@tyvm/knowhow 0.0.158 → 0.0.160

package/src/clients/gemini.ts

@@ -1,1169 +0,0 @@
-import {
-  GoogleGenAI,
-  Part,
-  Content,
-  Tool as GoogleTool, // Rename to avoid conflict with your Tool type
-  FunctionDeclaration,
-  FunctionCallingConfigMode,
-  GenerationConfig,
-  ToolConfig,
-  UsageMetadata,
-} from "@google/genai";
-import * as os from "os";
-import * as fsSync from "fs";
-import * as pathSync from "path";
-import { wait } from "../utils";
-import {
-  EmbeddingModels,
-  Models,
-  GoogleThinkingLevelModels,
-  GoogleThinkingBudgetModels,
-  GoogleImageModels,
-  GoogleVideoModels,
-  GoogleTTSModels,
-  GoogleEmbeddingModelsList,
-  GoogleReasoningModels,
-} from "../types";
-import { GeminiTextPricing } from "./pricing";
-import { ContextLimits } from "./contextLimits";
-import { ModelModality, TokenUsage } from "./types";
-
-import {
-  GenericClient,
-  CompletionOptions,
-  CompletionResponse,
-  EmbeddingOptions,
-  EmbeddingResponse,
-  Tool,
-  Message,
-  MessageContent,
-  ToolCall,
-  OutputMessage,
-  AudioTranscriptionOptions,
-  AudioTranscriptionResponse,
-  AudioGenerationOptions,
-  AudioGenerationResponse,
-  ImageGenerationOptions,
-  ImageGenerationResponse,
-  VideoGenerationOptions,
-  VideoGenerationResponse,
-  VideoStatusOptions,
-  VideoStatusResponse,
-  FileUploadOptions,
-  FileUploadResponse,
-  FileDownloadOptions,
-  FileDownloadResponse,
-} from "./types";
-
-function getMimeTypeFromUrl(url: string): string {
-  if (url.endsWith(".png")) return "image/png";
-  if (url.endsWith(".gif")) return "image/gif";
-  if (url.endsWith(".webp")) return "image/webp";
-  return "image/jpeg";
-}
-
-function getVideoMimeTypeFromUrl(url: string): string {
-  if (url.endsWith(".mp4")) return "video/mp4";
-  if (url.endsWith(".webm")) return "video/webm";
-  if (url.endsWith(".mov")) return "video/quicktime";
-  if (url.endsWith(".avi")) return "video/x-msvideo";
-  return "video/mp4";
-}
-
-/**
- * Converts raw PCM audio data to WAV format by prepending a WAV header.
- * Gemini TTS returns raw 16-bit PCM (audio/L16) which needs a WAV header to be playable.
- */
-function pcmToWav(
-  pcmData: Buffer,
-  sampleRate: number = 24000,
-  numChannels: number = 1,
-  bitsPerSample: number = 16
-): Buffer {
-  const dataSize = pcmData.length;
-  const headerSize = 44;
-  const wavBuffer = Buffer.alloc(headerSize + dataSize);
-
-  // RIFF header
-  wavBuffer.write("RIFF", 0);
-  wavBuffer.writeUInt32LE(36 + dataSize, 4); // file size - 8
-  wavBuffer.write("WAVE", 8);
-
-  // fmt chunk
-  wavBuffer.write("fmt ", 12);
-  wavBuffer.writeUInt32LE(16, 16); // chunk size
-  wavBuffer.writeUInt16LE(1, 20); // PCM format
-  wavBuffer.writeUInt16LE(numChannels, 22);
-  wavBuffer.writeUInt32LE(sampleRate, 24);
-  wavBuffer.writeUInt32LE(sampleRate * numChannels * (bitsPerSample / 8), 28); // byte rate
-  wavBuffer.writeUInt16LE(numChannels * (bitsPerSample / 8), 32); // block align
-  wavBuffer.writeUInt16LE(bitsPerSample, 34);
-
-  // data chunk
-  wavBuffer.write("data", 36);
-  wavBuffer.writeUInt32LE(dataSize, 40);
-  pcmData.copy(wavBuffer, 44);
-
-  return wavBuffer;
-}
-
-export class GenericGeminiClient implements GenericClient {
-  private client: GoogleGenAI;
-  private apiKey?: string;
-
-  constructor(apiKey?: string) {
-    this.setKey(apiKey || process.env.GEMINI_API_KEY || "");
-  }
-
-  setKey(apiKey: string) {
-    this.apiKey = apiKey;
-    this.client = new GoogleGenAI({
-      apiKey: apiKey || process.env.GEMINI_API_KEY,
-    });
-  }
-
-  /**
-   * Transforms a generic MessageContent array or string into Google GenAI ContentPart array.
