Token Tunnel detects a sleeping computer in seconds

When a run used a model served from your own computer through Token Tunnel and that computer slept or lost its network, the run could wait 90 seconds to 4 minutes before failing. Fallback then tried every other model on the same unreachable machine before reaching a cloud model. In a typical week, about 96% of device disconnects were only caught by a server ping, mostly computers that fell asleep with the tunnel still marked connected.

How it works

• Faster detection. The desktop app closes its connection before the computer sleeps, which fails open requests in about a second, and a dropped connection is detected within 15 seconds instead of 30 to 45.
• Disconnects aren't model faults. Timers are tuned per provider and can never fire before dead-connection detection, so an unreachable computer is always reported as a disconnect.
• No retrying a dead machine. When the device is unreachable, fallback skips every other model it serves and goes straight to the next option.
• Native endpoints. Claude runs over the Messages API and ChatGPT's Codex models over the Responses API, with tool inputs streamed as they're written. A 32 KB file write used to mean 99 seconds of silence; the longest gap is now 1.6 seconds.
• Thinking isn't a stall. The desktop proxy sends a keepalive while a model thinks quietly, so long reasoning is no longer cut off.

Also shipped

• Workflows pinned to a subscription model are no longer disabled when one account hits a rate limit.