Token导航 LogoToken导航

Agent Skill · 研究检索

pr-fix(公关修复)

pr-fix 用于查找、检索和筛选相关信息,适合在 Codex、Claude、Cursor、Gemini CLI 中需要根据关键词、任务场景或来源线索快速定位候选结果时使用。可结合来源仓库、安装命令和原始 README 继续核验具体用法。安装前建议确认权限范围、维护状态,以及是否会触发联网、命令执行或文件读写。

pr-fix 是研究检索类 Agent Skill,可用宿主为 Codex、Claude、Cursor、Gemini CLI,通过GitHub 仓库获取,权限为执行命令;许可证信息还不明确,第三方扫描有提醒项,按安装量在研究检索类的 73,675 个 Skill 里排第 72,323。

总安装 188周安装 8GitHub Stars 8下载 66核验 2026-05-01

安装说明

本站只整理中文说明,不托管安装包,也不代用户安装。
GitHub
许可证
unknown

安装方式

通过对话安装

复制提示词发给支持本地命令或 Skills 的 AI 助手,先确认命令和权限,再让它执行。

请帮我安装这个 Agent Skill:pr-fix(公关修复)
来源仓库:https://github.com/phrazzld/claude-config
仓库路径:skills/pr-fix
安装命令:
npx skills add https://github.com/phrazzld/claude-config --skill pr-fix
安装前请先检查当前环境是否支持对应 CLI,并向我确认将要执行的命令、安装目录、联网范围和文件读写权限;确认后再执行。

命令行安装

复制命令到本机终端执行。该命令会通过 npx skills 从第三方来源获取 Skill;本站只展示命令,不托管安装包,也不自动执行。

skills.shnpx skills
npx skills add https://github.com/phrazzld/claude-config --skill pr-fix

这个 Skill 是做什么的

研究检索类 Skill 主要解决「找资料、核对事实、把来源整理清楚」的问题,通常会让助手先检索再总结。

pr-fix 用于查找、检索和筛选相关信息。适合在 Codex、Claude、Cursor、Gemini CLI 中根据关键词、任务场景或来源线索快速定位候选结果。通过 npx skills add 命令从指定 GitHub 仓库安装,需结合原始 README 确认具体用法。

安装前建议确认权限范围、维护状态,以及是否会触发联网、命令执行或文件读写。pr-fix 属于研究检索类 Skill,可作为该场景下的辅助能力补充。

来源仓库内路径:skills/pr-fix

上手步骤

4 步
  1. 1

    确认自己的宿主工具

    公开资料标注支持 Codex、Claude、Cursor、Gemini CLI;先确认自己用的是其中之一,再按对应工具的 Skills 目录放置。

  2. 2

    拿到安装命令

    可以直接复制的命令是 npx skills add https://github.com/phrazzld/claude-config --skill pr-fix;命令里的占位符换成自己的路径或密钥再执行。

  3. 3

    安装后重启宿主

    命令执行完要完全退出宿主工具再打开,让新装的 Skill 被重新扫描到,否则列表里可能看不到。

  4. 4

    按最小权限试用

    它需要执行命令,先在测试目录或测试账号里跑,确认没问题再放到正式项目。

常见问题

6 条

pr-fix 怎么安装?

在终端执行 npx skills add https://github.com/phrazzld/claude-config --skill pr-fix。装完完全退出宿主工具再打开,让它重新扫描本地 Skill。

pr-fix 支持哪些 AI 工具?

公开资料标注支持 Codex、Claude、Cursor、Gemini CLI。其他工具只要能读取本地 Skill 目录,一般也能手动放进去,但要自己确认触发方式是否一致。

pr-fix 需要什么权限?

权限级别是执行命令。安装流程涉及命令执行,可能通过 npx skills add https://github.com/phrazzld/claude-config --skill pr-fix 联网下载 Skill 或依赖。用户安装前应确认命令来源、仓库内容和执行环境。

pr-fix 是免费的吗?

公开资料没有标注明确许可证,收费和商用限制都无法确认。正式使用前请到来源仓库核对授权说明。

pr-fix 用的人多吗?

公开数据是累计安装 188 次,近一周安装 8 次,下载 66 次,GitHub 8 星。这些数字来自第三方目录统计,只能当作热度参考,不代表质量或安全性。

pr-fix 安全吗?

