three-brain

To prevent ineffective self-review during code review, this Skill detects review verbs like "check" or "verify" and automatically routes Claude's just-produced work to external expert models like Codex for an objective, adversarial analysis instead of a blind self-audit.

Adversarial testing on sensitive code paths is triggered automatically when an active edit touches high-blast-radius areas such as auth, billing, secrets, or deployment configurations, routing the changes to Codex for specialized risk-path detection and review.

Multimodal analysis handles video and PDF preprocessing by routing media-heavy requests to Gemini, applying explicit pipelines that cap duration or pages, extract audio, and enforce structured, evidence-first outputs with timestamps, page numbers, or file/line citations.

Yes, you can escalate tasks to a rescue model after repeated failures; the Skill uses deterministic failure counters to monitor repeated unsuccessful attempts on the same operation and automatically triggers Codex rescue after defined retries.

Yes, high-stakes consensus mode is supported when explicitly requested, running Claude, Codex, and Gemini in parallel to analyze the same task and enforcing a structured adjudication template to resolve differences and verify results.

You should avoid automated expert reviewer routing for simple, low-blast-radius edits that do not touch sensitive code paths, as the overhead of invoking external models is unnecessary when standard self-review or basic repo scanning suffices.