@tanstack/ai-gemini 0.18.1 → 0.18.3

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

@@ -0,0 +1 @@
+{"version":3,"file":"text.js","sources":["../../../src/adapters/text.ts"],"sourcesContent":["import { FinishReason } from '@google/genai'\nimport { EventType, normalizeSystemPrompts } from '@tanstack/ai'\nimport { toRunErrorRawEvent } from '@tanstack/ai/adapter-internals'\nimport { BaseTextAdapter } from '@tanstack/ai/adapters'\nimport { convertToolsToProviderFormat } from '../tools/tool-converter'\nimport { buildGeminiUsage } from '../usage'\nimport {\n  createGeminiClient,\n  generateId,\n  getGeminiApiKeyFromEnv,\n} from '../utils'\nimport { GEMINI_COMBINED_TOOLS_AND_SCHEMA_MODELS } from '../model-meta'\nimport type {\n  GEMINI_MODELS,\n  GeminiChatModelProviderOptionsByName,\n  GeminiChatModelToolCapabilitiesByName,\n  GeminiModelInputModalitiesByName,\n} from '../model-meta'\nimport type {\n  StructuredOutputOptions,\n  StructuredOutputResult,\n} from '@tanstack/ai/adapters'\nimport type { InternalLogger } from '@tanstack/ai/adapter-internals'\nimport type {\n  Content,\n  GenerateContentParameters,\n  GenerateContentResponse,\n  GoogleGenAI,\n  Part,\n  ThinkingLevel,\n} from '@google/genai'\nimport type {\n  ContentPart,\n  Modality,\n  ModelMessage,\n  StreamChunk,\n  TextOptions,\n} from '@tanstack/ai'\nimport type { ExternalTextProviderOptions } from '../text/text-provider-options'\nimport type {\n  GeminiMessageMetadataByModality,\n  GeminiToolCallMetadata,\n} from '../message-types'\nimport type { GeminiClientConfig } from '../utils'\n\n/**\n * Configuration for Gemini text adapter\n */\nexport interface GeminiTextConfig extends GeminiClientConfig {}\n\n/**\n * Gemini-specific provider options for text/chat\n */\nexport type GeminiTextProviderOptions = ExternalTextProviderOptions\n\n// ===========================\n// Type Resolution Helpers\n// ===========================\n\n/**\n * Resolve provider options for a specific model.\n * If the model has explicit options in the map, use those; otherwise use base options.\n */\ntype ResolveProviderOptions<TModel extends string> =\n  TModel extends keyof GeminiChatModelProviderOptionsByName\n    ? GeminiChatModelProviderOptionsByName[TModel]\n    : GeminiTextProviderOptions\n\n/**\n * Resolve input modalities for a specific model.\n * If the model has explicit modalities in the map, use those; otherwise use all modalities.\n */\ntype ResolveInputModalities<TModel extends string> =\n  TModel extends keyof GeminiModelInputModalitiesByName\n    ? GeminiModelInputModalitiesByName[TModel]\n    : readonly ['text', 'image', 'audio', 'video', 'document']\n\n/**\n * Resolve tool capabilities for a specific model.\n * If the model has explicit tools in the map, use those; otherwise use empty tuple.\n */\ntype ResolveToolCapabilities<TModel extends string> =\n  TModel extends keyof GeminiChatModelToolCapabilitiesByName\n    ? NonNullable<GeminiChatModelToolCapabilitiesByName[TModel]>\n    : readonly []\n\n// ===========================\n// Adapter Implementation\n// ===========================\n\n/**\n * Gemini Text (Chat) Adapter\n *\n * Tree-shakeable adapter for Gemini chat/text completion functionality.\n * Import only what you need for smaller bundle sizes.\n */\nexport class GeminiTextAdapter<\n  TModel extends (typeof GEMINI_MODELS)[number],\n  TProviderOptions extends Record<string, any> = ResolveProviderOptions<TModel>,\n  TInputModalities extends ReadonlyArray<Modality> =\n    ResolveInputModalities<TModel>,\n  TToolCapabilities extends ReadonlyArray<string> =\n    ResolveToolCapabilities<TModel>,\n> extends BaseTextAdapter<\n  TModel,\n  TProviderOptions,\n  TInputModalities,\n  GeminiMessageMetadataByModality,\n  TToolCapabilities,\n  GeminiToolCallMetadata\n> {\n  override readonly kind = 'text' as const\n  readonly name = 'gemini' as const\n\n  private readonly client: GoogleGenAI\n\n  constructor(config: GeminiTextConfig, model: TModel) {\n    super({}, model)\n    this.client = createGeminiClient(config)\n  }\n\n  async *chatStream(\n    options: TextOptions<GeminiTextProviderOptions>,\n  ): AsyncIterable<StreamChunk> {\n    const mappedOptions = this.mapCommonOptionsToGemini(options)\n    const { logger } = options\n\n    try {\n      logger.request(\n        `activity=chat provider=gemini model=${this.model} messages=${options.messages.length} tools=${options.tools?.length ?? 0} stream=true`,\n        { provider: 'gemini', model: this.model },\n      )\n      const result =\n        await this.client.models.generateContentStream(mappedOptions)\n\n      yield* this.processStreamChunks(result, options, logger)\n    } catch (error) {\n      const rawEvent = toRunErrorRawEvent(error)\n      logger.errors('gemini.chatStream fatal', {\n        error,\n        source: 'gemini.chatStream',\n      })\n      yield {\n        type: EventType.RUN_ERROR,\n        model: options.model,\n        timestamp: Date.now(),\n        message:\n          error instanceof Error\n            ? error.message\n            : 'An unknown error occurred during the chat stream.',\n        // Forward the provider's structured error body when present (see\n        // toRunErrorRawEvent); omitted otherwise.\n        ...(rawEvent !== undefined && { rawEvent }),\n        error: {\n          message:\n            error instanceof Error\n              ? error.message\n              : 'An unknown error occurred during the chat stream.',\n        },\n      }\n    }\n  }\n\n  /**\n   * Generate structured output using Gemini's native JSON response format.\n   * Uses responseMimeType: 'application/json' and responseSchema for structured output.\n   * The outputSchema is already JSON Schema (converted in the ai layer).\n   */\n  async structuredOutput(\n    options: StructuredOutputOptions<GeminiTextProviderOptions>,\n  ): Promise<StructuredOutputResult<unknown>> {\n    const { chatOptions, outputSchema } = options\n    const { logger } = chatOptions\n\n    const mappedOptions = this.mapCommonOptionsToGemini(chatOptions)\n\n    try {\n      logger.request(\n        `activity=chat provider=gemini model=${this.model} messages=${chatOptions.messages.length} tools=${chatOptions.tools?.length ?? 0} stream=false`,\n        { provider: 'gemini', model: this.model },\n      )\n      // Add structured output configuration\n      const result = await this.client.models.generateContent({\n        ...mappedOptions,\n        config: {\n          ...mappedOptions.config,\n          responseMimeType: 'application/json',\n          responseSchema: outputSchema,\n        },\n      })\n\n      // Extract text content from the response\n      const rawText = this.extractTextFromResponse(result)\n\n      // Parse the JSON response\n      let parsed: unknown\n      try {\n        parsed = JSON.parse(rawText)\n      } catch {\n        throw new Error(\n          `Failed to parse structured output as JSON. Content: ${rawText.slice(0, 200)}${rawText.length > 200 ? '...' : ''}`,\n        )\n      }\n\n      return {\n        data: parsed,\n        rawText,\n        usage: result.usageMetadata\n          ? buildGeminiUsage(result.usageMetadata)\n          : undefined,\n      }\n    } catch (error) {\n      logger.errors('gemini.structuredOutput fatal', {\n        error,\n        source: 'gemini.structuredOutput',\n      })\n      throw new Error(\n        error instanceof Error\n          ? error.message\n          : 'An unknown error occurred during structured output generation.',\n      )\n    }\n  }\n\n  /**\n   * Extract text content from a non-streaming response\n   */\n  private extractTextFromResponse(response: GenerateContentResponse): string {\n    let textContent = ''\n\n    if (response.candidates?.[0]?.content?.parts) {\n      for (const part of response.candidates[0].content.parts) {\n        if (part.text) {\n          textContent += part.text\n        }\n      }\n    }\n\n    return textContent\n  }\n\n  private async *processStreamChunks(\n    result: AsyncGenerator<GenerateContentResponse, unknown, unknown>,\n    options: TextOptions<GeminiTextProviderOptions>,\n    logger: InternalLogger,\n  ): AsyncIterable<StreamChunk> {\n    const model = options.model\n    let accumulatedContent = ''\n    let accumulatedThinking = ''\n    const toolCallMap = new Map<\n      string,\n      {\n        name: string\n        args: string\n        index: number\n        started: boolean\n        thoughtSignature?: string\n      }\n    >()\n    let nextToolIndex = 0\n\n    // AG-UI lifecycle tracking\n    const runId = options.runId ?? generateId(this.name)\n    const threadId = options.threadId ?? generateId(this.name)\n    const messageId = generateId(this.name)\n    let stepId: string | null = null\n    let reasoningMessageId: string | null = null\n    let hasClosedReasoning = false\n    let hasEmittedRunStarted = false\n    let hasEmittedTextMessageStart = false\n    let hasEmittedStepStarted = false\n\n    for await (const chunk of result) {\n      logger.provider(`provider=gemini`, { chunk })\n      // Emit RUN_STARTED on first chunk\n      if (!hasEmittedRunStarted) {\n        hasEmittedRunStarted = true\n        yield {\n          type: EventType.RUN_STARTED,\n          runId,\n          threadId,\n          model,\n          timestamp: Date.now(),\n          parentRunId: options.parentRunId,\n        }\n      }\n\n      if (chunk.candidates?.