Fix thread detail subscription race

[Quicksaver]

Summary

Fixes a race in the thread-detail WebSocket subscription path where live events could be emitted while the server was still loading the initial snapshot and snapshot sequence. subscribeThread now attaches its live domain-event subscription before snapshot loading, still emits the loaded snapshot first, then emits only matching thread-detail events newer than the loaded snapshot sequence.

What Changed

• Exposed OrchestrationEngine.subscribeDomainEvents so callers that need gap-free live observation can acquire a scoped PubSub subscription before starting dependent work, while preserving streamDomainEvents as the existing live stream surface.
• Added a WebSocket helper that builds a live domain-event stream from the scoped subscription when available, with a fallback to the existing streamDomainEvents implementation for older or test engines.
• Updated subscribeThread to attach the live stream before loading the thread snapshot and snapshot sequence.
• Preserved the snapshot-first stream contract by emitting each loaded snapshot first, then draining queued/live events whose sequence is greater than the loaded snapshot sequence.
• Added a server regression test that delays the initial thread snapshot, emits a live user-message event during that delay, and verifies subscribers receive the snapshot followed by the live event.

Why

This addresses an intermittent first-prompt rendering bug: when sending the first prompt, the first user message could briefly appear in the conversation and then disappear, leaving only the rest of the conversation. In that failure mode, the first visible conversation item could become the Working for... label instead of the user's submitted message.

Validation

• pnpm exec vp test run apps/server/src/server.test.ts -t "buffers thread detail events emitted while loading the initial thread snapshot" --reporter=dot passed on head 8568992f.
• pnpm exec vp check passed on head 8568992f, with existing lint warnings outside the PR diff.
• pnpm exec vp run typecheck passed on head 8568992f, with existing desktop Effect suggestions outside the PR diff.
• After pushing head 8568992f, GitHub reports the PR is open, mergeable, and approved; CodeRabbit, Vercel, PR Vouch, PR Size, Cursor Bugbot, Macroscope Approvability, Macroscope Correctness, and Macroscope Effect Service Conventions are successful or skipped where appropriate.

Proof

• No proof files are attached.

---

Note

Medium Risk
Changes real-time delivery semantics on a core conversation WebSocket path; mistakes could duplicate or omit messages, though new tests target the race.

Overview
Fixes a race where subscribeThread could drop live thread-detail events that arrived while the server was still loading the initial snapshot or finishing persisted catch-up replay.

OrchestrationEngine now exposes subscribeDomainEvents, a scoped PubSub subscription so callers can attach a dedicated live listener before slower work runs. streamDomainEvents is unchanged for existing consumers.

The WebSocket path starts that subscription first, filters to thread-detail events, and drains them into a scoped unbounded queue. After the snapshot (or replay through afterSequence), it emits only events with sequence greater than the loaded/replayed cursor so overlaps are not duplicated.

Regression tests cover delayed snapshot load and replay overlapping with live publishes.

^{Reviewed by Cursor Bugbot for commit b516f9c. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Fix race condition in subscribeThread WebSocket handler by buffering events during snapshot load

• The subscribeThread handler in ws.ts previously used a direct stream from orchestrationEngine.streamDomainEvents, which could drop events that arrived while the initial snapshot or catch-up replay was loading.
• A new subscribeDomainEvents method is added to OrchestrationEngineShape that returns a scoped, independent PubSub subscription per caller.
• Live thread events are now buffered into an unbounded queue before snapshot/replay begins; the tail stream deduplicates by filtering out events at or below the last snapshot or catch-up sequence number.
• Two new tests in server.test.ts validate ordered delivery and deduplication for both the initial snapshot and afterSequence replay flows.
• Behavioral Change: events arriving during snapshot load or catch-up replay are now buffered and replayed rather than dropped, changing delivery semantics for the subscribeThread stream.

^{Macroscope summarized b516f9c.}

[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: 2fcf3957-67c4-48b5-befa-97119c882a2e

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: Approved

This is a targeted bug fix for a race condition in thread detail subscriptions. The fix acquires the live event subscription and buffers events before loading snapshots, then deduplicates by sequence number. The change is well-tested with two scenarios covering the race conditions, and follows existing patterns in the codebase.

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

[Quicksaver]

Comment audit complete: I refreshed the tracked PR comments and found no actionable review feedback. Current comments are informational only: CodeRabbit skipped automated review because reviews are disabled, and Macroscope marked the PR approved. There are no inline review threads to resolve.

[Quicksaver]

Comment audit refreshed: I rechecked the PR issue comments, review comments, and review threads. There is still no actionable review feedback to address. Current comments remain informational only: CodeRabbit skipped automated review because repository reviews are disabled, Macroscope marked the PR approved, and there are no inline review threads to resolve.

Dismissing prior approval to re-evaluate 7a1c76d

[macroscopeapp]

Effect service review: one convention violation found in the new desktop backend advertisement error. See inline comment.

^{Posted via Macroscope — Effect Service Conventions}

[macroscopeapp]

Effect service convention review found 3 issues. See inline comments.

^{Posted via Macroscope — Effect Service Conventions}

Dismissing prior approval to re-evaluate 2f3b463

Dismissing prior approval to re-evaluate ecffa75

Dismissing prior approval to re-evaluate 8568992

Dismissing prior approval to re-evaluate 414eafb