[macOS, 26.903.61454, Edge] Extension discovery succeeds but tab enumeration times out, including in a fresh task

App and Platform

• ChatGPT desktop / Codex-integrated app and bundled browser components: 26.903.61454.
• macOS: 26.0.1, build 25A362, Apple Silicon.
• Microsoft Edge: 152.0.4191.66.
• OpenAI ChatGPT Edge extension: 1.26.901.11451, installed and enabled in the selected browser profile.

Problem

The official Computer Use integration can discover and select the Edge extension browser, but tab enumeration fails after approximately 30 seconds with:

Error: nodeRepl.fetch request failed

This prevents browser automation before any target website content is read. The failure was reproduced in both an existing task and a newly created diagnostic task. No sensitive site was used in the diagnostic task.

Observed Results

One diagnostic timing wrapper had a local undeclared-variable error after browser selection returned its documentation. This was kept separate from the connection failure: the subsequent cua.listTabs call used the browser ID from the successful live discovery directly, without relying on the failed variable assignment.

Reproduction

1. Run the desktop app and Edge with the OpenAI extension installed and enabled in the selected profile.
2. Open the official ChatGPT Edge Add-ons listing; no private website is needed.
3. Initialize the official Computer Use tool using its documented entry point.
4. Run standalone cua.listBrowsers() and obtain the live Edge browser ID.
5. Select that Edge browser using cua.getBrowser({id: observedEdgeId}).
6. Independently run cua.listTabs({browser: observedEdgeId, emit:false}). It fails after approximately 30 seconds with the error above.
7. In a newly created diagnostic task, repeat live discovery and the single enumeration call. Discovery still succeeds and enumeration still fails.

Expected Behavior

After discovery and selection, tab enumeration should succeed or return an actionable error identifying the failing component. A fresh task should not reproduce an opaque connection error if the issue is solely stale state in the old task.

Diagnostics and Recovery Already Tested

• The bundled official running-browser check passed.
• The installed-browser and extension checks passed; the extension was enabled in the selected profile.
• The native-host manifest check passed, including allowed extension origins.
• The running native-host version matched the current desktop/browser components.
• Strict code-signature verification passed and identified the official OpenAI signer. No manual re-signing or host repair was performed.
• A new normal Edge window in the same profile did not restore tab enumeration.
• A full normal Edge quit/relaunch also restarted its native host. The same profile and extension instance remained, but naming and tab enumeration still failed.
• A new task was tested following the official troubleshooting guidance and reproduced successful discovery followed by failed enumeration.
• Cookies, site permissions and safety settings were not changed; no alternate automation transport was used.

Related Report and Request

Related: https://github.com/openai/codex/issues/43787 reports similar behavior on an earlier app version with Chrome. This report adds a reproduction on app version 26.903.61454 with Edge and a fresh-task test. A shared root cause has not been established.

Please advise on a supported diagnostic or recovery for the tab-enumeration/control channel, without disabling safety checks, changing origin permissions or manually modifying the native host.

This report excludes account identifiers, email addresses, financial data, local paths, extension-instance identifiers, task IDs, credentials, screenshots, conversation transcripts and raw logs.

Start with the documented Computer Use entry points used in the report: run standalone cua.listBrowsers(), select the returned Edge browser with cua.getBrowser(), then run cua.listTabs({browser: observedEdgeId, emit:false}) in a fresh task. Compare discovery and tab-enumeration behavior across the existing and fresh tasks, and identify the failing control channel. Done means enumeration succeeds or returns an actionable error naming the failing component.