[macOS] Opening a completed subagent crashes the Codex App UI
What version of the Codex App are you using (From “About Codex” dialog)?
26.727.51351 (build 6119)
What subscription do you have?
ChatGPT Pro (100$)
What platform is your computer?
Darwin 24.6.0 arm64 arm (macOS 15.7.3, Apple silicon M4 Max)
What issue are you seeing?
Opening the detail view of a completed subagent causes the entire Codex App UI to fail.
The subagent panel initially opens and its completed output is displayed correctly. Shortly afterward, the whole application UI is replaced by the generic “Oops, something went wrong” error screen.
The parent task itself remains accessible after retrying, but opening the same completed subagent causes the failure again.
What steps can reproduce the bug?
- Open an existing Codex Desktop task that used one or more subagents.
- Open the Subagents panel.
- Select a completed subagent to view its transcript.
- Wait briefly after the child transcript is rendered.
- Observe that the entire Codex UI is replaced by the generic error screen.
- Click “Try again” and reopen the parent task.
- Open the same completed subagent again.
- Observe the same application-level error.
In the attached recording, the issue reproduces twice.
What is the expected behavior?
Selecting a completed subagent should open its transcript without affecting the parent task or the rest of the application.
If the child transcript cannot be rendered, Codex should show a localized error inside the subagent panel and keep the main application usable.
Additional information
Video reproduction:
https://hbj2cpjgdu.ufs.sh/f/0IdehRjdCTn3TBox0DrO1pyjWv62JDLralBht8GXosMAdEbf
(UploadThing)
Relevant timestamps:
- Around 00:09: the completed subagent transcript is visible.
- Around 00:10: the whole application switches to the generic error screen.
- Around 00:13: the parent task is accessible again after retrying.
- Around 00:16: opening the subagent reproduces the same error.
The error screen is localized as “Oops, something went wrong” and offers “Update ChatGPT” and “Try again”, even though the failure occurs inside Codex.
The affected subagent was performing a security/XSS review. Its transcript contains the literal string javascript: in inline code. I do not know whether the failure is caused by that particular transcript content or by completed-subagent rendering in general.
No stack trace or more specific error message is shown in the UI.
Affected Codex thread ID:
019fc7f9-b09d-76b2-9770-36e19954d32b
I can provide a redacted rollout transcript if it would help with debugging.
---
A separate UI issue is also visible in this workflow: after Codex Security has been used in a conversation, its panel automatically reopens every time I return to that conversation.
Closing the panel only appears to dismiss it temporarily. Codex should remember that the panel was closed and should not reopen it automatically unless the user explicitly requests it.