feat: refresh models from a hosted catalog

.github/workflows/refresh-model-catalog.yml

@@ -0,0 +1,61 @@
+name: Refresh hosted model catalog
+
+on:
+  schedule:
+    # Daily, off the top of the hour to avoid the cron rush.
+    - cron: "17 5 * * *"
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+concurrency:
+  group: refresh-hosted-model-catalog
+  cancel-in-progress: false
+
+jobs:
+  publish:
+    name: Generate, validate, and publish
+    runs-on: ubuntu-latest
+    timeout-minutes: 20
+    steps:
+      - name: Checkout default branch
+        uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7.0.0
+        with:
+          ref: ${{ github.event.repository.default_branch }}
+          persist-credentials: false
+
+      - name: Setup Node.js
+        uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020 # v7.0.0
+        with:
+          node-version: 22
+          cache: npm
+
+      - name: Install dependencies
+        run: npm ci --ignore-scripts
+
+      - name: Generate aggregate catalog
+        working-directory: packages/ai
+        env:
+          PRIME_AGENT_MODEL_CATALOG_OUTPUT: /tmp/model-catalog.json
+          PRIME_AGENT_MODEL_CATALOG_STRICT: "1"
+        run: npm run generate-models
+
+      - name: Validate catalog
+        working-directory: packages/ai
+        run: npm run validate-model-catalog -- /tmp/model-catalog.json
+
+      - name: Publish catalog to R2
+        env:
+          AWS_ACCESS_KEY_ID: ${{ secrets.R2_ACCESS_KEY_ID }}
+          AWS_DEFAULT_REGION: auto
+          AWS_SECRET_ACCESS_KEY: ${{ secrets.R2_SECRET_ACCESS_KEY }}
+          R2_BUCKET: ${{ secrets.R2_BUCKET }}
+          R2_ENDPOINT_URL: ${{ secrets.R2_ENDPOINT_URL }}
+        run: |
+          test -n "$R2_BUCKET"
+          test -n "$R2_ENDPOINT_URL"
+          aws s3 cp /tmp/model-catalog.json "s3://${R2_BUCKET}/model-catalog.json" \
+            --endpoint-url "$R2_ENDPOINT_URL" \
+            --content-type application/json \
+            --cache-control 'public, max-age=3600, must-revalidate'

packages/ai/.changes/hosted-model-catalog.md

@@ -0,0 +1 @@
+- Added a validated aggregate model catalog for daily publication from live provider catalogs.

packages/ai/package.json

@@ -65,6 +65,7 @@
 	"scripts": {
 		"clean": "shx rm -rf dist",
 		"generate-models": "npx tsx scripts/generate-models.ts",
+		"validate-model-catalog": "npx tsx scripts/validate-model-catalog.ts",
 		"build": "npm run generate-models && tsgo -p tsconfig.build.json",
 		"dev": "tsgo -p tsconfig.build.json --watch --preserveWatchOutput",
 		"dev:tsc": "tsgo -p tsconfig.build.json --watch --preserveWatchOutput",

packages/ai/scripts/generate-models.ts

@@ -6,6 +6,7 @@ import { dirname, join } from "path";
 import { fileURLToPath } from "url";
 import { getAnthropicCacheCosts } from "../src/cache-pricing.js";
 import { getOpenRouterReasoningCapabilities } from "../src/openrouter-reasoning.js";
+import { createModelCatalog } from "../src/model-catalog.js";
 import {
 	CLOUDFLARE_AI_GATEWAY_ANTHROPIC_BASE_URL,
 	CLOUDFLARE_AI_GATEWAY_COMPAT_BASE_URL,
@@ -76,6 +77,7 @@ const KIMI_STATIC_HEADERS = {
 
