@tanstack/ai-gemini 0.18.1 → 0.18.3

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

@@ -0,0 +1,92 @@
+import { BaseAudioAdapter } from '@tanstack/ai/adapters';
+import { GEMINI_AUDIO_MODELS } from '../model-meta.js';
+import { AudioGenerationOptions, AudioGenerationResult } from '@tanstack/ai';
+import { GeminiClientConfig } from '../utils.js';
+/**
+ * Provider options for Gemini Lyria music generation.
+ *
+ * Notes on the Lyria 3 surface area:
+ * - `lyria-3-clip-preview` always returns MP3 (30-second clips). It does
+ *   not accept `responseMimeType`, and duration is fixed at 30 seconds —
+ *   the generic `duration` option on `AudioActivityOptions` is ignored.
+ * - `lyria-3-pro-preview` returns MP3 by default. Duration is controlled
+ *   via the natural-language prompt, not a separate SDK field, so the
+ *   generic `duration` option is similarly ignored.
+ * - `negativePrompt` is NOT accepted by `GenerateContentConfig` and has
+ *   therefore been removed from this surface to avoid giving callers a
+ *   silently-dropped knob.
+ *
+ * @see https://ai.google.dev/gemini-api/docs/music-generation
+ */
+export interface GeminiAudioProviderOptions {
+    /**
+     * Seed for deterministic generation.
+     */
+    seed?: number;
+}
+export interface GeminiAudioConfig extends GeminiClientConfig {
+}
+/** Model type for Gemini Lyria audio generation */
+export type GeminiAudioModel = (typeof GEMINI_AUDIO_MODELS)[number];
+/**
+ * Gemini Lyria Music Generation Adapter.
+ *
+ * Tree-shakeable adapter for Google Lyria music generation via the Gemini API.
+ *
+ * Models:
+ * - `lyria-3-pro-preview` — flagship model, full-length songs with verses,
+ *   choruses, and bridges. Outputs MP3 or WAV at 48 kHz stereo.
+ * - `lyria-3-clip-preview` — 30-second clips in MP3.
+ *
+ * @see https://ai.google.dev/gemini-api/docs/music-generation
+ *
+ * @example
+ * ```typescript
+ * const adapter = geminiAudio('lyria-3-pro-preview')
+ * const result = await generateAudio({
+ *   adapter,
+ *   prompt: 'An upbeat jazz track with saxophone and drums',
+ * })
+ * ```
+ */
+export declare class GeminiAudioAdapter<TModel extends GeminiAudioModel> extends BaseAudioAdapter<TModel, GeminiAudioProviderOptions> {
+    readonly name: "gemini";
+    private readonly client;
+    constructor(config: GeminiAudioConfig, model: TModel);
+    generateAudio(options: AudioGenerationOptions<GeminiAudioProviderOptions>): Promise<AudioGenerationResult>;
+}
+/**
+ * Creates a Gemini Lyria audio adapter with an explicit API key.
+ *
+ * @param model - The Lyria model name (e.g., 'lyria-3-pro-preview')
+ * @param apiKey - Your Google API key
+ * @param config - Optional additional configuration
+ *
+ * @example
+ * ```typescript
+ * const adapter = createGeminiAudio('lyria-3-pro-preview', 'your-api-key')
+ * const result = await generateAudio({
+ *   adapter,
+ *   prompt: 'Ambient electronic music with soft pads',
+ * })
+ * ```
+ */
+export declare function createGeminiAudio<TModel extends GeminiAudioModel>(model: TModel, apiKey: string, config?: Omit<GeminiAudioConfig, 'apiKey'>): GeminiAudioAdapter<TModel>;
+/**
+ * Creates a Gemini Lyria audio adapter with automatic API key detection.
+ *
+ * Looks for `GOOGLE_API_KEY` or `GEMINI_API_KEY` in the environment.
