Code Review Agents

Reverse-engineered approximation of Anthropic's Claude Code Review multi-agent system.

Architecture

GitHub Webhook
    │
    ▼
┌──────────────┐
│ ORCHESTRATOR │ ── reads CLAUDE.md, REVIEW.md, builds context
└──────┬───────┘
       │
       ├──► Logic Agent ──────────┐
       ├──► Security Agent ───────┤
       ├──► Edge Case Agent ──────┤  parallel
       ├──► Regression Agent ─────┤
       └──► Convention Agent ─────┘
                                  │
                                  ▼
                          ┌──────────────┐
                          │  VERIFIER    │ ── false-positive filter
                          └──────┬───────┘
                                 │
                                 ▼
                          ┌──────────────┐
                          │ DEDUPLICATOR │ ── merge, rank, format
                          └──────┬───────┘
                                 │
                                 ▼
                          GitHub Inline Comments

Agents

Agent	File	Role
Orchestrator	orchestrator.md	Entry point. Fetches context, fans out to review agents, collects results.
Logic	agents/logic.md	Finds logical errors, incorrect conditions, wrong return values, broken control flow.
Security	agents/security.md	Identifies injection, auth bypass, data exposure, unsafe deserialization, etc.
Edge Cases	agents/edge-cases.md	Catches unhandled nulls, empty inputs, boundary conditions, race conditions.
Regression	agents/regression.md	Detects changes that break existing behavior or violate implicit contracts.
Convention	agents/convention.md	Enforces project-specific style rules from REVIEW.md and CLAUDE.md.
Verifier	agents/verifier.md	Cross-checks each candidate finding against actual code to eliminate false positives.
Deduplicator	agents/deduplicator.md	Merges overlapping findings, assigns final severity, formats for GitHub posting.

Context Schema

See context-schema.md for the shared data structures passed between agents.