@tanstack/ai-gemini 0.16.0 → 0.17.0

package/dist/esm/adapters/image.d.ts

@@ -1,6 +1,6 @@
 import { BaseImageAdapter } from '@tanstack/ai/adapters';
 import { GEMINI_IMAGE_MODELS } from '../model-meta.js';
-import { GeminiImageModelProviderOptionsByName, GeminiImageModelSizeByName, GeminiImageProviderOptions } from '../image/image-provider-options.js';
+import { GeminiImageModelInputModalitiesByName, GeminiImageModelProviderOptionsByName, GeminiImageModelSizeByName, GeminiImageProviderOptions } from '../image/image-provider-options.js';
 import { ImageGenerationOptions, ImageGenerationResult } from '@tanstack/ai';
 import { GeminiClientConfig } from '../utils.js';
 /**
@@ -24,19 +24,32 @@ export type GeminiImageModel = (typeof GEMINI_IMAGE_MODELS)[number];
  * - Watermark options
  * - Extended resolution tiers (Nano Banana 2)
  */
-export declare class GeminiImageAdapter<TModel extends GeminiImageModel> extends BaseImageAdapter<TModel, GeminiImageProviderOptions, GeminiImageModelProviderOptionsByName, GeminiImageModelSizeByName> {
+export declare class GeminiImageAdapter<TModel extends GeminiImageModel> extends BaseImageAdapter<TModel, GeminiImageProviderOptions, GeminiImageModelProviderOptionsByName, GeminiImageModelSizeByName, GeminiImageModelInputModalitiesByName> {
     readonly kind: "image";
     readonly name: "gemini";
     '~types': {
         providerOptions: GeminiImageProviderOptions;
         modelProviderOptionsByName: GeminiImageModelProviderOptionsByName;
         modelSizeByName: GeminiImageModelSizeByName;
+        modelInputModalitiesByName: GeminiImageModelInputModalitiesByName;
     };
     private readonly client;
     constructor(config: GeminiImageConfig, model: TModel);
     generateImages(options: ImageGenerationOptions<GeminiImageProviderOptions>): Promise<ImageGenerationResult>;
     private isGeminiImageModel;
     private generateWithGeminiApi;
+    /**
+     * Build the multimodal `contents` payload. Text-only prompts pass through
+     * as a plain string (the SDK accepts it directly); prompts with image
+     * parts become a single user `Content` whose `parts` mirror the prompt's
+     * interleaved order — position is meaningful to Gemini ("not like this
+     * *(image)*, more like this *(image)*").
+     *
+     * The generateContent API has no numberOfImages parameter, so when more
+     * than one image is requested a trailing instruction is appended.
+     */
+    private buildContents;
+    private imagePartToGeminiPart;
     private transformGeminiResponse;
     private buildImagenConfig;
     private transformImagenResponse;

package/dist/esm/adapters/image.js

@@ -1,4 +1,6 @@
+import { resolveMediaPrompt } from "@tanstack/ai";
 import { BaseImageAdapter } from "@tanstack/ai/adapters";
+import { arrayBufferToBase64 } from "@tanstack/ai-utils";
 import { createGeminiClient, generateId, getGeminiApiKeyFromEnv } from "../utils/client.js";
 import { buildGeminiUsage } from "../usage.js";
 import { validatePrompt, validateImageSize, validateNumberOfImages, parseNativeImageSize, sizeToAspectRatio } from "../image/image-provider-options.js";
@@ -11,7 +13,7 @@ class GeminiImageAdapter extends BaseImageAdapter {
     this.client = createGeminiClient(config);
   }
   async generateImages(options) {
-    const { model, prompt, logger } = options;
+    const { model, logger } = options;
     logger.request(
       `activity=generateImage provider=gemini model=${this.model}`,
       {
@@ -20,16 +22,34 @@ class GeminiImageAdapter extends BaseImageAdapter {
       }
     );
     try {
-      validatePrompt({ prompt, model });
+      const resolved = resolveMediaPrompt(options.prompt);
+      if (resolved.images.length === 0) {
+        validatePrompt({ prompt: resolved.text, model });
+      }
+      if (resolved.videos.length > 0) {
+        throw new Error(
+          `${this.name}.generateImages does not support video prompt parts (model: ${model}).`
+        );
+      }
+      if (resolved.audios.length > 0) {
+        throw new Error(
+          `${this.name}.generateImages does not support audio prompt parts (model: ${model}).`
+        );
+      }
       if (this.isGeminiImageModel(model)) {
-        return await this.generateWithGeminiApi(options);
+        return await this.generateWithGeminiApi(options, resolved);
+      }
+      if (resolved.images.length > 0) {
+        throw new Error(
+          `${this.name}: model "${model}" (Imagen) does not support image prompt parts. Use a Gemini-native image model (e.g. gemini-2.5-flash-image, "nano-banana") for image-conditioned generation.`
+        );
       }
       validateImageSize(model, options.size);
       validateNumberOfImages(model, options.numberOfImages);
       const config = this.buildImagenConfig(options);
       const response = await this.client.models.generateImages({
         model,
-        prompt,
+        prompt: resolved.text,
         config
       });
       return this.transformImagenResponse(model, response);
@@ -44,10 +64,9 @@ class GeminiImageAdapter extends BaseImageAdapter {
   isGeminiImageModel(model) {
     return model.startsWith("gemini-");
   }
-  async generateWithGeminiApi(options) {
-    const { model, prompt, size, numberOfImages, modelOptions } = options;
+  async generateWithGeminiApi(options, resolved) {
+    const { model, size, numberOfImages, modelOptions } = options;
     const parsedSize = size ? parseNativeImageSize(size) : void 0;
-    const augmentedPrompt = numberOfImages && numberOfImages > 1 ? `${prompt} Generate ${numberOfImages} distinct images.` : prompt;
     const nativeConfig = {};
     if (modelOptions?.seed !== void 0) {
       nativeConfig.seed = modelOptions.seed;
@@ -69,13 +88,82 @@ class GeminiImageAdapter extends BaseImageAdapter {
         }
       }
     };
+    const contents = await this.buildContents(resolved, numberOfImages);
     const response = await this.client.models.generateContent({
       model,
-      contents: augmentedPrompt,
+      contents,
       config
     });
     return this.transformGeminiResponse(model, response);
   }
+  /**
+   * Build the multimodal `contents` payload. Text-only prompts pass through
+   * as a plain string (the SDK accepts it directly); prompts with image
+   * parts become a single user `Content` whose `parts` mirror the prompt's
+   * interleaved order — position is meaningful to Gemini ("not like this
+   * *(image)*, more like this *(image)*").
+   *
+   * The generateContent API has no numberOfImages parameter, so when more
+   * than one image is requested a trailing instruction is appended.
+   */
+  async buildContents(resolved, numberOfImages) {
+    const countInstruction = numberOfImages && numberOfImages > 1 ? `Generate ${numberOfImages} distinct images.` : void 0;
+    if (resolved.images.length === 0) {
+      return countInstruction ? `${resolved.text} ${countInstruction}` : resolved.text;
+    }
+    const parts = await Promise.all(
+      resolved.parts.map((part) => {
+        if (part.type === "text") {
+          return Promise.resolve({ text: part.content });
+        }
+        if (part.type === "image") {
+          return this.imagePartToGeminiPart(part);
+        }
+        throw new Error(
+          `gemini: unsupported prompt part type "${part.type}" in image generation.`
+        );
+      })
+    );
+    if (countInstruction) {
+      parts.push({ text: countInstruction });
+    }
+    return [{ role: "user", parts }];
+  }
+  async imagePartToGeminiPart(part) {
+    if (part.source.type === "data") {
+      return {
+        inlineData: {
+          mimeType: part.source.mimeType || "image/png",
+          data: part.source.value
+        }
+      };
+    }
+    if (part.source.value.startsWith("gs://") || /^https?:\/\/generativelanguage\.googleapis\.com\//.test(
+      part.source.value
+    )) {
+      return {
+        fileData: {
+          fileUri: part.source.value,
+          ...part.source.mimeType && { mimeType: part.source.mimeType }
+        }
+      };
+    }
+    const response = await fetch(part.source.value);
+    if (!response.ok) {
+      throw new Error(
+        `Failed to fetch image input (${response.status} ${response.statusText}): ${part.source.value}`
+      );
+    }
+    const blob = await response.blob();
+    const buffer = await blob.arrayBuffer();
+    const base64 = arrayBufferToBase64(buffer);
+    return {
+      inlineData: {
+        mimeType: part.source.mimeType || blob.type || "image/png",
+        data: base64
+      }
+    };
+  }
   transformGeminiResponse(model, response) {
     const images = [];
     const textParts = [];

package/dist/esm/adapters/image.js.map

@@ -1 +1 @@
-{"version":3,"file":"image.js","sources":["../../../src/adapters/image.ts"],"sourcesContent":["import { BaseImageAdapter } from '@tanstack/ai/adapters'\nimport {\n  createGeminiClient,\n  generateId,\n  getGeminiApiKeyFromEnv,\n} from '../utils'\nimport { buildGeminiUsage } from '../usage'\nimport {\n  parseNativeImageSize,\n  sizeToAspectRatio,\n  validateImageSize,\n  validateNumberOfImages,\n  validatePrompt,\n} from '../image/image-provider-options'\nimport type { GEMINI_IMAGE_MODELS } from '../model-meta'\nimport type {\n  GeminiImageModelProviderOptionsByName,\n  GeminiImageModelSizeByName,\n  GeminiImageProviderOptions,\n} from '../image/image-provider-options'\nimport type {\n  GeneratedImage,\n  ImageGenerationOptions,\n  ImageGenerationResult,\n} from '@tanstack/ai'\nimport type {\n  GenerateContentConfig,\n  GenerateContentResponse,\n  GenerateImagesConfig,\n  GenerateImagesResponse,\n  GoogleGenAI,\n} from '@google/genai'\nimport type { GeminiClientConfig } from '../utils'\n\n/**\n * Configuration for Gemini image adapter\n */\nexport interface GeminiImageConfig extends GeminiClientConfig {}\n\n/** Model type for Gemini Image */\nexport type GeminiImageModel = (typeof GEMINI_IMAGE_MODELS)[number]\n\n/**\n * Gemini Image Generation Adapter\n *\n * Tree-shakeable adapter for Gemini image generation functionality.\n * Supports Imagen 3/4 models (via generateImages API) and Gemini native\n * image models like Nano Banana 2 (via generateContent API).\n *\n * Features:\n * - Aspect ratio-based image sizing\n * - Person generation controls\n * - Safety filtering\n * - Watermark options\n * - Extended resolution tiers (Nano Banana 2)\n */\nexport class GeminiImageAdapter<\n  TModel extends GeminiImageModel,\n> extends BaseImageAdapter<\n  TModel,\n  GeminiImageProviderOptions,\n  GeminiImageModelProviderOptionsByName,\n  GeminiImageModelSizeByName\n> {\n  override readonly kind = 'image' as const\n  readonly name = 'gemini' as const\n\n  // Type-only property - never assigned at runtime\n  declare '~types': {\n    providerOptions: GeminiImageProviderOptions\n    modelProviderOptionsByName: GeminiImageModelProviderOptionsByName\n    modelSizeByName: GeminiImageModelSizeByName\n  }\n\n  private readonly client: GoogleGenAI\n\n  constructor(config: GeminiImageConfig, model: TModel) {\n    super(model, config)\n    this.client = createGeminiClient(config)\n  }\n\n  async generateImages(\n    options: ImageGenerationOptions<GeminiImageProviderOptions>,\n  ): Promise<ImageGenerationResult> {\n    const { model, prompt, logger } = options\n\n    logger.request(\n      `activity=generateImage provider=gemini model=${this.model}`,\n      {\n        provider: 'gemini',\n        model: this.model,\n      },\n    )\n\n    try {\n      validatePrompt({ prompt, model })\n\n      if (this.isGeminiImageModel(model)) {\n        return await this.generateWithGeminiApi(options)\n      }\n\n      // Imagen models path (generateImages API)\n      validateImageSize(model, options.size)\n      validateNumberOfImages(model, options.numberOfImages)\n\n      const config = this.buildImagenConfig(options)\n\n      const response = await this.client.models.generateImages({\n        model,\n        prompt,\n        config,\n      })\n\n      return this.transformImagenResponse(model, response)\n    } catch (error) {\n      logger.errors('gemini.generateImage fatal', {\n        error,\n        source: 'gemini.generateImage',\n      })\n      throw error\n    }\n  }\n\n  private isGeminiImageModel(model: string): boolean {\n    return model.startsWith('gemini-')\n  }\n\n  private async generateWithGeminiApi(\n    options: ImageGenerationOptions<GeminiImageProviderOptions>,\n  ): Promise<ImageGenerationResult> {\n    const { model, prompt, size, numberOfImages, modelOptions } = options\n\n    const parsedSize = size ? parseNativeImageSize(size) : undefined\n\n    // The generateContent API has no numberOfImages parameter.\n    // Instead, augment the prompt to request multiple images when needed.\n    const augmentedPrompt =\n      numberOfImages && numberOfImages > 1\n        ? `${prompt} Generate ${numberOfImages} distinct images.`\n        : prompt\n\n    // GeminiImageProviderOptions is Imagen-shaped — most fields\n    // (personGeneration, safetyFilterLevel, addWatermark, outputMimeType,\n    // outputCompressionQuality, guidanceScale, enhancePrompt,\n    // includeSafetyAttributes, includeRaiReason, outputGcsUri, labels,\n    // negativePrompt, language) are only valid on GenerateImagesConfig and\n    // would be rejected by the Gemini-native generateContent path. Pick only\n    // the fields that are valid on GenerateContentConfig instead of spreading\n    // the whole options object.\n    const nativeConfig: GenerateContentConfig = {}\n    if (modelOptions?.seed !== undefined) {\n      nativeConfig.seed = modelOptions.seed\n    }\n\n    const config: GenerateContentConfig = {\n      ...nativeConfig,\n      // Include TEXT so the model can interleave descriptions between images.\n      // IMPORTANT: responseModalities is a protected default — set it AFTER\n      // nativeConfig so nothing can silently disable image output.\n      responseModalities: ['TEXT', 'IMAGE'],\n      ...(parsedSize && {\n        imageConfig: {\n          ...(parsedSize.aspectRatio && {\n            aspectRatio: parsedSize.aspectRatio,\n          }),\n          ...(parsedSize.resolution && {\n            imageSize: parsedSize.resolution,\n          }),\n        },\n      }),\n    }\n\n    const response = await this.client.models.generateContent({\n      model,\n      contents: augmentedPrompt,\n      config,\n    })\n\n    return this.transformGeminiResponse(model, response)\n  }\n\n  private transformGeminiResponse(\n    model: string,\n    response: GenerateContentResponse,\n  ): ImageGenerationResult {\n    const images: Array<GeneratedImage> = []\n    const textParts: Array<string> = []\n    const parts = response.candidates?.[0]?.content?.parts ?? []\n\n    for (const part of parts) {\n      if (\n        part.inlineData?.data &&\n        typeof part.inlineData.data === 'string' &&\n        part.inlineData.data.length > 0\n      ) {\n        images.push({ b64Json: part.inlineData.data })\n      } else if (typeof part.text === 'string' && part.text.length > 0) {\n        textParts.push(part.text)\n      }\n    }\n\n    // If the model returned only text parts (for example a safety refusal\n    // or a \"can't do that\" message), surface the text instead of silently\n    // resolving to an empty images array — otherwise callers can't tell a\n    // generation failure apart from a genuine empty response.\n    if (images.length === 0) {\n      const reason =\n        textParts.length > 0\n          ? `: ${textParts.join(' ').trim()}`\n          : ' (no inline image or text parts were returned).'\n      throw new Error(`Gemini ${model} returned no images${reason}`)\n    }\n\n    return {\n      id: generateId(this.name),\n      model,\n      images,\n      // Surface token usage (with per-modality breakdown) when the model\n      // reports it (e.g. Nano Banana via generateContent). Conditionally spread\n      // to satisfy exactOptionalPropertyTypes — only include usage when\n      // present. See #330.\n      ...(response.usageMetadata\n        ? { usage: buildGeminiUsage(response.usageMetadata) }\n        : {}),\n    }\n  }\n\n  private buildImagenConfig(\n    options: ImageGenerationOptions<GeminiImageProviderOptions>,\n  ): GenerateImagesConfig {\n    const { size, numberOfImages, modelOptions } = options\n\n    // Build with conditional spreads — under exactOptionalPropertyTypes the\n    // vendor `GenerateImagesConfig` fields are `field?: T` (no `| undefined`),\n    // so we can only assign the property when we actually have a value.\n    const sizeAspectRatio = size ? sizeToAspectRatio(size) : undefined\n    return {\n      numberOfImages: numberOfImages ?? 1,\n      // Map size to aspect ratio if provided (modelOptions.aspectRatio will override)\n      ...(sizeAspectRatio !== undefined && { aspectRatio: sizeAspectRatio }),\n      ...modelOptions,\n    }\n  }\n\n  private transformImagenResponse(\n    model: string,\n    response: GenerateImagesResponse,\n  ): ImageGenerationResult {\n    const entries = response.generatedImages ?? []\n    const images: Array<GeneratedImage> = []\n    const filterReasons: Array<string> = []\n\n    for (const item of entries) {\n      const b64Json = item.image?.imageBytes\n      if (b64Json) {\n        images.push({\n          b64Json,\n          ...(item.enhancedPrompt !== undefined && {\n            revisedPrompt: item.enhancedPrompt,\n          }),\n        })\n        continue\n      }\n      // Imagen can drop individual entries with a raiFilteredReason when\n      // Responsible-AI filters fire. Preserve the reason so callers can\n      // surface it instead of silently getting back fewer images.\n      const reason = (item as { raiFilteredReason?: string }).raiFilteredReason\n      if (reason) {\n        filterReasons.push(reason)\n      }\n    }\n\n    // Every entry was filtered — no usable images to return. Throw rather\n    // than resolve to an empty array so the caller is forced to handle the\n    // failure mode explicitly.\n    if (entries.length > 0 && images.length === 0) {\n      const joined = filterReasons.length > 0 ? filterReasons.join('; ') : ''\n      throw new Error(\n        `Imagen ${model} returned no images: all ${entries.length} generated image(s) were filtered by Responsible-AI${joined ? ` (${joined})` : ''}.`,\n      )\n    }\n\n    // Partial filter: surface via console.warn since ImageGenerationResult\n    // has no warnings field. Callers that care can still inspect the count\n    // mismatch between requested and returned images.\n    if (filterReasons.length > 0 && typeof console !== 'undefined') {\n      console.warn(\n        `[gemini-image] ${filterReasons.length} of ${entries.length} images from ${model} were filtered by Responsible-AI: ${filterReasons.join('; ')}`,\n      )\n    }\n\n    return {\n      id: generateId(this.name),\n      model,\n      images,\n    }\n  }\n}\n\n/**\n * Creates a Gemini image adapter with explicit API key.