cartethyia 2.5.9 → 2.6.0

package/src/providers/bundled-discovery-providers.ts

@@ -1,122 +1,122 @@
-import type { ProviderModule } from "./provider-registry";
-import {
-  GENERIC_API_KEY_SPECS,
-  VERSIONED_ROOT_API_KEY_SPECS,
-} from "./configured-openai-providers";
-import { createApiKeyAdapter, type ApiKeyProviderSpec } from "./configured-provider";
-
-/**
- * Providers whose adapter and catalog discovery are the same bearer-auth
- * OpenAI-compatible contract. Providers with static catalogs or additional
- * upstream behavior remain declared beside their special handling.
- */
-export const DISCOVERY_BACKED_PROVIDER_IDS = [
-  "groq",
-  "together",
-  "deepinfra",
-  "siliconflow",
-  "siliconflowcn",
-  "novita",
-  "larprouter",
-  "venice",
-  "nanogpt",
-  "baseten",
-  "synthetic",
-  "stepfun",
-  "sakana",
-  "helmcode",
-  "coralbricks",
-  "aiand",
-  "wafer",
-  "firepass",
-  "zhipucoding",
-  "dashscope",
-  "qianfan",
-] as const;
-
-export type DiscoveryBackedProviderId = (typeof DISCOVERY_BACKED_PROVIDER_IDS)[number];
-
-type DiscoveryBackedCapabilities = Pick<
-  ProviderModule,
-  "endpointPathsByWireFamily" | "loadAdapter" | "loadModelDiscovery"
->;
-
-type ModelDiscoveryFactory = NonNullable<ProviderModule["loadModelDiscovery"]>;
-type OpenAIModelDiscoveryFactory = (
-  providerId: string,
-  options?: {
-    readonly headers?: (credential: string) => Readonly<Record<string, string>>;
-  },
-) => ModelDiscoveryFactory;
-
-const DISCOVERY_BACKED_PROVIDER_SPECS = {
-  groq: VERSIONED_ROOT_API_KEY_SPECS.groq,
-  larprouter: GENERIC_API_KEY_SPECS.larprouter,
-  together: GENERIC_API_KEY_SPECS.together,
-  deepinfra: VERSIONED_ROOT_API_KEY_SPECS.deepinfra,
-  siliconflow: GENERIC_API_KEY_SPECS.siliconflow,
-  siliconflowcn: GENERIC_API_KEY_SPECS.siliconflowcn,
-  novita: VERSIONED_ROOT_API_KEY_SPECS.novita,
-  venice: VERSIONED_ROOT_API_KEY_SPECS.venice,
-  nanogpt: GENERIC_API_KEY_SPECS.nanogpt,
-  baseten: GENERIC_API_KEY_SPECS.baseten,
-  synthetic: GENERIC_API_KEY_SPECS.synthetic,
-  stepfun: GENERIC_API_KEY_SPECS.stepfun,
-  sakana: GENERIC_API_KEY_SPECS.sakana,
-  helmcode: GENERIC_API_KEY_SPECS.helmcode,
-  coralbricks: GENERIC_API_KEY_SPECS.coralbricks,
-  aiand: GENERIC_API_KEY_SPECS.aiand,
-  wafer: GENERIC_API_KEY_SPECS.wafer,
-  firepass: VERSIONED_ROOT_API_KEY_SPECS.firepass,
-  zhipucoding: VERSIONED_ROOT_API_KEY_SPECS.zhipucoding,
-  dashscope: GENERIC_API_KEY_SPECS.dashscope,
-  qianfan: VERSIONED_ROOT_API_KEY_SPECS.qianfan,
-} satisfies Readonly<Record<DiscoveryBackedProviderId, ApiKeyProviderSpec>>;
-
-function discoveryBackedCapabilities(
-  providerId: DiscoveryBackedProviderId,
-  spec: ApiKeyProviderSpec,
-  openAIModelDiscovery: OpenAIModelDiscoveryFactory,
-): DiscoveryBackedCapabilities {
-  return {
-    ...(spec.endpoint_paths_by_wire_family === undefined
-      ? {}
-      : { endpointPathsByWireFamily: spec.endpoint_paths_by_wire_family }),
-    loadAdapter: async () => createApiKeyAdapter(spec),
-    loadModelDiscovery: openAIModelDiscovery(providerId, {
-      headers: (credential) => ({ authorization: `Bearer ${credential}` }),
-    }),
-  };
-}
-
-/**
- * Creates the standard provider capability table using the registry's shared
- * discovery loader, keeping its lazy import and authentication behavior intact.
- */
-export function createDiscoveryBackedProviderCapabilities(
-  openAIModelDiscovery: OpenAIModelDiscoveryFactory,
-): Readonly<Record<DiscoveryBackedProviderId, DiscoveryBackedCapabilities>> {
-  return {
-    groq: discoveryBackedCapabilities("groq", DISCOVERY_BACKED_PROVIDER_SPECS.groq, openAIModelDiscovery),
-    together: discoveryBackedCapabilities("together", DISCOVERY_BACKED_PROVIDER_SPECS.together, openAIModelDiscovery),
-    deepinfra: discoveryBackedCapabilities("deepinfra", DISCOVERY_BACKED_PROVIDER_SPECS.deepinfra, openAIModelDiscovery),
-    siliconflow: discoveryBackedCapabilities("siliconflow", DISCOVERY_BACKED_PROVIDER_SPECS.siliconflow, openAIModelDiscovery),
-    siliconflowcn: discoveryBackedCapabilities("siliconflowcn", DISCOVERY_BACKED_PROVIDER_SPECS.siliconflowcn, openAIModelDiscovery),
-    novita: discoveryBackedCapabilities("novita", DISCOVERY_BACKED_PROVIDER_SPECS.novita, openAIModelDiscovery),
-    larprouter: discoveryBackedCapabilities("larprouter", DISCOVERY_BACKED_PROVIDER_SPECS.larprouter, openAIModelDiscovery),
-    venice: discoveryBackedCapabilities("venice", DISCOVERY_BACKED_PROVIDER_SPECS.venice, openAIModelDiscovery),
-    nanogpt: discoveryBackedCapabilities("nanogpt", DISCOVERY_BACKED_PROVIDER_SPECS.nanogpt, openAIModelDiscovery),
-    baseten: discoveryBackedCapabilities("baseten", DISCOVERY_BACKED_PROVIDER_SPECS.baseten, openAIModelDiscovery),
-    synthetic: discoveryBackedCapabilities("synthetic", DISCOVERY_BACKED_PROVIDER_SPECS.synthetic, openAIModelDiscovery),
-    stepfun: discoveryBackedCapabilities("stepfun", DISCOVERY_BACKED_PROVIDER_SPECS.stepfun, openAIModelDiscovery),
-    sakana: discoveryBackedCapabilities("sakana", DISCOVERY_BACKED_PROVIDER_SPECS.sakana, openAIModelDiscovery),
-    helmcode: discoveryBackedCapabilities("helmcode", DISCOVERY_BACKED_PROVIDER_SPECS.helmcode, openAIModelDiscovery),
-    coralbricks: discoveryBackedCapabilities("coralbricks", DISCOVERY_BACKED_PROVIDER_SPECS.coralbricks, openAIModelDiscovery),
-    aiand: discoveryBackedCapabilities("aiand", DISCOVERY_BACKED_PROVIDER_SPECS.aiand, openAIModelDiscovery),
-    wafer: discoveryBackedCapabilities("wafer", DISCOVERY_BACKED_PROVIDER_SPECS.wafer, openAIModelDiscovery),
-    firepass: discoveryBackedCapabilities("firepass", DISCOVERY_BACKED_PROVIDER_SPECS.firepass, openAIModelDiscovery),
-    zhipucoding: discoveryBackedCapabilities("zhipucoding", DISCOVERY_BACKED_PROVIDER_SPECS.zhipucoding, openAIModelDiscovery),
-    dashscope: discoveryBackedCapabilities("dashscope", DISCOVERY_BACKED_PROVIDER_SPECS.dashscope, openAIModelDiscovery),
-    qianfan: discoveryBackedCapabilities("qianfan", DISCOVERY_BACKED_PROVIDER_SPECS.qianfan, openAIModelDiscovery),
-  };
-}
+import type { ProviderModule } from "./provider-registry";
+import {
+  GENERIC_API_KEY_SPECS,
+  VERSIONED_ROOT_API_KEY_SPECS,
+} from "./configured-openai-providers";
+import { createApiKeyAdapter, type ApiKeyProviderSpec } from "./configured-provider";
+
+/**
+ * Providers whose adapter and catalog discovery are the same bearer-auth
+ * OpenAI-compatible contract. Providers with static catalogs or additional
+ * upstream behavior remain declared beside their special handling.