[0]?.content?.parts) {\n        const parts = chunk.candidates[0].content.parts\n\n        for (const part of parts) {\n          if (part.text) {\n            if (part.thought) {\n              // Emit STEP_STARTED and REASONING events on first thinking content\n              if (!hasEmittedStepStarted) {\n                hasEmittedStepStarted = true\n                stepId = generateId(this.name)\n                reasoningMessageId = generateId(this.name)\n\n                // Spec REASONING events\n                yield {\n                  type: EventType.REASONING_START,\n                  messageId: reasoningMessageId,\n                  model,\n                  timestamp: Date.now(),\n                }\n                yield {\n                  type: EventType.REASONING_MESSAGE_START,\n                  messageId: reasoningMessageId,\n                  role: 'reasoning' as const,\n                  model,\n                  timestamp: Date.now(),\n                }\n\n                // Legacy STEP events (kept during transition)\n                yield {\n                  type: EventType.STEP_STARTED,\n                  stepName: stepId,\n                  stepId,\n                  model,\n                  timestamp: Date.now(),\n                  stepType: 'thinking',\n                }\n              }\n\n              accumulatedThinking += part.text\n\n              // Spec REASONING content event — reasoningMessageId is set in the\n              // hasEmittedStepStarted block above (entered on the same `part.thought` path)\n              if (!reasoningMessageId) continue\n              yield {\n                type: EventType.REASONING_MESSAGE_CONTENT,\n                messageId: reasoningMessageId,\n                delta: part.text,\n                model,\n                timestamp: Date.now(),\n              }\n\n              // Legacy STEP event\n              yield {\n                type: EventType.STEP_FINISHED,\n                stepName: stepId || generateId(this.name),\n                stepId: stepId || generateId(this.name),\n                model,\n                timestamp: Date.now(),\n                delta: part.text,\n                content: accumulatedThinking,\n              }\n            } else if (part.text.trim()) {\n              // Close reasoning before text starts\n              if (reasoningMessageId && !hasClosedReasoning) {\n                hasClosedReasoning = true\n                yield {\n                  type: EventType.REASONING_MESSAGE_END,\n                  messageId: reasoningMessageId,\n                  model,\n                  timestamp: Date.now(),\n                }\n                yield {\n                  type: EventType.REASONING_END,\n                  messageId: reasoningMessageId,\n                  model,\n                  timestamp: Date.now(),\n                }\n              }\n\n              // Skip whitespace-only text parts (e.g. \"\\n\" during auto-continuation)\n              // Emit TEXT_MESSAGE_START on first text content\n              if (!hasEmittedTextMessageStart) {\n                hasEmittedTextMessageStart = true\n                yield {\n                  type: EventType.TEXT_MESSAGE_START,\n                  messageId,\n                  model,\n                  timestamp: Date.now(),\n                  role: 'assistant',\n                }\n              }\n\n              accumulatedContent += part.text\n              yield {\n                type: EventType.TEXT_MESSAGE_CONTENT,\n                messageId,\n                model,\n                timestamp: Date.now(),\n                delta: part.text,\n                content: accumulatedContent,\n              }\n            }\n          }\n\n          const functionCall = part.functionCall\n          if (functionCall) {\n            const toolCallId =\n              functionCall.id ||\n              `${functionCall.name}_${Date.now()}_${nextToolIndex}`\n            const functionArgs = functionCall.args || {}\n\n            // Gemini emits thoughtSignature as a Part-level sibling of\n            // functionCall (per @google/genai Part type), not nested inside\n            // functionCall itself.\n            const partThoughtSignature = part.thoughtSignature || undefined\n\n            let toolCallData = toolCallMap.get(toolCallId)\n            if (!toolCallData) {\n              toolCallData = {\n                name: functionCall.name || '',\n                args:\n                  typeof functionArgs === 'string'\n                    ? functionArgs\n                    : JSON.stringify(functionArgs),\n                index: nextToolIndex++,\n                started: false,\n                // Only set thoughtSignature when present — under EOPT, the\n                // optional field cannot accept an explicit `undefined`.\n                ...(partThoughtSignature !== undefined && {\n                  thoughtSignature: partThoughtSignature,\n                }),\n              }\n              toolCallMap.set(toolCallId, toolCallData)\n            } else {\n              if (!toolCallData.thoughtSignature && partThoughtSignature) {\n                toolCallData.thoughtSignature = partThoughtSignature\n              }\n              try {\n                const existingArgs = JSON.parse(toolCallData.args)\n                const newArgs =\n                  typeof functionArgs === 'string'\n                    ? JSON.parse(functionArgs)\n                    : functionArgs\n                const mergedArgs = { ...existingArgs, ...newArgs }\n                toolCallData.args = JSON.stringify(mergedArgs)\n              } catch {\n                toolCallData.args =\n                  typeof functionArgs === 'string'\n                    ? functionArgs\n                    : JSON.stringify(functionArgs)\n              }\n            }\n\n            // Emit TOOL_CALL_START if not already started\n            if (!toolCallData.started) {\n              toolCallData.started = true\n              yield {\n                type: EventType.TOOL_CALL_START,\n                toolCallId,\n                toolCallName: toolCallData.name,\n                toolName: toolCallData.name,\n                parentMessageId: messageId,\n                model,\n                timestamp: Date.now(),\n                index: toolCallData.index,\n                ...(toolCallData.thoughtSignature && {\n                  metadata: {\n                    thoughtSignature: toolCallData.thoughtSignature,\n                  } satisfies GeminiToolCallMetadata,\n                }),\n              }\n            }\n\n            // Emit TOOL_CALL_ARGS\n            yield {\n              type: EventType.TOOL_CALL_ARGS,\n              toolCallId,\n              model,\n              timestamp: Date.now(),\n              delta: toolCallData.args,\n              args: toolCallData.args,\n            }\n          }\n        }\n      } else if (chunk.data && chunk.data.trim()) {\n        // Skip whitespace-only data (e.g. \"\\n\" during auto-continuation)\n        // Emit TEXT_MESSAGE_START on first text content\n        if (!hasEmittedTextMessageStart) {\n          hasEmittedTextMessageStart = true\n          yield {\n            type: EventType.TEXT_MESSAGE_START,\n            messageId,\n            model,\n            timestamp: Date.now(),\n            role: 'assistant',\n          }\n        }\n\n        accumulatedContent += chunk.data\n        yield {\n          type: EventType.TEXT_MESSAGE_CONTENT,\n          messageId,\n          model,\n          timestamp: Date.now(),\n          delta: chunk.data,\n          content: accumulatedContent,\n        }\n      }\n\n      if (chunk.candidates?.[0]?.finishReason) {\n        const finishReason = chunk.candidates[0].finishReason\n\n        if (finishReason === FinishReason.UNEXPECTED_TOOL_CALL) {\n          if (chunk.candidates[0].content?.parts) {\n            for (const part of chunk.candidates[0].content.parts) {\n              const functionCall = part.functionCall\n              if (functionCall) {\n                const toolCallId =\n                  functionCall.id ||\n                  `${functionCall.name}_${Date.now()}_${nextToolIndex}`\n                const functionArgs = functionCall.args || {}\n\n                const argsString =\n                  typeof functionArgs === 'string'\n                    ? functionArgs\n                    : JSON.stringify(functionArgs)\n\n                toolCallMap.set(toolCallId, {\n                  name: functionCall.name || '',\n                  args: argsString,\n                  index: nextToolIndex++,\n                  started: true,\n                })\n\n                // Emit TOOL_CALL_START\n                yield {\n                  type: EventType.TOOL_CALL_START,\n                  toolCallId,\n                  toolCallName: functionCall.name || '',\n                  toolName: functionCall.name || '',\n                  parentMessageId: messageId,\n                  model,\n                  timestamp: Date.now(),\n                  index: nextToolIndex - 1,\n                }\n\n                // Emit TOOL_CALL_END with parsed input\n                let parsedInput: unknown = {}\n                try {\n                  const parsed =\n                    typeof functionArgs === 'string'\n                      ? JSON.parse(functionArgs)\n                      : functionArgs\n                  parsedInput =\n                    parsed && typeof parsed === 'object' ? parsed : {}\n                } catch {\n                  parsedInput = {}\n                }\n\n                yield {\n                  type: EventType.TOOL_CALL_END,\n                  toolCallId,\n                  toolCallName: functionCall.name || '',\n                  toolName: functionCall.name || '',\n                  model,\n                  timestamp: Date.now(),\n                  input: parsedInput,\n                }\n              }\n            }\n          }\n        }\n\n        // Emit TOOL_CALL_END for all tracked tool calls\n        for (const [toolCallId, toolCallData] of toolCallMap.entries()) {\n          let parsedInput: unknown = {}\n          try {\n            const parsed = JSON.parse(toolCallData.args)\n            parsedInput = parsed && typeof parsed === 'object' ? parsed : {}\n          } catch {\n            parsedInput = {}\n          }\n\n          yield {\n            type: EventType.TOOL_CALL_END,\n            toolCallId,\n            toolCallName: toolCallData.name,\n            toolName: toolCallData.name,\n            model,\n            timestamp: Date.now(),\n            input: parsedInput,\n          }\n        }\n\n        // Reset so a new TEXT_MESSAGE_START is emitted if text follows tool calls\n        if (toolCallMap.size > 0) {\n          hasEmittedTextMessageStart = false\n        }\n\n        if (finishReason === FinishReason.MAX_TOKENS) {\n          yield {\n            type: EventType.RUN_ERROR,\n            runId,\n            model,\n            timestamp: Date.now(),\n            message:\n              'The response was cut off because the maximum token limit was reached.',\n            code: 'max_tokens',\n            error: {\n              message:\n                'The response was cut off because the maximum token limit was reached.',\n              code: 'max_tokens',\n            },\n          }\n        }\n\n        // Close reasoning events if still open\n        if (reasoningMessageId && !hasClosedReasoning) {\n          hasClosedReasoning = true\n          yield {\n            type: EventType.REASONING_MESSAGE_END,\n            messageId: reasoningMessageId,\n            model,\n            timestamp: Date.now(),\n          }\n          yield {\n            type: EventType.REASONING_END,\n            messageId: reasoningMessageId,\n            model,\n            timestamp: Date.now(),\n          }\n        }\n\n        // Emit TEXT_MESSAGE_END if we had text content\n        if (hasEmittedTextMessageStart) {\n          yield {\n            type: EventType.TEXT_MESSAGE_END,\n            messageId,\n            model,\n            timestamp: Date.now(),\n          }\n        }\n\n        yield {\n          type: EventType.RUN_FINISHED,\n          runId,\n          threadId,\n          model,\n          timestamp: Date.now(),\n          finishReason: toolCallMap.size > 0 ? 'tool_calls' : 'stop',\n          // RunFinishedEvent.usage is `usage?: {...}` (no `| undefined`) under\n          // exactOptionalPropertyTypes; only include it when usageMetadata is\n          // present rather than assigning an explicit `undefined`.\n          ...(chunk.usageMetadata && {\n            usage: buildGeminiUsage(chunk.usageMetadata),\n          }),\n        }\n      }\n    }\n  }\n\n  private convertContentPartToGemini(part: ContentPart): Part {\n    switch (part.type) {\n      case 'text':\n        return { text: part.content }\n      case 'image':\n      case 'audio':\n      case 'video':\n      case 'document': {\n        if (part.source.type === 'data') {\n          return {\n            inlineData: {\n              data: part.source.value,\n              mimeType: part.source.mimeType,\n            },\n          }\n        } else {\n          // For URL sources, use provided mimeType or fall back to reasonable defaults\n          const defaultMimeType = {\n            image: 'image/jpeg',\n            audio: 'audio/mp3',\n            video: 'video/mp4',\n            document: 'application/pdf',\n          }[part.type]\n\n          return {\n            fileData: {\n              fileUri: part.source.value,\n              mimeType: part.source.mimeType ?? defaultMimeType,\n            },\n          }\n        }\n      }\n      default: {\n        const _exhaustiveCheck: never = part\n        throw new Error(\n          `Unsupported content part type: ${(_exhaustiveCheck as ContentPart).type}`,\n        )\n      }\n    }\n  }\n\n  private formatMessages(\n    messages: Array<ModelMessage>,\n  ): GenerateContentParameters['contents'] {\n    // Build a lookup from toolCallId → function name so functionResponse uses the\n    // correct name instead of the raw call ID.\n    const toolCallIdToName = new Map<string, string>()\n    for (const msg of messages) {\n      if (msg.role === 'assistant' && msg.toolCalls) {\n        for (const tc of msg.toolCalls) {\n          toolCallIdToName.set(tc.id, tc.function.name)\n        }\n      }\n    }\n\n    const formatted = messages.map((msg) => {\n      const role: 'user' | 'model' = msg.role === 'assistant' ? 'model' : 'user'\n      const parts: Array<Part> = []\n\n      if (Array.isArray(msg.content)) {\n        for (const contentPart of msg.content) {\n          parts.push(this.convertContentPartToGemini(contentPart))\n        }\n      } else if (msg.content && msg.role !== 'tool') {\n        parts.push({ text: msg.content })\n      }\n\n      if (msg.role === 'assistant' && msg.toolCalls?.length) {\n        for (const toolCall of msg.toolCalls) {\n          let parsedArgs: Record<string, unknown> = {}\n          try {\n            parsedArgs = toolCall.function.arguments\n              ? (JSON.parse(toolCall.function.arguments) as Record<\n                  string,\n                  unknown\n                >)\n              : {}\n          } catch {\n            parsedArgs = {}\n          }\n\n          const thoughtSignature = (\n            toolCall.metadata as GeminiToolCallMetadata | undefined\n          )?.thoughtSignature\n          // Gemini requires thoughtSignature at the Part level (sibling of\n          // functionCall), not nested inside functionCall. Nesting it causes\n          // the API to reject the next turn with\n          // \"Function call is missing a thought_signature\".\n          const part: Part = {\n            functionCall: {\n              id: toolCall.id,\n              name: toolCall.function.name,\n              args: parsedArgs,\n            },\n          }\n          if (thoughtSignature) {\n            part.thoughtSignature = thoughtSignature\n          }\n          parts.push(part)\n        }\n      }\n\n      if (msg.role === 'tool' && msg.toolCallId) {\n        const functionName =\n          toolCallIdToName.get(msg.toolCallId) || msg.toolCallId\n        const toolContent = msg.content\n        if (Array.isArray(toolContent)) {\n          const textChunks: Array<string> = []\n          const mediaParts: Array<Part> = []\n          for (const part of toolContent) {\n            if (part.type === 'text') {\n              textChunks.push(part.content)\n            } else if (part.source.type === 'data') {\n              mediaParts.push({\n                inlineData: {\n                  data: part.source.value,\n                  mimeType: part.source.mimeType,\n                },\n              })\n            } else {\n              const defaultMimeType = {\n                image: 'image/jpeg',\n                audio: 'audio/mp3',\n                video: 'video/mp4',\n                document: 'application/pdf',\n              }[part.type]\n              mediaParts.push({\n                fileData: {\n                  fileUri: part.source.value,\n                  mimeType: part.source.mimeType ?? defaultMimeType,\n                },\n              })\n            }\n          }\n          parts.push({\n            functionResponse: {\n              id: msg.toolCallId,\n              name: functionName,\n              response: { content: textChunks.join('\\n') },\n              ...(mediaParts.length > 0 && { parts: mediaParts }),\n            },\n          })\n        } else {\n          parts.push({\n            functionResponse: {\n              id: msg.toolCallId,\n              name: functionName,\n              response: { content: toolContent || '' },\n            },\n          })\n        }\n      }\n\n      return {\n        role,\n        parts: parts.length > 0 ? parts : [{ text: '' }],\n      }\n    })\n\n    // Post-process: Gemini requires strictly alternating user/model roles.\n    // Tool results are mapped to role:'user', which can create consecutive\n    // user messages when followed by a new user message. Merge them.\n    return this.mergeConsecutiveSameRoleMessages(formatted)\n  }\n\n  /**\n   * Merge consecutive messages of the same role into a single message.\n   * Gemini's API requires strictly alternating user/model roles.\n   * Tool results are mapped to role:'user', which can collide with actual\n   * user messages in multi-turn conversations.\n   *\n   * Also filters out empty model messages (e.g., from a previous failed request)\n   * and deduplicates functionResponse parts with the same name (tool call ID).\n   */\n  private mergeConsecutiveSameRoleMessages(\n    messages: Array<Content>,\n  ): Array<Content> {\n    const merged: Array<Content> = []\n\n    for (const msg of messages) {\n      const parts = msg.parts || []\n\n      // Skip empty model messages (no parts or only empty text)\n      if (msg.role === 'model') {\n        const hasContent =\n          parts.length > 0 &&\n          !parts.every(\n            (p) => 'text' in p && (p as { text: string }).text === '',\n          )\n        if (!hasContent) {\n          continue\n        }\n      }\n\n      const prev = merged[merged.length - 1]\n      if (prev && prev.role === msg.role) {\n        // Merge parts arrays\n        prev.parts = [...(prev.parts || []), ...parts]\n      } else {\n        merged.push({ ...msg, parts: [...parts] })\n      }\n    }\n\n    // Deduplicate functionResponse parts with the same name (tool call ID)\n    for (const msg of merged) {\n      if (!msg.parts) continue\n      const seenFunctionResponseNames = new Set<string>()\n      msg.parts = msg.parts.filter((part) => {\n        if ('functionResponse' in part && part.functionResponse?.name) {\n          if (seenFunctionResponseNames.has(part.functionResponse.name)) {\n            return false\n          }\n          seenFunctionResponseNames.add(part.functionResponse.name)\n        }\n        return true\n      })\n    }\n\n    return merged\n  }\n\n  private mapCommonOptionsToGemini(\n    options: TextOptions<GeminiTextProviderOptions>,\n  ) {\n    // Separate `thinkingConfig` from the other model options so the loose\n    // local `thinkingLevel?: keyof typeof ThinkingLevel` type doesn't leak\n    // into the SDK config object via the `...modelOpts` spread — we re-add a\n    // properly-typed `ThinkingConfig` below.