AgentSkillsCN

code-review

向多家 AI 提供商（GitHub Copilot、Anthropic Claude、OpenAI Codex、Google Gemini）提交代码审查请求，并获取修复补丁。当用户说“代码审查”、“审查这段代码”、“获取修复补丁”或需要 AI 生成的统一代码差异以完成修复时，可使用此技能。

SKILL.md
--- frontmatter
name: code-review
description: >
  Submit code review requests to multiple AI providers (GitHub Copilot, Anthropic Claude,
  OpenAI Codex, Google Gemini) and get patches back. Use when user says "code review",
  "review this code", "get a patch for", or needs AI-generated unified diffs for code fixes.
allowed-tools: Bash, Read
triggers:
  - code review
  - review this code
  - get a patch
  - copilot review
  - generate diff
  - review request
  - full review
metadata:
  short-description: Multi-provider AI code review CLI

Code Review Skill

Submit structured code review requests to multiple AI providers and get unified diffs back.

Supported Providers

ProviderCLIDefault ModelNotes
githubcopilotgpt-5GitHub Copilot (default)
anthropicclaudesonnetClaude CLI
openaicodexgpt-5.2-codexOpenAI Codex (high reasoning default)
googlegeminigemini-2.5-flashGemini CLI

Prerequisites

bash
# Check provider CLI availability
python .agents/skills/code-review/code_review.py check
python .agents/skills/code-review/code_review.py check --provider anthropic
python .agents/skills/code-review/code_review.py check --provider openai
python .agents/skills/code-review/code_review.py check --provider google

Quick Start

bash
# Single-step review (default: github/copilot)
python .agents/skills/code-review/code_review.py review --file request.md

# Use different provider
python .agents/skills/code-review/code_review.py review --file request.md --provider anthropic

# OpenAI with high reasoning
python .agents/skills/code-review/code_review.py review --file request.md --provider openai --reasoning high

# Include uncommitted local files via workspace
python .agents/skills/code-review/code_review.py review --file request.md --workspace ./src --workspace ./tests

# Full 3-step pipeline (generate -> judge -> finalize)
python .agents/skills/code-review/code_review.py review-full --file request.md

Commands

check

Verify provider CLI and authentication.

OptionShortDescription
--provider-PProvider to check (default: github)

review

Submit a single code review request.

OptionShortDescription
--file-fMarkdown request file (required)
--provider-PProvider: github, anthropic, openai, google
--model-mModel (provider-specific, uses default if not set)
--add-dir-dAdd directory for file access
--workspace-wCopy local paths to temp workspace (for uncommitted files)
--reasoning-RReasoning effort: low, medium, high (openai only)
--rawOutput raw response without JSON
--extract-diffExtract only the diff block

review-full

Run iterative code review pipeline.

Pipeline (per round):

  1. •Generate - Initial review with diff and clarifying questions
  2. •Judge - Reviews output, answers questions, provides feedback
  3. •Finalize - Regenerates diff incorporating feedback
OptionShortDescription
--file-fMarkdown request file (required)
--provider-PProvider: github, anthropic, openai, google
--model-mModel for all steps
--add-dir-dAdd directory for file access
--workspace-wCopy local paths to temp workspace
--reasoning-RReasoning effort: low, medium, high (openai only)
--rounds-rIteration rounds (default: 2)
--save-intermediate-sSave step outputs
--output-dir-oDirectory for output files
bash
# Full pipeline with intermediate files saved
python .agents/skills/code-review/code_review.py review-full \
  --file request.md \
  --save-intermediate \
  --output-dir ./reviews
# Creates: round1_step1.md, round1_step2.md, round1_final.md, round1.patch

build

Build a request markdown file from options.

OptionShortDescription
--title-tTitle describing the fix (required)
--repo-rRepository owner/repo (required)
--branch-bBranch name (required)
--path-pPaths of interest (repeatable)
--summary-sProblem summary
--objective-oObjectives (repeatable)
--acceptance-aAcceptance criteria (repeatable)
--touchKnown touch points (repeatable)
--outputWrite to file instead of stdout

bundle

Bundle request for copy/paste into GitHub Copilot web.

IMPORTANT: Copilot web can only see committed & pushed changes!

OptionShortDescription
--file-fMarkdown request file (required)
--repo-dir-dRepository directory (for git status check)
--output-oOutput file (default: stdout)
--clipboard-cCopy to clipboard (xclip/pbcopy)
--skip-git-checkSkip git status verification

models

List available models for a provider.

OptionShortDescription
--provider-PProvider to list models for

template

Print the example review request template.

find

Search for review request markdown files.

Workspace Feature

The --workspace flag copies local files to a temporary directory that providers can access. This is useful when you have uncommitted changes that aren't visible to remote-based providers.

bash
# Copy src/ and tests/ to temp workspace
python .agents/skills/code-review/code_review.py review \
  --file request.md \
  --workspace ./src \
  --workspace ./tests

The workspace is automatically cleaned up after the review completes.

Provider-Specific Notes

GitHub Copilot (github)

  • •Requires gh CLI authenticated
  • •Supports --continue for session continuity
  • •Models: gpt-5, claude-sonnet-4, claude-sonnet-4.5, claude-haiku-4.5

Anthropic Claude (anthropic)

  • •Requires claude CLI
  • •Supports --continue for session continuity
  • •Models: opus, sonnet, haiku, opus-4.5, sonnet-4.5, sonnet-4

OpenAI Codex (openai)

  • •Requires codex CLI
  • •Default reasoning: high (best results)
  • •Models: gpt-5, gpt-5.2, gpt-5.2-codex, o3, o3-mini
  • •Does NOT support --continue (session context lost between rounds)

Google Gemini (google)

  • •Requires gemini CLI
  • •Models: gemini-3-pro, gemini-3-flash, gemini-2.5-pro, gemini-2.5-flash, auto
  • •Does NOT support --continue (use /chat save/resume in interactive mode)

Project Agent Workflow (Recommended)

The intended workflow is for the project agent (Claude) to interpret and apply suggestions:

code
User asks for code review
        |
Project agent creates review request (follows template)
        |
review/review-full sends to provider CLI
        |
Project agent parses output, decides what to apply
        |
Project agent uses Edit tool to apply changes it concurs with
        |
If unclear, agent asks provider (another round) or the user

Why this approach:

  • •Patch format may be malformed (won't git apply)
  • •Project agent can exercise judgment on suggestions
  • •Agent can ask clarifying questions back to provider
  • •Same workflow humans use, but automated

Dependencies

bash
pip install typer rich