+ */
+export const DISCOVERY_BACKED_PROVIDER_IDS = [
+  "groq",
+  "together",
+  "deepinfra",
+  "siliconflow",
+  "siliconflowcn",
+  "novita",
+  "larprouter",
+  "venice",
+  "nanogpt",
+  "baseten",
+  "synthetic",
+  "stepfun",
+  "sakana",
+  "helmcode",
+  "coralbricks",
+  "aiand",
+  "wafer",
+  "firepass",
+  "zhipucoding",
+  "dashscope",
+  "qianfan",
+] as const;
+
+export type DiscoveryBackedProviderId = (typeof DISCOVERY_BACKED_PROVIDER_IDS)[number];
+
+type DiscoveryBackedCapabilities = Pick<
+  ProviderModule,
+  "endpointPathsByWireFamily" | "loadAdapter" | "loadModelDiscovery"
+>;
+
+type ModelDiscoveryFactory = NonNullable<ProviderModule["loadModelDiscovery"]>;
+type OpenAIModelDiscoveryFactory = (
+  providerId: string,
+  options?: {
+    readonly headers?: (credential: string) => Readonly<Record<string, string>>;
+  },
+) => ModelDiscoveryFactory;
+
+const DISCOVERY_BACKED_PROVIDER_SPECS = {
+  groq: VERSIONED_ROOT_API_KEY_SPECS.groq,
+  larprouter: GENERIC_API_KEY_SPECS.larprouter,
+  together: GENERIC_API_KEY_SPECS.together,
+  deepinfra: VERSIONED_ROOT_API_KEY_SPECS.deepinfra,
+  siliconflow: GENERIC_API_KEY_SPECS.siliconflow,
+  siliconflowcn: GENERIC_API_KEY_SPECS.siliconflowcn,
+  novita: VERSIONED_ROOT_API_KEY_SPECS.novita,
+  venice: VERSIONED_ROOT_API_KEY_SPECS.venice,
+  nanogpt: GENERIC_API_KEY_SPECS.nanogpt,
+  baseten: GENERIC_API_KEY_SPECS.baseten,
+  synthetic: GENERIC_API_KEY_SPECS.synthetic,
+  stepfun: GENERIC_API_KEY_SPECS.stepfun,
+  sakana: GENERIC_API_KEY_SPECS.sakana,
+  helmcode: GENERIC_API_KEY_SPECS.helmcode,
+  coralbricks: GENERIC_API_KEY_SPECS.coralbricks,
+  aiand: GENERIC_API_KEY_SPECS.aiand,
+  wafer: GENERIC_API_KEY_SPECS.wafer,
+  firepass: VERSIONED_ROOT_API_KEY_SPECS.firepass,
+  zhipucoding: VERSIONED_ROOT_API_KEY_SPECS.zhipucoding,
+  dashscope: GENERIC_API_KEY_SPECS.dashscope,
+  qianfan: VERSIONED_ROOT_API_KEY_SPECS.qianfan,
+} satisfies Readonly<Record<DiscoveryBackedProviderId, ApiKeyProviderSpec>>;
+
+function discoveryBackedCapabilities(
+  providerId: DiscoveryBackedProviderId,
+  spec: ApiKeyProviderSpec,
+  openAIModelDiscovery: OpenAIModelDiscoveryFactory,
+): DiscoveryBackedCapabilities {
+  return {
+    ...(spec.endpoint_paths_by_wire_family === undefined
+      ? {}
+      : { endpointPathsByWireFamily: spec.endpoint_paths_by_wire_family }),
+    loadAdapter: async () => createApiKeyAdapter(spec),
+    loadModelDiscovery: openAIModelDiscovery(providerId, {
+      headers: (credential) => ({ authorization: `Bearer ${credential}` }),
+    }),
+  };
+}
+
+/**
+ * Creates the standard provider capability table using the registry's shared
+ * discovery loader, keeping its lazy import and authentication behavior intact.