\n    const { thinkingConfig, ...modelOpts } = options.modelOptions ?? {}\n    // Build the thinkingConfig payload only when the caller actually supplied\n    // one. Our local `thinkingLevel` is typed as `keyof typeof ThinkingLevel`\n    // (string union) so users can pass plain strings; the SDK target is the\n    // `ThinkingLevel` enum, and every field is `field?: T` under EOPT — so we\n    // re-emit fields via conditional spreads.\n    const mappedThinkingConfig = thinkingConfig\n      ? {\n          ...(thinkingConfig.includeThoughts !== undefined && {\n            includeThoughts: thinkingConfig.includeThoughts,\n          }),\n          ...(thinkingConfig.thinkingBudget !== undefined && {\n            thinkingBudget: thinkingConfig.thinkingBudget,\n          }),\n          ...(thinkingConfig.thinkingLevel\n            ? {\n                thinkingLevel: thinkingConfig.thinkingLevel as ThinkingLevel,\n              }\n            : {}),\n        }\n      : undefined\n\n    const normalizedPrompts = normalizeSystemPrompts(options.systemPrompts)\n    const systemInstruction =\n      normalizedPrompts.length > 0\n        ? normalizedPrompts.map((p) => p.content).join('\\n')\n        : undefined\n\n    // Native combined mode (issue #605): when the engine threads\n    // `outputSchema` through TextOptions, the adapter declared\n    // `supportsCombinedToolsAndSchema` (Gemini 3.x only). The schema is\n    // already JSON Schema (pre-converted at the activity boundary). Wire\n    // it into `config.responseSchema` + `responseMimeType: 'application/json'`\n    // alongside any `tools` — the model emits function calls during the\n    // agent loop and the schema-constrained JSON on its natural final\n    // turn, so the engine can harvest it without the separate\n    // `structuredOutput` finalization round-trip.\n    const combinedSchema = options.outputSchema as\n      | Record<string, unknown>\n      | undefined\n    const combinedSchemaConfig = combinedSchema\n      ? {\n          responseMimeType: 'application/json' as const,\n          responseSchema: combinedSchema,\n        }\n      : undefined\n\n    // Vendor `GenerateContentConfig` fields are `field?: T` (no `| undefined`)\n    // under EOPT, so spread each common option only when present rather than\n    // emitting `field: undefined`s into the wire payload.\n    const requestOptions: GenerateContentParameters = {\n      model: options.model,\n      contents: this.formatMessages(options.messages),\n      config: {\n        ...modelOpts,\n        ...(mappedThinkingConfig !== undefined && {\n          thinkingConfig: mappedThinkingConfig,\n        }),\n        ...(systemInstruction !== undefined && { systemInstruction }),\n        tools: convertToolsToProviderFormat(options.tools),\n        ...(combinedSchemaConfig ?? {}),\n      },\n    }\n\n    return requestOptions\n  }\n\n  /**\n   * Gemini 3.x natively combines `tools` + `responseSchema` in a single\n   * streaming `generateContentStream` call (issue #605). Gemini 2.x is\n   * documented as brittle for the combination and keeps the engine's\n   * legacy finalization path.\n   */\n  supportsCombinedToolsAndSchema(): boolean {\n    return GEMINI_COMBINED_TOOLS_AND_SCHEMA_MODELS.has(this.model)\n  }\n}\n\n/**\n * Creates a Gemini text adapter with explicit API key.\n * Type resolution happens here at the call site.\n */\nexport function createGeminiChat<TModel extends (typeof GEMINI_MODELS)[number]>(\n  model: TModel,\n  apiKey: string,\n  config?: Omit<GeminiTextConfig, 'apiKey'>,\n): GeminiTextAdapter<\n  TModel,\n  ResolveProviderOptions<TModel>,\n  ResolveInputModalities<TModel>,\n  ResolveToolCapabilities<TModel>\n> {\n  return new GeminiTextAdapter({ apiKey, ...config }, model)\n}\n\n/**\n * Creates a Gemini text adapter with automatic API key detection.\n * Type resolution happens here at the call site.\n */\nexport function geminiText<TModel extends (typeof GEMINI_MODELS)[number]>(\n  model: TModel,\n  config?: Omit<GeminiTextConfig, 'apiKey'>,\n): GeminiTextAdapter<\n  TModel,\n  ResolveProviderOptions<TModel>,\n  ResolveInputModalities<TModel>,\n  ResolveToolCapabilities<TModel>\n> {\n  const apiKey = getGeminiApiKeyFromEnv()\n  return createGeminiChat(model, apiKey, config)\n}\n"],"names":[],"mappings":";;;;;;;;AAgGO,MAAM,0BAOH,gBAOR;AAAA,EACkB,OAAO;AAAA,EAChB,OAAO;AAAA,EAEC;AAAA,EAEjB,YAAY,QAA0B,OAAe;AACnD,UAAM,CAAA,GAAI,KAAK;AACf,SAAK,SAAS,mBAAmB,MAAM;AAAA,EACzC;AAAA,EAEA,OAAO,WACL,SAC4B;AAC5B,UAAM,gBAAgB,KAAK,yBAAyB,OAAO;AAC3D,UAAM,EAAE,WAAW;AAEnB,QAAI;AACF,aAAO;AAAA,QACL,uCAAuC,KAAK,KAAK,aAAa,QAAQ,SAAS,MAAM,UAAU,QAAQ,OAAO,UAAU,CAAC;AAAA,QACzH,EAAE,UAAU,UAAU,OAAO,KAAK,MAAA;AAAA,MAAM;AAE1C,YAAM,SACJ,MAAM,KAAK,OAAO,OAAO,sBAAsB,aAAa;AAE9D,aAAO,KAAK,oBAAoB,QAAQ,SAAS,MAAM;AAAA,IACzD,SAAS,OAAO;AACd,YAAM,WAAW,mBAAmB,KAAK;AACzC,aAAO,OAAO,2BAA2B;AAAA,QACvC;AAAA,QACA,QAAQ;AAAA,MAAA,CACT;AACD,YAAM;AAAA,QACJ,MAAM,UAAU;AAAA,QAChB,OAAO,QAAQ;AAAA,QACf,WAAW,KAAK,IAAA;AAAA,QAChB,SACE,iBAAiB,QACb,MAAM,UACN;AAAA;AAAA;AAAA,QAGN,GAAI,aAAa,UAAa,EAAE,SAAA;AAAA,QAChC,OAAO;AAAA,UACL,SACE,iBAAiB,QACb,MAAM,UACN;AAAA,QAAA;AAAA,MACR;AAAA,IAEJ;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOA,MAAM,iBACJ,SAC0C;AAC1C,UAAM,EAAE,aAAa,aAAA,IAAiB;AACtC,UAAM,EAAE,WAAW;AAEnB,UAAM,gBAAgB,KAAK,yBAAyB,WAAW;AAE/D,QAAI;AACF,aAAO;AAAA,QACL,uCAAuC,KAAK,KAAK,aAAa,YAAY,SAAS,MAAM,UAAU,YAAY,OAAO,UAAU,CAAC;AAAA,QACjI,EAAE,UAAU,UAAU,OAAO,KAAK,MAAA;AAAA,MAAM;AAG1C,YAAM,SAAS,MAAM,KAAK,OAAO,OAAO,gBAAgB;AAAA,QACtD,GAAG;AAAA,QACH,QAAQ;AAAA,UACN,GAAG,cAAc;AAAA,UACjB,kBAAkB;AAAA,UAClB,gBAAgB;AAAA,QAAA;AAAA,MAClB,CACD;AAGD,YAAM,UAAU,KAAK,wBAAwB,MAAM;AAGnD,UAAI;AACJ,UAAI;AACF,iBAAS,KAAK,MAAM,OAAO;AAAA,MAC7B,QAAQ;AACN,cAAM,IAAI;AAAA,UACR,uDAAuD,QAAQ,MAAM,GAAG,GAAG,CAAC,GAAG,QAAQ,SAAS,MAAM,QAAQ,EAAE;AAAA,QAAA;AAAA,MAEpH;AAEA,aAAO;AAAA,QACL,MAAM;AAAA,QACN;AAAA,QACA,OAAO,OAAO,gBACV,iBAAiB,OAAO,aAAa,IACrC;AAAA,MAAA;AAAA,IAER,SAAS,OAAO;AACd,aAAO,OAAO,iCAAiC;AAAA,QAC7C;AAAA,QACA,QAAQ;AAAA,MAAA,CACT;AACD,YAAM,IAAI;AAAA,QACR,iBAAiB,QACb,MAAM,UACN;AAAA,MAAA;AAAA,IAER;AAAA,EACF;AAAA;AAAA;AAAA;AAAA,EAKQ,wBAAwB,UAA2C;AACzE,QAAI,cAAc;AAElB,QAAI,SAAS,aAAa,CAAC,GAAG,SAAS,OAAO;AAC5C,iBAAW,QAAQ,SAAS,WAAW,CAAC,EAAE,QAAQ,OAAO;AACvD,YAAI,KAAK,MAAM;AACb,yBAAe,KAAK;AAAA,QACtB;AAAA,MACF;AAAA,IACF;AAEA,WAAO;AAAA,EACT;AAAA,EAEA,OAAe,oBACb,QACA,SACA,QAC4B;AAC5B,UAAM,QAAQ,QAAQ;AACtB,QAAI,qBAAqB;AACzB,QAAI,sBAAsB;AAC1B,UAAM,kCAAkB,IAAA;AAUxB,QAAI,gBAAgB;AAGpB,UAAM,QAAQ,QAAQ,SAAS,WAAW,KAAK,IAAI;AACnD,UAAM,WAAW,QAAQ,YAAY,WAAW,KAAK,IAAI;AACzD,UAAM,YAAY,WAAW,KAAK,IAAI;AACtC,QAAI,SAAwB;AAC5B,QAAI,qBAAoC;AACxC,QAAI,qBAAqB;AACzB,QAAI,uBAAuB;AAC3B,QAAI,6BAA6B;AACjC,QAAI,wBAAwB;AAE5B,qBAAiB,SAAS,QAAQ;AAChC,aAAO,SAAS,mBAAmB,EAAE,MAAA,CAAO;AAE5C,UAAI,CAAC,sBAAsB;AACzB,+BAAuB;AACvB,cAAM;AAAA,UACJ,MAAM,UAAU;AAAA,UAChB;AAAA,UACA;AAAA,UACA;AAAA,UACA,WAAW,KAAK,IAAA;AAAA,UAChB,aAAa,QAAQ;AAAA,QAAA;AAAA,MAEzB;AAEA,UAAI,MAAM,aAAa,CAAC,GAAG,SAAS,OAAO;AACzC,cAAM,QAAQ,MAAM,WAAW,CAAC,EAAE,QAAQ;AAE1C,mBAAW,QAAQ,OAAO;AACxB,cAAI,KAAK,MAAM;AACb,gBAAI,KAAK,SAAS;AAEhB,kBAAI,CAAC,uBAAuB;AAC1B,wCAAwB;AACxB,yBAAS,WAAW,KAAK,IAAI;AAC7B,qCAAqB,WAAW,KAAK,IAAI;AAGzC,sBAAM;AAAA,kBACJ,MAAM,UAAU;AAAA,kBAChB,WAAW;AAAA,kBACX;AAAA,kBACA,WAAW,KAAK,IAAA;AAAA,gBAAI;AAEtB,sBAAM;AAAA,kBACJ,MAAM,UAAU;AAAA,kBAChB,WAAW;AAAA,kBACX,MAAM;AAAA,kBACN;AAAA,kBACA,WAAW,KAAK,IAAA;AAAA,gBAAI;AAItB,sBAAM;AAAA,kBACJ,MAAM,UAAU;AAAA,kBAChB,UAAU;AAAA,kBACV;AAAA,kBACA;AAAA,kBACA,WAAW,KAAK,IAAA;AAAA,kBAChB,UAAU;AAAA,gBAAA;AAAA,cAEd;AAEA,qCAAuB,KAAK;AAI5B,kBAAI,CAAC,mBAAoB;AACzB,oBAAM;AAAA,gBACJ,MAAM,UAAU;AAAA,gBAChB,WAAW;AAAA,gBACX,OAAO,KAAK;AAAA,gBACZ;AAAA,gBACA,WAAW,KAAK,IAAA;AAAA,cAAI;AAItB,oBAAM;AAAA,gBACJ,MAAM,UAAU;AAAA,gBAChB,UAAU,UAAU,WAAW,KAAK,IAAI;AAAA,gBACxC,QAAQ,UAAU,WAAW,KAAK,IAAI;AAAA,gBACtC;AAAA,gBACA,WAAW,KAAK,IAAA;AAAA,gBAChB,OAAO,KAAK;AAAA,gBACZ,SAAS;AAAA,cAAA;AAAA,YAEb,WAAW,KAAK,KAAK,KAAA,GAAQ;AAE3B,kBAAI,sBAAsB,CAAC,oBAAoB;AAC7C,qCAAqB;AACrB,sBAAM;AAAA,kBACJ,MAAM,UAAU;AAAA,kBAChB,WAAW;AAAA,kBACX;AAAA,kBACA,WAAW,KAAK,IAAA;AAAA,gBAAI;AAEtB,sBAAM;AAAA,kBACJ,MAAM,UAAU;AAAA,kBAChB,WAAW;AAAA,kBACX;AAAA,kBACA,WAAW,KAAK,IAAA;AAAA,gBAAI;AAAA,cAExB;AAIA,kBAAI,CAAC,4BAA4B;AAC/B,6CAA6B;AAC7B,sBAAM;AAAA,kBACJ,MAAM,UAAU;AAAA,kBAChB;AAAA,kBACA;AAAA,kBACA,WAAW,KAAK,IAAA;AAAA,kBAChB,MAAM;AAAA,gBAAA;AAAA,cAEV;AAEA,oCAAsB,KAAK;AAC3B,oBAAM;AAAA,gBACJ,MAAM,UAAU;AAAA,gBAChB;AAAA,gBACA;AAAA,gBACA,WAAW,KAAK,IAAA;AAAA,gBAChB,OAAO,KAAK;AAAA,gBACZ,SAAS;AAAA,cAAA;AAAA,YAEb;AAAA,UACF;AAEA,gBAAM,eAAe,KAAK;AAC1B,cAAI,cAAc;AAChB,kBAAM,aACJ,aAAa,MACb,GAAG,aAAa,IAAI,IAAI,KAAK,IAAA,CAAK,IAAI,aAAa;AACrD,kBAAM,eAAe,aAAa,QAAQ,CAAA;AAK1C,kBAAM,uBAAuB,KAAK,oBAAoB;AAEtD,gBAAI,eAAe,YAAY,IAAI,UAAU;AAC7C,gBAAI,CAAC,cAAc;AACjB,6BAAe;AAAA,gBACb,MAAM,aAAa,QAAQ;AAAA,gBAC3B,MACE,OAAO,iBAAiB,WACpB,eACA,KAAK,UAAU,YAAY;AAAA,gBACjC,OAAO;AAAA,gBACP,SAAS;AAAA;AAAA;AAAA,gBAGT,GAAI,yBAAyB,UAAa;AAAA,kBACxC,kBAAkB;AAAA,gBAAA;AAAA,cACpB;AAEF,0BAAY,IAAI,YAAY,YAAY;AAAA,YAC1C,OAAO;AACL,kBAAI,CAAC,aAAa,oBAAoB,sBAAsB;AAC1D,6BAAa,mBAAmB;AAAA,cAClC;AACA,kBAAI;AACF,sBAAM,eAAe,KAAK,MAAM,aAAa,IAAI;AACjD,sBAAM,UACJ,OAAO,iBAAiB,WACpB,KAAK,MAAM,YAAY,IACvB;AACN,sBAAM,aAAa,EAAE,GAAG,cAAc,GAAG,QAAA;AACzC,6BAAa,OAAO,KAAK,UAAU,UAAU;AAAA,cAC/C,QAAQ;AACN,6BAAa,OACX,OAAO,iBAAiB,WACpB,eACA,KAAK,UAAU,YAAY;AAAA,cACnC;AAAA,YACF;AAGA,gBAAI,CAAC,aAAa,SAAS;AACzB,2BAAa,UAAU;AACvB,oBAAM;AAAA,gBACJ,MAAM,UAAU;AAAA,gBAChB;AAAA,gBACA,cAAc,aAAa;AAAA,gBAC3B,UAAU,aAAa;AAAA,gBACvB,iBAAiB;AAAA,gBACjB;AAAA,gBACA,WAAW,KAAK,IAAA;AAAA,gBAChB,OAAO,aAAa;AAAA,gBACpB,GAAI,aAAa,oBAAoB;AAAA,kBACnC,UAAU;AAAA,oBACR,kBAAkB,aAAa;AAAA,kBAAA;AAAA,gBACjC;AAAA,cACF;AAAA,YAEJ;AAGA,kBAAM;AAAA,cACJ,MAAM,UAAU;AAAA,cAChB;AAAA,cACA;AAAA,cACA,WAAW,KAAK,IAAA;AAAA,cAChB,OAAO,aAAa;AAAA,cACpB,MAAM,aAAa;AAAA,YAAA;AAAA,UAEvB;AAAA,QACF;AAAA,MACF,WAAW,MAAM,QAAQ,MAAM,KAAK,QAAQ;AAG1C,YAAI,CAAC,4BAA4B;AAC/B,uCAA6B;AAC7B,gBAAM;AAAA,YACJ,MAAM,UAAU;AAAA,YAChB;AAAA,YACA;AAAA,YACA,WAAW,KAAK,IAAA;AAAA,YAChB,MAAM;AAAA,UAAA;AAAA,QAEV;AAEA,8BAAsB,MAAM;AAC5B,cAAM;AAAA,UACJ,MAAM,UAAU;AAAA,UAChB;AAAA,UACA;AAAA,UACA,WAAW,KAAK,IAAA;AAAA,UAChB,OAAO,MAAM;AAAA,UACb,SAAS;AAAA,QAAA;AAAA,MAEb;AAEA,UAAI,MAAM,aAAa,CAAC,GAAG,cAAc;AACvC,cAAM,eAAe,MAAM,WAAW,CAAC,EAAE;AAEzC,YAAI,iBAAiB,aAAa,sBAAsB;AACtD,cAAI,MAAM,WAAW,CAAC,EAAE,SAAS,OAAO;AACtC,uBAAW,QAAQ,MAAM,WAAW,CAAC,EAAE,QAAQ,OAAO;AACpD,oBAAM,eAAe,KAAK;AAC1B,kBAAI,cAAc;AAChB,sBAAM,aACJ,aAAa,MACb,GAAG,aAAa,IAAI,IAAI,KAAK,IAAA,CAAK,IAAI,aAAa;AACrD,sBAAM,eAAe,aAAa,QAAQ,CAAA;AAE1C,sBAAM,aACJ,OAAO,iBAAiB,WACpB,eACA,KAAK,UAAU,YAAY;AAEjC,4BAAY,IAAI,YAAY;AAAA,kBAC1B,MAAM,aAAa,QAAQ;AAAA,kBAC3B,MAAM;AAAA,kBACN,OAAO;AAAA,kBACP,SAAS;AAAA,gBAAA,CACV;AAGD,sBAAM;AAAA,kBACJ,MAAM,UAAU;AAAA,kBAChB;AAAA,kBACA,cAAc,aAAa,QAAQ;AAAA,kBACnC,UAAU,aAAa,QAAQ;AAAA,kBAC/B,iBAAiB;AAAA,kBACjB;AAAA,kBACA,WAAW,KAAK,IAAA;AAAA,kBAChB,OAAO,gBAAgB;AAAA,gBAAA;AAIzB,oBAAI,cAAuB,CAAA;AAC3B,oBAAI;AACF,wBAAM,SACJ,OAAO,iBAAiB,WACpB,KAAK,MAAM,YAAY,IACvB;AACN,gCACE,UAAU,OAAO,WAAW,WAAW,SAAS,CAAA;AAAA,gBACpD,QAAQ;AACN,gCAAc,CAAA;AAAA,gBAChB;AAEA,sBAAM;AAAA,kBACJ,MAAM,UAAU;AAAA,kBAChB;AAAA,kBACA,cAAc,aAAa,QAAQ;AAAA,kBACnC,UAAU,aAAa,QAAQ;AAAA,kBAC/B;AAAA,kBACA,WAAW,KAAK,IAAA;AAAA,kBAChB,OAAO;AAAA,gBAAA;AAAA,cAEX;AAAA,YACF;AAAA,UACF;AAAA,QACF;AAGA,mBAAW,CAAC,YAAY,YAAY,KAAK,YAAY,WAAW;AAC9D,cAAI,cAAuB,CAAA;AAC3B,cAAI;AACF,kBAAM,SAAS,KAAK,MAAM,aAAa,IAAI;AAC3C,0BAAc,UAAU,OAAO,WAAW,WAAW,SAAS,CAAA;AAAA,UAChE,QAAQ;AACN,0BAAc,CAAA;AAAA,UAChB;AAEA,gBAAM;AAAA,YACJ,MAAM,UAAU;AAAA,YAChB;AAAA,YACA,cAAc,aAAa;AAAA,YAC3B,UAAU,aAAa;AAAA,YACvB;AAAA,YACA,WAAW,KAAK,IAAA;AAAA,YAChB,OAAO;AAAA,UAAA;AAAA,QAEX;AAGA,YAAI,YAAY,OAAO,GAAG;AACxB,uCAA6B;AAAA,QAC/B;AAEA,YAAI,iBAAiB,aAAa,YAAY;AAC5C,gBAAM;AAAA,YACJ,MAAM,UAAU;AAAA,YAChB;AAAA,YACA;AAAA,YACA,WAAW,KAAK,IAAA;AAAA,YAChB,SACE;AAAA,YACF,MAAM;AAAA,YACN,OAAO;AAAA,cACL,SACE;AAAA,cACF,MAAM;AAAA,YAAA;AAAA,UACR;AAAA,QAEJ;AAGA,YAAI,sBAAsB,CAAC,oBAAoB;AAC7C,+BAAqB;AACrB,gBAAM;AAAA,YACJ,MAAM,UAAU;AAAA,YAChB,WAAW;AAAA,YACX;AAAA,YACA,WAAW,KAAK,IAAA;AAAA,UAAI;AAEtB,gBAAM;AAAA,YACJ,MAAM,UAAU;AAAA,YAChB,WAAW;AAAA,YACX;AAAA,YACA,WAAW,KAAK,IAAA;AAAA,UAAI;AAAA,QAExB;AAGA,YAAI,4BAA4B;AAC9B,gBAAM;AAAA,YACJ,MAAM,UAAU;AAAA,YAChB;AAAA,YACA;AAAA,YACA,WAAW,KAAK,IAAA;AAAA,UAAI;AAAA,QAExB;AAEA,cAAM;AAAA,UACJ,MAAM,UAAU;AAAA,UAChB;AAAA,UACA;AAAA,UACA;AAAA,UACA,WAAW,KAAK,IAAA;AAAA,UAChB,cAAc,YAAY,OAAO,IAAI,eAAe;AAAA;AAAA;AAAA;AAAA,UAIpD,GAAI,MAAM,iBAAiB;AAAA,YACzB,OAAO,iBAAiB,MAAM,aAAa;AAAA,UAAA;AAAA,QAC7C;AAAA,MAEJ;AAAA,IACF;AAAA,EACF;AAAA,EAEQ,2BAA2B,MAAyB;AAC1D,YAAQ,KAAK,MAAA;AAAA,MACX,KAAK;AACH,eAAO,EAAE,MAAM,KAAK,QAAA;AAAA,MACtB,KAAK;AAAA,MACL,KAAK;AAAA,MACL,KAAK;AAAA,MACL,KAAK,YAAY;AACf,YAAI,KAAK,OAAO,SAAS,QAAQ;AAC/B,iBAAO;AAAA,YACL,YAAY;AAAA,cACV,MAAM,KAAK,OAAO;AAAA,cAClB,UAAU,KAAK,OAAO;AAAA,YAAA;AAAA,UACxB;AAAA,QAEJ,OAAO;AAEL,gBAAM,kBAAkB;AAAA,YACtB,OAAO;AAAA,YACP,OAAO;AAAA,YACP,OAAO;AAAA,YACP,UAAU;AAAA,UAAA,EACV,KAAK,IAAI;AAEX,iBAAO;AAAA,YACL,UAAU;AAAA,cACR,SAAS,KAAK,OAAO;AAAA,cACrB,UAAU,KAAK,OAAO,YAAY;AAAA,YAAA;AAAA,UACpC;AAAA,QAEJ;AAAA,MACF;AAAA,MACA,SAAS;AACP,cAAM,mBAA0B;AAChC,cAAM,IAAI;AAAA,UACR,kCAAmC,iBAAiC,IAAI;AAAA,QAAA;AAAA,MAE5E;AAAA,IAAA;AAAA,EAEJ;AAAA,EAEQ,eACN,UACuC;AAGvC,UAAM,uCAAuB,IAAA;AAC7B,eAAW,OAAO,UAAU;AAC1B,UAAI,IAAI,SAAS,eAAe,IAAI,WAAW;AAC7C,mBAAW,MAAM,IAAI,WAAW;AAC9B,2BAAiB,IAAI,GAAG,IAAI,GAAG,SAAS,IAAI;AAAA,QAC9C;AAAA,MACF;AAAA,IACF;AAEA,UAAM,YAAY,SAAS,IAAI,CAAC,QAAQ;AACtC,YAAM,OAAyB,IAAI,SAAS,cAAc,UAAU;AACpE,YAAM,QAAqB,CAAA;AAE3B,UAAI,MAAM,QAAQ,IAAI,OAAO,GAAG;AAC9B,mBAAW,eAAe,IAAI,SAAS;AACrC,gBAAM,KAAK,KAAK,2BAA2B,WAAW,CAAC;AAAA,QACzD;AAAA,MACF,WAAW,IAAI,WAAW,IAAI,SAAS,QAAQ;AAC7C,cAAM,KAAK,EAAE,MAAM,IAAI,SAAS;AAAA,MAClC;AAEA,UAAI,IAAI,SAAS,eAAe,IAAI,WAAW,QAAQ;AACrD,mBAAW,YAAY,IAAI,WAAW;AACpC,cAAI,aAAsC,CAAA;AAC1C,cAAI;AACF,yBAAa,SAAS,SAAS,YAC1B,KAAK,MAAM,SAAS,SAAS,SAAS,IAIvC,CAAA;AAAA,UACN,QAAQ;AACN,yBAAa,CAAA;AAAA,UACf;AAEA,gBAAM,mBACJ,SAAS,UACR;AAKH,gBAAM,OAAa;AAAA,YACjB,cAAc;AAAA,cACZ,IAAI,SAAS;AAAA,cACb,MAAM,SAAS,SAAS;AAAA,cACxB,MAAM;AAAA,YAAA;AAAA,UACR;AAEF,cAAI,kBAAkB;AACpB,iBAAK,mBAAmB;AAAA,UAC1B;AACA,gBAAM,KAAK,IAAI;AAAA,QACjB;AAAA,MACF;AAEA,UAAI,IAAI,SAAS,UAAU,IAAI,YAAY;AACzC,cAAM,eACJ,iBAAiB,IAAI,IAAI,UAAU,KAAK,IAAI;AAC9C,cAAM,cAAc,IAAI;AACxB,YAAI,MAAM,QAAQ,WAAW,GAAG;AAC9B,gBAAM,aAA4B,CAAA;AAClC,gBAAM,aAA0B,CAAA;AAChC,qBAAW,QAAQ,aAAa;AAC9B,gBAAI,KAAK,SAAS,QAAQ;AACxB,yBAAW,KAAK,KAAK,OAAO;AAAA,YAC9B,WAAW,KAAK,OAAO,SAAS,QAAQ;AACtC,yBAAW,KAAK;AAAA,gBACd,YAAY;AAAA,kBACV,MAAM,KAAK,OAAO;AAAA,kBAClB,UAAU,KAAK,OAAO;AAAA,gBAAA;AAAA,cACxB,CACD;AAAA,YACH,OAAO;AACL,oBAAM,kBAAkB;AAAA,gBACtB,OAAO;AAAA,gBACP,OAAO;AAAA,gBACP,OAAO;AAAA,gBACP,UAAU;AAAA,cAAA,EACV,KAAK,IAAI;AACX,yBAAW,KAAK;AAAA,gBACd,UAAU;AAAA,kBACR,SAAS,KAAK,OAAO;AAAA,kBACrB,UAAU,KAAK,OAAO,YAAY;AAAA,gBAAA;AAAA,cACpC,CACD;AAAA,YACH;AAAA,UACF;AACA,gBAAM,KAAK;AAAA,YACT,kBAAkB;AAAA,cAChB,IAAI,IAAI;AAAA,cACR,MAAM;AAAA,cACN,UAAU,EAAE,SAAS,WAAW,KAAK,IAAI,EAAA;AAAA,cACzC,GAAI,WAAW,SAAS,KAAK,EAAE,OAAO,WAAA;AAAA,YAAW;AAAA,UACnD,CACD;AAAA,QACH,OAAO;AACL,gBAAM,KAAK;AAAA,YACT,kBAAkB;AAAA,cAChB,IAAI,IAAI;AAAA,cACR,MAAM;AAAA,cACN,UAAU,EAAE,SAAS,eAAe,GAAA;AAAA,YAAG;AAAA,UACzC,CACD;AAAA,QACH;AAAA,MACF;AAEA,aAAO;AAAA,QACL;AAAA,QACA,OAAO,MAAM,SAAS,IAAI,QAAQ,CAAC,EAAE,MAAM,GAAA,CAAI;AAAA,MAAA;AAAA,IAEnD,CAAC;AAKD,WAAO,KAAK,iCAAiC,SAAS;AAAA,EACxD;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAWQ,iCACN,UACgB;AAChB,UAAM,SAAyB,CAAA;AAE/B,eAAW,OAAO,UAAU;AAC1B,YAAM,QAAQ,IAAI,SAAS,CAAA;AAG3B,UAAI,IAAI,SAAS,SAAS;AACxB,cAAM,aACJ,MAAM,SAAS,KACf,CAAC,MAAM;AAAA,UACL,CAAC,MAAM,UAAU,KAAM,EAAuB,SAAS;AAAA,QAAA;AAE3D,YAAI,CAAC,YAAY;AACf;AAAA,QACF;AAAA,MACF;AAEA,YAAM,OAAO,OAAO,OAAO,SAAS,CAAC;AACrC,UAAI,QAAQ,KAAK,SAAS,IAAI,MAAM;AAElC,aAAK,QAAQ,CAAC,GAAI,KAAK,SAAS,CAAA,GAAK,GAAG,KAAK;AAAA,MAC/C,OAAO;AACL,eAAO,KAAK,EAAE,GAAG,KAAK,OAAO,CAAC,GAAG,KAAK,GAAG;AAAA,MAC3C;AAAA,IACF;AAGA,eAAW,OAAO,QAAQ;AACxB,UAAI,CAAC,IAAI,MAAO;AAChB,YAAM,gDAAgC,IAAA;AACtC,UAAI,QAAQ,IAAI,MAAM,OAAO,CAAC,SAAS;AACrC,YAAI,sBAAsB,QAAQ,KAAK,kBAAkB,MAAM;AAC7D,cAAI,0BAA0B,IAAI,KAAK,iBAAiB,IAAI,GAAG;AAC7D,mBAAO;AAAA,UACT;AACA,oCAA0B,IAAI,KAAK,iBAAiB,IAAI;AAAA,QAC1D;AACA,eAAO;AAAA,MACT,CAAC;AAAA,IACH;AAEA,WAAO;AAAA,EACT;AAAA,EAEQ,yBACN,SACA;AAKA,UAAM,EAAE,gBAAgB,GAAG,cAAc,QAAQ,gBAAgB,CAAA;AAMjE,UAAM,uBAAuB,iBACzB;AAAA,MACE,GAAI,eAAe,oBAAoB,UAAa;AAAA,QAClD,iBAAiB,eAAe;AAAA,MAAA;AAAA,MAElC,GAAI,eAAe,mBAAmB,UAAa;AAAA,QACjD,gBAAgB,eAAe;AAAA,MAAA;AAAA,MAEjC,GAAI,eAAe,gBACf;AAAA,QACE,eAAe,eAAe;AAAA,MAAA,IAEhC,CAAA;AAAA,IAAC,IAEP;AAEJ,UAAM,oBAAoB,uBAAuB,QAAQ,aAAa;AACtE,UAAM,oBACJ,kBAAkB,SAAS,IACvB,kBAAkB,IAAI,CAAC,MAAM,EAAE,OAAO,EAAE,KAAK,IAAI,IACjD;AAWN,UAAM,iBAAiB,QAAQ;AAG/B,UAAM,uBAAuB,iBACzB;AAAA,MACE,kBAAkB;AAAA,MAClB,gBAAgB;AAAA,IAAA,IAElB;AAKJ,UAAM,iBAA4C;AAAA,MAChD,OAAO,QAAQ;AAAA,MACf,UAAU,KAAK,eAAe,QAAQ,QAAQ;AAAA,MAC9C,QAAQ;AAAA,QACN,GAAG;AAAA,QACH,GAAI,yBAAyB,UAAa;AAAA,UACxC,gBAAgB;AAAA,QAAA;AAAA,QAElB,GAAI,sBAAsB,UAAa,EAAE,kBAAA;AAAA,QACzC,OAAO,6BAA6B,QAAQ,KAAK;AAAA,QACjD,GAAI,wBAAwB,CAAA;AAAA,MAAC;AAAA,IAC/B;AAGF,WAAO;AAAA,EACT;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,iCAA0C;AACxC,WAAO,wCAAwC,IAAI,KAAK,KAAK;AAAA,EAC/D;AACF;AAMO,SAAS,iBACd,OACA,QACA,QAMA;AACA,SAAO,IAAI,kBAAkB,EAAE,QAAQ,GAAG,OAAA,GAAU,KAAK;AAC3D;AAMO,SAAS,WACd,OACA,QAMA;AACA,QAAM,SAAS,uBAAA;AACf,SAAO,iBAAiB,OAAO,QAAQ,MAAM;AAC/C;"}

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

@@ -0,0 +1,165 @@
+import { BaseTTSAdapter } from '@tanstack/ai/adapters';
+import { GEMINI_TTS_MODELS, GeminiTTSVoice } from '../model-meta.js';
+import { TTSOptions, TTSResult } from '@tanstack/ai';
+import { GeminiClientConfig } from '../utils.js';
+/**
+ * Configuration for a single speaker in a multi-speaker dialogue.
+ * Supported by Gemini 3.1 Flash TTS Preview and the 2.5 TTS models.
+ */
+export interface GeminiSpeakerVoiceConfig {
+    /** A name used in the prompt to refer to this speaker */
+    speaker: string;
+    /** Voice configuration for this speaker */
+    voiceConfig: {
+        prebuiltVoiceConfig: {
+            voiceName: GeminiTTSVoice;
+        };
+    };
+}
+/**
+ * Provider-specific options for Gemini TTS
+ *
+ * @experimental Gemini TTS is an experimental feature.
+ * @see https://ai.google.dev/gemini-api/docs/speech-generation
+ * @see https://ai.google.dev/gemini-api/docs/models/gemini-3.1-flash-tts-preview
+ */
+export interface GeminiTTSProviderOptions {
+    /**
+     * Voice configuration for single-speaker TTS.
+     * Choose from 30 available voices with different characteristics.
+     *
+     * Use `multiSpeakerVoiceConfig` instead for dialogues.
+     */
+    voiceConfig?: {
+        prebuiltVoiceConfig?: {
+            /**
+             * The voice name to use for speech synthesis.
+             * @see https://ai.google.dev/gemini-api/docs/speech-generation#voices
+             */
+            voiceName?: GeminiTTSVoice;
+        };
+    };
+    /**
+     * Multi-speaker voice configuration (up to 2 speakers).
+     * Supported by Gemini 3.1 Flash TTS Preview and the 2.5 TTS models.
+     *
+     * Each speaker's lines in the prompt are prefixed with the name defined
+     * here, e.g.:
+     *
+     * ```text
+     * Joe: Hey, how's it going?
+     * Jane: Not bad, you?
+     * ```
+     */
+    multiSpeakerVoiceConfig?: {
+        speakerVoiceConfigs: Array<GeminiSpeakerVoiceConfig>;
+    };
+    /**
+     * System instruction for controlling speech style.
+     * Use natural language to describe the desired speaking style,
+     * pace, tone, accent, or other characteristics.
+     *
+     * With Gemini 3.1 Flash TTS, you can also use inline audio tags like
+     * `[whispering]`, `[laughs]`, `[excited]` directly in the input text
+     * to control delivery.
