Codex gave contradictory stop/restart advice for a live cloud-agent run from an environment label

What issue are you seeing?

Codex gave contradictory operational advice while I had a live Claude Cloud research run open.

I asked whether I was running a Claude research prompt in the correct place and provided a screenshot. The screenshot showed:

• Task title: AI reviewer supervision burden research
• Environment label: Default Cloud Environment · operator-kit
• A live run doing web/command work
• A message that it would read a shipped /loop skill locally

Codex first told me:

You are in the wrong Claude environment. Stop this run.

It also claimed the research was “contaminated” by the attached repository and should not be used.

After I challenged this, Codex reversed itself without performing any verification:

You followed the prompt correctly: run it in Claude Code Cloud. Do not stop or restart it. Let the current run finish.

It then made a second unsupported claim: that the run would be contaminated only if the final report cited private repository files.

The environment label alone did not establish what private files were read, what influenced the research, whether anything was modified, or whether the result was usable. No tool call inspected the active session's file accesses, instruction context, or completed output between the contradictory answers.

The original prompt was itself ambiguous about whether “Claude cloud research chat” meant a regular Claude chat or Claude Code Cloud. Codex should have identified that ambiguity instead of telling me to stop a live run.

What steps can reproduce the bug?

1. Give Codex a prompt intended for another AI product that says to use a “cloud” research chat but does not clearly distinguish ordinary chat from a cloud coding environment.
2. Show Codex a screenshot of a live run with an environment or repository label.
3. Ask where the prompt should run.
4. Observe that Codex may infer that the run is wrong or contaminated solely from the label and recommend cancellation/restart.
5. Challenge that inference.
6. Observe that Codex may reverse the advice without inspecting the actual accessed files, runtime inputs, or output.

What is the expected behavior?

Codex should:

• Distinguish visible screenshot facts from unknown runtime behavior.
• Read the exact prompt and state when the destination is ambiguous.
• Avoid directing a user to interrupt or restart live work without evidence tied to the actual scope or behavior.
• Explain what additional evidence would be needed to assess research independence.
• If it corrects itself, identify the unsupported inference and what new evidence supports the revised advice. User pushback alone is not verification.
• Include a short practical run guide when it provides a prompt for another product: exact surface, model/mode, inputs, paste/send step, expected output, return route, and stop condition.

Additional information

• Platform: Codex Desktop on macOS; exact version, Codex model selector, and reasoning effort were not captured.
• The live screenshot showed Claude controls Opus 5 and Max; those do not identify the Codex model that produced the advice.
• I do not know whether I stopped or restarted the run, whether it accessed private files, or whether it caused measurable cost. Those outcomes remain unknown.
• I searched existing openai/codex issues for operator-kit and Stop this run before filing and found no matching issue.
• No credentials, private repository contents, or account identifiers are included here.

No implementation files or tests are named. Start by reproducing the ambiguous cloud-environment prompt in Codex Desktop and inspect how screenshot facts, unknown runtime behavior, and user pushback affect the response. Done means the behavior distinguishes evidence from inference, avoids unsupported stop/restart advice, and explains what verification is needed.