Codex App stops rendering assistant messages after realtime voice chat

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

26.908.9136.0

What subscription do you have?

ChatGPT Pro $200

What platform is your computer?

Microsoft Windows NT 10.0.26200.0 x64

What issue are you seeing?

After using realtime voice chat in an existing Codex task and then ending the voice session, the task can continue receiving later typed prompts and performing the requested work, but assistant commentary and final responses are not rendered visibly in the chat.

The agent is clearly still active: it runs tools, updates external records, executes commands, and completes the requested operations. However, the user sees no corresponding assistant text and cannot tell whether the task is working, blocked, or finished. The only evidence may be the side effects of the work.

This is not the normal voice transcription feature. It occurs after using the interactive realtime voice-chat mode. I encountered it again in the task associated with the Feedback ID below, and I saw similar behavior a few weeks earlier after a realtime voice session.

What steps can reproduce the bug?

Feedback ID: 01a04a6e-ccdb-7cb1-9f3e-86fd102d35af

1. Open an existing Codex App task with a substantial conversation history.
2. Start realtime voice chat in that task.
3. Have several back-and-forth voice exchanges, including requests that cause the agent to inspect data or use tools.
4. End the realtime voice session.
5. Continue in the same task by sending a normal typed prompt that requires tool use or several minutes of work.
6. Observe that the agent receives the prompt and performs the work, but its progress commentary and/or final response do not appear in the visible chat.
7. Confirm the work occurred by inspecting its external effects, such as updated Linear records, commands that ran, or newly created tasks.

Observed in this occurrence:

• The typed request was successfully executed after realtime voice ended.
• Multiple tool-backed actions completed.
• The user did not receive visible status or completion text and had to ask why the agent had done substantial work without saying anything.
• No visible error, disconnected-state warning, or indication that assistant messages had failed to render appeared.

The issue appears intermittent, but the common factor in both occurrences was continuing the same task after realtime voice chat.

What is the expected behavior?

After realtime voice chat ends, the task should return cleanly to normal text-chat behavior.

Every subsequent assistant commentary update, clarification, error, and final response should render in the visible conversation as it normally would. If message delivery or rendering fails, the app should show a clear recoverable error instead of allowing the agent to continue performing hidden work.

Additional information

This is especially problematic for long-running agent tasks because the backend appears healthy and continues making changes while the user has no visible progress, decision request, or completion summary.

The failure seems more likely to involve the Codex App conversation/rendering state after realtime voice ends than model execution itself, because the agent continues following instructions and producing tool side effects.

Please correlate client and conversation events using Feedback ID 01a04a6e-ccdb-7cb1-9f3e-86fd102d35af. The recurrence a few weeks apart suggests this may be an existing intermittent regression rather than a one-time network interruption.

I intentionally have not included private repository contents, credentials, or screenshots. The report contains the observable behavior and reproduction sequence in text.

The problematic task has self diagnosed and filled out this entire form, though still provides no visible chat feedback to the user.

Confirmed workaround

Starting another realtime voice-chat session in the affected task, confirming the assistant responds in voice, and then ending that session restored normal visible text responses. Afterward, a typed message in the same task received a visible assistant response again.

This workaround recovered the task in this occurrence, but it does not resolve the underlying intermittent bug. It suggests the additional voice-session lifecycle reset or repaired the conversation's text-rendering state.

Start by reproducing the sequence with Feedback ID 01a04a6e-ccdb-7cb1-9f3e-86fd102d35af: use realtime voice chat in an existing task, end it, then send a typed prompt requiring tool work. Compare the affected task with the confirmed workaround of starting and ending another voice session. Done means subsequent commentary, errors, and final responses render visibly, or a clear recoverable error appears.