0.7.1

On 31 July the service your AI client connects to stopped responding for several hours. Connectors looked offline, and adding a new one failed with a sign-in error that pointed at the wrong thing. This was ours, not your AI client's, and there was nothing to fix at your end. Your connections and the credentials behind them were never at risk, and no data was exposed.

What happened

When an AI client stops using a connection, we tidy it up about an hour later. If the client had already gone away, and after an idle hour it usually has, that tidy-up hit an error nothing was listening for, and the whole service stopped instead of just that one connection. It restarted, met the same problem, and eventually stopped restarting.

What changed

Cleanup now expects a client that has already left, because that is the normal case rather than the exception. Separately, a background error can no longer take the service down with it: it gets recorded and everything else keeps running.

We also fixed how we tell requests apart. Traffic reaches us through a proxy, and we were reading the proxy's address rather than the sender's, so our anti-abuse limits were shared across everyone instead of applied per sender. That could turn away a legitimate attempt to add a connector, which is the error some of you saw.