openai / openai/codex

Codex desktop agent regresses approved image regions and reports visual defects fixed without verification

Open
#45,979 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

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

Description

Product and environment

Windows ChatGPT desktop app, Codex/Work local task. The agent controlled a local third-party Stable Diffusion WebUI Forge-Neo installation. This report concerns Codex agent planning and verification, not an OpenAI image-generation model or a confirmed Forge-Neo bug.

Reproduction pattern
  1. Give the agent an existing full-body illustration and instruct it to edit one local region while preserving all other regions.
  2. Specify concrete accepted constraints for footwear: exactly two straps per boot; two buckles visible on the viewer-left outer side; buckles on the far outer side of the viewer-right boot must not be visible; no extra hardware. Require natural ankle continuity and full-image quality.
  3. Let the agent perform several local inpaint/reference-image iterations and inspect each result.
  4. Point out defects that remain visible and ask it to correct them.
Observed

Across repeated attempts, the agent sometimes described a defect as fixed while the output still visibly contained it (for example, a misplaced buckle or unwanted chin redness). It also reused an older full-image base, causing already-approved shirt-knot and accessory edits to regress while trying to fix the boots. Local improvements frequently introduced full-image failures such as broken ankle seams, floating straps, color mottling, or rough paint. Later runs did record failures more honestly, so this report does not assert that every run was misreported.

Expected

The agent should maintain a per-region accepted-reference ledger; avoid silently reverting approved regions; compare each output against explicit numeric/spatial constraints; inspect both the edited crop and the full composition before saying “fixed”; and report uncertainty or failure when visual validation does not pass.

Scope and privacy

The source illustration, character/project identity, personal paths, and conversation transcript are intentionally omitted. I can provide a redacted minimal reproduction if maintainers specify a private submission route. This is separate from a billing support request.

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 file, test, or implementation entry point is named; start by obtaining the redacted minimal reproduction and tracing the Codex agent planning and verification flow involved in local image-editing tasks. Done means approved regions are retained, crop and full-image constraints are checked, and the agent reports failure or uncertainty when visual validation does not pass.

Written by the indexing model from the issue text.

Assessment

Tech stack
rust
Domain
ai, devtools
Issue type
Bug
Difficulty
5/5
Estimated time
Over a week
Activity status
Active
Clarity
Needs clarification
Newbie friendliness
30/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.