[Bug] Codex App repeatedly disconnects and cannot complete tasks on macOS

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

26.803.61601 — Powered by Codex & OWL, released Aug 8, 2026

What subscription do you have?

ChatGPT Plus

What platform is your computer?

Darwin 25.3.0 x86_64 i386

What issue are you seeing?

Since August 10, 2026, I have been unable to complete any task in the Codex App.

The app opens normally and can access and read my project files. However, every task disconnects before completion. The app repeatedly switches between disconnected and reconnecting states, creating an endless disconnect/reconnect loop. The visible status message is “Disconnected.”

This happens consistently across different tasks and models. Switching models, refreshing, retrying the task, and restarting the Codex App many times have not resolved the problem.

My internet connection appears to be working normally. On the same MacBook Pro and network, Claude Code and Claude Cowork respond immediately, and the ChatGPT website also works normally. The problem appears to be specific to task execution in the Codex App.

The task in which the issue originally occurred was automatically refreshed/reloaded by Codex. As a result, the original task state and complete error details are no longer visible, so I am unable to copy the full original error output. However, the same disconnection behavior continues to occur on every new task.

What steps can reproduce the bug?

1. Open Codex App version 26.803.61601 on macOS.
2. Sign in with a ChatGPT Plus account.
3. Open any local project or working folder.
4. Start a new task using any available model.
5. Ask Codex to perform a normal task that requires reading the project files and completing work.
6. Codex starts the task and can access and read the files.
7. Before completing the task, the app displays “Disconnected.”
8. The app then repeatedly disconnects and attempts to reconnect, creating an endless disconnect/reconnect loop.
9. Refresh the task, retry it, restart the app, or switch to another model.
10. The same problem occurs again, and the task never completes.

The issue has occurred on every task since August 10, 2026.

No code snippet is applicable because this is a Codex Desktop connection and task-execution issue, not a bug in project code.

The affected task was automatically refreshed/reloaded by Codex, so its original session ID and complete error details are no longer visible. Token-limit and context-window usage are also unavailable. The issue occurs even on newly created tasks before a large context has accumulated.

What is the expected behavior?

The Codex App should maintain a stable connection after starting a task, continue working after reading the project files, and complete the requested task successfully.

Switching models or starting a new task should not cause a repeated disconnect/reconnect loop. If a task fails, the original task state, session ID, and complete error details should remain available for troubleshooting.

Additional information

Environment:

• Codex App version: 26.803.61601
• Build information: Powered by Codex & OWL
• Release date shown in the app: Aug 8, 2026
• Subscription: ChatGPT Plus
• Computer: MacBook Pro
• Platform: Darwin 25.3.0 x86_64 i386
• Issue began: Aug 10, 2026
• Reproducibility: Every task

I have restarted the Codex App many times, refreshed and retried the tasks, and switched between different models. None of these actions resolved the problem.

On the same MacBook Pro and network, Claude Code and Claude Cowork respond immediately, and the ChatGPT website works normally. This suggests that the issue is specific to task execution in the Codex Desktop App rather than a general internet connectivity problem on my computer.

GitHub suggested the following possible duplicates:

• #37992 is very similar, but it affects Windows. My issue occurs on macOS and affects every task.
• #37285 concerns remote conversations failing to load when moving from an Intel Mac to an Apple Silicon Mac, which does not match my issue.
• #33094 concerns disconnections specifically while analyzing images, while my issue affects all tasks.

The original affected task was automatically refreshed/reloaded by Codex, so I can no longer retrieve its session ID or full error output.

Start by reproducing the disconnect/reconnect loop in the Codex App on macOS using a new task and any available model, then inspect the task-execution connection path and available app diagnostics. Done means tasks remain connected through completion, retries and model switching do not reproduce the loop, and the failed task retains its session details and error output.