@warlock.js/ai-google 5.30.0 → 5.32.0

package/esm/sdk.d.mts

@@ -1,6 +1,5 @@
 import { GoogleEmbedderConfig, GoogleImageConfig, GoogleModelConfig, GoogleSDKConfig } from "./config.type.mjs";
 import { EmbedderContract, ImageModelContract, ModelContract, SDKAdapterContract } from "@warlock.js/ai";
-
 //#region ../../packages/ai-google/src/sdk.d.ts
 /**
  * Google Gemini-backed implementation of `SDKAdapterContract`.
@@ -26,7 +25,7 @@ import { EmbedderContract, ImageModelContract, ModelContract, SDKAdapterContract
  * const model = google.model({ name: "gemini-2.5-flash", temperature: 0.7 });
  * const embedder = google.embedder({ name: "gemini-embedding-001" });
  */
-declare class GoogleSDK implements SDKAdapterContract {
+export declare class GoogleSDK implements SDKAdapterContract {
   private readonly ai;
   private readonly provider;
   private readonly pricing?;
@@ -37,8 +36,9 @@ declare class GoogleSDK implements SDKAdapterContract {
    * `GoogleGenAI` client. The SDK's `provider` label is forwarded.
    *
    * Pricing resolution: per-model `config.pricing` wins; otherwise the
-   * SDK-level registry entry keyed by `config.name`; otherwise
-   * `undefined` (no cost computed).
+   * SDK-level registry entry keyed by `config.name`; otherwise the
+   * built-in list price (`GOOGLE_MODEL_PRICING`); otherwise `undefined`
+   * (no cost computed).
    */
   model(config: GoogleModelConfig): ModelContract;
   /**
@@ -82,5 +82,4 @@ declare class GoogleSDK implements SDKAdapterContract {
   image(config: GoogleImageConfig): ImageModelContract;
 }
 //#endregion
-export { GoogleSDK };
 //# sourceMappingURL=sdk.d.mts.map

package/esm/sdk.d.mts.map

@@ -1 +1 @@
-{"version":3,"file":"sdk.d.mts","names":[],"sources":["../../../../../../../packages/ai-google/src/sdk.ts"],"mappings":";;;;;;AA2EA;;;;;;;;;;;;;;;;;;;;;;cAAa,SAAA,YAAqB,kBAAA;EAAA,iBACf,EAAA;EAAA,iBACA,QAAA;EAAA,iBACA,OAAA;cAEE,MAAA,EAAQ,eAAA;EA+BM;;;;;;;;;EAd1B,KAAA,CAAM,MAAA,EAAQ,iBAAA,GAAoB,aAAA;EAoDkB;AAAA;;;;;EAtC9C,KAAA,CAAM,IAAA,UAAc,MAAA,YAAkB,OAAA;;;;;;;;EAW5C,QAAA,CAAS,MAAA,EAAQ,oBAAA,GAAuB,gBAAA;;;;;;;;;;;;;;;;;;;;;;;;EA2BxC,KAAA,CAAM,MAAA,EAAQ,iBAAA,GAAoB,kBAAA;AAAA"}
+{"version":3,"file":"sdk.d.mts","names":[],"sources":["../../../../../../../packages/ai-google/src/sdk.ts"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;qBA+Ea,qBAAqB;mBACf;mBACA;mBACA;EAEjB,YAAmB,QAAQ;;;;;;;;;;;EAkBpB,MAAM,QAAQ,oBAAoB;;;;;;;EAe5B,MAAM,cAAc,kBAAkB;;;;;;;;EAW5C,SAAS,QAAQ,uBAAuB;;;;;;;;;;;;;;;;;;;;;;;;EA2BxC,MAAM,QAAQ,oBAAoB"}

package/esm/sdk.mjs

@@ -1,6 +1,7 @@
 import { GoogleEmbedder } from "./embedder.mjs";
 import { GeminiImageModel } from "./gemini-image.mjs";
 import { GoogleImageModel } from "./image.mjs";
+import { googlePricingFor } from "./known-pricing.mjs";
 import { GoogleModel } from "./model.mjs";
 import { GoogleGenAI } from "@google/genai";
 import { approximateTokenCount } from "@warlock.js/ai";
@@ -73,11 +74,12 @@ var GoogleSDK = class {
 	* `GoogleGenAI` client. The SDK's `provider` label is forwarded.
 	*
 	* Pricing resolution: per-model `config.pricing` wins; otherwise the
-	* SDK-level registry entry keyed by `config.name`; otherwise
-	* `undefined` (no cost computed).
+	* SDK-level registry entry keyed by `config.name`; otherwise the
+	* built-in list price (`GOOGLE_MODEL_PRICING`); otherwise `undefined`
+	* (no cost computed).
 	*/
 	model(config) {
-		const resolvedPricing = config.pricing ?? this.pricing?.[config.name];
+		const resolvedPricing = config.pricing ?? this.pricing?.[config.name] ?? googlePricingFor(config.name);
 		const resolvedConfig = resolvedPricing === config.pricing ? config : {
 			...config,
 			pricing: resolvedPricing

