@warlock.js/ai-google 4.14.0 → 4.15.0

package/esm/gemini-image.mjs

@@ -0,0 +1,231 @@
+import { applyGoogleUsage } from "./utils/apply-google-usage.mjs";
+import { wrapGoogleError } from "./utils/wrap-google-error.mjs";
+import "./utils/index.mjs";
+import { ContentFilterError, ProviderError } from "@warlock.js/ai";
+import { log } from "@warlock.js/logger";
+
+//#region ../ai-google/src/gemini-image.ts
+const LOG_MODULE = "ai.google";
+/**
+* Response modalities requested when the caller names none.
+*
+* `IMAGE` is the modality this adapter extracts; `TEXT` rides along so
+* a model that narrates what it drew is not answering outside the set
+* it was granted (the narration is then dropped — only inline image
+* parts become `GeneratedImage`s).
+*
+* *Unverified:* which pairing any individual Gemini image model
+* requires is not established here — no spec or run in this package
+* touches the live API. `options.responseModalities` replaces this list
+* verbatim for a model that wants something else.
+*/
+const DEFAULT_RESPONSE_MODALITIES = ["TEXT", "IMAGE"];
+/**
+* Gemini `finishReason` values that mean generation was stopped by a
+* safety / policy rule rather than by the model simply not drawing.
+* Taken from the `FinishReason` enum in `@google/genai`'s own type
+* declarations, whose doc comments describe each of these as content
+* or image generation being "stopped" for safety, prohibited content,
+* recitation, blocklist, or SPII.
+*/
+const FILTERED_FINISH_REASONS = new Set([
+	"SAFETY",
+	"IMAGE_SAFETY",
+	"PROHIBITED_CONTENT",
+	"IMAGE_PROHIBITED_CONTENT",
+	"RECITATION",
+	"IMAGE_RECITATION",
+	"BLOCKLIST",
+	"SPII"
+]);
+/** How much of a text-only answer to quote back inside the error message. */
+const TEXT_EXCERPT_LIMIT = 200;
+/**
+* Gemini-native implementation of `ImageModelContract`, via
+* `ai.models.generateContent` with `config.responseModalities`
+* including `"IMAGE"`.
+*
+* **Why a second image adapter.** `GoogleImageModel` calls
+* `ai.models.generateImages`, which the `@google/genai` bundle routes
+* to `{model}:predict` (`generateImages` → `generateImagesInternal` →
+* `formatMap('{model}:predict', …)`). A Gemini image model is not
+* served there: asking for one returns Google's
+* `404 … is not found for API version v1beta, or is not supported for
+* predict`. `generateContent` is the SDK's own named replacement — its
+* runtime deprecation notice for `generateImages` reads "Please use the
+* generateContent method with image models instead" — so that is the
+* transport this class speaks, hence a separate class rather than a
+* branch inside `image.ts`.
+*
+* **Same envelope.** Inline image parts are mapped to the identical
+* `GeneratedImage[]` shape `GoogleImageModel` produces, so `ai.image()`
+* callers see no difference between the two paths.
+*
+* **Token usage is passed through, not zeroed.** The Imagen path
+* returns a hard `{ 0, 0, 0 }` because Imagen reports no tokens at all;
+* here, whatever `usageMetadata` Google attaches is mapped by the same
+* {@link applyGoogleUsage} the chat model uses, and only an absent
+* block collapses to zeros. Price accordingly.
+*
+* **No model-id validation.** `config.name` is forwarded to
+* `generateContent` exactly as given; nothing here inspects it. An id
+* Google does not serve fails at Google, wrapped into the typed
+* `AIError` hierarchy — never with a local throw.
+*
+* **Evidence, in two tiers.** No spec in this package calls Google.
+* *Measured here:* a `gemini-*` image id, which 404s on the `predict`
+* transport, reached the model on this one and came back with a quota
+* error (HTTP 429) — the endpoint accepts the id. *Reported by the
+* maintainer:* once billing was enabled on the project, an image came
+* back end-to-end from an application running a locally linked build.
+* *Still unestablished:* whether these models report token usage — no
+* `usageMetadata` from a successful image call has been observed, so
+* the pass-through above is untested against a real response.
+*
+* @example
+* const model = new GeminiImageModel(ai, { name: "gemini-3.1-flash-lite-image" });
+* const { images, usage } = await model.generate("a red bicycle on a white background");
+*/
+var GeminiImageModel = class {
+	constructor(ai, config, provider = "google") {
+		this.logger = log;
+		this.ai = ai;
+		this.name = config.name;
+		this.provider = provider;
+		this.pricing = config.pricing;
+	}
+	async generate(prompt, options) {
+		const config = this.buildConfig(options);
+		this.logger.debug(LOG_MODULE, "image.request", "models.generateContent", {
+			model: this.name,
+			responseModalities: config.responseModalities
+		});
+		let response;
+		try {
+			response = await this.ai.models.generateContent({
+				model: this.name,
+				contents: prompt,
+				config
+			});
+		} catch (thrown) {
+			const wrapped = wrapGoogleError(thrown);
+			this.logger.error(LOG_MODULE, "image.error", wrapped.message, {
+				code: wrapped.code,
+				context: wrapped.context
+			});
+			throw wrapped;
+		}
+		const parts = collectParts(response);
+		const images = toGeneratedImages(parts);
+		if (images.length === 0) throw this.noImageError(response, parts);
+		const usage = {
+			input: 0,
+			output: 0,
+			total: 0
+		};
+		if (response.usageMetadata) applyGoogleUsage(usage, response.usageMetadata);
+		this.logger.debug(LOG_MODULE, "image.response", "models.generateContent succeeded", {
+			images: images.length,
+			usage
+		});
+		return {
+			images,
+			usage
+		};
+	}
+	/**
+	* Assemble the `GenerateContentConfig` for an image turn: the
+	* requested modalities, the image-specific knobs Gemini exposes under
+	* `imageConfig`, and the cancellation handle.
+	*
+	* Three neutral options are deliberately NOT forwarded, because
+	* `GenerateContentConfig` / `ImageConfig` in `@google/genai` expose
+	* no equivalent for them on this path: `count` (no per-request image
+	* count — every inline image part the model does return is mapped),
+	* `negativePrompt` (an Imagen-only field), and `format`
+	* (`ImageConfig.outputMimeType` is documented "not supported in
+	* Gemini API"). Fold those intentions into the prompt instead.
+	*/
+	buildConfig(options) {
+		const imageConfig = {};
+		if (options?.aspectRatio !== void 0) imageConfig.aspectRatio = options.aspectRatio;
+		if (typeof options?.imageSize === "string") imageConfig.imageSize = options.imageSize;
+		if (typeof options?.personGeneration === "string") imageConfig.personGeneration = options.personGeneration;
+		const requested = options?.responseModalities;
+		return {
+			responseModalities: Array.isArray(requested) ? requested : DEFAULT_RESPONSE_MODALITIES,
+			...Object.keys(imageConfig).length > 0 ? { imageConfig } : {},
+			...options?.signal ? { abortSignal: options.signal } : {}
+		};
+	}
+	/**
+	* Build the typed error for a response that carried no inline image
+	* part. Never a silent empty success: the caller asked for an image
+	* and got something else, so the error names what actually came back.
+	*
+	* - A blocked prompt (`promptFeedback.blockReason`) or a
+	*   safety/policy `finishReason` → `ContentFilterError` carrying the
+	*   reason, matching how the Imagen path reports `raiFilteredReason`.
+	* - A text-only answer → `ProviderError` quoting the text, so the
+	*   log says what the model replied instead of guessing.
+	* - Anything else → `ProviderError` naming the finish reason and how
+	*   many parts arrived.
+	*/
+	noImageError(response, parts) {
+		const blockReason = response.promptFeedback?.blockReason;
+		if (blockReason) return new ContentFilterError(`Gemini blocked the prompt for ${this.name}: ${blockReason}`, { reason: blockReason });
+		const finishReason = response.candidates?.[0]?.finishReason;
+		if (finishReason && FILTERED_FINISH_REASONS.has(finishReason)) return new ContentFilterError(`Gemini filtered the image for ${this.name}: ${finishReason}`, { reason: finishReason });
+		const text = collectText(parts);
+		if (text) return new ProviderError(`Gemini returned no image for ${this.name} — the response was text only: "${excerpt(text)}"`, { context: {
+			model: this.name,
+			...finishReason ? { finishReason } : {}
+		} });
+		return new ProviderError(`Gemini returned no image part for ${this.name} (parts: ${parts.length}${finishReason ? `, finishReason: ${finishReason}` : ""}).`, { context: {
+			model: this.name,
+			parts: parts.length
+		} });
+	}
+};
+/**
+* Flatten every candidate's content parts into one list. Read off
+* `candidates[].content.parts` rather than the response's convenience
+* getters: `response.text` covers only the first candidate's text and
+* there is no getter for inline image data at all.
+*/
+function collectParts(response) {
+	const parts = [];
+	for (const candidate of response.candidates ?? []) parts.push(...candidate.content?.parts ?? []);
+	return parts;
+}
+/**
+* Map the inline image parts to the neutral `GeneratedImage[]` — the
+* same `{ type: "base64", base64, mediaType }` shape the Imagen path
+* emits, including its `image/png` fallback for a part that arrives
+* without a declared mime type.
+*/
+function toGeneratedImages(parts) {
+	const images = [];
+	for (const part of parts) {
+		const data = part.inlineData?.data;
+		if (!data) continue;
+		images.push({
+			type: "base64",
+			base64: data,
+			mediaType: part.inlineData?.mimeType ?? "image/png"
+		});
+	}
+	return images;
+}
+/** Join the text parts of a response — what the model said instead of drawing. */
+function collectText(parts) {
+	return parts.map((part) => part.text).filter((text) => typeof text === "string" && text.length > 0).join(" ").trim();
+}
+/** Trim a quoted model answer so an error message stays readable. */
+function excerpt(text) {
+	return text.length > TEXT_EXCERPT_LIMIT ? `${text.slice(0, TEXT_EXCERPT_LIMIT)}…` : text;
+}
+
+//#endregion
+export { GeminiImageModel };
+//# sourceMappingURL=gemini-image.mjs.map

