Since Nov 8, my two self-hosted workers (agent worker start, CLI version 2026.11.05-86047df) on the same Linux machine keep losing their bridge connection. They stay running, but the bridge goes silent for anywhere from a few minutes to hours, then the worker logs “No frames from bridge; treating stream as half-open and reconnecting.” Earlier attempts also showed HTTP/2 PING timeouts and rpc_unavailable.
Both workers drop at the exact same moments. Today that happened at 5:25, 6:02, 9:54 and 11:27 AM EST, which points to something shared rather than a problem in one repository. Reconnecting the worker didn’t help, the status page shows no incident, and turning off traffic routing through my home PC changed nothing. Because of this, no long task can finish.
I opened support ticket T-J51275 but haven’t gotten a substantive answer yet. Has anyone else seen this, or can staff check the bridge side?
Hi @Will4, from what we can see on our side, the drops are coming from the network path between your machine and us, not from the bridge.
Each time it happens, all of your workers’ connections are cut in the same second even though they’re attached to different bridge servers. Those connections also reach us through a VPN or relay provider’s network rather than directly from your ISP, so something on that machine or its network is still routing the traffic after you turned off the home PC routing.
Could you try this with one worker?
On the Linux machine, run env | grep -i proxy and check for any VPN, relay or proxy client that’s running.
Stop one worker, turn off whatever you find (or unset the proxy variables in that shell), and start it again with agent worker start.
Leave it running for an hour or two and let us know whether it still logs “No frames from bridge”.
Thanks Colin. Following your suggestion, I ran a test worker on a separate DigitalOcean droplet (Ubuntu, no proxy, direct internet, same Cursor account, CLI 2026.11.05-86047df), outside the bot computer’s network relay.
Result: it never connected at all. From 12:31 PM to 1:22 PM EST it logged about 800 `ConnectError: [invalid_argument]` errors with failureReason “rpc_other”, retrying continuously. It had zero half-open, PING timeout or rpc_unavailable errors. The CLI login was valid the whole time (`agent status` OK), and memory was fine.
In the same window, the two workers on the original machine kept connecting and then dropping (half-open/closed at 12:19, 12:25, 12:39, 12:58 and 1:11 PM EST).
So with the relay out of the picture, the worker still fails, just with a different error that comes back from your side. For what it’s worth, all of this started right after the recent update. Could you look into the invalid_argument rejection for this account/worker? I can leave the test worker running if that helps.
Hi @Will4, the droplet test turned up something different from the drops. Our side can’t read the git remote of the folder that test worker serves, so it turns the connection away as an invalid repository URL on every attempt. It isn’t a problem with your account.
This usually means origin points at a local path or at the other machine instead of the hosted repo. To fix it on the droplet:
Open a shell in the folder the test worker serves.
Run git remote get-url origin.
If that prints a local path or anything other than your hosted repo URL (for example https://github.com/<owner>/<repo>.git or [email protected]:<owner>/<repo>.git), run git remote set-url origin <that URL>.
Stop the test worker and start it again with agent worker start.