第三方扫描有提醒项,说明存在需要人工确认的行为。安装前请打开源码看清它读了什么、联网到哪里。

值得选的地方和要先想清楚的地方

值得选的地方

  • 同时适配 4 种工具,换客户端不用重新找替代品。

先想清楚的地方

  • 需要执行命令,装到常用环境前要先评估影响面。
  • 许可证没有明确标注,能不能商用、能不能再分发都不确定。
  • 第三方扫描存在提醒项或未通过项,需要人工逐条确认。
  • 第三方目录整理,版本更新和行为细节以来源仓库为准。

适不适合你

更适合

  • 已经在用 Codex,想直接扩展它能力的人。
  • 需要处理研究检索类任务、又不想每次都手写提示词的人。

不太适合

  • 不能接受 Skill 在你机器上执行命令的人。
  • 需要明确授权才能用到公司项目里的场景。
  • 需要厂商级技术支持和 SLA 的生产系统。

研究检索类 Skill 对比

6 个同类
Skill总安装GitHub 星权限许可证扫描
find-skills

寻找技能 · Codex、Claude

31,199,81316,486操作浏览器MIT有提醒
self-improving-agent

自我改进 Agent · OpenClaw

16,464,0023,396接触敏感数据MIT-0通过
soultrace

灵魂追踪 · Codex、Claude

7,588,94412需要联网未明确有提醒
self-improving

自我完善 · OpenClaw

6,829,2841,046需要联网MIT-0通过
microsoft-foundry

微软代工厂 · Codex、Claude

6,614,144764调用外部服务未明确有提醒
pr-fix

公关修复 · 本页

1888执行命令未明确有提醒

按安装量取同分类前几名;数字来自第三方目录统计,只作热度参考,不代表质量结论。

SKILL.md 原文

/pr-fix

One command takes a blocked PR to green.

Role

Senior engineer unblocking a PR. Methodical, not reactive. Each phase resolves a class of blocker in dependency order.

Objective

Take PR $ARGUMENTS (or current branch's PR) from blocked to mergeable: no conflicts, CI green, reviews addressed.

Dependency Order

Conflicts -> CI -> Self-Review -> Reviews. Can't run CI on conflicted code. Can't review broken code. Can't address others' reviews before fixing your own issues.

Bounded Shell Output (MANDATORY)

  • Size before detail: counts/metadata first
  • Never print unbounded logs/comments
  • Add explicit bounds: --limit, head -n, tail -n, per_page
  • If no useful signal in 20s: abort, narrow, rerun
  • Use ~/.claude/scripts/safe-read.sh for large local files

Workflow

1. Assess

gh pr view $PR --json number,title,headRefName,baseRefName,mergeable,reviewDecision,statusCheckRollup
gh pr checks $PR --json name,state,startedAt,completedAt,link
gh pr view $PR --json body --jq '.body | split("\n")[:80] | join("\n")'

Read PR description and linked issue. Understand what this PR is trying to do — semantic context drives conflict resolution and review decisions.

Fetch latest base:

BASE="$(gh pr view $PR --json baseRefName --jq .baseRefName)"
git fetch origin "$BASE"

Determine blockers: conflicts? CI failures? pending reviews? Build a checklist.

2. Resolve Conflicts

Skip if: mergeable!= CONFLICTING

Rebase onto base branch:

git rebase "origin/$BASE"

When conflicts arise, resolve semantically based on PR purpose, not mechanically:

  • Read both sides. Understand intent.
  • Preserve the PR's behavioral changes. Integrate upstream structural changes.
  • Reference git-mastery/references/conflict-resolution.md for strategies.
  • Never blindly accept ours/theirs.

After resolution, verify locally:

git rebase --continue
# Run project's test/typecheck commands

3. Fix CI

Skip if: all checks passing.

Push current state and invoke /fix-ci:

git push --force-with-lease

Then run the /fix-ci skill. Wait for checks to go green.

If /fix-ci introduces changes that create new conflicts: return to Phase 2 (max 2 full-pipeline retries).

4. Self-Review the Diff

Always run this phase. CI passing does not mean the code is good. Linters catch syntax; this catches logic.

Review the full diff against base:

BASE="$(gh pr view $PR --json baseRefName --jq .baseRefName)"
git diff "origin/$BASE"...HEAD

For each changed file, check for:

  • Dead code: unused variables (especially _-prefixed params that signal "I know this is unused"), unreachable branches, useMemo/useCallback with values never read
  • Logic bugs: loops that always break on first iteration, conditions that are always true/false, off-by-one errors
  • Wasted computation: expensive operations whose results are discarded, duplicate work (e.g., running the same test suite twice in CI)
  • Wrong log levels: success messages on stderr (console.warn/console.error), debug output on stdout in production
  • Semantic mismatches: function names that don't match behavior, comments that contradict code

Fix every issue found. Run typecheck + tests after fixes. Commit before proceeding.