-   * Handles text and image_url types.
-   * @param content The generic message content.
-   * @returns An array of Google GenAI ContentPart.
-   */
-  transformContentParts(content: string | MessageContent[]): Part[] {
-    if (typeof content === "string") {
-      return [{ text: content }];
-    }
-
-    return content
-      .map((part) => {
-        if (part.type === "text") {
-          return { text: part.text };
-        }
-        if (part.type === "image_url") {
-          const url = part.image_url.url;
-          if (url.startsWith("data:")) {
-            const [header, base64Data] = url.split(",");
-            const mimeType = header.split(":")[1].split(";")[0];
-            return {
-              inlineData: {
-                data: base64Data,
-                mimeType,
-              },
-            };
-          }
-
-          // If it's a File API URI
-          if (url.startsWith("https://generativelanguage.googleapis.com")) {
-            return {
-              fileData: {
-                fileUri: url,
-                mimeType: getMimeTypeFromUrl(url),
-              },
-            };
-          }
-        }
-        if (part.type === "video_url") {
-          const mimeType = getVideoMimeTypeFromUrl(part.video_url.url);
-          return {
-            fileData: {
-              fileUri: part.video_url.url,
-              mimeType,
-            },
-          };
-        }
-        // Handle other potential generic message content types if necessary
-        // For now, only text and image_url are explicitly handled.
-        console.warn(
-          `Unsupported generic message content part type: ${part.type}`
-        );
-        return { text: `[Unsupported content type: ${part.type}]` };
-      })
-      .filter((part) => !!part); // Filter out any null/undefined parts if transformation fails
-  }
-
-  /**
-   * Transforms a generic Message array into a Google GenAI Content array.
-   * Extracts system messages separately. Maps user, assistant, and tool roles.
-   * Maps OpenAI-style tool messages to Google's functionResponse parts within user roles.
-   * Maps OpenAI-style assistant messages with tool_calls to Google's tool_use parts within assistant roles.
-   * @param messages The generic message array.
-   * @returns An object containing the system instruction (if any) and the Content array for the API call.
-   */
-  transformMessages(messages: Message[]): {
-    systemInstruction: string | undefined;
-    contents: Content[];
-  } {
-    const googleContents: Content[] = [];
-    let systemInstruction: string | undefined;
-
-    // Temporary storage for assistant tool calls keyed by ID, needed to map tool results
-    const assistantToolCalls: { [id: string]: ToolCall } = {};
-
-    for (const msg of messages) {
-      if (msg.role === "system") {
-        // System messages go into the systemInstruction field
-        if (typeof msg.content === "string") {
-          systemInstruction =
-            (systemInstruction ? systemInstruction + "\n" : "") + msg.content;
-        } else {
-          // System message content is typically string, handle array as text parts?
-          // Google's systemInstruction is string, so concatenate text parts if array.