+     *
+     * @example "Speak slowly and calmly, as if telling a bedtime story"
+     * @example "Use an upbeat, enthusiastic tone with moderate pace"
+     * @example "Speak with a British accent"
+     */
+    systemInstruction?: string;
+    /**
+     * Language code hint for the speech synthesis.
+     * Gemini 3.1 Flash TTS supports 70+ languages with auto-detection;
+     * the 2.5 TTS models support 24 languages.
+     *
+     * @example "en-US" for American English
+     * @example "es-ES" for Spanish (Spain)
+     * @example "ja-JP" for Japanese
+     */
+    languageCode?: string;
+}
+/**
+ * Configuration for Gemini TTS adapter
+ *
+ * @experimental Gemini TTS is an experimental feature.
+ */
+export interface GeminiTTSConfig extends GeminiClientConfig {
+}
+/** Model type for Gemini TTS */
+export type GeminiTTSModel = (typeof GEMINI_TTS_MODELS)[number];
+/**
+ * Gemini Text-to-Speech Adapter
+ *
+ * Tree-shakeable adapter for Gemini TTS functionality.
+ *
+ * **IMPORTANT**: Gemini TTS uses the Live API (WebSocket-based) which requires
+ * different handling than traditional REST APIs. This adapter provides a
+ * simplified interface but may have limitations.
+ *
+ * @experimental Gemini TTS is an experimental feature and may change.
+ *
+ * Models:
+ * - gemini-2.5-flash-preview-tts
+ */
+export declare class GeminiTTSAdapter<TModel extends GeminiTTSModel> extends BaseTTSAdapter<TModel, GeminiTTSProviderOptions> {
+    readonly name: "gemini";
+    private readonly client;
+    constructor(config: GeminiTTSConfig, model: TModel);
+    /**
+     * Generate speech from text using Gemini's TTS model.
+     *
+     * @experimental This implementation is experimental and may change.
+     * @see https://ai.google.dev/gemini-api/docs/speech-generation
+     */
+    generateSpeech(options: TTSOptions<GeminiTTSProviderOptions>): Promise<TTSResult>;
+}
+/**
+ * Creates a Gemini TTS adapter with explicit API key.
+ * Type resolution happens here at the call site.
+ *
+ * @experimental Gemini TTS is an experimental feature and may change.
+ *
+ * @param model - The model name (e.g., 'gemini-2.5-flash-preview-tts')
+ * @param apiKey - Your Google API key
+ * @param config - Optional additional configuration
+ * @returns Configured Gemini TTS adapter instance with resolved types
+ *
+ * @example
+ * ```typescript
+ * const adapter = createGeminiSpeech('gemini-2.5-flash-preview-tts', "your-api-key");
+ *
+ * const result = await generateSpeech({
+ *   adapter,
+ *   text: 'Hello, world!'
+ * });
+ * ```
+ */
+export declare function createGeminiSpeech<TModel extends GeminiTTSModel>(model: TModel, apiKey: string, config?: Omit<GeminiTTSConfig, 'apiKey'>): GeminiTTSAdapter<TModel>;
+/**
+ * Creates a Gemini speech adapter with automatic API key detection from environment variables.
+ * Type resolution happens here at the call site.
+ *
+ * @experimental Gemini TTS is an experimental feature and may change.
+ *
+ * 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., 'gemini-2.5-flash-preview-tts')
+ * @param config - Optional configuration (excluding apiKey which is auto-detected)
+ * @returns Configured Gemini speech 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 = geminiSpeech('gemini-2.5-flash-preview-tts');
+ *
+ * const result = await generateSpeech({
+ *   adapter,
+ *   text: 'Welcome to TanStack AI!'
+ * });
+ * ```
+ */
+export declare function geminiSpeech<TModel extends GeminiTTSModel>(model: TModel, config?: Omit<GeminiTTSConfig, 'apiKey'>): GeminiTTSAdapter<TModel>;

package/dist/esm/adapters/tts.js

@@ -0,0 +1,192 @@
+import { BaseTTSAdapter } from "@tanstack/ai/adapters";
+import { createGeminiClient, generateId, getGeminiApiKeyFromEnv } from "../utils/client.js";
+import { GEMINI_TTS_VOICES } from "../model-meta.js";
+import { buildGeminiUsage } from "../usage.js";
+class GeminiTTSAdapter extends BaseTTSAdapter {
+  name = "gemini";
+  client;
+  constructor(config, model) {
+    super(model, config);
+    this.client = createGeminiClient(config);
+  }
+  /**
+   * Generate speech from text using Gemini's TTS model.
+   *
+   * @experimental This implementation is experimental and may change.
+   * @see https://ai.google.dev/gemini-api/docs/speech-generation
+   */
+  async generateSpeech(options) {
+    const { model, text, modelOptions, voice, logger } = options;
+    logger.request(`activity=generateSpeech provider=gemini model=${model}`, {
+      provider: "gemini",
+      model
+    });
+    const speechConfig = {};
+    if (modelOptions?.multiSpeakerVoiceConfig) {
+      const speakerConfigs = modelOptions.multiSpeakerVoiceConfig.speakerVoiceConfigs;
+      if (!Array.isArray(speakerConfigs) || speakerConfigs.length < 1 || speakerConfigs.length > 2) {
+        throw new Error(
+          `Gemini TTS multiSpeakerVoiceConfig.speakerVoiceConfigs must contain 1 or 2 speakers; received ${Array.isArray(speakerConfigs) ? speakerConfigs.length : "non-array"}.`
+        );
+      }
+      speechConfig.multiSpeakerVoiceConfig = modelOptions.multiSpeakerVoiceConfig;
+    } else {
+      if (voice !== void 0 && !GEMINI_TTS_VOICES.includes(voice)) {
+        throw new Error(
+          `Invalid Gemini TTS voice "${voice}". Valid voices are: ${GEMINI_TTS_VOICES.join(", ")}.`
+        );
+      }
+      const defaultVoiceName = voice ?? "Kore";
+      const supplied = modelOptions?.voiceConfig;
+      const resolvedVoiceName = supplied?.prebuiltVoiceConfig?.voiceName ?? defaultVoiceName;
+      speechConfig.voiceConfig = {
+        prebuiltVoiceConfig: { voiceName: resolvedVoiceName }
+      };
+    }
+    if (modelOptions?.languageCode) {
+      speechConfig.languageCode = modelOptions.languageCode;
+    }
+    try {
+      const response = await this.client.models.generateContent({
+        model,
+        contents: [
+          {
+            role: "user",
+            parts: [{ text }]
+          }
+        ],
+        config: {
+          responseModalities: ["AUDIO"],
+          speechConfig,
+          // systemInstruction belongs inside `config` per the @google/genai
+          // contract — matches sibling Gemini adapters (summarize, text).
+          ...modelOptions?.systemInstruction && {
+            systemInstruction: modelOptions.systemInstruction
+          }
+        }
+      });
+      const candidate = response.candidates?.[0];
+      const parts = candidate?.content?.parts;
+      if (!parts || parts.length === 0) {
+        throw new Error("No audio output received from Gemini TTS");
+      }
+      const audioPart = parts.find(
+        (part) => part.inlineData?.mimeType?.startsWith("audio/")
+      );
+      if (!audioPart || !audioPart.inlineData || !audioPart.inlineData.data) {
+        throw new Error("No audio data in Gemini TTS response");
+      }
+      const audioBase64 = audioPart.inlineData.data;
+      const mimeType = audioPart.inlineData.mimeType;
+      const usageField = response.usageMetadata ? { usage: buildGeminiUsage(response.usageMetadata) } : {};
+      const pcm = parsePcmMimeType(mimeType);
+      if (pcm) {
+        const wavBase64 = wrapPcmBase64AsWav(
+          audioBase64,
+          pcm.sampleRate,
+          pcm.channels,
+          pcm.bitsPerSample
+        );
+        return {
+          id: generateId(this.name),
+          model,
+          audio: wavBase64,
+          format: "wav",
+          contentType: "audio/wav",
+          ...usageField
+        };
+      }
+      const format = (mimeType.split(";")[0] ?? "").split("/")[1] || "wav";
+      return {
+        id: generateId(this.name),
+        model,
+        audio: audioBase64,
+        format,
+        contentType: mimeType,
+        ...usageField
+      };
+    } catch (error) {
+      logger.errors("gemini.generateSpeech fatal", {
+        error,
+        source: "gemini.generateSpeech"
+      });
+      throw error;
+    }
+  }
+}
+function parsePcmMimeType(mimeType) {
+  const normalized = mimeType.toLowerCase();
+  const subtype = (normalized.split(";")[0] ?? "").split("/")[1] ?? "";
+  if (subtype.includes("wav")) return void 0;
+  const bitDepthMatch = /^audio\/l(\d+)/.exec(normalized);
+  const isPcm = bitDepthMatch !== null || normalized.startsWith("audio/pcm") || normalized.startsWith("audio/x-pcm") || normalized.includes("codec=pcm");
+  if (!isPcm) return void 0;
+  const rateMatch = /rate=(\d+)/.exec(normalized);
+  const channelsMatch = /channels=(\d+)/.exec(normalized);
+  const bitsPerSample = bitDepthMatch ? Number(bitDepthMatch[1]) : 16;
+  return {
+    sampleRate: rateMatch ? Number(rateMatch[1]) : 24e3,
+    channels: channelsMatch ? Number(channelsMatch[1]) : 1,
+    bitsPerSample
+  };
+}
+function wrapPcmBase64AsWav(pcmBase64, sampleRate, channels = 1, bitsPerSample = 16) {
+  if (bitsPerSample !== 16) {
+    throw new Error(
+      `Unsupported PCM bit depth ${bitsPerSample}: only 16-bit PCM can be wrapped as WAV.`
+    );
+  }
+  const pcmBytes = typeof Buffer !== "undefined" ? new Uint8Array(Buffer.from(pcmBase64, "base64")) : decodeBase64(pcmBase64);
+  const byteRate = sampleRate * channels * bitsPerSample / 8;
+  const blockAlign = channels * bitsPerSample / 8;
+  const dataSize = pcmBytes.byteLength;
+  const buffer = new ArrayBuffer(44 + dataSize);
+  const view = new DataView(buffer);
+  writeAscii(view, 0, "RIFF");
+  view.setUint32(4, 36 + dataSize, true);
+  writeAscii(view, 8, "WAVE");
+  writeAscii(view, 12, "fmt ");
+  view.setUint32(16, 16, true);
+  view.setUint16(20, 1, true);
+  view.setUint16(22, channels, true);
+  view.setUint32(24, sampleRate, true);
+  view.setUint32(28, byteRate, true);
+  view.setUint16(32, blockAlign, true);
+  view.setUint16(34, bitsPerSample, true);
+  writeAscii(view, 36, "data");
+  view.setUint32(40, dataSize, true);
+  new Uint8Array(buffer, 44).set(pcmBytes);
+  if (typeof Buffer !== "undefined") {
+    return Buffer.from(buffer).toString("base64");
+  }
+  let binary = "";
+  const bytes = new Uint8Array(buffer);
+  for (const byte of bytes) {
+    binary += String.fromCharCode(byte);
+  }
+  return btoa(binary);
+}
