Windows: Codex desktop loses VS Code connection after first message, main "New chat" button stops working, app-server exits after plugin 403
What version of the Codex App are you using (From “About Codex” dialog)?
26.406.31014
What subscription do you have?
Plus
What platform is your computer?
Microsoft Windows NT 10.0.26200.0 x64
What issue are you seeing?
On Windows, Codex desktop became stuck in a broken state after a VS Code .ipynb / Jupyter-triggered extension host crash.
After that:
- a brand-new empty chat shows the VS Code button in the top-right
- the moment I send the first message, the VS Code button disappears
- the main
New chatbutton does nothing - the
New chatbutton next toThreadsstill works - Codex desktop starts with zero discovered connections
- Codex app-server later exits after plugin bootstrap hits Cloudflare
403 Forbidden - When I close Codex and I try opening it from the "Hidden Icons" it doesn't come up. I have to manually close it and reopen the app.
This survives reboot, Codex reinstall/reset, VS Code extension reinstall, deleting Codex local package data, and testing on a different network/hotspot.
Original trigger
The first trigger appears to be VS Code notebook/Jupyter related.
Steps:
- Open a
.ipynbin VS Code - VS Code starts selecting a kernel
- It appears to find one, then loses it and retries
- VS Code then shows:
Extension host terminated unexpectedly 3 times within the last 5 minutes
A VS Code extension bisect pointed to Jupyter as the trigger for that original crash.
What steps can reproduce the bug?
- Launch Codex desktop
- Open a brand-new empty chat
- Observe the VS Code button is visible in the top-right
- Send the first message
- Observe the VS Code button disappears
- Try the main
New chatbutton - Observe it does nothing
What is the expected behavior?
- Codex desktop should keep the VS Code connection available after the first message
- the main
New chatbutton should create a new chat - a VS Code/Jupyter crash should not leave Codex desktop permanently unable to discover connections
Additional information
Actual behavior
- Codex desktop repeatedly reports zero available connections
- the VS Code button disappears as soon as the first message is sent
- the main
New chatbutton is nonfunctional - the issue persists across reset/reinstall/reboot
- testing on a phone hotspot did not resolve it
Diagnostics
1. Codex desktop boots with zero connections
Desktop logs repeatedly show:
discoveredConnectionCount=0totalConnectionCount=0shared_object_synced connectionCount=0availableConnectionCount=0selectedRemoteHostId=null
This suggests the app is not starting with a healthy editor connection and then losing it. It already believes there are no attachable connections.
2. VS Code Codex bridge appears to be in a bad IPC state
VS Code Codex logs show warnings like:
Received broadcast but no handler is configured method=client-status-changed- repeated
thread-stream-state-changed - repeated
thread-read-state-changed
This looks like an internal Codex bridge / event subscription issue, not just a Jupyter kernel failure.
3. Codex plugin bootstrap is failing with Cloudflare 403
Codex logs show plugin bootstrap requests hitting Cloudflare challenge pages:
https://chatgpt.com/backend-api/plugins/listhttps://chatgpt.com/backend-api/plugins/featured
Later the desktop log shows:
- app-server transport stopping
- worker exit
code=1
A different network/hotspot did not change this.
What I already tried
- rebooted PC
- reinstalled VS Code OpenAI extension
- reinstalled Codex desktop
- used Windows app
Terminate,Repair, andReset - deleted Codex local package data
- disabled suspect VS Code extensions
- used a fresh VS Code profile
- tested on a different network / hotspot
None of these resolved the issue.
The Jupyter/Notebook crash appears to be the original trigger, but the persistent failure afterward looks like a Codex desktop recovery / connection-management / UI-state bug.
6 Comments
Potential duplicates detected. Please review them and close your issue if it is a duplicate.
Powered by Codex Action
Additional note: Right before the bug triggered, I had 80% weekly usage left. When the bug started and I reinstalled the app, my usage suddenly went back up to 100%. Weirdly enough, my usage shows 100% on the website too.
On this exact same version if you go to File -> Setting the "Back to app" button does nothing also.
@DorukG I figured out the usage left. It's not how much you used until now, it's how much you have left. They inverted the number..
Posted this on the wrong issue initially.
This comment was intended for the separate Windows sidebar
New Chat/ hit-testing problem, not the VS Code connection-loss flow described here.The specific symptom I meant to report is:
New chatbutton does not respond to clickCtrl+Nstill creates a new chat successfullyThat points more toward a UI click / hit-testing bug than a broken new-chat action itself.
Reproducing on the latest extension (26.527.31454) — with a root-cause trace pointing at the webview↔host RPC bridge
Confirming this on a clean, current setup, and adding diagnostic detail that may explain the
availableConnectionCount=0/ dead "New chat" state — same "one New Chat path works, the other doesn't" split described here.What I'm seeing
~/.codex/sessionsholds saved rollout files andsession_index.jsonllists them.BODY.clickfires) but sends no backend request (noConversation createdin the extension host log) and throws no app-level exception.Root-cause trace (webview DevTools, "pause on caught exceptions", with VS Code core
workbench.desktop.main.jsignore-listed). During webview mount the host↔webview RPC bridge aborts:no such export ID: 1means the root remote-API proxy reference is gone, so the awaited connection loop (gl) aborts — leaving thread create/list inert while the independent composer survives. This is consistent with thediscoveredConnectionCount=0/availableConnectionCount=0and theReceived broadcast but no handler is configured method=client-status-changed/…=thread-stream-state-changedwarnings already reported here (I see the same ones).Ruled out (to save others time)
chatgpt.com/ces/v1/rgstr(CloudflareCF-Mitigated: challenge) — those are caught by Statsig's own ErrorBoundary; clicking New Chat and waiting >20s (past the 10s flush timeout) still does nothing.~/.codex; new threads still persist to disk.auth_elicitation— that startup-32600was fixed between 26.519 and 26.527.Steps to reproduce
Expected behavior
New Chat in the agent panel should create a new thread (as the sidebar does), and the history list should show existing threads from
~/.codex/sessions.Workaround: use the Codex sidebar ("New Thread in Codex Sidebar") for new chats; the panel composer still works for the current thread.
This looks like the same underlying connection/IPC failure behind the history-loading reports (#18993, #19137). Happy to provide full
app-serverlogs or a complete bridge trace if useful — just let me know what format helps most.