-          systemInstruction =
-            (systemInstruction ? systemInstruction + "\n" : "") +
-            this.transformContentParts(msg.content)
-              .filter((p) => "text" in p && typeof p.text === "string")
-              .map((p) => p.text)
-              .join("\n");
-        }
-      } else if (msg.role === "user" || msg.role === "assistant") {
-        const parts = msg.content
-          ? this.transformContentParts(msg.content)
-          : [];
-
-        // Add tool_use parts if the assistant message has tool_calls
-        if (
-          msg.role === "assistant" &&
-          msg.tool_calls &&
-          msg.tool_calls.length > 0
-        ) {
-          // Look up any captured thoughtSignatures for this assistant message so
-          // we can re-attach them to the matching functionCall part (required by
-          // Gemini 3 to preserve reasoning continuity across tool-use turns).
-          const details = (msg as any)._reasoning_details;
-          const sigByCallId: { [id: string]: string } = {};
-          const standaloneThoughtSigs: string[] = [];
-          if (details && details.provider === "gemini" && Array.isArray(details.items)) {
-            for (const item of details.items) {
-              if (item?.kind === "functionCall" && item.callId) {
-                sigByCallId[item.callId] = item.thoughtSignature;
-              } else if (item?.kind === "thoughtSignature" && item.thoughtSignature) {
-                standaloneThoughtSigs.push(item.thoughtSignature);
-              }
-            }
-          }
-          for (const toolCall of msg.tool_calls) {
-            const callPart: any = {
-              functionCall: {
-                name: toolCall.function.name,
-                // Google expects arguments as a parsed object, not a string
-                args: JSON.parse(toolCall.function.arguments || "{}"),
-              },
-            };
-            if (sigByCallId[toolCall.id]) {
-              callPart.thoughtSignature = sigByCallId[toolCall.id];
-            }
-            parts.push(callPart);
-            // Store the tool call to potentially link with a future tool response message
-            assistantToolCalls[toolCall.id] = toolCall;
-          }
-          // Re-attach any standalone thought-part signatures onto the leading
-          // text part(s) so reasoning-only thoughts are also round-tripped.
-          for (let s = 0; s < standaloneThoughtSigs.length; s++) {
-            const p: any = parts[s];
-            if (p && "text" in p && !p.thoughtSignature) {
-              p.thoughtSignature = standaloneThoughtSigs[s];
-            }
-          }
-        }
-
-        if (parts.length > 0) {
-          googleContents.push({
-            role: msg.role === "user" ? "user" : "model",
-            parts,
-          });
-        }
-      } else if (msg.role === "tool") {
-        // OpenAI tool messages represent the *result* of a tool call.
-        // Google represents this as a 'functionResponse' part within a 'user' role message.
-        // The content of the tool message is the tool output (usually a string).
-        // The tool_call_id links it back to the assistant's tool_use part.
-
-        if (!msg.tool_call_id) {
-          throw new Error("Tool message must have a tool_call_id.");
-        }
-
-        // Ensure content is treated as string for functionResponse
-        const toolOutputContent =
-          typeof msg.content === "string"
-            ? msg.content
-            : JSON.stringify(msg.content); // Coerce array content to string representation if necessary
-
-        // Find the matching tool call name from the stored assistant tool calls
-        // This is needed for the functionResponse part's name field in Google's API.
-        const matchingToolCall = assistantToolCalls[msg.tool_call_id];
-        const functionName = matchingToolCall
-          ? matchingToolCall.function.name
-          : "unknown_function";
-
-        if (!matchingToolCall) {
-          console.warn(
-            `Matching assistant tool call not found for tool_call_id: ${msg.tool_call_id}. Using name '${functionName}'.`,
-            msg
-          );
-        }
-
-        // Add the user message with the functionResponse part
-        // Google's API expects the user role for tool results.
-        googleContents.push({
-          role: "user",
-          parts: [
-            {
-              functionResponse: {
-                name: functionName, // Google API requires the function name here
-                response: {
-                  result: toolOutputContent,
-                },
-              },
-            },
-          ],
-        });
-      } else {
-        console.warn(
-          `Unsupported generic message role: ${msg.role}, skipping.`
-        );
-      }
-    }
-
-    return { systemInstruction, contents: googleContents };
-  }
-
-  /**
-   * Recursively cleans a JSON schema to remove properties not supported by Gemini API.
