[macOS][Desktop] Task started from Codex is persisted as ChatGPT and loses Codex/browser skills

What version of the Codex App are you using (From “About Codex” dialog)?

Version 26.825.32147 (build 7303)

What subscription do you have?

Signed-in ChatGPT paid account (the exact tier is not exposed in the task diagnostics)

What platform is your computer?

Darwin 25.5.0 arm64 arm (macOS 26.5.2, build 25F84)

What issue are you seeing?

A new conversation that the user started from the Codex surface of the unified desktop app was persisted as a ChatGPT-backed conversation instead of a Codex task.

The app's read-only task listing categorized the affected conversation as ChatGPT rather than Codex and showed no associated local working directory.

Because it was routed as ChatGPT, it did not receive the Codex skill/tool catalog. In particular, browser:control-in-app-browser and its callable browser-control tools were absent even though the signed-in in-app browser was open. The assistant therefore repeatedly said it could see only the ambient URL and could not click or type.

A control task created in the same desktop app was categorized as Codex and had a generated local working directory.

That control task received browser:control-in-app-browser, claimed the same open browser tab, inspected it, typed into the AWS console search, and navigated successfully. This isolates the failure to the conversation/session routing and capability provisioning, not the browser tab, authentication, or the Browser plugin itself.

The mode mismatch is not visible enough to the user. From the user's perspective this was a Codex task, but its backing kind and capabilities were ChatGPT. Typing / also showed almost none of the expected Codex capabilities, which made the repeated browser-control refusal appear arbitrary.

What steps can reproduce the bug?

1. In the unified ChatGPT desktop app on macOS, select the Codex surface.
2. Start a new task/conversation from that surface.
3. Open the in-app browser and ask the task to inspect or click the current tab.
4. Observe that the task says browser control is unavailable and only the ambient browser URL is present.
5. Compare the affected conversation with a fresh Codex task using the app's task listing.
6. Observe that the affected conversation is categorized as ChatGPT with no working directory, despite having been started as a Codex task.
7. Start a fresh control task that is correctly reported as kind: codex and repeat the browser request.
8. Observe that the Browser skill and callable browser-control tools are present and the request succeeds.

This was observed on August 28, 2026. It has not yet been reproduced on every new task, so the trigger may involve stale mode state or an incorrect route selected during task creation.

What is the expected behavior?

A task started from the Codex surface should be persisted as a Codex task and receive the configured Codex skills and tools. If the app is intentionally creating a ChatGPT Chat/Work conversation, that mode should be explicit before submission and the conversation should not visually present itself as a Codex task.

Additional information

Related issues were checked before filing:

• #33910 reports a Codex CLI session appearing under ChatGPT Work. It is the same general attribution boundary, but its trigger is CLI-to-desktop synchronization rather than desktop task creation.
• #39489 reports a new chat from a Local Project becoming a global chat and losing project skills. This report is not tied to a Local Project; the confirmed mismatch is ChatGPT versus Codex task classification.
• #33079 requests a persistent Work/Codex badge. Better labeling would reduce confusion, but it would not correct the wrong backing kind or missing capability provisioning reported here.

No session identifiers, AWS URLs, account details, screenshots, logs, or transcript content are included. A private session reference can be supplied through the in-app feedback channel if maintainers need it.

Start at Codex task creation from the unified desktop app on macOS and reproduce the mismatch by comparing the affected conversation with a correctly categorized control task in the task listing. Trace how the selected Codex surface is persisted and how skills, tools, and the local working directory are provisioned; done means new Codex tasks are categorized as Codex and receive the expected browser capability.