package/esm/sdk.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"sdk.mjs","names":[],"sources":["../../../../../../../packages/ai-google/src/sdk.ts"],"sourcesContent":["import { GoogleGenAI } from \"@google/genai\";\nimport type {\n  EmbedderContract,\n  ImageModelContract,\n  ModelContract,\n  ModelPricing,\n  SDKAdapterContract,\n} from \"@warlock.js/ai\";\nimport { approximateTokenCount } from \"@warlock.js/ai\";\nimport type {\n  GoogleEmbedderConfig,\n  GoogleImageConfig,\n  GoogleModelConfig,\n  GoogleSDKConfig,\n} from \"./config.type\";\nimport { GoogleEmbedder } from \"./embedder\";\nimport { GeminiImageModel } from \"./gemini-image\";\nimport { GoogleImageModel } from \"./image\";\nimport { GoogleModel } from \"./model\";\n\n/**\n * Pick the transport for an image model id.\n *\n * `ai.models.generateImages` calls `{model}:predict`, and a `gemini-`\n * id sent there comes back `404 … is not supported for predict`\n * (observed verbatim from Google). `generateContent` is what the SDK\n * itself points `generateImages` users at — its deprecation notice\n * reads \"Please use the generateContent method with image models\n * instead\" — so the id has to choose the transport.\n *\n * Runs in this package establish where a `gemini-` id is ACCEPTED, not\n * what it returns: on this transport such an id got as far as a quota\n * error (HTTP 429) instead of the 404. That an image\n * comes back end-to-end once billing is enabled is reported by the\n * maintainer from a locally linked build, not measured here. Whether\n * these models report token usage is still unknown.\n *\n * This is ROUTING, not validation — no id is refused here. An id this\n * function does not recognize takes the `generateImages` route, the\n * only route that existed before Gemini image support landed, so every\n * id that reached Google before still reaches Google the same way and\n * still fails (or succeeds) at the provider.\n *\n * A leading `models/` resource prefix is tolerated, matching the id\n * shapes `inferVisionCapability` already accepts\n * (`models/gemini-1.5-flash-001`).\n */\nfunction usesGeminiImageTransport(name: string): boolean {\n  return name.toLowerCase().replace(/^models\\//, \"\").startsWith(\"gemini-\");\n}\n\n/**\n * Google Gemini-backed implementation of `SDKAdapterContract`.\n *\n * **Role.** The package entry point for Gemini models via the\n * `@google/genai` SDK. A single `GoogleSDK` holds one live\n * `GoogleGenAI` client, shared by every `ModelContract` /\n * `EmbedderContract` it produces. Construct one SDK per\n * account/project and reuse it everywhere.\n *\n * **Responsibility.**\n * - Owns: a long-lived `GoogleGenAI` client (auth, Vertex vs Gemini\n *   API) and its lifetime. Factory for `GoogleModel` /\n *   `GoogleEmbedder` instances sharing that client.\n * - Does NOT own: anything per-call — those live in `GoogleModel` /\n *   `GoogleEmbedder` and the agent runtime.\n *\n * Modeled as a class (see §4.2 of code-style.md — \"long-lived state\n * across many calls\"), fronted by FP usage like the other adapters.\n *\n * @example\n * const google = new GoogleSDK({ apiKey: process.env.GEMINI_API_KEY! });\n * const model = google.model({ name: \"gemini-2.5-flash\", temperature: 0.7 });\n * const embedder = google.embedder({ name: \"gemini-embedding-001\" });\n */\nexport class GoogleSDK implements SDKAdapterContract {\n  private readonly ai: GoogleGenAI;\n  private readonly provider: string;\n  private readonly pricing?: Record<string, ModelPricing>;\n\n  public constructor(config: GoogleSDKConfig) {\n    const { provider, pricing, ...clientOptions } = config;\n\n    this.ai = new GoogleGenAI(clientOptions);\n    this.provider = provider ?? \"google\";\n    this.pricing = pricing;\n  }\n\n  /**\n   * Build a `GoogleModel` bound to this SDK's client. Each call\n   * returns a fresh instance; all instances share the underlying\n   * `GoogleGenAI` client. The SDK's `provider` label is forwarded.\n   *\n   * Pricing resolution: per-model `config.pricing` wins; otherwise the\n   * SDK-level registry entry keyed by `config.name`; otherwise\n   * `undefined` (no cost computed).\n   */\n  public model(config: GoogleModelConfig): ModelContract {\n    const resolvedPricing = config.pricing ?? this.pricing?.[config.name];\n    const resolvedConfig: GoogleModelConfig =\n      resolvedPricing === config.pricing ? config : { ...config, pricing: resolvedPricing };\n\n    return new GoogleModel(this.ai, resolvedConfig, this.provider);\n  }\n\n  /**\n   * Rough token-count estimate. Uses the character-heuristic\n   * (`approximateTokenCount`) from the core package — Gemini's\n   * `countTokens` is a network round-trip; `count()` is intentionally\n   * offline. Good for budgeting/quota guards, not billing.\n   */\n  public async count(text: string, _model?: string): Promise<number> {\n    return approximateTokenCount(text);\n  }\n\n  /**\n   * Build a `GoogleEmbedder` bound to this SDK's client.\n   *\n   * @example\n   * const embedder = google.embedder({ name: \"gemini-embedding-001\" });\n   * const { vector } = await embedder.embed(\"Hello world\");\n   */\n  public embedder(config: GoogleEmbedderConfig): EmbedderContract {\n    return new GoogleEmbedder(this.ai, config, this.provider);\n  }\n\n  /**\n   * Build an image model bound to this SDK's client for use with\n   * `ai.image({ model, prompt })`. `config.name` decides the transport\n   * (see {@link usesGeminiImageTransport}) — a `gemini-` id gets the\n   * `generateContent` implementation, everything else the Imagen\n   * `generateImages` one. No id is rejected locally either way, so an\n   * unsupported model fails at Google, not here.\n   *\n   * The two differ in what usage they can report, which is what the\n   * caller must price for: the Imagen path always returns a zero token\n   * `Usage` (Imagen reports none — price with `{ perImage }`), while the\n   * Gemini path passes through whatever `usageMetadata` Google attaches\n   * (price with `{ input, output }` when tokens come back).\n   *\n   * Pricing resolution mirrors `model()`: per-model `config.pricing`\n   * wins, otherwise the SDK-level registry entry keyed by `config.name`,\n   * otherwise `undefined`.\n   *\n   * @example\n   * const imagen = google.image({ name: \"imagen-4.0-generate-001\" });\n   * const gemini = google.image({ name: \"gemini-3.1-flash-lite-image\" });\n   * const { data } = await ai.image({ model: gemini, prompt: \"a red bicycle\" });\n   */\n  public image(config: GoogleImageConfig): ImageModelContract {\n    const resolvedPricing = config.pricing ?? this.pricing?.[config.name];\n    const resolvedConfig: GoogleImageConfig =\n      resolvedPricing === config.pricing ? config : { ...config, pricing: resolvedPricing };\n\n    if (usesGeminiImageTransport(config.name)) {\n      return new GeminiImageModel(this.ai, resolvedConfig, this.provider);\n    }\n\n    return new GoogleImageModel(this.ai, resolvedConfig, this.provider);\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA+CA,SAAS,yBAAyB,MAAuB;CACvD,OAAO,KAAK,YAAY,CAAC,CAAC,QAAQ,aAAa,EAAE,CAAC,CAAC,WAAW,SAAS;AACzE;;;;;;;;;;;;;;;;;;;;;;;;;AA0BA,IAAa,YAAb,MAAqD;CAKnD,AAAO,YAAY,QAAyB;EAC1C,MAAM,EAAE,UAAU,SAAS,GAAG,kBAAkB;EAEhD,KAAK,KAAK,IAAI,YAAY,aAAa;EACvC,KAAK,WAAW,YAAY;EAC5B,KAAK,UAAU;CACjB;;;;;;;;;;CAWA,AAAO,MAAM,QAA0C;EACrD,MAAM,kBAAkB,OAAO,WAAW,KAAK,UAAU,OAAO;EAChE,MAAM,iBACJ,oBAAoB,OAAO,UAAU,SAAS;GAAE,GAAG;GAAQ,SAAS;EAAgB;EAEtF,OAAO,IAAI,YAAY,KAAK,IAAI,gBAAgB,KAAK,QAAQ;CAC/D;;;;;;;CAQA,MAAa,MAAM,MAAc,QAAkC;EACjE,OAAO,sBAAsB,IAAI;CACnC;;;;;;;;CASA,AAAO,SAAS,QAAgD;EAC9D,OAAO,IAAI,eAAe,KAAK,IAAI,QAAQ,KAAK,QAAQ;CAC1D;;;;;;;;;;;;;;;;;;;;;;;;CAyBA,AAAO,MAAM,QAA+C;EAC1D,MAAM,kBAAkB,OAAO,WAAW,KAAK,UAAU,OAAO;EAChE,MAAM,iBACJ,oBAAoB,OAAO,UAAU,SAAS;GAAE,GAAG;GAAQ,SAAS;EAAgB;EAEtF,IAAI,yBAAyB,OAAO,IAAI,GACtC,OAAO,IAAI,iBAAiB,KAAK,IAAI,gBAAgB,KAAK,QAAQ;EAGpE,OAAO,IAAI,iBAAiB,KAAK,IAAI,gBAAgB,KAAK,QAAQ;CACpE;AACF"}
+{"version":3,"file":"sdk.mjs","names":[],"sources":["../../../../../../../packages/ai-google/src/sdk.ts"],"sourcesContent":["import { GoogleGenAI } from \"@google/genai\";\nimport type {\n  EmbedderContract,\n  ImageModelContract,\n  ModelContract,\n  ModelPricing,\n  SDKAdapterContract,\n} from \"@warlock.js/ai\";\nimport { approximateTokenCount } from \"@warlock.js/ai\";\nimport type {\n  GoogleEmbedderConfig,\n  GoogleImageConfig,\n  GoogleModelConfig,\n  GoogleSDKConfig,\n} from \"./config.type\";\nimport { GoogleEmbedder } from \"./embedder\";\nimport { GeminiImageModel } from \"./gemini-image\";\nimport { GoogleImageModel } from \"./image\";\nimport { googlePricingFor } from \"./known-pricing\";\nimport { GoogleModel } from \"./model\";\n\n/**\n * Pick the transport for an image model id.\n *\n * `ai.models.generateImages` calls `{model}:predict`, and a `gemini-`\n * id sent there comes back `404 … is not supported for predict`\n * (observed verbatim from Google). `generateContent` is what the SDK\n * itself points `generateImages` users at — its deprecation notice\n * reads \"Please use the generateContent method with image models\n * instead\" — so the id has to choose the transport.\n *\n * Runs in this package establish where a `gemini-` id is ACCEPTED, not\n * what it returns: on this transport such an id got as far as a quota\n * error (HTTP 429) instead of the 404. That an image\n * comes back end-to-end once billing is enabled is reported by the\n * maintainer from a locally linked build, not measured here. Whether\n * these models report token usage is still unknown.\n *\n * This is ROUTING, not validation — no id is refused here. An id this\n * function does not recognize takes the `generateImages` route, the\n * only route that existed before Gemini image support landed, so every\n * id that reached Google before still reaches Google the same way and\n * still fails (or succeeds) at the provider.\n *\n * A leading `models/` resource prefix is tolerated, matching the id\n * shapes `inferVisionCapability` already accepts\n * (`models/gemini-1.5-flash-001`).\n */\nfunction usesGeminiImageTransport(name: string): boolean {\n  return name\n    .toLowerCase()\n    .replace(/^models\\//, \"\")\n    .startsWith(\"gemini-\");\n}\n\n/**\n * Google Gemini-backed implementation of `SDKAdapterContract`.\n *\n * **Role.** The package entry point for Gemini models via the\n * `@google/genai` SDK. A single `GoogleSDK` holds one live\n * `GoogleGenAI` client, shared by every `ModelContract` /\n * `EmbedderContract` it produces. Construct one SDK per\n * account/project and reuse it everywhere.\n *\n * **Responsibility.**\n * - Owns: a long-lived `GoogleGenAI` client (auth, Vertex vs Gemini\n *   API) and its lifetime. Factory for `GoogleModel` /\n *   `GoogleEmbedder` instances sharing that client.\n * - Does NOT own: anything per-call — those live in `GoogleModel` /\n *   `GoogleEmbedder` and the agent runtime.\n *\n * Modeled as a class (see §4.2 of code-style.md — \"long-lived state\n * across many calls\"), fronted by FP usage like the other adapters.\n *\n * @example\n * const google = new GoogleSDK({ apiKey: process.env.GEMINI_API_KEY! });\n * const model = google.model({ name: \"gemini-2.5-flash\", temperature: 0.7 });\n * const embedder = google.embedder({ name: \"gemini-embedding-001\" });\n */\nexport class GoogleSDK implements SDKAdapterContract {\n  private readonly ai: GoogleGenAI;\n  private readonly provider: string;\n  private readonly pricing?: Record<string, ModelPricing>;\n\n  public constructor(config: GoogleSDKConfig) {\n    const { provider, pricing, ...clientOptions } = config;\n\n    this.ai = new GoogleGenAI(clientOptions);\n    this.provider = provider ?? \"google\";\n    this.pricing = pricing;\n  }\n\n  /**\n   * Build a `GoogleModel` bound to this SDK's client. Each call\n   * returns a fresh instance; all instances share the underlying\n   * `GoogleGenAI` client. The SDK's `provider` label is forwarded.\n   *\n   * Pricing resolution: per-model `config.pricing` wins; otherwise the\n   * SDK-level registry entry keyed by `config.name`; otherwise the\n   * built-in list price (`GOOGLE_MODEL_PRICING`); otherwise `undefined`\n   * (no cost computed).\n   */\n  public model(config: GoogleModelConfig): ModelContract {\n    const resolvedPricing =\n      config.pricing ?? this.pricing?.[config.name] ?? googlePricingFor(config.name);\n    const resolvedConfig: GoogleModelConfig =\n      resolvedPricing === config.pricing ? config : { ...config, pricing: resolvedPricing };\n\n    return new GoogleModel(this.ai, resolvedConfig, this.provider);\n  }\n\n  /**\n   * Rough token-count estimate. Uses the character-heuristic\n   * (`approximateTokenCount`) from the core package — Gemini's\n   * `countTokens` is a network round-trip; `count()` is intentionally\n   * offline. Good for budgeting/quota guards, not billing.\n   */\n  public async count(text: string, _model?: string): Promise<number> {\n    return approximateTokenCount(text);\n  }\n\n  /**\n   * Build a `GoogleEmbedder` bound to this SDK's client.\n   *\n   * @example\n   * const embedder = google.embedder({ name: \"gemini-embedding-001\" });\n   * const { vector } = await embedder.embed(\"Hello world\");\n   */\n  public embedder(config: GoogleEmbedderConfig): EmbedderContract {\n    return new GoogleEmbedder(this.ai, config, this.provider);\n  }\n\n  /**\n   * Build an image model bound to this SDK's client for use with\n   * `ai.image({ model, prompt })`. `config.name` decides the transport\n   * (see {@link usesGeminiImageTransport}) — a `gemini-` id gets the\n   * `generateContent` implementation, everything else the Imagen\n   * `generateImages` one. No id is rejected locally either way, so an\n   * unsupported model fails at Google, not here.\n   *\n   * The two differ in what usage they can report, which is what the\n   * caller must price for: the Imagen path always returns a zero token\n   * `Usage` (Imagen reports none — price with `{ perImage }`), while the\n   * Gemini path passes through whatever `usageMetadata` Google attaches\n   * (price with `{ input, output }` when tokens come back).\n   *\n   * Pricing resolution mirrors `model()`: per-model `config.pricing`\n   * wins, otherwise the SDK-level registry entry keyed by `config.name`,\n   * otherwise `undefined`.\n   *\n   * @example\n   * const imagen = google.image({ name: \"imagen-4.0-generate-001\" });\n   * const gemini = google.image({ name: \"gemini-3.1-flash-lite-image\" });\n   * const { data } = await ai.image({ model: gemini, prompt: \"a red bicycle\" });\n   */\n  public image(config: GoogleImageConfig): ImageModelContract {\n    const resolvedPricing = config.pricing ?? this.pricing?.[config.name];\n    const resolvedConfig: GoogleImageConfig =\n      resolvedPricing === config.pricing ? config : { ...config, pricing: resolvedPricing };\n\n    if (usesGeminiImageTransport(config.name)) {\n      return new GeminiImageModel(this.ai, resolvedConfig, this.provider);\n    }\n\n    return new GoogleImageModel(this.ai, resolvedConfig, this.provider);\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAgDA,SAAS,yBAAyB,MAAuB;CACvD,OAAO,KACJ,YAAY,CAAC,CACb,QAAQ,aAAa,EAAE,CAAC,CACxB,WAAW,SAAS;AACzB;;;;;;;;;;;;;;;;;;;;;;;;;AA0BA,IAAa,YAAb,MAAqD;CAKnD,AAAO,YAAY,QAAyB;EAC1C,MAAM,EAAE,UAAU,SAAS,GAAG,kBAAkB;EAEhD,KAAK,KAAK,IAAI,YAAY,aAAa;EACvC,KAAK,WAAW,YAAY;EAC5B,KAAK,UAAU;CACjB;;;;;;;;;;;CAYA,AAAO,MAAM,QAA0C;EACrD,MAAM,kBACJ,OAAO,WAAW,KAAK,UAAU,OAAO,SAAS,iBAAiB,OAAO,IAAI;EAC/E,MAAM,iBACJ,oBAAoB,OAAO,UAAU,SAAS;GAAE,GAAG;GAAQ,SAAS;EAAgB;EAEtF,OAAO,IAAI,YAAY,KAAK,IAAI,gBAAgB,KAAK,QAAQ;CAC/D;;;;;;;CAQA,MAAa,MAAM,MAAc,QAAkC;EACjE,OAAO,sBAAsB,IAAI;CACnC;;;;;;;;CASA,AAAO,SAAS,QAAgD;EAC9D,OAAO,IAAI,eAAe,KAAK,IAAI,QAAQ,KAAK,QAAQ;CAC1D;;;;;;;;;;;;;;;;;;;;;;;;CAyBA,AAAO,MAAM,QAA+C;EAC1D,MAAM,kBAAkB,OAAO,WAAW,KAAK,UAAU,OAAO;EAChE,MAAM,iBACJ,oBAAoB,OAAO,UAAU,SAAS;GAAE,GAAG;GAAQ,SAAS;EAAgB;EAEtF,IAAI,yBAAyB,OAAO,IAAI,GACtC,OAAO,IAAI,iBAAiB,KAAK,IAAI,gBAAgB,KAAK,QAAQ;EAGpE,OAAO,IAAI,iBAAiB,KAAK,IAAI,gBAAgB,KAAK,QAAQ;CACpE;AACF"}