+ *
+ * @param model - The Lyria model name (e.g., 'lyria-3-pro-preview')
+ * @param config - Optional configuration (excluding apiKey)
+ *
+ * @example
+ * ```typescript
+ * const adapter = geminiAudio('lyria-3-pro-preview')
+ * const result = await generateAudio({
+ *   adapter,
+ *   prompt: 'An orchestral piece with strings and brass',
+ * })
+ * ```
+ */
+export declare function geminiAudio<TModel extends GeminiAudioModel>(model: TModel, config?: Omit<GeminiAudioConfig, 'apiKey'>): GeminiAudioAdapter<TModel>;

package/dist/esm/adapters/audio.js

@@ -0,0 +1,66 @@
+import { BaseAudioAdapter } from "@tanstack/ai/adapters";
+import { createGeminiClient, generateId, getGeminiApiKeyFromEnv } from "../utils/client.js";
+import { buildGeminiUsage } from "../usage.js";
+class GeminiAudioAdapter extends BaseAudioAdapter {
+  name = "gemini";
+  client;
+  constructor(config, model) {
+    super(model, config);
+    this.client = createGeminiClient(config);
+  }
+  async generateAudio(options) {
+    const { model, prompt, modelOptions, logger } = options;
+    logger.request(`activity=generateAudio provider=gemini model=${model}`, {
+      provider: "gemini",
+      model
+    });
+    try {
+      const response = await this.client.models.generateContent({
+        model,
+        contents: [{ role: "user", parts: [{ text: prompt }] }],
+        config: {
+          responseModalities: ["AUDIO", "TEXT"],
+          ...modelOptions?.seed != null ? { seed: modelOptions.seed } : {}
+        }
+      });
+      const parts = response.candidates?.[0]?.content?.parts ?? [];
+      const audioPart = parts.find(
+        (part) => part.inlineData?.mimeType?.startsWith("audio/")
+      );
+      if (!audioPart?.inlineData?.data) {
+        throw new Error("No audio data in Gemini Lyria response");
+      }
+      const contentType = audioPart.inlineData.mimeType;
+      return {
+        id: generateId(this.name),
+        model,
+        audio: {
+          b64Json: audioPart.inlineData.data,
+          ...contentType !== void 0 && { contentType }
+        },
+        // Surface token usage (with per-modality breakdown) when Gemini reports
+        // it. Spread conditionally for exactOptionalPropertyTypes.
+        ...response.usageMetadata ? { usage: buildGeminiUsage(response.usageMetadata) } : {}
+      };
+    } catch (error) {
+      logger.errors("gemini.generateAudio fatal", {
+        error,
+        source: "gemini.generateAudio"
+      });
+      throw error;
+    }
+  }
+}
+function createGeminiAudio(model, apiKey, config) {
+  return new GeminiAudioAdapter({ ...config, apiKey }, model);
+}
+function geminiAudio(model, config) {
+  const apiKey = getGeminiApiKeyFromEnv();
+  return createGeminiAudio(model, apiKey, config);
+}
+export {
+  GeminiAudioAdapter,
+  createGeminiAudio,
+  geminiAudio
+};