+function decodeBase64(b64) {
+  const binary = atob(b64);
+  const out = new Uint8Array(binary.length);
+  for (let i = 0; i < binary.length; i += 1) out[i] = binary.charCodeAt(i);
+  return out;
+}
+function writeAscii(view, offset, text) {
+  for (let i = 0; i < text.length; i += 1) {
+    view.setUint8(offset + i, text.charCodeAt(i));
+  }
+}
+function createGeminiSpeech(model, apiKey, config) {
+  return new GeminiTTSAdapter({ ...config, apiKey }, model);
+}
+function geminiSpeech(model, config) {
+  const apiKey = getGeminiApiKeyFromEnv();
+  return createGeminiSpeech(model, apiKey, config);
+}
+export {
+  GeminiTTSAdapter,
+  createGeminiSpeech,
+  geminiSpeech
+};
+//# sourceMappingURL=tts.js.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"tts.js","sources":["../../../src/adapters/tts.ts"],"sourcesContent":["import { BaseTTSAdapter } from '@tanstack/ai/adapters'\nimport {\n  createGeminiClient,\n  generateId,\n  getGeminiApiKeyFromEnv,\n} from '../utils'\nimport { GEMINI_TTS_VOICES } from '../model-meta'\nimport { buildGeminiUsage } from '../usage'\nimport type { GEMINI_TTS_MODELS, GeminiTTSVoice } from '../model-meta'\nimport type { TTSOptions, TTSResult } from '@tanstack/ai'\nimport type { GoogleGenAI, SpeechConfig } from '@google/genai'\nimport type { GeminiClientConfig } from '../utils'\n\n/**\n * Configuration for a single speaker in a multi-speaker dialogue.\n * Supported by Gemini 3.1 Flash TTS Preview and the 2.5 TTS models.\n */\nexport interface GeminiSpeakerVoiceConfig {\n  /** A name used in the prompt to refer to this speaker */\n  speaker: string\n  /** Voice configuration for this speaker */\n  voiceConfig: {\n    prebuiltVoiceConfig: {\n      voiceName: GeminiTTSVoice\n    }\n  }\n}\n\n/**\n * Provider-specific options for Gemini TTS\n *\n * @experimental Gemini TTS is an experimental feature.\n * @see https://ai.google.dev/gemini-api/docs/speech-generation\n * @see https://ai.google.dev/gemini-api/docs/models/gemini-3.1-flash-tts-preview\n */\nexport interface GeminiTTSProviderOptions {\n  /**\n   * Voice configuration for single-speaker TTS.\n   * Choose from 30 available voices with different characteristics.\n   *\n   * Use `multiSpeakerVoiceConfig` instead for dialogues.\n   */\n  voiceConfig?: {\n    prebuiltVoiceConfig?: {\n      /**\n       * The voice name to use for speech synthesis.\n       * @see https://ai.google.dev/gemini-api/docs/speech-generation#voices\n       */\n      voiceName?: GeminiTTSVoice\n    }\n  }\n\n  /**\n   * Multi-speaker voice configuration (up to 2 speakers).\n   * Supported by Gemini 3.1 Flash TTS Preview and the 2.5 TTS models.\n   *\n   * Each speaker's lines in the prompt are prefixed with the name defined\n   * here, e.g.:\n   *\n   * ```text\n   * Joe: Hey, how's it going?\n   * Jane: Not bad, you?\n   * ```\n   */\n  multiSpeakerVoiceConfig?: {\n    speakerVoiceConfigs: Array<GeminiSpeakerVoiceConfig>\n  }\n\n  /**\n   * System instruction for controlling speech style.\n   * Use natural language to describe the desired speaking style,\n   * pace, tone, accent, or other characteristics.\n   *\n   * With Gemini 3.1 Flash TTS, you can also use inline audio tags like\n   * `[whispering]`, `[laughs]`, `[excited]` directly in the input text\n   * to control delivery.\n   *\n   * @example \"Speak slowly and calmly, as if telling a bedtime story\"\n   * @example \"Use an upbeat, enthusiastic tone with moderate pace\"\n   * @example \"Speak with a British accent\"\n   */\n  systemInstruction?: string\n\n  /**\n   * Language code hint for the speech synthesis.\n   * Gemini 3.1 Flash TTS supports 70+ languages with auto-detection;\n   * the 2.5 TTS models support 24 languages.\n   *\n   * @example \"en-US\" for American English\n   * @example \"es-ES\" for Spanish (Spain)\n   * @example \"ja-JP\" for Japanese\n   */\n  languageCode?: string\n}\n\n/**\n * Configuration for Gemini TTS adapter\n *\n * @experimental Gemini TTS is an experimental feature.\n */\nexport interface GeminiTTSConfig extends GeminiClientConfig {}\n\n/** Model type for Gemini TTS */\nexport type GeminiTTSModel = (typeof GEMINI_TTS_MODELS)[number]\n\n/**\n * Gemini Text-to-Speech Adapter\n *\n * Tree-shakeable adapter for Gemini TTS functionality.\n *\n * **IMPORTANT**: Gemini TTS uses the Live API (WebSocket-based) which requires\n * different handling than traditional REST APIs. This adapter provides a\n * simplified interface but may have limitations.\n *\n * @experimental Gemini TTS is an experimental feature and may change.\n *\n * Models:\n * - gemini-2.5-flash-preview-tts\n */\nexport class GeminiTTSAdapter<\n  TModel extends GeminiTTSModel,\n> extends BaseTTSAdapter<TModel, GeminiTTSProviderOptions> {\n  readonly name = 'gemini' as const\n\n  private readonly client: GoogleGenAI\n\n  constructor(config: GeminiTTSConfig, model: TModel) {\n    super(model, config)\n    this.client = createGeminiClient(config)\n  }\n\n  /**\n   * Generate speech from text using Gemini's TTS model.\n   *\n   * @experimental This implementation is experimental and may change.\n   * @see https://ai.google.dev/gemini-api/docs/speech-generation\n   */\n  async generateSpeech(\n    options: TTSOptions<GeminiTTSProviderOptions>,\n  ): Promise<TTSResult> {\n    const { model, text, modelOptions, voice, logger } = options\n\n    logger.request(`activity=generateSpeech provider=gemini model=${model}`, {\n      provider: 'gemini',\n      model,\n    })\n\n    const speechConfig: SpeechConfig = {}\n\n    if (modelOptions?.multiSpeakerVoiceConfig) {\n      // Validate multi-speaker config: 1 or 2 speakers allowed.\n      const speakerConfigs =\n        modelOptions.multiSpeakerVoiceConfig.speakerVoiceConfigs\n      if (\n        !Array.isArray(speakerConfigs) ||\n        speakerConfigs.length < 1 ||\n        speakerConfigs.length > 2\n      ) {\n        throw new Error(\n          `Gemini TTS multiSpeakerVoiceConfig.speakerVoiceConfigs must contain 1 or 2 speakers; received ${Array.isArray(speakerConfigs) ? speakerConfigs.length : 'non-array'}.`,\n        )\n      }\n      speechConfig.multiSpeakerVoiceConfig =\n        modelOptions.multiSpeakerVoiceConfig\n    } else {\n      // Honor the standard TTSOptions.voice (used by every other TTS adapter)\n      // as a fallback for the prebuilt voice name. If an explicit\n      // modelOptions.voiceConfig is supplied its values win — but we still\n      // fall back to `voice` / 'Kore' if the supplied voiceConfig is missing\n      // prebuiltVoiceConfig.voiceName.\n      if (\n        voice !== undefined &&\n        !(GEMINI_TTS_VOICES as ReadonlyArray<string>).includes(voice)\n      ) {\n        throw new Error(\n          `Invalid Gemini TTS voice \"${voice}\". Valid voices are: ${GEMINI_TTS_VOICES.join(', ')}.`,\n        )\n      }\n      const defaultVoiceName = (voice as GeminiTTSVoice | undefined) ?? 'Kore'\n      const supplied = modelOptions?.voiceConfig\n      const resolvedVoiceName =\n        supplied?.prebuiltVoiceConfig?.voiceName ?? defaultVoiceName\n      speechConfig.voiceConfig = {\n        prebuiltVoiceConfig: { voiceName: resolvedVoiceName },\n      }\n    }\n\n    if (modelOptions?.languageCode) {\n      speechConfig.languageCode = modelOptions.languageCode\n    }\n\n    try {\n      const response = await this.client.models.generateContent({\n        model,\n        contents: [\n          {\n            role: 'user',\n            parts: [{ text }],\n          },\n        ],\n        config: {\n          responseModalities: ['AUDIO'],\n          speechConfig,\n          // systemInstruction belongs inside `config` per the @google/genai\n          // contract — matches sibling Gemini adapters (summarize, text).\n          ...(modelOptions?.systemInstruction && {\n            systemInstruction: modelOptions.systemInstruction,\n          }),\n        },\n      })\n\n      // Extract audio data from response\n      const candidate = response.candidates?.[0]\n      const parts = candidate?.content?.parts\n\n      if (!parts || parts.length === 0) {\n        throw new Error('No audio output received from Gemini TTS')\n      }\n\n      // Look for inline data (audio)\n      const audioPart = parts.find((part: any) =>\n        part.inlineData?.mimeType?.startsWith('audio/'),\n      )\n\n      if (!audioPart || !audioPart.inlineData || !audioPart.inlineData.data) {\n        throw new Error('No audio data in Gemini TTS response')\n      }\n\n      const audioBase64 = audioPart.inlineData.data\n      // mime is guaranteed by the `startsWith('audio/')` find predicate above.\n      const mimeType = audioPart.inlineData.mimeType as string\n\n      // Surface token usage (with per-modality breakdown) when Gemini reports\n      // it. Spread conditionally for exactOptionalPropertyTypes — shared by both\n      // the PCM→WAV and pass-through return paths below.\n      const usageField = response.usageMetadata\n        ? { usage: buildGeminiUsage(response.usageMetadata) }\n        : {}\n\n      // Gemini TTS models return raw 16-bit LE PCM with a mime type like\n      // `audio/L16;codec=pcm;rate=24000`. That isn't playable in an <audio>\n      // element and the bare string isn't a usable file extension, so we\n      // prepend a RIFF/WAV header here and normalize the result to audio/wav.\n      const pcm = parsePcmMimeType(mimeType)\n      if (pcm) {\n        const wavBase64 = wrapPcmBase64AsWav(\n          audioBase64,\n          pcm.sampleRate,\n          pcm.channels,\n          pcm.bitsPerSample,\n        )\n        return {\n          id: generateId(this.name),\n          model,\n          audio: wavBase64,\n          format: 'wav',\n          contentType: 'audio/wav',\n          ...usageField,\n        }\n      }\n\n      // Strip any mime parameters (e.g. `audio/ogg;codec=opus`) before pulling\n      // the subtype out as the file format. `String.split` always returns at\n      // least one element, so [0] is defined.