package/esm/utils/apply-google-usage.mjs

@@ -14,8 +14,13 @@
 *
 * `total` falls back to `input + output` when Google omits
 * `totalTokenCount`.
+*
+* `promptTokenCount` "includes the number of tokens in the cached
+* content" (https://ai.google.dev/api/generate-content#UsageMetadata), so
+* the usage is marked `inputIncludesCache: true`.
 */
 function applyGoogleUsage(usage, raw) {
+	usage.inputIncludesCache = true;
 	usage.input = raw.promptTokenCount ?? usage.input;
 	usage.output = raw.candidatesTokenCount ?? usage.output;
 	usage.total = raw.totalTokenCount ?? usage.input + usage.output;

package/esm/utils/apply-google-usage.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"apply-google-usage.mjs","names":[],"sources":["../../../../../../../../packages/ai-google/src/utils/apply-google-usage.ts"],"sourcesContent":["import type { Usage } from \"@warlock.js/ai\";\nimport type { GenerateContentResponse } from \"@google/genai\";\n\n/** Gemini's per-response token accounting block, as the SDK types it. */\nexport type GoogleUsageMetadata = NonNullable<GenerateContentResponse[\"usageMetadata\"]>;\n\n/**\n * Fold a Gemini `usageMetadata` block into a running neutral `Usage`\n * accumulator. Shared by every `generateContent`-backed surface — the\n * chat model's `complete()`, its streaming loop (where the final chunk\n * carries cumulative totals), and the Gemini image model — so one\n * mapping decides what a Gemini token report means package-wide.\n *\n * Cache-read hits (`cachedContentTokenCount`, implicit or explicit\n * context caching) surface as `cachedTokens`; the thinking-phase tokens\n * of a reasoning model (`thoughtsTokenCount`) surface as\n * `reasoningTokens`. Both are emitted only when reported `> 0` so an\n * absent channel leaves the field undefined rather than a false zero.\n *\n * `total` falls back to `input + output` when Google omits\n * `totalTokenCount`.\n */\nexport function applyGoogleUsage(usage: Usage, raw: GoogleUsageMetadata): void {\n  usage.input = raw.promptTokenCount ?? usage.input;\n  usage.output = raw.candidatesTokenCount ?? usage.output;\n  usage.total = raw.totalTokenCount ?? usage.input + usage.output;\n\n  const cached = raw.cachedContentTokenCount;\n\n  if (cached && cached > 0) {\n    usage.cachedTokens = cached;\n  }\n\n  const reasoning = raw.thoughtsTokenCount;\n\n  if (reasoning && reasoning > 0) {\n    usage.reasoningTokens = reasoning;\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;AAsBA,SAAgB,iBAAiB,OAAc,KAAgC;CAC7E,MAAM,QAAQ,IAAI,oBAAoB,MAAM;CAC5C,MAAM,SAAS,IAAI,wBAAwB,MAAM;CACjD,MAAM,QAAQ,IAAI,mBAAmB,MAAM,QAAQ,MAAM;CAEzD,MAAM,SAAS,IAAI;CAEnB,IAAI,UAAU,SAAS,GACrB,MAAM,eAAe;CAGvB,MAAM,YAAY,IAAI;CAEtB,IAAI,aAAa,YAAY,GAC3B,MAAM,kBAAkB;AAE5B"}
+{"version":3,"file":"apply-google-usage.mjs","names":[],"sources":["../../../../../../../../packages/ai-google/src/utils/apply-google-usage.ts"],"sourcesContent":["import type { Usage } from \"@warlock.js/ai\";\nimport type { GenerateContentResponse } from \"@google/genai\";\n\n/** Gemini's per-response token accounting block, as the SDK types it. */\nexport type GoogleUsageMetadata = NonNullable<GenerateContentResponse[\"usageMetadata\"]>;\n\n/**\n * Fold a Gemini `usageMetadata` block into a running neutral `Usage`\n * accumulator. Shared by every `generateContent`-backed surface — the\n * chat model's `complete()`, its streaming loop (where the final chunk\n * carries cumulative totals), and the Gemini image model — so one\n * mapping decides what a Gemini token report means package-wide.\n *\n * Cache-read hits (`cachedContentTokenCount`, implicit or explicit\n * context caching) surface as `cachedTokens`; the thinking-phase tokens\n * of a reasoning model (`thoughtsTokenCount`) surface as\n * `reasoningTokens`. Both are emitted only when reported `> 0` so an\n * absent channel leaves the field undefined rather than a false zero.\n *\n * `total` falls back to `input + output` when Google omits\n * `totalTokenCount`.\n *\n * `promptTokenCount` \"includes the number of tokens in the cached\n * content\" (https://ai.google.dev/api/generate-content#UsageMetadata), so\n * the usage is marked `inputIncludesCache: true`.\n */\nexport function applyGoogleUsage(usage: Usage, raw: GoogleUsageMetadata): void {\n  usage.inputIncludesCache = true;\n  usage.input = raw.promptTokenCount ?? usage.input;\n  usage.output = raw.candidatesTokenCount ?? usage.output;\n  usage.total = raw.totalTokenCount ?? usage.input + usage.output;\n\n  const cached = raw.cachedContentTokenCount;\n\n  if (cached && cached > 0) {\n    usage.cachedTokens = cached;\n  }\n\n  const reasoning = raw.thoughtsTokenCount;\n\n  if (reasoning && reasoning > 0) {\n    usage.reasoningTokens = reasoning;\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;AA0BA,SAAgB,iBAAiB,OAAc,KAAgC;CAC7E,MAAM,qBAAqB;CAC3B,MAAM,QAAQ,IAAI,oBAAoB,MAAM;CAC5C,MAAM,SAAS,IAAI,wBAAwB,MAAM;CACjD,MAAM,QAAQ,IAAI,mBAAmB,MAAM,QAAQ,MAAM;CAEzD,MAAM,SAAS,IAAI;CAEnB,IAAI,UAAU,SAAS,GACrB,MAAM,eAAe;CAGvB,MAAM,YAAY,IAAI;CAEtB,IAAI,aAAa,YAAY,GAC3B,MAAM,kBAAkB;AAE5B"}

