Complex models

An executor can route to several targets, enrich a request, and preserve both streaming and non-streaming behavior. Keep this logic in the model source, not in client applications.

complex-model.ts

import { prependSystemPrompt } from "@neutrome/lil-engine";
import type { Executor } from "@neutrome/lilsdk";
import { fallback, retry } from "@neutrome/lilsdk/loops";

const upstream = fallback([
  retry("openai/gpt-5", { attempts: 2 }),
  "openai/gpt-5-mini",
]);

const executor: Executor = {
  async execute(request, ctx) {
    return ctx.invoke(
      upstream,
      prependSystemPrompt(request, "Answer clearly and cite uncertainty."),
    );
  },

  async *stream(request, ctx) {
    yield* ctx.invokeStream(
      upstream,
      prependSystemPrompt(request, "Answer clearly and cite uncertainty."),
    );
  },
};

export default executor;

prependSystemPrompt() returns a new program. It does not mutate the client request. retry() retries only before a stream emits content; fallback() tries the next target only when the prior target fails before emitting content.

Composition guidelines

• Use provider-qualified targets such as openai/gpt-5.
• Use another workspace model ID when composing your own models.
• Keep execute and stream behavior equivalent unless the model is intentionally streaming-only.
• Attach tools in the dashboard; a deployed attachment owns the tool loop.

Use the loops reference and engine reference when you need the complete option contracts.