+//# sourceMappingURL=audio.js.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"audio.js","sources":["../../../src/adapters/audio.ts"],"sourcesContent":["import { BaseAudioAdapter } from '@tanstack/ai/adapters'\nimport {\n  createGeminiClient,\n  generateId,\n  getGeminiApiKeyFromEnv,\n} from '../utils'\nimport { buildGeminiUsage } from '../usage'\nimport type { GEMINI_AUDIO_MODELS } from '../model-meta'\nimport type {\n  AudioGenerationOptions,\n  AudioGenerationResult,\n} from '@tanstack/ai'\nimport type { GoogleGenAI } from '@google/genai'\nimport type { GeminiClientConfig } from '../utils'\n\n/**\n * Provider options for Gemini Lyria music generation.\n *\n * Notes on the Lyria 3 surface area:\n * - `lyria-3-clip-preview` always returns MP3 (30-second clips). It does\n *   not accept `responseMimeType`, and duration is fixed at 30 seconds —\n *   the generic `duration` option on `AudioActivityOptions` is ignored.\n * - `lyria-3-pro-preview` returns MP3 by default. Duration is controlled\n *   via the natural-language prompt, not a separate SDK field, so the\n *   generic `duration` option is similarly ignored.\n * - `negativePrompt` is NOT accepted by `GenerateContentConfig` and has\n *   therefore been removed from this surface to avoid giving callers a\n *   silently-dropped knob.\n *\n * @see https://ai.google.dev/gemini-api/docs/music-generation\n */\nexport interface GeminiAudioProviderOptions {\n  /**\n   * Seed for deterministic generation.\n   */\n  seed?: number\n}\n\nexport interface GeminiAudioConfig extends GeminiClientConfig {}\n\n/** Model type for Gemini Lyria audio generation */\nexport type GeminiAudioModel = (typeof GEMINI_AUDIO_MODELS)[number]\n\n/**\n * Gemini Lyria Music Generation Adapter.\n *\n * Tree-shakeable adapter for Google Lyria music generation via the Gemini API.\n *\n * Models:\n * - `lyria-3-pro-preview` — flagship model, full-length songs with verses,\n *   choruses, and bridges. Outputs MP3 or WAV at 48 kHz stereo.\n * - `lyria-3-clip-preview` — 30-second clips in MP3.\n *\n * @see https://ai.google.dev/gemini-api/docs/music-generation\n *\n * @example\n * ```typescript\n * const adapter = geminiAudio('lyria-3-pro-preview')\n * const result = await generateAudio({\n *   adapter,\n *   prompt: 'An upbeat jazz track with saxophone and drums',\n * })\n * ```\n */\nexport class GeminiAudioAdapter<\n  TModel extends GeminiAudioModel,\n> extends BaseAudioAdapter<TModel, GeminiAudioProviderOptions> {\n  readonly name = 'gemini' as const\n\n  private readonly client: GoogleGenAI\n\n  constructor(config: GeminiAudioConfig, model: TModel) {\n    super(model, config)\n    this.client = createGeminiClient(config)\n  }\n\n  async generateAudio(\n    options: AudioGenerationOptions<GeminiAudioProviderOptions>,\n  ): Promise<AudioGenerationResult> {\n    const { model, prompt, modelOptions, logger } = options\n\n    logger.request(`activity=generateAudio provider=gemini model=${model}`, {\n      provider: 'gemini',\n      model,\n    })\n\n    try {\n      // FIXME (SDK audit): Lyria 3 music generation may not belong on\n      // generateContent at all — @google/genai exposes a `LiveMusicSession`\n      // (`ai.live.music.connect`) with a `musicGenerationConfig` object.