subagent-routing

Injects a subagent model-routing policy into the system prompt each turn (for the frontier pi profile). The orchestrator plans, specs, and judges on its own live model, fans implementation out to other configured models, then spawns a reviewer subagent to grade the actual diff — not the worker's self-report — before accepting it. Every spawn carries a maxTurns budget (a stuck agent dies and escalates instead of spinning), reviews run on the mid tier for mechanical work, and two failed reviews of one task stop the ladder and report rather than climbing further.

Every configured, authenticated model is a candidate, minus local (mlx/*) models and the orchestrator's own. The ladder is rebuilt each turn from the model registry: models are deduped by id across providers (cheapest provider wins), sorted by price and split into terciles — cheap, mid, frontier — rendered compactly, one line per provider per tier.

The mode shapes the decision procedure, not the menu:

• cost (default) — route each task to the cheapest tier plausibly adequate for it; escalate one tier on a failed review; frontier only for genuinely hard tasks.
• performance — route each task to the tier most likely to pass review first try; mechanical work still goes cheap, and the priciest model is never the default.

Price is the capability proxy (the registry carries no benchmark data), so the config takes per-model tier overrides for the models price misplaces — e.g. a $0 subscription model that is actually frontier-class:

{
  "mode": "cost",
  "tiers": { "kimi-coding/*": "frontier" }
}

Keys are provider/id exact or provider/*; exact beats glob. Config lives at <agentDir>/subagent-routing.json and is re-read every turn, so hand-edits apply on the next turn. Switch modes with /subagent-routing cost|performance. The policy also licenses the orchestrator to trust its own knowledge of a model over its price-derived tier.

Nothing names the orchestrator model: spawn a subagent with no model override and it inherits whatever you are running. The judge model is chosen per task — mid tier for mechanical work, the live model for disputes, second failures, and security-sensitive diffs.

Pairs with the active-model extension (which tells the orchestrator its live identity), the global reviewer agent type (~/.pi/agent/agents/reviewer.md), and @tintinweb/pi-subagents (which provides the Agent tool this policy drives).

Install

pi install npm:@guygrigsby/pi-subagent-routing