robertoecf/adversarial-review
11 stars · Last commit 2026-09-30
Review triad plugin for Claude Code: prompt optimization, adversarial red-team review, plan validation, and intelligent orchestration.
README preview
# adversarial-review **v0.9.9**: Cross-host adversarial review for coding workflows. Works in **Claude Code, Codex, Pi, and Grok** from the same source: detects which agent host you're running in and routes the heavy critique to the OTHER agent. Critique discipline (v0.8.1): break confidence rather than validate; expensive attack surfaces first; material-only findings; mandatory simplicity counterfactuals; steerable focus; terse ship/no-ship summary. Lessons adapted from OpenAI Codex Companion and Matt Pocock's engineering skills. Routing uses three tiers: T1 Claude and Codex review each other, T2 is the Pi chain (opencode-go DeepSeek V4.1 Flash, GLM 5.3, then Grok 4.7), and T3 is the Gemini Antigravity ladder as last resort. Reviewers get read-only tools. On host Claude, Codex-authored work is reviewed by the Claude main loop itself. Outside host Claude the Codex reviewer defaults to GPT-6 Astra medium; the architect selects high only for auth, sensitive data, migrations, concurrency, or changes spanning multiple modules. The Gemini ladder is `gemini-3.8-flash-high`, `gemini-3.7-flash-high`, then `gemini-3.1-pro-high`.