-
Notifications
You must be signed in to change notification settings - Fork 33
Description
The problem
Claude Code sometimes gets stuck in an iteration, running many turns and triggering compaction repeatedly and never arriving at a solution until usage limits are hit. However, this can often be fixed by killing the iteration and restarting the same task with a fresh iteration.
Proposed solutions
In config.yaml, could we add support for either:
- An iteration timeout in minutes, or
- Support for the
--max-turnsflag documented here?
In either case, when the limit is reached, perhaps the agent process could be killed, the Chief iteration count be reduced by 1, and a new agent process could spawned on the same user story.
I think this should be done without git reset to clean up the files from the previous iteration. In my testing thus far, an iteration is already capable of recognizing that the task is partially complete and taking it across the finish line from there.
What do you think?