\n * Type resolution happens here at the call site.\n *\n * @param model - The model name (e.g., 'imagen-3.0-generate-002')\n * @param apiKey - Your Google API key\n * @param config - Optional additional configuration\n * @returns Configured Gemini image adapter instance with resolved types\n *\n * @example\n * ```typescript\n * const adapter = createGeminiImage('imagen-3.0-generate-002', \"your-api-key\");\n *\n * const result = await generateImage({\n *   adapter,\n *   prompt: 'A cute baby sea otter'\n * });\n * ```\n */\nexport function createGeminiImage<TModel extends GeminiImageModel>(\n  model: TModel,\n  apiKey: string,\n  config?: Omit<GeminiImageConfig, 'apiKey'>,\n): GeminiImageAdapter<TModel> {\n  return new GeminiImageAdapter({ apiKey, ...config }, model)\n}\n\n/**\n * Creates a Gemini image adapter with automatic API key detection from environment variables.\n * Type resolution happens here at the call site.\n *\n * Looks for `GOOGLE_API_KEY` or `GEMINI_API_KEY` in:\n * - `process.env` (Node.js)\n * - `window.env` (Browser with injected env)\n *\n * @param model - The model name (e.g., 'imagen-4.0-generate-001')\n * @param config - Optional configuration (excluding apiKey which is auto-detected)\n * @returns Configured Gemini image adapter instance with resolved types\n * @throws Error if GOOGLE_API_KEY or GEMINI_API_KEY is not found in environment\n *\n * @example\n * ```typescript\n * // Automatically uses GOOGLE_API_KEY from environment\n * const adapter = geminiImage('imagen-4.0-generate-001');\n *\n * const result = await generateImage({\n *   adapter,\n *   prompt: 'A beautiful sunset over mountains'\n * });\n * ```\n */\nexport function geminiImage<TModel extends GeminiImageModel>(\n  model: TModel,\n  config?: Omit<GeminiImageConfig, 'apiKey'>,\n): GeminiImageAdapter<TModel> {\n  const apiKey = getGeminiApiKeyFromEnv()\n  return createGeminiImage(model, apiKey, config)\n}\n"],"names":[],"mappings":";;;;AAwDO,MAAM,2BAEH,iBAKR;AAAA,EACkB,OAAO;AAAA,EAChB,OAAO;AAAA,EASC;AAAA,EAEjB,YAAY,QAA2B,OAAe;AACpD,UAAM,OAAO,MAAM;AACnB,SAAK,SAAS,mBAAmB,MAAM;AAAA,EACzC;AAAA,EAEA,MAAM,eACJ,SACgC;AAChC,UAAM,EAAE,OAAO,QAAQ,OAAA,IAAW;AAElC,WAAO;AAAA,MACL,gDAAgD,KAAK,KAAK;AAAA,MAC1D;AAAA,QACE,UAAU;AAAA,QACV,OAAO,KAAK;AAAA,MAAA;AAAA,IACd;AAGF,QAAI;AACF,qBAAe,EAAE,QAAQ,OAAO;AAEhC,UAAI,KAAK,mBAAmB,KAAK,GAAG;AAClC,eAAO,MAAM,KAAK,sBAAsB,OAAO;AAAA,MACjD;AAGA,wBAAkB,OAAO,QAAQ,IAAI;AACrC,6BAAuB,OAAO,QAAQ,cAAc;AAEpD,YAAM,SAAS,KAAK,kBAAkB,OAAO;AAE7C,YAAM,WAAW,MAAM,KAAK,OAAO,OAAO,eAAe;AAAA,QACvD;AAAA,QACA;AAAA,QACA;AAAA,MAAA,CACD;AAED,aAAO,KAAK,wBAAwB,OAAO,QAAQ;AAAA,IACrD,SAAS,OAAO;AACd,aAAO,OAAO,8BAA8B;AAAA,QAC1C;AAAA,QACA,QAAQ;AAAA,MAAA,CACT;AACD,YAAM;AAAA,IACR;AAAA,EACF;AAAA,EAEQ,mBAAmB,OAAwB;AACjD,WAAO,MAAM,WAAW,SAAS;AAAA,EACnC;AAAA,EAEA,MAAc,sBACZ,SACgC;AAChC,UAAM,EAAE,OAAO,QAAQ,MAAM,gBAAgB,iBAAiB;AAE9D,UAAM,aAAa,OAAO,qBAAqB,IAAI,IAAI;AAIvD,UAAM,kBACJ,kBAAkB,iBAAiB,IAC/B,GAAG,MAAM,aAAa,cAAc,sBACpC;AAUN,UAAM,eAAsC,CAAA;AAC5C,QAAI,cAAc,SAAS,QAAW;AACpC,mBAAa,OAAO,aAAa;AAAA,IACnC;AAEA,UAAM,SAAgC;AAAA,MACpC,GAAG;AAAA;AAAA;AAAA;AAAA,MAIH,oBAAoB,CAAC,QAAQ,OAAO;AAAA,MACpC,GAAI,cAAc;AAAA,QAChB,aAAa;AAAA,UACX,GAAI,WAAW,eAAe;AAAA,YAC5B,aAAa,WAAW;AAAA,UAAA;AAAA,UAE1B,GAAI,WAAW,cAAc;AAAA,YAC3B,WAAW,WAAW;AAAA,UAAA;AAAA,QACxB;AAAA,MACF;AAAA,IACF;AAGF,UAAM,WAAW,MAAM,KAAK,OAAO,OAAO,gBAAgB;AAAA,MACxD;AAAA,MACA,UAAU;AAAA,MACV;AAAA,IAAA,CACD;AAED,WAAO,KAAK,wBAAwB,OAAO,QAAQ;AAAA,EACrD;AAAA,EAEQ,wBACN,OACA,UACuB;AACvB,UAAM,SAAgC,CAAA;AACtC,UAAM,YAA2B,CAAA;AACjC,UAAM,QAAQ,SAAS,aAAa,CAAC,GAAG,SAAS,SAAS,CAAA;AAE1D,eAAW,QAAQ,OAAO;AACxB,UACE,KAAK,YAAY,QACjB,OAAO,KAAK,WAAW,SAAS,YAChC,KAAK,WAAW,KAAK,SAAS,GAC9B;AACA,eAAO,KAAK,EAAE,SAAS,KAAK,WAAW,MAAM;AAAA,MAC/C,WAAW,OAAO,KAAK,SAAS,YAAY,KAAK,KAAK,SAAS,GAAG;AAChE,kBAAU,KAAK,KAAK,IAAI;AAAA,MAC1B;AAAA,IACF;AAMA,QAAI,OAAO,WAAW,GAAG;AACvB,YAAM,SACJ,UAAU,SAAS,IACf,KAAK,UAAU,KAAK,GAAG,EAAE,KAAA,CAAM,KAC/B;AACN,YAAM,IAAI,MAAM,UAAU,KAAK,sBAAsB,MAAM,EAAE;AAAA,IAC/D;AAEA,WAAO;AAAA,MACL,IAAI,WAAW,KAAK,IAAI;AAAA,MACxB;AAAA,MACA;AAAA;AAAA;AAAA;AAAA;AAAA,MAKA,GAAI,SAAS,gBACT,EAAE,OAAO,iBAAiB,SAAS,aAAa,MAChD,CAAA;AAAA,IAAC;AAAA,EAET;AAAA,EAEQ,kBACN,SACsB;AACtB,UAAM,EAAE,MAAM,gBAAgB,aAAA,IAAiB;AAK/C,UAAM,kBAAkB,OAAO,kBAAkB,IAAI,IAAI;AACzD,WAAO;AAAA,MACL,gBAAgB,kBAAkB;AAAA;AAAA,MAElC,GAAI,oBAAoB,UAAa,EAAE,aAAa,gBAAA;AAAA,MACpD,GAAG;AAAA,IAAA;AAAA,EAEP;AAAA,EAEQ,wBACN,OACA,UACuB;AACvB,UAAM,UAAU,SAAS,mBAAmB,CAAA;AAC5C,UAAM,SAAgC,CAAA;AACtC,UAAM,gBAA+B,CAAA;AAErC,eAAW,QAAQ,SAAS;AAC1B,YAAM,UAAU,KAAK,OAAO;AAC5B,UAAI,SAAS;AACX,eAAO,KAAK;AAAA,UACV;AAAA,UACA,GAAI,KAAK,mBAAmB,UAAa;AAAA,YACvC,eAAe,KAAK;AAAA,UAAA;AAAA,QACtB,CACD;AACD;AAAA,MACF;AAIA,YAAM,SAAU,KAAwC;AACxD,UAAI,QAAQ;AACV,sBAAc,KAAK,MAAM;AAAA,MAC3B;AAAA,IACF;AAKA,QAAI,QAAQ,SAAS,KAAK,OAAO,WAAW,GAAG;AAC7C,YAAM,SAAS,cAAc,SAAS,IAAI,cAAc,KAAK,IAAI,IAAI;AACrE,YAAM,IAAI;AAAA,QACR,UAAU,KAAK,4BAA4B,QAAQ,MAAM,sDAAsD,SAAS,KAAK,MAAM,MAAM,EAAE;AAAA,MAAA;AAAA,IAE/I;AAKA,QAAI,cAAc,SAAS,KAAK,OAAO,YAAY,aAAa;AAC9D,cAAQ;AAAA,QACN,kBAAkB,cAAc,MAAM,OAAO,QAAQ,MAAM,gBAAgB,KAAK,qCAAqC,cAAc,KAAK,IAAI,CAAC;AAAA,MAAA;AAAA,IAEjJ;AAEA,WAAO;AAAA,MACL,IAAI,WAAW,KAAK,IAAI;AAAA,MACxB;AAAA,MACA;AAAA,IAAA;AAAA,EAEJ;AACF;AAqBO,SAAS,kBACd,OACA,QACA,QAC4B;AAC5B,SAAO,IAAI,mBAAmB,EAAE,QAAQ,GAAG,OAAA,GAAU,KAAK;AAC5D;AA0BO,SAAS,YACd,OACA,QAC4B;AAC5B,QAAM,SAAS,uBAAA;AACf,SAAO,kBAAkB,OAAO,QAAQ,MAAM;AAChD;"}
+{"version":3,"file":"image.js","sources":["../../../src/adapters/image.ts"],"sourcesContent":["import { resolveMediaPrompt } from '@tanstack/ai'\nimport { BaseImageAdapter } from '@tanstack/ai/adapters'\nimport { arrayBufferToBase64 } from '@tanstack/ai-utils'\nimport {\n  createGeminiClient,\n  generateId,\n  getGeminiApiKeyFromEnv,\n} from '../utils'\nimport { buildGeminiUsage } from '../usage'\nimport {\n  parseNativeImageSize,\n  sizeToAspectRatio,\n  validateImageSize,\n  validateNumberOfImages,\n  validatePrompt,\n} from '../image/image-provider-options'\nimport type { GEMINI_IMAGE_MODELS } from '../model-meta'\nimport type {\n  GeminiImageModelInputModalitiesByName,\n  GeminiImageModelProviderOptionsByName,\n  GeminiImageModelSizeByName,\n  GeminiImageProviderOptions,\n} from '../image/image-provider-options'\nimport type {\n  GeneratedImage,\n  ImageGenerationOptions,\n  ImageGenerationResult,\n  ImagePart,\n  MediaInputMetadata,\n  ResolvedMediaPrompt,\n} from '@tanstack/ai'\nimport type {\n  Content,\n  GenerateContentConfig,\n  GenerateContentResponse,\n  GenerateImagesConfig,\n  GenerateImagesResponse,\n  GoogleGenAI,\n  Part,\n} from '@google/genai'\nimport type { GeminiClientConfig } from '../utils'\n\n/**\n * Configuration for Gemini image adapter\n */\nexport interface GeminiImageConfig extends GeminiClientConfig {}\n\n/** Model type for Gemini Image */\nexport type GeminiImageModel = (typeof GEMINI_IMAGE_MODELS)[number]\n\n/**\n * Gemini Image Generation Adapter\n *\n * Tree-shakeable adapter for Gemini image generation functionality.\n * Supports Imagen 3/4 models (via generateImages API) and Gemini native\n * image models like Nano Banana 2 (via generateContent API).\n *\n * Features:\n * - Aspect ratio-based image sizing\n * - Person generation controls\n * - Safety filtering\n * - Watermark options\n * - Extended resolution tiers (Nano Banana 2)\n */\nexport class GeminiImageAdapter<\n  TModel extends GeminiImageModel,\n> extends BaseImageAdapter<\n  TModel,\n  GeminiImageProviderOptions,\n  GeminiImageModelProviderOptionsByName,\n  GeminiImageModelSizeByName,\n  GeminiImageModelInputModalitiesByName\n> {\n  override readonly kind = 'image' as const\n  readonly name = 'gemini' as const\n\n  // Type-only property - never assigned at runtime\n  declare '~types': {\n    providerOptions: GeminiImageProviderOptions\n    modelProviderOptionsByName: GeminiImageModelProviderOptionsByName\n    modelSizeByName: GeminiImageModelSizeByName\n    modelInputModalitiesByName: GeminiImageModelInputModalitiesByName\n  }\n\n  private readonly client: GoogleGenAI\n\n  constructor(config: GeminiImageConfig, model: TModel) {\n    super(model, config)\n    this.client = createGeminiClient(config)\n  }\n\n  async generateImages(\n    options: ImageGenerationOptions<GeminiImageProviderOptions>,\n  ): Promise<ImageGenerationResult> {\n    const { model, logger } = options\n\n    logger.request(\n      `activity=generateImage provider=gemini model=${this.model}`,\n      {\n        provider: 'gemini',\n        model: this.model,\n      },\n    )\n\n    try {\n      const resolved = resolveMediaPrompt(options.prompt)\n\n      // Image-only prompts are allowed (the image inputs carry the intent);\n      // a prompt with neither text nor images is always an error.\n      if (resolved.images.length === 0) {\n        validatePrompt({ prompt: resolved.text, model })\n      }\n\n      if (resolved.videos.length > 0) {\n        throw new Error(\n          `${this.name}.generateImages does not support video prompt parts (model: ${model}).`,\n        )\n      }\n      if (resolved.audios.length > 0) {\n        throw new Error(\n          `${this.name}.generateImages does not support audio prompt parts (model: ${model}).`,\n        )\n      }\n\n      if (this.isGeminiImageModel(model)) {\n        return await this.generateWithGeminiApi(options, resolved)\n      }\n\n      // Imagen does not accept image inputs — it's strictly text-to-image.\n      if (resolved.images.length > 0) {\n        throw new Error(\n          `${this.name}: model \"${model}\" (Imagen) does not support image prompt parts. ` +\n            `Use a Gemini-native image model (e.g. gemini-2.5-flash-image, \"nano-banana\") for image-conditioned generation.