Codex Desktop removes the feedback affordance when a user stops an unsafe tool turn

Summary

When a user stops or rejects an active Codex Desktop turn, the interrupted turn does not retain the normal negative-feedback/report affordance.

This creates an accountability gap: the user can either allow a problematic action to finish and retain the ability to rate it, or intervene successfully and lose the most direct reporting mechanism for the behavior that required intervention.

Environment

• Codex Desktop: 26.707.8479.0
• Bundled Codex CLI: 0.137.0
• OS: Windows 11 x64 (Microsoft Windows NT 10.0.26200.0)
• Native Windows Desktop session
• Observed: July 13, 2026

Steps observed

1. Codex begins a tool-driven implementation workflow.
2. The workflow attempts an action the user did not authorize.
3. A technical approval/escalation prompt appears.
4. The user rejects or stops the action.
5. The task is marked interrupted or aborted.
6. The normal negative-feedback control is unavailable for the offending turn.

This was observed once. The UI behavior appears deterministic, but broader reproduction has not been attempted.

Actual behavior

Stopping the turn prevents execution but also removes the response-level feedback target.

The user cannot directly associate negative feedback with:

• the assistant’s interpretation of the request;
• the attempted tool action;
• the escalation request; or
• the sequence that led to intervention.

Feedback can only be placed on a later response, which separates the report from the originating event and may omit the relevant execution trace.

Expected behavior

Interrupted and rejected turns should remain visible and reportable.

The retained turn record should include:

• the assistant reasoning or commentary shown to the user;
• attempted tool actions;
• the approval request and its requested scope;
• the user’s rejection or stop action;
• confirmation of whether any mutation occurred; and
• the normal negative-feedback/report control.

The user should not need to let an unwanted action complete in order to report it accurately.

Impact

• Safety-relevant interventions generate less feedback than completed turns.
• Product telemetry may undercount boundary violations that users successfully stop.
• Model-quality reviewers lose the most relevant turn-level context.
• Users must reconstruct incidents manually using screenshots and later messages.
• The feedback system creates survivorship bias toward actions that were allowed to finish.

Privacy

The public report excludes the raw task transcript, screenshots, local paths, account information, and task identifiers.

Unredacted evidence can be supplied privately through an authorized OpenAI feedback or support channel.