\n      // `seed` is valid on GenerateContentConfig, and Lyria always returns\n      // MP3 today, so we don't forward `responseMimeType` either.\n      // The runtime test `emits only GenerateContentConfig-valid fields`\n      // asserts the config shape so a later SDK audit can catch regressions.\n      const response = await this.client.models.generateContent({\n        model,\n        contents: [{ role: 'user', parts: [{ text: prompt }] }],\n        config: {\n          responseModalities: ['AUDIO', 'TEXT'],\n          ...(modelOptions?.seed != null ? { seed: modelOptions.seed } : {}),\n        },\n      })\n\n      const parts = response.candidates?.[0]?.content?.parts ?? []\n      const audioPart = parts.find((part: any) =>\n        part.inlineData?.mimeType?.startsWith('audio/'),\n      )\n\n      if (!audioPart?.inlineData?.data) {\n        throw new Error('No audio data in Gemini Lyria response')\n      }\n\n      // audioPart was selected because mimeType.startsWith('audio/') was\n      // truthy, so the mime type is guaranteed to be a string here. Trust the\n      // value Gemini returned rather than inventing a non-standard\n      // `audio/mp3` fallback (IANA is `audio/mpeg`).\n      const contentType = audioPart.inlineData.mimeType\n\n      return {\n        id: generateId(this.name),\n        model,\n        audio: {\n          b64Json: audioPart.inlineData.data,\n          ...(contentType !== undefined && { contentType }),\n        },\n        // Surface token usage (with per-modality breakdown) when Gemini reports\n        // it. Spread conditionally for exactOptionalPropertyTypes.\n        ...(response.usageMetadata\n          ? { usage: buildGeminiUsage(response.usageMetadata) }\n          : {}),\n      }\n    } catch (error) {\n      logger.errors('gemini.generateAudio fatal', {\n        error,\n        source: 'gemini.generateAudio',\n      })\n      throw error\n    }\n  }\n}\n\n/**\n * Creates a Gemini Lyria audio adapter with an explicit API key.\n *\n * @param model - The Lyria model name (e.g., 'lyria-3-pro-preview')\n * @param apiKey - Your Google API key\n * @param config - Optional additional configuration\n *\n * @example\n * ```typescript\n * const adapter = createGeminiAudio('lyria-3-pro-preview', 'your-api-key')\n * const result = await generateAudio({\n *   adapter,\n *   prompt: 'Ambient electronic music with soft pads',\n * })\n * ```\n */\nexport function createGeminiAudio<TModel extends GeminiAudioModel>(\n  model: TModel,\n  apiKey: string,\n  config?: Omit<GeminiAudioConfig, 'apiKey'>,\n): GeminiAudioAdapter<TModel> {\n  // Put apiKey LAST so caller-supplied config can't silently override the\n  // explicit argument.\n  return new GeminiAudioAdapter({ ...config, apiKey }, model)\n}\n\n/**\n * Creates a Gemini Lyria audio adapter with automatic API key detection.\n *\n * Looks for `GOOGLE_API_KEY` or `GEMINI_API_KEY` in the environment.\n *\n * @param model - The Lyria model name (e.g., 'lyria-3-pro-preview')\n * @param config - Optional configuration (excluding apiKey)\n *\n * @example\n * ```typescript\n * const adapter = geminiAudio('lyria-3-pro-preview')\n * const result = await generateAudio({\n *   adapter,\n *   prompt: 'An orchestral piece with strings and brass',\n * })\n * ```\n */\nexport function geminiAudio<TModel extends GeminiAudioModel>(\n  model: TModel,\n  config?: Omit<GeminiAudioConfig, 'apiKey'>,\n): GeminiAudioAdapter<TModel> {\n  const apiKey = getGeminiApiKeyFromEnv()\n  return createGeminiAudio(model, apiKey, config)\n}\n"],"names":[],"mappings":";;;AAgEO,MAAM,2BAEH,iBAAqD;AAAA,EACpD,OAAO;AAAA,EAEC;AAAA,EAEjB,YAAY,QAA2B,OAAe;AACpD,UAAM,OAAO,MAAM;AACnB,SAAK,SAAS,mBAAmB,MAAM;AAAA,EACzC;AAAA,EAEA,MAAM,cACJ,SACgC;AAChC,UAAM,EAAE,OAAO,QAAQ,cAAc,WAAW;AAEhD,WAAO,QAAQ,gDAAgD,KAAK,IAAI;AAAA,MACtE,UAAU;AAAA,MACV;AAAA,IAAA,CACD;AAED,QAAI;AAQF,YAAM,WAAW,MAAM,KAAK,OAAO,OAAO,gBAAgB;AAAA,QACxD;AAAA,QACA,UAAU,CAAC,EAAE,MAAM,QAAQ,OAAO,CAAC,EAAE,MAAM,OAAA,CAAQ,GAAG;AAAA,QACtD,QAAQ;AAAA,UACN,oBAAoB,CAAC,SAAS,MAAM;AAAA,UACpC,GAAI,cAAc,QAAQ,OAAO,EAAE,MAAM,aAAa,SAAS,CAAA;AAAA,QAAC;AAAA,MAClE,CACD;AAED,YAAM,QAAQ,SAAS,aAAa,CAAC,GAAG,SAAS,SAAS,CAAA;AAC1D,YAAM,YAAY,MAAM;AAAA,QAAK,CAAC,SAC5B,KAAK,YAAY,UAAU,WAAW,QAAQ;AAAA,MAAA;AAGhD,UAAI,CAAC,WAAW,YAAY,MAAM;AAChC,cAAM,IAAI,MAAM,wCAAwC;AAAA,MAC1D;AAMA,YAAM,cAAc,UAAU,WAAW;AAEzC,aAAO;AAAA,QACL,IAAI,WAAW,KAAK,IAAI;AAAA,QACxB;AAAA,QACA,OAAO;AAAA,UACL,SAAS,UAAU,WAAW;AAAA,UAC9B,GAAI,gBAAgB,UAAa,EAAE,YAAA;AAAA,QAAY;AAAA;AAAA;AAAA,QAIjD,GAAI,SAAS,gBACT,EAAE,OAAO,iBAAiB,SAAS,aAAa,MAChD,CAAA;AAAA,MAAC;AAAA,IAET,SAAS,OAAO;AACd,aAAO,OAAO,8BAA8B;AAAA,QAC1C;AAAA,QACA,QAAQ;AAAA,MAAA,CACT;AACD,YAAM;AAAA,IACR;AAAA,EACF;AACF;AAkBO,SAAS,kBACd,OACA,QACA,QAC4B;AAG5B,SAAO,IAAI,mBAAmB,EAAE,GAAG,QAAQ,OAAA,GAAU,KAAK;AAC5D;AAmBO,SAAS,YACd,OACA,QAC4B;AAC5B,QAAM,SAAS,uBAAA;AACf,SAAO,kBAAkB,OAAO,QAAQ,MAAM;AAChD;"}

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

@@ -0,0 +1,101 @@
+import { BaseImageAdapter } from '@tanstack/ai/adapters';
+import { GEMINI_IMAGE_MODELS } from '../model-meta.js';
+import { GeminiImageModelInputModalitiesByName, GeminiImageModelProviderOptionsByName, GeminiImageModelSizeByName, GeminiImageProviderOptions } from '../image/image-provider-options.js';
+import { ImageGenerationOptions, ImageGenerationResult } from '@tanstack/ai';
+import { GeminiClientConfig } from '../utils.js';
+/**
+ * Configuration for Gemini image adapter
+ */
+export interface GeminiImageConfig extends GeminiClientConfig {
+}
+/** Model type for Gemini Image */
+export type GeminiImageModel = (typeof GEMINI_IMAGE_MODELS)[number];
+/**
+ * Gemini Image Generation Adapter
+ *
+ * Tree-shakeable adapter for Gemini image generation functionality.
+ * Supports Imagen 3/4 models (via generateImages API) and Gemini native
+ * image models like Nano Banana 2 (via generateContent API).
+ *
+ * Features:
+ * - Aspect ratio-based image sizing
+ * - Person generation controls
+ * - Safety filtering
+ * - Watermark options
+ * - Extended resolution tiers (Nano Banana 2)
+ */
+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;
+}
+/**
+ * Creates a Gemini image adapter with explicit API key.
+ * Type resolution happens here at the call site.
