[Codex Desktop] Agent falsely reports feedback unavailable despite callable feedback/upload

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

Unified ChatGPT/Codex Desktop 26.825.31414 (build 7287)

What subscription do you have?

NOT_PROVED

What platform is your computer?

Darwin 25.6.0 arm64 arm

What issue are you seeing?

After an explicit request to submit two in-product feedback reports, Codex searched the current task-tool catalog, found no high-level feedback writer, and returned SUBMISSION_SURFACE_NOT_EXPOSED. After correction, the same task inspected the bundled app-server schema, found the documented feedback/upload method, and successfully submitted both reports with verified receipts.

The dynamic task-tool catalog did not expose a semantic feedback writer, but that was incorrectly treated as proof that the Codex host had no feedback capability.

The false unavailable classification persisted even though the user had explicitly supplied the bundled OpenAI Docs skill, pointed to the native OpenAI Developer Docs MCP integration as a discovery lead, and provided official documentation links. The agent still did not follow those leads to the documented app-server feedback/upload method before returning the unavailable state.

What steps can reproduce the bug?

1. In a Projectless Codex Desktop task, ask Codex to submit an authorized in-product feedback report.
2. Supply the bundled OpenAI Docs skill, identify the native OpenAI Developer Docs MCP integration, and provide the official app-server documentation link as capability-discovery leads.
3. Let the agent inspect only its currently advertised task tools instead of following those leads.
4. Observe that it may return SUBMISSION_SURFACE_NOT_EXPOSED when no semantic feedback tool appears.
5. Direct the same task to inspect the bundled app-server schema or official documentation.
6. Observe that it discovers feedback/upload and can submit the report successfully.

No session ID, task ID, local path, logs, or private content is included.

What is the expected behavior?

Before claiming feedback is unavailable, Codex should follow explicit user-supplied skill, MCP, and documentation leads and exhaust supported capability-discovery paths, including official documentation, /feedback, the bundled app-server schema, and any supported UI control. Absence from the dynamic task-tool catalog must not be treated as proof that the host capability does not exist.

Additional information

Feedback ID: no-active-thread-01a04672-fe94-7523-bbf2-64dbf93cbe97

Official documentation: https://learn.chatgpt.com/docs/app-server

Related, not duplicate: https://github.com/openai/codex/discussions/40128

Community corroboration: https://community.openai.com/t/demonstration-creating-an-openai-feedback-id/1392745

Diagnostics, logs, traces, screenshots, and attachments: NONE.

Start by tracing the Projectless Codex Desktop capability-discovery path from the dynamic task-tool catalog, then compare it with the bundled app-server schema and official app-server documentation. Reproduce the authorized feedback request and inspect how the absence of a semantic feedback writer becomes SUBMISSION_SURFACE_NOT_EXPOSED. Done means supported discovery leads, including feedback/upload, are considered before reporting feedback unavailable.