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。
直接命令不会经过审查 Prompt;运行前请先检查来源。
npx skills add https://github.com/benjamcalvin/bootstraps --skill implement-address命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地?可下载 SkillsMP 当前能够提供的文件。
基于 SOC 职业分类
正在显示 SKILL.md
| name | implement-address |
| description | Address filtered review findings for implement workflow (runs as subagent) |
| license | MIT |
| metadata | {"version":"2.0.1","tags":["implement","address","subagent"],"author":"benjamcalvin"} |
Suite capability: focused-only. Run focused tests, lint, builds, and acceptance commands only. Do not execute or consume the target repository's authoritative verification command or ordered command plan. Final verification owns that evidence.
Parse the PR number, round identifier, and findings-file path from the invocation input:
$ARGUMENTS
If the current client leaves $ARGUMENTS literal, use the delegation prompt instead.
At runtime, fetch the PR metadata and comments, then read the supplied findings file before editing.
You are the addresser for the /implement workflow. You fix issues identified by the review, run tests, and push fixes. The filtered findings above contain only findings the referee accepted — address these and only these.
Guard: If the findings file is missing, unreadable, or contains no findings, stop immediately and report the issue to the orchestrator. Do not proceed with an empty or absent findings list.
Use the current client's task or plan tracker when available.
Read the filtered findings in the Context section above. For each finding:
For each finding, take one of these actions:
Apply — The finding is correct. Make the change.
Partially apply — The core insight is right but the suggested fix isn't optimal.
Reject with justification — The finding is incorrect or doesn't apply after deeper investigation.
Escalate — You're unsure whether the finding is valid.
Preserve test evidence when compressing. If a finding asks you to tighten, dedupe, or compress tests, do not drop the assertions that actually bind the behavior — rollback seams, exhaustive comparators, boundary checks, and regression coverage. Compressing away binding assertions re-opens the exact gaps the review caught and drives another round. Prefer removing redundant/duplicate assertions over removing the strongest one.
After addressing all findings:
fix: address review round <round> — <description>git push
Post your summary to the PR:
gh pr comment <pr-number> --body "<summary>"
Return a summary table:
| # | Finding | Action | Details |
|---|---|---|---|
| 1 | Applied / Partially applied / Rejected / Escalated |
Tests: — Commits: