openai / openai/codex

[macOS][Computer Use] Inconsistent review of an explicitly approved recoverable grocery-cart edit

Open
#46,679 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

app bug computer-use sandbox
Dominant language
Rust
Stars
125k
Forks
19.4k
PR merge metrics
PR metrics pending

Description

Inconsistent approval review for an explicitly approved recoverable grocery-cart edit

Environment

  • Codex desktop / ChatGPT app for macOS, version 26.911.61220, build 9647 (verified from the installed app bundle).
  • macOS 27.0, build 26A428.
  • Computer-use browser control with automatic approval review.
  • Observed September 19, 2026.

Observed sequence

  1. The assistant identified the exact existing cart lines to clear and presented that action as option A.
  2. The user explicitly selected that option; the approved scope remained unchanged.
  3. The first quantity-decrement click was denied by automatic review despite its own high authorization assessment.
  4. After a read-only check of the approval context and public review decision, one retry through the same tool and browser succeeded. The UI exposed an Undo removal control.

This is one observed deny-then-allow sequence; deterministic reproduction has not been established.

Expected

Carry forward explicit, unchanged-scope user approval for the identified grocery-cart edits. Apply the current distinction between permanent deletion and recoverable removal.

Actual

The automatic reviewer recorded user_authorization=high, acknowledged that the user had approved clearing the identified 20 product lines, but denied the first decrement click because it claimed fresh confirmation was absent. The current policy supplied to the reviewer also instructed against repeated confirmation when scope and risk remain unchanged.

After inspecting the approval context and public review decision, the assistant retried the same UI action through the same tool/browser. It succeeded and Kroger exposed Undo removal. No approval settings, policy, or browser route were changed. The requested cart preparation then completed.

Impact

Repeated user handoffs and misleading claims of missing consent despite an adequate task-specific approval.

Investigation requested

Check consistency between authorization assessment and action-time confirmation handling, treatment of recoverable shopping-cart removal, and propagation of existing approval across turns. Clarify whether supported pre-tool context reaches the independent reviewer and how task-scoped approval can be reliably carried without repetitive user confirmation.

Privacy

This report omits account identity, addresses, grocery history, task IDs, and full transcripts. Session logs have not been submitted. Any optional log attachment requires a separate privacy review.

Related issue

https://github.com/openai/codex/issues/42696 concerns scheduled final submissions requiring action-time confirmation. This report concerns a recoverable cart edit during an active, explicitly approved task, where the same route first denied and then allowed the action. It does not request disabling mandatory confirmations or other safeguards.

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

No files or tests are named. Start with the observed deny-then-allow sequence, the approval context, and the public review decision described in the report. Check whether unchanged task-scoped approval reaches the independent reviewer and whether recoverable removal is treated consistently; done means a reproducible explanation or a clearly verified correction without unnecessary reconfirmation.

Written by the indexing model from the issue text.

Assessment

Tech stack
macos, rust
Domain
authorization, desktop
Issue type
Bug
Difficulty
5/5
Estimated time
Over a week
Activity status
Active
Clarity
Mostly clear
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.