Add Autonomous Legal War Game — adversarial multi-agent M&A stress-tester

[fmalik100]

Summary

• Adds legal_warroom/ — a standalone Python tool that stress-tests legal documents (M&A agreements, NDAs, etc.) using an adversarial multi-agent simulation
• Multi-round automatic agent interaction — a Plaintiff Agent (Red Team) attacks each clause, a Defense Agent (Blue Team) hardens the language, then the Plaintiff re-attacks the hardened rewrite; this loops until convergence or a configurable round cap
• Dual provider support — runs on Anthropic (claude-opus-4-6 with adaptive thinking) or 100% locally via Ollama (no API costs) with any compatible model (qwen2.5:14b, llama3.1:8b, etc.)

Architecture

legal_warroom/
├── main.py                          CLI (typer) — --provider, --model, --rounds, --parallel, --html
├── warroom/
│   ├── providers/
│   │   ├── base.py                  LLMProvider protocol + make_provider() factory
│   │   ├── anthropic_p.py           Anthropic SDK backend
│   │   └── ollama_p.py              Ollama backend (OpenAI-compatible API)
│   ├── agents/
│   │   ├── plaintiff.py             Red Team — hunts ambiguity, liability gaps, edge cases
│   │   └── defense.py               Blue Team — precision redrafting, preserves business intent
│   ├── loop/
│   │   └── adversarial.py           Multi-round loop with convergence detection
│   ├── models/schemas.py            Pydantic structured outputs + IterativeSegmentReport
│   ├── document/processor.py        PDF/TXT ingestion, section-aware segmentation
│   ├── orchestrator.py              Sequential / parallel segment processing
│   └── report/generator.py          Terminal + JSON + HTML report output

Test plan

• python main.py sample.txt --max-segments 1 — single segment dry run (Anthropic)
• python main.py sample.txt --provider ollama --model llama3.1:8b --max-segments 1 — Ollama dry run
• python main.py sample.txt --rounds 2 --html — multi-round with HTML report
• python main.py sample.txt --parallel --max-segments 3 — parallel processing
• Verify JSON report written to output/
• Verify exit code 2 when CRITICAL vulnerabilities remain

🤖 Generated with Claude Code