nyldn/claude-octopus v4.7.0

v4.7.0 - Adversarial Cross-Model Review

on GitHub

🤼 Crossfire: Adversarial Cross-Model Review

This release introduces two new commands that leverage adversarial AI-vs-AI review to catch more issues than single-model review. Different models have different blind spots—Crossfire forces them to critique each other.

New Commands

grapple - Adversarial Debate

Codex vs Gemini wrestling match until consensus:

./scripts/orchestrate.sh grapple "implement password reset API"
./scripts/orchestrate.sh grapple --principles security "implement JWT auth"

Flow:

1. Round 1: Both models propose solutions independently
2. Round 2: Cross-critique (Gemini critiques Codex, Codex critiques Gemini)
3. Round 3: Synthesis determines winner and final implementation

squeeze - Red Team Security Review

Blue Team defends, Red Team attacks:

./scripts/orchestrate.sh squeeze "review auth.ts for vulnerabilities"

Flow:

1. Blue Team (Codex): Implements secure solution
2. Red Team (Gemini): Finds vulnerabilities with exploit proofs
3. Remediation: Fixes all found issues
4. Validation: Verifies all vulnerabilities are fixed

Constitutional Principles

Grapple supports domain-specific critique principles via --principles:

Auto-Routing

The auto command now detects crossfire intents:

./scripts/orchestrate.sh auto "security audit the auth module"  # → squeeze
./scripts/orchestrate.sh auto "have both models debate the design"  # → grapple

Cost Estimate

Both are more expensive than single-agent but catch 2-3x more issues.

Full Changelog: v4.6.0...v4.7.0