`,\n        )\n      }\n\n      // Imagen models path (generateImages API)\n      validateImageSize(model, options.size)\n      validateNumberOfImages(model, options.numberOfImages)\n\n      const config = this.buildImagenConfig(options)\n\n      const response = await this.client.models.generateImages({\n        model,\n        prompt: resolved.text,\n        config,\n      })\n\n      return this.transformImagenResponse(model, response)\n    } catch (error) {\n      logger.errors('gemini.generateImage fatal', {\n        error,\n        source: 'gemini.generateImage',\n      })\n      throw error\n    }\n  }\n\n  private isGeminiImageModel(model: string): boolean {\n    return model.startsWith('gemini-')\n  }\n\n  private async generateWithGeminiApi(\n    options: ImageGenerationOptions<GeminiImageProviderOptions>,\n    resolved: ResolvedMediaPrompt,\n  ): Promise<ImageGenerationResult> {\n    const { model, size, numberOfImages, modelOptions } = options\n\n    const parsedSize = size ? parseNativeImageSize(size) : undefined\n\n    // GeminiImageProviderOptions is Imagen-shaped — most fields\n    // (personGeneration, safetyFilterLevel, addWatermark, outputMimeType,\n    // outputCompressionQuality, guidanceScale, enhancePrompt,\n    // includeSafetyAttributes, includeRaiReason, outputGcsUri, labels,\n    // negativePrompt, language) are only valid on GenerateImagesConfig and\n    // would be rejected by the Gemini-native generateContent path. Pick only\n    // the fields that are valid on GenerateContentConfig instead of spreading\n    // the whole options object.\n    const nativeConfig: GenerateContentConfig = {}\n    if (modelOptions?.seed !== undefined) {\n      nativeConfig.seed = modelOptions.seed\n    }\n\n    const config: GenerateContentConfig = {\n      ...nativeConfig,\n      // Include TEXT so the model can interleave descriptions between images.\n      // IMPORTANT: responseModalities is a protected default — set it AFTER\n      // nativeConfig so nothing can silently disable image output.\n      responseModalities: ['TEXT', 'IMAGE'],\n      ...(parsedSize && {\n        imageConfig: {\n          ...(parsedSize.aspectRatio && {\n            aspectRatio: parsedSize.aspectRatio,\n          }),\n          ...(parsedSize.resolution && {\n            imageSize: parsedSize.resolution,\n          }),\n        },\n      }),\n    }\n\n    const contents = await this.buildContents(resolved, numberOfImages)\n\n    const response = await this.client.models.generateContent({\n      model,\n      contents,\n      config,\n    })\n\n    return this.transformGeminiResponse(model, response)\n  }\n\n  /**\n   * Build the multimodal `contents` payload. Text-only prompts pass through\n   * as a plain string (the SDK accepts it directly); prompts with image\n   * parts become a single user `Content` whose `parts` mirror the prompt's\n   * interleaved order — position is meaningful to Gemini (\"not like this\n   * *(image)*, more like this *(image)*\").\n   *\n   * The generateContent API has no numberOfImages parameter, so when more\n   * than one image is requested a trailing instruction is appended.\n   */\n  private async buildContents(\n    resolved: ResolvedMediaPrompt,\n    numberOfImages: number | undefined,\n  ): Promise<string | Array<Content>> {\n    const countInstruction =\n      numberOfImages && numberOfImages > 1\n        ? `Generate ${numberOfImages} distinct images.`\n        : undefined\n\n    if (resolved.images.length === 0) {\n      return countInstruction\n        ? `${resolved.text} ${countInstruction}`\n        : resolved.text\n    }\n\n    const parts: Array<Part> = await Promise.all(\n      resolved.parts.map((part) => {\n        if (part.type === 'text') {\n          return Promise.resolve<Part>({ text: part.content })\n        }\n        if (part.type === 'image') {\n          return this.imagePartToGeminiPart(part)\n        }\n        // Video / audio parts were rejected in generateImages above.\n        throw new Error(\n          `gemini: unsupported prompt part type \"${part.type}\" in image generation.`,\n        )\n      }),\n    )\n    if (countInstruction) {\n      parts.push({ text: countInstruction })\n    }\n    return [{ role: 'user', parts }]\n  }\n\n  private async imagePartToGeminiPart(\n    part: ImagePart<MediaInputMetadata>,\n  ): Promise<Part> {\n    if (part.source.type === 'data') {\n      return {\n        inlineData: {\n          mimeType: part.source.mimeType || 'image/png',\n          data: part.source.value,\n        },\n      }\n    }\n    // For URL sources, prefer passing the URL through as `fileData` when it\n    // looks like a Google Files API URI; otherwise fetch and inline as base64.\n    if (\n      part.source.value.startsWith('gs://') ||\n      /^https?:\\/\\/generativelanguage\\.googleapis\\.com\\//.test(\n        part.source.value,\n      )\n    ) {\n      return {\n        fileData: {\n          fileUri: part.source.value,\n          ...(part.source.mimeType && { mimeType: part.source.mimeType }),\n        },\n      }\n    }\n    const response = await fetch(part.source.value)\n    if (!response.ok) {\n      throw new Error(\n        `Failed to fetch image input (${response.status} ${response.statusText}): ${part.source.value}`,\n      )\n    }\n    const blob = await response.blob()\n    const buffer = await blob.arrayBuffer()\n    const base64 = arrayBufferToBase64(buffer)\n    return {\n      inlineData: {\n        mimeType: part.source.mimeType || blob.type || 'image/png',\n        data: base64,\n      },\n    }\n  }\n\n  private transformGeminiResponse(\n    model: string,\n    response: GenerateContentResponse,\n  ): ImageGenerationResult {\n    const images: Array<GeneratedImage> = []\n    const textParts: Array<string> = []\n    const parts = response.candidates?.[0]?.content?.parts ?? []\n\n    for (const part of parts) {\n      if (\n        part.inlineData?.data &&\n        typeof part.inlineData.data === 'string' &&\n        part.inlineData.data.length > 0\n      ) {\n        images.push({ b64Json: part.inlineData.data })\n      } else if (typeof part.text === 'string' && part.text.length > 0) {\n        textParts.push(part.text)\n      }\n    }\n\n    // If the model returned only text parts (for example a safety refusal\n    // or a \"can't do that\" message), surface the text instead of silently\n    // resolving to an empty images array — otherwise callers can't tell a\n    // generation failure apart from a genuine empty response.\n    if (images.length === 0) {\n      const reason =\n        textParts.length > 0\n          ? `: ${textParts.join(' ').trim()}`\n          : ' (no inline image or text parts were returned).'\n      throw new Error(`Gemini ${model} returned no images${reason}`)\n    }\n\n    return {\n      id: generateId(this.name),\n      model,\n      images,\n      // Surface token usage (with per-modality breakdown) when the model\n      // reports it (e.g. Nano Banana via generateContent). Conditionally spread\n      // to satisfy exactOptionalPropertyTypes — only include usage when\n      // present. See #330.\n      ...(response.usageMetadata\n        ? { usage: buildGeminiUsage(response.usageMetadata) }\n        : {}),\n    }\n  }\n\n  private buildImagenConfig(\n    options: ImageGenerationOptions<GeminiImageProviderOptions>,\n  ): GenerateImagesConfig {\n    const { size, numberOfImages, modelOptions } = options\n\n    // Build with conditional spreads — under exactOptionalPropertyTypes the\n    // vendor `GenerateImagesConfig` fields are `field?: T` (no `| undefined`),\n    // so we can only assign the property when we actually have a value.\n    const sizeAspectRatio = size ? sizeToAspectRatio(size) : undefined\n    return {\n      numberOfImages: numberOfImages ?? 1,\n      // Map size to aspect ratio if provided (modelOptions.aspectRatio will override)\n      ...(sizeAspectRatio !== undefined && { aspectRatio: sizeAspectRatio }),\n      ...modelOptions,\n    }\n  }\n\n  private transformImagenResponse(\n    model: string,\n    response: GenerateImagesResponse,\n  ): ImageGenerationResult {\n    const entries = response.generatedImages ?? []\n    const images: Array<GeneratedImage> = []\n    const filterReasons: Array<string> = []\n\n    for (const item of entries) {\n      const b64Json = item.image?.imageBytes\n      if (b64Json) {\n        images.push({\n          b64Json,\n          ...(item.enhancedPrompt !== undefined && {\n            revisedPrompt: item.enhancedPrompt,\n          }),\n        })\n        continue\n      }\n      // Imagen can drop individual entries with a raiFilteredReason when\n      // Responsible-AI filters fire. Preserve the reason so callers can\n      // surface it instead of silently getting back fewer images.\n      const reason = (item as { raiFilteredReason?: string }).raiFilteredReason\n      if (reason) {\n        filterReasons.push(reason)\n      }\n    }\n\n    // Every entry was filtered — no usable images to return. Throw rather\n    // than resolve to an empty array so the caller is forced to handle the\n    // failure mode explicitly.\n    if (entries.length > 0 && images.length === 0) {\n      const joined = filterReasons.length > 0 ? filterReasons.join('; ') : ''\n      throw new Error(\n        `Imagen ${model} returned no images: all ${entries.length} generated image(s) were filtered by Responsible-AI${joined ? ` (${joined})` : ''}.