macOS Codex 26.422.21637: OAuth/ChatGPT auth triggers AppRoutes re-mount loop (regression of #12754)

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

• #19220
• #19075

Powered by Codex Action

No, it is my personal pro account.

Pulled more data from a 15-second launch capture. Tagging @HeMuling since this may overlap with #19075's timeline (their Windows hang also showed app/list failures).

Method-call distribution in a 15-second window (renderer in loop):

| method | count | notes |
|---|---|---|
| experimentalFeature/enablement/set | 121 | ~13 calls per renderer mount, fired repeatedly |
| account/read | 66 | once per re-mount |
| experimentalFeature/list | 38 | |
| config/read | 6 | |
| thread/read | 4 | (pinned threads, only loaded once) |
| thread/list | 1 | only during first mount |
| model/list | 1 | only during first mount |
| configRequirements/read | 1 | only during first mount |
| collaborationMode/list | 1 | only during first mount |
| app/list | 1, durationMs=5046 | every other call is <100 ms |

The app/list call stands out — it takes 5 seconds where everything else is sub-100 ms. Ordinary init calls (thread/list, model/list, configRequirements/read, collaborationMode/list) fire exactly once, while experimentalFeature/enablement/set and account/read loop forever.

First renderer mount is normal (T+2948 ms), the loop starts at the second mount (T+3213 ms):

[renderer] app routes mounted after 2948ms          ← first mount
experimentalFeature/enablement/set × 13              ← writes feature flags
[renderer] app routes mounted after 3213ms          ← re-mount (loop starts)
[renderer] app routes mounted after 3269ms
[renderer] app routes mounted after 3341ms
[renderer] app routes mounted after 3395ms
[renderer] app routes mounted after 3448ms
[renderer] app routes mounted after 3518ms
[renderer] app routes mounted after 3599ms
... continues every ~50–80 ms for the whole 15 s window

Same webContentsId=1 throughout — React is repeatedly remounting the route tree inside a live renderer (not a process respawn).

Hypothesis: something inside the renderer's handling of the initial experimentalFeature/enablement/set batch (or the app/list response that arrives ~5 s in) triggers a state change that React reacts to by re-mounting <AppRoutes>; on re-mount the same side-effect fires, infinite loop.

Account is an individual Personal ChatGPT Plus — no team/business workspace, so #19075's "deactivated workspace" doesn't apply here, but app/list taking 5 seconds on this single-workspace account suggests it may still be where the renderer gets wedged.

Workaround still stands: API-key auth.json (~215 B) → 2 mounts, no loop. Any OAuth JWT → 141 mounts.

Same here. No reference to "workspace_dependencies" in the config.toml; just totally bricked as of today.

However, I updated the CLI (had to rm the Brew cask manually), logged out and then back in. All good now.

Thanks for the tag. Small clarification: my related issue is #18556, not #19075.

I think #18556 is adjacent at the startup-state reconciliation level, but probably not the same direct failure mode as this issue. In #18556, the app restored a stale discovered SSH host/project state and then tried to route requests to a hostId that no longer had an AppServerManager. The key error was No AppServerManager registered for hostId: remote-ssh-discovered:....

This issue looks more auth/app-list/feature-enablement related, especially given the OAuth auth.json flip and the repeated AppRoutes remount loop. So I would treat #18556 as the same broad class of “invalid persisted state bricks startup,” but not as a duplicate unless logs show a stale remote hostId involved.

#18974 looks much closer to #18556.

I don't think they test much before release!

My other Mac happened too. I try to upgrade mac os to 26.4, no help here.