Follow-up cleanups for recent chat and voice PRs

[threepointone]

Thank you @whoiskatrin for the very good PRs here. I spent some time reviewing the recent fixes in detail and I am adding a handful of cleanups on top because I am, unfortunately, a little anal about tightening the nearby edge cases once I see them. Sorry for the extra follow-up churn, but these are all meant to preserve the direction of the landed PRs while hardening the corners around them.

Summary

• Adds follow-up robustness for the recently landed voice STT, voice streaming, useVoiceAgent, agent-tool recovery, chat resume negotiation, and ai-chat socket disconnect fixes.
• Keeps the fixes scoped to the reviewed behavior, with regression tests for the edge cases we found during review.
• Updates relevant changesets/docs/readmes where the behavior is user-facing.

Commit Details

bd1348a4 - fix(voice): harden Workers AI STT turn handling

Builds on PR #1458.

This expands the Workers AI STT fix beyond the original late/empty final transcript issue:

• Tracks the current Flux transcript across Update, StartOfTurn, and TurnResumed, so an empty EndOfTurn can still emit the best known utterance.
• Uses StartOfTurn as a model-driven speech-start signal for server-side barge-in.
• Adds playback_interrupt handling so the server can tell the client to stop playback when user speech starts during assistant audio.
• Adds a playback generation guard in VoiceClient so a stale decodeAudioData() task cannot start playing after an interrupt has already arrived.
• Adds server/client regression coverage for Flux turn handling, Nova late messages, barge-in, playback interruption, and decode races.
• Updates the voice docs/readmes to describe model-driven barge-in behavior.

f7b1b344 - fix(chat): harden stream resume negotiation close races

Builds on PR #1463.

This extends the sendIfOpen pattern from the original replay fallback fix to the rest of the resume negotiation paths:

• Wraps bare resume protocol sends in @cloudflare/think and @cloudflare/ai-chat so TypeError: WebSocket send() after close does not bubble out of close races.
• Only records pending resume connections when the STREAM_RESUMING notification send actually succeeds.
• Applies the same close-safe behavior to ContinuationState.sendResumeNone() in the shared agents chat continuation path.
• Adds focused tests for closed connection handling and non-close send errors.
• Updates changesets to reflect the broader resume negotiation hardening.

2954bd35 - fix(voice): parse raw NDJSON text streams

Builds on PR #1462.

This fills the NDJSON half of the documented iterateText() byte-stream support:

• Parses raw newline-delimited JSON as well as SSE data: lines.
• Handles OpenAI-style choices[0].delta.content, response, split byte chunks, final unterminated lines, raw [DONE], and malformed raw JSON lines.
• Accepts valid SSE data:{...} lines without requiring a space after data:.
• Preserves the earlier AI SDK textStream custom async iterator preference from PR Fix voice TTS for AI SDK textStream responses #1462.

5aefa984 - fix(agents): defer recovered agent-tool finish hooks (#1476)

Builds on PR #1476.

This addresses recovery lifecycle gaps found during review of agent-tool finalization:

• Defers recovered onAgentToolFinish hooks until after the agent's user onStart has completed, so user startup/mirror initialization runs before recovered finish hooks execute.
• Keeps durable internal state and terminal client events finalized during recovery before user finish hooks run.
• Isolates errors from deferred finish hooks through onError, so one failed hook does not prevent later recovered runs from draining or agent startup from completing.
• Adds coverage for completed, still-running, uninspectable, replay-failure, deferred-hook ordering, and throwing-hook recovery scenarios.

527c5ba2 - test(voice): cover useVoiceAgent enabled lifecycle (#1478)

Builds on PR #1478.

This adds follow-up tests/docs around the new useVoiceAgent({ enabled }) gate:

• Verifies that enabling after a disabled mount connects using the latest query/capability token.
• Verifies onReconnect fires for real connection identity changes while enabled, but not for first enable.
• Documents that disabled action callbacks are safe no-ops and that first enable is treated as an initial connection rather than a reconnect.

b1497715 - fix(ai-chat): close resumed streams on disconnect (#1487)

Builds on PR #1487.

PR #1487 correctly closed the original sendMessages() stream when the socket closes before a terminal frame. This extends the same transport-owned stream cleanup to the other WebSocket-fed paths:

• Closes tool continuation streams if the socket closes before the resume handshake completes.
• Closes and cleans request bookkeeping if the socket closes mid tool-continuation stream.
• Closes resumed replay/live streams and removes their active request ids if the socket closes before done: true.
• Adds focused regression tests for those disconnect paths.

Test Plan

Ran focused verification while developing the commits:

• npm run build -w agents
• npm run build -w @cloudflare/ai-chat
• npm --workspace @cloudflare/voice run build
• npm run test:workers -w @cloudflare/ai-chat -- src/tests/agent-tools.test.ts
• npm run test:workers -w @cloudflare/ai-chat -- src/tests/ws-transport-resume.test.ts
• npm run test:react -w @cloudflare/ai-chat -- use-agent-chat.test.tsx
• npm run test:workers -w @cloudflare/think -- agent-tools.test.ts
• npm --workspace @cloudflare/voice run test:react -- useVoiceAgent.test.tsx
• npx tsc -p packages/ai-chat/src/tests/tsconfig.json --noEmit
• Focused oxfmt --check / lint checks for edited files

Made with Cursor

---

[devin-ai-integration]

Devin Review found 1 potential issue.

View 6 additional findings in Devin Review.

[devin-ai-integration]

🟡 Client-side interrupt detection bypasses #playbackGeneration guard, allowing stale audio to play after interrupt

The PR introduces #playbackGeneration and #stopPlayback() (at packages/voice/src/voice-client.ts:766-779) to prevent a race where a pending decodeAudioData resolves after playback is stopped and starts unwanted audio. #stopPlayback() increments #playbackGeneration, and #playAudio() checks it at lines 726 and 731. However, the client-side interrupt path in #processAudioLevel still uses the old manual pattern — it clears #activeSource, #playbackQueue, and #isPlaying without incrementing #playbackGeneration. If decodeAudioData is pending when the client-side interrupt fires, the decode resolves, the generation check passes (generation hasn't changed), and unwanted audio plays after the user interrupted.

(Refers to lines 863-867)

---

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

[pkg-pr-new]

Open in StackBlitz

agents

npm i https://pkg.pr.new/agents@1497

@cloudflare/ai-chat

npm i https://pkg.pr.new/@cloudflare/ai-chat@1497

@cloudflare/codemode

npm i https://pkg.pr.new/@cloudflare/codemode@1497

hono-agents

npm i https://pkg.pr.new/hono-agents@1497

@cloudflare/shell

npm i https://pkg.pr.new/@cloudflare/shell@1497

@cloudflare/think

npm i https://pkg.pr.new/@cloudflare/think@1497

@cloudflare/voice

npm i https://pkg.pr.new/@cloudflare/voice@1497

@cloudflare/worker-bundler

npm i https://pkg.pr.new/@cloudflare/worker-bundler@1497

commit: 71cd902

[changeset-bot]

⚠️ No Changeset found

Latest commit: 71cd902

Merging this PR will not cause a version bump for any packages. If these changes should not result in a new version, you're good to go. If these changes should result in a version bump, you need to add a changeset.

This PR includes changesets to release 4 packages

Name	Type
agents	Patch
@cloudflare/ai-chat	Patch
@cloudflare/voice	Patch
@cloudflare/think	Patch

Click here to learn what changesets are, and how to add one.

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