`,\n      )\n    }\n\n    // Partial filter: surface via console.warn since ImageGenerationResult\n    // has no warnings field. Callers that care can still inspect the count\n    // mismatch between requested and returned images.\n    if (filterReasons.length > 0 && typeof console !== 'undefined') {\n      console.warn(\n        `[gemini-image] ${filterReasons.length} of ${entries.length} images from ${model} were filtered by Responsible-AI: ${filterReasons.join('; ')}`,\n      )\n    }\n\n    return {\n      id: generateId(this.name),\n      model,\n      images,\n    }\n  }\n}\n\n/**\n * Creates a Gemini image adapter with explicit API key.\n * Type resolution happens here at the call site.\n *\n * @param model - The model name (e.g., 'imagen-3.0-generate-002')\n * @param apiKey - Your Google API key\n * @param config - Optional additional configuration\n * @returns Configured Gemini image adapter instance with resolved types\n *\n * @example\n * ```typescript\n * const adapter = createGeminiImage('imagen-3.0-generate-002', \"your-api-key\");\n *\n * const result = await generateImage({\n *   adapter,\n *   prompt: 'A cute baby sea otter'\n * });\n * ```\n */\nexport function createGeminiImage<TModel extends GeminiImageModel>(\n  model: TModel,\n  apiKey: string,\n  config?: Omit<GeminiImageConfig, 'apiKey'>,\n): GeminiImageAdapter<TModel> {\n  return new GeminiImageAdapter({ apiKey, ...config }, model)\n}\n\n/**\n * Creates a Gemini image adapter with automatic API key detection from environment variables.\n * Type resolution happens here at the call site.\n *\n * Looks for `GOOGLE_API_KEY` or `GEMINI_API_KEY` in:\n * - `process.env` (Node.js)\n * - `window.env` (Browser with injected env)\n *\n * @param model - The model name (e.g., 'imagen-4.0-generate-001')\n * @param config - Optional configuration (excluding apiKey which is auto-detected)\n * @returns Configured Gemini image adapter instance with resolved types\n * @throws Error if GOOGLE_API_KEY or GEMINI_API_KEY is not found in environment\n *\n * @example\n * ```typescript\n * // Automatically uses GOOGLE_API_KEY from environment\n * const adapter = geminiImage('imagen-4.0-generate-001');\n *\n * const result = await generateImage({\n *   adapter,\n *   prompt: 'A beautiful sunset over mountains'\n * });\n * ```\n */\nexport function geminiImage<TModel extends GeminiImageModel>(\n  model: TModel,\n  config?: Omit<GeminiImageConfig, 'apiKey'>,\n): GeminiImageAdapter<TModel> {\n  const apiKey = getGeminiApiKeyFromEnv()\n  return createGeminiImage(model, apiKey, config)\n}\n"],"names":[],"mappings":";;;;;;AAgEO,MAAM,2BAEH,iBAMR;AAAA,EACkB,OAAO;AAAA,EAChB,OAAO;AAAA,EAUC;AAAA,EAEjB,YAAY,QAA2B,OAAe;AACpD,UAAM,OAAO,MAAM;AACnB,SAAK,SAAS,mBAAmB,MAAM;AAAA,EACzC;AAAA,EAEA,MAAM,eACJ,SACgC;AAChC,UAAM,EAAE,OAAO,OAAA,IAAW;AAE1B,WAAO;AAAA,MACL,gDAAgD,KAAK,KAAK;AAAA,MAC1D;AAAA,QACE,UAAU;AAAA,QACV,OAAO,KAAK;AAAA,MAAA;AAAA,IACd;AAGF,QAAI;AACF,YAAM,WAAW,mBAAmB,QAAQ,MAAM;AAIlD,UAAI,SAAS,OAAO,WAAW,GAAG;AAChC,uBAAe,EAAE,QAAQ,SAAS,MAAM,OAAO;AAAA,MACjD;AAEA,UAAI,SAAS,OAAO,SAAS,GAAG;AAC9B,cAAM,IAAI;AAAA,UACR,GAAG,KAAK,IAAI,+DAA+D,KAAK;AAAA,QAAA;AAAA,MAEpF;AACA,UAAI,SAAS,OAAO,SAAS,GAAG;AAC9B,cAAM,IAAI;AAAA,UACR,GAAG,KAAK,IAAI,+DAA+D,KAAK;AAAA,QAAA;AAAA,MAEpF;AAEA,UAAI,KAAK,mBAAmB,KAAK,GAAG;AAClC,eAAO,MAAM,KAAK,sBAAsB,SAAS,QAAQ;AAAA,MAC3D;AAGA,UAAI,SAAS,OAAO,SAAS,GAAG;AAC9B,cAAM,IAAI;AAAA,UACR,GAAG,KAAK,IAAI,YAAY,KAAK;AAAA,QAAA;AAAA,MAGjC;AAGA,wBAAkB,OAAO,QAAQ,IAAI;AACrC,6BAAuB,OAAO,QAAQ,cAAc;AAEpD,YAAM,SAAS,KAAK,kBAAkB,OAAO;AAE7C,YAAM,WAAW,MAAM,KAAK,OAAO,OAAO,eAAe;AAAA,QACvD;AAAA,QACA,QAAQ,SAAS;AAAA,QACjB;AAAA,MAAA,CACD;AAED,aAAO,KAAK,wBAAwB,OAAO,QAAQ;AAAA,IACrD,SAAS,OAAO;AACd,aAAO,OAAO,8BAA8B;AAAA,QAC1C;AAAA,QACA,QAAQ;AAAA,MAAA,CACT;AACD,YAAM;AAAA,IACR;AAAA,EACF;AAAA,EAEQ,mBAAmB,OAAwB;AACjD,WAAO,MAAM,WAAW,SAAS;AAAA,EACnC;AAAA,EAEA,MAAc,sBACZ,SACA,UACgC;AAChC,UAAM,EAAE,OAAO,MAAM,gBAAgB,iBAAiB;AAEtD,UAAM,aAAa,OAAO,qBAAqB,IAAI,IAAI;AAUvD,UAAM,eAAsC,CAAA;AAC5C,QAAI,cAAc,SAAS,QAAW;AACpC,mBAAa,OAAO,aAAa;AAAA,IACnC;AAEA,UAAM,SAAgC;AAAA,MACpC,GAAG;AAAA;AAAA;AAAA;AAAA,MAIH,oBAAoB,CAAC,QAAQ,OAAO;AAAA,MACpC,GAAI,cAAc;AAAA,QAChB,aAAa;AAAA,UACX,GAAI,WAAW,eAAe;AAAA,YAC5B,aAAa,WAAW;AAAA,UAAA;AAAA,UAE1B,GAAI,WAAW,cAAc;AAAA,YAC3B,WAAW,WAAW;AAAA,UAAA;AAAA,QACxB;AAAA,MACF;AAAA,IACF;AAGF,UAAM,WAAW,MAAM,KAAK,cAAc,UAAU,cAAc;AAElE,UAAM,WAAW,MAAM,KAAK,OAAO,OAAO,gBAAgB;AAAA,MACxD;AAAA,MACA;AAAA,MACA;AAAA,IAAA,CACD;AAED,WAAO,KAAK,wBAAwB,OAAO,QAAQ;AAAA,EACrD;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAYA,MAAc,cACZ,UACA,gBACkC;AAClC,UAAM,mBACJ,kBAAkB,iBAAiB,IAC/B,YAAY,cAAc,sBAC1B;AAEN,QAAI,SAAS,OAAO,WAAW,GAAG;AAChC,aAAO,mBACH,GAAG,SAAS,IAAI,IAAI,gBAAgB,KACpC,SAAS;AAAA,IACf;AAEA,UAAM,QAAqB,MAAM,QAAQ;AAAA,MACvC,SAAS,MAAM,IAAI,CAAC,SAAS;AAC3B,YAAI,KAAK,SAAS,QAAQ;AACxB,iBAAO,QAAQ,QAAc,EAAE,MAAM,KAAK,SAAS;AAAA,QACrD;AACA,YAAI,KAAK,SAAS,SAAS;AACzB,iBAAO,KAAK,sBAAsB,IAAI;AAAA,QACxC;AAEA,cAAM,IAAI;AAAA,UACR,yCAAyC,KAAK,IAAI;AAAA,QAAA;AAAA,MAEtD,CAAC;AAAA,IAAA;AAEH,QAAI,kBAAkB;AACpB,YAAM,KAAK,EAAE,MAAM,iBAAA,CAAkB;AAAA,IACvC;AACA,WAAO,CAAC,EAAE,MAAM,QAAQ,OAAO;AAAA,EACjC;AAAA,EAEA,MAAc,sBACZ,MACe;AACf,QAAI,KAAK,OAAO,SAAS,QAAQ;AAC/B,aAAO;AAAA,QACL,YAAY;AAAA,UACV,UAAU,KAAK,OAAO,YAAY;AAAA,UAClC,MAAM,KAAK,OAAO;AAAA,QAAA;AAAA,MACpB;AAAA,IAEJ;AAGA,QACE,KAAK,OAAO,MAAM,WAAW,OAAO,KACpC,oDAAoD;AAAA,MAClD,KAAK,OAAO;AAAA,IAAA,GAEd;AACA,aAAO;AAAA,QACL,UAAU;AAAA,UACR,SAAS,KAAK,OAAO;AAAA,UACrB,GAAI,KAAK,OAAO,YAAY,EAAE,UAAU,KAAK,OAAO,SAAA;AAAA,QAAS;AAAA,MAC/D;AAAA,IAEJ;AACA,UAAM,WAAW,MAAM,MAAM,KAAK,OAAO,KAAK;AAC9C,QAAI,CAAC,SAAS,IAAI;AAChB,YAAM,IAAI;AAAA,QACR,gCAAgC,SAAS,MAAM,IAAI,SAAS,UAAU,MAAM,KAAK,OAAO,KAAK;AAAA,MAAA;AAAA,IAEjG;AACA,UAAM,OAAO,MAAM,SAAS,KAAA;AAC5B,UAAM,SAAS,MAAM,KAAK,YAAA;AAC1B,UAAM,SAAS,oBAAoB,MAAM;AACzC,WAAO;AAAA,MACL,YAAY;AAAA,QACV,UAAU,KAAK,OAAO,YAAY,KAAK,QAAQ;AAAA,QAC/C,MAAM;AAAA,MAAA;AAAA,IACR;AAAA,EAEJ;AAAA,EAEQ,wBACN,OACA,UACuB;AACvB,UAAM,SAAgC,CAAA;AACtC,UAAM,YAA2B,CAAA;AACjC,UAAM,QAAQ,SAAS,aAAa,CAAC,GAAG,SAAS,SAAS,CAAA;AAE1D,eAAW,QAAQ,OAAO;AACxB,UACE,KAAK,YAAY,QACjB,OAAO,KAAK,WAAW,SAAS,YAChC,KAAK,WAAW,KAAK,SAAS,GAC9B;AACA,eAAO,KAAK,EAAE,SAAS,KAAK,WAAW,MAAM;AAAA,MAC/C,WAAW,OAAO,KAAK,SAAS,YAAY,KAAK,KAAK,SAAS,GAAG;AAChE,kBAAU,KAAK,KAAK,IAAI;AAAA,MAC1B;AAAA,IACF;AAMA,QAAI,OAAO,WAAW,GAAG;AACvB,YAAM,SACJ,UAAU,SAAS,IACf,KAAK,UAAU,KAAK,GAAG,EAAE,KAAA,CAAM,KAC/B;AACN,YAAM,IAAI,MAAM,UAAU,KAAK,sBAAsB,MAAM,EAAE;AAAA,IAC/D;AAEA,WAAO;AAAA,MACL,IAAI,WAAW,KAAK,IAAI;AAAA,MACxB;AAAA,MACA;AAAA;AAAA;AAAA;AAAA;AAAA,MAKA,GAAI,SAAS,gBACT,EAAE,OAAO,iBAAiB,SAAS,aAAa,MAChD,CAAA;AAAA,IAAC;AAAA,EAET;AAAA,EAEQ,kBACN,SACsB;AACtB,UAAM,EAAE,MAAM,gBAAgB,aAAA,IAAiB;AAK/C,UAAM,kBAAkB,OAAO,kBAAkB,IAAI,IAAI;AACzD,WAAO;AAAA,MACL,gBAAgB,kBAAkB;AAAA;AAAA,MAElC,GAAI,oBAAoB,UAAa,EAAE,aAAa,gBAAA;AAAA,MACpD,GAAG;AAAA,IAAA;AAAA,EAEP;AAAA,EAEQ,wBACN,OACA,UACuB;AACvB,UAAM,UAAU,SAAS,mBAAmB,CAAA;AAC5C,UAAM,SAAgC,CAAA;AACtC,UAAM,gBAA+B,CAAA;AAErC,eAAW,QAAQ,SAAS;AAC1B,YAAM,UAAU,KAAK,OAAO;AAC5B,UAAI,SAAS;AACX,eAAO,KAAK;AAAA,UACV;AAAA,UACA,GAAI,KAAK,mBAAmB,UAAa;AAAA,YACvC,eAAe,KAAK;AAAA,UAAA;AAAA,QACtB,CACD;AACD;AAAA,MACF;AAIA,YAAM,SAAU,KAAwC;AACxD,UAAI,QAAQ;AACV,sBAAc,KAAK,MAAM;AAAA,MAC3B;AAAA,IACF;AAKA,QAAI,QAAQ,SAAS,KAAK,OAAO,WAAW,GAAG;AAC7C,YAAM,SAAS,cAAc,SAAS,IAAI,cAAc,KAAK,IAAI,IAAI;AACrE,YAAM,IAAI;AAAA,QACR,UAAU,KAAK,4BAA4B,QAAQ,MAAM,sDAAsD,SAAS,KAAK,MAAM,MAAM,EAAE;AAAA,MAAA;AAAA,IAE/I;AAKA,QAAI,cAAc,SAAS,KAAK,OAAO,YAAY,aAAa;AAC9D,cAAQ;AAAA,QACN,kBAAkB,cAAc,MAAM,OAAO,QAAQ,MAAM,gBAAgB,KAAK,qCAAqC,cAAc,KAAK,IAAI,CAAC;AAAA,MAAA;AAAA,IAEjJ;AAEA,WAAO;AAAA,MACL,IAAI,WAAW,KAAK,IAAI;AAAA,MACxB;AAAA,MACA;AAAA,IAAA;AAAA,EAEJ;AACF;AAqBO,SAAS,kBACd,OACA,QACA,QAC4B;AAC5B,SAAO,IAAI,mBAAmB,EAAE,QAAQ,GAAG,OAAA,GAAU,KAAK;AAC5D;AA0BO,SAAS,YACd,OACA,QAC4B;AAC5B,QAAM,SAAS,uBAAA;AACf,SAAO,kBAAkB,OAAO,QAAQ,MAAM;AAChD;"}

package/dist/esm/adapters/video.d.ts

@@ -0,0 +1,108 @@
+import { BaseVideoAdapter, DurationOptions } from '@tanstack/ai/adapters';
+import { VideoGenerationOptions, VideoJobResult, VideoStatusResult, VideoUrlResult } from '@tanstack/ai';
+import { GoogleGenAI } from '@google/genai';
+import { GeminiVideoModel, GeminiVideoModelDurationByName, GeminiVideoModelInputModalitiesByName, GeminiVideoModelProviderOptionsByName, GeminiVideoModelSizeByName, GeminiVideoProviderOptions, GeminiVideoSize } from '../video/video-provider-options.js';
+import { GeminiClientConfig } from '../utils.js';
+/**
+ * Configuration for Gemini video adapter.
+ *
+ * @experimental Video generation is an experimental feature and may change.
+ */
+export interface GeminiVideoConfig extends GeminiClientConfig {
+}
+/**
+ * Gemini Veo Video Generation Adapter
+ *
+ * Tree-shakeable adapter for Google Veo video generation. Veo runs as a
+ * long-running operation: `createVideoJob` starts the operation via the
+ * `:predictLongRunning` endpoint, `getVideoStatus` polls it, and
+ * `getVideoUrl` extracts the generated video's URI once it completes.
+ *
+ * Image prompt parts are routed by `metadata.role`:
+ * - `'start_frame'` (or the first un-roled image) → the input image the
+ *   video starts from
+ * - `'end_frame'` → `lastFrame` (the frame the video ends on)
+ * - `'reference'` / `'character'` → `referenceImages` (asset references,
+ *   Veo 3.1)
+ *
+ * Note: the returned video URI is served by the Gemini Files API and
+ * requires the API key (`x-goog-api-key` header or `?key=` query
+ * parameter) to download.
+ *
+ * @experimental Video generation is an experimental feature and may change.
