What problem does it solve? Setting up external AI model providers (Codex, GitHub Copilot, Gemini) for code review and design feedback involves hidden failure modes: policy gates, token precedence, model-name churn, and same-family reviewers that silently agree with the author. This Skill walks through setup, verification, and correct pinning so a QA gate never silently passes. ## Core Features & Use Cases - Provider setup and verification: Step-by-step checks for Codex CLI, GitHub Copilot CLI (three ordered gates: policy, token, real call), and Gemini (CLI or direct API), each ending in a real probe call rather than a PATH check. - Model family interlock: Enforces the rule that a reviewer must come from a different model family than the author, with role pins, alternates, and an announced claude-sonnet-5 fallback. - Cost and privacy guidance: Explains Copilot AI Credits multipliers, Gemini free-tier data training trade-offs, and local-model options (Ollama, Kimi open weights) for code that must stay on the machine. - Use Case: A team wants an independent reviewer on every diff. Use this Skill to install Codex, pin gpt-5.6-luna for QA review, confirm Copilot's org policy allows the CLI, and verify each provider with a real round-trip call before trusting the gate. ## Quick Start Ask the assistant to set up Codex as a QA reviewer and verify it with a real probe call before wiring it into the review workflow.
