cartethyia 2.4.5 → 2.5.0

package/src/providers/integrations/openrouter/provider.ts

@@ -1,281 +1,139 @@
-import type { ApiKeyProviderSpec } from "../../configured-provider";
-import type { ModelDefinition } from "../../provider-model-definition";
-import { fetchOpenAICompatibleModels } from "../../discovery/openai-model-discovery";
-import { providerBaseUrl } from "../../provider-metadata";
-import type { DiscoveryInput } from "../../discovery/discovery-types";
-import { defineModel } from "../../provider-model-definition";
-
-const OPENROUTER_PROVIDER_ID = "openrouter" as const;
-
-/** OpenRouter requires these on every request for ranking/attribution. */
-const OPENROUTER_REQUIRED_HEADERS: Readonly<Record<string, string>> = {
-  "HTTP-Referer": "https://endpoint-proxy.local",
-  "X-Title": "Endpoint Proxy",
-};
-
-/**
- * OpenRouter's non-chat service models.
- *
- * OpenRouter fronts the System One decision API (`/api/v1/systemone`) as
- * `typesafe/jev-1.13`, which is not a chat model: it answers `{answers}` to a
- * `{state, questions}` body. `/api/v1/models` does not list it under a wire the
- * generic fetcher can classify, so it is declared here — a `serviceKind` row
- * served by the native System One route, never by the chat surface.
- */
-export const OPENROUTER_MODELS: readonly ModelDefinition[] = [
-  defineModel({
-    id: "typesafe/jev-1.13",
-    providerId: "openrouter",
-    serviceKind: "systemone",
-    wireFamily: "chat",
-    endpoint: "/systemone",
-    ctx: 200_000,
-    out: 8_192,
-    toolCall: false,
-  }),
-];
-
-
-/**
- * OpenRouter's `:free` tier, as a default catalog.
- *
- * OpenRouter runs live `/api/v1/models` discovery, so this is a floor rather
- * than the only source — but without it the provider lists nothing at all
- * until a discovery run happens to fetch it, and the free tier is the one part
- * of OpenRouter an operator can use with no spend.
- *
- * Fetched 2026-10-12 from `https://openrouter.ai/api/v1/models`, which served
- * 15 `:free` rows; the 14 that the base catalog also files under `openrouter`
- * are listed here. `apodex/apodex-1.1-mini:free` is absent because the base
- * catalog has no row for it under any provider, so its limits and price would
- * have to be guessed — and a published guess reads as a measurement. Discovery
- * still picks it up, which is the correct path for a row with no catalog data.
- *
- * Every row is marked `freeTier` as well as `free`, because the two answer
- * different questions: `free` says this bills zero per token, and `freeTier` is
- * what the model list groups on so the free roster is labelled as such rather
- * than mixed in with the billed catalog.
- */
-export const OPENROUTER_FREE_MODELS: readonly ModelDefinition[] = [
-  defineModel({
-    id: "cohere/north-mini-code:free",
-    providerId: OPENROUTER_PROVIDER_ID,
-    reasoning: true,
-    toolCall: true,
-    free: true,
-    freeTier: true,
-  }),
-  defineModel({
-    id: "dots-studio/dots-3-note-preview:free",
-    providerId: OPENROUTER_PROVIDER_ID,
-    reasoning: true,
-    toolCall: true,
-    free: true,
-    freeTier: true,
-  }),
-  defineModel({
-    id: "google/gemma-4-26b-a4b-it:free",
-    providerId: OPENROUTER_PROVIDER_ID,
-    reasoning: true,
-    toolCall: true,
-    free: true,
-    freeTier: true,
-  }),
-  defineModel({
-    id: "google/gemma-4-31b-it:free",
-    providerId: OPENROUTER_PROVIDER_ID,
-    reasoning: true,
-    toolCall: true,
-    free: true,
-    freeTier: true,
-  }),
-  defineModel({
-    id: "inclusionai/ling-3.0-flash-fin:free",