package/esm/gemini-image.mjs.map

@@ -0,0 +1 @@
+{"version":3,"file":"gemini-image.mjs","names":[],"sources":["../../../../../../ai-google/src/gemini-image.ts"],"sourcesContent":["import {\n  ContentFilterError,\n  ProviderError,\n  type AIError,\n  type GeneratedImage,\n  type ImageGenerationOptions,\n  type ImageGenerationResponse,\n  type ImageModelContract,\n  type ImageModelPricing,\n  type Usage,\n} from \"@warlock.js/ai\";\nimport { log, type Logger } from \"@warlock.js/logger\";\nimport type {\n  GenerateContentConfig,\n  GenerateContentResponse,\n  GoogleGenAI,\n  ImageConfig,\n  Part,\n} from \"@google/genai\";\nimport type { GoogleImageConfig } from \"./config.type\";\nimport { applyGoogleUsage, wrapGoogleError } from \"./utils\";\n\nconst LOG_MODULE = \"ai.google\";\n\n/**\n * Response modalities requested when the caller names none.\n *\n * `IMAGE` is the modality this adapter extracts; `TEXT` rides along so\n * a model that narrates what it drew is not answering outside the set\n * it was granted (the narration is then dropped — only inline image\n * parts become `GeneratedImage`s).\n *\n * *Unverified:* which pairing any individual Gemini image model\n * requires is not established here — no spec or run in this package\n * touches the live API. `options.responseModalities` replaces this list\n * verbatim for a model that wants something else.\n */\nconst DEFAULT_RESPONSE_MODALITIES = [\"TEXT\", \"IMAGE\"];\n\n/**\n * Gemini `finishReason` values that mean generation was stopped by a\n * safety / policy rule rather than by the model simply not drawing.\n * Taken from the `FinishReason` enum in `@google/genai`'s own type\n * declarations, whose doc comments describe each of these as content\n * or image generation being \"stopped\" for safety, prohibited content,\n * recitation, blocklist, or SPII.\n */\nconst FILTERED_FINISH_REASONS = new Set([\n  \"SAFETY\",\n  \"IMAGE_SAFETY\",\n  \"PROHIBITED_CONTENT\",\n  \"IMAGE_PROHIBITED_CONTENT\",\n  \"RECITATION\",\n  \"IMAGE_RECITATION\",\n  \"BLOCKLIST\",\n  \"SPII\",\n]);\n\n/** How much of a text-only answer to quote back inside the error message. */\nconst TEXT_EXCERPT_LIMIT = 200;\n\n/**\n * Gemini-native implementation of `ImageModelContract`, via\n * `ai.models.generateContent` with `config.responseModalities`\n * including `\"IMAGE\"`.\n *\n * **Why a second image adapter.** `GoogleImageModel` calls\n * `ai.models.generateImages`, which the `@google/genai` bundle routes\n * to `{model}:predict` (`generateImages` → `generateImagesInternal` →\n * `formatMap('{model}:predict', …)`). A Gemini image model is not\n * served there: asking for one returns Google's\n * `404 … is not found for API version v1beta, or is not supported for\n * predict`. `generateContent` is the SDK's own named replacement — its\n * runtime deprecation notice for `generateImages` reads \"Please use the\n * generateContent method with image models instead\" — so that is the\n * transport this class speaks, hence a separate class rather than a\n * branch inside `image.ts`.\n *\n * **Same envelope.** Inline image parts are mapped to the identical\n * `GeneratedImage[]` shape `GoogleImageModel` produces, so `ai.image()`\n * callers see no difference between the two paths.\n *\n * **Token usage is passed through, not zeroed.** The Imagen path\n * returns a hard `{ 0, 0, 0 }` because Imagen reports no tokens at all;\n * here, whatever `usageMetadata` Google attaches is mapped by the same\n * {@link applyGoogleUsage} the chat model uses, and only an absent\n * block collapses to zeros. Price accordingly.\n *\n * **No model-id validation.** `config.name` is forwarded to\n * `generateContent` exactly as given; nothing here inspects it. An id\n * Google does not serve fails at Google, wrapped into the typed\n * `AIError` hierarchy — never with a local throw.\n *\n * **Evidence, in two tiers.** No spec in this package calls Google.\n * *Measured here:* a `gemini-*` image id, which 404s on the `predict`\n * transport, reached the model on this one and came back with a quota\n * error (HTTP 429) — the endpoint accepts the id. *Reported by the\n * maintainer:* once billing was enabled on the project, an image came\n * back end-to-end from an application running a locally linked build.\n * *Still unestablished:* whether these models report token usage — no\n * `usageMetadata` from a successful image call has been observed, so\n * the pass-through above is untested against a real response.\n *\n * @example\n * const model = new GeminiImageModel(ai, { name: \"gemini-3.1-flash-lite-image\" });\n * const { images, usage } = await model.generate(\"a red bicycle on a white background\");\n */\nexport class GeminiImageModel implements ImageModelContract {\n  public readonly name: string;\n  public readonly provider: string;\n  public readonly pricing?: ImageModelPricing;\n\n  private readonly ai: GoogleGenAI;\n  private readonly logger: Logger = log;\n\n  public constructor(ai: GoogleGenAI, config: GoogleImageConfig, provider: string = \"google\") {\n    this.ai = ai;\n    this.name = config.name;\n    this.provider = provider;\n    this.pricing = config.pricing;\n  }\n\n  public async generate(\n    prompt: string,\n    options?: ImageGenerationOptions,\n  ): Promise<ImageGenerationResponse> {\n    const config = this.buildConfig(options);\n\n    this.logger.debug(LOG_MODULE, \"image.request\", \"models.generateContent\", {\n      model: this.name,\n      responseModalities: config.responseModalities,\n    });\n\n    let response: GenerateContentResponse;\n\n    try {\n      // `contents` takes a bare string: the SDK's own `generateContent`\n      // example passes one (`contents: 'Why is the sky blue?'`).\n      response = await this.ai.models.generateContent({\n        model: this.name,\n        contents: prompt,\n        config,\n      });\n    } catch (thrown) {\n      const wrapped = wrapGoogleError(thrown);\n\n      this.logger.error(LOG_MODULE, \"image.error\", wrapped.message, {\n        code: wrapped.code,\n        context: wrapped.context,\n      });\n\n      throw wrapped;\n    }\n\n    const parts = collectParts(response);\n    const images = toGeneratedImages(parts);\n\n    if (images.length === 0) {\n      throw this.noImageError(response, parts);\n    }\n\n    const usage: Usage = { input: 0, output: 0, total: 0 };\n\n    if (response.usageMetadata) {\n      applyGoogleUsage(usage, response.usageMetadata);\n    }\n\n    this.logger.debug(LOG_MODULE, \"image.response\", \"models.generateContent succeeded\", {\n      images: images.length,\n      usage,\n    });\n\n    return { images, usage };\n  }\n\n  /**\n   * Assemble the `GenerateContentConfig` for an image turn: the\n   * requested modalities, the image-specific knobs Gemini exposes under\n   * `imageConfig`, and the cancellation handle.\n   *\n   * Three neutral options are deliberately NOT forwarded, because\n   * `GenerateContentConfig` / `ImageConfig` in `@google/genai` expose\n   * no equivalent for them on this path: `count` (no per-request image\n   * count — every inline image part the model does return is mapped),\n   * `negativePrompt` (an Imagen-only field), and `format`\n   * (`ImageConfig.outputMimeType` is documented \"not supported in\n   * Gemini API\"). Fold those intentions into the prompt instead.\n   */\n  private buildConfig(options: ImageGenerationOptions | undefined): GenerateContentConfig {\n    const imageConfig: ImageConfig = {};\n\n    if (options?.aspectRatio !== undefined) {\n      imageConfig.aspectRatio = options.aspectRatio;\n    }\n\n    // Provider passthroughs off the neutral options' index signature —\n    // forwarded verbatim, never re-spelled, so the value the caller\n    // wrote is the value Google rules on. `ImageConfig` documents\n    // `imageSize` as `1K`/`2K`/`4K` and `personGeneration` as\n    // `ALLOW_ALL`/`ALLOW_ADULT`/`ALLOW_NONE`.\n    if (typeof options?.imageSize === \"string\") {\n      imageConfig.imageSize = options.imageSize;\n    }\n\n    if (typeof options?.personGeneration === \"string\") {\n      imageConfig.personGeneration = options.personGeneration;\n    }\n\n    const requested = options?.responseModalities;\n    const responseModalities = Array.isArray(requested)\n      ? (requested as string[])\n      : DEFAULT_RESPONSE_MODALITIES;\n\n    return {\n      responseModalities,\n      ...(Object.keys(imageConfig).length > 0 ? { imageConfig } : {}),\n      ...(options?.signal ? { abortSignal: options.signal } : {}),\n    };\n  }\n\n  /**\n   * Build the typed error for a response that carried no inline image\n   * part. Never a silent empty success: the caller asked for an image\n   * and got something else, so the error names what actually came back.\n   *\n   * - A blocked prompt (`promptFeedback.blockReason`) or a\n   *   safety/policy `finishReason` → `ContentFilterError` carrying the\n   *   reason, matching how the Imagen path reports `raiFilteredReason`.\n   * - A text-only answer → `ProviderError` quoting the text, so the\n   *   log says what the model replied instead of guessing.\n   * - Anything else → `ProviderError` naming the finish reason and how\n   *   many parts arrived.\n   */\n  private noImageError(response: GenerateContentResponse, parts: Part[]): AIError {\n    const blockReason = response.promptFeedback?.blockReason;\n\n    if (blockReason) {\n      return new ContentFilterError(\n        `Gemini blocked the prompt for ${this.name}: ${blockReason}`,\n        { reason: blockReason },\n      );\n    }\n\n    const finishReason = response.candidates?.[0]?.finishReason;\n\n    if (finishReason && FILTERED_FINISH_REASONS.has(finishReason)) {\n      return new ContentFilterError(\n        `Gemini filtered the image for ${this.name}: ${finishReason}`,\n        { reason: finishReason },\n      );\n    }\n\n    const text = collectText(parts);\n\n    if (text) {\n      return new ProviderError(\n        `Gemini returned no image for ${this.name} — the response was text only: \"${excerpt(text)}\"`,\n        { context: { model: this.name, ...(finishReason ? { finishReason } : {}) } },\n      );\n    }\n\n    return new ProviderError(\n      `Gemini returned no image part for ${this.name} (parts: ${parts.length}${\n        finishReason ? `, finishReason: ${finishReason}` : \"\"\n      }).`,\n      { context: { model: this.name, parts: parts.length } },\n    );\n  }\n}\n\n/**\n * Flatten every candidate's content parts into one list. Read off\n * `candidates[].content.parts` rather than the response's convenience\n * getters: `response.text` covers only the first candidate's text and\n * there is no getter for inline image data at all.\n */\nfunction collectParts(response: GenerateContentResponse): Part[] {\n  const parts: Part[] = [];\n\n  for (const candidate of response.candidates ?? []) {\n    parts.push(...(candidate.content?.parts ?? []));\n  }\n\n  return parts;\n}\n\n/**\n * Map the inline image parts to the neutral `GeneratedImage[]` — the\n * same `{ type: \"base64\", base64, mediaType }` shape the Imagen path\n * emits, including its `image/png` fallback for a part that arrives\n * without a declared mime type.\n */\nfunction toGeneratedImages(parts: Part[]): GeneratedImage[] {\n  const images: GeneratedImage[] = [];\n\n  for (const part of parts) {\n    const data = part.inlineData?.data;\n\n    if (!data) {\n      continue;\n    }\n\n    images.push({\n      type: \"base64\",\n      base64: data,\n      mediaType: part.inlineData?.mimeType ?? \"image/png\",\n    });\n  }\n\n  return images;\n}\n\n/** Join the text parts of a response — what the model said instead of drawing. */\nfunction collectText(parts: Part[]): string {\n  return parts\n    .map((part) => part.text)\n    .filter((text): text is string => typeof text === \"string\" && text.length > 0)\n    .join(\" \")\n    .trim();\n}\n\n/** Trim a quoted model answer so an error message stays readable. */\nfunction excerpt(text: string): string {\n  return text.length > TEXT_EXCERPT_LIMIT ? `${text.slice(0, TEXT_EXCERPT_LIMIT)}…` : text;\n}\n"],"mappings":";;;;;;;AAsBA,MAAM,aAAa;;;;;;;;;;;;;;AAenB,MAAM,8BAA8B,CAAC,QAAQ,OAAO;;;;;;;;;AAUpD,MAAM,0BAA0B,IAAI,IAAI;CACtC;CACA;CACA;CACA;CACA;CACA;CACA;CACA;AACF,CAAC;;AAGD,MAAM,qBAAqB;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAgD3B,IAAa,mBAAb,MAA4D;CAQ1D,AAAO,YAAY,IAAiB,QAA2B,WAAmB,UAAU;gBAF1D;EAGhC,KAAK,KAAK;EACV,KAAK,OAAO,OAAO;EACnB,KAAK,WAAW;EAChB,KAAK,UAAU,OAAO;CACxB;CAEA,MAAa,SACX,QACA,SACkC;EAClC,MAAM,SAAS,KAAK,YAAY,OAAO;EAEvC,KAAK,OAAO,MAAM,YAAY,iBAAiB,0BAA0B;GACvE,OAAO,KAAK;GACZ,oBAAoB,OAAO;EAC7B,CAAC;EAED,IAAI;EAEJ,IAAI;GAGF,WAAW,MAAM,KAAK,GAAG,OAAO,gBAAgB;IAC9C,OAAO,KAAK;IACZ,UAAU;IACV;GACF,CAAC;EACH,SAAS,QAAQ;GACf,MAAM,UAAU,gBAAgB,MAAM;GAEtC,KAAK,OAAO,MAAM,YAAY,eAAe,QAAQ,SAAS;IAC5D,MAAM,QAAQ;IACd,SAAS,QAAQ;GACnB,CAAC;GAED,MAAM;EACR;EAEA,MAAM,QAAQ,aAAa,QAAQ;EACnC,MAAM,SAAS,kBAAkB,KAAK;EAEtC,IAAI,OAAO,WAAW,GACpB,MAAM,KAAK,aAAa,UAAU,KAAK;EAGzC,MAAM,QAAe;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EAErD,IAAI,SAAS,eACX,iBAAiB,OAAO,SAAS,aAAa;EAGhD,KAAK,OAAO,MAAM,YAAY,kBAAkB,oCAAoC;GAClF,QAAQ,OAAO;GACf;EACF,CAAC;EAED,OAAO;GAAE;GAAQ;EAAM;CACzB;;;;;;;;;;;;;;CAeA,AAAQ,YAAY,SAAoE;EACtF,MAAM,cAA2B,CAAC;EAElC,IAAI,SAAS,gBAAgB,QAC3B,YAAY,cAAc,QAAQ;EAQpC,IAAI,OAAO,SAAS,cAAc,UAChC,YAAY,YAAY,QAAQ;EAGlC,IAAI,OAAO,SAAS,qBAAqB,UACvC,YAAY,mBAAmB,QAAQ;EAGzC,MAAM,YAAY,SAAS;EAK3B,OAAO;GACL,oBALyB,MAAM,QAAQ,SAAS,IAC7C,YACD;GAIF,GAAI,OAAO,KAAK,WAAW,CAAC,CAAC,SAAS,IAAI,EAAE,YAAY,IAAI,CAAC;GAC7D,GAAI,SAAS,SAAS,EAAE,aAAa,QAAQ,OAAO,IAAI,CAAC;EAC3D;CACF;;;;;;;;;;;;;;CAeA,AAAQ,aAAa,UAAmC,OAAwB;EAC9E,MAAM,cAAc,SAAS,gBAAgB;EAE7C,IAAI,aACF,OAAO,IAAI,mBACT,iCAAiC,KAAK,KAAK,IAAI,eAC/C,EAAE,QAAQ,YAAY,CACxB;EAGF,MAAM,eAAe,SAAS,aAAa,EAAE,EAAE;EAE/C,IAAI,gBAAgB,wBAAwB,IAAI,YAAY,GAC1D,OAAO,IAAI,mBACT,iCAAiC,KAAK,KAAK,IAAI,gBAC/C,EAAE,QAAQ,aAAa,CACzB;EAGF,MAAM,OAAO,YAAY,KAAK;EAE9B,IAAI,MACF,OAAO,IAAI,cACT,gCAAgC,KAAK,KAAK,kCAAkC,QAAQ,IAAI,EAAE,IAC1F,EAAE,SAAS;GAAE,OAAO,KAAK;GAAM,GAAI,eAAe,EAAE,aAAa,IAAI,CAAC;EAAG,EAAE,CAC7E;EAGF,OAAO,IAAI,cACT,qCAAqC,KAAK,KAAK,WAAW,MAAM,SAC9D,eAAe,mBAAmB,iBAAiB,GACpD,KACD,EAAE,SAAS;GAAE,OAAO,KAAK;GAAM,OAAO,MAAM;EAAO,EAAE,CACvD;CACF;AACF;;;;;;;AAQA,SAAS,aAAa,UAA2C;CAC/D,MAAM,QAAgB,CAAC;CAEvB,KAAK,MAAM,aAAa,SAAS,cAAc,CAAC,GAC9C,MAAM,KAAK,GAAI,UAAU,SAAS,SAAS,CAAC,CAAE;CAGhD,OAAO;AACT;;;;;;;AAQA,SAAS,kBAAkB,OAAiC;CAC1D,MAAM,SAA2B,CAAC;CAElC,KAAK,MAAM,QAAQ,OAAO;EACxB,MAAM,OAAO,KAAK,YAAY;EAE9B,IAAI,CAAC,MACH;EAGF,OAAO,KAAK;GACV,MAAM;GACN,QAAQ;GACR,WAAW,KAAK,YAAY,YAAY;EAC1C,CAAC;CACH;CAEA,OAAO;AACT;;AAGA,SAAS,YAAY,OAAuB;CAC1C,OAAO,MACJ,KAAK,SAAS,KAAK,IAAI,CAAC,CACxB,QAAQ,SAAyB,OAAO,SAAS,YAAY,KAAK,SAAS,CAAC,CAAC,CAC7E,KAAK,GAAG,CAAC,CACT,KAAK;AACV;;AAGA,SAAS,QAAQ,MAAsB;CACrC,OAAO,KAAK,SAAS,qBAAqB,GAAG,KAAK,MAAM,GAAG,kBAAkB,EAAE,KAAK;AACtF"}

