Claude 3.5 Sonnet Hierarchical Multi-Agent Orchestrator

You are a Principal AI Agent Architect specializing in LangGraph, Anthropic Tool Use, and Autonomous Multi-Agent Swarms.

Architect a production hierarchical multi-agent system for {Automated Pull Request Code Review, Static Security Audit, and GitHub PR Comment Generation} using {LangGraph (TypeScript / JavaScript)}.

### MANDATORY MULTI-AGENT INVARIANTS:
1. **Supervisor Router Pattern**: A dedicated Supervisor LLM analyzes incoming tasks and routes execution to specialized subagents via structured function calling (`next_agent: 'researcher' | 'coder' | 'auditor' | '__end__'`).
2. **Tool Scoping & Least Privilege**: Subagents must only possess tools strictly required for their specific domain. Never give all tools to all agents.
3. **Shared State Graph (TypedDict / Zod)**: State updates must be explicit and immutable. Use reducer functions to append messages and update shared artifacts.
4. **Human-in-the-Loop (HITL) Gate**: Insert a mandatory pause/approval checkpoint before any destructive actions (e.g., executing transactions, deleting data, pushing to production).
5. **Infinite Loop Prevention**: Enforce a hard recursion limit (max 15 hops) with graceful degradation to supervisor explanation.

Subagent Roles: {1) DiffParserAgent (tools: fetch_pr_diff, list_files), 2) SecurityAuditorAgent (tools: scan_semgrep_ast, check_cve), 3) CommentFormatterAgent (tools: post_github_review)}
Shared State Schema: {pr_id, repo, diff_chunks, security_vulnerabilities, audit_score, approved_by_human}

Generate complete project code:
1. Shared State Definition
2. Supervisor Node & Worker Subagent definitions with tool bindings
3. Graph Compilation with conditional edges and HITL interrupt points