AGENT SKILL

report-review-run

Use when the PR review harness starts or ends a phase or loop, records a pushed fix, or posts the end-of-run GitHub report comment, or when the user says /report_review_run.

License 许可
MIT
所属仓库
sazlin/loadout
最近更新
2026-11-14
Stars
★ 0
Forks
1

01 前言 Front Matter

name
report-review-run
metadata
{ "loadout.sha": "49eb9bd", "loadout.source": "skills/report-review-run/SKILL.md", "loadout.managed": "true" }
description
Use when the PR review harness starts or ends a phase or loop, records a pushed fix, or posts the end-of-run GitHub report comment, or when the user says /report_review_run.

02 资源文件 1

03 SKILL.md 正文

Report review run

Record wall-clock timings in REVIEW_RUN.json and render the final PR comment from that log. Do not invent durations. Do not hand-write the report body.

When to use

  • review_orchestrator starts or ends a panel loop, resolve wave, verify loop, risk classification, or merge wrap-up
  • After issue_resolver commits land on PR head
  • At the end of a harness run (ok, blocked, or aborted)
  • The user asks /report_review_run

Do not use to review, fix, or merge. Do not commit REVIEW_RUN.json.

Steps

  1. Script: python3 .claude/skills/report-review-run/scripts/review_run_report.py. If .claude/ is missing, use skills/report-review-run/scripts/review_run_report.py. Run --help for flags. Default file is REVIEW_RUN.json in the project root.
  2. Start a step before the work: begin --section "Panel Review" --label "loop 1". Sections: Panel Review, Resolve Issues, Verifiers, Risk Classification, Merge. Labels: loop N, wave N, or a TASK-00N id. begin closes any still-open step.
  3. End a step when that work returns: end.
  4. Record a pushed fix after a clean cherry-pick/push: change --sha <short> --task TASK-00N --summary "..." --path <file>:+N,-M (repeat --path). Take N/M from git show --numstat --format= <sha>. Source files only. Never log TASKS_TO_RESOLVE-*.md or REVIEW_HISTORY.md.
  5. Stage cells before render, matching the Decision/Aborted table: stage --panel "✅|2 loops" --resolve "✅|3 tasks" --verifiers "✅|4/4" --risk "🟢|\low`" --merge "✅|done"`.
  6. Dashboard: dashboard https://cursor.com/agents/<id> once run-info returns. Never invent an id.
  7. Render and post at the end of the run, including abort: render --out /tmp/review-run-report.md then gh pr comment <n> --body-file /tmp/review-run-report.md. Never --edit-last. Never paste a hand-written stand-in.
  8. Delete REVIEW_RUN.json after the comment is posted (or after max comment retries). Missing file is a no-op.

Guardrails

  • Never invent timestamps or durations
  • Never skip begin/end around a phase or loop
  • Never post a Run report that did not come from render
  • Never commit REVIEW_RUN.json
  • The mermaid fence in render stdout is already closed. Do not wrap it in another mermaid fence

04 同仓库更多 Skills