 const AI_GATEWAY_MODELS_URL = "https://ai-gateway.vercel.sh/v1";
 const AI_GATEWAY_BASE_URL = "https://ai-gateway.vercel.sh";
+const STRICT_MODEL_CATALOG_REFRESH = process.env.PRIME_AGENT_MODEL_CATALOG_STRICT === "1";
 const ZAI_TOOL_STREAM_UNSUPPORTED_MODELS = new Set(["glm-4.5", "glm-4.5-air", "glm-4.5-flash", "glm-4.5v"]);
 const EAGER_TOOL_INPUT_STREAMING_UNSUPPORTED_ANTHROPIC_MODELS = new Set([
 	"github-copilot:claude-haiku-4.5",
@@ -651,20 +653,25 @@ async function fetchPrimeInferenceModels(): Promise<Model<"openai-completions">[
 		const response = await fetch(`${PRIME_INFERENCE_BASE_URL}/models`, {
 			headers: getPrimeInferenceHeaders(apiKey, teamId),
 		});
+		if (!response.ok) throw new Error(`Prime Inference catalog request failed with status ${response.status}`);
 		catalog = parsePrimeInferenceCatalog(await response.json());
+		if (catalog.length === 0) throw new Error("Prime Inference catalog is empty or invalid");
 	} catch (error) {
 		console.error("Failed to fetch Prime Inference models:", error);
+		if (STRICT_MODEL_CATALOG_REFRESH) throw error;
 	}
 
 	let openRouterIndex = new Map<string, PrimeInferenceOpenRouterMetadata>();
 	try {
 		openRouterIndex = buildPrimeInferenceOpenRouterIndex(await fetchOpenRouterCatalog());
 	} catch (error) {
 		console.error("Failed to fetch OpenRouter catalog for Prime Inference metadata:", error);
+		if (STRICT_MODEL_CATALOG_REFRESH) throw error;
 	}
 	if (openRouterIndex.size === 0) {
 		// Without OpenRouter metadata every model would regress to the defaults;
 		// keep the previous snapshot instead.
+		if (STRICT_MODEL_CATALOG_REFRESH) throw new Error("OpenRouter catalog has no Prime Inference metadata");
 		console.error("OpenRouter catalog unavailable; keeping snapshot Prime Inference models");
 		return getExistingPrimeInferenceModels();
 	}
@@ -743,8 +750,10 @@ function fetchOpenRouterCatalog(): Promise<any[]> {
 	openRouterCatalogPromise ??= (async () => {
 		console.log("Fetching models from OpenRouter API...");
 		const response = await fetch("https://openrouter.ai/api/v1/models");
+		if (!response.ok) throw new Error(`OpenRouter catalog request failed with status ${response.status}`);
 		const data = await response.json();
-		return Array.isArray(data?.data) ? data.data : [];
+		if (!Array.isArray(data?.data) || data.data.length === 0) throw new Error("OpenRouter catalog is empty or invalid");
+		return data.data;
 	})();
 	return openRouterCatalogPromise;
 }
@@ -805,10 +814,12 @@ async function fetchOpenRouterModels(): Promise<Model<any>[]> {
 			models.push(normalizedModel);
 		}
 
+		if (models.length === 0) throw new Error("OpenRouter catalog has no tool-capable models");
 		console.log(`Fetched ${models.length} tool-capable models from OpenRouter`);
 		return models;
 	} catch (error) {
 		console.error("Failed to fetch OpenRouter models:", error);
+		if (STRICT_MODEL_CATALOG_REFRESH) throw error;
 		return [];
 	}
 }
@@ -817,7 +828,9 @@ async function fetchAiGatewayModels(): Promise<Model<any>[]> {
 	try {
 		console.log("Fetching models from Vercel AI Gateway API...");
 		const response = await fetch(`${AI_GATEWAY_MODELS_URL}/models`);
+		if (!response.ok) throw new Error(`Vercel AI Gateway catalog request failed with status ${response.status}`);
 		const data = await response.json();
+		if (!Array.isArray(data?.data)) throw new Error("Vercel AI Gateway catalog is invalid");
 		const models: Model<any>[] = [];
 
 		const toNumber = (value: string | number | undefined): number => {
@@ -864,10 +877,12 @@ async function fetchAiGatewayModels(): Promise<Model<any>[]> {
 			});
 		}
 
+		if (models.length === 0) throw new Error("Vercel AI Gateway catalog has no tool-capable models");
 		console.log(`Fetched ${models.length} tool-capable models from Vercel AI Gateway`);
 		return models;
 	} catch (error) {
 		console.error("Failed to fetch Vercel AI Gateway models:", error);
+		if (STRICT_MODEL_CATALOG_REFRESH) throw error;
 		return [];
 	}
 }
@@ -876,7 +891,9 @@ async function loadModelsDevData(): Promise<Model<any>[]> {
 	try {
 		console.log("Fetching models from models.dev API...");
 		const response = await fetch("https://models.dev/api.json");
+		if (!response.ok) throw new Error(`models.dev catalog request failed with status ${response.status}`);
 		const data = await response.json();
+		if (!isRecord(data)) throw new Error("models.dev catalog is invalid");
 
 		const models: Model<any>[] = [];
 