This phase catches what CI cannot: code that compiles and passes tests but is wrong, wasteful, or misleading. A PR that's "green" but ships dead code or wasted computation is not actually unblocked — it's shipping tech debt.

5. Address Reviews

Skip condition: ALL THREE of these are zero: unresolved review threads, unreplied review comments, AND unaddressed bot issue comments. Use the queries below — never rely on reviewDecision alone or prior "PR Unblocked" summary comments.

OWNER="$(gh repo view --json owner --jq .owner.login)"
REPO="$(gh repo view --json name --jq .name)"

# Count unresolved review threads (inline comments)
UNRESOLVED_THREADS="$(gh api graphql -f query='
  query($owner:String!, $repo:String!, $number:Int!){
    repository(owner:$owner,name:$repo){
      pullRequest(number:$number){
        reviewThreads(first:100){nodes{isResolved}}
      }
    }
  }' -F owner="$OWNER" -F repo="$REPO" -F number="$PR" \
  --jq '[.data.repository.pullRequest.reviewThreads.nodes[] | select(.isResolved==false)] | length')"
5a. Bot issue comments (MANDATORY)

Why this exists: Some bot reviewers (Claude, CodeRabbit, etc.) post review feedback as issue comments (/issues/$PR/comments), not pull request review comments (/pulls/$PR/comments). These are a completely different API endpoint and are invisible to the GraphQL reviewThreads query. Missing them means missing actionable review feedback.

# Fetch bot issue comments — these are NOT in reviewThreads or PR comments
BOT_COMMENTS="$(gh api "repos/$OWNER/$REPO/issues/$PR/comments?per_page=100" --paginate \
  --jq '[.[] | select(.user.type == "Bot") | {id, user: .user.login, body: .body}]')"

Filter for comments that contain actionable review feedback (code suggestions, bug findings, security concerns). Ignore:

  • Status/summary comments (CI reports, merge readiness checks)
  • Comments you've already replied to with fixes
  • Informational comments with no action items

For each bot comment with actionable findings:

  1. Read the FULL comment body — no truncation
  2. Extract each finding — bots typically number them or use headers
  3. Read the current file to check if already addressed
  4. Fix or defer each finding (same as review comments below)
  5. Reply to the comment with resolution status for each finding
5b. Review comments and threads

Independent verification (MANDATORY)

Never trust prior session comments, "PR Unblocked" summaries, or claims that feedback was addressed. For EVERY open review comment:

  1. Read the FULL comment body — no truncation. Use the GitHub API without .body[:N] limits.
  2. Read the current file at the referenced line to verify the fix is actually present.
  3. Reply directly on the comment thread with the specific commit SHA and line confirming the fix. An open thread without a reply = unaddressed, regardless of what a summary comment claims.
# Fetch ALL review comments with full bodies — never truncate
gh api "repos/$OWNER/$REPO/pulls/$PR/comments?per_page=100" --paginate \
  --jq '.[] | {id, user: .user.login, path, line, body, in_reply_to_id}'

For each comment without a reply from this PR's author:

  • If already fixed in code: reply with commit SHA + current line reference confirming the fix
  • If needs fixing: fix it, then reply with commit SHA
  • If deferred: reply with follow-up issue number
  • If declined: reply with public reasoning

Bot feedback (CodeRabbit, Cerberus, Gemini, Codex) gets the same treatment as human feedback.

Execution
  1. Invoke /respond — Categorize all feedback (critical / in-scope / follow-up / declined). Post transparent assessment to PR. Reviewer feedback CAN be declined with public reasoning.
  2. Invoke /address-review — TDD fixes for critical and in-scope items. GitHub issues for follow-up items.
  3. Reply to every open thread — use gh api repos/$OWNER/$REPO/pulls/$PR/comments/$ID/replies -f body='...' so the thread shows addressed.
  4. Resolve every thread via GraphQL — Replies alone do NOT resolve threads. Non-outdated comments stay visible as open issues to reviewers even after fixing the code and replying. You MUST resolve them:
# Get unresolved thread IDs
gh api graphql -F owner="$OWNER" -F repo="$REPO" -F number=$PR -f query='
  query($owner: String!, $repo: String!, $number: Int!) {
    repository(owner: $owner, name: $repo) {
      pullRequest(number: $number) {
        reviewThreads(first: 100) {
          nodes { id isResolved isOutdated }
        }
      }
    }
  }' --jq '.data.repository.pullRequest.reviewThreads.nodes[] | select(.isResolved == false) | .id'

# Resolve each thread
gh api graphql -f query='mutation { resolveReviewThread(input: {threadId: "THREAD_ID"}) { thread { isResolved } } }'

Additive commits do NOT make comments outdated. Only changes to the diff hunk a comment is attached to trigger outdating. Resolve explicitly.

