Codex app repeatedly surfaces self-inspection warning when Computer Use targets Codex

Open 💬 2 comments Opened May 27, 2026 by omarpinarecords

💡 Likely answer: A maintainer (github-actions[bot], contributor) responded on this thread — see the highlighted reply below.

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

Version 26.519.81530 (3178)

What subscription do you have?

Pro 20x

What platform is your computer?

Darwin 25.5.0 arm64 arm

What issue are you seeing?

The Codex app repeatedly surfaced a self-inspection limitation message after Computer Use was asked to inspect the Codex app itself.

User-facing message:

Computer Use cannot inspect the Codex app itself from inside Codex, so I cannot visually confirm the sidebar.

The underlying Computer Use tool refusal is:

Computer Use is not allowed to use the app 'com.openai.codex' for safety reasons.

The safety boundary makes sense. The issue is that Codex appears to allow the agent path to reach this known-denied target and then surfaces a noisy, low-action explanation to the user instead of preflighting/suppressing the attempt or presenting a single clear fallback.

Feedback ID / conversation ID:

019e6834-96ee-7241-8849-243ae305b74d

What steps can reproduce the bug?

Open a Codex app thread.
Ask Codex to use Computer Use to inspect or visually confirm something in the Codex app UI itself, such as the sidebar.
Computer Use refuses to inspect com.openai.codex.
Codex surfaces the self-inspection limitation message to the user.

I also uploaded feedback/trace from the app for this occurrence.

Feedback ID:

019e6834-96ee-7241-8849-243ae305b74d

What is the expected behavior?

Codex should preflight or suppress Computer Use attempts against com.openai.codex, since that target is intentionally blocked.

Expected behavior would be one of:

Avoid selecting Computer Use for the Codex app itself.
Show a single clear fallback message explaining that Codex cannot inspect its own UI.
Continue with non-visual/local-state verification when available, without repeatedly surfacing the self-inspection warning.

Additional information

Environment:

Codex app: Version 26.519.81530 (3178)
Platform: Darwin 25.5.0 arm64 arm
Subscription: Pro 20x

Trace/performance evidence was uploaded through the app. During local repro, a short performance sample of the running Codex app showed the app mostly idle in the AppKit/Electron event loop during the sample window. The related app log showed feedback/upload completing for the feedback/conversation ID above.

View original on GitHub ↗

2 Comments

github-actions[bot] contributor · 3 months ago

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

#23452

Powered by Codex Action

omarpinarecords · 3 months ago

Reviewed #23452. It appears related, but this issue is tracking a narrower downstream UX/preflight problem rather than requesting that Computer Use be allowed to inspect the Codex app.

#23452 covers the underlying self-targeting refusal:

Computer Use is not allowed to use the app 'com.openai.codex' for safety reasons.

This issue is about Codex surfacing that known-denied path back to the user as a repeated self-inspection warning, instead of preflighting/suppressing the attempt or showing one clear fallback.

To be clear: if the intended design is that Computer Use must not inspect com.openai.codex for safety/security reasons, that is reasonable. The bug is that Codex still appears to select or attempt that blocked path, then reports the tool-level refusal back to the user. The safer behavior would be for the agent/app to recognize the blocked target before tool use and immediately fall back to logs, local settings/state, screenshots supplied by the user, or a single clear unsupported-action message.

If maintainers prefer to track both under #23452, I can close this as duplicate, but I’m leaving it open for now because the requested fix here is specifically the Codex app/agent UX around the denial.