Let him cook🚀
one claude session is great. four repos, multiple sessions, and parallel work is where things get messy.
as a session gets longer, it starts filling up with code, logs, retries, and context.
two sessions working on the same repo step on each other, so most work ends up running sequentially. a fresh session retries the same local run failures without knowing what already failed.
i shipped shiploop, a harness for claude code where an advisor breaks work into tasks and sends them to the right-sized workers.
you just give it a multi-repo project.
- n parallel agents in git worktrees
- failure traces are shared so agents don't repeat each other's mistakes
- related tasks share exploration instead of rediscovering the same things
- context persists across sessions, failures and learnings carry forward
shiploop. ship more, in one conversation.