package/esm/utils/grounding.mjs

@@ -0,0 +1,30 @@
+//#region ../../packages/ai-google/src/utils/grounding.ts
+/** Candidate snapshots are replaced, not added: streamed metadata may repeat.
+* Installed @google/genai 2.19.0 GroundingMetadata.webSearchQueries; REST schema:
+* https://ai.google.dev/api/generate-content#GroundingMetadata
+* Billing ignores empty queries and counts unique queries for Gemini 3+:
+* https://ai.google.dev/gemini-api/docs/google-search#pricing
+* Gemini 2.5 and older bill per grounded prompt; this counter is NOT that unit. */
+function applyGrounding(usage, snapshots, response) {
+	for (const [position, candidate] of (response.candidates ?? []).entries()) if (candidate.groundingMetadata) snapshots.set(candidate.index ?? position, candidate.groundingMetadata);
+	const reported = [...snapshots.values()].filter((metadata) => metadata.webSearchQueries !== void 0);
+	if (reported.length > 0) {
+		const queries = reported.flatMap((metadata) => metadata.webSearchQueries ?? []);
+		usage.serverToolUse = {
+			...usage.serverToolUse,
+			webSearchRequests: new Set(queries.filter((query) => query !== "")).size,
+			googleWebSearchQueries: queries.length
+		};
+	}
+}
+/** Retain chunks, supports, search entry point and queries without flattening them. */
+function groundingMetadata(snapshots) {
+	return snapshots.size ? { google: { groundingMetadata: [...snapshots].map(([index, metadata]) => ({
+		index,
+		...metadata
+	})) } } : void 0;
+}
+
+//#endregion
+export { applyGrounding, groundingMetadata };
+//# sourceMappingURL=grounding.mjs.map

package/esm/utils/grounding.mjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"grounding.mjs","names":[],"sources":["../../../../../../../../packages/ai-google/src/utils/grounding.ts"],"sourcesContent":["import type { GenerateContentResponse, GroundingMetadata } from \"@google/genai\";\nimport type { Usage } from \"@warlock.js/ai\";\n\n/** Candidate snapshots are replaced, not added: streamed metadata may repeat.\n * Installed @google/genai 2.19.0 GroundingMetadata.webSearchQueries; REST schema:\n * https://ai.google.dev/api/generate-content#GroundingMetadata\n * Billing ignores empty queries and counts unique queries for Gemini 3+:\n * https://ai.google.dev/gemini-api/docs/google-search#pricing\n * Gemini 2.5 and older bill per grounded prompt; this counter is NOT that unit. */\nexport function applyGrounding(\n  usage: Usage,\n  snapshots: Map<number, GroundingMetadata>,\n  response: GenerateContentResponse,\n): void {\n  for (const [position, candidate] of (response.candidates ?? []).entries()) {\n    if (candidate.groundingMetadata)\n      snapshots.set(candidate.index ?? position, candidate.groundingMetadata);\n  }\n  const reported = [...snapshots.values()].filter(\n    (metadata) => metadata.webSearchQueries !== undefined,\n  );\n  if (reported.length > 0) {\n    const queries = reported.flatMap((metadata) => metadata.webSearchQueries ?? []);\n    usage.serverToolUse = {\n      ...usage.serverToolUse,\n      webSearchRequests: new Set(queries.filter((query) => query !== \"\")).size,\n      googleWebSearchQueries: queries.length,\n    };\n  }\n}\n\n/** Retain chunks, supports, search entry point and queries without flattening them. */\nexport function groundingMetadata(\n  snapshots: Map<number, GroundingMetadata>,\n): Record<string, unknown> | undefined {\n  return snapshots.size\n    ? {\n        google: {\n          groundingMetadata: [...snapshots].map(([index, metadata]) => ({ index, ...metadata })),\n        },\n      }\n    : undefined;\n}\n"],"mappings":";;;;;;;AASA,SAAgB,eACd,OACA,WACA,UACM;CACN,KAAK,MAAM,CAAC,UAAU,eAAe,SAAS,cAAc,CAAC,EAAC,CAAE,QAAQ,GACtE,IAAI,UAAU,mBACZ,UAAU,IAAI,UAAU,SAAS,UAAU,UAAU,iBAAiB;CAE1E,MAAM,WAAW,CAAC,GAAG,UAAU,OAAO,CAAC,CAAC,CAAC,QACtC,aAAa,SAAS,qBAAqB,MAC9C;CACA,IAAI,SAAS,SAAS,GAAG;EACvB,MAAM,UAAU,SAAS,SAAS,aAAa,SAAS,oBAAoB,CAAC,CAAC;EAC9E,MAAM,gBAAgB;GACpB,GAAG,MAAM;GACT,mBAAmB,IAAI,IAAI,QAAQ,QAAQ,UAAU,UAAU,EAAE,CAAC,CAAC,CAAC;GACpE,wBAAwB,QAAQ;EAClC;CACF;AACF;;AAGA,SAAgB,kBACd,WACqC;CACrC,OAAO,UAAU,OACb,EACE,QAAQ,EACN,mBAAmB,CAAC,GAAG,SAAS,CAAC,CAAC,KAAK,CAAC,OAAO,eAAe;EAAE;EAAO,GAAG;CAAS,EAAE,EACvF,EACF,IACA;AACN"}

package/esm/utils/index.mjs

@@ -4,4 +4,4 @@ import { toGoogleContents } from "./to-google-contents.mjs";
 import { toGoogleTools } from "./to-google-tools.mjs";
 import { wrapGoogleError } from "./wrap-google-error.mjs";
 
-export {  };
+export { applyGoogleUsage, mapFinishReason, toGoogleContents, toGoogleTools, wrapGoogleError };

package/esm/utils/to-google-contents.mjs

@@ -15,6 +15,12 @@ import { InvalidRequestError, safeJsonParse } from "@warlock.js/ai";
 * 4. **Tool calls are `functionCall` parts.** An assistant message
 *    with `toolCalls` becomes a `"model"` content: an optional leading
 *    `text` part followed by one `functionCall` part per call.
+* 5. **Adjacent same-role turns merge.** Consecutive contents with the
+*    same role become one content with their parts in order, so the
+*    results of parallel tool calls travel as one `user` turn (as the
+*    `@google/genai` SDK's own automatic function calling sends them)
+*    and a compaction summary next to a same-role turn does not break
+*    the user/model alternation.
 *
 * @example
 * const { systemInstruction, contents } = toGoogleContents([
@@ -74,10 +80,36 @@ function toGoogleContents(messages) {
 	}
 	return {
 		systemInstruction: systemParts.length > 0 ? systemParts.join("\n\n") : void 0,
-		contents
+		contents: mergeAdjacentContents(contents)
 	};
 }
 /**
+* Join consecutive contents that share a role into one, keeping part
+* order: parallel `functionResponse` parts stay in their `functionCall`
+* order, and each `functionCall` keeps its own `thoughtSignature`.
+* Returns new content objects; the input contents are not modified.
+*
+* Source for the single-turn shape: `@google/genai` 2.19.0
+* `Models.generateContent` (automatic function calling) sends every
+* function response of one model turn as a single `{ role: "user",
+* parts }` content, in call order.
+*/
+function mergeAdjacentContents(contents) {
+	const merged = [];
+	for (const content of contents) {
+		const previous = merged.at(-1);
+		if (previous && previous.role === content.role) {
+			merged[merged.length - 1] = {
+				...previous,
+				parts: [...previous.parts ?? [], ...content.parts ?? []]
+			};
+			continue;
+		}
+		merged.push(content);
+	}
+	return merged;
+}
+/**
 * Multipart content is only meaningful on user messages — for any
 * other role collapse a `ContentPart[]` to concatenated text. Plain
 * strings pass through unchanged.

package/esm/utils/to-google-contents.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"to-google-contents.mjs","names":[],"sources":["../../../../../../../../packages/ai-google/src/utils/to-google-contents.ts"],"sourcesContent":["import { InvalidRequestError, safeJsonParse, type ContentPart, type Message } from \"@warlock.js/ai\";\nimport type { Content, Part } from \"@google/genai\";\n\n/**\n * Result of splitting a vendor-neutral `Message[]` for Gemini's\n * `generateContent`: the system prompt is hoisted to a separate\n * `systemInstruction` string (Gemini has no `\"system\"` role — content\n * roles must be `\"user\"` or `\"model\"`), and the remaining turns map to\n * `Content[]`.\n */\nexport type GoogleContents = {\n  systemInstruction: string | undefined;\n  contents: Content[];\n};\n\n/**\n * Convert vendor-neutral `Message[]` into Gemini's request shape.\n *\n * Gemini specifics this function absorbs:\n *\n * 1. **No `system` role.** System messages concatenate into the\n *    separate `systemInstruction` config field.\n * 2. **Role names differ.** Neutral `assistant` → Gemini `\"model\"`;\n *    `user` stays `\"user\"`.\n * 3. **Tool results are `user` turns.** A neutral `tool` message\n *    becomes a `\"user\"` content with a single `functionResponse` part.\n * 4. **Tool calls are `functionCall` parts.** An assistant message\n *    with `toolCalls` becomes a `\"model\"` content: an optional leading\n *    `text` part followed by one `functionCall` part per call.\n *\n * @example\n * const { systemInstruction, contents } = toGoogleContents([\n *   { role: \"system\", content: \"Be concise.\" },\n *   { role: \"user\", content: \"Hi\" },\n * ]);\n */\nexport function toGoogleContents(messages: Message[]): GoogleContents {\n  const systemParts: string[] = [];\n  const contents: Content[] = [];\n\n  for (const message of messages) {\n    if (message.role === \"system\") {\n      systemParts.push(stringifyContent(message.content));\n\n      continue;\n    }\n\n    if (message.role === \"tool\") {\n      contents.push({\n        role: \"user\",\n        parts: [\n          {\n            // Gemini matches a `functionResponse` to its `functionCall`\n            // by `name` (the Developer API has no call ids). `name` is\n            // the neutral `toolCallId`, which `GoogleModel` set to the\n            // function name. The wire `id` is intentionally omitted —\n            // an empty/synthetic id is rejected as an invalid argument.\n            functionResponse: {\n              name: message.toolCallId ?? \"\",\n              response: toResponseObject(stringifyContent(message.content)),\n            },\n          },\n        ],\n      });\n\n      continue;\n    }\n\n    if (message.role === \"assistant\" && message.toolCalls && message.toolCalls.length > 0) {\n      const parts: Part[] = [];\n      const text = stringifyContent(message.content);\n\n      if (text) {\n        parts.push({ text });\n      }\n\n      for (const toolCall of message.toolCalls) {\n        // Replay the opaque `thoughtSignature` Gemini attached to this\n        // function call on the original turn. Thinking models reject\n        // the follow-up request with a 400 if the signature is missing\n        // from the echoed `functionCall` part. Captured by\n        // `GoogleModel.partToToolCall` into `providerMetadata`.\n        const thoughtSignature = toolCall.providerMetadata?.thoughtSignature;\n\n        parts.push({\n          ...(typeof thoughtSignature === \"string\" ? { thoughtSignature } : {}),\n          // `id` omitted deliberately — Gemini Developer API function\n          // calls have no ids; echoing an empty/synthetic one is\n          // rejected as an invalid argument. Matched by `name`.\n          functionCall: {\n            name: toolCall.name,\n            args: (toolCall.input ?? {}) as Record<string, unknown>,\n          },\n        });\n      }\n\n      contents.push({ role: \"model\", parts });\n\n      continue;\n    }\n\n    if (message.role === \"user\" && Array.isArray(message.content)) {\n      contents.push({ role: \"user\", parts: message.content.map(toGooglePart) });\n\n      continue;\n    }\n\n    contents.push({\n      role: message.role === \"assistant\" ? \"model\" : \"user\",\n      parts: [{ text: stringifyContent(message.content) }],\n    });\n  }\n\n  return {\n    systemInstruction: systemParts.length > 0 ? systemParts.join(\"\\n\\n\") : undefined,\n    contents,\n  };\n}\n\n/**\n * Multipart content is only meaningful on user messages — for any\n * other role collapse a `ContentPart[]` to concatenated text. Plain\n * strings pass through unchanged.\n */\nfunction stringifyContent(content: string | ContentPart[]): string {\n  if (typeof content === \"string\") {\n    return content;\n  }\n\n  return content\n    .filter((part): part is { type: \"text\"; text: string } => part.type === \"text\")\n    .map((part) => part.text)\n    .join(\"\");\n}\n\n/**\n * Gemini's `functionResponse.response` must be a JSON object. Tool\n * results arrive as a string (usually stringified JSON) — parse it\n * when it is a JSON object, otherwise wrap the raw string under a\n * `result` key so the model always receives a well-formed object.\n */\nfunction toResponseObject(raw: string): Record<string, unknown> {\n  const parsed = safeJsonParse<unknown>(raw, undefined);\n\n  if (parsed !== null && typeof parsed === \"object\" && !Array.isArray(parsed)) {\n    return parsed as Record<string, unknown>;\n  }\n\n  return { result: raw };\n}\n\n/**\n * Map a resolved `ContentPart` to a Gemini `Part`. All binary\n * modalities — **image, PDF, and audio** — go to a single\n * `inlineData: { mimeType, data }` block; Gemini's multimodal input is\n * media-agnostic and keys off the IANA `mimeType` (`image/png`,\n * `application/pdf`, `audio/mpeg`, …), so one mapping covers every part\n * type the model's capabilities admit. PDF and audio reach this point\n * only when the model declares the matching capability (`google.model`\n * infers `pdf` / `audio` from the multimodal Gemini families); the\n * agent's modality gate throws upfront otherwise, so capability and\n * behavior stay in lockstep.\n *\n * Gemini's `generateContent` does not fetch arbitrary remote URLs (only\n * Files API / GCS URIs via `fileData`), so a neutral `{ url }` source\n * surfaces a typed `InvalidRequestError` upfront — for any modality —\n * rather than a downstream Gemini fault. The agent resolves attachments\n * before this point, so nothing is read or fetched here.\n */\nfunction toGooglePart(part: ContentPart): Part {\n  if (part.type === \"text\") {\n    return { text: part.text };\n  }\n\n  if (\"url\" in part.source) {\n    throw new InvalidRequestError(\n      `Gemini generateContent cannot fetch remote-URL ${part.type} media; supply base64 bytes instead.`,\n    );\n  }\n\n  return {\n    inlineData: { mimeType: part.source.mediaType, data: part.source.base64 },\n  };\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;AAoCA,SAAgB,iBAAiB,UAAqC;CACpE,MAAM,cAAwB,CAAC;CAC/B,MAAM,WAAsB,CAAC;CAE7B,KAAK,MAAM,WAAW,UAAU;EAC9B,IAAI,QAAQ,SAAS,UAAU;GAC7B,YAAY,KAAK,iBAAiB,QAAQ,OAAO,CAAC;GAElD;EACF;EAEA,IAAI,QAAQ,SAAS,QAAQ;GAC3B,SAAS,KAAK;IACZ,MAAM;IACN,OAAO,CACL,EAME,kBAAkB;KAChB,MAAM,QAAQ,cAAc;KAC5B,UAAU,iBAAiB,iBAAiB,QAAQ,OAAO,CAAC;IAC9D,EACF,CACF;GACF,CAAC;GAED;EACF;EAEA,IAAI,QAAQ,SAAS,eAAe,QAAQ,aAAa,QAAQ,UAAU,SAAS,GAAG;GACrF,MAAM,QAAgB,CAAC;GACvB,MAAM,OAAO,iBAAiB,QAAQ,OAAO;GAE7C,IAAI,MACF,MAAM,KAAK,EAAE,KAAK,CAAC;GAGrB,KAAK,MAAM,YAAY,QAAQ,WAAW;IAMxC,MAAM,mBAAmB,SAAS,kBAAkB;IAEpD,MAAM,KAAK;KACT,GAAI,OAAO,qBAAqB,WAAW,EAAE,iBAAiB,IAAI,CAAC;KAInE,cAAc;MACZ,MAAM,SAAS;MACf,MAAO,SAAS,SAAS,CAAC;KAC5B;IACF,CAAC;GACH;GAEA,SAAS,KAAK;IAAE,MAAM;IAAS;GAAM,CAAC;GAEtC;EACF;EAEA,IAAI,QAAQ,SAAS,UAAU,MAAM,QAAQ,QAAQ,OAAO,GAAG;GAC7D,SAAS,KAAK;IAAE,MAAM;IAAQ,OAAO,QAAQ,QAAQ,IAAI,YAAY;GAAE,CAAC;GAExE;EACF;EAEA,SAAS,KAAK;GACZ,MAAM,QAAQ,SAAS,cAAc,UAAU;GAC/C,OAAO,CAAC,EAAE,MAAM,iBAAiB,QAAQ,OAAO,EAAE,CAAC;EACrD,CAAC;CACH;CAEA,OAAO;EACL,mBAAmB,YAAY,SAAS,IAAI,YAAY,KAAK,MAAM,IAAI;EACvE;CACF;AACF;;;;;;AAOA,SAAS,iBAAiB,SAAyC;CACjE,IAAI,OAAO,YAAY,UACrB,OAAO;CAGT,OAAO,QACJ,QAAQ,SAAiD,KAAK,SAAS,MAAM,CAAC,CAC9E,KAAK,SAAS,KAAK,IAAI,CAAC,CACxB,KAAK,EAAE;AACZ;;;;;;;AAQA,SAAS,iBAAiB,KAAsC;CAC9D,MAAM,SAAS,cAAuB,KAAK,MAAS;CAEpD,IAAI,WAAW,QAAQ,OAAO,WAAW,YAAY,CAAC,MAAM,QAAQ,MAAM,GACxE,OAAO;CAGT,OAAO,EAAE,QAAQ,IAAI;AACvB;;;;;;;;;;;;;;;;;;;AAoBA,SAAS,aAAa,MAAyB;CAC7C,IAAI,KAAK,SAAS,QAChB,OAAO,EAAE,MAAM,KAAK,KAAK;CAG3B,IAAI,SAAS,KAAK,QAChB,MAAM,IAAI,oBACR,kDAAkD,KAAK,KAAK,qCAC9D;CAGF,OAAO,EACL,YAAY;EAAE,UAAU,KAAK,OAAO;EAAW,MAAM,KAAK,OAAO;CAAO,EAC1E;AACF"}
+{"version":3,"file":"to-google-contents.mjs","names":[],"sources":["../../../../../../../../packages/ai-google/src/utils/to-google-contents.ts"],"sourcesContent":["import { InvalidRequestError, safeJsonParse, type ContentPart, type Message } from \"@warlock.js/ai\";\nimport type { Content, Part } from \"@google/genai\";\n\n/**\n * Result of splitting a vendor-neutral `Message[]` for Gemini's\n * `generateContent`: the system prompt is hoisted to a separate\n * `systemInstruction` string (Gemini has no `\"system\"` role — content\n * roles must be `\"user\"` or `\"model\"`), and the remaining turns map to\n * `Content[]`.\n */\nexport type GoogleContents = {\n  systemInstruction: string | undefined;\n  contents: Content[];\n};\n\n/**\n * Convert vendor-neutral `Message[]` into Gemini's request shape.\n *\n * Gemini specifics this function absorbs:\n *\n * 1. **No `system` role.** System messages concatenate into the\n *    separate `systemInstruction` config field.\n * 2. **Role names differ.** Neutral `assistant` → Gemini `\"model\"`;\n *    `user` stays `\"user\"`.\n * 3. **Tool results are `user` turns.** A neutral `tool` message\n *    becomes a `\"user\"` content with a single `functionResponse` part.\n * 4. **Tool calls are `functionCall` parts.** An assistant message\n *    with `toolCalls` becomes a `\"model\"` content: an optional leading\n *    `text` part followed by one `functionCall` part per call.\n * 5. **Adjacent same-role turns merge.** Consecutive contents with the\n *    same role become one content with their parts in order, so the\n *    results of parallel tool calls travel as one `user` turn (as the\n *    `@google/genai` SDK's own automatic function calling sends them)\n *    and a compaction summary next to a same-role turn does not break\n *    the user/model alternation.\n *\n * @example\n * const { systemInstruction, contents } = toGoogleContents([\n *   { role: \"system\", content: \"Be concise.\" },\n *   { role: \"user\", content: \"Hi\" },\n * ]);\n */\nexport function toGoogleContents(messages: Message[]): GoogleContents {\n  const systemParts: string[] = [];\n  const contents: Content[] = [];\n\n  for (const message of messages) {\n    if (message.role === \"system\") {\n      systemParts.push(stringifyContent(message.content));\n\n      continue;\n    }\n\n    if (message.role === \"tool\") {\n      contents.push({\n        role: \"user\",\n        parts: [\n          {\n            // Gemini matches a `functionResponse` to its `functionCall`\n            // by `name` (the Developer API has no call ids). `name` is\n            // the neutral `toolCallId`, which `GoogleModel` set to the\n            // function name. The wire `id` is intentionally omitted —\n            // an empty/synthetic id is rejected as an invalid argument.\n            functionResponse: {\n              name: message.toolCallId ?? \"\",\n              response: toResponseObject(stringifyContent(message.content)),\n            },\n          },\n        ],\n      });\n\n      continue;\n    }\n\n    if (message.role === \"assistant\" && message.toolCalls && message.toolCalls.length > 0) {\n      const parts: Part[] = [];\n      const text = stringifyContent(message.content);\n\n      if (text) {\n        parts.push({ text });\n      }\n\n      for (const toolCall of message.toolCalls) {\n        // Replay the opaque `thoughtSignature` Gemini attached to this\n        // function call on the original turn. Thinking models reject\n        // the follow-up request with a 400 if the signature is missing\n        // from the echoed `functionCall` part. Captured by\n        // `GoogleModel.partToToolCall` into `providerMetadata`.\n        const thoughtSignature = toolCall.providerMetadata?.thoughtSignature;\n\n        parts.push({\n          ...(typeof thoughtSignature === \"string\" ? { thoughtSignature } : {}),\n          // `id` omitted deliberately — Gemini Developer API function\n          // calls have no ids; echoing an empty/synthetic one is\n          // rejected as an invalid argument. Matched by `name`.\n          functionCall: {\n            name: toolCall.name,\n            args: (toolCall.input ?? {}) as Record<string, unknown>,\n          },\n        });\n      }\n\n      contents.push({ role: \"model\", parts });\n\n      continue;\n    }\n\n    if (message.role === \"user\" && Array.isArray(message.content)) {\n      contents.push({ role: \"user\", parts: message.content.map(toGooglePart) });\n\n      continue;\n    }\n\n    contents.push({\n      role: message.role === \"assistant\" ? \"model\" : \"user\",\n      parts: [{ text: stringifyContent(message.content) }],\n    });\n  }\n\n  return {\n    systemInstruction: systemParts.length > 0 ? systemParts.join(\"\\n\\n\") : undefined,\n    contents: mergeAdjacentContents(contents),\n  };\n}\n\n/**\n * Join consecutive contents that share a role into one, keeping part\n * order: parallel `functionResponse` parts stay in their `functionCall`\n * order, and each `functionCall` keeps its own `thoughtSignature`.\n * Returns new content objects; the input contents are not modified.\n *\n * Source for the single-turn shape: `@google/genai` 2.19.0\n * `Models.generateContent` (automatic function calling) sends every\n * function response of one model turn as a single `{ role: \"user\",\n * parts }` content, in call order.\n */\nfunction mergeAdjacentContents(contents: Content[]): Content[] {\n  const merged: Content[] = [];\n\n  for (const content of contents) {\n    const previous = merged.at(-1);\n\n    if (previous && previous.role === content.role) {\n      merged[merged.length - 1] = {\n        ...previous,\n        parts: [...(previous.parts ?? []), ...(content.parts ?? [])],\n      };\n\n      continue;\n    }\n\n    merged.push(content);\n  }\n\n  return merged;\n}\n\n/**\n * Multipart content is only meaningful on user messages — for any\n * other role collapse a `ContentPart[]` to concatenated text. Plain\n * strings pass through unchanged.\n */\nfunction stringifyContent(content: string | ContentPart[]): string {\n  if (typeof content === \"string\") {\n    return content;\n  }\n\n  return content\n    .filter((part): part is { type: \"text\"; text: string } => part.type === \"text\")\n    .map((part) => part.text)\n    .join(\"\");\n}\n\n/**\n * Gemini's `functionResponse.response` must be a JSON object. Tool\n * results arrive as a string (usually stringified JSON) — parse it\n * when it is a JSON object, otherwise wrap the raw string under a\n * `result` key so the model always receives a well-formed object.\n */\nfunction toResponseObject(raw: string): Record<string, unknown> {\n  const parsed = safeJsonParse<unknown>(raw, undefined);\n\n  if (parsed !== null && typeof parsed === \"object\" && !Array.isArray(parsed)) {\n    return parsed as Record<string, unknown>;\n  }\n\n  return { result: raw };\n}\n\n/**\n * Map a resolved `ContentPart` to a Gemini `Part`. All binary\n * modalities — **image, PDF, and audio** — go to a single\n * `inlineData: { mimeType, data }` block; Gemini's multimodal input is\n * media-agnostic and keys off the IANA `mimeType` (`image/png`,\n * `application/pdf`, `audio/mpeg`, …), so one mapping covers every part\n * type the model's capabilities admit. PDF and audio reach this point\n * only when the model declares the matching capability (`google.model`\n * infers `pdf` / `audio` from the multimodal Gemini families); the\n * agent's modality gate throws upfront otherwise, so capability and\n * behavior stay in lockstep.\n *\n * Gemini's `generateContent` does not fetch arbitrary remote URLs (only\n * Files API / GCS URIs via `fileData`), so a neutral `{ url }` source\n * surfaces a typed `InvalidRequestError` upfront — for any modality —\n * rather than a downstream Gemini fault. The agent resolves attachments\n * before this point, so nothing is read or fetched here.\n */\nfunction toGooglePart(part: ContentPart): Part {\n  if (part.type === \"text\") {\n    return { text: part.text };\n  }\n\n  if (\"url\" in part.source) {\n    throw new InvalidRequestError(\n      `Gemini generateContent cannot fetch remote-URL ${part.type} media; supply base64 bytes instead.`,\n    );\n  }\n\n  return {\n    inlineData: { mimeType: part.source.mediaType, data: part.source.base64 },\n  };\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA0CA,SAAgB,iBAAiB,UAAqC;CACpE,MAAM,cAAwB,CAAC;CAC/B,MAAM,WAAsB,CAAC;CAE7B,KAAK,MAAM,WAAW,UAAU;EAC9B,IAAI,QAAQ,SAAS,UAAU;GAC7B,YAAY,KAAK,iBAAiB,QAAQ,OAAO,CAAC;GAElD;EACF;EAEA,IAAI,QAAQ,SAAS,QAAQ;GAC3B,SAAS,KAAK;IACZ,MAAM;IACN,OAAO,CACL,EAME,kBAAkB;KAChB,MAAM,QAAQ,cAAc;KAC5B,UAAU,iBAAiB,iBAAiB,QAAQ,OAAO,CAAC;IAC9D,EACF,CACF;GACF,CAAC;GAED;EACF;EAEA,IAAI,QAAQ,SAAS,eAAe,QAAQ,aAAa,QAAQ,UAAU,SAAS,GAAG;GACrF,MAAM,QAAgB,CAAC;GACvB,MAAM,OAAO,iBAAiB,QAAQ,OAAO;GAE7C,IAAI,MACF,MAAM,KAAK,EAAE,KAAK,CAAC;GAGrB,KAAK,MAAM,YAAY,QAAQ,WAAW;IAMxC,MAAM,mBAAmB,SAAS,kBAAkB;IAEpD,MAAM,KAAK;KACT,GAAI,OAAO,qBAAqB,WAAW,EAAE,iBAAiB,IAAI,CAAC;KAInE,cAAc;MACZ,MAAM,SAAS;MACf,MAAO,SAAS,SAAS,CAAC;KAC5B;IACF,CAAC;GACH;GAEA,SAAS,KAAK;IAAE,MAAM;IAAS;GAAM,CAAC;GAEtC;EACF;EAEA,IAAI,QAAQ,SAAS,UAAU,MAAM,QAAQ,QAAQ,OAAO,GAAG;GAC7D,SAAS,KAAK;IAAE,MAAM;IAAQ,OAAO,QAAQ,QAAQ,IAAI,YAAY;GAAE,CAAC;GAExE;EACF;EAEA,SAAS,KAAK;GACZ,MAAM,QAAQ,SAAS,cAAc,UAAU;GAC/C,OAAO,CAAC,EAAE,MAAM,iBAAiB,QAAQ,OAAO,EAAE,CAAC;EACrD,CAAC;CACH;CAEA,OAAO;EACL,mBAAmB,YAAY,SAAS,IAAI,YAAY,KAAK,MAAM,IAAI;EACvE,UAAU,sBAAsB,QAAQ;CAC1C;AACF;;;;;;;;;;;;AAaA,SAAS,sBAAsB,UAAgC;CAC7D,MAAM,SAAoB,CAAC;CAE3B,KAAK,MAAM,WAAW,UAAU;EAC9B,MAAM,WAAW,OAAO,GAAG,EAAE;EAE7B,IAAI,YAAY,SAAS,SAAS,QAAQ,MAAM;GAC9C,OAAO,OAAO,SAAS,KAAK;IAC1B,GAAG;IACH,OAAO,CAAC,GAAI,SAAS,SAAS,CAAC,GAAI,GAAI,QAAQ,SAAS,CAAC,CAAE;GAC7D;GAEA;EACF;EAEA,OAAO,KAAK,OAAO;CACrB;CAEA,OAAO;AACT;;;;;;AAOA,SAAS,iBAAiB,SAAyC;CACjE,IAAI,OAAO,YAAY,UACrB,OAAO;CAGT,OAAO,QACJ,QAAQ,SAAiD,KAAK,SAAS,MAAM,CAAC,CAC9E,KAAK,SAAS,KAAK,IAAI,CAAC,CACxB,KAAK,EAAE;AACZ;;;;;;;AAQA,SAAS,iBAAiB,KAAsC;CAC9D,MAAM,SAAS,cAAuB,KAAK,MAAS;CAEpD,IAAI,WAAW,QAAQ,OAAO,WAAW,YAAY,CAAC,MAAM,QAAQ,MAAM,GACxE,OAAO;CAGT,OAAO,EAAE,QAAQ,IAAI;AACvB;;;;;;;;;;;;;;;;;;;AAoBA,SAAS,aAAa,MAAyB;CAC7C,IAAI,KAAK,SAAS,QAChB,OAAO,EAAE,MAAM,KAAK,KAAK;CAG3B,IAAI,SAAS,KAAK,QAChB,MAAM,IAAI,oBACR,kDAAkD,KAAK,KAAK,qCAC9D;CAGF,OAAO,EACL,YAAY;EAAE,UAAU,KAAK,OAAO;EAAW,MAAM,KAAK,OAAO;CAAO,EAC1E;AACF"}

package/esm/utils/wrap-google-error.mjs

@@ -1,5 +1,5 @@
 import { ApiError } from "@google/genai";
-import { AIError, ContextLengthExceededError, InvalidRequestError, ProviderAuthError, ProviderError, ProviderRateLimitError, ProviderTimeoutError } from "@warlock.js/ai";
+import { AIError, ContextLengthExceededError, InvalidRequestError, ProviderAuthError, ProviderError, ProviderOverloadedError, ProviderRateLimitError, ProviderTimeoutError } from "@warlock.js/ai";
 
 //#region ../../packages/ai-google/src/utils/wrap-google-error.ts
 /**
@@ -28,35 +28,48 @@ function wrapGoogleError(thrown) {
 	const shape = toShape(thrown);
 	const context = buildContext(shape);
 	const message = shape.message ?? (thrown instanceof Error ? thrown.message : String(thrown));
+	const billed = shape.status === 400 || shape.status === 500 ? false : void 0;
 	if (isTimeout(shape)) return new ProviderTimeoutError(message, {
 		cause: thrown,
-		context
+		context,
+		billed
 	});
 	if (shape.status === 401 || shape.status === 403 || /permission_denied|api key not valid|unauthenticated/i.test(message)) return new ProviderAuthError(message, {
 		cause: thrown,
-		context
+		context,
+		billed
 	});
 	if (shape.status === 429 || /resource_exhausted|quota/i.test(message)) return new ProviderRateLimitError(message, {
 		cause: thrown,
-		context
+		context,
+		billed
 	});
 	if (shape.status === 400) {
 		if (/token count|context length|exceeds the maximum|input is too long/i.test(message)) return new ContextLengthExceededError(message, {
 			cause: thrown,
-			context
+			context,
+			billed
 		});
 		return new InvalidRequestError(message, {
 			cause: thrown,
-			context
+			context,
+			billed
 		});
 	}
 	if (shape.status === 404 || isClientStatus(shape.status)) return new InvalidRequestError(message, {
 		cause: thrown,
-		context
+		context,
+		billed
+	});
+	if (shape.status === 503 || /\bUNAVAILABLE\b/.test(message) || /overloaded/i.test(message)) return new ProviderOverloadedError(message, {
+		cause: thrown,
+		context,
+		billed
 	});
 	return new ProviderError(message, {
 		cause: thrown,
-		context
+		context,
+		billed
 	});
 }
 /**

package/esm/utils/wrap-google-error.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"wrap-google-error.mjs","names":[],"sources":["../../../../../../../../packages/ai-google/src/utils/wrap-google-error.ts"],"sourcesContent":["import {\n  AIError,\n  ContextLengthExceededError,\n  InvalidRequestError,\n  ProviderAuthError,\n  ProviderError,\n  ProviderRateLimitError,\n  ProviderTimeoutError,\n} from \"@warlock.js/ai\";\nimport { ApiError } from \"@google/genai\";\n\n/**\n * Raw-error fields the wrapper reads off a Gemini SDK error.\n * `@google/genai`'s `ApiError` exposes `status` (HTTP code) +\n * `message`; transport aborts surface as `AbortError` / `ETIMEDOUT`.\n * We duck-type so proxied / re-thrown errors still classify.\n */\ntype GoogleErrorShape = {\n  status?: number;\n  message?: string;\n  name?: string;\n  code?: string;\n};\n\n/**\n * Wrap any thrown value caught inside the Gemini adapter into the\n * appropriate `@warlock.js/ai` `AIError` subclass.\n *\n * **Dispatch strategy.** Gemini has no machine error `code`; the\n * signals are the HTTP `status` and the canonical status phrase Google\n * embeds in `message` (`PERMISSION_DENIED`, `RESOURCE_EXHAUSTED`,\n * `INVALID_ARGUMENT`, …). Dispatch keys on `status`, using the message\n * phrase as the tie-breaker for the two 400 sub-cases\n * (context-length vs generic) and for status-less auth/quota errors.\n *\n * `AIError` instances pass through unchanged so `catch/throw wrap(e)`\n * pipelines never double-wrap.\n *\n * @example\n * try {\n *   return await this.ai.models.generateContent(...);\n * } catch (thrown) {\n *   throw wrapGoogleError(thrown);\n * }\n */\nexport function wrapGoogleError(thrown: unknown): AIError {\n  if (thrown instanceof AIError) {\n    return thrown;\n  }\n\n  const shape = toShape(thrown);\n  const context = buildContext(shape);\n  const message = shape.message ?? (thrown instanceof Error ? thrown.message : String(thrown));\n\n  if (isTimeout(shape)) {\n    return new ProviderTimeoutError(message, { cause: thrown, context });\n  }\n\n  if (\n    shape.status === 401 ||\n    shape.status === 403 ||\n    /permission_denied|api key not valid|unauthenticated/i.test(message)\n  ) {\n    return new ProviderAuthError(message, { cause: thrown, context });\n  }\n\n  if (shape.status === 429 || /resource_exhausted|quota/i.test(message)) {\n    return new ProviderRateLimitError(message, { cause: thrown, context });\n  }\n\n  if (shape.status === 400) {\n    if (/token count|context length|exceeds the maximum|input is too long/i.test(message)) {\n      return new ContextLengthExceededError(message, { cause: thrown, context });\n    }\n\n    return new InvalidRequestError(message, { cause: thrown, context });\n  }\n\n  if (shape.status === 404 || isClientStatus(shape.status)) {\n    return new InvalidRequestError(message, { cause: thrown, context });\n  }\n\n  return new ProviderError(message, { cause: thrown, context });\n}\n\n/**\n * Read the raw error shape. The Gemini SDK's `ApiError` carries a\n * numeric `status`; flattened/proxied errors may carry it (or `code`)\n * loosely.\n */\nfunction toShape(thrown: unknown): GoogleErrorShape {\n  if (thrown instanceof ApiError) {\n    return { status: thrown.status, message: thrown.message, name: thrown.name };\n  }\n\n  if (typeof thrown === \"object\" && thrown !== null) {\n    const raw = thrown as Record<string, unknown>;\n\n    return {\n      status: typeof raw.status === \"number\" ? raw.status : undefined,\n      message: typeof raw.message === \"string\" ? raw.message : undefined,\n      name: typeof raw.name === \"string\" ? raw.name : undefined,\n      code: typeof raw.code === \"string\" ? raw.code : undefined,\n    };\n  }\n\n  return {};\n}\n\n/**\n * Decide whether the error is a timeout. Gemini maps gateway timeouts\n * to HTTP 504 (`DEADLINE_EXCEEDED`); transport aborts surface as\n * `AbortError` / `ETIMEDOUT` / `ECONNABORTED`.\n */\nfunction isTimeout(shape: GoogleErrorShape): boolean {\n  if (shape.status === 504) {\n    return true;\n  }\n\n  if (shape.name === \"AbortError\" || /deadline_exceeded/i.test(shape.message ?? \"\")) {\n    return true;\n  }\n\n  return shape.code === \"ETIMEDOUT\" || shape.code === \"ECONNABORTED\";\n}\n\n/** True for HTTP 4xx — a client-side request problem, not a server fault. */\nfunction isClientStatus(status: number | undefined): boolean {\n  return typeof status === \"number\" && status >= 400 && status < 500;\n}\n\n/** Attach the diagnostic fields to `error.context`. */\nfunction buildContext(shape: GoogleErrorShape): Record<string, unknown> {\n  const context: Record<string, unknown> = {};\n\n  if (shape.status !== undefined) {\n    context.status = shape.status;\n  }\n\n  if (shape.name) {\n    context.code = shape.name;\n  }\n\n  return context;\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;AA6CA,SAAgB,gBAAgB,QAA0B;CACxD,IAAI,kBAAkB,SACpB,OAAO;CAGT,MAAM,QAAQ,QAAQ,MAAM;CAC5B,MAAM,UAAU,aAAa,KAAK;CAClC,MAAM,UAAU,MAAM,YAAY,kBAAkB,QAAQ,OAAO,UAAU,OAAO,MAAM;CAE1F,IAAI,UAAU,KAAK,GACjB,OAAO,IAAI,qBAAqB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGrE,IACE,MAAM,WAAW,OACjB,MAAM,WAAW,OACjB,uDAAuD,KAAK,OAAO,GAEnE,OAAO,IAAI,kBAAkB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGlE,IAAI,MAAM,WAAW,OAAO,4BAA4B,KAAK,OAAO,GAClE,OAAO,IAAI,uBAAuB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGvE,IAAI,MAAM,WAAW,KAAK;EACxB,IAAI,oEAAoE,KAAK,OAAO,GAClF,OAAO,IAAI,2BAA2B,SAAS;GAAE,OAAO;GAAQ;EAAQ,CAAC;EAG3E,OAAO,IAAI,oBAAoB,SAAS;GAAE,OAAO;GAAQ;EAAQ,CAAC;CACpE;CAEA,IAAI,MAAM,WAAW,OAAO,eAAe,MAAM,MAAM,GACrD,OAAO,IAAI,oBAAoB,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;CAGpE,OAAO,IAAI,cAAc,SAAS;EAAE,OAAO;EAAQ;CAAQ,CAAC;AAC9D;;;;;;AAOA,SAAS,QAAQ,QAAmC;CAClD,IAAI,kBAAkB,UACpB,OAAO;EAAE,QAAQ,OAAO;EAAQ,SAAS,OAAO;EAAS,MAAM,OAAO;CAAK;CAG7E,IAAI,OAAO,WAAW,YAAY,WAAW,MAAM;EACjD,MAAM,MAAM;EAEZ,OAAO;GACL,QAAQ,OAAO,IAAI,WAAW,WAAW,IAAI,SAAS;GACtD,SAAS,OAAO,IAAI,YAAY,WAAW,IAAI,UAAU;GACzD,MAAM,OAAO,IAAI,SAAS,WAAW,IAAI,OAAO;GAChD,MAAM,OAAO,IAAI,SAAS,WAAW,IAAI,OAAO;EAClD;CACF;CAEA,OAAO,CAAC;AACV;;;;;;AAOA,SAAS,UAAU,OAAkC;CACnD,IAAI,MAAM,WAAW,KACnB,OAAO;CAGT,IAAI,MAAM,SAAS,gBAAgB,qBAAqB,KAAK,MAAM,WAAW,EAAE,GAC9E,OAAO;CAGT,OAAO,MAAM,SAAS,eAAe,MAAM,SAAS;AACtD;;AAGA,SAAS,eAAe,QAAqC;CAC3D,OAAO,OAAO,WAAW,YAAY,UAAU,OAAO,SAAS;AACjE;;AAGA,SAAS,aAAa,OAAkD;CACtE,MAAM,UAAmC,CAAC;CAE1C,IAAI,MAAM,WAAW,QACnB,QAAQ,SAAS,MAAM;CAGzB,IAAI,MAAM,MACR,QAAQ,OAAO,MAAM;CAGvB,OAAO;AACT"}
+{"version":3,"file":"wrap-google-error.mjs","names":[],"sources":["../../../../../../../../packages/ai-google/src/utils/wrap-google-error.ts"],"sourcesContent":["import {\n  AIError,\n  ContextLengthExceededError,\n  InvalidRequestError,\n  ProviderAuthError,\n  ProviderError,\n  ProviderOverloadedError,\n  ProviderRateLimitError,\n  ProviderTimeoutError,\n} from \"@warlock.js/ai\";\nimport { ApiError } from \"@google/genai\";\n\n/**\n * Raw-error fields the wrapper reads off a Gemini SDK error.\n * `@google/genai`'s `ApiError` exposes `status` (HTTP code) +\n * `message`; transport aborts surface as `AbortError` / `ETIMEDOUT`.\n * We duck-type so proxied / re-thrown errors still classify.\n */\ntype GoogleErrorShape = {\n  status?: number;\n  message?: string;\n  name?: string;\n  code?: string;\n};\n\n/**\n * Wrap any thrown value caught inside the Gemini adapter into the\n * appropriate `@warlock.js/ai` `AIError` subclass.\n *\n * **Dispatch strategy.** Gemini has no machine error `code`; the\n * signals are the HTTP `status` and the canonical status phrase Google\n * embeds in `message` (`PERMISSION_DENIED`, `RESOURCE_EXHAUSTED`,\n * `INVALID_ARGUMENT`, …). Dispatch keys on `status`, using the message\n * phrase as the tie-breaker for the two 400 sub-cases\n * (context-length vs generic) and for status-less auth/quota errors.\n *\n * `AIError` instances pass through unchanged so `catch/throw wrap(e)`\n * pipelines never double-wrap.\n *\n * @example\n * try {\n *   return await this.ai.models.generateContent(...);\n * } catch (thrown) {\n *   throw wrapGoogleError(thrown);\n * }\n */\nexport function wrapGoogleError(thrown: unknown): AIError {\n  if (thrown instanceof AIError) {\n    return thrown;\n  }\n\n  const shape = toShape(thrown);\n  const context = buildContext(shape);\n  const message = shape.message ?? (thrown instanceof Error ? thrown.message : String(thrown));\n  // \"If your request fails with a 400 or 500 error, you won't be charged\n  // for the tokens used.\" — https://ai.google.dev/gemini-api/docs/billing\n  // (fetched 2026-10-17). Every other failure keeps the default.\n  const billed = shape.status === 400 || shape.status === 500 ? false : undefined;\n\n  if (isTimeout(shape)) {\n    return new ProviderTimeoutError(message, { cause: thrown, context, billed });\n  }\n\n  if (\n    shape.status === 401 ||\n    shape.status === 403 ||\n    /permission_denied|api key not valid|unauthenticated/i.test(message)\n  ) {\n    return new ProviderAuthError(message, { cause: thrown, context, billed });\n  }\n\n  if (shape.status === 429 || /resource_exhausted|quota/i.test(message)) {\n    return new ProviderRateLimitError(message, { cause: thrown, context, billed });\n  }\n\n  if (shape.status === 400) {\n    if (/token count|context length|exceeds the maximum|input is too long/i.test(message)) {\n      return new ContextLengthExceededError(message, { cause: thrown, context, billed });\n    }\n\n    return new InvalidRequestError(message, { cause: thrown, context, billed });\n  }\n\n  if (shape.status === 404 || isClientStatus(shape.status)) {\n    return new InvalidRequestError(message, { cause: thrown, context, billed });\n  }\n\n  // 503 UNAVAILABLE — \"temporarily overloaded or down\"\n  // (https://ai.google.dev/gemini-api/docs/api-errors).\n  if (shape.status === 503 || /\\bUNAVAILABLE\\b/.test(message) || /overloaded/i.test(message)) {\n    return new ProviderOverloadedError(message, { cause: thrown, context, billed });\n  }\n\n  return new ProviderError(message, { cause: thrown, context, billed });\n}\n\n/**\n * Read the raw error shape. The Gemini SDK's `ApiError` carries a\n * numeric `status`; flattened/proxied errors may carry it (or `code`)\n * loosely.\n */\nfunction toShape(thrown: unknown): GoogleErrorShape {\n  if (thrown instanceof ApiError) {\n    return { status: thrown.status, message: thrown.message, name: thrown.name };\n  }\n\n  if (typeof thrown === \"object\" && thrown !== null) {\n    const raw = thrown as Record<string, unknown>;\n\n    return {\n      status: typeof raw.status === \"number\" ? raw.status : undefined,\n      message: typeof raw.message === \"string\" ? raw.message : undefined,\n      name: typeof raw.name === \"string\" ? raw.name : undefined,\n      code: typeof raw.code === \"string\" ? raw.code : undefined,\n    };\n  }\n\n  return {};\n}\n\n/**\n * Decide whether the error is a timeout. Gemini maps gateway timeouts\n * to HTTP 504 (`DEADLINE_EXCEEDED`); transport aborts surface as\n * `AbortError` / `ETIMEDOUT` / `ECONNABORTED`.\n */\nfunction isTimeout(shape: GoogleErrorShape): boolean {\n  if (shape.status === 504) {\n    return true;\n  }\n\n  if (shape.name === \"AbortError\" || /deadline_exceeded/i.test(shape.message ?? \"\")) {\n    return true;\n  }\n\n  return shape.code === \"ETIMEDOUT\" || shape.code === \"ECONNABORTED\";\n}\n\n/** True for HTTP 4xx — a client-side request problem, not a server fault. */\nfunction isClientStatus(status: number | undefined): boolean {\n  return typeof status === \"number\" && status >= 400 && status < 500;\n}\n\n/** Attach the diagnostic fields to `error.context`. */\nfunction buildContext(shape: GoogleErrorShape): Record<string, unknown> {\n  const context: Record<string, unknown> = {};\n\n  if (shape.status !== undefined) {\n    context.status = shape.status;\n  }\n\n  if (shape.name) {\n    context.code = shape.name;\n  }\n\n  return context;\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;AA8CA,SAAgB,gBAAgB,QAA0B;CACxD,IAAI,kBAAkB,SACpB,OAAO;CAGT,MAAM,QAAQ,QAAQ,MAAM;CAC5B,MAAM,UAAU,aAAa,KAAK;CAClC,MAAM,UAAU,MAAM,YAAY,kBAAkB,QAAQ,OAAO,UAAU,OAAO,MAAM;CAI1F,MAAM,SAAS,MAAM,WAAW,OAAO,MAAM,WAAW,MAAM,QAAQ;CAEtE,IAAI,UAAU,KAAK,GACjB,OAAO,IAAI,qBAAqB,SAAS;EAAE,OAAO;EAAQ;EAAS;CAAO,CAAC;CAG7E,IACE,MAAM,WAAW,OACjB,MAAM,WAAW,OACjB,uDAAuD,KAAK,OAAO,GAEnE,OAAO,IAAI,kBAAkB,SAAS;EAAE,OAAO;EAAQ;EAAS;CAAO,CAAC;CAG1E,IAAI,MAAM,WAAW,OAAO,4BAA4B,KAAK,OAAO,GAClE,OAAO,IAAI,uBAAuB,SAAS;EAAE,OAAO;EAAQ;EAAS;CAAO,CAAC;CAG/E,IAAI,MAAM,WAAW,KAAK;EACxB,IAAI,oEAAoE,KAAK,OAAO,GAClF,OAAO,IAAI,2BAA2B,SAAS;GAAE,OAAO;GAAQ;GAAS;EAAO,CAAC;EAGnF,OAAO,IAAI,oBAAoB,SAAS;GAAE,OAAO;GAAQ;GAAS;EAAO,CAAC;CAC5E;CAEA,IAAI,MAAM,WAAW,OAAO,eAAe,MAAM,MAAM,GACrD,OAAO,IAAI,oBAAoB,SAAS;EAAE,OAAO;EAAQ;EAAS;CAAO,CAAC;CAK5E,IAAI,MAAM,WAAW,OAAO,kBAAkB,KAAK,OAAO,KAAK,cAAc,KAAK,OAAO,GACvF,OAAO,IAAI,wBAAwB,SAAS;EAAE,OAAO;EAAQ;EAAS;CAAO,CAAC;CAGhF,OAAO,IAAI,cAAc,SAAS;EAAE,OAAO;EAAQ;EAAS;CAAO,CAAC;AACtE;;;;;;AAOA,SAAS,QAAQ,QAAmC;CAClD,IAAI,kBAAkB,UACpB,OAAO;EAAE,QAAQ,OAAO;EAAQ,SAAS,OAAO;EAAS,MAAM,OAAO;CAAK;CAG7E,IAAI,OAAO,WAAW,YAAY,WAAW,MAAM;EACjD,MAAM,MAAM;EAEZ,OAAO;GACL,QAAQ,OAAO,IAAI,WAAW,WAAW,IAAI,SAAS;GACtD,SAAS,OAAO,IAAI,YAAY,WAAW,IAAI,UAAU;GACzD,MAAM,OAAO,IAAI,SAAS,WAAW,IAAI,OAAO;GAChD,MAAM,OAAO,IAAI,SAAS,WAAW,IAAI,OAAO;EAClD;CACF;CAEA,OAAO,CAAC;AACV;;;;;;AAOA,SAAS,UAAU,OAAkC;CACnD,IAAI,MAAM,WAAW,KACnB,OAAO;CAGT,IAAI,MAAM,SAAS,gBAAgB,qBAAqB,KAAK,MAAM,WAAW,EAAE,GAC9E,OAAO;CAGT,OAAO,MAAM,SAAS,eAAe,MAAM,SAAS;AACtD;;AAGA,SAAS,eAAe,QAAqC;CAC3D,OAAO,OAAO,WAAW,YAAY,UAAU,OAAO,SAAS;AACjE;;AAGA,SAAS,aAAa,OAAkD;CACtE,MAAM,UAAmC,CAAC;CAE1C,IAAI,MAAM,WAAW,QACnB,QAAQ,SAAS,MAAM;CAGzB,IAAI,MAAM,MACR,QAAQ,OAAO,MAAM;CAGvB,OAAO;AACT"}