Labsco
microsoft logo

review-pr

✓ Official★ 52

by microsoft · part of microsoft/agent365-devtools

Generate structured PR review comments using Claude Code agents and post them to GitHub. No API key required - uses Claude Code's existing authentication.

🔥🔥🔥✓ VerifiedAccount requiredNeeds API keys
#automation#cloud#devops#git#security
🧩 One of 2 skills in the microsoft/agent365-devtools package — works on its own, and pairs well with its siblings.

Generate structured PR review comments using Claude Code agents and post them to GitHub. No API key required - uses Claude Code's existing authentication.

Inspect the full instructions your agent will receiveExpand

This is the exact playbook injected into your agent when the skill activates — shown here so you can audit it before installing. You don't need to read it to use the skill.

by microsoft

Generate structured PR review comments using Claude Code agents and post them to GitHub. No API key required - uses Claude Code's existing authentication. npx skills add https://github.com/microsoft/agent365-devtools --skill review-pr Download ZIPGitHub52

PR Review Skill

Generate and post AI-powered PR review comments to GitHub following engineering best practices.

What this skill does

Step 1: Generate (/review-pr <number>)

  • Fetches PR details from GitHub using the gh CLI

  • Performs architectural review (NEW!): Questions design decisions, checks for scope creep, validates use cases

  • Analyzes changes for security, testing, design patterns, and code quality issues

  • Differentiates contexts: CLI code vs GitHub Actions code (different standards)

  • Creates actionable feedback: Specific refactoring suggestions based on file names and patterns

  • Generates structured review comments in an editable YAML file

  • Shows preview of all generated comments

Step 2: Post (/review-pr <number> --post)

  • Reads the YAML file you reviewed/edited

  • Posts to GitHub: Submits all enabled comments to the PR

  • Automatic fallback: If GitHub API posting fails (e.g., Enterprise Managed User restrictions), automatically generates a markdown file with formatted comments for manual copy/paste

Engineering Review Principles

This skill enforces the following principles:

Architectural Review (NEW!)

  • Design Decision Validation: Questions "why" before reviewing "how"

  • Scope Creep Detection: Flags expansions beyond Agent365 deployment/management

  • Use Case Validation: Requires concrete scenarios for new features

  • Overlap Detection: Identifies duplication with existing tools (Azure CLI, Portal)

  • YAGNI Enforcement: Questions features without documented need

Architecture & Patterns

  • .NET architect patterns: Reviews follow .NET best practices

  • Azure CLI alignment: Ensures consistency with az cli patterns and conventions

  • Cross-platform compatibility: Validates Windows, Linux, and macOS compatibility (for CLI code)

Design Patterns

  • KISS (Keep It Simple, Stupid): Prefers simple, straightforward solutions

  • DRY (Don't Repeat Yourself): Identifies code duplication

  • SOLID principles: Especially Single Responsibility Principle

  • YAGNI (You Aren't Gonna Need It): Avoids over-engineering

  • One class per file: Enforces clean code organization

Code Quality

  • No large files: Flags files over 500 additions

  • Function reuse: Encourages reusing functions across commands

  • No special characters: Avoids emojis in logs/output (Windows compatibility)

  • Self-documenting code: Prefers clear code over excessive comments

  • Crisp comments (pr-code-reviewer #30): Flags added comments that run past 1-2 lines, restate the code, or narrate design history — a comment says why in one line; long-form reasoning belongs in the commit/PR.

  • Release-note-ready CHANGELOG (pr-code-reviewer #31): Flags CHANGELOG.md entries that name internals, explain mechanism, or run multiple sentences — each entry is one crisp consumer-facing sentence (it ships verbatim to nuget.org release notes).

  • Minimal changes: Makes only necessary changes to solve the problem

Testing Standards

  • Framework: xUnit, FluentAssertions, NSubstitute for .NET; pytest/unittest for Python

  • Quality over quantity: Focus on critical paths and edge cases

  • CLI reliability: CLI code without tests is BLOCKING

  • GitHub Actions tests: Strongly recommended (HIGH severity) but not blocking

  • Mock external dependencies: Proper mocking patterns

Security

  • No hardcoded secrets: Use environment variables or Azure Key Vault

  • Credential management: Follow az cli patterns for CLI code; use GitHub Secrets for Actions

Context Awareness

The skill differentiates between:

  • CLI code (strict requirements): Cross-platform, reliable, must have tests

  • GitHub Actions code (GitHub-specific): Linux-only is acceptable, tests strongly recommended

Review Comments Output

Generated comments are saved to:

Copy & paste — that's it
C:\Users\ \AppData\Local\Temp\pr-reviews\pr- -review.yaml

You can edit this file to:

  • Disable comments by setting enabled: false

  • Modify comment text

  • Adjust severity levels (blocking, high, medium, low, info)

  • Add or remove comments

Implementation

The skill uses Claude Code directly for semantic code analysis (inspired by Agent365-dotnet). No separate API key required!

Generate mode (default):

  • Claude Code reads .claude/agents/pr-code-reviewer.md for review process guidelines. Read the working-tree (PR) version of this file and of .github/copilot-instructions.md and CLAUDE.md — not the base-branch copy. When the PR under review adds or changes a review rule (as PR #461 did with rules #30/#31), the new rule must be applied to that same PR in the same run; reading the base copy would skip it.

  • Claude Code reads .github/copilot-instructions.md for coding standards

  • Claude Code fetches PR details: gh pr view <number> --json ...

  • Claude Code analyzes actual code changes: gh pr diff <number>

  • Claude Code performs semantic analysis using its own capabilities

  • Claude Code identifies specific issues with line numbers and code references

  • Claude Code writes YAML file to C:\Users\<username>\AppData\Local\Temp\pr-reviews\pr-<number>-review.yaml

Post mode (with --post flag):

  • Python script reads the YAML file

  • Python script posts comments to GitHub using gh pr comment

  • If posting fails (API permissions), automatically generates markdown file for manual copy/paste

Key Advantages:

  • ✅ No ANTHROPIC_API_KEY required - uses Claude Code's existing authentication

  • ✅ Better semantic analysis - Claude Code has full context and conversation history

  • ✅ Simpler Python script - only handles posting logic (~240 lines vs ~1500 lines)

  • ✅ Easier to maintain and debug

Workflow

Generate review: /review-pr 180

  • Fetches PR details from GitHub

  • Analyzes code and generates review comments

  • Saves to YAML file (shows path in output)

Review and edit: Open the YAML file

  • Review all generated comments

  • Edit comment text if needed

  • Disable comments by setting enabled: false

  • Add your own comments if desired

Post to GitHub: /review-pr 180 --post

  • Reads the YAML file

  • Posts all enabled comments to the PR

  • If API posting fails, automatically generates a markdown file for manual copy/paste

See Also