@swoosh-dev/judge

Dynamic routing policies for @swoosh-dev/router. Classify each request's prompt with an LLM judge (structured output), then route by the verdict — cheap models for small talk, top models for hard problems, search-capable models for current events.

npm install @swoosh-dev/judge

import { ModelRouter } from "@swoosh-dev/router";
import { byBenchmark } from "@swoosh-dev/router";
import { llmJudgePolicy } from "@swoosh-dev/judge";

type Kind = "chat" | "coding" | "research";

const smart = llmJudgePolicy<Kind>({
  // Your structured-output classifier — route the judge through a cheap model.
  classify: ({ prompt }) => judge.generateObject({ schema: KindSchema, prompt }),
  route: {
    chat: "cheapest",
    coding: byBenchmark("swe_bench"),
    research: ({ candidates }) => candidates.filter((c) => c.capability.features.includes("web_search")),
  },
  fallback: "balanced",     // used if classify throws or returns an unmapped verdict
});

await router.run({ ...request, preference: smart });

How it works

llmJudgePolicy returns an ordinary RoutingPolicy — async, since it awaits your classifier. The verdict picks a base preference/policy from route; the result is still a ranked candidate list, so fallback, constraints, and plan() inspection all work unchanged. plan.preference reports "custom".

• You supply the judge. classify is your call (sync or async). The package adds no LLM dependency — route the judge itself through a cheap swoosh model if you like.
• Verdicts are cached by prompt signature (the judge costs latency + tokens). On by default; cache: false to disable, or cache: { key, max } to tune.
• Best-effort prompt. The classifier receives { prompt, request }, where prompt is request.prompt if present, else a stringified request.input — so it works with plan, run, and runText.

Options

| Field | Type | Notes | | --- | --- | --- | | classify | (input) => Verdict \| Promise<Verdict> | Your structured-output classifier. | | route | Record<Verdict, RoutingPreference \| RoutingPolicy> | Verdict → base ranking. | | fallback | RoutingPreference \| RoutingPolicy | On error/unmapped verdict. Defaults to "balanced". | | cache | boolean \| { key?, max? } | Cache verdicts by prompt. Defaults to on. |