Codex Feedback Report: Voice Dictation Can Produce Partial Transcript Without Warning

Status: draft Created: 2026-05-06 Product area: Codex desktop app, in-app voice dictation, audio processing, window/thread navigation Severity: high user-impact partial capture loss Submitted: pending GitHub issue: pending

Subject

Codex voice dictation should not silently truncate the end of a recording, especially if the user switches windows while audio is still processing.

Message

I am reporting a Codex desktop app voice-dictation failure mode that produced a partial transcript without a clear warning.

On May 6, 2026 around 5:50pm PT, I recorded an audio message in one Codex window. The recording was approximately six minutes and change. Codex transcribed the majority of the message correctly, but the final chunk was missing.

I may have navigated away from that Codex window and reopened another Codex window on screen before the audio was fully processed. I do not know whether that caused the truncation, and I am not going to run repeated audio tests to reproduce it because each test risks losing important dictated material.

If switching away from a Codex window while audio is still processing can interrupt, truncate, or corrupt the transcript, Codex should explicitly warn the user before allowing the window switch, or should preserve the source audio and finish processing safely in the background.

Why This Matters

Partial transcription failure is especially dangerous because it can look like success. The user sees a transcript that is mostly correct and may not immediately notice that the ending is missing.

For high-context work, the final part of a voice message often contains the conclusion, decision, correction, action item, or clarification. Silent truncation can therefore cause operational errors, not just inconvenience.

Voice capture is a critical workflow for complex, nonlinear, high-detail instruction. Users need to be able to trust that recorded audio is either fully transcribed, explicitly marked as incomplete, or recoverable from a saved source audio file.

Observed Behavior

1. I recorded a voice message in a Codex window.
2. The recording was approximately six minutes and change.
3. Codex transcribed most of the message correctly.
4. The transcript silently cut off the final chunk.
5. I may have switched to another Codex window while the original recording was still processing, but this is not confirmed as the cause.
6. I did not receive a warning that leaving the window could interrupt audio processing or cause transcript loss.

Expected Behavior

Codex should not silently produce a truncated transcript without warning.

Safer behavior would be one of:

1. Continue processing the full audio safely even if the user switches windows.
2. Prevent navigation away from the processing window until the audio is safely captured.
3. Warn the user: "If you leave this window now, you could lose your audio. Are you sure you want to switch?"
4. Save the source audio as a recoverable draft before allowing navigation.
5. Mark the transcript as incomplete if processing stops early.
6. Provide a retry path from preserved source audio when transcription is incomplete.

Requested Product Changes

Please consider:

• preserving source audio until transcription has completed and been delivered;
• warning users before navigation that could interrupt active audio processing;
• adding incomplete-transcript detection and visible error states;
• allowing users to retry transcription from saved audio when only a partial transcript is produced;
• testing Codex voice dictation across multi-window navigation and thread switching.

Related Issues

This is related to, but distinct from, two earlier Codex audio reports:

1. Long dictation failed transcription and original audio could not be recovered: https://github.com/openai/codex/issues/21408

2. Permission prompt interrupted active voice dictation and discarded audio: https://github.com/openai/codex/issues/21445

This report is about successful-but-partial transcription, possibly associated with switching away from the window during audio processing.

Desired Outcome

Codex voice dictation should be trustworthy for critical work.

At minimum, Codex should never silently truncate a transcript. If processing fails or is interrupted, the user should be warned, the source audio should be recoverable, and the transcript should be clearly marked incomplete.