ChatGPT Project local AGENTS.md edits can silently diverge from Project UI instructions

Problem

A Codex task opened from an account-backed ChatGPT Project receives a local mirror containing AGENTS.md and source files. A user or agent can change the local AGENTS.md, but the Project instructions shown in the desktop app's Project settings remain unchanged. The two copies can then disagree. It is easy to assume that editing the visible local instructions updated the Project, especially when the assistant reports that it "updated the instructions."

Steps to reproduce

1. Open a Codex task from a ChatGPT Project that has Project instructions.
2. Edit the task's local mirror AGENTS.md to change a project rule.
3. Reopen the Project instructions field in the app.
4. Observe that the field still contains the earlier rule.

Actual result

The local file and Project UI remain separate without an obvious in-flow indication of which is authoritative or whether the local edit will carry into future project tasks. This can leave conflicting guidance and cause users to believe a change persisted when it did not.

Expected result

Make the one-way mirror and source of truth clear at the point of editing or reviewing project instructions. For example, label the generated AGENTS.md as a local snapshot, warn when it is modified, and offer a clear path to update the Project instructions field. A two-way update could also solve it if supported, but the main request is to prevent silent divergence.

This is a usability and documentation report, not a claim that the current one-way behavior is an unintended sync bug. No personal project data is included.

Start by tracing how a Codex task from a ChatGPT Project creates the local AGENTS.md mirror, then inspect the Project instructions field in the desktop app. Compare the local-file and Project settings flows. Done means users can identify the source of truth, recognize a local snapshot, and find a clear path to update Project instructions.