review-local-changes

针对本地未提交/暂存的更改，开展多智能体代码评审——无需 GitLab MR。适用于提交前或推送前的代码审查。智能体包括：架构（Claude）、安全（Gemini）、性能（Claude）、测试（Gemini）。

SKILL.md

--- frontmatter

name: review-local-changes
description: Multi-agent code review for local uncommitted/staged changes — no GitLab MR required. Use for pre-commit or pre-push reviews. Agents: Architecture (Claude), Security (Gemini), Performance (Claude), Testing (Gemini).

Review Local Changes

Review local changes using Claude + Gemini agents. No GitLab required.

Inputs

Input	Type	Default	Purpose
`mode`	string	staged	staged, unstaged, all, commit, branch
`commit_sha`	string	-	For mode=commit
`base_branch`	string	main	For mode=branch
`repo`	string	.	Repo path
`agents`	string	architecture,security,performance	Comma-separated
`model`	string	sonnet	sonnet, opus, haiku
`files`	string	-	Comma-separated files to review
`run_tests`	bool	false	Run tests before review
`run_precommit`	bool	false	Run pre-commit hooks

Modes

•staged: git diff --cached
•unstaged: git diff
•all: git diff HEAD
•commit: git show {sha}
•branch: git diff {base}...HEAD

Workflow

1. Get Diff

•Run git command based on mode
•Limit diff to ~15k chars for LLM context
•Get --stat for summary
•Abort if no changes

2. Optional Quality Checks

•If run_tests: test_run(repo)
•If run_precommit: precommit_run(repo)

3. Parse Agents

•Enabled: architecture, security, performance, testing, documentation, style
•Config: architecture→claude, security→gemini, performance→claude, testing→gemini

4. Run Agents (parallel)

•Each agent: prompt with diff, focus area
•Prompts: architecture (SOLID, design), security (auth, injection), performance (complexity, caching), testing (coverage)
•Output format: [CRITICAL], [WARNING], [SUGGESTION]

5. Synthesize Review

•Combine agent outputs
•Claude synthesis: PASS / WARN / BLOCK
•If CRITICAL → BLOCK; if WARNING only → WARN; else PASS

6. Memory

•memory_session_log("Local code review ({mode})", "Verdict: {verdict}, Agents: {count}")

Agent Prompts (concise)

•Architecture: design patterns, SOLID, modularity
•Security: input validation, auth, injection
•Performance: complexity, resource usage, caching
•Testing: coverage, edge cases

Output

•Mode, lines, verdict (PASS/WARN/BLOCK)
•Changes stat
•Synthesized review
•Agent results list
•Ready to commit / Consider warnings / Fix critical

Tools

•git diff (via shell or git tools)
•test_run, precommit_run (if enabled)
•Claude/Gemini CLI for agent reviews
•memory_session_log