mirror of
https://github.com/openai/codex.git
synced 2026-11-02 19:43:24 +00:00
## Why Starting a voice turn while the speaker is idle suppresses response audio until assistant captions arrive, preventing audio that arrives first from being accepted. ## What changed Make speaker suppression conditional when user transcripts start a new turn. Keep idle output ready for audio while preserving interruption of active assistant output and older, uncaptioned speech that is queued or accepted. ## Testing Add regression coverage for quiet turns with incremental or final-only user transcripts and for interruption of older pending speech across queue and caption states. GitOrigin-RevId: 288d2071766a1d0c1b8c7a10c0ab434dc4d7cee1