Leerie
Leerie is an autonomous task driver for Claude Code. One prompt. Finished, committed, validated code. No steering mid-run, no polishing when it's done.
Most tools that call themselves autonomous still require you: to confirm a direction, catch a hallucination, or clean up the result before it's usable. Leerie doesn't. It classifies the task, decomposes it, implements each piece in parallel isolated worktrees, validates the integrated result, and merges — beginning to end, unattended.
It runs entirely on the Claude Code CLI and your existing subscription — no Anthropic API key, no per-call billing. If you have Claude Code installed and logged in, you have everything it needs.
Why it actually finishes without you:
Most AI "orchestrators" let the model pilot: the model decides what to do next, declares when it's done, and judges whether it succeeded. That's where drift, hallucinated completion, and silent failures come from — and why you end up steering.
Leerie inverts the relationship. The model writes code. The program runs everything else. Phases, wave scheduling, retries, caps, merge logic, and success-criteria enforcement are ordinary Python — real loops and conditionals that cannot drift.
- No silent failures. Every worker output is JSON-schema-validated before the orchestrator acts on it. A worker cannot, by malformed output or confident hallucination, cause the system to do something undefined.
- Confidence is the only hard gate. The implementer self-gates on evidence-anchored confidence in
root_cause and solution (≥9 on both, see DESIGN.md §8) — falsifiers tested, contradictions reconciled, gaps named with concrete artifacts. A worker that cannot justify the score exits blocked with the gap analysis. Everything else — tests passing, lint clean, build green, per-criterion satisfaction — is best-effort: surfaced as advisory warnings on the subtask result, never escalated to failed or blocked by the orchestrator. The criteria file is the implementer's working note, not a gate.
- Workers must justify confidence with evidence, not feelings. Before writing code, an implementer clears domain-specific evidence gates — file-and-line citations, reproductions, falsification attempts. A self-reported score without hard artifacts doesn't clear the bar.
- Parallel work that's actually safe. Each implementer gets an isolated git worktree. Parallel writes never collide. Conflicts surface one wave at a time, close to the work that caused them.
- Resumable by design. A reboot, network blip, budget cap, the Claude Code subscription rate-limit, Ctrl-C, or an external kill (SIGTERM from CI / systemd / a closed terminal) all lose nothing — the run branch is the durable record, worktrees are torn down, and
--resume picks up from the last completed wave. When the subscription rate-limit hits and the reset time is unambiguously parseable, leerie even auto-resumes after the reset window without manual intervention. The explicit "throw this away" gesture is scripts/cleanup.sh --run-id <id> --branches, not Ctrl-C.
- Parallel-safe across runs. Multiple
./leerie invocations in the same repository each get a unique run_id (a derived branch + state directory). Their branches, worktrees, and per-run state directories never collide. Launch a fix and a feature in parallel without coordination.
How it works
The orchestrator is a Python program — not an in-session agent. It shells out
to claude -p (headless mode) for each unit of LLM work. Each call is a
separate process, so there is no subagent nesting anywhere. Control flow lives
in real Python: for loops, if statements, counters. It cannot drift.
