feat(voice): speak assistant text while it streams

[johnmatthewtennant]

Summary

Berd feeds incremental ACP assistant text into the active Pocket utterance instead of waiting for the response to finish. Speech begins once a complete sentence is available, and later text continues in the same utterance as it arrives.

Assistant text around a tool call remains one voice turn. Starting the tool flushes buffered speakable text, and each visible text block receives its own speaking and terminal delivery status. Internal synthesis chunks are not exposed as transcript blocks.

Reviewer-reproducible examples

• With a model that emits incremental text, request a long response. Confirm speech begins before the full response has arrived and the transcript continues to grow.
• Request a response containing assistant text before and after a tool call. Confirm the pre-tool text is delivered before the tool completes, post-tool text continues in the same voice turn, and both visible text blocks receive the appropriate delivery status.

[loganj]

🤖 Carl follow-up review at current head 35c78af9c21890f0c624d25971a55d75d722ae6e: the blocking status-owner defect from 62ab90b is addressed. This head clears my blocker.

The fix changes ActiveUtterance from one target to a deduplicated targets collection and retains the current utterance status (nativeAssistantSpeech.ts:15-24,99-104,259-291). When post-tool text arrives, its target is added; if playback already started, the new block immediately inherits speaking (:260-269). Every terminal path now projects status to all participating targets: failed, completed, interrupted event, and local barge-in/stop (:106-123,141-190). This resolves both failure modes from my original review: normal completion no longer marks only the pre-tool block, and interruption no longer attributes post-tool interruption solely to that first block.

The new regression (nativeAssistantSpeech.test.ts:233-291) exercises text → tool → text, verifies the post-tool block inherits speaking, then verifies both blocks become spoken on completion. The existing interruption test remains single-block; adding a cross-tool interruption assertion would make the regression proof symmetric, but the shared setUtteranceStatus() path is the same for completion, stream interruption, failure, and local interruption, so I do not consider that omission blocking.

Exact-head CI is green: frontend checks/unit tests, clean-room frozen install, Rust checks on macOS/Linux/Windows, transcript virtualization, frontend build smoke, and DCO (workflow run 32393632587; install run 32393632586). I inspected the exact head on fresh read-only Blox workstation carl-berd135-rereview2 (ID 1908348) and did not rerun broad suites.

Updated verdict: approve / ship 35c78af9c21890f0c624d25971a55d75d722ae6e.

[morgmart]

🤖 Automated code review

Reviewed the complete exact three-dot comparison 6405382...92fc2b2 using static inspection only, after running the project code-review and user wes-review workflows. Final self-check covered every changed voice-streaming flow, accessibility, i18n, navigation/consent, async failure and never-completes behavior, lifecycle/races, test honesty, project/design-system rules, duplicate overlap, and blocking evidence/user effect. Found one blocking response-loss race. Supplied GitHub evidence was inspected: Frontend checks and unit tests is failing; macOS/Linux/Windows Rust checks, clean-room install, and DCO pass, while two downstream frontend jobs are skipped. Recommendation: REQUEST_CHANGES; required checks separately govern merge readiness.

Deterministic publication result: 1 blocking and 0 non-blocking finding(s) publishable; 1 duplicate(s) suppressed.

[morgmart]

🤖 Automated code review

Reviewed all nine changed files in the complete exact three-dot comparison 6405382...04f4874 using static inspection only, after explicitly running the project code-review and user wes-review workflows. The final self-check covered every changed voice-streaming flow; accessibility; i18n/localization; navigation and consent guards; async errors, never-completes behavior, lifecycle and races; test honesty; design-system and project laws; duplicate overlap; and concrete evidence plus user effect for blocking severity. No publishable findings remain. A same-inspection variant of reply loss is suppressed because it is the same underlying finishing-playback issue as a resolved automation thread with a substantive human-account reply. Supplied GitHub evidence is structurally valid and was inspected: Frontend checks and unit tests is failing, Windows Rust checks is in progress, macOS/Linux Rust checks, clean-room install, and DCO pass, and two downstream frontend jobs are skipped. Recommendation: APPROVE under the publication policy; required checks separately govern merge readiness.

Deterministic publication result: 0 blocking and 0 non-blocking finding(s) publishable; 1 duplicate(s) suppressed.

Completed checks currently reporting failure:

• Frontend checks and unit tests: failure

Pending checks: 1 check(s) are not complete.

This approval reflects the completed code review only; merge readiness remains governed by the repository's required checks.

[morgmart]

🤖 Automated code review

Reviewed all nine changed files in the complete exact three-dot comparison 6405382...168968a using static inspection only, after explicitly running the project code-review and user wes-review workflows. The final self-check covered every changed voice-streaming flow; accessibility; i18n/localization; navigation and consent guards; async errors, never-completes behavior, lifecycle and races; test honesty; design-system and project laws; duplicate overlap; and concrete evidence plus user effect for blocking severity. No publishable findings remain. A same-inspection variant of reply loss is suppressed because it is the same underlying finishing-playback issue as a resolved automation thread with a substantive human-account reply. Supplied GitHub evidence is structurally valid and was inspected: DCO, transcript virtualization, frontend build smoke, frontend checks and unit tests, macOS/Linux/Windows Rust checks, and clean-room install all completed successfully. Recommendation: APPROVE under the publication policy; required checks separately govern merge readiness.

Deterministic publication result: 0 blocking and 0 non-blocking finding(s) publishable; 1 duplicate(s) suppressed.

Pending checks: 1 check(s) are not complete.

This approval reflects the completed code review only; merge readiness remains governed by the repository's required checks.