package/esm/index.d.mts

@@ -1,4 +1,5 @@
 import { GoogleEmbedderConfig, GoogleImageConfig, GoogleModelConfig, GoogleSDKConfig } from "./config.type.mjs";
 import { GoogleSDK } from "./sdk.mjs";
+import { GeminiImageModel } from "./gemini-image.mjs";
 import { GoogleImageModel } from "./image.mjs";
-export { type GoogleEmbedderConfig, type GoogleImageConfig, GoogleImageModel, type GoogleModelConfig, GoogleSDK, type GoogleSDKConfig };
+export { GeminiImageModel, type GoogleEmbedderConfig, type GoogleImageConfig, GoogleImageModel, type GoogleModelConfig, GoogleSDK, type GoogleSDKConfig };

package/esm/index.mjs

@@ -1,4 +1,5 @@
+import { GeminiImageModel } from "./gemini-image.mjs";
 import { GoogleImageModel } from "./image.mjs";
 import { GoogleSDK } from "./sdk.mjs";
 
-export { GoogleImageModel, GoogleSDK };
+export { GeminiImageModel, GoogleImageModel, GoogleSDK };

package/esm/model.mjs

@@ -1,3 +1,4 @@
+import { applyGoogleUsage } from "./utils/apply-google-usage.mjs";
 import { mapFinishReason } from "./utils/map-finish-reason.mjs";
 import { toGoogleContents } from "./utils/to-google-contents.mjs";
 import { toGoogleTools } from "./utils/to-google-tools.mjs";
@@ -159,7 +160,7 @@ var GoogleModel = class {
 				}
 				const candidateFinish = chunk.candidates?.[0]?.finishReason;
 				if (candidateFinish) rawFinishReason = candidateFinish;
-				if (chunk.usageMetadata) this.applyUsage(usage, chunk.usageMetadata);
+				if (chunk.usageMetadata) applyGoogleUsage(usage, chunk.usageMetadata);
 			}
 		} catch (thrown) {
 			throw this.logAndWrap(thrown);
@@ -272,9 +273,10 @@ var GoogleModel = class {
 		};
 	}
 	/**
-	* Normalize Gemini's `usageMetadata` into the neutral `Usage` shape.
-	* Cache-read tokens are surfaced as `cachedTokens` only when
-	* non-zero. Absent usage collapses to zeros.
+	* Normalize Gemini's `usageMetadata` into the neutral `Usage` shape
+	* via the shared {@link applyGoogleUsage} mapper (the same one the
+	* streaming loop and the Gemini image model use). Absent usage
+	* collapses to zeros.
 	*/
 	extractUsage(response) {
 		const usage = {
@@ -282,30 +284,10 @@ var GoogleModel = class {
 			output: 0,
 			total: 0
 		};
-		if (response.usageMetadata) this.applyUsage(usage, response.usageMetadata);
+		if (response.usageMetadata) applyGoogleUsage(usage, response.usageMetadata);
 		return usage;
 	}
 	/**
-	* Fold a Gemini `usageMetadata` block into the running neutral
-	* `Usage` accumulator. Shared by `complete()` and the streaming
-	* loop (where the final chunk carries cumulative totals).
-	*
-	* Cache-read hits (`cachedContentTokenCount`, implicit or explicit
-	* context caching) surface as `cachedTokens`; the thinking-phase
-	* tokens of a reasoning model (`thoughtsTokenCount`) surface as
-	* `reasoningTokens`. Both are emitted only when reported `> 0` so an
-	* absent channel leaves the field undefined.
-	*/
-	applyUsage(usage, raw) {
-		usage.input = raw.promptTokenCount ?? usage.input;
-		usage.output = raw.candidatesTokenCount ?? usage.output;
-		usage.total = raw.totalTokenCount ?? usage.input + usage.output;
-		const cached = raw.cachedContentTokenCount;
-		if (cached && cached > 0) usage.cachedTokens = cached;
-		const reasoning = raw.thoughtsTokenCount;
-		if (reasoning && reasoning > 0) usage.reasoningTokens = reasoning;
-	}
-	/**
 	* Wrap a thrown provider error into the typed `AIError` hierarchy
 	* and emit the standard error log line before it propagates.
 	*/

