菜单
Multi-agent PR review that spawns parallel domain-specific subagents (python-expert, rust-expert, dbt-expert, dotnet-expert, kedro-expert, d2-tala-expert, ui-designer) based on changed file types, posts inline review comments on GitHub, GitLab, and Azure DevOps.
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
SQL engineering standards for writing correct, safe, cross-dialect SQL across PostgreSQL, MySQL, SQL Server, SQLite, DuckDB, and Oracle. Use when writing SQL queries, exploring schemas, running migrations, optimizing queries, transpiling between dialects, or connecting to databases via native CLIs.
MongoDB engineering standards for querying collections, building aggregation pipelines, designing document schemas, managing indexes, and administering deployments via mongosh. Use when writing MongoDB queries, exploring collections, designing schemas, optimizing performance, or managing replica sets and sharding.
Engineering planning standards for creating design docs, DRI working docs, ADRs, and weekly updates. Use when planning projects, writing design docs, creating RFCs, tracking risks, writing status updates, or reviewing engineering plans. Covers structure, operational tracking, context engineering, and scaling from startup to enterprise.
dbt engineering standards for building clean, modular, well-tested data transformations. Use when writing dbt models, configuring sources, designing project structure, writing tests, managing materializations, or preparing for dbt Fusion. Covers dbt Core 1.8–1.11, Fusion v2.0, SQL style, testing, governance, and deployment.
Langfuse observability platform standards for querying traces, managing prompts, debugging LLM applications, and accessing Langfuse data programmatically. Use when interacting with Langfuse, querying traces or sessions, managing prompts, instrumenting applications, or looking up Langfuse documentation. Covers CLI-based API access (via npx) and documentation retrieval.
.NET/C# engineering standards for writing clean, scalable, production-ready code. Use when writing C# code, implementing features, designing APIs, or reviewing code quality. Covers modern .NET 10+, C# 14, clean architecture, CQRS, EF Core, testing, and project configuration.
| name | pr-review |
| description | Multi-agent PR review that spawns parallel domain-specific subagents (python-expert, rust-expert, dbt-expert, dotnet-expert, kedro-expert, d2-tala-expert, ui-designer) based on changed file types, posts inline review comments on GitHub, GitLab, and Azure DevOps. |
| disable-model-invocation | true |
| argument-hint | [PR-number] |
You are orchestrating a multi-agent code review. You MUST spawn parallel domain-specific subagents to review the PR, then collect their findings and post inline review comments.
git remote get-url origin 2>/dev/nullgh pr view $ARGUMENTS --json number,title,baseRefName,headRefName,headRefOid,url 2>/dev/null || glab mr view $ARGUMENTS --output json 2>/dev/null || echo "FETCH_FAILED: Manually detect platform and fetch PR metadata"gh pr diff $ARGUMENTS --name-only 2>/dev/null || glab mr diff $ARGUMENTS --name-only 2>/dev/null || echo "FETCH_FAILED: Manually fetch changed file list"If any pre-fetch shows FETCH_FAILED, detect the platform from the remote URL and fetch manually:
gh pr view <number> --json number,title,baseRefName,headRefName,headRefOid,urlglab mr view <number> --output jsonaz repos pr show --id <number># GitHub
gh pr diff <number>
# GitLab
glab mr diff <number>
# Azure DevOps
git diff origin/<target-branch>...HEAD
You need the full diff to extract relevant sections for each subagent.
Map each changed file to the appropriate expert using the routing table in routing.md.
CRITICAL: You MUST spawn ALL matched expert subagents in a SINGLE message with multiple Agent tool calls. Use run_in_background: true for each one. This is the core value of this skill — parallel multi-agent review, exactly like /simplify spawns 3 agents in parallel.
For each matched expert, use this prompt template:
Review the following files from PR #<number> for code quality, security,
performance, and adherence to best practices.
PR title: <title>
PR branch: <head> -> <base>
Changed files in your domain:
- path/to/file1.ext (modified)
- path/to/file2.ext (new file)
Full diff for these files:
<paste ONLY the diff sections for files in this expert's domain>
Return your findings as a structured list. For EACH issue use this EXACT format:
---
FILE: path/to/file.ext
LINE: <line number in the NEW file>
SEVERITY: critical | warning | suggestion
FINDING: Brief description of the issue
RECOMMENDATION: How to fix it
---
If no issues found, state "No issues found" explicitly.
Do NOT modify any files — this is a read-only review.
Rules:
subagent_type for each expert (e.g., python-expert, rust-expert)For files that don't match any expert agent, review them yourself using these checklists:
Security: No hardcoded secrets, proper input validation, no injection vectors Code Quality: Clear naming, no duplication, proper error handling, single responsibility Performance: No N+1 queries, appropriate data structures, no resource leaks Testing: New code has tests, edge cases covered
Use the same structured findings format (FILE/LINE/SEVERITY/FINDING/RECOMMENDATION).
Wait for ALL background subagents to complete. Then collect and classify all findings:
| Severity | Meaning | Example |
|---|---|---|
| critical | Security vulnerability, data loss risk, broken logic | SQL injection, hardcoded secrets, null dereference |
| warning | Code smell, performance issue, missing validation | N+1 queries, no error handling, duplicated logic |
| suggestion | Style, readability, minor optimization | Naming, documentation, simplification |
Determine the review verdict:
critical findings -> REQUEST_CHANGESwarning + suggestion -> COMMENTAPPROVEBefore posting anything, show the user a complete review summary:
## Code Review Summary
**PR**: #<number> — <title>
**Reviewed by**: pr-review (orchestrated multi-agent review)
**Experts consulted**: <list of experts spawned>
### Findings
| Severity | Count |
|----------|-------|
| Critical | N |
| Warning | N |
| Suggestion | N |
### By Domain
| Expert | Files Reviewed | Findings |
|--------|---------------|----------|
| python-expert | N files | N warnings, N suggestions |
| self-review | N files | N suggestions |
### Verdict: **<verdict>**
IMPORTANT: Ask the user for confirmation before posting. Especially for REQUEST_CHANGES.
See platforms.md for the platform-specific API calls to post inline review comments.
Always:
REQUEST_CHANGES without explicit user approval