openai / openai/codex

Desktop auto-review cannot verify direct owner authorization relayed to an engineer task

Open
#46,431 2 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

app bug sandbox subagent windows-os
Dominant language
Rust
Stars
125k
Forks
19.4k
PR merge metrics
PR metrics pending

Description

Problem

In Codex Desktop on Windows, the user gave a coordinating task a direct, explicit instruction to make a permanent, system-wide change to local scheduling code and its governing instructions. The coordinator relayed that exact scope to an existing engineer task working on the scheduler. Automatic approval review rejected the broader edit twice because it could not verify the engineer task's authority for the system-wide change. The edit remained unapplied even though the owner had already authorized it in the coordinating task.

The owner does not want to repeat the same approval in each execution task. The repeated review is blocking a time-sensitive, authorized local operation. We did not bypass the rejection or claim the change was complete.

Observed sequence

  1. The owner explicitly established the permanent rule and directed it to be applied across the system.
  2. The coordinator passed the owner's scope and wording to an existing engineer task.
  3. The engineer prepared the broader local rule/source revision. Automatic review rejected the action, citing inability to verify authority in that task; a second attempt with the owner's exact wording was also rejected.
  4. A narrower local edit for one scheduled item succeeded, leaving the system-wide rule only partially implemented.
  5. The owner asked why prior authorization was not enough and explicitly declined another approval loop.

Expected behavior

A supported coordinator-to-engineer handoff should carry verifiable provenance and scope of a direct user instruction, or the product should provide a clear way to complete authorization from the original task. Review feedback should distinguish a hard policy restriction from missing cross-task provenance, so the agent can resolve the actual blocker. The owner should not have to restate the same decision repeatedly.

This report does not request a silent bypass of review. It asks for a workable path for authorized cross-task edits and an accurate explanation when review blocks one.

Related: #46159 and #45422. Private project names, user identity, file paths, task IDs, customer details, and logs are intentionally omitted.

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

Start by reviewing the coordinator-to-engineer task handoff and automatic approval review paths in Codex Desktop on Windows. Reproduce a direct authorization relayed to an engineer task, then verify that completion preserves authorization scope or provides a clear path from the original task. Done feedback should distinguish a hard policy restriction from missing cross-task provenance.

Written by the indexing model from the issue text.

Assessment

Tech stack
rust
Domain
authorization, desktop, security
Issue type
Feature
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.