-    providerId: OPENROUTER_PROVIDER_ID,
-    reasoning: true,
-    toolCall: true,
-    free: true,
-    freeTier: true,
-  }),
-  defineModel({
-    id: "inclusionai/ling-3.0-flash-sante:free",
-    providerId: OPENROUTER_PROVIDER_ID,
-    reasoning: true,
-    toolCall: true,
-    free: true,
-    freeTier: true,
-  }),
-  defineModel({
-    id: "liquid/lfm-2.5-2.6b:free",
-    providerId: OPENROUTER_PROVIDER_ID,
-    reasoning: true,
-    toolCall: true,
-    free: true,
-    freeTier: true,
-  }),
-  defineModel({
-    id: "nex-agi/nex-n2.5-mini:free",
-    providerId: OPENROUTER_PROVIDER_ID,
-    reasoning: true,
-    toolCall: true,
-    free: true,
-    freeTier: true,
-  }),
-  defineModel({
-    id: "nex-agi/nex-n2.5-pro:free",
-    providerId: OPENROUTER_PROVIDER_ID,
-    reasoning: true,
-    toolCall: true,
-    free: true,
-    freeTier: true,
-  }),
-  defineModel({
-    id: "nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free",
-    providerId: OPENROUTER_PROVIDER_ID,
-    audio: true,
-    reasoning: true,
-    toolCall: true,
-    free: true,
-    freeTier: true,
-  }),
-  defineModel({
-    id: "nvidia/nemotron-3-super-120b-a12b:free",
-    providerId: OPENROUTER_PROVIDER_ID,
-    reasoning: true,
-    toolCall: true,
-    free: true,
-    freeTier: true,
-  }),
-  defineModel({
-    id: "nvidia/nemotron-3-ultra-550b-a55b:free",
-    providerId: OPENROUTER_PROVIDER_ID,
-    reasoning: true,
-    toolCall: true,
-    free: true,
-    freeTier: true,
-  }),
-  defineModel({
-    id: "nvidia/nemotron-3.5-content-safety:free",
-    providerId: OPENROUTER_PROVIDER_ID,
-    reasoning: true,
-    // A content-safety classifier, not a chat model: it is published on the
-    // same free tier and serves no tool calls.
-    toolCall: false,
-    free: true,
-    freeTier: true,
-  }),
-  defineModel({
-    id: "nvidia/nemotron-3.5-lightning:free",
-    providerId: OPENROUTER_PROVIDER_ID,
-    reasoning: true,
-    toolCall: true,
-    free: true,
-    freeTier: true,
-  }),
-  defineModel({
-    id: "poolside/laguna-s-2.1:free",
-    providerId: OPENROUTER_PROVIDER_ID,
-    reasoning: true,
-    toolCall: true,
-    free: true,
-    freeTier: true,
-  }),
-  defineModel({
-    id: "poolside/laguna-xs-2.1:free",
-    providerId: OPENROUTER_PROVIDER_ID,
-    reasoning: true,
-    toolCall: true,
-    free: true,
-    freeTier: true,
-  }),
-  defineModel({
-    id: "qwen/qwen3.8-27b:free",
-    providerId: OPENROUTER_PROVIDER_ID,
-    reasoning: true,
-    toolCall: true,
-    free: true,
-    freeTier: true,
-  }),
-  defineModel({
-    id: "thinkingmachines/inkling-small:free",
-    providerId: OPENROUTER_PROVIDER_ID,
-    reasoning: true,
-    toolCall: true,
-    free: true,
-    freeTier: true,
-  }),
-  defineModel({
-    id: "thinkingmachines/inkling:free",
-    providerId: OPENROUTER_PROVIDER_ID,
-    reasoning: true,
-    toolCall: true,
-    free: true,
-    freeTier: true,
-  }),
-  defineModel({
-    id: "z-ai/glm-5.2:free",
-    providerId: OPENROUTER_PROVIDER_ID,
-    reasoning: true,
-    toolCall: true,
-    free: true,
-    freeTier: true,
-  }),
-];
-
-
-/**
- * OpenRouter's `:free` rows, discovered live.
- *
- * OpenRouter rotates its free tier constantly — models are added and pulled
- * within days — so a committed list is only ever a snapshot of one afternoon.
- * This is the same shape as the OpenCode Zen free fetch: filter the provider's
- * own live listing down to the free rows and mark each one `freeTier`, which is
- * what the model list groups under "Free models (auto)".
