openai / openai/codex

Codex became confused, defensive, and stopped executing a straightforward coding task

Open
#43,747 1 comment 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

bug model-behavior
Dominant language
Rust
Stars
125k
Forks
19.4k
PR merge metrics
PR metrics pending

Description

I asked Codex to add a reusable visual style to an existing stepper
component and use it on the pricing page. During the task, Codex
behaved in several seriously unhelpful ways:

  1. It overcomplicated a focused UI change by creating an extensive
    architecture process instead of first checking the component’s actual
    API and implementation.

  2. It modified many UI-system layers and claimed the feature was
    implemented, but did not ensure the application was consuming the
    updated compiled package. The visible UI therefore remained unchanged.

  3. When I reported that the supplied prop was not accepted, Codex did
    not calmly reconcile the architecture, source contract, and generated
    package. It oscillated between three contradictory conclusions:

    • presentation should be added;
    • presentation is invalid and should become variant;
    • neither presentation nor variant belongs in the API.
  4. It renamed the API from presentation to variant during an
    interrupted correction, even though its own architecture record
    explicitly required presentation. This left the repository in a more
    inconsistent state.

  5. It repeatedly narrated policy constraints and requested additional
    authorization instead of performing the coding work already authorized.
    It treated a necessary package compilation step as a reason to stop.

  6. It responded to corrections with more analysis rather than
    inspecting the exact current state and applying the smallest repair.
    The user had to repeatedly explain facts that Codex had already
    established earlier in the same conversation.

  7. It claimed completion without runtime or visual evidence. The
    worktree ended with approximately 18 modified files, an inconsistent
    public contract, stale generated output, and no visible improvement.

  8. After losing track of the task, it produced a large handoff instead
    of recovering and completing the straightforward correction.

The expected behavior was:

  • Inspect the existing stepper API and its canonical recipe/layout
    owners.
  • Read the already-created architecture once.
  • Make the smallest consistent contract and implementation change.
  • Update the pricing consumer with the exact accepted API.
  • Ensure the package consumed by the application is synchronized.
  • Verify the visible result before claiming completion.
  • Preserve unrelated user edits.
  • If a genuine authority conflict exists, state the exact conflict once
    and ask one focused question.
  • Never improvise successive API names or replace implementation with
    defensive process narration.

The central failure was not lack of technical capability. It was
failure to maintain a stable understanding of repository state and
accepted authority while executing a simple, already-authorized change.
Codex’s excessive process adherence, contradictory reasoning, and
premature completion claims made it actively counterproductive.

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 source files, tests, or entry points are named. Start by locating the agent workflow responsible for repository-state tracking, authorization handling, package synchronization, and completion claims, then compare its behavior with the expected sequence in the issue. Done means a focused authorized UI task preserves the API, updates the consuming package, verifies the visible result, and avoids unrelated edits.

Written by the indexing model from the issue text.

Assessment

Tech stack
rust
Domain
ai-infra-agents, devtools
Issue type
Bug
Difficulty
5/5
Estimated time
Over a week
Activity status
Active
Clarity
Needs clarification
Newbie friendliness
25/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.