[Codex Web] Cloud tasks stall at “Running setup scripts” before custom setup script begins

What issue are you seeing?

Codex Cloud tasks repeatedly stall at Running setup scripts before the first observable command in the custom setup script executes.

The setup script begins with:

set -Eeuo pipefail
echo "[content setup] begin"

The task log completes Codex’s built-in language-runtime configuration and prints:

Running setup scripts...

It then stops. The [content setup] begin marker never appears, so the custom setup script does not appear to have started.

I reproduced this with:

• Post-setup caching enabled
• Post-setup caching disabled
• An explicitly reset environment cache
• An interactive environment terminal
• A minimal dummy task that did not inspect or modify files

The controlled dummy prompt was:

Do not inspect or edit files. After environment setup completes, run printf 'dummy task passed\n' and reply exactly: dummy task passed

The dummy task remained at Running setup scripts without reaching the first setup checkpoint:

https://chatgpt.com/codex/cloud/tasks/task_e_6a5116edf030832ea23eb397a97bbf8e

What steps can reproduce the bug?

1. Create or use a Codex Cloud environment with a manual setup script.
2. Put an observable marker at the beginning of the script:

``bash
echo "[content setup] begin"
``

1. Start a minimal task that does not require repository inspection or modification:

``text
Do not inspect or edit files. After environment setup completes, run printf 'dummy task passed\n' and reply exactly: dummy task passed
``

1. Open the task’s environment setup log.
2. Observe that Codex completes its built-in language-runtime configuration and displays Running setup scripts....
3. Observe that the custom setup marker never appears and the task remains active indefinitely.
4. Cancel the task manually.

The behavior persisted after resetting the environment cache and also occurred with post-setup caching disabled.

What is the expected behavior?

After Codex finishes configuring its built-in runtimes, it should invoke the custom setup script and display the first checkpoint promptly.

If the container or setup runner cannot start, the task should fail with an actionable error rather than remaining indefinitely at Running setup scripts.

After setup succeeds, the dummy task should run its one command and respond:

dummy task passed

Additional information

This appears to be a regression of, or closely related to, #3941:

https://github.com/openai/codex/issues/3941

The environment’s setup commands themselves have run successfully in another real Codex Cloud container. That run produced:

[content dev preflight] portable chromium unpacked
[content dev preflight] chromium launch and DOM render ok

Evidence task:

https://chatgpt.com/codex/cloud/tasks/task_e_6a51078bff7c832e89cc8f79632a57a7

This demonstrates that the setup script and provisioned browser can work when the custom script is actually invoked.

Other observations:

• Repository: 3mdistal/agent-native (public)
• Subscription: ChatGPT Pro
• At the time of testing, 96% of the five-hour usage allowance and 99% of the weekly allowance remained.
• The setup script uses quiet package-manager reporters and does not use set -x, ruling out excessive custom-script output in the controlled reproduction.
• Multiple diagnostic tasks were cancelled manually after they failed to reach the first custom setup checkpoint.
• Reproduced on July 10, 2026.