i’ve lost maybe 40 requests between yesterday and today just for stuff like this. This is especially happening on 3.7-thinking, but normal 3.7 also aborts sometimes on its own. It would be less aggravating if it counted as a server or cursor-side error but it counts it as a user abort so I’ve been using multiple reuqests per request just because of this, pushing me toward the 500 limit.
Reproduction details: mostly on tool calls. Thinking model in particular will try to edit a file and the api stream call will be aborted mid tool call and pretend I aborted it myself, even if I’m on another window at the time.
Another issue with this is that unlike server errors, this one is not allowing the Resume button! This means that these calls have to be made from scratch or negotiated again with the agent, which loses even more time. It also sometimes charges each tool use as one full fast call, so by implementing maybe 2 featurs today I got from 400 to 490 requests between fails and half-successes.
These things need to be fixed before release man, gonna have to switch on pay per message because you even disabled 3.7 for slow users, so this update gets me out of fast requests, and then removes the ability to use the model lol.
sample req id: 7e675461-7b7b-49fa-a4ac-88c9ce7d4434
