Codex mobile remote access can reuse revoked enrollment and connect to the wrong Mac

Summary

Codex mobile remote access can enter a cross-host stale state with two Macs on the same ChatGPT account/workspace:

• Mac A shows Codex mobile / remote access as connected or successfully set up.
• The iPhone ChatGPT app shows or connects to Mac B instead.
• Re-running mobile setup / scanning a new QR code on Mac A does not replace the stale enrollment.
• Codex Desktop keeps reusing a revoked remote-control client and repeatedly reconnects to the wrong remote-control host.

This makes Mac A impossible to pair from the UI even though setup appears successful.

Environment

• Codex App for macOS: 26.513.31313
• macOS: 26.5, build 25F71
• ChatGPT mobile: latest available from App Store as of 2026-05-19
• Account/workspace: same ChatGPT account and workspace on both Macs and the iPhone
• Setup: two Mac hosts, both used with Codex mobile remote access

Expected behavior

Each Mac should have an independent Codex mobile remote-control enrollment / host identity.

When setting up mobile access on Mac A:

• Mac A should generate or use an enrollment for Mac A.
• The iPhone should show Mac A as the connected host.
• If a previous enrollment is revoked, Codex should discard it and force a fresh setup.
• A stale or revoked enrollment from Mac B should not make Mac A appear connected.

Actual behavior

Mac A shows mobile access as connected / set up, but the iPhone shows the connected host as Mac B.

Mac A repeatedly tries to use a stale enrollment and wrong remote-control:<env> host. The local state had:

• electron-remote-control-client-enrollments
• added-remote-control-env-ids
• selected-remote-host-id
• remote-connection-auto-connect-by-host-id
• remote-projects
• codex-mobile-has-connected-device
• electron-local-remote-control-installation-id

The remote project shown on Mac A was from Mac B.

Logs

Identifiers are redacted, but the important failure pattern is:

[remote-control-websocket] remote_control_client_refresh_start_request hasExistingEnrollment=true localEnrollmentClientId=<redacted stale client id> ...
[AppServerConnection] app_server_connection.transport_connect_failed errorMessage="Remote-control client has been revoked" hostId=remote-control:<redacted stale env id> transport=websocket
[remote-connections/window-context] connect_from_config_apply_failed errorMessage="Remote-control client has been revoked" hostId=remote-control:<redacted stale env id>

This repeats on app startup and QR setup retries.

Recovery attempts

Clearing the remote-control keys from ~/.codex/.codex-global-state.json while Codex is running is not sufficient because the running app rewrites the stale state back to disk.

Clearing both ~/.codex/.codex-global-state.json and ~/.codex/.codex-global-state.json.bak while Codex is fully stopped removes the local stale state. However, there is no obvious mobile UI to remove the incorrect/stale host from the iPhone side, and the Mac UI can still present misleading connected state before cleanup.

Why this looks like a product bug

The app treats a revoked remote-control enrollment as an existing valid local enrollment (hasExistingEnrollment=true) and keeps retrying it instead of invalidating it.

The UI also appears to conflate or reuse remote-control host state between two Macs on the same account/workspace. This produces a cross-host mismatch: Mac A reports success, while the iPhone points at Mac B.

Suggested fixes

• When the relay returns "Remote-control client has been revoked", clear the corresponding local enrollment and remote host mapping automatically.
• Do not display mobile access as connected if the stored enrollment is revoked.
• Ensure electron-local-remote-control-installation-id / host identity cannot be cloned or reused across Macs without re-enrollment.
• Add a visible "Remove this Mac / disconnect host" action in ChatGPT mobile Codex.
• Add a "Reset Codex mobile pairing" action in Codex App Settings > Connections.