+ */
+export declare class GeminiVideoAdapter<TModel extends GeminiVideoModel> extends BaseVideoAdapter<TModel, GeminiVideoProviderOptions, GeminiVideoModelProviderOptionsByName, GeminiVideoModelSizeByName, GeminiVideoModelInputModalitiesByName, GeminiVideoModelDurationByName> {
+    readonly name: "gemini";
+    protected client: GoogleGenAI;
+    constructor(config: GeminiVideoConfig, model: TModel);
+    createVideoJob(options: VideoGenerationOptions<GeminiVideoProviderOptions, GeminiVideoSize, GeminiVideoModelDurationByName[TModel]>): Promise<VideoJobResult>;
+    /**
+     * Route image prompt parts onto Veo's request fields by `metadata.role`.
+     */
+    private routeImageParts;
+    getVideoStatus(jobId: string): Promise<VideoStatusResult>;
+    getVideoUrl(jobId: string): Promise<VideoUrlResult>;
+    availableDurations(): DurationOptions<GeminiVideoModelDurationByName[TModel]>;
+    snapDuration(seconds: number): GeminiVideoModelDurationByName[TModel] | undefined;
+    /**
+     * Fetch the long-running operation by name. The SDK's
+     * `operations.getVideosOperation` needs a real `GenerateVideosOperation`
+     * instance (it calls `_fromAPIResponse` on it), so reconstruct one from
+     * the job ID rather than passing an object literal.
+     */
+    private getOperation;
+}
+/**
+ * Creates a Gemini video adapter with an explicit API key.
+ * Type resolution happens here at the call site.
+ *
+ * @experimental Video generation is an experimental feature and may change.
+ *
+ * @param model - The model name (e.g., 'veo-3.1-generate-preview')
+ * @param apiKey - Your Google API key
+ * @param config - Optional additional configuration
+ * @returns Configured Gemini video adapter instance with resolved types
+ *
+ * @example
+ * ```typescript
+ * const adapter = createGeminiVideo('veo-3.1-generate-preview', 'your-api-key');
+ *
+ * const { jobId } = await generateVideo({
+ *   adapter,
+ *   prompt: 'A beautiful sunset over the ocean',
+ *   duration: adapter.snapDuration(7), // → 6
+ * });
+ * ```
+ */
+export declare function createGeminiVideo<TModel extends GeminiVideoModel>(model: TModel, apiKey: string, config?: Omit<GeminiVideoConfig, 'apiKey'>): GeminiVideoAdapter<TModel>;
+/**
+ * Creates a Gemini video adapter with automatic API key detection from environment variables.
+ * Type resolution happens here at the call site.
+ *
+ * Looks for `GOOGLE_API_KEY` or `GEMINI_API_KEY` in:
+ * - `process.env` (Node.js)
+ * - `window.env` (Browser with injected env)
+ *
+ * @experimental Video generation is an experimental feature and may change.
+ *
+ * @param model - The model name (e.g., 'veo-3.1-generate-preview')
+ * @param config - Optional configuration (excluding apiKey which is auto-detected)
+ * @returns Configured Gemini video adapter instance with resolved types
+ * @throws Error if GOOGLE_API_KEY or GEMINI_API_KEY is not found in environment
+ *
+ * @example
+ * ```typescript
+ * // Automatically uses GOOGLE_API_KEY from environment
+ * const adapter = geminiVideo('veo-3.1-generate-preview');
+ *
+ * // Create a video generation job
+ * const { jobId } = await generateVideo({
+ *   adapter,
+ *   prompt: 'A cat playing piano'
+ * });
+ *
+ * // Poll for status
+ * const status = await getVideoJobStatus({ adapter, jobId });
+ * ```
+ */
+export declare function geminiVideo<TModel extends GeminiVideoModel>(model: TModel, config?: Omit<GeminiVideoConfig, 'apiKey'>): GeminiVideoAdapter<TModel>;

package/dist/esm/adapters/video.js

@@ -0,0 +1,227 @@
+import { VideoGenerationReferenceType, GenerateVideosOperation } from "@google/genai";
+import { resolveMediaPrompt } from "@tanstack/ai";
+import { BaseVideoAdapter, snapToDurationOption } from "@tanstack/ai/adapters";
+import { arrayBufferToBase64 } from "@tanstack/ai-utils";
+import { createGeminiClient, getGeminiApiKeyFromEnv } from "../utils/client.js";
+import { getGeminiVideoDurationOptions } from "../video/video-provider-options.js";
+function operationErrorMessage(error) {
+  if (typeof error.message === "string" && error.message.length > 0) {
+    return error.message;
+  }
+  return JSON.stringify(error);
+}
+async function imagePartToVeoImage(part) {
+  if (part.source.type === "data") {
+    return {
+      imageBytes: part.source.value,
+      mimeType: part.source.mimeType || "image/png"
+    };
+  }
+  const url = part.source.value;
+  if (url.startsWith("gs://")) {
+    return {
+      gcsUri: url,
+      ...part.source.mimeType && { mimeType: part.source.mimeType }
+    };
+  }
+  if (url.startsWith("data:")) {
+    const match = url.match(/^data:([^;,]+)?(;base64)?,(.*)$/);
+    if (!match || !match[2]) {
+      throw new Error(
+        "gemini: only base64 data: URIs are supported for video image inputs."
+      );
+    }
+    return {
+      imageBytes: match[3] ?? "",
+      mimeType: match[1] || part.source.mimeType || "image/png"
+    };
+  }
+  const response = await fetch(url);
+  if (!response.ok) {
+    throw new Error(
+      `Failed to fetch image input (${response.status} ${response.statusText}): ${url}`
+    );
+  }
+  const blob = await response.blob();
+  const buffer = await blob.arrayBuffer();
+  return {
+    imageBytes: arrayBufferToBase64(buffer),
+    mimeType: part.source.mimeType || blob.type || "image/png"
+  };
+}
+class GeminiVideoAdapter extends BaseVideoAdapter {
+  name = "gemini";
+  client;
+  constructor(config, model) {
+    super({}, model);
+    this.client = createGeminiClient(config);
+  }
+  async createVideoJob(options) {
+    const { prompt, size, duration, modelOptions, logger } = options;
+    logger.request(
+      `activity=video.create provider=${this.name} model=${this.model} size=${size ?? "default"} duration=${duration ?? "default"}`,
+      { provider: this.name, model: this.model }
+    );
+    try {
+      const resolved = resolveMediaPrompt(prompt);
+      if (resolved.videos.length > 0) {
+        throw new Error(
+          `${this.name}.createVideoJob does not support video prompt parts (model: ${this.model}).`
+        );
+      }
+      if (resolved.audios.length > 0) {
+        throw new Error(
+          `${this.name}.createVideoJob does not support audio prompt parts (model: ${this.model}).`
+        );
+      }
+      const { image, lastFrame, referenceImages } = await this.routeImageParts(
+        resolved.images
+      );
+      const config = {
+        ...modelOptions,
+        ...size !== void 0 && { aspectRatio: size },
+        ...duration !== void 0 && { durationSeconds: duration },
+        ...lastFrame && { lastFrame },
+        ...referenceImages.length > 0 && { referenceImages }
+      };
+      const operation = await this.client.models.generateVideos({
+        model: this.model,
+        prompt: resolved.text,
+        ...image && { image },
+        config
+      });
+      if (!operation.name) {
+        throw new Error(
+          "Veo did not return an operation name for the video generation job."
+        );
+      }
+      return { jobId: operation.name, model: this.model };
+    } catch (error) {
+      logger.errors(`${this.name}.createVideoJob fatal`, {
+        error,
+        source: `${this.name}.createVideoJob`
+      });
+      throw error;
+    }
+  }
+  /**
+   * Route image prompt parts onto Veo's request fields by `metadata.role`.
+   */
+  async routeImageParts(parts) {
+    let image;
+    let lastFrame;
+    const referenceImages = [];
+    for (const part of parts) {
+      const role = part.metadata?.role;
+      switch (role) {
+        case "end_frame": {
+          if (lastFrame) {
+            throw new Error(
+              `${this.name}: Veo accepts at most one 'end_frame' image.`
+            );
+          }
+          lastFrame = await imagePartToVeoImage(part);
+          break;
+        }
+        case "reference":
+        case "character": {
+          referenceImages.push({
+            image: await imagePartToVeoImage(part),
+            referenceType: VideoGenerationReferenceType.ASSET
+          });
+          break;
+        }
+        case "start_frame":
+        case void 0: {
+          if (image) {
+            throw new Error(
+              `${this.name}: Veo accepts at most one starting image; received multiple 'start_frame'/un-roled images. Use metadata.role ('end_frame', 'reference') to disambiguate the others.`
+            );
+          }
+          image = await imagePartToVeoImage(part);
+          break;
+        }
+        case "mask":
+        case "control":
+          throw new Error(
+            `${this.name}: unsupported image role "${role}" for Veo video generation.`
+          );
+      }
+    }
+    return { image, lastFrame, referenceImages };
+  }
+  async getVideoStatus(jobId) {
+    const operation = await this.getOperation(jobId);
+    if (!operation.done) {
+      return { jobId, status: "processing" };
+    }
+    if (operation.error) {
+      return {
+        jobId,
+        status: "failed",
+        error: operationErrorMessage(operation.error)
+      };
+    }
+    const videos = operation.response?.generatedVideos ?? [];
+    if (videos.length === 0) {
+      const reasons = operation.response?.raiMediaFilteredReasons;
+      return {
+        jobId,
+        status: "failed",
+        error: reasons?.length ? `Video was filtered by Responsible-AI: ${reasons.join("; ")}` : "Veo returned no generated videos."
+      };
+    }
+    return { jobId, status: "completed" };
+  }
+  async getVideoUrl(jobId) {
+    const operation = await this.getOperation(jobId);
+    if (!operation.done) {
+      throw new Error(
+        `Video is not ready yet. Check status first. Job ID: ${jobId}`
+      );
+    }
+    if (operation.error) {
+      throw new Error(
+        `Video generation failed: ${operationErrorMessage(operation.error)}`
+      );
+    }
+    const uri = operation.response?.generatedVideos?.[0]?.video?.uri;
+    if (!uri) {
+      const reasons = operation.response?.raiMediaFilteredReasons;
+      throw new Error(
+        reasons?.length ? `Video was filtered by Responsible-AI: ${reasons.join("; ")}` : `Video URL not found in operation response. Job ID: ${jobId}`
+      );
+    }
+    return { jobId, url: uri };
+  }
+  availableDurations() {
+    return getGeminiVideoDurationOptions(this.model);
+  }
+  snapDuration(seconds) {
+    return snapToDurationOption(seconds, this.availableDurations());
+  }
+  /**
+   * Fetch the long-running operation by name. The SDK's
+   * `operations.getVideosOperation` needs a real `GenerateVideosOperation`
+   * instance (it calls `_fromAPIResponse` on it), so reconstruct one from
+   * the job ID rather than passing an object literal.
