feat(ai): infer model provider options (#39493)

packages/ai/test/llm-option-types.types.ts

@@ -0,0 +1,62 @@
+import { Schema } from "effect"
+import { LLM, Model, type ModelProviderOptions, type ProviderOptions } from "../src"
+import { OpenAIChat } from "../src/protocols"
+
+interface ExampleOptions {
+  readonly [key: string]: unknown
+  readonly mode?: "fast" | "thorough"
+}
+
+type ExampleProviderOptions = ProviderOptions & {
+  readonly example?: ExampleOptions
+}
+
+const model = OpenAIChat.route
+  .with({ endpoint: { baseURL: "https://example.com/v1" } })
+  .model<ExampleProviderOptions>({ id: "example" })
+
+LLM.request({ model, prompt: "Hello", providerOptions: { example: { mode: "fast" } } })
+LLM.request({ model, prompt: "Hello", providerOptions: { future: { option: true } } })
+
+LLM.request({
+  model,
+  prompt: "Hello",
+  // @ts-expect-error Known provider options preserve their value types.
+  providerOptions: { example: { mode: "slow" } },
+})
+
+LLM.generateObject({
+  model,
+  prompt: "Hello",
+  schema: Schema.Struct({ answer: Schema.String }),
+  providerOptions: { example: { mode: "thorough" } },
+})
+
+LLM.generateObject({
+  model,
+  prompt: "Hello",
+  jsonSchema: { type: "object" },
+  // @ts-expect-error Dynamic object generation uses the selected model's provider options.
+  providerOptions: { example: { mode: false } },
+})
+
+declare const generic: Model
+LLM.request({ model: generic, prompt: "Hello", providerOptions: { arbitrary: { option: true } } })
+
+const options: ModelProviderOptions<typeof model> = { example: { mode: "fast" } }
+void options
+
+model.route.model<ExampleProviderOptions>({
+  id: "example-with-defaults",
+  defaults: {
+    // @ts-expect-error Low-level model defaults preserve known provider option types.
+    providerOptions: { example: { mode: 1 } },
+  },
+})
+
+Model.update(model, {
+  defaults: {
+    // @ts-expect-error Updating a model cannot contradict its provider option type.
+    providerOptions: { example: { mode: "slow" } },
+  },
+})