[johnmatthewtennant]

Summary

Voice calls currently wait for finalized recognition before the agent can begin answering. Berd now prepares a private draft when the user pauses, enabled by default, without adding visual states. Voice settings includes an “Early delivery (experimental)” switch, on by default; explicitly saved off remains off. Disabling it immediately discards private preparation and leaves final-only delivery unchanged. Preparation uses a session hidden at creation, with the same configured provider/model and backend-enforced chat-only mode, so provisional recognition cannot execute tools or steer an unrelated run.

Only the final transcript enters the normal chat with execution authority. Preparation receives only the provisional transcript, not copied conversation history. A completed draft is assistant-only context, never a substitute for answering or carrying out the final request. Final delivery waits at most 750 ms for preparation; stalled or unsupported preparation is discarded without blocking later speech. Late native finals retain their original recognition IDs even after recognition holds expire.

The change preserves the queue ordering and dispatch requirements in LAWS/CHAT.md and configured-agent requirement in LAWS/AGENTS.md.

Related issue

None found.

Testing

Rendered-app scenarios exercise the production controller and ACP paths with native recognition and the model substituted at their boundaries:

• Provisional input prompts only a hidden session after chat-only mode succeeds. Mode rejection fails closed; private text never appears in the chat or plays aloud.
• Unchanged and corrected finals enter the normal chat exactly once, with optional draft context that does not suppress requested actions. The same behavior holds while an unrelated main-session run is active.
• A stalled private prompt cannot hold final delivery or a later utterance. Cancellation targets only the private session, and late private output remains hidden.
• Recognition ending without a final discards preparation and permits the next utterance.
• The Voice settings switch defaults on, persists off across reloads, prevents private preparation while off without blocking final delivery, and permits preparation again after re-enabling.

Mutation checks reject invoking provisional inference without chat-only mode. Native regressions cover hold expiry, a new provisional utterance, then old and new finals retaining their matching IDs.

The isolated-draft implementation has not been exercised with a live macOS microphone.

Screenshots

[morgmart]

🤖 Automated code review

Engineering found three blocking issues in the exact full PR comparison: corrected speech can block indefinitely behind a speculative run, provisional recognition retains ordinary tool authority before final confirmation, and late finals can be assigned to a newer utterance. The supplied GitHub evidence is structurally valid and all ten captured checks completed successfully, but required checks still independently govern merge readiness. Recommendation: REQUEST_CHANGES.

Deterministic publication result: 3 blocking and 0 non-blocking inline finding(s) publishable; 0 duplicate(s) suppressed; 0 blocking screenshot-evidence requirement(s) in this review body.