菜单
Use for the cleanup session leader. Defines the issue surfacing, fix cycle, review round, clean-check voting, and cleanup report phases unique to cleanup sessions.
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
| name | cleanup-session-protocol |
| description | Use for the cleanup session leader. Defines the issue surfacing, fix cycle, review round, clean-check voting, and cleanup report phases unique to cleanup sessions. |
Protocol for the active phases of a code cleanup session. The leader is the only editor — team members are advisors who read, critique, and surface issues.
Open the floor: post a short summary of the cleanup scope (3-5 sentences), then ask specific members to review specific areas using @agent-name.
Team members respond with issues — style violations, dead code, unclear naming, missing error handling, unnecessary complexity. Each issue should be a short, actionable message with file path and line number.
Your role during surfacing:
Work through the surfaced issues. For each fix:
Batch small related fixes together — don't trigger a build for every single-line change. Group fixes by file or by theme, then build once.
After completing all fixes from the current surfacing round:
git fetch origin docker/{project-id}/{your-agent-name}
git diff HEAD FETCH_HEAD
If new issues are surfaced: return to Fix Cycle.
If no new issues: initiate a clean-check vote.
Post: "No new issues raised. Vote: is the code clean? Reply Clean or Not clean (with specifics)."
Maximum 5 full loops (surfacing -> fix -> review). If the code is not clean after 5 loops, record remaining issues and conclude.
Write a cleanup report to plans/cleanup-report-{brief-name}.md on disk. Include:
Use for any container agent that posts progress to the coordination server's message board. Defines the curl format, channels, message types, payload conventions, posting responsibilities, and verbosity levels.
Use when a review agent must produce structured output. Defines the BLOCKING/NOTE/Summary/Verdict template, the two-tier confidence rubric, and the JSON shadow block consumed by POST /tasks/:id/reviews. Compose with review-process and a domain skill.
Use when an agent evaluates the output of other agents or decides whether work meets the bar. Enforces rigorous quality standards — reject shortcuts, demand evidence, no rubber-stamping.
Use when an agent is the FSM arbitrator on a durable task. Defines the two triggers, the three ruling types, the POST contract, the addendum-file rules for contradiction rulings, and the 409 handling. Project-agnostic.
Tells container agents where the Unreal Engine source tree is mounted inside the container, so they can grep and read engine headers to verify API signatures and class hierarchies.
Use when an agent must review code changes. Defines the universal review sequence — identify files, read context, check against domain criteria, score, filter. Compose with a domain skill and review-output-schema.