+ */
+export function createDiscoveryBackedProviderCapabilities(
+  openAIModelDiscovery: OpenAIModelDiscoveryFactory,
+): Readonly<Record<DiscoveryBackedProviderId, DiscoveryBackedCapabilities>> {
+  return {
+    groq: discoveryBackedCapabilities("groq", DISCOVERY_BACKED_PROVIDER_SPECS.groq, openAIModelDiscovery),
+    together: discoveryBackedCapabilities("together", DISCOVERY_BACKED_PROVIDER_SPECS.together, openAIModelDiscovery),
+    deepinfra: discoveryBackedCapabilities("deepinfra", DISCOVERY_BACKED_PROVIDER_SPECS.deepinfra, openAIModelDiscovery),
+    siliconflow: discoveryBackedCapabilities("siliconflow", DISCOVERY_BACKED_PROVIDER_SPECS.siliconflow, openAIModelDiscovery),
+    siliconflowcn: discoveryBackedCapabilities("siliconflowcn", DISCOVERY_BACKED_PROVIDER_SPECS.siliconflowcn, openAIModelDiscovery),
+    novita: discoveryBackedCapabilities("novita", DISCOVERY_BACKED_PROVIDER_SPECS.novita, openAIModelDiscovery),
+    larprouter: discoveryBackedCapabilities("larprouter", DISCOVERY_BACKED_PROVIDER_SPECS.larprouter, openAIModelDiscovery),
+    venice: discoveryBackedCapabilities("venice", DISCOVERY_BACKED_PROVIDER_SPECS.venice, openAIModelDiscovery),
+    nanogpt: discoveryBackedCapabilities("nanogpt", DISCOVERY_BACKED_PROVIDER_SPECS.nanogpt, openAIModelDiscovery),
+    baseten: discoveryBackedCapabilities("baseten", DISCOVERY_BACKED_PROVIDER_SPECS.baseten, openAIModelDiscovery),
+    synthetic: discoveryBackedCapabilities("synthetic", DISCOVERY_BACKED_PROVIDER_SPECS.synthetic, openAIModelDiscovery),
+    stepfun: discoveryBackedCapabilities("stepfun", DISCOVERY_BACKED_PROVIDER_SPECS.stepfun, openAIModelDiscovery),
+    sakana: discoveryBackedCapabilities("sakana", DISCOVERY_BACKED_PROVIDER_SPECS.sakana, openAIModelDiscovery),
+    helmcode: discoveryBackedCapabilities("helmcode", DISCOVERY_BACKED_PROVIDER_SPECS.helmcode, openAIModelDiscovery),
+    coralbricks: discoveryBackedCapabilities("coralbricks", DISCOVERY_BACKED_PROVIDER_SPECS.coralbricks, openAIModelDiscovery),
+    aiand: discoveryBackedCapabilities("aiand", DISCOVERY_BACKED_PROVIDER_SPECS.aiand, openAIModelDiscovery),
+    wafer: discoveryBackedCapabilities("wafer", DISCOVERY_BACKED_PROVIDER_SPECS.wafer, openAIModelDiscovery),
+    firepass: discoveryBackedCapabilities("firepass", DISCOVERY_BACKED_PROVIDER_SPECS.firepass, openAIModelDiscovery),
+    zhipucoding: discoveryBackedCapabilities("zhipucoding", DISCOVERY_BACKED_PROVIDER_SPECS.zhipucoding, openAIModelDiscovery),
+    dashscope: discoveryBackedCapabilities("dashscope", DISCOVERY_BACKED_PROVIDER_SPECS.dashscope, openAIModelDiscovery),
+    qianfan: discoveryBackedCapabilities("qianfan", DISCOVERY_BACKED_PROVIDER_SPECS.qianfan, openAIModelDiscovery),
+  };
+}

package/src/providers/default-registry.ts

@@ -1,3 +1,7 @@
+// dynamic-import-ok: lazy provider registry. Every entry is a
+// `() => await import("./integrations/<name>/...")` so loading a provider
+// pays only when the first request routes to it. Static imports here would
+// drag every integration into memory at process start.
 import {
   GENERIC_API_KEY_SPECS,
 } from "./configured-openai-providers";
@@ -234,6 +238,11 @@ export const PROVIDER_CAPABILITIES = {
     loadModels: async () => (await import("./integrations/codex/codex")).CODEX_MODELS,
     loadAuthentication: oauthCapability(() => import("./integrations/codex/codex-oauth"), "codexOAuthClient", { withRefresher: true }),
     loadQuotaCollector: quotaCapability(() => import("./integrations/codex/codex-quota"), "fetchCodexQuota"),
+    loadModelDiscovery: async () => async ({ credential, fetcher }) =>
+      (await import("./integrations/codex/codex")).discoverCodexModels({
+        credential,
+        ...(fetcher === undefined ? {} : { fetcher }),
+      }),
   },
   grok: {
     loadAdapter: async () => (await import("./integrations/grok/grok")).grokAdapter,
@@ -523,7 +532,12 @@ export const PROVIDER_CAPABILITIES = {
   },
   inferhub: {
     loadAdapter: async () => (await import("./integrations/inferhub")).createInferhubAdapter(),
+    // Live `/v1/models` is the primary source: the marketplace republishes its
+    // whole roster and it moves faster than a committed array can. The static
+    // list stays as the offline fallback when the fetch fails.
     loadModels: async () => (await import("./integrations/inferhub")).INFERHUB_MODELS,
+    loadModelDiscovery: async () =>
+      (await import("./integrations/inferhub")).inferhubModelDiscovery()(),
     loadQuotaCollector: quotaCapability(() => import("./integrations/inferhub"), "fetchInferhubQuota"),
   },
   hermes: {