+   */
+  async getOperation(jobId) {
+    const operation = new GenerateVideosOperation();
+    operation.name = jobId;
+    return await this.client.operations.getVideosOperation({ operation });
+  }
+}
+function createGeminiVideo(model, apiKey, config) {
+  return new GeminiVideoAdapter({ apiKey, ...config }, model);
+}
+function geminiVideo(model, config) {
+  const apiKey = getGeminiApiKeyFromEnv();
+  return createGeminiVideo(model, apiKey, config);
+}
+export {
+  GeminiVideoAdapter,
+  createGeminiVideo,
+  geminiVideo
+};
+//# sourceMappingURL=video.js.map

package/dist/esm/adapters/video.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"video.js","sources":["../../../src/adapters/video.ts"],"sourcesContent":["import {\n  GenerateVideosOperation,\n  VideoGenerationReferenceType,\n} from '@google/genai'\nimport { resolveMediaPrompt } from '@tanstack/ai'\nimport { BaseVideoAdapter, snapToDurationOption } from '@tanstack/ai/adapters'\nimport { arrayBufferToBase64 } from '@tanstack/ai-utils'\nimport { createGeminiClient, getGeminiApiKeyFromEnv } from '../utils'\nimport { getGeminiVideoDurationOptions } from '../video/video-provider-options'\nimport type { DurationOptions } from '@tanstack/ai/adapters'\nimport type {\n  ImagePart,\n  MediaInputMetadata,\n  VideoGenerationOptions,\n  VideoJobResult,\n  VideoStatusResult,\n  VideoUrlResult,\n} from '@tanstack/ai'\nimport type {\n  GenerateVideosConfig,\n  GoogleGenAI,\n  Image,\n  VideoGenerationReferenceImage,\n} from '@google/genai'\nimport type {\n  GeminiVideoModel,\n  GeminiVideoModelDurationByName,\n  GeminiVideoModelInputModalitiesByName,\n  GeminiVideoModelProviderOptionsByName,\n  GeminiVideoModelSizeByName,\n  GeminiVideoProviderOptions,\n  GeminiVideoSize,\n} from '../video/video-provider-options'\nimport type { GeminiClientConfig } from '../utils'\n\n/**\n * Configuration for Gemini video adapter.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport interface GeminiVideoConfig extends GeminiClientConfig {}\n\n/**\n * Extract a human-readable message from a long-running operation's error,\n * which the SDK types as `Record<string, unknown>` (a google.rpc.Status).\n */\nfunction operationErrorMessage(error: Record<string, unknown>): string {\n  if (typeof error.message === 'string' && error.message.length > 0) {\n    return error.message\n  }\n  return JSON.stringify(error)\n}\n\n/**\n * Convert a TanStack image prompt part into the genai `Image` shape Veo\n * accepts: base64 `imageBytes` (data sources, data: URIs, fetched HTTP\n * URLs) or a `gcsUri` passthrough for Cloud Storage references.\n */\nasync function imagePartToVeoImage(\n  part: ImagePart<MediaInputMetadata>,\n): Promise<Image> {\n  if (part.source.type === 'data') {\n    return {\n      imageBytes: part.source.value,\n      mimeType: part.source.mimeType || 'image/png',\n    }\n  }\n  const url = part.source.value\n  if (url.startsWith('gs://')) {\n    return {\n      gcsUri: url,\n      ...(part.source.mimeType && { mimeType: part.source.mimeType }),\n    }\n  }\n  if (url.startsWith('data:')) {\n    const match = url.match(/^data:([^;,]+)?(;base64)?,(.*)$/)\n    if (!match || !match[2]) {\n      throw new Error(\n        'gemini: only base64 data: URIs are supported for video image inputs.',\n      )\n    }\n    return {\n      imageBytes: match[3] ?? '',\n      mimeType: match[1] || part.source.mimeType || 'image/png',\n    }\n  }\n  const response = await fetch(url)\n  if (!response.ok) {\n    throw new Error(\n      `Failed to fetch image input (${response.status} ${response.statusText}): ${url}`,\n    )\n  }\n  const blob = await response.blob()\n  const buffer = await blob.arrayBuffer()\n  return {\n    imageBytes: arrayBufferToBase64(buffer),\n    mimeType: part.source.mimeType || blob.type || 'image/png',\n  }\n}\n\n/**\n * Gemini Veo Video Generation Adapter\n *\n * Tree-shakeable adapter for Google Veo video generation. Veo runs as a\n * long-running operation: `createVideoJob` starts the operation via the\n * `:predictLongRunning` endpoint, `getVideoStatus` polls it, and\n * `getVideoUrl` extracts the generated video's URI once it completes.\n *\n * Image prompt parts are routed by `metadata.role`:\n * - `'start_frame'` (or the first un-roled image) → the input image the\n *   video starts from\n * - `'end_frame'` → `lastFrame` (the frame the video ends on)\n * - `'reference'` / `'character'` → `referenceImages` (asset references,\n *   Veo 3.1)\n *\n * Note: the returned video URI is served by the Gemini Files API and\n * requires the API key (`x-goog-api-key` header or `?key=` query\n * parameter) to download.\n *\n * @experimental Video generation is an experimental feature and may change.\n */\nexport class GeminiVideoAdapter<\n  TModel extends GeminiVideoModel,\n> extends BaseVideoAdapter<\n  TModel,\n  GeminiVideoProviderOptions,\n  GeminiVideoModelProviderOptionsByName,\n  GeminiVideoModelSizeByName,\n  GeminiVideoModelInputModalitiesByName,\n  GeminiVideoModelDurationByName\n> {\n  readonly name = 'gemini' as const\n\n  protected client: GoogleGenAI\n\n  constructor(config: GeminiVideoConfig, model: TModel) {\n    super({}, model)\n    this.client = createGeminiClient(config)\n  }\n\n  async createVideoJob(\n    options: VideoGenerationOptions<\n      GeminiVideoProviderOptions,\n      GeminiVideoSize,\n      GeminiVideoModelDurationByName[TModel]\n    >,\n  ): Promise<VideoJobResult> {\n    const { prompt, size, duration, modelOptions, logger } = options\n\n    logger.request(\n      `activity=video.create provider=${this.name} model=${this.model} size=${size ?? 'default'} duration=${duration ?? 'default'}`,\n      { provider: this.name, model: this.model },\n    )\n\n    try {\n      const resolved = resolveMediaPrompt(prompt)\n\n      if (resolved.videos.length > 0) {\n        throw new Error(\n          `${this.name}.createVideoJob does not support video prompt parts (model: ${this.model}).`,\n        )\n      }\n      if (resolved.audios.length > 0) {\n        throw new Error(\n          `${this.name}.createVideoJob does not support audio prompt parts (model: ${this.model}).`,\n        )\n      }\n\n      const { image, lastFrame, referenceImages } = await this.routeImageParts(\n        resolved.images,\n      )\n\n      const config: GenerateVideosConfig = {\n        ...modelOptions,\n        ...(size !== undefined && { aspectRatio: size }),\n        ...(duration !== undefined && { durationSeconds: duration }),\n        ...(lastFrame && { lastFrame }),\n        ...(referenceImages.length > 0 && { referenceImages }),\n      }\n\n      const operation = await this.client.models.generateVideos({\n        model: this.model,\n        prompt: resolved.text,\n        ...(image && { image }),\n        config,\n      })\n\n      if (!operation.name) {\n        throw new Error(\n          'Veo did not return an operation name for the video generation job.',\n        )\n      }\n\n      return { jobId: operation.name, model: this.model }\n    } catch (error) {\n      logger.errors(`${this.name}.createVideoJob fatal`, {\n        error,\n        source: `${this.name}.createVideoJob`,\n      })\n      throw error\n    }\n  }\n\n  /**\n   * Route image prompt parts onto Veo's request fields by `metadata.role`.\n   */\n  private async routeImageParts(\n    parts: Array<ImagePart<MediaInputMetadata>>,\n  ): Promise<{\n    image: Image | undefined\n    lastFrame: Image | undefined\n    referenceImages: Array<VideoGenerationReferenceImage>\n  }> {\n    let image: Image | undefined\n    let lastFrame: Image | undefined\n    const referenceImages: Array<VideoGenerationReferenceImage> = []\n\n    for (const part of parts) {\n      const role = part.metadata?.role\n      switch (role) {\n        case 'end_frame': {\n          if (lastFrame) {\n            throw new Error(\n              `${this.name}: Veo accepts at most one 'end_frame' image.`,\n            )\n          }\n          lastFrame = await imagePartToVeoImage(part)\n          break\n        }\n        case 'reference':\n        case 'character': {\n          referenceImages.push({\n            image: await imagePartToVeoImage(part),\n            referenceType: VideoGenerationReferenceType.ASSET,\n          })\n          break\n        }\n        case 'start_frame':\n        case undefined: {\n          if (image) {\n            throw new Error(\n              `${this.name}: Veo accepts at most one starting image; received multiple 'start_frame'/un-roled images. Use metadata.role ('end_frame', 'reference') to disambiguate the others.`,\n            )\n          }\n          image = await imagePartToVeoImage(part)\n          break\n        }\n        case 'mask':\n        case 'control':\n          throw new Error(\n            `${this.name}: unsupported image role \"${role}\" for Veo video generation.`,\n          )\n      }\n    }\n\n    return { image, lastFrame, referenceImages }\n  }\n\n  async getVideoStatus(jobId: string): Promise<VideoStatusResult> {\n    const operation = await this.getOperation(jobId)\n\n    if (!operation.done) {\n      return { jobId, status: 'processing' }\n    }\n\n    if (operation.error) {\n      return {\n        jobId,\n        status: 'failed',\n        error: operationErrorMessage(operation.error),\n      }\n    }\n\n    // The operation can finish \"successfully\" with every sample dropped by\n    // Responsible-AI filters — surface that as a failure instead of letting\n    // getVideoUrl() throw on an empty response.