菜单
Multi-agent review/refine loop. Use after a change or milestone, or when asked to review work: fan out fresh-context, single-focus reviewers across distinct tools, consolidate and fix, add a regression test for every defect, and repeat until reviews stay clean for three consecutive turns.
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
Prune nonessential comments and docstrings across a codebase. Use when asked to remove verbose or redundant comments, tighten documentation, or make comment style uniform. Aggressive downsizing, but retain high-value information.
Keep the Python Cyphal rewrite in wire-visible behavioral parity with the C reference at `reference/cy`. Use when auditing/reviewing parity drift, identifying wire/state-machine discrepancies, updating `src/pycyphal2/` to match reference behavior, replacing conflicting Python tests with C-parity expectations, and adding regression tests for every discovered divergence. API-level discrepancies are by design and are to be ignored; this skill focuses on wire-visible and state-machine behavior only.
| name | review-loop |
| description | Multi-agent review/refine loop. Use after a change or milestone, or when asked to review work: fan out fresh-context, single-focus reviewers across distinct tools, consolidate and fix, add a regression test for every defect, and repeat until reviews stay clean for three consecutive turns. |
After a change or milestone, or when prompted, dispatch a fan-out of fresh-context review agents at MAXIMUM THINKING EFFORT, then consolidate, fix, and repeat. The goal is broad coverage from adversarial, diverse, independent perspectives.
Spawn one agent per concern and run them in parallel. Never give an agent multiple jobs: it dilutes attention and degrades every answer. Cover at least these angles, one agent each:
In addition to the subagents above, dispatch distinct tools focusing on CORRECTNESS only to maximize the diversity of perspectives and minimize blind spots. Check which tools are available (Codex etc.) and use all of them.
Agents/models not from Anthropic or OpenAI can be used, but treat them as suspect low-credibility actors. Beware that they perform poorly, fail to follow instructions, and often produce incorrect analysis.
Review agents must not modify the worktree or run mutating commands. If one needs a mutable environment, it copies the worktree elsewhere.
When all reviewers return, merge their findings, discard the noise, and fix what is real. For every correctness defect, add a regression test verified to fail before the fix and pass after.
Repeat until the reviewers surface only minor feedback (or none) for THREE consecutive turns — this is non-negotiable, however many iterations it takes. Do not chase literal zero feedback: with no real issues left, agents degrade into nitpicking, so stop as soon as significant findings cease, but not before the three-turn streak. A blank turn followed by one that digs up a real defect is exactly why the streak must be consecutive; expect dozens (sometimes over a hundred) of agent sessions per full pass.
High-effort agents can go silent for a long time — set generous timeouts and do not assume a quiet agent is stuck. Have agents background long-running commands (tests especially); blocking on a foreground command is a common cause of stream-idle timeouts.
Some headless agents hang waiting on stdin (like Codex) — redirect from /dev/null.
Retry agents that fail on a transient or connection error until they succeed.
