macOS realtime voice: response-start delays become much higher late in a long conversation

What version of the Codex App are you using?

Current installed desktop app: 26.903.71938 (build 8576).

What subscription do you have?

Not disclosed in this public report.

What platform is your computer?

macOS, ChatGPT/Codex desktop app, realtime voice conversation.

What issue are you seeing?

During a realtime voice conversation lasting approximately 90 minutes, the delay between finishing a short spoken sentence and receiving a meaningful assistant response became noticeably and substantially longer toward the end of the conversation.

The delay was not measured precisely. Estimates such as one or two minutes should not be treated as timing measurements or a fixed threshold; the reliable observation is a large, obvious increase in perceived response-start latency late in the session.

What steps can reproduce the bug?

Observed sequence, not yet an independently reproduced test:

1. Start a realtime voice conversation in the desktop app.
2. Continue using the same conversation for an extended period (about 90 minutes in this observation).
3. Toward the end, speak a short sentence and observe the much longer wait before the assistant begins a meaningful reply.

What is the expected behavior?

Ordinary conversational replies should remain responsive throughout a long voice conversation. If work requires additional time, the conversation should not appear stalled without a timely, useful response.

Additional information

The user did not observe network instability, unusually high computer load, or a buildup of background work at the time. These are user observations, not instrumented exclusions of possible causes.

The underlying delayed stage has not been established: transcription, voice turn-taking, backend reasoning, or delivery could require separate timing checks. This report does not claim a fixed session-duration limit or a confirmed root cause.

Related but not necessarily the same issue: #35508 reports progressive realtime transcription slowdown. This report concerns the user's experienced wait for the assistant to respond; transcription timing was not measured.

Submitted with the user's explicit permission. No conversation transcript, audio, private task details, local paths, identifiers, credentials, or raw logs are attached.

No file, test, or entry point is named. Start by reproducing a long realtime voice conversation in the macOS desktop app and measure transcription, turn-taking, backend reasoning, and delivery separately; compare the observation with related issue #35508. Done means identifying the delayed stage, reproducing the late-session latency increase, and adding coverage or diagnostics for the confirmed cause.