+ *
+ * @param model - The model name (e.g., 'imagen-3.0-generate-002')
+ * @param apiKey - Your Google API key
+ * @param config - Optional additional configuration
+ * @returns Configured Gemini image adapter instance with resolved types
+ *
+ * @example
+ * ```typescript
+ * const adapter = createGeminiImage('imagen-3.0-generate-002', "your-api-key");
+ *
+ * const result = await generateImage({
+ *   adapter,
+ *   prompt: 'A cute baby sea otter'
+ * });
+ * ```
+ */
+export declare function createGeminiImage<TModel extends GeminiImageModel>(model: TModel, apiKey: string, config?: Omit<GeminiImageConfig, 'apiKey'>): GeminiImageAdapter<TModel>;
+/**
+ * Creates a Gemini image 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)
+ *
+ * @param model - The model name (e.g., 'imagen-4.0-generate-001')
+ * @param config - Optional configuration (excluding apiKey which is auto-detected)
+ * @returns Configured Gemini image 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 = geminiImage('imagen-4.0-generate-001');
+ *
+ * const result = await generateImage({
+ *   adapter,
+ *   prompt: 'A beautiful sunset over mountains'
+ * });
+ * ```
+ */
+export declare function geminiImage<TModel extends GeminiImageModel>(model: TModel, config?: Omit<GeminiImageConfig, 'apiKey'>): GeminiImageAdapter<TModel>;

package/dist/esm/adapters/image.js

@@ -0,0 +1,253 @@
+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";
+class GeminiImageAdapter extends BaseImageAdapter {
+  kind = "image";
+  name = "gemini";
+  client;
+  constructor(config, model) {
+    super(model, config);
+    this.client = createGeminiClient(config);
+  }
+  async generateImages(options) {
+    const { model, logger } = options;
+    logger.request(
+      `activity=generateImage provider=gemini model=${this.model}`,
+      {
+        provider: "gemini",
+        model: this.model
+      }
+    );
+    try {
+      const resolved = resolveMediaPrompt(options.prompt);
+      if (resolved.images.length === 0) {
+        validatePrompt({ prompt: resolved.text, model });
+      }
+      if (resolved.videos.length > 0) {
+        throw new Error(
+          `${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: ${model}).`
+        );
+      }
+      if (this.isGeminiImageModel(model)) {
+        return await this.generateWithGeminiApi(options, resolved);
+      }
+      if (resolved.images.length > 0) {
+        throw new Error(
+          `${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: resolved.text,
+        config
+      });
+      return this.transformImagenResponse(model, response);
+    } catch (error) {
+      logger.errors("gemini.generateImage fatal", {
+        error,
+        source: "gemini.generateImage"
+      });
+      throw error;
+    }
+  }
+  isGeminiImageModel(model) {
+    return model.startsWith("gemini-");
+  }
+  async generateWithGeminiApi(options, resolved) {
+    const { model, size, numberOfImages, modelOptions } = options;
+    const parsedSize = size ? parseNativeImageSize(size) : void 0;
+    const nativeConfig = {};
+    if (modelOptions?.seed !== void 0) {
+      nativeConfig.seed = modelOptions.seed;
+    }
+    const config = {
+      ...nativeConfig,
+      // Include TEXT so the model can interleave descriptions between images.
+      // IMPORTANT: responseModalities is a protected default — set it AFTER
+      // nativeConfig so nothing can silently disable image output.
+      responseModalities: ["TEXT", "IMAGE"],
+      ...parsedSize && {
+        imageConfig: {
+          ...parsedSize.aspectRatio && {
+            aspectRatio: parsedSize.aspectRatio
+          },
+          ...parsedSize.resolution && {
+            imageSize: parsedSize.resolution
+          }
+        }
+      }
+    };
+    const contents = await this.buildContents(resolved, numberOfImages);
+    const response = await this.client.models.generateContent({
+      model,
+      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 = [];
+    const parts = response.candidates?.[0]?.content?.parts ?? [];
+    for (const part of parts) {
+      if (part.inlineData?.data && typeof part.inlineData.data === "string" && part.inlineData.data.length > 0) {
+        images.push({ b64Json: part.inlineData.data });
+      } else if (typeof part.text === "string" && part.text.length > 0) {
+        textParts.push(part.text);
+      }
+    }
+    if (images.length === 0) {
+      const reason = textParts.length > 0 ? `: ${textParts.join(" ").trim()}` : " (no inline image or text parts were returned).";
+      throw new Error(`Gemini ${model} returned no images${reason}`);
+    }
+    return {
+      id: generateId(this.name),
+      model,
+      images,
+      // Surface token usage (with per-modality breakdown) when the model
+      // reports it (e.g. Nano Banana via generateContent). Conditionally spread
+      // to satisfy exactOptionalPropertyTypes — only include usage when
+      // present. See #330.
