菜单
Drive one pi-mono parity gap end to end in the pipy repo (select gap, plan, different-family review, implement, docs, review until CLEAN, commit). Applies only when the working directory is the pipy repo (~/projects/pipy).
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Run a bounded, fail-closed Claude Code review gate over the current git worktree with a configurable Claude model and strict read-only isolation. Use for fresh-context different-family reviews before commit, including requests for Opus, Fable, Sonnet, or an explicitly selected Claude model; drive fix and re-review rounds until the configured reviewer returns CLEAN.
Use before committing an implementation slice when a configurable different-family reviewer should inspect the slice, or when coordinating up to three fix/re-review cycles with Claude, Codex, or Pi. Claude reviews delegate to the supervised claude-review-loop harness.
Use when you want a fresh-context code-review gate before committing — runs Pi only with the approved OpenAI Codex GPT-5.6 Sol model at high reasoning over the current git diff in a bounded, observable loop, and only proceeds when Pi returns CLEAN. Never falls back to Claude/Anthropic, local models, OpenRouter, or another provider. Drives review → fix → re-review up to a round cap. Triggers: "have pi review this", "pi review before commit", "run the pi review loop".
Use before committing an implementation slice when a configurable different-family reviewer should inspect the slice, or when coordinating up to three fix/re-review cycles with Claude, Codex, or Pi. Claude reviews delegate to the supervised claude-review-loop harness.
Use when the user asks for a compact goal prompt, goal condition, create_goal objective, continuation objective, or second-agent goal statement, especially when the destination has a strict character limit such as 4000 characters. Generate a concise, context-specific goal condition from the current worktree, session context, and relevant planning docs without producing a full implementation or review handoff.
Legacy compatibility alias for claude-review-loop. Use only when an existing workflow explicitly invokes opus-review-loop; delegate the review to the canonical configurable Claude gate with the Opus model.
| name | pipy-parity-loop |
| description | Drive one pi-mono parity gap end to end in the pipy repo (select gap, plan, different-family review, implement, docs, review until CLEAN, commit). Applies only when the working directory is the pipy repo (~/projects/pipy). |
This skill applies only when your working directory is the pipy repo
(~/projects/pipy). If you are anywhere else, do not use it.
Follow the canonical workflow in docs/parity-loop/skill-body.md — resolve that
relative path against the pipy repo root and read it now. Drive exactly one
parity gap end to end; run the phases inline.
Honor the body's hard rules:
claude-review-loop (Opus by default, or the
explicitly configured Claude model).main); never weaken or delete tests to pass a gate.The body's Phase 0 also drains the lesson backlog and Phase 9 captures lessons via
scripts/parity_lessons.py; to apply accumulated lessons use the
parity-improve skill.