Model invents speculative edge case and reverses recommendation instead of asking

Summary

In a Codex desktop conversation, the agent had verified that a storage device was empty and had strong context that the workflow was Mac-only. When asked whether the device should be formatted, the agent introduced hypothetical Windows compatibility and recommended waiting, despite having no evidence that Windows support was relevant.

This is feedback about agent reasoning behavior, not disk-formatting functionality.

Reproduction

1. Establish that an external drive is verified empty and is being considered for use in a Mac-centric workflow.
2. Ask: "Should I format it?"
3. The agent answers, in substance: "Not yet; the future role determines the format. Use APFS for Mac-only storage or exFAT if Windows compatibility is required."
4. Ask why Windows compatibility would be required.
5. The agent admits there is no evidence it would be required and switches to recommending APFS.
6. Ask why formatting should wait.
7. The agent admits there was no reason to wait and reverses the original recommendation.

Actual behavior

The model elevated an unverified, generic edge case into a decision blocker. That produced the opposite of the justified recommendation and forced the user to interrogate the model twice to uncover that its stated uncertainty had no factual basis.

The answer sounded cautious, but the caution was fabricated rather than evidence-based.

Expected behavior

When a missing fact would materially change the recommendation, the agent should either:

• ask one direct, decision-changing clarification, such as whether the device needs Windows compatibility; or
• infer the answer from strong context, label the inference, and give the recommendation.

It should not invent a hypothetical requirement and then use that speculation to recommend inaction.

Environment

• ChatGPT desktop app: 26.825.51511 (build 7377)
• Bundled Codex CLI: 0.151.0-alpha.7.2
• macOS

Reproduce the behavior in a Codex desktop conversation using the Mac-only, verified-empty-device scenario described in the issue. Compare the model's response with the expected clarification or context-based recommendation, and verify that it does not elevate unsupported Windows compatibility into a blocker or reverse its recommendation without evidence.