+      ...response.usageMetadata ? { usage: buildGeminiUsage(response.usageMetadata) } : {}
+    };
+  }
+  buildImagenConfig(options) {
+    const { size, numberOfImages, modelOptions } = options;
+    const sizeAspectRatio = size ? sizeToAspectRatio(size) : void 0;
+    return {
+      numberOfImages: numberOfImages ?? 1,
+      // Map size to aspect ratio if provided (modelOptions.aspectRatio will override)
+      ...sizeAspectRatio !== void 0 && { aspectRatio: sizeAspectRatio },
+      ...modelOptions
+    };
+  }
+  transformImagenResponse(model, response) {
+    const entries = response.generatedImages ?? [];
+    const images = [];
+    const filterReasons = [];
+    for (const item of entries) {
+      const b64Json = item.image?.imageBytes;
+      if (b64Json) {
+        images.push({
+          b64Json,
+          ...item.enhancedPrompt !== void 0 && {
+            revisedPrompt: item.enhancedPrompt
+          }
+        });
+        continue;
+      }
+      const reason = item.raiFilteredReason;
+      if (reason) {
+        filterReasons.push(reason);
+      }
+    }
+    if (entries.length > 0 && images.length === 0) {
+      const joined = filterReasons.length > 0 ? filterReasons.join("; ") : "";
+      throw new Error(
+        `Imagen ${model} returned no images: all ${entries.length} generated image(s) were filtered by Responsible-AI${joined ? ` (${joined})` : ""}.`
+      );
+    }
+    if (filterReasons.length > 0 && typeof console !== "undefined") {
+      console.warn(
+        `[gemini-image] ${filterReasons.length} of ${entries.length} images from ${model} were filtered by Responsible-AI: ${filterReasons.join("; ")}`
+      );
+    }
+    return {
+      id: generateId(this.name),
+      model,
+      images
+    };
+  }
+}
+function createGeminiImage(model, apiKey, config) {
+  return new GeminiImageAdapter({ apiKey, ...config }, model);
+}
+function geminiImage(model, config) {
+  const apiKey = getGeminiApiKeyFromEnv();
+  return createGeminiImage(model, apiKey, config);
+}
+export {
+  GeminiImageAdapter,
+  createGeminiImage,
+  geminiImage
+};
+//# sourceMappingURL=image.js.map

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

@@ -0,0 +1 @@
+{"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/summarize.d.ts

@@ -0,0 +1,33 @@
+import { ChatStreamSummarizeAdapter, InferTextProviderOptions } from '@tanstack/ai/adapters';
+import { GeminiTextAdapter } from './text.js';
+import { GEMINI_MODELS } from '../model-meta.js';
+import { GeminiClientConfig } from '../utils.js';
+/**
+ * Configuration for Gemini summarize adapter
+ */
+export interface GeminiSummarizeConfig extends GeminiClientConfig {
+}
+export type GeminiSummarizeModel = (typeof GEMINI_MODELS)[number];
+/**
+ * Creates a Gemini summarize adapter with explicit API key and model.
+ *
+ * Note: keeps the historical (apiKey, model, config) argument order to
+ * avoid breaking existing callers.
+ *
+ * @example
+ * ```typescript
+ * const adapter = createGeminiSummarize('AIza...', 'gemini-2.5-flash');
+ * ```
+ */
+export declare function createGeminiSummarize<TModel extends GeminiSummarizeModel>(apiKey: string, model: TModel, config?: Omit<GeminiSummarizeConfig, 'apiKey'>): ChatStreamSummarizeAdapter<TModel, InferTextProviderOptions<GeminiTextAdapter<TModel>>>;
+/**
+ * Creates a Gemini summarize adapter with API key from `GOOGLE_API_KEY` /
+ * `GEMINI_API_KEY` environment variables.
