菜单
Run a consensus-loop audit manually — Codex reviews pending trigger_tag items in the watch file. Use when you want to trigger an audit without editing the watch file, re-run a failed audit, or test the audit prompt.
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
Design tasks into tracks with work breakdowns and execution order. Also writes and maintains PRDs (Product Requirements Documents) — when a user requests a feature, analyzes it, adds it to the PRD, then decomposes into work breakdowns. Use for new feature planning, PRD writing, architecture changes, multi-track decomposition, requirements analysis, or adjusting existing execution plans. Trigger when user says things like 'add feature X', 'I need Y', 'plan Z', or describes any product requirement.
Run all done-criteria checks (CQ/T/CC/CL/S/I/FV/CV) and produce a pass/fail verification report. Use after implementing code, before submitting evidence to the quorum audit.
Squash-merge a worktree branch into the target branch with a structured commit message. Use after audit consensus and retrospective completion.
Session orchestrator for the consensus-loop — reads handoff, picks unblocked tasks, distributes to parallel workers, tracks agent assignments, manages correction cycles via SendMessage. Use when starting a work session with pending tasks, distributing implementation work, or reviewing completed worker output.
Extract learnings from audit history and conversation, manage memories, clean up stale entries. Use after completing a track, during retrospective (③ memory step), at end of session, or anytime the user wants memory maintenance. Triggers on 'what did we learn', 'memory cleanup', 'review learnings', 'retrospective', 'update memories', '회고', '메모리 정리'.
Show current consensus-loop status — pending reviews, audit state, retro marker, agent assignments. Use to check what's happening before starting work or after returning from a break.
| name | consensus-loop:audit |
| description | Run a consensus-loop audit manually — Codex reviews pending trigger_tag items in the watch file. Use when you want to trigger an audit without editing the watch file, re-run a failed audit, or test the audit prompt. |
| argument-hint | [--dry-run | --no-resume | --auto-fix | --model <name>] |
Run the audit process manually.
node ${CLAUDE_PLUGIN_ROOT}/audit.mjs {{ arguments }}
Read the respond file (gpt.md) and summarize:
| Flag | Effect |
|---|---|
--dry-run | Print audit prompt without executing |
--no-resume | Fresh session (don't resume previous) |
--auto-fix | Auto-correct via Claude CLI after audit |
--model <name> | Model override (default: gpt-5.4) |
--reset-session | Delete saved session before running |
--watch-file <path> | Override watch file path (worktree support) |