Autonomous improvement loop for the Orca CLI. Inspired by karpathy/autoresearch. One run produces one branch with all accepted changes; create one PR from it when done.

The main process is the orchestrator: it runs CI, finds issues (or improvement ideas when CI passes), delegates to a worker to fix one thing, then validates and keeps or discards. Workers can run either in-process (direct Claude) or as orca workers in separate worktrees.

# Run the loop (Ctrl+C to stop) — direct Claude in-process
uv run autoimprove/loop.py

# Use orca: one worker per iteration (worktree, merge, validate)
uv run autoimprove/loop.py --use-orca

# Orca + sprint-team coder role in the task
uv run autoimprove/loop.py --use-orca --sprint-team

# Run 5 iterations then stop
uv run autoimprove/loop.py --max-iters 5

# Dry run — show what would happen (limited to 3 iters if no --max-iters)
uv run autoimprove/loop.py --dry-run

Requirements: uv, Claude Code CLI (claude in PATH). For --use-orca: orca on PATH and daemon (auto-started on first spawn).

• The script always runs on an autoimprove branch. If the current branch is not autoimprove/*, it creates autoimprove/<date> (e.g. autoimprove/mar19) and checks it out.
• Without --use-orca: each iteration runs claude in the project dir; changes are committed there and CI is re-run; on failure the last commit is reset.
• With --use-orca: each iteration spawns one orca worker (e.g. ai-iter-1) with --base-branch set to the current autoimprove branch. The worker runs in a worktree (.worktrees/ai-iter-N). When the worker is done, the loop merges the worktree’s HEAD into the main branch, runs CI, keeps or discards the merge, then kills the worker and removes the worktree.

LOOP:
  1. Run CI (fmt, clippy, test)
  2. If failures → extract error output; if all pass → build “find an improvement” prompt (+ optional clippy hints)
  3. Invoke worker (direct claude OR orca spawn with task)
  4. Worker fixes ONE thing and commits (in worktree if orca)
  5. If orca: merge worktree commit into current branch
  6. Re-run CI to validate
  7. Keep if CI passes, discard (git reset) if not
  8. Log to results.tsv (status: keep | discard | skip | no_improvement)
  9. GOTO 1

All iterations contribute to the same branch (e.g. autoimprove/mar19). Workers may push and create PRs from their worktrees if they want; the orchestrator merges each accepted change into its branch. That branch is the single clear destination: one PR with all features for the human. After the loop finishes (or you stop it), push and open that one PR:

git push -u origin autoimprove/mar19
gh pr create --base main --fill

When the loop exits (normally or Ctrl+C), all worktrees under .worktrees/ are removed so only the main working tree remains.

With --sprint-team, the task passed to the worker is prefixed with the coder role from .agents/skills/sprint-team/references/coder.md (with {{project_dir}}, {{base_branch}}, and the current issues substituted). Use this when you want the worker to follow the sprint coder workflow (TDD, coverage, etc.).

Edit program.md to change what the worker focuses on. The loop itself (loop.py) is just plumbing.