+ *
+ * @example
+ * ```typescript
+ * const adapter = geminiSummarize('gemini-2.5-flash');
+ * await summarize({ adapter, text: 'Long article text...' });
+ * ```
+ */
+export declare function geminiSummarize<TModel extends GeminiSummarizeModel>(model: TModel, config?: Omit<GeminiSummarizeConfig, 'apiKey'>): ChatStreamSummarizeAdapter<TModel, InferTextProviderOptions<GeminiTextAdapter<TModel>>>;

package/dist/esm/adapters/summarize.js

@@ -0,0 +1,18 @@
+import { ChatStreamSummarizeAdapter } from "@tanstack/ai/adapters";
+import { getGeminiApiKeyFromEnv } from "../utils/client.js";
+import { GeminiTextAdapter } from "./text.js";
+function createGeminiSummarize(apiKey, model, config) {
+  return new ChatStreamSummarizeAdapter(
+    new GeminiTextAdapter({ ...config, apiKey }, model),
+    model,
+    "gemini"
+  );
+}
+function geminiSummarize(model, config) {
+  return createGeminiSummarize(getGeminiApiKeyFromEnv(), model, config);
+}
+export {
+  createGeminiSummarize,
+  geminiSummarize
+};
+//# sourceMappingURL=summarize.js.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"summarize.js","sources":["../../../src/adapters/summarize.ts"],"sourcesContent":["import { ChatStreamSummarizeAdapter } from '@tanstack/ai/adapters'\nimport { getGeminiApiKeyFromEnv } from '../utils'\nimport { GeminiTextAdapter } from './text'\nimport type { InferTextProviderOptions } from '@tanstack/ai/adapters'\nimport type { GEMINI_MODELS } from '../model-meta'\nimport type { GeminiClientConfig } from '../utils'\n\n/**\n * Configuration for Gemini summarize adapter\n */\nexport interface GeminiSummarizeConfig extends GeminiClientConfig {}\n\nexport type GeminiSummarizeModel = (typeof GEMINI_MODELS)[number]\n\n/**\n * Creates a Gemini summarize adapter with explicit API key and model.\n *\n * Note: keeps the historical (apiKey, model, config) argument order to\n * avoid breaking existing callers.\n *\n * @example\n * ```typescript\n * const adapter = createGeminiSummarize('AIza...', 'gemini-2.5-flash');\n * ```\n */\nexport function createGeminiSummarize<TModel extends GeminiSummarizeModel>(\n  apiKey: string,\n  model: TModel,\n  config?: Omit<GeminiSummarizeConfig, 'apiKey'>,\n): ChatStreamSummarizeAdapter<\n  TModel,\n  InferTextProviderOptions<GeminiTextAdapter<TModel>>\n> {\n  return new ChatStreamSummarizeAdapter(\n    new GeminiTextAdapter({ ...config, apiKey }, model),\n    model,\n    'gemini',\n  )\n}\n\n/**\n * Creates a Gemini summarize adapter with API key from `GOOGLE_API_KEY` /\n * `GEMINI_API_KEY` environment variables.\n *\n * @example\n * ```typescript\n * const adapter = geminiSummarize('gemini-2.5-flash');\n * await summarize({ adapter, text: 'Long article text...' });\n * ```\n */\nexport function geminiSummarize<TModel extends GeminiSummarizeModel>(\n  model: TModel,\n  config?: Omit<GeminiSummarizeConfig, 'apiKey'>,\n): ChatStreamSummarizeAdapter<\n  TModel,\n  InferTextProviderOptions<GeminiTextAdapter<TModel>>\n> {\n  return createGeminiSummarize(getGeminiApiKeyFromEnv(), model, config)\n}\n"],"names":[],"mappings":";;;AAyBO,SAAS,sBACd,QACA,OACA,QAIA;AACA,SAAO,IAAI;AAAA,IACT,IAAI,kBAAkB,EAAE,GAAG,QAAQ,OAAA,GAAU,KAAK;AAAA,IAClD;AAAA,IACA;AAAA,EAAA;AAEJ;AAYO,SAAS,gBACd,OACA,QAIA;AACA,SAAO,sBAAsB,0BAA0B,OAAO,MAAM;AACtE;"}