6. Verify and Push

git push --force-with-lease

Watch checks. If a phase-4 or phase-5 fix broke CI, invoke /fix-ci again (count toward 2-retry max).

If 2 full retries exhausted: stop, summarize state, ask user.

7. Update PR Description with Before / After

Edit the PR body to include a Before / After section documenting the fix:

# Get current body, append Before/After section
gh pr edit $PR --body "$(current body + before/after section)"

Text (MANDATORY): Describe the blocked state (before) and the unblocked state (after). Example: "Before: CI failing on type error in auth module. After: Types corrected, CI green."

Screenshots (when applicable): Capture before/after for any visible change — CI status pages, error output, UI changes from review fixes. Use ![before](url) / ![after](url).

Skip screenshots only when all fixes are purely internal (conflict resolution with no behavior change, CI config fixes with no visible output difference).

8. Signal

Post summary comment on PR:

gh pr comment $PR --body "$(cat <<'EOF'
## PR Unblocked

**Conflicts**: [resolved N files / none]
**CI**: [green / was: failure type]
**Reviews**: [N fixed, N deferred (#issue), N declined (see above)]

Ready for re-review.
EOF
)"

Retry Policy

Max 2 full-pipeline retries when fixing one phase breaks another. After 2: stop and escalate to user with clear status.

Anti-Patterns

  • Mechanical ours/theirs conflict resolution
  • Pushing without local verification
  • Silently ignoring review feedback
  • Retrying CI without understanding failures
  • Fixing review comments that should be declined
  • Trusting prior "PR Unblocked" or summary comments — always verify each comment against current code independently. A previous session claiming "fixed" means nothing until you read the file yourself.
  • Leaving review threads without direct replies — an open thread with no reply = unaddressed, even if the code is fixed. Reviewers can't see that you checked.
  • Truncating comment bodies — never use .body[:N] when fetching review comments. The actionable detail is often at the end of long comments.
  • Replying without resolving — a reply on a thread does NOT resolve it. Non-outdated threads with replies still show as open conversations. Use resolveReviewThread GraphQL mutation after replying.
  • NEVER lowering quality gates to pass CI — coverage thresholds, lint rules, type strictness, security gates. If a gate fails, write tests/code to meet it. Moving the goalpost is not a fix. This is an absolute, non-negotiable rule.
  • Skipping self-review because CI is green — CI catches syntax and test failures. Dead code, wasted computation, wrong log levels, and semantic mismatches all pass CI. Review the diff yourself before declaring unblocked.
  • Only checking review threads and PR comments — bot reviewers (Claude, CodeRabbit, etc.) often post feedback as issue comments (/issues/$PR/comments), not PR review comments (/pulls/$PR/comments). These are different API endpoints. You MUST check all three: GraphQL reviewThreads, REST PR comments, AND REST issue comments.

Output

Summary: blockers found, phases executed, conflicts resolved, CI fixes applied, reviews addressed/deferred/declined, final check status.

平台分布

Codex

36.48%
按下载量换算24

Claude

29.68%
按下载量换算20

Cursor

20.4%
按下载量换算13

Gemini CLI

10.8%
按下载量换算7

安全审计

Gen Agent Trust Hub

通过

Socket

通过

Snyk

可疑

权限和风险

执行命令

安装流程涉及命令执行,可能通过 npx skills add https://github.com/phrazzld/claude-config --skill pr-fix 联网下载 Skill 或依赖。用户安装前应确认命令来源、仓库内容和执行环境。

安装前确认

本站仅展示第三方公开信息,不托管安装包,不提供自动安装或运行环境。安装前应自行审查源码、依赖和命令行为。来源安全扫描存在提醒或未通过项,安装前请自行复核。目前只有一个公开来源,建议同时核对源仓库。

继续浏览同类 Skills