\n      const format = (mimeType.split(';')[0] ?? '').split('/')[1] || 'wav'\n\n      return {\n        id: generateId(this.name),\n        model,\n        audio: audioBase64,\n        format,\n        contentType: mimeType,\n        ...usageField,\n      }\n    } catch (error) {\n      logger.errors('gemini.generateSpeech fatal', {\n        error,\n        source: 'gemini.generateSpeech',\n      })\n      throw error\n    }\n  }\n}\n\nfunction parsePcmMimeType(\n  mimeType: string,\n): { sampleRate: number; channels: number; bitsPerSample: number } | undefined {\n  const normalized = mimeType.toLowerCase()\n  const subtype = (normalized.split(';')[0] ?? '').split('/')[1] ?? ''\n  // Exclude containerized wav (e.g. `audio/wav;codec=pcm`) — those already\n  // carry a RIFF header and must not be re-wrapped.\n  if (subtype.includes('wav')) return undefined\n\n  // Accept the variants Gemini and other providers actually emit:\n  //   - audio/L16;codec=pcm;rate=24000 (IANA PCM with bit depth in the type)\n  //   - audio/L24 and friends\n  //   - audio/pcm and audio/x-pcm\n  //   - anything else that explicitly tags codec=pcm and isn't wav-containered\n  const bitDepthMatch = /^audio\\/l(\\d+)/.exec(normalized)\n  const isPcm =\n    bitDepthMatch !== null ||\n    normalized.startsWith('audio/pcm') ||\n    normalized.startsWith('audio/x-pcm') ||\n    normalized.includes('codec=pcm')\n  if (!isPcm) return undefined\n\n  const rateMatch = /rate=(\\d+)/.exec(normalized)\n  const channelsMatch = /channels=(\\d+)/.exec(normalized)\n  // Default to 16-bit when the mime type doesn't specify — matches Gemini's\n  // audio/L16;codec=pcm;rate=24000 response.\n  const bitsPerSample = bitDepthMatch ? Number(bitDepthMatch[1]) : 16\n  return {\n    sampleRate: rateMatch ? Number(rateMatch[1]) : 24000,\n    channels: channelsMatch ? Number(channelsMatch[1]) : 1,\n    bitsPerSample,\n  }\n}\n\nfunction wrapPcmBase64AsWav(\n  pcmBase64: string,\n  sampleRate: number,\n  channels = 1,\n  bitsPerSample = 16,\n): string {\n  // The WAV writer below emits a 16-bit PCM fmt chunk. If the source claims a\n  // different bit depth we'd be lying about the payload, so bail out loudly\n  // rather than producing a corrupt file.\n  if (bitsPerSample !== 16) {\n    throw new Error(\n      `Unsupported PCM bit depth ${bitsPerSample}: only 16-bit PCM can be wrapped as WAV.`,\n    )\n  }\n\n  const pcmBytes =\n    typeof Buffer !== 'undefined'\n      ? new Uint8Array(Buffer.from(pcmBase64, 'base64'))\n      : decodeBase64(pcmBase64)\n\n  const byteRate = (sampleRate * channels * bitsPerSample) / 8\n  const blockAlign = (channels * bitsPerSample) / 8\n  const dataSize = pcmBytes.byteLength\n  const buffer = new ArrayBuffer(44 + dataSize)\n  const view = new DataView(buffer)\n\n  writeAscii(view, 0, 'RIFF')\n  view.setUint32(4, 36 + dataSize, true)\n  writeAscii(view, 8, 'WAVE')\n  writeAscii(view, 12, 'fmt ')\n  view.setUint32(16, 16, true)\n  view.setUint16(20, 1, true)\n  view.setUint16(22, channels, true)\n  view.setUint32(24, sampleRate, true)\n  view.setUint32(28, byteRate, true)\n  view.setUint16(32, blockAlign, true)\n  view.setUint16(34, bitsPerSample, true)\n  writeAscii(view, 36, 'data')\n  view.setUint32(40, dataSize, true)\n  new Uint8Array(buffer, 44).set(pcmBytes)\n\n  if (typeof Buffer !== 'undefined') {\n    return Buffer.from(buffer).toString('base64')\n  }\n  let binary = ''\n  const bytes = new Uint8Array(buffer)\n  for (const byte of bytes) {\n    binary += String.fromCharCode(byte)\n  }\n  return btoa(binary)\n}\n\nfunction decodeBase64(b64: string): Uint8Array {\n  const binary = atob(b64)\n  const out = new Uint8Array(binary.length)\n  for (let i = 0; i < binary.length; i += 1) out[i] = binary.charCodeAt(i)\n  return out\n}\n\nfunction writeAscii(view: DataView, offset: number, text: string): void {\n  for (let i = 0; i < text.length; i += 1) {\n    view.setUint8(offset + i, text.charCodeAt(i))\n  }\n}\n\n/**\n * Creates a Gemini TTS adapter with explicit API key.\n * Type resolution happens here at the call site.\n *\n * @experimental Gemini TTS is an experimental feature and may change.\n *\n * @param model - The model name (e.g., 'gemini-2.5-flash-preview-tts')\n * @param apiKey - Your Google API key\n * @param config - Optional additional configuration\n * @returns Configured Gemini TTS adapter instance with resolved types\n *\n * @example\n * ```typescript\n * const adapter = createGeminiSpeech('gemini-2.5-flash-preview-tts', \"your-api-key\");\n *\n * const result = await generateSpeech({\n *   adapter,\n *   text: 'Hello, world!'\n * });\n * ```\n */\nexport function createGeminiSpeech<TModel extends GeminiTTSModel>(\n  model: TModel,\n  apiKey: string,\n  config?: Omit<GeminiTTSConfig, 'apiKey'>,\n): GeminiTTSAdapter<TModel> {\n  // Put apiKey LAST so caller-supplied config can't silently override the\n  // explicit argument.\n  return new GeminiTTSAdapter({ ...config, apiKey }, model)\n}\n\n/**\n * Creates a Gemini speech adapter with automatic API key detection from environment variables.\n * Type resolution happens here at the call site.\n *\n * @experimental Gemini TTS is an experimental feature and may change.\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., 'gemini-2.5-flash-preview-tts')\n * @param config - Optional configuration (excluding apiKey which is auto-detected)\n * @returns Configured Gemini speech 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 = geminiSpeech('gemini-2.5-flash-preview-tts');\n *\n * const result = await generateSpeech({\n *   adapter,\n *   text: 'Welcome to TanStack AI!'\n * });\n * ```\n */\nexport function geminiSpeech<TModel extends GeminiTTSModel>(\n  model: TModel,\n  config?: Omit<GeminiTTSConfig, 'apiKey'>,\n): GeminiTTSAdapter<TModel> {\n  const apiKey = getGeminiApiKeyFromEnv()\n  return createGeminiSpeech(model, apiKey, config)\n}\n"],"names":[],"mappings":";;;;AAuHO,MAAM,yBAEH,eAAiD;AAAA,EAChD,OAAO;AAAA,EAEC;AAAA,EAEjB,YAAY,QAAyB,OAAe;AAClD,UAAM,OAAO,MAAM;AACnB,SAAK,SAAS,mBAAmB,MAAM;AAAA,EACzC;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAQA,MAAM,eACJ,SACoB;AACpB,UAAM,EAAE,OAAO,MAAM,cAAc,OAAO,WAAW;AAErD,WAAO,QAAQ,iDAAiD,KAAK,IAAI;AAAA,MACvE,UAAU;AAAA,MACV;AAAA,IAAA,CACD;AAED,UAAM,eAA6B,CAAA;AAEnC,QAAI,cAAc,yBAAyB;AAEzC,YAAM,iBACJ,aAAa,wBAAwB;AACvC,UACE,CAAC,MAAM,QAAQ,cAAc,KAC7B,eAAe,SAAS,KACxB,eAAe,SAAS,GACxB;AACA,cAAM,IAAI;AAAA,UACR,iGAAiG,MAAM,QAAQ,cAAc,IAAI,eAAe,SAAS,WAAW;AAAA,QAAA;AAAA,MAExK;AACA,mBAAa,0BACX,aAAa;AAAA,IACjB,OAAO;AAML,UACE,UAAU,UACV,CAAE,kBAA4C,SAAS,KAAK,GAC5D;AACA,cAAM,IAAI;AAAA,UACR,6BAA6B,KAAK,wBAAwB,kBAAkB,KAAK,IAAI,CAAC;AAAA,QAAA;AAAA,MAE1F;AACA,YAAM,mBAAoB,SAAwC;AAClE,YAAM,WAAW,cAAc;AAC/B,YAAM,oBACJ,UAAU,qBAAqB,aAAa;AAC9C,mBAAa,cAAc;AAAA,QACzB,qBAAqB,EAAE,WAAW,kBAAA;AAAA,MAAkB;AAAA,IAExD;AAEA,QAAI,cAAc,cAAc;AAC9B,mBAAa,eAAe,aAAa;AAAA,IAC3C;AAEA,QAAI;AACF,YAAM,WAAW,MAAM,KAAK,OAAO,OAAO,gBAAgB;AAAA,QACxD;AAAA,QACA,UAAU;AAAA,UACR;AAAA,YACE,MAAM;AAAA,YACN,OAAO,CAAC,EAAE,KAAA,CAAM;AAAA,UAAA;AAAA,QAClB;AAAA,QAEF,QAAQ;AAAA,UACN,oBAAoB,CAAC,OAAO;AAAA,UAC5B;AAAA;AAAA;AAAA,UAGA,GAAI,cAAc,qBAAqB;AAAA,YACrC,mBAAmB,aAAa;AAAA,UAAA;AAAA,QAClC;AAAA,MACF,CACD;AAGD,YAAM,YAAY,SAAS,aAAa,CAAC;AACzC,YAAM,QAAQ,WAAW,SAAS;AAElC,UAAI,CAAC,SAAS,MAAM,WAAW,GAAG;AAChC,cAAM,IAAI,MAAM,0CAA0C;AAAA,MAC5D;AAGA,YAAM,YAAY,MAAM;AAAA,QAAK,CAAC,SAC5B,KAAK,YAAY,UAAU,WAAW,QAAQ;AAAA,MAAA;AAGhD,UAAI,CAAC,aAAa,CAAC,UAAU,cAAc,CAAC,UAAU,WAAW,MAAM;AACrE,cAAM,IAAI,MAAM,sCAAsC;AAAA,MACxD;AAEA,YAAM,cAAc,UAAU,WAAW;AAEzC,YAAM,WAAW,UAAU,WAAW;AAKtC,YAAM,aAAa,SAAS,gBACxB,EAAE,OAAO,iBAAiB,SAAS,aAAa,EAAA,IAChD,CAAA;AAMJ,YAAM,MAAM,iBAAiB,QAAQ;AACrC,UAAI,KAAK;AACP,cAAM,YAAY;AAAA,UAChB;AAAA,UACA,IAAI;AAAA,UACJ,IAAI;AAAA,UACJ,IAAI;AAAA,QAAA;AAEN,eAAO;AAAA,UACL,IAAI,WAAW,KAAK,IAAI;AAAA,UACxB;AAAA,UACA,OAAO;AAAA,UACP,QAAQ;AAAA,UACR,aAAa;AAAA,UACb,GAAG;AAAA,QAAA;AAAA,MAEP;AAKA,YAAM,UAAU,SAAS,MAAM,GAAG,EAAE,CAAC,KAAK,IAAI,MAAM,GAAG,EAAE,CAAC,KAAK;AAE/D,aAAO;AAAA,QACL,IAAI,WAAW,KAAK,IAAI;AAAA,QACxB;AAAA,QACA,OAAO;AAAA,QACP;AAAA,QACA,aAAa;AAAA,QACb,GAAG;AAAA,MAAA;AAAA,IAEP,SAAS,OAAO;AACd,aAAO,OAAO,+BAA+B;AAAA,QAC3C;AAAA,QACA,QAAQ;AAAA,MAAA,CACT;AACD,YAAM;AAAA,IACR;AAAA,EACF;AACF;AAEA,SAAS,iBACP,UAC6E;AAC7E,QAAM,aAAa,SAAS,YAAA;AAC5B,QAAM,WAAW,WAAW,MAAM,GAAG,EAAE,CAAC,KAAK,IAAI,MAAM,GAAG,EAAE,CAAC,KAAK;AAGlE,MAAI,QAAQ,SAAS,KAAK,EAAG,QAAO;AAOpC,QAAM,gBAAgB,iBAAiB,KAAK,UAAU;AACtD,QAAM,QACJ,kBAAkB,QAClB,WAAW,WAAW,WAAW,KACjC,WAAW,WAAW,aAAa,KACnC,WAAW,SAAS,WAAW;AACjC,MAAI,CAAC,MAAO,QAAO;AAEnB,QAAM,YAAY,aAAa,KAAK,UAAU;AAC9C,QAAM,gBAAgB,iBAAiB,KAAK,UAAU;AAGtD,QAAM,gBAAgB,gBAAgB,OAAO,cAAc,CAAC,CAAC,IAAI;AACjE,SAAO;AAAA,IACL,YAAY,YAAY,OAAO,UAAU,CAAC,CAAC,IAAI;AAAA,IAC/C,UAAU,gBAAgB,OAAO,cAAc,CAAC,CAAC,IAAI;AAAA,IACrD;AAAA,EAAA;AAEJ;AAEA,SAAS,mBACP,WACA,YACA,WAAW,GACX,gBAAgB,IACR;AAIR,MAAI,kBAAkB,IAAI;AACxB,UAAM,IAAI;AAAA,MACR,6BAA6B,aAAa;AAAA,IAAA;AAAA,EAE9C;AAEA,QAAM,WACJ,OAAO,WAAW,cACd,IAAI,WAAW,OAAO,KAAK,WAAW,QAAQ,CAAC,IAC/C,aAAa,SAAS;AAE5B,QAAM,WAAY,aAAa,WAAW,gBAAiB;AAC3D,QAAM,aAAc,WAAW,gBAAiB;AAChD,QAAM,WAAW,SAAS;AAC1B,QAAM,SAAS,IAAI,YAAY,KAAK,QAAQ;AAC5C,QAAM,OAAO,IAAI,SAAS,MAAM;AAEhC,aAAW,MAAM,GAAG,MAAM;AAC1B,OAAK,UAAU,GAAG,KAAK,UAAU,IAAI;AACrC,aAAW,MAAM,GAAG,MAAM;AAC1B,aAAW,MAAM,IAAI,MAAM;AAC3B,OAAK,UAAU,IAAI,IAAI,IAAI;AAC3B,OAAK,UAAU,IAAI,GAAG,IAAI;AAC1B,OAAK,UAAU,IAAI,UAAU,IAAI;AACjC,OAAK,UAAU,IAAI,YAAY,IAAI;AACnC,OAAK,UAAU,IAAI,UAAU,IAAI;AACjC,OAAK,UAAU,IAAI,YAAY,IAAI;AACnC,OAAK,UAAU,IAAI,eAAe,IAAI;AACtC,aAAW,MAAM,IAAI,MAAM;AAC3B,OAAK,UAAU,IAAI,UAAU,IAAI;AACjC,MAAI,WAAW,QAAQ,EAAE,EAAE,IAAI,QAAQ;AAEvC,MAAI,OAAO,WAAW,aAAa;AACjC,WAAO,OAAO,KAAK,MAAM,EAAE,SAAS,QAAQ;AAAA,EAC9C;AACA,MAAI,SAAS;AACb,QAAM,QAAQ,IAAI,WAAW,MAAM;AACnC,aAAW,QAAQ,OAAO;AACxB,cAAU,OAAO,aAAa,IAAI;AAAA,EACpC;AACA,SAAO,KAAK,MAAM;AACpB;AAEA,SAAS,aAAa,KAAyB;AAC7C,QAAM,SAAS,KAAK,GAAG;AACvB,QAAM,MAAM,IAAI,WAAW,OAAO,MAAM;AACxC,WAAS,IAAI,GAAG,IAAI,OAAO,QAAQ,KAAK,EAAG,KAAI,CAAC,IAAI,OAAO,WAAW,CAAC;AACvE,SAAO;AACT;AAEA,SAAS,WAAW,MAAgB,QAAgB,MAAoB;AACtE,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK,GAAG;AACvC,SAAK,SAAS,SAAS,GAAG,KAAK,WAAW,CAAC,CAAC;AAAA,EAC9C;AACF;AAuBO,SAAS,mBACd,OACA,QACA,QAC0B;AAG1B,SAAO,IAAI,iBAAiB,EAAE,GAAG,QAAQ,OAAA,GAAU,KAAK;AAC1D;AA4BO,SAAS,aACd,OACA,QAC0B;AAC1B,QAAM,SAAS,uBAAA;AACf,SAAO,mBAAmB,OAAO,QAAQ,MAAM;AACjD;"}

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>;