\n    const videos = operation.response?.generatedVideos ?? []\n    if (videos.length === 0) {\n      const reasons = operation.response?.raiMediaFilteredReasons\n      return {\n        jobId,\n        status: 'failed',\n        error: reasons?.length\n          ? `Video was filtered by Responsible-AI: ${reasons.join('; ')}`\n          : 'Veo returned no generated videos.',\n      }\n    }\n\n    return { jobId, status: 'completed' }\n  }\n\n  async getVideoUrl(jobId: string): Promise<VideoUrlResult> {\n    const operation = await this.getOperation(jobId)\n\n    if (!operation.done) {\n      throw new Error(\n        `Video is not ready yet. Check status first. Job ID: ${jobId}`,\n      )\n    }\n\n    if (operation.error) {\n      throw new Error(\n        `Video generation failed: ${operationErrorMessage(operation.error)}`,\n      )\n    }\n\n    const uri = operation.response?.generatedVideos?.[0]?.video?.uri\n    if (!uri) {\n      const reasons = operation.response?.raiMediaFilteredReasons\n      throw new Error(\n        reasons?.length\n          ? `Video was filtered by Responsible-AI: ${reasons.join('; ')}`\n          : `Video URL not found in operation response. Job ID: ${jobId}`,\n      )\n    }\n\n    return { jobId, url: uri }\n  }\n\n  override availableDurations(): DurationOptions<\n    GeminiVideoModelDurationByName[TModel]\n  > {\n    return getGeminiVideoDurationOptions(this.model)\n  }\n\n  override snapDuration(\n    seconds: number,\n  ): GeminiVideoModelDurationByName[TModel] | undefined {\n    return snapToDurationOption(seconds, this.availableDurations())\n  }\n\n  /**\n   * Fetch the long-running operation by name. The SDK's\n   * `operations.getVideosOperation` needs a real `GenerateVideosOperation`\n   * instance (it calls `_fromAPIResponse` on it), so reconstruct one from\n   * the job ID rather than passing an object literal.\n   */\n  private async getOperation(jobId: string): Promise<GenerateVideosOperation> {\n    const operation = new GenerateVideosOperation()\n    operation.name = jobId\n    return await this.client.operations.getVideosOperation({ operation })\n  }\n}\n\n/**\n * Creates a Gemini video adapter with an explicit API key.\n * Type resolution happens here at the call site.\n *\n * @experimental Video generation is an experimental feature and may change.\n *\n * @param model - The model name (e.g., 'veo-3.1-generate-preview')\n * @param apiKey - Your Google API key\n * @param config - Optional additional configuration\n * @returns Configured Gemini video adapter instance with resolved types\n *\n * @example\n * ```typescript\n * const adapter = createGeminiVideo('veo-3.1-generate-preview', 'your-api-key');\n *\n * const { jobId } = await generateVideo({\n *   adapter,\n *   prompt: 'A beautiful sunset over the ocean',\n *   duration: adapter.snapDuration(7), // → 6\n * });\n * ```\n */\nexport function createGeminiVideo<TModel extends GeminiVideoModel>(\n  model: TModel,\n  apiKey: string,\n  config?: Omit<GeminiVideoConfig, 'apiKey'>,\n): GeminiVideoAdapter<TModel> {\n  return new GeminiVideoAdapter({ apiKey, ...config }, model)\n}\n\n/**\n * Creates a Gemini video adapter with automatic API key detection from environment variables.\n * Type resolution happens here at the call site.\n *\n * Looks for `GOOGLE_API_KEY` or `GEMINI_API_KEY` in:\n * - `process.env` (Node.js)\n * - `window.env` (Browser with injected env)\n *\n * @experimental Video generation is an experimental feature and may change.\n *\n * @param model - The model name (e.g., 'veo-3.1-generate-preview')\n * @param config - Optional configuration (excluding apiKey which is auto-detected)\n * @returns Configured Gemini video adapter instance with resolved types\n * @throws Error if GOOGLE_API_KEY or GEMINI_API_KEY is not found in environment\n *\n * @example\n * ```typescript\n * // Automatically uses GOOGLE_API_KEY from environment\n * const adapter = geminiVideo('veo-3.1-generate-preview');\n *\n * // Create a video generation job\n * const { jobId } = await generateVideo({\n *   adapter,\n *   prompt: 'A cat playing piano'\n * });\n *\n * // Poll for status\n * const status = await getVideoJobStatus({ adapter, jobId });\n * ```\n */\nexport function geminiVideo<TModel extends GeminiVideoModel>(\n  model: TModel,\n  config?: Omit<GeminiVideoConfig, 'apiKey'>,\n): GeminiVideoAdapter<TModel> {\n  const apiKey = getGeminiApiKeyFromEnv()\n  return createGeminiVideo(model, apiKey, config)\n}\n"],"names":[],"mappings":";;;;;;AA8CA,SAAS,sBAAsB,OAAwC;AACrE,MAAI,OAAO,MAAM,YAAY,YAAY,MAAM,QAAQ,SAAS,GAAG;AACjE,WAAO,MAAM;AAAA,EACf;AACA,SAAO,KAAK,UAAU,KAAK;AAC7B;AAOA,eAAe,oBACb,MACgB;AAChB,MAAI,KAAK,OAAO,SAAS,QAAQ;AAC/B,WAAO;AAAA,MACL,YAAY,KAAK,OAAO;AAAA,MACxB,UAAU,KAAK,OAAO,YAAY;AAAA,IAAA;AAAA,EAEtC;AACA,QAAM,MAAM,KAAK,OAAO;AACxB,MAAI,IAAI,WAAW,OAAO,GAAG;AAC3B,WAAO;AAAA,MACL,QAAQ;AAAA,MACR,GAAI,KAAK,OAAO,YAAY,EAAE,UAAU,KAAK,OAAO,SAAA;AAAA,IAAS;AAAA,EAEjE;AACA,MAAI,IAAI,WAAW,OAAO,GAAG;AAC3B,UAAM,QAAQ,IAAI,MAAM,iCAAiC;AACzD,QAAI,CAAC,SAAS,CAAC,MAAM,CAAC,GAAG;AACvB,YAAM,IAAI;AAAA,QACR;AAAA,MAAA;AAAA,IAEJ;AACA,WAAO;AAAA,MACL,YAAY,MAAM,CAAC,KAAK;AAAA,MACxB,UAAU,MAAM,CAAC,KAAK,KAAK,OAAO,YAAY;AAAA,IAAA;AAAA,EAElD;AACA,QAAM,WAAW,MAAM,MAAM,GAAG;AAChC,MAAI,CAAC,SAAS,IAAI;AAChB,UAAM,IAAI;AAAA,MACR,gCAAgC,SAAS,MAAM,IAAI,SAAS,UAAU,MAAM,GAAG;AAAA,IAAA;AAAA,EAEnF;AACA,QAAM,OAAO,MAAM,SAAS,KAAA;AAC5B,QAAM,SAAS,MAAM,KAAK,YAAA;AAC1B,SAAO;AAAA,IACL,YAAY,oBAAoB,MAAM;AAAA,IACtC,UAAU,KAAK,OAAO,YAAY,KAAK,QAAQ;AAAA,EAAA;AAEnD;AAuBO,MAAM,2BAEH,iBAOR;AAAA,EACS,OAAO;AAAA,EAEN;AAAA,EAEV,YAAY,QAA2B,OAAe;AACpD,UAAM,CAAA,GAAI,KAAK;AACf,SAAK,SAAS,mBAAmB,MAAM;AAAA,EACzC;AAAA,EAEA,MAAM,eACJ,SAKyB;AACzB,UAAM,EAAE,QAAQ,MAAM,UAAU,cAAc,WAAW;AAEzD,WAAO;AAAA,MACL,kCAAkC,KAAK,IAAI,UAAU,KAAK,KAAK,SAAS,QAAQ,SAAS,aAAa,YAAY,SAAS;AAAA,MAC3H,EAAE,UAAU,KAAK,MAAM,OAAO,KAAK,MAAA;AAAA,IAAM;AAG3C,QAAI;AACF,YAAM,WAAW,mBAAmB,MAAM;AAE1C,UAAI,SAAS,OAAO,SAAS,GAAG;AAC9B,cAAM,IAAI;AAAA,UACR,GAAG,KAAK,IAAI,+DAA+D,KAAK,KAAK;AAAA,QAAA;AAAA,MAEzF;AACA,UAAI,SAAS,OAAO,SAAS,GAAG;AAC9B,cAAM,IAAI;AAAA,UACR,GAAG,KAAK,IAAI,+DAA+D,KAAK,KAAK;AAAA,QAAA;AAAA,MAEzF;AAEA,YAAM,EAAE,OAAO,WAAW,gBAAA,IAAoB,MAAM,KAAK;AAAA,QACvD,SAAS;AAAA,MAAA;AAGX,YAAM,SAA+B;AAAA,QACnC,GAAG;AAAA,QACH,GAAI,SAAS,UAAa,EAAE,aAAa,KAAA;AAAA,QACzC,GAAI,aAAa,UAAa,EAAE,iBAAiB,SAAA;AAAA,QACjD,GAAI,aAAa,EAAE,UAAA;AAAA,QACnB,GAAI,gBAAgB,SAAS,KAAK,EAAE,gBAAA;AAAA,MAAgB;AAGtD,YAAM,YAAY,MAAM,KAAK,OAAO,OAAO,eAAe;AAAA,QACxD,OAAO,KAAK;AAAA,QACZ,QAAQ,SAAS;AAAA,QACjB,GAAI,SAAS,EAAE,MAAA;AAAA,QACf;AAAA,MAAA,CACD;AAED,UAAI,CAAC,UAAU,MAAM;AACnB,cAAM,IAAI;AAAA,UACR;AAAA,QAAA;AAAA,MAEJ;AAEA,aAAO,EAAE,OAAO,UAAU,MAAM,OAAO,KAAK,MAAA;AAAA,IAC9C,SAAS,OAAO;AACd,aAAO,OAAO,GAAG,KAAK,IAAI,yBAAyB;AAAA,QACjD;AAAA,QACA,QAAQ,GAAG,KAAK,IAAI;AAAA,MAAA,CACrB;AACD,YAAM;AAAA,IACR;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAKA,MAAc,gBACZ,OAKC;AACD,QAAI;AACJ,QAAI;AACJ,UAAM,kBAAwD,CAAA;AAE9D,eAAW,QAAQ,OAAO;AACxB,YAAM,OAAO,KAAK,UAAU;AAC5B,cAAQ,MAAA;AAAA,QACN,KAAK,aAAa;AAChB,cAAI,WAAW;AACb,kBAAM,IAAI;AAAA,cACR,GAAG,KAAK,IAAI;AAAA,YAAA;AAAA,UAEhB;AACA,sBAAY,MAAM,oBAAoB,IAAI;AAC1C;AAAA,QACF;AAAA,QACA,KAAK;AAAA,QACL,KAAK,aAAa;AAChB,0BAAgB,KAAK;AAAA,YACnB,OAAO,MAAM,oBAAoB,IAAI;AAAA,YACrC,eAAe,6BAA6B;AAAA,UAAA,CAC7C;AACD;AAAA,QACF;AAAA,QACA,KAAK;AAAA,QACL,KAAK,QAAW;AACd,cAAI,OAAO;AACT,kBAAM,IAAI;AAAA,cACR,GAAG,KAAK,IAAI;AAAA,YAAA;AAAA,UAEhB;AACA,kBAAQ,MAAM,oBAAoB,IAAI;AACtC;AAAA,QACF;AAAA,QACA,KAAK;AAAA,QACL,KAAK;AACH,gBAAM,IAAI;AAAA,YACR,GAAG,KAAK,IAAI,6BAA6B,IAAI;AAAA,UAAA;AAAA,MAC/C;AAAA,IAEN;AAEA,WAAO,EAAE,OAAO,WAAW,gBAAA;AAAA,EAC7B;AAAA,EAEA,MAAM,eAAe,OAA2C;AAC9D,UAAM,YAAY,MAAM,KAAK,aAAa,KAAK;AAE/C,QAAI,CAAC,UAAU,MAAM;AACnB,aAAO,EAAE,OAAO,QAAQ,aAAA;AAAA,IAC1B;AAEA,QAAI,UAAU,OAAO;AACnB,aAAO;AAAA,QACL;AAAA,QACA,QAAQ;AAAA,QACR,OAAO,sBAAsB,UAAU,KAAK;AAAA,MAAA;AAAA,IAEhD;AAKA,UAAM,SAAS,UAAU,UAAU,mBAAmB,CAAA;AACtD,QAAI,OAAO,WAAW,GAAG;AACvB,YAAM,UAAU,UAAU,UAAU;AACpC,aAAO;AAAA,QACL;AAAA,QACA,QAAQ;AAAA,QACR,OAAO,SAAS,SACZ,yCAAyC,QAAQ,KAAK,IAAI,CAAC,KAC3D;AAAA,MAAA;AAAA,IAER;AAEA,WAAO,EAAE,OAAO,QAAQ,YAAA;AAAA,EAC1B;AAAA,EAEA,MAAM,YAAY,OAAwC;AACxD,UAAM,YAAY,MAAM,KAAK,aAAa,KAAK;AAE/C,QAAI,CAAC,UAAU,MAAM;AACnB,YAAM,IAAI;AAAA,QACR,uDAAuD,KAAK;AAAA,MAAA;AAAA,IAEhE;AAEA,QAAI,UAAU,OAAO;AACnB,YAAM,IAAI;AAAA,QACR,4BAA4B,sBAAsB,UAAU,KAAK,CAAC;AAAA,MAAA;AAAA,IAEtE;AAEA,UAAM,MAAM,UAAU,UAAU,kBAAkB,CAAC,GAAG,OAAO;AAC7D,QAAI,CAAC,KAAK;AACR,YAAM,UAAU,UAAU,UAAU;AACpC,YAAM,IAAI;AAAA,QACR,SAAS,SACL,yCAAyC,QAAQ,KAAK,IAAI,CAAC,KAC3D,sDAAsD,KAAK;AAAA,MAAA;AAAA,IAEnE;AAEA,WAAO,EAAE,OAAO,KAAK,IAAA;AAAA,EACvB;AAAA,EAES,qBAEP;AACA,WAAO,8BAA8B,KAAK,KAAK;AAAA,EACjD;AAAA,EAES,aACP,SACoD;AACpD,WAAO,qBAAqB,SAAS,KAAK,mBAAA,CAAoB;AAAA,EAChE;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,MAAc,aAAa,OAAiD;AAC1E,UAAM,YAAY,IAAI,wBAAA;AACtB,cAAU,OAAO;AACjB,WAAO,MAAM,KAAK,OAAO,WAAW,mBAAmB,EAAE,WAAW;AAAA,EACtE;AACF;AAwBO,SAAS,kBACd,OACA,QACA,QAC4B;AAC5B,SAAO,IAAI,mBAAmB,EAAE,QAAQ,GAAG,OAAA,GAAU,KAAK;AAC5D;AAgCO,SAAS,YACd,OACA,QAC4B;AAC5B,QAAM,SAAS,uBAAA;AACf,SAAO,kBAAkB,OAAO,QAAQ,MAAM;AAChD;"}