AgentSkillsCN

fpf-review

任务结束前的质量门。在结束任何会话之前，必须予以调用。核实任务完成情况，确保关键主张有充分证据支撑，并确认FPF相关成果（异常/决策/术语表/工作日志）均已更新。

SKILL.md
--- frontmatter
name: fpf-review
description: End-of-task quality gate. MUST be invoked before ending any session. Verify completion, ensure evidence exists for key claims, and confirm FPF artifacts (anomalies/decisions/glossary/worklog) are updated.

Goal

Before stopping, run a lightweight quality gate so we don't ship hand-wavy work.

Procedure

  1. •Restate the objective
  • •What was the user's request?
  • •What "done" means.
  1. •Check FPF artifacts
  • •Anomaly record exists if this was investigative/debugging work.
  • •DRR exists if we made a non-trivial decision.
  • •Evidence records exist for any "this works" claims.
  • •Glossary updated if new terms were introduced or naming drift was resolved.
  • •Worklog updated for this session.
  1. •Run the minimum viable verification
  • •Run the smallest test suite or static checks that meaningfully reduce risk.
  • •If running is not possible, state what would need to be run and why.
  1. •Report remaining risks
  • •Any untested assumptions?
  • •Any incomplete migrations?
  • •Any evidence that might be stale soon (validity windows)?

Output

  • •Update .fpf/worklog/session-${CLAUDE_SESSION_ID}.md with a final summary.
  • •Provide the user with a concise checklist-style report:
    • •Completed items
    • •Evidence links
    • •Remaining risks / next steps

Session sentinel cleanup

After the review is complete, MUST delete the session sentinel:

code
Delete file: .fpf/.session-active

This signals that the session ended properly through the review gate. The SessionStart hook also cleans stale sentinels, but explicit cleanup is preferred.

Quality bar

  • •Conservative: if evidence is missing, do not claim high confidence.
  • •Explicit: list what would change your mind.