/route-three-brain

name: route-three-brain
description: The routing law for handing off between Claude (primary driver), Codex/GPT-5.5 (adversarial review, second opinions), and Gemini (long-context, repo-wide synthesis). Codifies when each model is the right tool and enforces the no-self-review rule. Trigger when the user says "three brain", "route this", "which model should", "delegate to codex / gemini", or when a substantial Claude output needs an independent check.

route-three-brain

**Family:** routing **Status:** Stable

Purpose

The decision layer for cross-model routing. Three models, three roles:

• **Claude** — primary driver for code, content, design generation, conversation.
• **Codex / GPT-5.5** — adversarial review, second-opinion debugging, deep

correctness/security checking.

• **Gemini** — large-context whole-repo analysis, multi-document synthesis, big

structured-data passes.

**Hard rule:** Claude never reviews Claude — same architecture, same blind spots. If the user says "check your work" / "review this" / "is this right" / "second opinion" / "sanity check," route to Codex via `route-codex-review`.

Triggers

• "three brain" / "tri-brain" / "route this" / "which model"
• "delegate to codex" / "delegate to gemini" / "best model for X"

Inputs

• Task description; the current output (if asking for review/handoff)
• Constraints (cost, time, privacy)

Steps

1. **Startup self-check (once per session).** Verify the tools exist: `codex --version` and `gemini --version`. If one is missing, announce it once and route only to what's available. 2. **Classify the task:**

• Build/generate/converse → **Claude** keeps it.
• Review / verify / debug a substantial output, or any risk-path edit (auth,

billing, migrations, deploy, secrets) → **Codex** via `route-codex-review`.

• Needs more context than one window (whole repo, many docs) → **Gemini** via

`route-gemini-context`. 3. **Announce forced routes** in one line before running (so the user can interrupt): `[three-brain] routing to Codex (review) — risk path: src/auth/`. 4. **Execute** the handoff via the execution skills and **integrate** the result — diff where models agree/disagree and adjudicate by evidence, not by averaging.

Outputs

• A routing decision with justification
• The integrated result of any executed handoff

Dependencies

• `route-codex-review` (Codex execution), `route-gemini-context` (Gemini execution)
• The `codex` and `gemini` CLIs installed (or the adjacent `openai-codex` /

`cc-gemini-plugin` tools); degrades to a recommendation when neither is present

Notes

The law layer; the two `route-*` skills are the execution layers. Under-firing is fine; over-firing breaks trust. When uncertain, recommend rather than auto-route.