| name | multi-brain |
| description | Multi-LLM auto-router. Your main assistant orchestrates and routes a sub-task to a different model ONLY when that model fits the job better; the user talks to one assistant the whole time. The roster is yours to fill in. Fires automatically when a routing rule matches.
NO-SELF-REVIEW LAW: when the user asks you to "check / review / verify / sanity-check / second-opinion" ANY work YOU just produced, route the review to a model from a DIFFERENT lineage if one is on the roster. Triggers: "check your work" · "review what you just did" · "is this right?" · "second opinion" · "sanity check" · "double-check" · "make sure this works".
|
Multi-Brain Auto-Router
Many brains, one terminal. The orchestrator drives and stays in front of the user; it hands a sub-task to whichever brain is genuinely better for it, then integrates the result. Routing should be invisible and rare — most turns stay with the main model.
Why different lineages matter: two models from the same company share training, so they share blind spots. Cross-architecture diversity (e.g. Anthropic + OpenAI + Google + a local open-weights model) is what actually catches errors.
THE NO-SELF-REVIEW LAW (read this first). When the user asks you to "check / review / verify / sanity-check / second-opinion" ANY work YOU just produced — code, writing, plan, anything — route the review to a model from a DIFFERENT lineage if one is on the roster. A model reviewing its own output has the same blind spots that produced the bug. If no second lineage is installed, say so, then re-derive the check from scratch instead of rubber-stamping.
Your roster — FILL THIS IN
Edit this table for your setup by hand: delete the rows for brains you don't have, add a row for each brain you do. Examples are placeholders.
| Brain | Lineage | Best at (role) | How to call | Cost |
|---|
| (main) — e.g. Claude | e.g. Anthropic | orchestrator, builder, driver | — | plan |
| example: Codex | OpenAI/GPT | code review, adversarial critic, rescue when stuck | codex exec "<prompt>" | plan |
| example: Gemini | Google | multimodal (video/audio/PDF), huge-context repo scans | gemini -p "<prompt>" @file | free tier |
| example: local model | open-weights (Ollama) | private/offline work, free bulk tasks | ollama run <model> "<prompt>" | free |
No second CLI installed yet? That's fine — the skill stays mostly asleep. The No-Self-Review Law still applies in its fallback form (re-derive, don't rubber-stamp). When you install a second brain, add the row and the routes light up.
Routing table — fire when:
- [MUST] review/check/verify work the main model just produced → different-lineage review (the Law above)
- "tear apart / stress test / poke holes" → adversarial review on a different lineage
- "I'm stuck / hand it off" OR the main model failed the same operation 2+ times in a row → rescue route to another brain (send full context: failing output, what was tried)
- video / audio / huge PDF, and a multimodal brain is on the roster → multimodal route
- "scan the whole repo / find every place X", and a huge-context brain is on the roster → whole-repo route
- "ask all the models / consensus" → consensus panel, capped at 3 lineages, each returning the same structured template (recommendation / risks / assumptions / confidence); the orchestrator adjudicates by evidence, not by averaging
Do NOT fire for (stay with the main model):
- "Explain / what is / how does" → direct answer
- "Write / draft / build / edit / plan" → direct (but if the user then says "now check your work" — the Law applies)
- Reviewing content the USER wrote (their notes, their draft) → direct
- Conversation, status, file ops, normal Q&A → direct
- "What did we decide?" → the decisions log and wiki, not another model
Announcement protocol
Whenever you route somewhere the user did NOT explicitly request — announce in one line BEFORE running, so they can stop you with one word:
[multi-brain] routing review to <brain> — different lineage, no self-review
[multi-brain] handing to <brain> rescue — failed the same step 2x
Failure-detection rule (deterministic, not a vibe)
2× same test failure, 2× same shell error, or 2× same edit with no progress → MUST route a rescue (if a second brain exists) or stop and tell the user. Reset the counter only when the step passes, the goal changes, or the user says "keep trying." This 2× counter is the same kind of loop brake taught in references/agent-loops.md — read it before wiring any run-until-done loop.
Cost discipline
Default to the main model. Route only when the specialist's strength is the actual bottleneck. One strong model usually matches an ensemble at a fraction of the cost — the ensemble's real win is catching errors on high-stakes work, not speed. Under-firing is fine; over-firing breaks trust and burns money.
