Allow Codex agents to read and update their current project's settings and instructions

What variant of Codex are you using?

Codex desktop app on macOS, in a task opened from an account-backed ChatGPT Project.

What feature would you like to see?

Provide a supported, project-scoped tool that lets the Codex agent read and modify its current project's own settings when the user requests it, starting with persistent Project instructions.

For example, a user should be able to say:

Add a project instruction that bug reports must be tagged #bug and feature requests must be tagged #feature.

The agent should be able to update the authoritative Project instructions directly, preserve the existing content, and read the saved result back. The change should be visible in Project settings and available to future tasks in that project.

Suggested behavior:

• Resolve the current project's identity without requiring the user to supply an internal ID.
• Read the authoritative settings and apply a targeted update to supported fields.
• Return the persisted settings so the agent can verify the result.
• Distinguish account-backed Project instructions from a repository AGENTS.md or a generated local mirror.
• Report unsupported settings and permission failures clearly, and preserve existing authorization requirements for sensitive settings.

Additional information

Observed workflow and workaround

In this session, I asked Codex to add a Rosetta reminder-tagging rule to the project's instructions.

1. The agent initially edited the local AGENTS.md under the managed ChatGPT-project mirror. That file was read-only; the agent temporarily enabled writing, appended the rule, restored the original permissions, and verified the local content.
2. The directory was not a Git repository, so there was no commit to make. More importantly, the file identified itself as a local project mirror whose contents could be replaced. Editing it had not updated the authoritative Project instructions.
3. When asked to update the source project settings, the agent found no project-settings update operation among its available tools.
4. Attempting to use native computer automation on the Codex app was blocked: "Computer Use is not allowed to use the app 'com.openai.codex' for safety reasons."
5. The successful workaround was browser automation in an authenticated Chrome session: open chatgpt.com, go to Projects, find the project, open its options menu, choose Project settings, append the rule in Instructions, and click Save.
6. The agent reopened Project settings and verified that both the original instructions and the new tagging rule were saved.

The local permission change above is a description of the initial attempt, not a recommended solution. The browser edit was what updated the authoritative project settings.

This routine project-maintenance request should have a direct supported tool path instead of requiring a separate browser session and UI automation. The request is for an authorized project-settings capability, not removal of the restriction on automating Codex's own window.

Related issues

• #45494 — describes local AGENTS.md edits diverging from Project UI instructions. This request specifically asks for an agent-callable read/update capability for the authoritative settings.
• #27604 — requests access to or reuse of ChatGPT Project instructions.
• #36575 — requests project-level instructions in Codex Desktop.

The observation above concerns the tools available in this session; it is not a claim that no internal project-settings API exists.

No repository source file or test is named. Start by tracing the tools available to a Codex desktop task and how they relate to the account-backed Project settings; treat the local AGENTS.md mirror as non-authoritative. Done means an authorized project-scoped read/update path persists supported instructions, preserves existing content, returns the saved settings, and reports unsupported fields or permission failures clearly.