@@ -1567,10 +1584,12 @@ async function loadModelsDevData(): Promise<Model<any>[]> {
 			}
 		}
 
+		if (models.length === 0) throw new Error("models.dev catalog has no tool-capable models");
 		console.log(`Loaded ${models.length} tool-capable models from models.dev`);
 		return models;
 	} catch (error) {
 		console.error("Failed to load models.dev data:", error);
+		if (STRICT_MODEL_CATALOG_REFRESH) throw error;
 		return [];
 	}
 }
@@ -2433,6 +2452,17 @@ export const MODELS = {
 	writeFileSync(join(packageRoot, "src/models.generated.ts"), output);
 	console.log("Generated src/models.generated.ts");
 
+	const catalogOutputPath = process.env.PRIME_AGENT_MODEL_CATALOG_OUTPUT?.trim();
+	if (catalogOutputPath) {
+		const uniqueModels = sortedProviderIds.flatMap((providerId) =>
+			Object.keys(providers[providerId])
+				.sort()
+				.map((modelId) => providers[providerId][modelId]),
+		);
+		writeFileSync(catalogOutputPath, `${JSON.stringify(createModelCatalog(uniqueModels), null, 2)}\n`);
+		console.log(`Generated ${catalogOutputPath}`);
+	}
+
 	// Print statistics
 	const totalModels = allModels.length;
 	const reasoningModels = allModels.filter(m => m.reasoning).length;
@@ -2447,4 +2477,7 @@ export const MODELS = {
 }
 
 // Run the generator
-generateModels().catch(console.error);
+generateModels().catch((error) => {
+	console.error(error);
+	process.exitCode = 1;
+});

packages/ai/scripts/validate-model-catalog.ts

@@ -0,0 +1,9 @@
+#!/usr/bin/env tsx
+
+import { readFileSync } from "node:fs";
+import { parseModelCatalog } from "../src/model-catalog.js";
+
+const [candidatePath] = process.argv.slice(2);
+if (!candidatePath) throw new Error("Usage: validate-model-catalog.ts <candidate-path>");
+const candidate = parseModelCatalog(JSON.parse(readFileSync(candidatePath, "utf8")) as unknown);
+console.log(`Validated ${candidate.models.length} models`);

packages/ai/src/index.ts

@@ -4,6 +4,8 @@ export { Type } from "typebox";
 export * from "./api-registry.js";
 export * from "./env-api-keys.js";
 export * from "./log.js";
+export * from "./model-catalog.js";
+export { ProviderCompatSchema } from "./model-compat-schema.js";
 export * from "./models.js";
 export type { BedrockOptions, BedrockThinkingDisplay } from "./providers/amazon-bedrock.js";
 export type { AnthropicEffort, AnthropicOptions, AnthropicThinkingDisplay } from "./providers/anthropic.js";

