[Codex Desktop][Realtime Voice] Stale assistant response replays after “Stop” and interleaves with new turns

Summary

During a Codex Desktop realtime voice session, an assistant sentence from an earlier topic unexpectedly replayed/interleaved after the conversation had moved on. Saying “stop” did not reliably clear the stale output, and fragments of the live transcript also appeared duplicated across realtime updates.

The exact stale sentence was:

“Grassroots collective intelligence is the quintessential human technology: our shared ability to work together, combine knowledge, and solve problems.”

The sentence had appeared earlier in the same session during a separate discussion. Its later reappearance was unexpected and out of context; this report is about the replay behavior, not the content of the sentence.

Observed sequence

1. A realtime voice conversation was active in Codex Desktop.
2. The assistant produced the sentence quoted above during one topic.
3. The conversation moved to unrelated work.
4. The earlier sentence later reappeared/interleaved with new speech.
5. The user said “stop,” but stale content continued or returned.
6. Parts of user and assistant speech also appeared duplicated in successive transcript updates.

Expected behavior

• Each assistant response should be delivered once.
• Moving to a new turn should not replay content from an earlier turn.
• “Stop” should immediately cancel and clear pending spoken output.
• Each user utterance should be reconciled once without duplicated transcript fragments.

Actual behavior

• An earlier assistant sentence unexpectedly replayed in a later context.
• New and stale content appeared interleaved.
• “Stop” did not reliably clear the stale output.
• Transcript fragments appeared more than once across realtime updates.

Frequency and impact

This occurred on 2026-08-30. The user reports approximately 6 months of Codex use and 18 months of ChatGPT use and has not previously encountered this behavior. It created the appearance that unrelated content had been inserted into the conversation and undermined confidence in realtime session integrity.

Notes

There is no evidence here of account compromise, hacking, or human interference, and this report makes no such claim. Possible areas to inspect include stale TTS/output queues, duplicated realtime or handoff events, turn reconciliation, and cancellation handling after “stop.”

Related but not identical:

• #35517 — late same-turn realtime transcript events can cause duplicated transcript content
• #41254 — delivered voice turns can remain actionable and later be retried/re-injected

No files, tests, or entry points are named. Start by reproducing the realtime voice sequence, then inspect the stale TTS/output queues, turn reconciliation, duplicated realtime or handoff events, and cancellation handling identified in the report; compare with #35517 and #41254. Done means each response and transcript fragment is delivered once, and “stop” clears pending output without later replay.