fix(sdk): preserve stopped chat boundaries across successor sends

[gtremper]

Stop acknowledgments can close a later response, and lost Stop state can replay old tool output after a reload.
The SDK now records Stop before the request, retains its boundary through hydration, and requires transcript reload for accepted responses without sequence correlation.

Checklist

• The PR title follows the contribution convention.
• The changes include regression tests and a patch changeset.
• Maintainer CI and reference-project validation.

Testing

• Published SDK 4.6.3: 12 of the 15 new HTTP regressions fail
• Corrected source: all 15 cases pass with real HTTP, SSE parsing, and the AI SDK reader
• Full SDK suite: 740 tests pass across 81 files
• Core and SDK builds, repository formatting, and lint pass
• Knip: one existing unused export, RedisCacheStoreConfig, in unchanged internal-packages/cache/src/stores/redis.ts:7
• Independent correctness and security reviews: no remaining findings
• Coverage: delayed acknowledgments, failed Stop, reloads, passive abort, repeated Stop, missing boundaries, session recreation, and blocked duplicate retries
• Compatibility: sequence-free responses without a stopped boundary

Changelog

New chat responses remain intact after Stop, including slow acknowledgments and page reloads.

Risk

• Failed Stop: the local reader closes, but the remote run can continue
• Persisted state: new optional stopped-boundary and reload fields
• Reference-project validation: not run locally

[changeset-bot]

🦋 Changeset detected

Latest commit: 4ca683d

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 27 packages

Name	Type
@trigger.dev/sdk	Patch
@trigger.dev/python	Patch
@internal/dashboard-agent	Patch
@trigger.dev/build	Patch
trigger.dev	Patch
@trigger.dev/core	Patch
@trigger.dev/react-hooks	Patch
@trigger.dev/redis-worker	Patch
@trigger.dev/rsc	Patch
@trigger.dev/schema-to-json	Patch
@trigger.dev/database	Patch
@trigger.dev/otlp-importer	Patch
@trigger.dev/rbac	Patch
@trigger.dev/sso	Patch
@internal/clickhouse	Patch
@internal/llm-model-catalog	Patch
@internal/metrics-pipeline	Patch
@internal/redis	Patch
@internal/replication	Patch
@internal/run-engine	Patch
@internal/run-store	Patch
@internal/schedule-engine	Patch
@internal/tracing	Patch
@internal/webhook-engine	Patch
@internal/webhook-sources	Patch
@internal/testcontainers	Patch
@internal/cache	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[github-actions]

Thanks for your contribution! We require all external PRs to be opened in draft status first so you can address CodeRabbit review comments and ensure CI passes before requesting a review. Please re-open this PR as a draft. See CONTRIBUTING.md for details.

[coderabbitai]

Note

Currently processing new changes in this PR. This may take a few minutes, please wait...

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Advanced

Run ID: 7fd3e8c8-dfc4-4f10-8f2e-82bb72abc08e

📥 Commits

Reviewing files that changed from the base of the PR and between 7e92a02 and 4ca683d.

📒 Files selected for processing (5)

• .changeset/chat-stop-successor-boundary.md
• packages/trigger-sdk/src/v3/chat-stop.test.ts
• packages/trigger-sdk/src/v3/chat.test.ts
• packages/trigger-sdk/src/v3/chat.ts
• packages/trigger-sdk/test/chat-transport-events.test.ts

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create a new PR

---

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.}

[devin-ai-integration]

Devin Review found 1 potential issue.

[devin-ai-integration]

🔴 Transcript reload leaves chat blocked

After requiresTranscriptReload is set, seedResumeCursor leaves it and the stopped boundary intact. Reloaded chats still reject every send and reconnect.

Learn more

The transport sets requiresTranscriptReload after the server accepts a successor without returning its input sequence. The public transcript-loading flow calls seedResumeCursor, which neither advances an existing cursor nor clears this flag or the supersede gate. assertTranscriptReady therefore keeps rejecting sends after the requested reload, and reconnectToStream also keeps returning null.

Example: A stopped chat accepts message B without a sequence, so message B throws the reload error. useLoadTranscript then loads a snapshot through B and calls seedResumeCursor(chatId, "20"). The session still has requiresTranscriptReload: true; sending message C throws the same error instead of continuing from event 20.

Recommended fix: Add an explicit transcript-reload operation that force-applies the loaded output cursor and clears requiresTranscriptReload, skipToTurnComplete, supersededInputSeq, and stale active-turn state. Make useLoadTranscript use that operation only after a successful fresh transcript load.

---

Was this helpful? React with 👍 or 👎 to provide feedback.