loope CLAUDE.md

CLAUDE.md

Project instructions for the loope issue-processing daemon.

Lessons learned

On failure, continue from existing state — don't delete and restart from zero

When a step fails (a git worktree add, a pipeline run, a resume), prefer recovering and continuing from whatever already exists over deleting it and starting from scratch. Blowing away state on every failure throws out partial progress and, worse, can create infinite retry loops when the failure recurs each cycle.

This is a general working principle, not just a failure-recovery rule: never remove and start from zero — always try to build on what already exists. Reusing existing state saves tokens and effort; recreating it from scratch wastes both. Reach for existing artifacts, partial results, and prior work before regenerating anything.

Concretely:

• If a worktree still exists on the path, reuse it and continue working on it. Only reclaim (delete) a leftover when it is truly unusable — e.g. a bare orphaned branch whose worktree is gone.
• Reclaim/force-delete only as a last resort, gated on the specific condition that makes reuse impossible, not as the default reaction to any error.
• Keep the happy path untouched: recovery logic belongs in the error branch, so a normal run never pays for it.

See Worktree.Create in worker/infra/worktree.go for the pattern: try the operation; on failure, reuse the existing worktree if present, otherwise reclaim the bare leftover and retry once.