[romil-bijarnia]

This is the review copy of upstream #36195, which is currently closed by the open-PR limit.

The tested base has dictation, but no two-way voice conversation or Aube picker. This adds experimental live voice to saved Codex conversations on web and desktop, using the ChatGPT account already connected to the selected provider. It does not require an API key or a separate voice login.

The server discovers the backend's V3 voice catalogue, owns call authorization and lifecycle, and sends spoken coding requests through normal T3 orchestration. Clients handle microphone selection, playback, mute, stop and captions. Startup reserves session IDs across threads before validation can suspend. Managed connections bind runtime resolution and credentials under one auth admission, reject profile-revision mismatches, and retain that admission for the socket lifetime. Client mute requests persist through microphone selection and permission and are applied before audio tracks are attached. Ending a call does not cancel coding tasks already dispatched. Native mobile dictation is unchanged.

This is an alternative implementation to upstream #29394, not a claim that no related work exists. Unlike that implementation's native-turn interruption/resubmission path, T3 owns the voice delegation before a coding turn starts. A separate repair to the existing implementation is available in jakeleventhal/t3code#7.

Verified on Linux with unmodified Codex 0.167.2 and an existing ChatGPT-backed T3 provider:

• The real client opened the live-voice dialog, loaded 25 discovered profiles and selected Aube. No physical microphone was opened for the screenshot check.
• An actual HTTP/WebRTC call sent synthetic speech to Aube, dispatched exactly one real gpt-6-luna task, saved its request and VOICE_HANDOFF_CONFIRMED reply in the same thread, and spoke the result back. No registry, projection or orchestrator mocks were used. Decoded audio was nonzero, no realtime errors occurred, and stop completed cleanly.
• vp run check:changed 2d8a843e6be9ca170f3bce98cc92ad76617b00c4 --concurrency 2 passed formatting, lint, both Knip passes, GraphQL, 38 affected-package typechecks, 36 tests in seven selected files and three React Compiler checks. This is not a repo-wide suite.
• Cancellation before microphone selection and stopping during thread validation were reproduced and fixed; the latter regression previously dispatched a task after the call ended.
• Six additional race regressions failed before the review fixes and now pass: three managed-account binding/lifetime cases, both pending-microphone mute phases, and a duplicate session ID across different threads. Account races use dummy profiles and the real auth admission; no actual user accounts were switched.

Limits: only Sol and Aube received real audio tests; catalogue discovery is not an entitlement guarantee for every profile. Managed SIWC token-sharing voice, native mobile calls, remote/relay/tunnel operation and physical microphone/speaker quality remain unverified. Rejected account routes report errors rather than switching billing, accounts or auth methods.

The before screenshot is current main rendering an isolated copy of the same harmless saved test conversation. All state and account fixtures were separate from the installed nightly and live chats. Screenshots and the credential-free acceptance result are linked below, not committed to source.

Updated credential-free end-to-end result, rerun after the review fixes: real Aube audio, exactly one new same-thread Codex run, saved reply and spoken result; no realtime errors.

Current main, same saved test conversation:

Alternative with Aube selected:

Discovered voice profiles: