Codex Desktop crashes after a subagent approval card becomes unresponsive

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

26.707.31428

What subscription do you have?

Pro x 5

What platform is your computer?

Darwin 25.5.0 arm64 arm

What issue are you seeing?

Title: Codex Desktop crashes after a subagent approval card becomes unresponsive

Expected behavior

• File-change approval controls should remain interactive after app focus changes.
• A single unresponsive subagent should not make its task conversation crash.
• The parent task should recover, fail the subagent cleanly, or allow the user to cancel/retry it without leaving the task permanently active.

Actual behavior

The approval UI became non-interactive, reopening the affected task crashed the desktop app, and the parent task became permanently stuck in an active state.

Environment

• Platform: macOS
• Surface: Codex Desktop app
• Workflow: multi-agent / subagent workflow
• Task type: long-running Codex Security deep scan
• Codex app version: [please fill in from “About Codex”]
• macOS version: [please fill in]

Additional context

The issue began after one worker requested approval for a file modification. Other workers completed successfully, suggesting the failure is isolated to the approval state or its interaction with subagent/task rendering.

What steps can reproduce the bug?

What happened

While running a long-lived multi-agent Codex Security task in the Codex Desktop app on macOS, one subagent displayed a file-change approval card.

After switching back to Codex from another application:

1. The approval controls on that subagent card (such as Allow and Deny) were no longer clickable.
2. Clicking the stuck subagent card also had no effect.
3. Other subagents in the same task continued to run and finish normally.
4. After switching to another conversation and then reopening the affected conversation, the Codex Desktop app crashed and displayed the Chinese error screen: “Oops, something went wrong”, with only “Update ChatGPT” and “Retry” available.
5. Once the other workers finished, the parent task remained active but made no further progress. It was stuck while attempting to close or replace the unresponsive worker for several hours.

What is the expected behavior?

_No response_

Additional information

_No response_