- *
- * The rows are the fetcher's own definitions, spread rather than rebuilt: it
- * has already resolved limits, pricing, and the wire family from the id, and
- * re-deriving that here would discard the resolution and let the two drift.
- */
-export async function discoverOpenrouterFreeModels(
-  input: DiscoveryInput,
-): Promise<readonly ModelDefinition[] | null> {
-  const listed = await discoverOpenrouterModels(input);
-  if (listed === null) return null;
-  return (
-    listed
-      // OpenRouter's free-tier spelling is a `:free` id suffix, matched exactly
-      // rather than as a substring so a model whose own name contains "free" is
-      // not swept in by accident.
-      .filter((model) => model.modelId.endsWith(":free"))
-      .map((model) => ({
-        ...model,
-        // `free` is a billing statement and `freeTier` a catalog one. The
-        // `:free` suffix answers both, so both are set; the cost is left as the
-        // fetcher resolved it because OpenRouter publishes a real zero price.
-        free: true,
-        freeTier: true,
-      }))
-  );
-}
-
-
-/** OpenRouter adapter spec; the registry builds it via `createApiKeyAdapter`. */
-export const OPENROUTER_SPEC: ApiKeyProviderSpec = {
-  provider_id: OPENROUTER_PROVIDER_ID,
-  endpoint_paths_by_wire_family: {},
-  extra_headers: OPENROUTER_REQUIRED_HEADERS,
-};
-
-export async function discoverOpenrouterModels(
-  input: DiscoveryInput,
-): Promise<readonly ModelDefinition[] | null> {
-  return fetchOpenAICompatibleModels({
-    baseUrl: providerBaseUrl("openrouter"),
-    headers: {
-      authorization: `Bearer ${input.credential}`,
-      ...OPENROUTER_REQUIRED_HEADERS,
-    },
-    ...(input.signal ? { signal: input.signal } : {}),
-    ...(input.fetcher ? { fetcher: input.fetcher } : {}),
-  });
-}
+import type { ApiKeyProviderSpec } from "../../configured-provider";
+import type { ModelDefinition } from "../../provider-model-definition";
+import { fetchOpenAICompatibleModels } from "../../discovery/openai-model-discovery";
+import { providerBaseUrl } from "../../provider-metadata";
+import type { DiscoveryInput } from "../../discovery/discovery-types";
+import { defineModel } from "../../provider-model-definition";
+import { openrouterFreeRoster } from "../../discovery/openrouter-free-roster";
+
+const OPENROUTER_PROVIDER_ID = "openrouter" as const;
+
+/** OpenRouter requires these on every request for ranking/attribution. */
+const OPENROUTER_REQUIRED_HEADERS: Readonly<Record<string, string>> = {
+  "HTTP-Referer": "https://endpoint-proxy.local",
+  "X-Title": "Endpoint Proxy",
+};
+
+/**
+ * OpenRouter's non-chat service models.
+ *
+ * OpenRouter fronts the System One decision API (`/api/v1/systemone`) as
+ * `typesafe/jev-1.13`, which is not a chat model: it answers `{answers}` to a
+ * `{state, questions}` body. `/api/v1/models` does not list it under a wire the
+ * generic fetcher can classify, so it is declared here — a `serviceKind` row
+ * served by the native System One route, never by the chat surface.
+ */
+export const OPENROUTER_MODELS: readonly ModelDefinition[] = [
+  // The router, and only the router. A hand-kept catalog of specific ids was a
+  // snapshot of one afternoon: OpenRouter adds and pulls models within days, so
+  // every committed row outside this one drifts out of truth. Discovery is the
+  // authority, and the free roster below is the offline floor.
+  defineModel({
+    id: "openrouter/auto",
+    providerId: OPENROUTER_PROVIDER_ID,
+    wireFamily: "chat",
+    endpoint: "/chat/completions",
+    // The catalog publishes no limits for a router (pricing is -1, meaning
+    // "decided at dispatch"), so this is the largest window it may select into.
+    ctx: 2_000_000,
+    out: 128_000,
+    reasoning: true,
+  }),
+  // Not a chat row: it answers `{answers}` to a `{state, questions}` body, and
+  // is served by the native System One route. Kept because deleting it removes
+  // that route, which is a capability, not a stale catalog entry.
+  defineModel({
+    id: "typesafe/jev-1.13",
+    providerId: "openrouter",
+    serviceKind: "systemone",
+    wireFamily: "chat",
+    endpoint: "/systemone",
+    ctx: 200_000,
+    out: 8_192,
+    toolCall: false,
+  }),
+];
+
+
+/**
+ * OpenRouter's `:free` tier, as a default catalog.
+ *
+ * OpenRouter runs live `/api/v1/models` discovery, so this is a floor rather
+ * than the only source — but without it the provider lists nothing at all
+ * until a discovery run happens to fetch it, and the free tier is the one part
+ * of OpenRouter an operator can use with no spend.
+ *
+ * Fetched 2026-10-12 from `https://openrouter.ai/api/v1/models`, which served
+ * 15 `:free` rows; the 14 that the base catalog also files under `openrouter`
+ * are listed here. `apodex/apodex-1.1-mini:free` is absent because the base
+ * catalog has no row for it under any provider, so its limits and price would
+ * have to be guessed — and a published guess reads as a measurement. Discovery
+ * still picks it up, which is the correct path for a row with no catalog data.
+ *
+ * Every row is marked `freeTier` as well as `free`, because the two answer
+ * different questions: `free` says this bills zero per token, and `freeTier` is
+ * what the model list groups on so the free roster is labelled as such rather
+ * than mixed in with the billed catalog.
+ */
+export const OPENROUTER_FREE_MODELS: readonly ModelDefinition[] = openrouterFreeRoster({
+  providerId: OPENROUTER_PROVIDER_ID,
+  endpoint: "/chat/completions",
+});
+
+
+/**
+ * OpenRouter's `:free` rows, discovered live.
+ *
+ * OpenRouter rotates its free tier constantly — models are added and pulled
+ * within days — so a committed list is only ever a snapshot of one afternoon.
+ * This is the same shape as the OpenCode Zen free fetch: filter the provider's
+ * own live listing down to the free rows and mark each one `freeTier`, which is
+ * what the model list groups under "Free models (auto)".
+ *
+ * The rows are the fetcher's own definitions, spread rather than rebuilt: it
+ * has already resolved limits, pricing, and the wire family from the id, and
+ * re-deriving that here would discard the resolution and let the two drift.
+ */
+export async function discoverOpenrouterFreeModels(
+  input: DiscoveryInput,
+): Promise<readonly ModelDefinition[] | null> {
+  const listed = await discoverOpenrouterModels(input);
+  if (listed === null) return null;
+  return (
+    listed
+      // OpenRouter's free-tier spelling is a `:free` id suffix, matched exactly
+      // rather than as a substring so a model whose own name contains "free" is
+      // not swept in by accident.
+      .filter((model) => model.modelId.endsWith(":free"))
+      .map((model) => ({
+        ...model,
+        // `free` is a billing statement and `freeTier` a catalog one. The
+        // `:free` suffix answers both, so both are set; the cost is left as the
+        // fetcher resolved it because OpenRouter publishes a real zero price.
+        free: true,
+        freeTier: true,
+      }))
+  );
+}
+
+
+/** OpenRouter adapter spec; the registry builds it via `createApiKeyAdapter`. */
+export const OPENROUTER_SPEC: ApiKeyProviderSpec = {
+  provider_id: OPENROUTER_PROVIDER_ID,
+  endpoint_paths_by_wire_family: {},
+  extra_headers: OPENROUTER_REQUIRED_HEADERS,
+};
+
+export async function discoverOpenrouterModels(
+  input: DiscoveryInput,
+): Promise<readonly ModelDefinition[] | null> {
+  return fetchOpenAICompatibleModels({
+    baseUrl: providerBaseUrl("openrouter"),
+    headers: {
+      authorization: `Bearer ${input.credential}`,
+      ...OPENROUTER_REQUIRED_HEADERS,
+    },
+    ...(input.signal ? { signal: input.signal } : {}),
+    ...(input.fetcher ? { fetcher: input.fetcher } : {}),
+  });
+}