freev0.1.0unverified
Agent Loops Starter — one loop, one file, green tests or it keeps going
A single self-terminating fix loop for Claude Code. One skill file, one exercise,
installable in under a minute. This is the free starter for the agent-loops
category; the paid bundle adds orchestrated multi-loop workflows.
Quick start
Claude Code: copy skills/test-loop/ into your project's .claude/skills/,
then run /test-loop in a repo with failing tests.
No Claude Code: open skills/test-loop/SKILL.md and paste its body as your
first message to any coding agent — the loop contract works anywhere.
What you get
| File | Job |
|---|---|
skills/test-loop/SKILL.md | the loop: run tests → fix smallest failure → repeat until green or stuck |
exercise.md | stop-and-attempt checkpoint: write the lint variant yourself before peeking |
The one primitive this teaches
An agent loop is three lines of contract, not a framework:
- A check the agent runs, not asserts — the test command's exit code decides, never the model's opinion.
- The smallest fix per iteration — one failure at a time; big diffs are where loops go off the rails.
- A stuck rule — same failure twice in a row → stop and report, don't thrash.
Everything fancier (multi-agent orchestration, review gates, budget dials) is this
primitive composed. That's the paid bundle.
Proof
Unbenchmarked — v0. First ledger entry lands after running against a seeded
failing-test repo (see engine/benchmark.py).
Changelog
- 0.1.0 (2026-07-03) — initial release, unverified.
One email when a pack ships or gets re-benchmarked on a new model. No noise.