菜单
Hours-long (3-12h) multi-agent code-architecture optimization — cross-examines antipatterns into consensus, then applies only the highest-ROI refactors. Not for routine refactors.
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。
launch the advisor loop
Long-running (6-30h) multi-agent documentation-architecture optimization — verifies every claim in every non-executable text (markdown, spec systems, comments) against the code, cross-examines discrepancies into consensus, then applies only the highest-ROI fixes. Code is never modified — defects are reported, not fixed. Not for touching up a single file.
Long-running (6-30h) multi-agent integrity-boundary optimization — hunts every reachable state the boundary (types, invariants, error definitions, tests) fails to contain, then absorbs only the highest-ROI set into it — code and docs both, pinned by tests. Not for fixing a known bug.
Answer the owner's questions about a running or finished loop from its records — read-only, in a separate session
Implement the planned change task by task, then spawn the independent verify stage until it passes
Archive the completed change — task gate, delta sync into main specs, revalidate
| name | code |
| description | Long-running (6-30h) multi-agent code-architecture optimization — cross-examines antipatterns into consensus, then applies only the highest-ROI refactors. Not for routine refactors. |
| argument-hint | [focus area] |
| effort | max |
아래를 실행해, 출력된 Workflow(...) 호출을 그대로 실행하라.
"${CLAUDE_PLUGIN_ROOT}/bin/refine-hook" bootstrap code "$ARGUMENTS"
중단 시: 산출물은 agoraPath에 남고 agent는 자기 기록에서 이어간다. resumeFromRunId는 끝난 agent의 결과를 cache에서 그대로 돌려준다. args가 바뀌면 둘 다 잃는다.