Codex Desktop over-reports completion and compresses AGENTS.md instructions despite explicit project rules

What version of the Codex App are you using (From “About Codex” dialog)?

26.623.13972.0

What subscription do you have?

ChatGPT Plus Lite

What platform is your computer?

Microsoft Windows NT 10.0.22631.0 x64 /Windows 11

What issue are you seeing?

Codex Desktop repeatedly over-reports completion and compresses project instructions from AGENTS.md, even when the project explicitly defines stricter completion criteria.

The issue is not just a single wrong answer. In a local project, the root AGENTS.md and task notes explicitly required Codex to distinguish between:

• production implementation
• preview-only implementation
• diagnostic-only implementation
• fixture-only implementation
• UI display only
• type/contract only
• legacy or Result V3 display layer
• partial implementation
• unimplemented
• unknown / unverified

Despite those rules, Codex repeatedly treated weak evidence as stronger than it was. For example, it treated the existence of code paths, DTOs, UI display, preview runners, diagnostic PASS results, legacy/Result V3 display layers, or task-note wording as evidence of implementation/completion.

Expected behavior:

Codex should not say a feature is complete, implemented, restored, equivalent, or production-ready unless the project-defined completion gate is satisfied.

In this project, completion required at least:

1. Preserving the user's original goal without narrowing it.
2. Checking the actual production consumer path.
3. Separating production code from preview-only, diagnostic-only, UI-only, type-only, fixture-only, and legacy display paths.
4. Marking unverified areas as unknown instead of softening them into "mostly implemented".
5. Treating task-note phrases such as "preview-only", "diagnostic-only", "production not changed", "deferred", or "not calibrated" as blocking evidence against completion.

Actual behavior:

Codex read the relevant files and task notes, but still compressed the instruction set and reported several items as implemented or mostly implemented when they should have been classified as partial, preview-only, diagnostic-only, legacy display only, unimplemented, or unknown.

Impact:

This makes completion reports unreliable in large projects that depend on AGENTS.md for operational rules. It can cause Codex to:

• produce misleading implementation reports
• ignore explicit project completion criteria
• treat scaffolding or preview paths as production behavior
• require repeated user correction
• make unsafe assumptions before code changes

Suggested improvement:

When a project has explicit AGENTS.md completion criteria, Codex Desktop should enforce a stricter internal reporting gate before using completion-like language.

If production consumer paths, direct consumers, or required verification are not checked, Codex should report "partial", "not implemented", or "unknown" instead of "complete" or "implemented".

I can provide a sanitized transcript and the relevant AGENTS.md checklist if needed.

What steps can reproduce the bug?

Reproduction steps:

1. Open Codex Desktop in a local repository that has a root AGENTS.md.

1. Add explicit project rules to AGENTS.md that define completion criteria. Example:

Do not report a feature as complete or implemented unless the production consumer path is verified.

Separate the following states:
- production implementation
- preview-only
- diagnostic-only
- fixture-only
- UI display only
- type/contract only
- legacy or Result V3 display layer
- partial implementation
- unimplemented
- unknown / unverified

Do not use preview runners, diagnostic PASS, DTO/type existence, UI display, or legacy display as completion evidence.
If production behavior is not verified, report partial, not implemented, or unknown.
Ask Codex to analyze a feature implementation status in the repository. Example prompt:
Analyze the implementation status for this feature using code only.
Do not rely on documentation.
Classify implemented, partially implemented, unimplemented, and unknown.
Do not count legacy or Result V3 display layers as implementation.
Then ask Codex to re-check only relevant task notes and update the report:
Search only task notes related to this feature.
Separate implemented, unimplemented, and partially implemented.
Classify legacy Result V3 report-related items as legacy and exclude them.
Observe that Codex still compresses the project rules and over-reports implementation status. It treats partial evidence such as code presence, DTOs, UI display, preview-only paths, diagnostic runners, or legacy display layers as stronger evidence than allowed by AGENTS.md.
Expected result:
Codex should follow the project-defined completion gate and report uncertain or unverified items as partial, unimplemented, or unknown.
Actual result:
Codex reports some items as implemented or mostly implemented even though the production consumer path was not verified and the project rules required preview/diagnostic/UI/type/legacy paths to be separated.
Session/context details:
Codex Desktop version: 26.623.13972.0
Platform: Microsoft Windows NT 10.0.22631.0 x64
Shell: PowerShell 7.6.3
Session id: 019f32e4-b7c0-7631-8247-048536ad1f99
Token/context usage: not visible in the Codex Desktop UI at the time of filing

### What is the expected behavior?

_No response_

### Additional information

_No response_