[Windows][App][Chrome] Same conversation enters ~22 Hz active-writer retry loop across desktop and Chrome side panel

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

OpenAI.Codex 26.810.4967.0

Chrome integration runtime: 26.810.41047

Google Chrome: 150.0.7871.125

What subscription do you have?

ChatGPT Pro (5x, US$100/month tier)

What platform is your computer?

Microsoft Windows NT 10.0.26200.0 x64 (Windows 11 Pro, build 26200)

What issue are you seeing?

Opening the same Codex/ChatGPT conversation in the Windows desktop app and the ChatGPT Chrome extension side panel causes a rapid cross-client ownership/retry loop. The contention is bidirectional: it occurs whether the desktop app opens the conversation first or the Chrome side panel opens it first.

The first client does not need to be actively viewing the conversation, generating a response, or controlling a browser. Once that client has opened the conversation, its retained ownership/activity state can continue to trigger contention when the other client later opens the same conversation. In repeated tests, this stale competing state persisted until the first client/app was restarted, even when the conversation was no longer being actively browsed there.

Both surfaces become visually unstable. The desktop repeatedly shows and hides this message:

已在另一个应用中打开，请先在那边关闭会话，才能在这里继续

English meaning: “This conversation is open in another app. Close it there before continuing here.”

The warning does not settle into a stable blocked or read-only state. Restarting the client that previously opened the conversation clears its retained competing state, but users should not need to restart an application merely to release a conversation that is no longer active there.

A synchronized 15-second video/log capture confirms that this is a high-frequency state-machine loop rather than only a rendering artifact. During the exact video interval, the desktop emitted:

• 338 thread_stream_role_changed events;
• 338 maybe_resume_started events;
• 338 browser-session route captures;
• 337 failed thread/resume requests with error code -32600;
• 337 request failures reporting thread [redacted] already has an active writer.

That is approximately 22.5 ownership/resume cycles per second for the full recording.

Representative redacted sequence:

response_routed ... errorCode=-32600 method=thread/resume
Request failed ... error={"code":-32600,"message":"thread [redacted] already has an active writer"}
maybe_resume_started ... previousResumeState=needs_resume previousStreamRole=null
Failed to resume conversation ... already has an active writer
thread_stream_role_changed ... previousRole=null role=owner
IAB_LIFECYCLE webview captured browser use session route

A 20.82-second process sample while the contention remained active showed:

• desktop codex.exe app-server: 19.91 CPU-seconds, approximately one logical core;
• desktop renderer: 20.83 CPU-seconds, approximately one logical core;
• Chrome extension codex.exe app-server: 0.02 CPU-seconds;
• Chrome extension-host.exe: effectively 0 CPU-seconds.

This points to an active retry/IPC loop in the desktop client after the writer conflict, rather than CPU activity in the Chrome extension host.

What steps can reproduce the bug?

Desktop first

1. Open an existing conversation in the Windows desktop app.
2. Stop actively using that conversation. It does not need to be generating, controlling a browser, or even remain the currently viewed conversation. Do not restart the desktop app.
3. Open the same conversation in the ChatGPT Chrome extension side panel.
4. Observe both clients continuously flashing and the desktop ownership warning repeatedly appearing and disappearing.

Chrome first

1. Open an existing conversation in the ChatGPT Chrome extension side panel.
2. Stop actively using that conversation, but leave Chrome/the extension runtime running without restarting it.
3. Open the same conversation in the Windows desktop app.
4. Observe the same repeated contention in the opposite opening order.

The issue has been reproduced multiple times in both directions. The trigger is retained per-conversation client state, not simultaneous active browsing: opening the conversation previously and keeping that client runtime alive is sufficient.

Trigger matrix

| Condition | Result | Evidence scope |
| --- | --- | --- |
| Desktop opened conversation first; Chrome opens it later | Reproduces | Repeated observation; one synchronized video/log capture |
| Chrome opened conversation first; desktop opens it later | Reproduces | Repeated observation |
| First client is not generating or controlling a browser | Reproduces | Repeated observation |
| Conversation was opened earlier but is not currently being viewed; client has not restarted | Reproduces | Repeated observation |
| Client retaining the competing state is restarted | Contention clears | Repeated observation |

What is the expected behavior?

Only one client should hold the active-writer role. When another client opens the same conversation, it should receive one rejection and remain in a stable non-owner/read-only or blocked state until an ownership-change event occurs.

The client should not immediately retry thread/resume, reassert role=owner, recapture the browser route, or repeatedly mount/unmount the warning UI.

When a client is no longer actively viewing or using a conversation, its writer/activity lease should be released without requiring the entire client to restart.

Additional information

The observable loop appears to be:

1. Desktop attempts thread/resume.
2. App-server rejects it because another active writer exists.
3. Desktop immediately returns to needs_resume.
4. Desktop reports itself as role=owner and recaptures the browser route.
5. Desktop retries approximately 22 times per second.

Suggested engineering checks:

• transition an already has an active writer response into one stable observer/blocked state;
• wait for an ownership-change event or use bounded backoff instead of immediate retries;
• make browser route capture/release idempotent when writer acquisition fails;
• verify that hiding, minimizing, navigating away from, or closing a conversation reliably updates its writer/activity lease;
• add a two-client integration test covering both opening orders and an idle/minimized first client.

Related issues found before filing:

• #37403 reports already has an active writer during Desktop ↔ Remote Control/CLI handoff on macOS.
• #37856 reports stale ownership between VS Code clients.

This report is distinct because it covers the Windows desktop app ↔ ChatGPT Chrome side panel, is bidirectionally reproducible, and includes synchronized evidence of a sustained ~22.5 Hz retry/render loop.

Attachments:

• public-safe 1080p/30 FPS video, 15 seconds; Chrome conversation body blurred;
• redacted log events covering the exact video interval;
• capture summary with counts, process comparison, timestamps, and hashes.

Evidence scope: the attached synchronized video captures one concrete contention direction and proves the sustained retry/render loop for that run. The reverse opening order and the persistence after merely opening a conversation earlier were each reproduced repeatedly but are described as reproduction observations rather than being claimed as simultaneously shown by this single video.

Privacy: conversation IDs, request IDs, local paths, account information, and Chrome conversation content are redacted from public attachments. Full-resolution evidence is retained privately if maintainers need it through a private support channel.