AgentSkillsCN

Matrix Code Review

当用户要求“审查这段代码”、“审查 PR”、“代码评审”、“审查暂存变更”、“分析影响范围”、“检查变更影响”时，或需要进行全面的上下文感知代码评审时，应使用此技能。

SKILL.md
--- frontmatter
name: Matrix Code Review
description: This skill should be used when the user asks to "review this code", "review PR", "code review", "review staged changes", "blast radius analysis", "check impact of changes", or needs comprehensive context-aware code review.
user-invocable: true
context: fork
allowed-tools:
  # Matrix Code Index
  - mcp__plugin_matrix_matrix__matrix_find_callers
  - mcp__plugin_matrix_matrix__matrix_find_definition
  - mcp__plugin_matrix_matrix__matrix_search_symbols
  - mcp__plugin_matrix_matrix__matrix_list_exports
  - mcp__plugin_matrix_matrix__matrix_get_imports
  - mcp__plugin_matrix_matrix__matrix_index_status
  # Matrix Memory
  - mcp__plugin_matrix_matrix__matrix_recall
  - mcp__plugin_matrix_matrix__matrix_store
  - mcp__plugin_matrix_matrix__matrix_reward
  - mcp__plugin_matrix_matrix__matrix_failure
  # Context7 - Library Documentation
  - mcp__plugin_matrix_context7__resolve-library-id
  - mcp__plugin_matrix_context7__query-docs
  # Standard Tools
  - Read
  - Grep
  - Glob
  - Bash

Matrix Code Review

Perform comprehensive, context-aware code review using 5-agent architecture with 95% false positive reduction.

Tip: This skill runs in a forked context for an unbiased perspective - similar to how a human reviewer would approach the code for the first time.

Architecture

code
ORCHESTRATOR (parses target, routes, aggregates)
     │
     ├── DETECTION AGENT → security, runtime, breaking, logic flaws
     ├── IMPACT AGENT → blast radius, transitive graph, test coverage
     ├── TRIAGE AGENT → tier assignment, confidence calibration, noise filter
     ├── REMEDIATION AGENT → context-aware fixes, regression checks
     └── VERIFICATION AGENT → build, test, lint validation

Usage

Parse user arguments from the skill invocation (text after the trigger phrase).

Expected format: <target> [mode]

  • •target: File path, PR number, or "staged" for staged changes
  • •mode (optional): default | lazy (default: from config or default)

Modes

Default Mode (Comprehensive)

Full 5-agent review pipeline with maximum index utilization:

  • •Detection: Security vulns, runtime issues, breaking changes
  • •Impact: Transitive blast radius (2-3 levels), service boundaries, test coverage
  • •Triage: Tier classification, >80% signal ratio target
  • •Remediation: Context-aware fixes matching codebase patterns
  • •Verification: Run build, test, lint commands automatically

Lazy Mode (Quick)

Detection agent only:

  • •Direct code inspection
  • •Critical issues only (Tier 1)
  • •~2-3 comments, no blast radius

Review Pipeline

Follow the 5-agent orchestration detailed in references/review-phases.md:

  1. •Orchestrator - Parse target, dispatch agents, aggregate results
  2. •Detection Agent - Find security, runtime, breaking, logic issues
  3. •Impact Agent - Calculate transitive blast radius, test coverage
  4. •Triage Agent - Classify tiers, calibrate confidence, filter noise
  5. •Remediation Agent - Generate fixes, check regression risk
  6. •Verification Agent - Run build, test, lint; report results

Early exit: If Detection finds nothing critical, skip Impact/Triage depth.

Examples

code
/matrix:review src/utils/auth.ts          # Default mode (comprehensive)
/matrix:review staged                     # Review staged changes
/matrix:review staged lazy                # Quick review of staged changes
/matrix:review 123                        # Review PR #123
/matrix:review 123 lazy                   # Quick review of PR #123

Additional Resources

Reference Files

  • •references/review-phases.md - Orchestrator + 5-agent pipeline
  • •references/agents/detection.md - Detection patterns and output format
  • •references/agents/impact.md - Blast radius algorithm
  • •references/agents/triage.md - Tier classification and signal ratio
  • •references/agents/remediation.md - Fix generation and regression checks
  • •references/agents/verification.md - Build/test/lint command detection and execution