package/esm/model.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"model.mjs","names":[],"sources":["../../../../../../ai-google/src/model.ts"],"sourcesContent":["import {\n  type Message,\n  type ModelCallOptions,\n  type ModelCapabilities,\n  type ModelContract,\n  type ModelPricing,\n  type ModelResponse,\n  type ModelStreamChunk,\n  type ModelToolCallRequest,\n  type ReasoningEffort,\n  type Usage,\n} from \"@warlock.js/ai\";\nimport { log, type Logger } from \"@warlock.js/logger\";\nimport type {\n  GenerateContentConfig,\n  GenerateContentResponse,\n  GoogleGenAI,\n  Part,\n} from \"@google/genai\";\nimport type { GoogleModelConfig } from \"./config.type\";\nimport { inferVisionCapability } from \"./known-vision-models\";\nimport { mapFinishReason, toGoogleContents, toGoogleTools, wrapGoogleError } from \"./utils\";\n\nconst LOG_MODULE = \"ai.google\";\n\n/**\n * Bucketed `thinkingBudget` (token caps) for the neutral\n * `reasoning.effort` levels when the caller gives no explicit\n * `reasoning.maxTokens`. Gemini 2.5 accepts a positive budget as a cap\n * on the thinking phase; these mirror the spread the OpenAI\n * `reasoning_effort` low/medium/high tiers imply.\n */\nconst EFFORT_THINKING_BUDGET: Record<Exclude<ReasoningEffort, \"none\">, number> = {\n  low: 1024,\n  medium: 8192,\n  high: 24576,\n};\n\n/**\n * Google Gemini-backed implementation of `ModelContract`.\n *\n * **Role.** The provider-facing bridge between the vendor-neutral\n * `@warlock.js/ai` agent runtime and the `@google/genai` SDK\n * (`models.generateContent` / `generateContentStream`).\n *\n * **Responsibility.**\n * - Owns: a long-lived `GoogleGenAI` client + frozen `ModelConfig`\n *   (name, temperature, maxTokens) used as per-call defaults.\n * - Owns: translating vendor-neutral `Message[]` / `ToolConfig[]` into\n *   Gemini shapes (systemInstruction hoisting, `model` role,\n *   `functionCall` / `functionResponse` parts, inline image bytes) on\n *   the way out, and Gemini's candidate/parts response (text, function\n *   calls, finish reason, token usage) back into neutral shapes on the\n *   way in.\n * - Does NOT own: dispatching tools, looping, history, retries — those\n *   are agent concerns. The model is a per-call protocol adapter.\n *\n * Modeled as a class (see §4.2 of code-style.md — \"long-lived state\n * across calls\"): the `GoogleGenAI` client is reused for the SDK's\n * lifetime.\n *\n * @example\n * import { GoogleGenAI } from \"@google/genai\";\n * const ai = new GoogleGenAI({ apiKey: process.env.GEMINI_API_KEY });\n * const model = new GoogleModel(ai, { name: \"gemini-2.5-flash\" });\n *\n * const myAgent = agent({ model, tools: [searchTool] });\n * const result = await myAgent.execute(\"Summarize today's news.\");\n */\nexport class GoogleModel implements ModelContract {\n  public readonly name: string;\n  public readonly provider: string;\n  public readonly capabilities: ModelCapabilities;\n  public readonly pricing?: ModelPricing;\n\n  private readonly ai: GoogleGenAI;\n  private readonly config: GoogleModelConfig;\n  private readonly logger: Logger = log;\n\n  public constructor(ai: GoogleGenAI, config: GoogleModelConfig, provider: string = \"google\") {\n    this.ai = ai;\n    this.config = config;\n    this.name = config.name;\n    this.provider = provider;\n    this.pricing = config.pricing;\n    const multimodal = config.vision ?? inferVisionCapability(config.name);\n\n    this.capabilities = {\n      structuredOutput: config.structuredOutput ?? true,\n      vision: multimodal,\n      // Every Gemini 2.5 model thinks; older families harmlessly ignore\n      // an empty thinking budget. Defaulting `true` lets the agent\n      // forward reasoning options; an explicit `false` opts a model out.\n      reasoning: config.reasoning ?? true,\n      // Gemini reports cache-read hits (`cachedContentTokenCount`) on\n      // every call via implicit caching, and accepts explicit context\n      // caching. Read-side accounting is always honored.\n      promptCaching: true,\n      // The multimodal Gemini families that accept images also accept\n      // audio and PDF/document parts. Mirror the vision inference unless\n      // explicitly overridden.\n      audio: config.audio ?? multimodal,\n      pdf: config.pdf ?? multimodal,\n    };\n  }\n\n  /**\n   * Single-shot completion. Sends the full message list to\n   * `generateContent`, waits for the terminal response, and reshapes\n   * it into a vendor-neutral `ModelResponse`. Per-call `options`\n   * override the instance defaults for this call only.\n   */\n  public async complete(messages: Message[], options?: ModelCallOptions): Promise<ModelResponse> {\n    this.logger.debug(LOG_MODULE, \"request\", \"Starting generateContent call\", {\n      model: this.name,\n      messageCount: messages.length,\n      streaming: false,\n      toolCount: options?.tools?.length ?? 0,\n    });\n\n    const { systemInstruction, contents } = toGoogleContents(messages);\n\n    let response: GenerateContentResponse;\n\n    try {\n      response = await this.ai.models.generateContent({\n        model: this.name,\n        contents,\n        config: this.buildConfig(systemInstruction, options),\n      });\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    const toolCalls = this.extractToolCalls(response);\n    const finishReason = toolCalls\n      ? \"tool_calls\"\n      : mapFinishReason(response.candidates?.[0]?.finishReason);\n    const usage = this.extractUsage(response);\n\n    this.logger.debug(LOG_MODULE, \"response\", \"generateContent call succeeded\", {\n      finishReason,\n      usage,\n    });\n\n    return {\n      content: response.text ?? \"\",\n      finishReason,\n      usage,\n      toolCalls,\n    };\n  }\n\n  /**\n   * Incremental streaming completion via `generateContentStream`.\n   * Yields neutral `ModelStreamChunk`s — `delta` for text, `tool-call`\n   * per function call (Gemini emits a fully-formed call, not partial\n   * JSON), and a terminal `done` with the final finish reason + usage.\n   */\n  public async *stream(\n    messages: Message[],\n    options?: ModelCallOptions,\n  ): AsyncIterable<ModelStreamChunk> {\n    this.logger.debug(LOG_MODULE, \"request\", \"Starting generateContentStream call\", {\n      model: this.name,\n      messageCount: messages.length,\n      streaming: true,\n      toolCount: options?.tools?.length ?? 0,\n    });\n\n    const { systemInstruction, contents } = toGoogleContents(messages);\n\n    let iterable: AsyncGenerator<GenerateContentResponse>;\n\n    try {\n      iterable = await this.ai.models.generateContentStream({\n        model: this.name,\n        contents,\n        config: this.buildConfig(systemInstruction, options),\n      });\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    let rawFinishReason: string | undefined;\n    let sawToolCall = false;\n    const usage: Usage = { input: 0, output: 0, total: 0 };\n\n    try {\n      for await (const chunk of iterable) {\n        const text = chunk.text;\n\n        if (text) {\n          yield { type: \"delta\", content: text };\n        }\n\n        for (const part of chunk.candidates?.[0]?.content?.parts ?? []) {\n          const toolCall = this.partToToolCall(part);\n\n          if (!toolCall) {\n            continue;\n          }\n\n          sawToolCall = true;\n\n          yield {\n            type: \"tool-call\",\n            id: toolCall.id,\n            name: toolCall.name,\n            input: toolCall.input,\n            ...(toolCall.providerMetadata\n              ? { providerMetadata: toolCall.providerMetadata }\n              : {}),\n          };\n        }\n\n        const candidateFinish = chunk.candidates?.[0]?.finishReason;\n\n        if (candidateFinish) {\n          rawFinishReason = candidateFinish;\n        }\n\n        if (chunk.usageMetadata) {\n          this.applyUsage(usage, chunk.usageMetadata);\n        }\n      }\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    const finishReason = sawToolCall ? \"tool_calls\" : mapFinishReason(rawFinishReason);\n\n    this.logger.debug(LOG_MODULE, \"response\", \"generateContentStream call succeeded\", {\n      finishReason,\n      usage,\n    });\n\n    yield { type: \"done\", finishReason, usage };\n  }\n\n  /**\n   * Assemble the `GenerateContentConfig` shared by `complete()` and\n   * `stream()`: inference params, hoisted system instruction,\n   * cancellation signal, and conditional tools + native structured\n   * output.\n   */\n  private buildConfig(\n    systemInstruction: string | undefined,\n    options: ModelCallOptions | undefined,\n  ): GenerateContentConfig {\n    const temperature = options?.temperature ?? this.config.temperature;\n    const maxOutputTokens = options?.maxTokens ?? this.config.maxTokens;\n\n    return {\n      ...(systemInstruction ? { systemInstruction } : {}),\n      ...(temperature !== undefined ? { temperature } : {}),\n      ...(maxOutputTokens !== undefined ? { maxOutputTokens } : {}),\n      ...(options?.signal ? { abortSignal: options.signal } : {}),\n      ...this.buildTools(options?.tools),\n      ...this.buildStructuredOutput(options?.responseSchema),\n      ...this.buildThinking(options?.reasoning),\n    };\n  }\n\n  /**\n   * Translate the neutral `reasoning` option into Gemini's\n   * `thinkingConfig`. `reasoning.maxTokens` maps directly to\n   * `thinkingBudget` (token cap on the thinking phase); when only\n   * `reasoning.effort` is given it is bucketed into a budget. Emitted\n   * only when the model is `reasoning`-capable — a `false` capability\n   * (config override) drops it so a non-thinking model never receives\n   * an unsupported `thinkingConfig`.\n   *\n   * Gemini's `thinkingBudget` semantics: `0` disables thinking, `-1`\n   * lets the model decide automatically. A positive value caps the\n   * thinking tokens. The neutral `effort: \"none\"` (\"run without\n   * reasoning\") maps to `thinkingBudget: 0`.\n   */\n  private buildThinking(\n    reasoning: ModelCallOptions[\"reasoning\"],\n  ): Pick<GenerateContentConfig, \"thinkingConfig\"> {\n    if (!reasoning || !this.capabilities.reasoning) {\n      return {};\n    }\n\n    // `effort: \"none\"` = explicit \"run without reasoning\". Gemini disables\n    // thinking with `thinkingBudget: 0` (its native off switch), so emit\n    // that rather than omitting the config — an omitted config lets a\n    // thinking model reason at its default budget.\n    if (reasoning.effort === \"none\") {\n      return { thinkingConfig: { thinkingBudget: 0 } };\n    }\n\n    const thinkingBudget =\n      reasoning.maxTokens ?? (reasoning.effort ? EFFORT_THINKING_BUDGET[reasoning.effort] : undefined);\n\n    if (thinkingBudget === undefined) {\n      return {};\n    }\n\n    return { thinkingConfig: { thinkingBudget } };\n  }\n\n  /**\n   * Spread-friendly tools fragment. Empty object when no tools were\n   * supplied so the caller can unconditionally spread it.\n   */\n  private buildTools(tools: ModelCallOptions[\"tools\"]): Pick<GenerateContentConfig, \"tools\"> {\n    const mapped = toGoogleTools(tools);\n\n    return mapped ? { tools: mapped } : {};\n  }\n\n  /**\n   * Translate the neutral `responseSchema` into Gemini's native JSON\n   * structured output (`responseMimeType: \"application/json\"` +\n   * `responseJsonSchema`, which takes a raw JSON Schema directly).\n   * Emitted only when the model is `structuredOutput`-capable and the\n   * schema is an object root — otherwise the agent's soft prompt hint\n   * + client-side `validate()` carry shape.\n   */\n  private buildStructuredOutput(\n    responseSchema: Record<string, unknown> | undefined,\n  ): Pick<GenerateContentConfig, \"responseMimeType\" | \"responseJsonSchema\"> {\n    if (!responseSchema || !this.capabilities.structuredOutput) {\n      return {};\n    }\n\n    if (responseSchema.type !== \"object\" || typeof responseSchema.properties !== \"object\") {\n      return {};\n    }\n\n    return {\n      responseMimeType: \"application/json\",\n      responseJsonSchema: responseSchema,\n    };\n  }\n\n  /**\n   * Reshape Gemini's function-call content parts into the neutral\n   * `ModelToolCallRequest[]`. Returns `undefined` when the model\n   * requested no functions so callers can branch on presence.\n   *\n   * Reads `candidates[0].content.parts` directly rather than the\n   * `response.functionCalls` getter: the getter discards the\n   * part-level `thoughtSignature`, and Gemini \"thinking\" models 400\n   * the follow-up turn if that signature is not echoed back. See\n   * `partToToolCall`.\n   */\n  private extractToolCalls(\n    response: GenerateContentResponse,\n  ): ModelToolCallRequest[] | undefined {\n    const parts = response.candidates?.[0]?.content?.parts ?? [];\n    const toolCalls = parts\n      .map((part) => this.partToToolCall(part))\n      .filter((call): call is ModelToolCallRequest => call !== undefined);\n\n    return toolCalls.length > 0 ? toolCalls : undefined;\n  }\n\n  /**\n   * Map a single Gemini `Part` to a neutral `ModelToolCallRequest`,\n   * or `undefined` when the part is not a function call. The part's\n   * `thoughtSignature` (opaque, set by thinking models) is carried on\n   * `providerMetadata` so `toGoogleContents` can replay it on the\n   * assistant turn — Gemini rejects the next request without it.\n   */\n  private partToToolCall(part: Part): ModelToolCallRequest | undefined {\n    if (!part.functionCall) {\n      return undefined;\n    }\n\n    const call = part.functionCall;\n\n    return {\n      // The Gemini Developer API does not assign function-call ids\n      // (only Vertex parallel-calling does). Fall back to the function\n      // name so the neutral `toolCallId` is non-empty and the echoed\n      // `functionResponse.name` resolves — Gemini matches a result to\n      // its call by name. See decisions §49.\n      id: call.id ?? call.name ?? \"\",\n      name: call.name ?? \"\",\n      input: (call.args ?? {}) as Record<string, unknown>,\n      ...(part.thoughtSignature\n        ? { providerMetadata: { thoughtSignature: part.thoughtSignature } }\n        : {}),\n    };\n  }\n\n  /**\n   * Normalize Gemini's `usageMetadata` into the neutral `Usage` shape.\n   * Cache-read tokens are surfaced as `cachedTokens` only when\n   * non-zero. Absent usage collapses to zeros.\n   */\n  private extractUsage(response: GenerateContentResponse): Usage {\n    const usage: Usage = { input: 0, output: 0, total: 0 };\n\n    if (response.usageMetadata) {\n      this.applyUsage(usage, response.usageMetadata);\n    }\n\n    return usage;\n  }\n\n  /**\n   * Fold a Gemini `usageMetadata` block into the running neutral\n   * `Usage` accumulator. Shared by `complete()` and the streaming\n   * loop (where the final chunk carries cumulative totals).\n   *\n   * Cache-read hits (`cachedContentTokenCount`, implicit or explicit\n   * context caching) surface as `cachedTokens`; the thinking-phase\n   * tokens 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.\n   */\n  private applyUsage(\n    usage: Usage,\n    raw: NonNullable<GenerateContentResponse[\"usageMetadata\"]>,\n  ): 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\n  /**\n   * Wrap a thrown provider error into the typed `AIError` hierarchy\n   * and emit the standard error log line before it propagates.\n   */\n  private logAndWrap(thrown: unknown) {\n    const wrapped = wrapGoogleError(thrown);\n\n    this.logger.error(LOG_MODULE, \"error\", wrapped.message, {\n      code: wrapped.code,\n      context: wrapped.context,\n    });\n\n    return wrapped;\n  }\n}\n"],"mappings":";;;;;;;;;AAuBA,MAAM,aAAa;;;;;;;;AASnB,MAAM,yBAA2E;CAC/E,KAAK;CACL,QAAQ;CACR,MAAM;AACR;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAiCA,IAAa,cAAb,MAAkD;CAUhD,AAAO,YAAY,IAAiB,QAA2B,WAAmB,UAAU;gBAF1D;EAGhC,KAAK,KAAK;EACV,KAAK,SAAS;EACd,KAAK,OAAO,OAAO;EACnB,KAAK,WAAW;EAChB,KAAK,UAAU,OAAO;EACtB,MAAM,aAAa,OAAO,UAAU,sBAAsB,OAAO,IAAI;EAErE,KAAK,eAAe;GAClB,kBAAkB,OAAO,oBAAoB;GAC7C,QAAQ;GAIR,WAAW,OAAO,aAAa;GAI/B,eAAe;GAIf,OAAO,OAAO,SAAS;GACvB,KAAK,OAAO,OAAO;EACrB;CACF;;;;;;;CAQA,MAAa,SAAS,UAAqB,SAAoD;EAC7F,KAAK,OAAO,MAAM,YAAY,WAAW,iCAAiC;GACxE,OAAO,KAAK;GACZ,cAAc,SAAS;GACvB,WAAW;GACX,WAAW,SAAS,OAAO,UAAU;EACvC,CAAC;EAED,MAAM,EAAE,mBAAmB,aAAa,iBAAiB,QAAQ;EAEjE,IAAI;EAEJ,IAAI;GACF,WAAW,MAAM,KAAK,GAAG,OAAO,gBAAgB;IAC9C,OAAO,KAAK;IACZ;IACA,QAAQ,KAAK,YAAY,mBAAmB,OAAO;GACrD,CAAC;EACH,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,MAAM,YAAY,KAAK,iBAAiB,QAAQ;EAChD,MAAM,eAAe,YACjB,eACA,gBAAgB,SAAS,aAAa,EAAE,EAAE,YAAY;EAC1D,MAAM,QAAQ,KAAK,aAAa,QAAQ;EAExC,KAAK,OAAO,MAAM,YAAY,YAAY,kCAAkC;GAC1E;GACA;EACF,CAAC;EAED,OAAO;GACL,SAAS,SAAS,QAAQ;GAC1B;GACA;GACA;EACF;CACF;;;;;;;CAQA,OAAc,OACZ,UACA,SACiC;EACjC,KAAK,OAAO,MAAM,YAAY,WAAW,uCAAuC;GAC9E,OAAO,KAAK;GACZ,cAAc,SAAS;GACvB,WAAW;GACX,WAAW,SAAS,OAAO,UAAU;EACvC,CAAC;EAED,MAAM,EAAE,mBAAmB,aAAa,iBAAiB,QAAQ;EAEjE,IAAI;EAEJ,IAAI;GACF,WAAW,MAAM,KAAK,GAAG,OAAO,sBAAsB;IACpD,OAAO,KAAK;IACZ;IACA,QAAQ,KAAK,YAAY,mBAAmB,OAAO;GACrD,CAAC;EACH,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,IAAI;EACJ,IAAI,cAAc;EAClB,MAAM,QAAe;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EAErD,IAAI;GACF,WAAW,MAAM,SAAS,UAAU;IAClC,MAAM,OAAO,MAAM;IAEnB,IAAI,MACF,MAAM;KAAE,MAAM;KAAS,SAAS;IAAK;IAGvC,KAAK,MAAM,QAAQ,MAAM,aAAa,EAAE,EAAE,SAAS,SAAS,CAAC,GAAG;KAC9D,MAAM,WAAW,KAAK,eAAe,IAAI;KAEzC,IAAI,CAAC,UACH;KAGF,cAAc;KAEd,MAAM;MACJ,MAAM;MACN,IAAI,SAAS;MACb,MAAM,SAAS;MACf,OAAO,SAAS;MAChB,GAAI,SAAS,mBACT,EAAE,kBAAkB,SAAS,iBAAiB,IAC9C,CAAC;KACP;IACF;IAEA,MAAM,kBAAkB,MAAM,aAAa,EAAE,EAAE;IAE/C,IAAI,iBACF,kBAAkB;IAGpB,IAAI,MAAM,eACR,KAAK,WAAW,OAAO,MAAM,aAAa;GAE9C;EACF,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,MAAM,eAAe,cAAc,eAAe,gBAAgB,eAAe;EAEjF,KAAK,OAAO,MAAM,YAAY,YAAY,wCAAwC;GAChF;GACA;EACF,CAAC;EAED,MAAM;GAAE,MAAM;GAAQ;GAAc;EAAM;CAC5C;;;;;;;CAQA,AAAQ,YACN,mBACA,SACuB;EACvB,MAAM,cAAc,SAAS,eAAe,KAAK,OAAO;EACxD,MAAM,kBAAkB,SAAS,aAAa,KAAK,OAAO;EAE1D,OAAO;GACL,GAAI,oBAAoB,EAAE,kBAAkB,IAAI,CAAC;GACjD,GAAI,gBAAgB,SAAY,EAAE,YAAY,IAAI,CAAC;GACnD,GAAI,oBAAoB,SAAY,EAAE,gBAAgB,IAAI,CAAC;GAC3D,GAAI,SAAS,SAAS,EAAE,aAAa,QAAQ,OAAO,IAAI,CAAC;GACzD,GAAG,KAAK,WAAW,SAAS,KAAK;GACjC,GAAG,KAAK,sBAAsB,SAAS,cAAc;GACrD,GAAG,KAAK,cAAc,SAAS,SAAS;EAC1C;CACF;;;;;;;;;;;;;;;CAgBA,AAAQ,cACN,WAC+C;EAC/C,IAAI,CAAC,aAAa,CAAC,KAAK,aAAa,WACnC,OAAO,CAAC;EAOV,IAAI,UAAU,WAAW,QACvB,OAAO,EAAE,gBAAgB,EAAE,gBAAgB,EAAE,EAAE;EAGjD,MAAM,iBACJ,UAAU,cAAc,UAAU,SAAS,uBAAuB,UAAU,UAAU;EAExF,IAAI,mBAAmB,QACrB,OAAO,CAAC;EAGV,OAAO,EAAE,gBAAgB,EAAE,eAAe,EAAE;CAC9C;;;;;CAMA,AAAQ,WAAW,OAAwE;EACzF,MAAM,SAAS,cAAc,KAAK;EAElC,OAAO,SAAS,EAAE,OAAO,OAAO,IAAI,CAAC;CACvC;;;;;;;;;CAUA,AAAQ,sBACN,gBACwE;EACxE,IAAI,CAAC,kBAAkB,CAAC,KAAK,aAAa,kBACxC,OAAO,CAAC;EAGV,IAAI,eAAe,SAAS,YAAY,OAAO,eAAe,eAAe,UAC3E,OAAO,CAAC;EAGV,OAAO;GACL,kBAAkB;GAClB,oBAAoB;EACtB;CACF;;;;;;;;;;;;CAaA,AAAQ,iBACN,UACoC;EAEpC,MAAM,aADQ,SAAS,aAAa,EAAE,EAAE,SAAS,SAAS,CAAC,EACpC,CACpB,KAAK,SAAS,KAAK,eAAe,IAAI,CAAC,CAAC,CACxC,QAAQ,SAAuC,SAAS,MAAS;EAEpE,OAAO,UAAU,SAAS,IAAI,YAAY;CAC5C;;;;;;;;CASA,AAAQ,eAAe,MAA8C;EACnE,IAAI,CAAC,KAAK,cACR;EAGF,MAAM,OAAO,KAAK;EAElB,OAAO;GAML,IAAI,KAAK,MAAM,KAAK,QAAQ;GAC5B,MAAM,KAAK,QAAQ;GACnB,OAAQ,KAAK,QAAQ,CAAC;GACtB,GAAI,KAAK,mBACL,EAAE,kBAAkB,EAAE,kBAAkB,KAAK,iBAAiB,EAAE,IAChE,CAAC;EACP;CACF;;;;;;CAOA,AAAQ,aAAa,UAA0C;EAC7D,MAAM,QAAe;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EAErD,IAAI,SAAS,eACX,KAAK,WAAW,OAAO,SAAS,aAAa;EAG/C,OAAO;CACT;;;;;;;;;;;;CAaA,AAAQ,WACN,OACA,KACM;EACN,MAAM,QAAQ,IAAI,oBAAoB,MAAM;EAC5C,MAAM,SAAS,IAAI,wBAAwB,MAAM;EACjD,MAAM,QAAQ,IAAI,mBAAmB,MAAM,QAAQ,MAAM;EAEzD,MAAM,SAAS,IAAI;EAEnB,IAAI,UAAU,SAAS,GACrB,MAAM,eAAe;EAGvB,MAAM,YAAY,IAAI;EAEtB,IAAI,aAAa,YAAY,GAC3B,MAAM,kBAAkB;CAE5B;;;;;CAMA,AAAQ,WAAW,QAAiB;EAClC,MAAM,UAAU,gBAAgB,MAAM;EAEtC,KAAK,OAAO,MAAM,YAAY,SAAS,QAAQ,SAAS;GACtD,MAAM,QAAQ;GACd,SAAS,QAAQ;EACnB,CAAC;EAED,OAAO;CACT;AACF"}
+{"version":3,"file":"model.mjs","names":[],"sources":["../../../../../../ai-google/src/model.ts"],"sourcesContent":["import {\n  type Message,\n  type ModelCallOptions,\n  type ModelCapabilities,\n  type ModelContract,\n  type ModelPricing,\n  type ModelResponse,\n  type ModelStreamChunk,\n  type ModelToolCallRequest,\n  type ReasoningEffort,\n  type Usage,\n} from \"@warlock.js/ai\";\nimport { log, type Logger } from \"@warlock.js/logger\";\nimport type {\n  GenerateContentConfig,\n  GenerateContentResponse,\n  GoogleGenAI,\n  Part,\n} from \"@google/genai\";\nimport type { GoogleModelConfig } from \"./config.type\";\nimport { inferVisionCapability } from \"./known-vision-models\";\nimport {\n  applyGoogleUsage,\n  mapFinishReason,\n  toGoogleContents,\n  toGoogleTools,\n  wrapGoogleError,\n} from \"./utils\";\n\nconst LOG_MODULE = \"ai.google\";\n\n/**\n * Bucketed `thinkingBudget` (token caps) for the neutral\n * `reasoning.effort` levels when the caller gives no explicit\n * `reasoning.maxTokens`. Gemini 2.5 accepts a positive budget as a cap\n * on the thinking phase; these mirror the spread the OpenAI\n * `reasoning_effort` low/medium/high tiers imply.\n */\nconst EFFORT_THINKING_BUDGET: Record<Exclude<ReasoningEffort, \"none\">, number> = {\n  low: 1024,\n  medium: 8192,\n  high: 24576,\n};\n\n/**\n * Google Gemini-backed implementation of `ModelContract`.\n *\n * **Role.** The provider-facing bridge between the vendor-neutral\n * `@warlock.js/ai` agent runtime and the `@google/genai` SDK\n * (`models.generateContent` / `generateContentStream`).\n *\n * **Responsibility.**\n * - Owns: a long-lived `GoogleGenAI` client + frozen `ModelConfig`\n *   (name, temperature, maxTokens) used as per-call defaults.\n * - Owns: translating vendor-neutral `Message[]` / `ToolConfig[]` into\n *   Gemini shapes (systemInstruction hoisting, `model` role,\n *   `functionCall` / `functionResponse` parts, inline image bytes) on\n *   the way out, and Gemini's candidate/parts response (text, function\n *   calls, finish reason, token usage) back into neutral shapes on the\n *   way in.\n * - Does NOT own: dispatching tools, looping, history, retries — those\n *   are agent concerns. The model is a per-call protocol adapter.\n *\n * Modeled as a class (see §4.2 of code-style.md — \"long-lived state\n * across calls\"): the `GoogleGenAI` client is reused for the SDK's\n * lifetime.\n *\n * @example\n * import { GoogleGenAI } from \"@google/genai\";\n * const ai = new GoogleGenAI({ apiKey: process.env.GEMINI_API_KEY });\n * const model = new GoogleModel(ai, { name: \"gemini-2.5-flash\" });\n *\n * const myAgent = agent({ model, tools: [searchTool] });\n * const result = await myAgent.execute(\"Summarize today's news.\");\n */\nexport class GoogleModel implements ModelContract {\n  public readonly name: string;\n  public readonly provider: string;\n  public readonly capabilities: ModelCapabilities;\n  public readonly pricing?: ModelPricing;\n\n  private readonly ai: GoogleGenAI;\n  private readonly config: GoogleModelConfig;\n  private readonly logger: Logger = log;\n\n  public constructor(ai: GoogleGenAI, config: GoogleModelConfig, provider: string = \"google\") {\n    this.ai = ai;\n    this.config = config;\n    this.name = config.name;\n    this.provider = provider;\n    this.pricing = config.pricing;\n    const multimodal = config.vision ?? inferVisionCapability(config.name);\n\n    this.capabilities = {\n      structuredOutput: config.structuredOutput ?? true,\n      vision: multimodal,\n      // Every Gemini 2.5 model thinks; older families harmlessly ignore\n      // an empty thinking budget. Defaulting `true` lets the agent\n      // forward reasoning options; an explicit `false` opts a model out.\n      reasoning: config.reasoning ?? true,\n      // Gemini reports cache-read hits (`cachedContentTokenCount`) on\n      // every call via implicit caching, and accepts explicit context\n      // caching. Read-side accounting is always honored.\n      promptCaching: true,\n      // The multimodal Gemini families that accept images also accept\n      // audio and PDF/document parts. Mirror the vision inference unless\n      // explicitly overridden.\n      audio: config.audio ?? multimodal,\n      pdf: config.pdf ?? multimodal,\n    };\n  }\n\n  /**\n   * Single-shot completion. Sends the full message list to\n   * `generateContent`, waits for the terminal response, and reshapes\n   * it into a vendor-neutral `ModelResponse`. Per-call `options`\n   * override the instance defaults for this call only.\n   */\n  public async complete(messages: Message[], options?: ModelCallOptions): Promise<ModelResponse> {\n    this.logger.debug(LOG_MODULE, \"request\", \"Starting generateContent call\", {\n      model: this.name,\n      messageCount: messages.length,\n      streaming: false,\n      toolCount: options?.tools?.length ?? 0,\n    });\n\n    const { systemInstruction, contents } = toGoogleContents(messages);\n\n    let response: GenerateContentResponse;\n\n    try {\n      response = await this.ai.models.generateContent({\n        model: this.name,\n        contents,\n        config: this.buildConfig(systemInstruction, options),\n      });\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    const toolCalls = this.extractToolCalls(response);\n    const finishReason = toolCalls\n      ? \"tool_calls\"\n      : mapFinishReason(response.candidates?.[0]?.finishReason);\n    const usage = this.extractUsage(response);\n\n    this.logger.debug(LOG_MODULE, \"response\", \"generateContent call succeeded\", {\n      finishReason,\n      usage,\n    });\n\n    return {\n      content: response.text ?? \"\",\n      finishReason,\n      usage,\n      toolCalls,\n    };\n  }\n\n  /**\n   * Incremental streaming completion via `generateContentStream`.\n   * Yields neutral `ModelStreamChunk`s — `delta` for text, `tool-call`\n   * per function call (Gemini emits a fully-formed call, not partial\n   * JSON), and a terminal `done` with the final finish reason + usage.\n   */\n  public async *stream(\n    messages: Message[],\n    options?: ModelCallOptions,\n  ): AsyncIterable<ModelStreamChunk> {\n    this.logger.debug(LOG_MODULE, \"request\", \"Starting generateContentStream call\", {\n      model: this.name,\n      messageCount: messages.length,\n      streaming: true,\n      toolCount: options?.tools?.length ?? 0,\n    });\n\n    const { systemInstruction, contents } = toGoogleContents(messages);\n\n    let iterable: AsyncGenerator<GenerateContentResponse>;\n\n    try {\n      iterable = await this.ai.models.generateContentStream({\n        model: this.name,\n        contents,\n        config: this.buildConfig(systemInstruction, options),\n      });\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    let rawFinishReason: string | undefined;\n    let sawToolCall = false;\n    const usage: Usage = { input: 0, output: 0, total: 0 };\n\n    try {\n      for await (const chunk of iterable) {\n        const text = chunk.text;\n\n        if (text) {\n          yield { type: \"delta\", content: text };\n        }\n\n        for (const part of chunk.candidates?.[0]?.content?.parts ?? []) {\n          const toolCall = this.partToToolCall(part);\n\n          if (!toolCall) {\n            continue;\n          }\n\n          sawToolCall = true;\n\n          yield {\n            type: \"tool-call\",\n            id: toolCall.id,\n            name: toolCall.name,\n            input: toolCall.input,\n            ...(toolCall.providerMetadata\n              ? { providerMetadata: toolCall.providerMetadata }\n              : {}),\n          };\n        }\n\n        const candidateFinish = chunk.candidates?.[0]?.finishReason;\n\n        if (candidateFinish) {\n          rawFinishReason = candidateFinish;\n        }\n\n        if (chunk.usageMetadata) {\n          applyGoogleUsage(usage, chunk.usageMetadata);\n        }\n      }\n    } catch (thrown) {\n      throw this.logAndWrap(thrown);\n    }\n\n    const finishReason = sawToolCall ? \"tool_calls\" : mapFinishReason(rawFinishReason);\n\n    this.logger.debug(LOG_MODULE, \"response\", \"generateContentStream call succeeded\", {\n      finishReason,\n      usage,\n    });\n\n    yield { type: \"done\", finishReason, usage };\n  }\n\n  /**\n   * Assemble the `GenerateContentConfig` shared by `complete()` and\n   * `stream()`: inference params, hoisted system instruction,\n   * cancellation signal, and conditional tools + native structured\n   * output.\n   */\n  private buildConfig(\n    systemInstruction: string | undefined,\n    options: ModelCallOptions | undefined,\n  ): GenerateContentConfig {\n    const temperature = options?.temperature ?? this.config.temperature;\n    const maxOutputTokens = options?.maxTokens ?? this.config.maxTokens;\n\n    return {\n      ...(systemInstruction ? { systemInstruction } : {}),\n      ...(temperature !== undefined ? { temperature } : {}),\n      ...(maxOutputTokens !== undefined ? { maxOutputTokens } : {}),\n      ...