Completed Windows Codex App dictation saves transcript locally but loses it from composer

What version of the Codex App are you using (From “About Codex” dialog)?

Codex Desktop 26.803.5235.0

What subscription do you have?

Not included in the local handoff for this incident.

What platform is your computer?

Microsoft Windows 11 Pro 10.0.26200, build 26200, x64

What issue are you seeing?

A completed Codex Desktop voice dictation was not left in the composer and was not sent to the task.

The dictation lasted about 9 minutes 55 seconds. After it was stopped, the completed transcript disappeared from the composer. The next submitted message contained only a complaint that the recording/transcript had vanished.

Local evidence shows that the dictation session itself completed successfully and that the transcript was saved locally:

• surface: composer
• status: completed
• MIME type: audio/webm;codecs=opus
• duration: 595033.9 ms
• audio size: 9581976 bytes
• audio chunks: 119
• saved transcript length: 7081 characters
• transcription-history.jsonl exists but is 0 bytes

The transcript text and the audio are private and are not included in this issue. The local metadata confirms that the text exists, but the text field is intentionally redacted here.

This seems different from an ASR failure: the transcript was produced and persisted locally, but it did not remain visible in the composer for review/editing/submission.

What steps can reproduce the bug?

One confirmed occurrence:

1. Open an existing text task in Codex Desktop on Windows
2. Start voice dictation from the composer
3. Dictate for about 10 minutes
4. Stop dictation
5. Observe whether the completed transcript remains in the composer

Observed result: the transcript was not present in the composer and was not sent to the task, even though the local dictation session status is completed.

Session ID: d8e23082-d0e1-4e56-abeb-538d83998719

The exact UI action between stopping dictation and the text disappearing is not yet proven. I am not including it as a reproduction step until it is reproduced or confirmed.

What is the expected behavior?

After voice dictation completes, the transcript should remain in the composer so the user can review, edit, and send it.

If composer insertion fails, or if the composer draft is reset after transcription completes, Codex should keep the completed dictation visible and recoverable instead of making it look lost.

Additional information

Sanitized local metadata, with the transcript text omitted:

{
  "id": "d8e23082-d0e1-4e56-abeb-538d83998719",
  "mimeType": "audio/webm;codecs=opus",
  "sizeBytes": 9581976,
  "status": "completed",
  "surface": "composer",
  "durationMs": 595033.9,
  "textLength": 7081,
  "chunkCount": 119
}

The relevant local session directory contains 119 audio chunk files and one metadata file. The audio and full transcript are not attached because they contain private user content.

Nearby issues I found before filing:

• #21129 reports Windows dictation recording audio but producing no visible transcript
• #28765 reports pending dictation loss after ASR failure and thread/window switching

This case appears narrower: the dictation status is completed and the transcript exists locally, but the completed text was not preserved in the composer.