cartethyia 2.5.2 → 2.5.4

package/src/protocol/response/gemini.ts

@@ -63,11 +63,25 @@ export function responseParts(parts: readonly Record<string, unknown>[]): {
 
 export function mapGeminiUsage(body: Record<string, unknown>): Record<string, unknown> {
   const usage = isRecord(body["usageMetadata"]) ? (body["usageMetadata"] as Record<string, unknown>) : {};
+  const prompt = nullableNumber(usage["promptTokenCount"]) ?? 0;
+  const candidates = nullableNumber(usage["candidatesTokenCount"]) ?? 0;
+  // Reasoning is billed output that Gemini reports in its own counter:
+  // `candidatesTokenCount` excludes `thoughtsTokenCount`, so dropping it here
+  // silently under-reported every thinking model's completion tokens. It is
+  // surfaced through `output_tokens_details` because that is the slot
+  // `normalizeUsage` reads for reasoning — keeping it separate from the
+  // candidate count it is added into, so downstream can still split them.
+  const thoughts = nullableNumber(usage["thoughtsTokenCount"]) ?? 0;
+  const output = candidates + thoughts;
+  const cached = nullableNumber(usage["cachedContentTokenCount"]) ?? 0;
   return {
-    input_tokens: nullableNumber(usage["promptTokenCount"]) ?? 0,
-    output_tokens: nullableNumber(usage["candidatesTokenCount"]) ?? 0,
-    total_tokens: nullableNumber(usage["totalTokenCount"]) ?? 0,
-    cached_tokens: nullableNumber(usage["cachedContentTokenCount"]) ?? 0,
+    input_tokens: prompt,
+    output_tokens: output,
+    total_tokens: nullableNumber(usage["totalTokenCount"]) ?? prompt + output,
+    cached_tokens: cached,
+    ...(thoughts > 0
+      ? { output_tokens_details: { reasoning_tokens: thoughts } }
+      : {}),
   };
 }
 

package/src/providers/adapters/compatible-adapter.ts

@@ -207,6 +207,9 @@ export class OpenAICompatibleAdapter extends BaseProviderAdapter {
       ...(this.config.chat_token_limit_field === undefined
         ? {}
         : { chatTokenLimitField: this.config.chat_token_limit_field }),
+      ...(context.image_clamp_max_bytes === undefined
+        ? {}
+        : { maxBytesOverride: context.image_clamp_max_bytes }),
     });
     if (
       wireFamily !== "messages" &&

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

@@ -1,119 +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",
-  "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,
-  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),
-    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/configured-openai-providers.ts

@@ -1,135 +1,137 @@
-import type { ApiKeyProviderSpec } from "./configured-provider";
-import type { ProviderId } from "./provider-contracts";
-
-/**
- * Builtin providers whose entire adapter contract is "bearer auth against the
- * provider's `BUNDLED_PROVIDER_METADATA` base URL, OpenAI-compatible chat +
- * responses, `/v1`-prefixed paths". They carry no catalog, no OAuth, and no
- * provider-specific payload hook, so they share one spec shape rather than
- * one module each.
- *
- * `base_url` is resolved per provider by `providerBaseUrl(provider_id)`.
- *
- * These providers list their models through live `/v1/models` discovery rather
- * than a static catalog: every one of them republishes an aggregator's whole
- * roster, which changes faster than a committed array can. Each registry row
- * pairs this spec with `openAIModelDiscovery`.
- */
-export const GENERIC_API_KEY_PROVIDER_IDS = [
-  "mistral",
-  "fireworks",
-  "nvidia",
-  "gmi",
-  "ollamacloud",
-  "together",
-  "siliconflow",
-  "siliconflowcn",
-  "nanogpt",
-  "baseten",
-  "synthetic",
-  "stepfun",
-  "sakana",
-  "helmcode",
-  "coralbricks",
-  "aiand",
-  "wafer",
-  "coreweave",
-  "qwen-portal",
-  "dashscope",
-  // `deepseek` is intentionally absent: it dispatches on the Anthropic Messages
-  // wire at its `/anthropic` root, not the OpenAI chat/responses wires, so it
-  // carries a dedicated entry in `default-registry.ts` instead.
-  //
-  // The aggregators below are absent for a different reason: their base URL is
-  // already rooted at a version segment that the `/v1`-prefixed path this shape
-  // pins would duplicate (`/openai/v1`, `/v1/openai`, `/v3/openai`, `/api/v1`,
-  // `/v2`, and `/inference/v1`), so `joinUrl` would produce
-  // `…/v1/v1/chat/completions`. They use `versionedRootApiKeySpec` instead,
-  // which pins the path below the version root — the same reason `cerebras`
-  // declares `/chat/completions` explicitly.
-] as const;
-
-/**
- * Ids whose base URL already ends at a version segment.
- *
- * `joinUrl` is plain concatenation and nothing rewrites a declared path, so a
- * base already rooted at a version plus a `/v1/…` path doubles the segment and
- * every request 404s. These specs therefore declare the path *below* the
- * version root. That is the provider stating the truth about its own wire, not
- * a correction applied to it: a wrong path now reaches the operator as the
- * wrong URL instead of being quietly rewritten into a different address.
- */
-export const VERSIONED_ROOT_API_KEY_PROVIDER_IDS = [
-  "groq",
-  "deepinfra",
-  "novita",
-  "venice",
-  "qianfan",
-  "zhipucoding",
-  "firepass",
-] as const;
-
-export type VersionedRootApiKeyProviderId =
-  (typeof VERSIONED_ROOT_API_KEY_PROVIDER_IDS)[number];
-
-function versionedRootApiKeySpec(providerId: ProviderId): ApiKeyProviderSpec {
-  return {
-    provider_id: providerId,
-    endpoint_paths_by_wire_family: {
-      chat: "/chat/completions",
-      responses: "/responses",
-    },
-  };
-}
-
-export const VERSIONED_ROOT_API_KEY_SPECS = {
-  groq: versionedRootApiKeySpec("groq"),
-  deepinfra: versionedRootApiKeySpec("deepinfra"),
-  novita: versionedRootApiKeySpec("novita"),
-  venice: versionedRootApiKeySpec("venice"),
-  qianfan: versionedRootApiKeySpec("qianfan"),
-  zhipucoding: versionedRootApiKeySpec("zhipucoding"),
-  firepass: versionedRootApiKeySpec("firepass"),
-} satisfies Readonly<Record<VersionedRootApiKeyProviderId, ApiKeyProviderSpec>>;
-
-/**
- * One id that has both a generic spec and a registry row.
- *
- * Exported so the registry can be checked against the spec list rather than
- * repeating the id as a string: `discoveryBackedProvider` takes this type, so a
- * row naming a provider with no spec is a compile error instead of an
- * `undefined` adapter at first dispatch.
- */
-export type GenericApiKeyProviderId = (typeof GENERIC_API_KEY_PROVIDER_IDS)[number];
-
-function genericApiKeySpec(providerId: GenericApiKeyProviderId): ApiKeyProviderSpec {
-  // These roots are *not* bare hosts — every one of them already ends at `/v1`
-  // — so the path is declared below that root. It used to be pinned from
-  // `BARE_ROOT_ENDPOINTS` and silently corrected at dispatch, which meant the
-  // declared path was wrong for all of them and nothing said so. Both spec
-  // shapes now declare the same thing, so this is the one implementation.
-  return versionedRootApiKeySpec(providerId);
-}
-
-export const GENERIC_API_KEY_SPECS = {
-  mistral: genericApiKeySpec("mistral"),
-  fireworks: genericApiKeySpec("fireworks"),
-  nvidia: genericApiKeySpec("nvidia"),
-  gmi: genericApiKeySpec("gmi"),
-  ollamacloud: genericApiKeySpec("ollamacloud"),
-  together: genericApiKeySpec("together"),
-  siliconflow: genericApiKeySpec("siliconflow"),
-  siliconflowcn: genericApiKeySpec("siliconflowcn"),
-  nanogpt: genericApiKeySpec("nanogpt"),
-  baseten: genericApiKeySpec("baseten"),
-  synthetic: genericApiKeySpec("synthetic"),
-  stepfun: genericApiKeySpec("stepfun"),
-  sakana: genericApiKeySpec("sakana"),
-  helmcode: genericApiKeySpec("helmcode"),
-  coralbricks: genericApiKeySpec("coralbricks"),
-  aiand: genericApiKeySpec("aiand"),
-  wafer: genericApiKeySpec("wafer"),
-  coreweave: genericApiKeySpec("coreweave"),
-  dashscope: genericApiKeySpec("dashscope"),
-  "qwen-portal": genericApiKeySpec("qwen-portal"),
+import type { ApiKeyProviderSpec } from "./configured-provider";
+import type { ProviderId } from "./provider-contracts";
+
+/**
+ * Builtin providers whose entire adapter contract is "bearer auth against the
+ * provider's `BUNDLED_PROVIDER_METADATA` base URL, OpenAI-compatible chat +
+ * responses, `/v1`-prefixed paths". They carry no catalog, no OAuth, and no
+ * provider-specific payload hook, so they share one spec shape rather than
+ * one module each.
+ *
+ * `base_url` is resolved per provider by `providerBaseUrl(provider_id)`.
+ *
+ * These providers list their models through live `/v1/models` discovery rather
+ * than a static catalog: every one of them republishes an aggregator's whole
+ * roster, which changes faster than a committed array can. Each registry row
+ * pairs this spec with `openAIModelDiscovery`.
+ */
+export const GENERIC_API_KEY_PROVIDER_IDS = [
+  "larprouter",
+  "mistral",
+  "fireworks",
+  "nvidia",
+  "gmi",
+  "ollamacloud",
+  "together",
+  "siliconflow",
+  "siliconflowcn",
+  "nanogpt",
+  "baseten",
+  "synthetic",
+  "stepfun",
+  "sakana",
+  "helmcode",
+  "coralbricks",
+  "aiand",
+  "wafer",
+  "coreweave",
+  "qwen-portal",
+  "dashscope",
+  // `deepseek` is intentionally absent: it dispatches on the Anthropic Messages
+  // wire at its `/anthropic` root, not the OpenAI chat/responses wires, so it
+  // carries a dedicated entry in `default-registry.ts` instead.
+  //
+  // The aggregators below are absent for a different reason: their base URL is
+  // already rooted at a version segment that the `/v1`-prefixed path this shape
+  // pins would duplicate (`/openai/v1`, `/v1/openai`, `/v3/openai`, `/api/v1`,
+  // `/v2`, and `/inference/v1`), so `joinUrl` would produce
+  // `…/v1/v1/chat/completions`. They use `versionedRootApiKeySpec` instead,
+  // which pins the path below the version root — the same reason `cerebras`
+  // declares `/chat/completions` explicitly.
+] as const;
+
+/**
+ * Ids whose base URL already ends at a version segment.
+ *
+ * `joinUrl` is plain concatenation and nothing rewrites a declared path, so a
+ * base already rooted at a version plus a `/v1/…` path doubles the segment and
+ * every request 404s. These specs therefore declare the path *below* the
+ * version root. That is the provider stating the truth about its own wire, not
+ * a correction applied to it: a wrong path now reaches the operator as the
+ * wrong URL instead of being quietly rewritten into a different address.
+ */
+export const VERSIONED_ROOT_API_KEY_PROVIDER_IDS = [
+  "groq",
+  "deepinfra",
+  "novita",
+  "venice",
+  "qianfan",
+  "zhipucoding",
+  "firepass",
+] as const;
+
+export type VersionedRootApiKeyProviderId =
+  (typeof VERSIONED_ROOT_API_KEY_PROVIDER_IDS)[number];
+
+function versionedRootApiKeySpec(providerId: ProviderId): ApiKeyProviderSpec {
+  return {
+    provider_id: providerId,
+    endpoint_paths_by_wire_family: {
+      chat: "/chat/completions",
+      responses: "/responses",
+    },
+  };
+}
+
+export const VERSIONED_ROOT_API_KEY_SPECS = {
+  groq: versionedRootApiKeySpec("groq"),
+  deepinfra: versionedRootApiKeySpec("deepinfra"),
+  novita: versionedRootApiKeySpec("novita"),
+  venice: versionedRootApiKeySpec("venice"),
+  qianfan: versionedRootApiKeySpec("qianfan"),
+  zhipucoding: versionedRootApiKeySpec("zhipucoding"),
+  firepass: versionedRootApiKeySpec("firepass"),
+} satisfies Readonly<Record<VersionedRootApiKeyProviderId, ApiKeyProviderSpec>>;
+
+/**
+ * One id that has both a generic spec and a registry row.
+ *
+ * Exported so the registry can be checked against the spec list rather than
+ * repeating the id as a string: `discoveryBackedProvider` takes this type, so a
+ * row naming a provider with no spec is a compile error instead of an
+ * `undefined` adapter at first dispatch.
+ */
+export type GenericApiKeyProviderId = (typeof GENERIC_API_KEY_PROVIDER_IDS)[number];
+
+function genericApiKeySpec(providerId: GenericApiKeyProviderId): ApiKeyProviderSpec {
+  // These roots are *not* bare hosts — every one of them already ends at `/v1`
+  // — so the path is declared below that root. It used to be pinned from
+  // `BARE_ROOT_ENDPOINTS` and silently corrected at dispatch, which meant the
+  // declared path was wrong for all of them and nothing said so. Both spec
+  // shapes now declare the same thing, so this is the one implementation.
+  return versionedRootApiKeySpec(providerId);
+}
+
+export const GENERIC_API_KEY_SPECS = {
+  larprouter: genericApiKeySpec("larprouter"),
+  mistral: genericApiKeySpec("mistral"),
+  fireworks: genericApiKeySpec("fireworks"),
+  nvidia: genericApiKeySpec("nvidia"),
+  gmi: genericApiKeySpec("gmi"),
+  ollamacloud: genericApiKeySpec("ollamacloud"),
+  together: genericApiKeySpec("together"),
+  siliconflow: genericApiKeySpec("siliconflow"),
+  siliconflowcn: genericApiKeySpec("siliconflowcn"),
+  nanogpt: genericApiKeySpec("nanogpt"),
+  baseten: genericApiKeySpec("baseten"),
+  synthetic: genericApiKeySpec("synthetic"),
+  stepfun: genericApiKeySpec("stepfun"),
+  sakana: genericApiKeySpec("sakana"),
+  helmcode: genericApiKeySpec("helmcode"),
+  coralbricks: genericApiKeySpec("coralbricks"),
+  aiand: genericApiKeySpec("aiand"),
+  wafer: genericApiKeySpec("wafer"),
+  coreweave: genericApiKeySpec("coreweave"),
+  dashscope: genericApiKeySpec("dashscope"),
+  "qwen-portal": genericApiKeySpec("qwen-portal"),
 } satisfies Readonly<Record<GenericApiKeyProviderId, ApiKeyProviderSpec>>;

package/src/providers/default-registry.ts

@@ -314,10 +314,7 @@ export const PROVIDER_CAPABILITIES = {
     loadAuthentication: oauthCapability(() => import("./integrations/cursor/cursor-auth"), "cursorOAuthClient", {
       withRefresher: true,
     }),
-    // No quota collector: no Cursor quota endpoint is established by the donor
-    // or the API evidence, and inventing one would publish a balance this
-    // gateway cannot measure.
-    //
+    loadQuotaCollector: quotaCapability(() => import("./integrations/cursor/cursor-quota"), "fetchCursorQuota"),
     // The roster is per-account and per-plan, and Cursor declares each model's
     // capabilities, so discovery is the catalog rather than a static list.
     loadModelDiscovery: async () => async ({ credential, fetcher }) =>
@@ -483,6 +480,7 @@ export const PROVIDER_CAPABILITIES = {
     // device flow, so the issued token is used exactly as issued until the
     // operator re-runs login — same shape as Kilo Code.
     loadAuthentication: oauthCapability(() => import("./integrations/qoder/qoder-oauth"), "qoderOAuthClient"),
+    loadQuotaCollector: quotaCapability(() => import("./integrations/qoder/qoder-quota"), "fetchQoderQuota"),
     // Discovery signed by the account's own COSY identity: the model list is
     // per-account and per-realm, and a GET without the signed body is 403.
     loadModelDiscovery: qoderDiscovery("intl", "qoder"),
@@ -498,6 +496,7 @@ export const PROVIDER_CAPABILITIES = {
     // `QODER_REALMS.cn` — its own login/token/userinfo hosts, client id, and
     // native redirect.
     loadAuthentication: oauthCapability(() => import("./integrations/qoder/qoder-oauth"), "qoderCnOAuthClient"),
+    loadQuotaCollector: quotaCapability(() => import("./integrations/qoder/qoder-quota"), "fetchQoderCnQuota"),
     loadModelDiscovery: qoderDiscovery("cn", "qodercn"),
   },
   trae: {
@@ -507,6 +506,7 @@ export const PROVIDER_CAPABILITIES = {
     // keeps it alive, so the refresher is registered rather than optional:
     // without it the account dies at first expiry.
     loadAuthentication: oauthCapability(() => import("./integrations/trae/trae-oauth"), "traeOAuthClient", { withRefresher: true }),
+    loadQuotaCollector: quotaCapability(() => import("./integrations/trae/trae-quota"), "fetchTraeQuota"),
     // Live model list is intl-unsupported: upstream answers `403 no schema
     // found`, so discovery returning nothing is the truthful result and the
     // static roster above stays authoritative.
@@ -518,6 +518,7 @@ export const PROVIDER_CAPABILITIES = {
     // The CN client, not the intl one: the exchange runs on the CN account
     // host, and a refresh posted to the intl host is refused.
     loadAuthentication: oauthCapability(() => import("./integrations/trae/trae-oauth"), "traeCnOAuthClient", { withRefresher: true }),
+    loadQuotaCollector: quotaCapability(() => import("./integrations/trae/trae-quota"), "fetchTraeCnQuota"),
     loadModelDiscovery: traeDiscovery("cn", "traecn"),
   },
   inferhub: {

package/src/providers/discovery/openai-model-discovery.ts

@@ -49,6 +49,11 @@ export async function fetchOpenAICompatibleModels(
       const parsedContext =
         positiveNumber(entry.context_length) ??
         positiveNumber((entry as Record<string, unknown>).input_token_limit) ??
+        // `context_window` is the spelling an OpenAI-shaped catalog uses for
+        // its own field. Unread, every model on such an upstream fell to the
+        // 200_000 default below and was advertised at the same wrong window no
+        // matter what the upstream reported — a 1M model shown as 200k.
+        positiveNumber((entry as Record<string, unknown>).context_window) ??
         positiveNumber(entry.contextWindow) ??
         positiveNumber(entry.max_input_tokens) ??
         modalityContextLimit(entry);