Long running tasks - claude was much more stable than gpt agents #482
Description
What version of Code is running?
0.6.17
Which model were you using?
claude models
What platform is your computer?
Debian 12 on WSL2 under windows 11
What steps can reproduce the bug?
I've had a lot of issues with longer running (3H+) auto drive sessions using the OpenAI models where the system would get extremely slow, like took minutes to get a key stroke in and basically the session became lost.
Last night I had a hunch that the slow return times from the models (5x-10x slower to answer compared to Claude models) might be part of the issue. So I told it to only use Claude models for this very large, long running task.
I just checked in on it and everything is humming along 10h in. So I think that might be a clue about where whatever the issue is, is originating.
What is the expected behavior?
No response
What do you see instead?
No response
Additional information
No response