBarrier so consumers flush queued events before terminal settlement. xAI _x.ai/session/prompt_complete is wired via makeXAiPromptCompletionRuntime to race the native prompt, track promptIds, and synthesize cancelled/missing-stop-reason responses.

GrokAdapter gains centralized settlePromptInFlight, interruptedTurnIds, turn-scoped interruptTurn, event draining before turn.completed, and drops streaming for cancelled turns. The mock ACP agent and effect-acp client tests cover hang, stale completion, load replay, and mixed extension notification batches.

Web: interrupt sends the session’s active running turnId; the timeline prefers runningTurnId over a regressed latestTurn so the live turn stays unfolded.

^{Reviewed by Cursor Bugbot for commit bf4b5b5. Bugbot is set up for automated code reviews on this repo. Configure here.}

[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: Repository UI

Review profile: CHILL

Plan: Pro

Run ID: e88888ef-ddd4-4d21-8038-d79d0f20e24e

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

---

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.}

[macroscopeapp]

Approvability

Verdict: Needs human review

1 blocking correctness issue found. Introduces significant runtime behavior changes to Grok session and prompt lifecycle management, including new prompt settlement logic, session resume replay-idle detection, and xAI prompt completion fallback. The complexity and cross-cutting nature of these infrastructure changes warrant human review.

^{You can customize Macroscope's approvability policy. Learn more.}