packages/ai/src/model-catalog.ts

@@ -0,0 +1,105 @@
+import { type TProperties, Type } from "typebox";
+import { Value } from "typebox/value";
+import { isModelCompat } from "./model-compat-schema.js";
+import type { Api, Model } from "./types.js";
+
+const MODEL_CATALOG_SCHEMA_VERSION = 1 as const;
+const MAX_MODEL_CATALOG_MODELS = 20_000;
+
+export interface ModelCatalogV1 {
+	schemaVersion: typeof MODEL_CATALOG_SCHEMA_VERSION;
+	generatedAt: string;
+	models: Model<Api>[];
+}
+
+function strictObject<T extends TProperties>(properties: T) {
+	return Type.Object(properties, { additionalProperties: false });
+}
+
+const ThinkingLevelValueSchema = Type.Union([Type.String({ minLength: 1, maxLength: 128 }), Type.Null()]);
+const ThinkingLevelMapSchema = strictObject({
+	off: Type.Optional(ThinkingLevelValueSchema),
+	minimal: Type.Optional(ThinkingLevelValueSchema),
+	low: Type.Optional(ThinkingLevelValueSchema),
+	medium: Type.Optional(ThinkingLevelValueSchema),
+	high: Type.Optional(ThinkingLevelValueSchema),
+	xhigh: Type.Optional(ThinkingLevelValueSchema),
+	max: Type.Optional(ThinkingLevelValueSchema),
+});
+const CostSchema = Type.Number({ minimum: 0, maximum: 1_000_000 });
+const CatalogModelSchema = strictObject({
+	id: Type.String({ minLength: 1, maxLength: 1_024 }),
+	name: Type.String({ minLength: 1, maxLength: 1_024 }),
+	api: Type.String({ minLength: 1, maxLength: 128 }),
+	provider: Type.String({ minLength: 1, maxLength: 128 }),
+	baseUrl: Type.String({ maxLength: 2_048 }),
+	reasoning: Type.Boolean(),
+	thinkingLevelMap: Type.Optional(ThinkingLevelMapSchema),
+	input: Type.Array(Type.Union([Type.Literal("text"), Type.Literal("image")]), { minItems: 1, maxItems: 2 }),
+	cost: Type.Object({
+		input: CostSchema,
+		output: CostSchema,
+		cacheRead: CostSchema,
+		cacheWrite: CostSchema,
+	}),
+	contextWindow: Type.Integer({ minimum: 1, maximum: 100_000_000 }),
+	maxTokens: Type.Integer({ minimum: 1, maximum: 100_000_000 }),
+	featured: Type.Optional(Type.Boolean()),
+	compat: Type.Optional(Type.Unknown()),
+});
+const ModelCatalogEnvelopeSchema = Type.Object({
+	schemaVersion: Type.Literal(MODEL_CATALOG_SCHEMA_VERSION),
+	generatedAt: Type.String(),
+	models: Type.Array(Type.Unknown(), { minItems: 1, maxItems: MAX_MODEL_CATALOG_MODELS }),
+});
+
+function isRecord(value: unknown): value is Record<string, unknown> {
+	return typeof value === "object" && value !== null && !Array.isArray(value);
+}
+
+export function createModelCatalog(models: readonly Model<Api>[], generatedAt = new Date()): ModelCatalogV1 {
+	return {
+		schemaVersion: MODEL_CATALOG_SCHEMA_VERSION,
+		generatedAt: generatedAt.toISOString(),
+		models: models
+			.map((model) => {
+				const catalogModel = structuredClone(model);
+				delete catalogModel.headers;
+				return catalogModel;
+			})
+			.sort((left, right) => left.provider.localeCompare(right.provider) || left.id.localeCompare(right.id)),
+	};
+}
+
+export function parseModelCatalog(value: unknown, options: { skipInvalidModels?: boolean } = {}): ModelCatalogV1 {
+	if (!isRecord(value) || value.schemaVersion !== MODEL_CATALOG_SCHEMA_VERSION) {
+		throw new Error("Unsupported model catalog schema version");
+	}
+	if (typeof value.generatedAt !== "string" || !Number.isFinite(Date.parse(value.generatedAt))) {
+		throw new Error("Invalid model catalog timestamp");
+	}
+	if (!Array.isArray(value.models) || value.models.length === 0 || value.models.length > MAX_MODEL_CATALOG_MODELS) {
+		throw new Error("Invalid model catalog model count");
+	}
+	if (!Value.Check(ModelCatalogEnvelopeSchema, value)) throw new Error("Invalid model catalog entry");
+
+	const models: Model<Api>[] = [];
+	const seen = new Set<string>();
+	for (const candidate of value.models) {
+		if (!Value.Check(CatalogModelSchema, candidate)) {
+			if (options.skipInvalidModels) continue;
+			throw new Error("Invalid model catalog entry");
+		}
+		const model = candidate as Model<Api>;
+		if (!isModelCompat(model.api, model.compat)) {
+			if (options.skipInvalidModels) continue;
+			throw new Error("Invalid model catalog entry");
+		}
+		const key = JSON.stringify([model.provider, model.id]);
+		if (seen.has(key)) throw new Error(`Duplicate model catalog entry ${key}`);
+		seen.add(key);
+		models.push(model);
+	}
+	if (models.length === 0) throw new Error("Model catalog has no compatible entries");
+	return { schemaVersion: MODEL_CATALOG_SCHEMA_VERSION, generatedAt: value.generatedAt, models };
+}