(options?.signal ? { abortSignal: options.signal } : {}),\n      ...this.buildTools(options?.tools),\n      ...this.buildStructuredOutput(options?.responseSchema),\n      ...this.buildThinking(options?.reasoning),\n    };\n  }\n\n  /**\n   * Translate the neutral `reasoning` option into Gemini's\n   * `thinkingConfig`. `reasoning.maxTokens` maps directly to\n   * `thinkingBudget` (token cap on the thinking phase); when only\n   * `reasoning.effort` is given it is bucketed into a budget. Emitted\n   * only when the model is `reasoning`-capable — a `false` capability\n   * (config override) drops it so a non-thinking model never receives\n   * an unsupported `thinkingConfig`.\n   *\n   * Gemini's `thinkingBudget` semantics: `0` disables thinking, `-1`\n   * lets the model decide automatically. A positive value caps the\n   * thinking tokens. The neutral `effort: \"none\"` (\"run without\n   * reasoning\") maps to `thinkingBudget: 0`.\n   */\n  private buildThinking(\n    reasoning: ModelCallOptions[\"reasoning\"],\n  ): Pick<GenerateContentConfig, \"thinkingConfig\"> {\n    if (!reasoning || !this.capabilities.reasoning) {\n      return {};\n    }\n\n    // `effort: \"none\"` = explicit \"run without reasoning\". Gemini disables\n    // thinking with `thinkingBudget: 0` (its native off switch), so emit\n    // that rather than omitting the config — an omitted config lets a\n    // thinking model reason at its default budget.\n    if (reasoning.effort === \"none\") {\n      return { thinkingConfig: { thinkingBudget: 0 } };\n    }\n\n    const thinkingBudget =\n      reasoning.maxTokens ?? (reasoning.effort ? EFFORT_THINKING_BUDGET[reasoning.effort] : undefined);\n\n    if (thinkingBudget === undefined) {\n      return {};\n    }\n\n    return { thinkingConfig: { thinkingBudget } };\n  }\n\n  /**\n   * Spread-friendly tools fragment. Empty object when no tools were\n   * supplied so the caller can unconditionally spread it.\n   */\n  private buildTools(tools: ModelCallOptions[\"tools\"]): Pick<GenerateContentConfig, \"tools\"> {\n    const mapped = toGoogleTools(tools);\n\n    return mapped ? { tools: mapped } : {};\n  }\n\n  /**\n   * Translate the neutral `responseSchema` into Gemini's native JSON\n   * structured output (`responseMimeType: \"application/json\"` +\n   * `responseJsonSchema`, which takes a raw JSON Schema directly).\n   * Emitted only when the model is `structuredOutput`-capable and the\n   * schema is an object root — otherwise the agent's soft prompt hint\n   * + client-side `validate()` carry shape.\n   */\n  private buildStructuredOutput(\n    responseSchema: Record<string, unknown> | undefined,\n  ): Pick<GenerateContentConfig, \"responseMimeType\" | \"responseJsonSchema\"> {\n    if (!responseSchema || !this.capabilities.structuredOutput) {\n      return {};\n    }\n\n    if (responseSchema.type !== \"object\" || typeof responseSchema.properties !== \"object\") {\n      return {};\n    }\n\n    return {\n      responseMimeType: \"application/json\",\n      responseJsonSchema: responseSchema,\n    };\n  }\n\n  /**\n   * Reshape Gemini's function-call content parts into the neutral\n   * `ModelToolCallRequest[]`. Returns `undefined` when the model\n   * requested no functions so callers can branch on presence.\n   *\n   * Reads `candidates[0].content.parts` directly rather than the\n   * `response.functionCalls` getter: the getter discards the\n   * part-level `thoughtSignature`, and Gemini \"thinking\" models 400\n   * the follow-up turn if that signature is not echoed back. See\n   * `partToToolCall`.\n   */\n  private extractToolCalls(\n    response: GenerateContentResponse,\n  ): ModelToolCallRequest[] | undefined {\n    const parts = response.candidates?.[0]?.content?.parts ?? [];\n    const toolCalls = parts\n      .map((part) => this.partToToolCall(part))\n      .filter((call): call is ModelToolCallRequest => call !== undefined);\n\n    return toolCalls.length > 0 ? toolCalls : undefined;\n  }\n\n  /**\n   * Map a single Gemini `Part` to a neutral `ModelToolCallRequest`,\n   * or `undefined` when the part is not a function call. The part's\n   * `thoughtSignature` (opaque, set by thinking models) is carried on\n   * `providerMetadata` so `toGoogleContents` can replay it on the\n   * assistant turn — Gemini rejects the next request without it.\n   */\n  private partToToolCall(part: Part): ModelToolCallRequest | undefined {\n    if (!part.functionCall) {\n      return undefined;\n    }\n\n    const call = part.functionCall;\n\n    return {\n      // The Gemini Developer API does not assign function-call ids\n      // (only Vertex parallel-calling does). Fall back to the function\n      // name so the neutral `toolCallId` is non-empty and the echoed\n      // `functionResponse.name` resolves — Gemini matches a result to\n      // its call by name. See decisions §49.\n      id: call.id ?? call.name ?? \"\",\n      name: call.name ?? \"\",\n      input: (call.args ?? {}) as Record<string, unknown>,\n      ...(part.thoughtSignature\n        ? { providerMetadata: { thoughtSignature: part.thoughtSignature } }\n        : {}),\n    };\n  }\n\n  /**\n   * Normalize Gemini's `usageMetadata` into the neutral `Usage` shape\n   * via the shared {@link applyGoogleUsage} mapper (the same one the\n   * streaming loop and the Gemini image model use). Absent usage\n   * collapses to zeros.\n   */\n  private extractUsage(response: GenerateContentResponse): Usage {\n    const usage: Usage = { input: 0, output: 0, total: 0 };\n\n    if (response.usageMetadata) {\n      applyGoogleUsage(usage, response.usageMetadata);\n    }\n\n    return usage;\n  }\n\n  /**\n   * Wrap a thrown provider error into the typed `AIError` hierarchy\n   * and emit the standard error log line before it propagates.\n   */\n  private logAndWrap(thrown: unknown) {\n    const wrapped = wrapGoogleError(thrown);\n\n    this.logger.error(LOG_MODULE, \"error\", wrapped.message, {\n      code: wrapped.code,\n      context: wrapped.context,\n    });\n\n    return wrapped;\n  }\n}\n"],"mappings":";;;;;;;;;;AA6BA,MAAM,aAAa;;;;;;;;AASnB,MAAM,yBAA2E;CAC/E,KAAK;CACL,QAAQ;CACR,MAAM;AACR;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAiCA,IAAa,cAAb,MAAkD;CAUhD,AAAO,YAAY,IAAiB,QAA2B,WAAmB,UAAU;gBAF1D;EAGhC,KAAK,KAAK;EACV,KAAK,SAAS;EACd,KAAK,OAAO,OAAO;EACnB,KAAK,WAAW;EAChB,KAAK,UAAU,OAAO;EACtB,MAAM,aAAa,OAAO,UAAU,sBAAsB,OAAO,IAAI;EAErE,KAAK,eAAe;GAClB,kBAAkB,OAAO,oBAAoB;GAC7C,QAAQ;GAIR,WAAW,OAAO,aAAa;GAI/B,eAAe;GAIf,OAAO,OAAO,SAAS;GACvB,KAAK,OAAO,OAAO;EACrB;CACF;;;;;;;CAQA,MAAa,SAAS,UAAqB,SAAoD;EAC7F,KAAK,OAAO,MAAM,YAAY,WAAW,iCAAiC;GACxE,OAAO,KAAK;GACZ,cAAc,SAAS;GACvB,WAAW;GACX,WAAW,SAAS,OAAO,UAAU;EACvC,CAAC;EAED,MAAM,EAAE,mBAAmB,aAAa,iBAAiB,QAAQ;EAEjE,IAAI;EAEJ,IAAI;GACF,WAAW,MAAM,KAAK,GAAG,OAAO,gBAAgB;IAC9C,OAAO,KAAK;IACZ;IACA,QAAQ,KAAK,YAAY,mBAAmB,OAAO;GACrD,CAAC;EACH,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,MAAM,YAAY,KAAK,iBAAiB,QAAQ;EAChD,MAAM,eAAe,YACjB,eACA,gBAAgB,SAAS,aAAa,EAAE,EAAE,YAAY;EAC1D,MAAM,QAAQ,KAAK,aAAa,QAAQ;EAExC,KAAK,OAAO,MAAM,YAAY,YAAY,kCAAkC;GAC1E;GACA;EACF,CAAC;EAED,OAAO;GACL,SAAS,SAAS,QAAQ;GAC1B;GACA;GACA;EACF;CACF;;;;;;;CAQA,OAAc,OACZ,UACA,SACiC;EACjC,KAAK,OAAO,MAAM,YAAY,WAAW,uCAAuC;GAC9E,OAAO,KAAK;GACZ,cAAc,SAAS;GACvB,WAAW;GACX,WAAW,SAAS,OAAO,UAAU;EACvC,CAAC;EAED,MAAM,EAAE,mBAAmB,aAAa,iBAAiB,QAAQ;EAEjE,IAAI;EAEJ,IAAI;GACF,WAAW,MAAM,KAAK,GAAG,OAAO,sBAAsB;IACpD,OAAO,KAAK;IACZ;IACA,QAAQ,KAAK,YAAY,mBAAmB,OAAO;GACrD,CAAC;EACH,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,IAAI;EACJ,IAAI,cAAc;EAClB,MAAM,QAAe;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EAErD,IAAI;GACF,WAAW,MAAM,SAAS,UAAU;IAClC,MAAM,OAAO,MAAM;IAEnB,IAAI,MACF,MAAM;KAAE,MAAM;KAAS,SAAS;IAAK;IAGvC,KAAK,MAAM,QAAQ,MAAM,aAAa,EAAE,EAAE,SAAS,SAAS,CAAC,GAAG;KAC9D,MAAM,WAAW,KAAK,eAAe,IAAI;KAEzC,IAAI,CAAC,UACH;KAGF,cAAc;KAEd,MAAM;MACJ,MAAM;MACN,IAAI,SAAS;MACb,MAAM,SAAS;MACf,OAAO,SAAS;MAChB,GAAI,SAAS,mBACT,EAAE,kBAAkB,SAAS,iBAAiB,IAC9C,CAAC;KACP;IACF;IAEA,MAAM,kBAAkB,MAAM,aAAa,EAAE,EAAE;IAE/C,IAAI,iBACF,kBAAkB;IAGpB,IAAI,MAAM,eACR,iBAAiB,OAAO,MAAM,aAAa;GAE/C;EACF,SAAS,QAAQ;GACf,MAAM,KAAK,WAAW,MAAM;EAC9B;EAEA,MAAM,eAAe,cAAc,eAAe,gBAAgB,eAAe;EAEjF,KAAK,OAAO,MAAM,YAAY,YAAY,wCAAwC;GAChF;GACA;EACF,CAAC;EAED,MAAM;GAAE,MAAM;GAAQ;GAAc;EAAM;CAC5C;;;;;;;CAQA,AAAQ,YACN,mBACA,SACuB;EACvB,MAAM,cAAc,SAAS,eAAe,KAAK,OAAO;EACxD,MAAM,kBAAkB,SAAS,aAAa,KAAK,OAAO;EAE1D,OAAO;GACL,GAAI,oBAAoB,EAAE,kBAAkB,IAAI,CAAC;GACjD,GAAI,gBAAgB,SAAY,EAAE,YAAY,IAAI,CAAC;GACnD,GAAI,oBAAoB,SAAY,EAAE,gBAAgB,IAAI,CAAC;GAC3D,GAAI,SAAS,SAAS,EAAE,aAAa,QAAQ,OAAO,IAAI,CAAC;GACzD,GAAG,KAAK,WAAW,SAAS,KAAK;GACjC,GAAG,KAAK,sBAAsB,SAAS,cAAc;GACrD,GAAG,KAAK,cAAc,SAAS,SAAS;EAC1C;CACF;;;;;;;;;;;;;;;CAgBA,AAAQ,cACN,WAC+C;EAC/C,IAAI,CAAC,aAAa,CAAC,KAAK,aAAa,WACnC,OAAO,CAAC;EAOV,IAAI,UAAU,WAAW,QACvB,OAAO,EAAE,gBAAgB,EAAE,gBAAgB,EAAE,EAAE;EAGjD,MAAM,iBACJ,UAAU,cAAc,UAAU,SAAS,uBAAuB,UAAU,UAAU;EAExF,IAAI,mBAAmB,QACrB,OAAO,CAAC;EAGV,OAAO,EAAE,gBAAgB,EAAE,eAAe,EAAE;CAC9C;;;;;CAMA,AAAQ,WAAW,OAAwE;EACzF,MAAM,SAAS,cAAc,KAAK;EAElC,OAAO,SAAS,EAAE,OAAO,OAAO,IAAI,CAAC;CACvC;;;;;;;;;CAUA,AAAQ,sBACN,gBACwE;EACxE,IAAI,CAAC,kBAAkB,CAAC,KAAK,aAAa,kBACxC,OAAO,CAAC;EAGV,IAAI,eAAe,SAAS,YAAY,OAAO,eAAe,eAAe,UAC3E,OAAO,CAAC;EAGV,OAAO;GACL,kBAAkB;GAClB,oBAAoB;EACtB;CACF;;;;;;;;;;;;CAaA,AAAQ,iBACN,UACoC;EAEpC,MAAM,aADQ,SAAS,aAAa,EAAE,EAAE,SAAS,SAAS,CAAC,EACpC,CACpB,KAAK,SAAS,KAAK,eAAe,IAAI,CAAC,CAAC,CACxC,QAAQ,SAAuC,SAAS,MAAS;EAEpE,OAAO,UAAU,SAAS,IAAI,YAAY;CAC5C;;;;;;;;CASA,AAAQ,eAAe,MAA8C;EACnE,IAAI,CAAC,KAAK,cACR;EAGF,MAAM,OAAO,KAAK;EAElB,OAAO;GAML,IAAI,KAAK,MAAM,KAAK,QAAQ;GAC5B,MAAM,KAAK,QAAQ;GACnB,OAAQ,KAAK,QAAQ,CAAC;GACtB,GAAI,KAAK,mBACL,EAAE,kBAAkB,EAAE,kBAAkB,KAAK,iBAAiB,EAAE,IAChE,CAAC;EACP;CACF;;;;;;;CAQA,AAAQ,aAAa,UAA0C;EAC7D,MAAM,QAAe;GAAE,OAAO;GAAG,QAAQ;GAAG,OAAO;EAAE;EAErD,IAAI,SAAS,eACX,iBAAiB,OAAO,SAAS,aAAa;EAGhD,OAAO;CACT;;;;;CAMA,AAAQ,WAAW,QAAiB;EAClC,MAAM,UAAU,gBAAgB,MAAM;EAEtC,KAAK,OAAO,MAAM,YAAY,SAAS,QAAQ,SAAS;GACtD,MAAM,QAAQ;GACd,SAAS,QAAQ;EACnB,CAAC;EAED,OAAO;CACT;AACF"}

