What problem does it solve? Same-model code review shares the blind spots of the model that wrote the code, so defects survive review. This Skill adds an independent second opinion from a structurally different LLM to catch correlated blind spots before merge. ## Core Features & Use Cases - Two-phase review: Runs the standard five-pass structural review first, then dispatches the diff to a different LLM family for adversarial findings only if phase one approves. - Multi-backend fallback chain: Supports OpenAI-compatible providers (ollama, vLLM, OpenRouter, NVIDIA, DeepSeek, Gemini) and native Anthropic, with ordered fallbacks, per-backend timeouts, deadlines, and context-window preflight. - Autonomous-run escalation: On unattended L3/L4 runs, a critical adversarial finding escalates the verdict to needs-human; on interactive runs findings are soft signals the implementor must disprove with evidence. - Use Case: Configure slots.review: faffter-dark-adversarial-review in .faffrc so an overnight autonomous delivery run gets an independent review from a local ollama model before merging agent-written code. ## Quick Start Configure the review slot in .faffrc to use faffter-dark-adversarial-review with an adversarial provider and model, then let the graft pipeline invoke it during its review step.
