win4r

Loopeng — Testing skill for Claude Code

Testing community

Agent-agnostic Loop Engineering runner: portable loop.yaml spec + deterministic verification gate + guardrails + auditable JSONL ledger.

How to install Loopeng

This entry records only its repository, not the path inside it, so there is no exact command to give. Open win4r/loopeng and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Loopeng does

Agent-agnostic Loop Engineering runner: portable loop.yaml spec + deterministic verification gate + guardrails + auditable JSONL ledger. Drives Claude Code / Codex / any shell-callable agent.

Alternatives in Testing

  • Test Runner — Run Python tests with pytest, unittest, or other testing frameworks 23.4k ★
  • Testing Strategy — This project is small, runs in a terminal, and is mostly deterministic 10.9k ★
  • Superpowers Evals — Behavioral eval lab (Quorum) for the superpowers project that drives real coding-agent CLIs (Claude, Codex, Ge 107 ★

README

loopeng

**English** · [简体中文](README.zh-CN.md)

An **agent-agnostic Loop Engineering runner**. It drives any shell-callable coding agent (Claude Code, Codex, or a plain script) through a bounded `act → verify → feed-back` loop, defined by a portable `loop.yaml` spec.

It's a small, well-tested core that combines ideas proven across recent loop-engineering tools: a **portable loop spec** (ralphify), **agent-neutral adapters** (loom), **guardrails + auditable stop conditions** (openloop), a **deterministic verification gate** (the load-bearing half of Spotify's "Honk" findings), and **git-friendly state** via an append-only ledger.

loopeng at a glance: loop.yaml spec → agent → blast-radius write-set gate → verifier (exit 0 = pass) → success, or feedback → retry within limits; every iteration appends to a resumable ledger with a live heartbeat for status/resume; platform layers (skills, worktree isolation, orchestration, watch/schedule triggers, MCP server and adapter plugins) compose on the same core; and the held-out feedback dogfood where the agent runs a public verifier while loopeng additionally runs a held-out test outside the agent's reach, so the held-out requirement reaches the agent only through feedback

The core act → verify → feed-back loop, the platform layers, and the held-out feedback dogfood. Animated SVG — the static diagram renders fully on its own.

Install

loopeng is **not published to PyPI** — install from the GitHub release or from source (Python ≥ 3.9; no runtime dependencies beyond the standard library + PyYAML):

# 1) From the latest GitHub release (no clone needed):
pip install https://github.com/win4r/loopeng/releases/download/v0.3.3/loopeng-0.3.3-py3-none-any.whl
#    …or download the wheel / .tar.gz from https://github.com/win4r/loopeng/releases and
#    install the local file:   pip install ./lo