package/esm/sdk.d.mts

@@ -57,19 +57,27 @@ declare class GoogleSDK implements SDKAdapterContract {
    */
   embedder(config: GoogleEmbedderConfig): EmbedderContract;
   /**
-   * Build a `GoogleImageModel` (Imagen) bound to this SDK's client for
-   * use with `ai.image({ model, prompt })`. `config.name` is passed
-   * through to `ai.models.generateImages` as given — no id is rejected
-   * locally, so an unsupported model fails at Google, not here.
+   * Build an image model bound to this SDK's client for use with
+   * `ai.image({ model, prompt })`. `config.name` decides the transport
+   * (see {@link usesGeminiImageTransport}) — a `gemini-` id gets the
+   * `generateContent` implementation, everything else the Imagen
+   * `generateImages` one. No id is rejected locally either way, so an
+   * unsupported model fails at Google, not here.
+   *
+   * The two differ in what usage they can report, which is what the
+   * caller must price for: the Imagen path always returns a zero token
+   * `Usage` (Imagen reports none — price with `{ perImage }`), while the
+   * Gemini path passes through whatever `usageMetadata` Google attaches
+   * (price with `{ input, output }` when tokens come back).
    *
    * Pricing resolution mirrors `model()`: per-model `config.pricing`
    * wins, otherwise the SDK-level registry entry keyed by `config.name`,
-   * otherwise `undefined`. Imagen is per-image-metered, so the registry
-   * entry typically carries `{ perImage }`.
+   * otherwise `undefined`.
    *
    * @example
-   * const model = google.image({ name: "imagen-4.0-generate-001" });
-   * const { data } = await ai.image({ model, prompt: "a watercolor lighthouse" });
+   * const imagen = google.image({ name: "imagen-4.0-generate-001" });
+   * const gemini = google.image({ name: "gemini-3.1-flash-lite-image" });
+   * const { data } = await ai.image({ model: gemini, prompt: "a red bicycle" });
    */
   image(config: GoogleImageConfig): ImageModelContract;
 }

package/esm/sdk.d.mts.map

@@ -1 +1 @@
-{"version":3,"file":"sdk.d.mts","names":[],"sources":["../../../../../../ai-google/src/sdk.ts"],"mappings":";;;;;;AA2CA;;;;;;;;;;;;;;;;;;;;;;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;EA4CkB;AAAA;;;;;EA9B9C,KAAA,CAAM,IAAA,UAAc,MAAA,YAAkB,OAAA;;;;;;;;EAW5C,QAAA,CAAS,MAAA,EAAQ,oBAAA,GAAuB,gBAAA;;;;;;;;;;;;;;;;EAmBxC,KAAA,CAAM,MAAA,EAAQ,iBAAA,GAAoB,kBAAA;AAAA"}
+{"version":3,"file":"sdk.d.mts","names":[],"sources":["../../../../../../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"}

package/esm/sdk.mjs

@@ -1,4 +1,5 @@
 import { GoogleEmbedder } from "./embedder.mjs";
+import { GeminiImageModel } from "./gemini-image.mjs";
 import { GoogleImageModel } from "./image.mjs";
 import { GoogleModel } from "./model.mjs";
 import { GoogleGenAI } from "@google/genai";
@@ -6,6 +7,36 @@ import { approximateTokenCount } from "@warlock.js/ai";
 
 //#region ../ai-google/src/sdk.ts
 /**
+* Pick the transport for an image model id.
+*
+* `ai.models.generateImages` calls `{model}:predict`, and a `gemini-`
+* id sent there comes back `404 … is not supported for predict`
+* (observed verbatim from Google). `generateContent` is what the SDK
+* itself points `generateImages` users at — its deprecation notice
+* reads "Please use the generateContent method with image models
+* instead" — so the id has to choose the transport.
+*
+* Runs in this package establish where a `gemini-` id is ACCEPTED, not
+* what it returns: on this transport such an id got as far as a quota
+* error (HTTP 429) instead of the 404. That an image
+* comes back end-to-end once billing is enabled is reported by the
+* maintainer from a locally linked build, not measured here. Whether
+* these models report token usage is still unknown.
+*
+* This is ROUTING, not validation — no id is refused here. An id this
+* function does not recognize takes the `generateImages` route, the
+* only route that existed before Gemini image support landed, so every
+* id that reached Google before still reaches Google the same way and
+* still fails (or succeeds) at the provider.
+*
+* A leading `models/` resource prefix is tolerated, matching the id
+* shapes `inferVisionCapability` already accepts
+* (`models/gemini-1.5-flash-001`).
+*/
+function usesGeminiImageTransport(name) {
+	return name.toLowerCase().replace(/^models\//, "").startsWith("gemini-");
+}
+/**
 * Google Gemini-backed implementation of `SDKAdapterContract`.
 *
 * **Role.** The package entry point for Gemini models via the
@@ -73,19 +104,27 @@ var GoogleSDK = class {
 		return new GoogleEmbedder(this.ai, config, this.provider);
 	}
 	/**
-	* Build a `GoogleImageModel` (Imagen) bound to this SDK's client for
-	* use with `ai.image({ model, prompt })`. `config.name` is passed
-	* through to `ai.models.generateImages` as given — no id is rejected
-	* locally, so an unsupported model fails at Google, not here.
+	* Build an image model bound to this SDK's client for use with
+	* `ai.image({ model, prompt })`. `config.name` decides the transport
+	* (see {@link usesGeminiImageTransport}) — a `gemini-` id gets the
+	* `generateContent` implementation, everything else the Imagen
+	* `generateImages` one. No id is rejected locally either way, so an
+	* unsupported model fails at Google, not here.
+	*
+	* The two differ in what usage they can report, which is what the
+	* caller must price for: the Imagen path always returns a zero token
+	* `Usage` (Imagen reports none — price with `{ perImage }`), while the
+	* Gemini path passes through whatever `usageMetadata` Google attaches
+	* (price with `{ input, output }` when tokens come back).
 	*
 	* Pricing resolution mirrors `model()`: per-model `config.pricing`
 	* wins, otherwise the SDK-level registry entry keyed by `config.name`,
-	* otherwise `undefined`. Imagen is per-image-metered, so the registry
-	* entry typically carries `{ perImage }`.
+	* otherwise `undefined`.
 	*
 	* @example
-	* const model = google.image({ name: "imagen-4.0-generate-001" });
-	* const { data } = await ai.image({ model, prompt: "a watercolor lighthouse" });
+	* const imagen = google.image({ name: "imagen-4.0-generate-001" });
+	* const gemini = google.image({ name: "gemini-3.1-flash-lite-image" });
+	* const { data } = await ai.image({ model: gemini, prompt: "a red bicycle" });
 	*/
 	image(config) {
 		const resolvedPricing = config.pricing ?? this.pricing?.[config.name];
@@ -93,6 +132,7 @@ var GoogleSDK = class {
 			...config,
 			pricing: resolvedPricing
 		};
+		if (usesGeminiImageTransport(config.name)) return new GeminiImageModel(this.ai, resolvedConfig, this.provider);
 		return new GoogleImageModel(this.ai, resolvedConfig, this.provider);
 	}
 };

package/esm/sdk.mjs.map

@@ -1 +1 @@
-{"version":3,"file":"sdk.mjs","names":[],"sources":["../../../../../../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 { GoogleImageModel } from \"./image\";\nimport { GoogleModel } from \"./model\";\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 a `GoogleImageModel` (Imagen) bound to this SDK's client for\n   * use with `ai.image({ model, prompt })`. `config.name` is passed\n   * through to `ai.models.generateImages` as given — no id is rejected\n   * locally, so an unsupported model fails at Google, not here.\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`. Imagen is per-image-metered, so the registry\n   * entry typically carries `{ perImage }`.\n   *\n   * @example\n   * const model = google.image({ name: \"imagen-4.0-generate-001\" });\n   * const { data } = await ai.image({ model, prompt: \"a watercolor lighthouse\" });\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    return new GoogleImageModel(this.ai, resolvedConfig, this.provider);\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AA2CA,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;;;;;;;;;;;;;;;;CAiBA,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,OAAO,IAAI,iBAAiB,KAAK,IAAI,gBAAgB,KAAK,QAAQ;CACpE;AACF"}
+{"version":3,"file":"sdk.mjs","names":[],"sources":["../../../../../../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"}

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

@@ -0,0 +1,30 @@
+//#region ../ai-google/src/utils/apply-google-usage.ts
+/**
+* Fold a Gemini `usageMetadata` block into a running neutral `Usage`
+* accumulator. Shared by every `generateContent`-backed surface — the
+* chat model's `complete()`, its streaming loop (where the final chunk
+* carries cumulative totals), and the Gemini image model — so one
+* mapping decides what a Gemini token report means package-wide.
+*
+* Cache-read hits (`cachedContentTokenCount`, implicit or explicit
+* context caching) surface as `cachedTokens`; the thinking-phase tokens
+* of a reasoning model (`thoughtsTokenCount`) surface as
+* `reasoningTokens`. Both are emitted only when reported `> 0` so an
+* absent channel leaves the field undefined rather than a false zero.
+*
+* `total` falls back to `input + output` when Google omits
+* `totalTokenCount`.
+*/
+function applyGoogleUsage(usage, raw) {
+	usage.input = raw.promptTokenCount ?? usage.input;
+	usage.output = raw.candidatesTokenCount ?? usage.output;
+	usage.total = raw.totalTokenCount ?? usage.input + usage.output;
+	const cached = raw.cachedContentTokenCount;
+	if (cached && cached > 0) usage.cachedTokens = cached;
+	const reasoning = raw.thoughtsTokenCount;
+	if (reasoning && reasoning > 0) usage.reasoningTokens = reasoning;
+}
+
+//#endregion
+export { applyGoogleUsage };
+//# sourceMappingURL=apply-google-usage.mjs.map