This is an unofficial, unsupported local UI customization. It is not an OpenAI product, extension API, sandbox, or security boundary. Do not use it on a machine where exposing browser debugging to other local processes is unacceptable.

The launcher opens the official app with a random Chromium debugging port bound to loopback. It checks the launched process's ownership of the port, validates the browser endpoint, and targets only the primary Codex renderer. The helper injects CSS, disconnects, and exits. It does not intentionally read or upload conversation content.

The debugging listener does not close when the helper exits. It remains available until you quit the official app. Another process on your computer that can reach that endpoint may control the renderer and access sensitive session data. A random port is not authentication. Do not forward or expose it to a network, and do not add wildcard debugging origins.

Quitting the app and reopening the ordinary official app removes the injected style and starts without this launcher's debugging flags. Uninstalling the launcher does not close an already-running app's debugging listener.

• The installer creates a user-local wrapper app and two JavaScript files.
• The official app must pass the launcher's OpenAI signature checks. Its bundle is not edited or re-signed.
• Only the wrapper is ad-hoc signed locally. It is not a notarized distribution.
• Icon artwork is copied from the user's local official installation; it is not distributed in this repository or covered by this repository's MIT license.
• No persistent watcher, login agent, reload hook, or signal-based inspector activation is installed. The named legacy helper is disabled and removed.

Review the scripts before installing. Do not run them with sudo. A Codex update can invalidate the assumptions this helper relies on; a successful CSS-variable probe alone is not proof that every layout element is correct.

Use a GitHub issue for a sanitized reproducible report. Never include tokens, cookies, conversation content, full diagnostic logs, or a live debugging URL. If you find a sensitive exploit, avoid publishing exploit details in a public issue; ask the repository owner for a private reporting channel first.

There is no guaranteed support or security response timeline. Bugs in OpenAI's official app should be reported to OpenAI through its own reporting channels.