@yielded/agent-ai-decision

Choose a native Effect language model once per thread with AutoModel. Follow-up runs retain that choice, while each child thread selects independently. The upstream DecisionModel service evaluates the application-approved catalog.

import { OpenAiLanguageModel } from "@effect/ai-openai";
import { AutoModel } from "@yielded/agent-ai-decision";

const ThreadModels = AutoModel.make({
  version: "profiles-v1",
  models: {
    routine: {
      description: "Low cost; routine tasks",
      model: OpenAiLanguageModel.model("gpt-6-luna"),
    },
    complex: {
      description: "Difficult reasoning and ambiguous requirements",
      model: OpenAiLanguageModel.model("gpt-6-astra"),
    },
  },
});

Provide ThreadModels, a native DecisionModel such as TypeSafeDecisionModel.model("jev-latest") from @effect/ai-typesafe, provider clients, and one shared AutoModel.layerMemory() around your agent and subagent handlers. At least two profiles are required. Selection does not acquire candidate models until generation starts.

Durable hosts provide AutoModel.SelectionStore to atomically retain version 2 selection records. Wrong-thread, missing-profile, catalog-version, and record-version mismatches fail without reselection or mutation. Explicit select, restore, and resolve support host-owned admission.

For ordinary assessments, import Decision and DecisionModel from effect/ai. Use LanguageModelDecisionModel.layer to answer those decisions with any native language model that supports structured output. Provide your chosen model Layer to the adapter; provider clients and configuration remain application-owned. Its probabilities are LLM estimates, not calibrated confidence scores. Native validation rejects invalid distributions; retries and failover are explicit. See the language-model example and provider setup.

See the reference for ownership, configuration, and migration details, or the runnable decision, direct client, and tool examples.