[Windows Desktop] Reproducible “ChatGPT hit a snag” immediately after a command is stopped at a sandbox/permission boundary

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

Latest Windows Codex Desktop version available to me as of 2026-08-31. [If needed, I can provide the exact About Codex version.]

What subscription do you have?

ChatGPT Pro($100)

What platform is your computer?

Windows 11 x64, build 26200.

What issue are you seeing?

Codex Desktop repeatedly enters the “ChatGPT hit a snag” fatal error state while working on one specific local development task.

The failure is reproducible after restarting the app and also reproduced after updating Codex and starting a fresh thread.

For internal correlation, the affected thread ID is:

01a05805-47a8-7b33-a3e0-a0c6a3b0cd4e

The important detail is that the failure happens around a sandbox / permission boundary.

Immediately before “ChatGPT hit a snag”, the visible task transcript ends with:

The existing machine-readable state is still present, and the previous run had already classified the recovery as HYG002_CRASH_RECOVERY_CONTINUE_EXISTING.

The release/repair branches are still at the accepted HEAD, while the unfinished implementation is preserved as uncommitted changes in 6 production modules plus dedicated tests/docs/state.

I will first run the dedicated tests without starting any real execution child.

Stopped command

The dedicated test was interrupted at the sandbox boundary. Normal repair/test/commit/release-promotion and conditional offline closure require write permission only to the specified repository and external evidence root.

I will request access limited to those two paths. I am not requesting permission for live execution, market operations, orders, broker access, scheduler, or other production actions.

Stopped command
At or immediately after the second Stopped command, the desktop app displays:
ChatGPT hit a snag
and the task cannot continue normally.

### What steps can reproduce the bug?

1. Open Codex Desktop on Windows.
2. Start a fresh local Codex thread against an existing Git repository/worktree.
3. Give the agent a substantial development task that:
   - first performs read-only repository/worktree inspection;
   - uses parallel/subagent analysis;
   - then runs a dedicated test or command;
   - reaches a sandbox/write-permission boundary and needs a narrowly scoped permission escalation.
4. The agent successfully performs the read-only inspection.
5. The transcript shows Stopped command when the test/command reaches the sandbox boundary.
6. The agent attempts to continue/request the necessary limited permission.
7. Codex Desktop displays ChatGPT hit a snag.
Restarting the app and retrying has reproduced the problem multiple times.

### What is the expected behavior?

A stopped command or sandbox permission boundary should be handled normally:
- either request permission from the user;
- or fail closed and return control to the conversation.
It should not terminate/corrupt the active Codex task or display ChatGPT hit a snag.
The task should remain resumable after a denied/stopped command.

### Additional information

This does not appear to be simple machine-wide resource exhaustion:
- other Codex development tasks can continue running normally on the same machine;
- another long-running development task has remained stable while this task repeatedly fails;
- the failure is strongly associated with this task reaching the stopped-command / sandbox-permission transition.
The affected repository itself remains intact. The crash occurs during the Codex orchestration/execution flow, not because the repository was deleted or reset.
Earlier diagnostics for the same general failure symptom on this machine also found no normal Windows Application Error/WER crash at the moment the UI showed ChatGPT hit a snag.
I can provide sanitized diagnostic logs or additional process/app-server information if useful. I have intentionally not attached repository contents, credentials, or full raw logs publicly.
Thread ID for OpenAI-side correlation:
01a05805-47a8-7b33-a3e0-a0c6a3b0cd4e

Start at the Windows Desktop execution flow and reproduce the transition from a stopped command at a sandbox or permission boundary to the “ChatGPT hit a snag” state, using a fresh thread and a dedicated command. Compare the behavior with a denied or stopped command in the same flow; done means the task either requests narrowly scoped permission or fails closed and remains resumable without terminating the conversation.