adversarial-review

Cross-host adversarial review of any artifact — implementation plans, code/diffs/configs, or prompts/skill definitions. Classifies the input by context, then routes the heavy critique to the agent that is NOT the host (Codex from Claude, Pi chain from Codex, etc.), cross-validates against independent host-side analysis, and returns unified critics with severity ratings and a verdict. Prompts are analyzed host-side across 6 dimensions. Falls back to Gemini via Antigravity, then degraded host-self with explicit warning.