-   * Removes: additionalProperties, $ref, and other unsupported fields.
-   * Converts type strings to uppercase as required by Gemini.
-   * @param schema The schema object to clean
-   * @returns A cleaned schema object compatible with Gemini API
-   */
-  private cleanSchemaForGemini(schema: any): any {
-    if (!schema || typeof schema !== "object") {
-      return schema;
-    }
-
-    // Handle arrays
-    if (Array.isArray(schema)) {
-      return schema.map((item) => this.cleanSchemaForGemini(item));
-    }
-
-    const cleaned: any = {};
-
-    for (const key in schema) {
-      if (!Object.prototype.hasOwnProperty.call(schema, key)) {
-        continue;
-      }
-
-      // Skip unsupported properties:
-      // - additionalProperties: not supported by Gemini
-      // - $ref: JSON Schema references not supported
-      // - $defs: JSON Schema definitions not supported
-      // - positional: internal knowhow property, not part of JSON Schema
-      if (
-        key === "additionalProperties" ||
-        key === "$ref" ||
-        key === "$defs" ||
-        key === "positional"
-      ) {
-        continue;
-      }
-
-      const value = schema[key];
-
-      // Convert type to uppercase if it's a string
-      if (key === "type" && typeof value === "string") {
-        cleaned[key] = value.toUpperCase();
-      }
-      // Handle type arrays (e.g., ["string", "null"])
-      else if (key === "type" && Array.isArray(value)) {
-        cleaned[key] = value.map((t: string) =>
-          typeof t === "string" ? t.toUpperCase() : t
-        );
-      }
-      // Recursively clean nested objects
-      else if (typeof value === "object" && value !== null) {
-        cleaned[key] = this.cleanSchemaForGemini(value);
-      }
-      // Copy primitive values as-is
-      else {
-        cleaned[key] = value;
-      }
-    }
-
-    return cleaned;
-  }
-
-  /**
-   * Transforms generic Tool array into Google GenAI tools format.
-   * @param tools The generic tool array.
-   * @returns An array of Google GenAI Tool objects, or undefined if no tools.
-   */
-  transformTools(tools?: Tool[]): GoogleTool[] | undefined {
-    if (!tools || tools.length === 0) {
-      return undefined;
-    }
-
-    const functionDeclarations: FunctionDeclaration[] = tools.map((tool) => {
-      // Clean the entire parameters schema to remove unsupported fields
-      const cleanedParameters = this.cleanSchemaForGemini(
-        tool.function.parameters
-      );
-
-      return {
-        name: tool.function.name,
-        description: tool.function.description || "",
-        parameters: cleanedParameters,
-      };
-    });
-
-    // Google's tools structure is an array of objects, where each object
-    // can contain 'functionDeclarations', 'googleSearch', 'codeExecution', etc.
-    // Based on the provided docs, function calling tools go under `functionDeclarations`.
-    return [{ functionDeclarations }];
-  }
-
-  /**
-   * Builds the thinkingConfig for Gemini models that support it.
-   * - Gemini 3.x models use thinkingLevel (supported levels vary by model)
-   * - Gemini 2.5 models use thinkingBudget: number (0 = off, -1 = dynamic)
-   *
-   * Maps CompletionOptions.reasoning_effort to provider-specific values.
-   */
-  buildThinkingConfig(
-    options: CompletionOptions
-  ): Record<string, unknown> | undefined {
-    const model = options.model;
-    const effort = options.reasoning_effort ?? "low";
-
-    // Gemini 3.x — use thinkingLevel
-    if (GoogleThinkingLevelModels.includes(model)) {
-      // Gemini 3.7 Flash rejects "minimal"; low is its least intensive level.
-      const minimumLevel = model === Models.google.Gemini_37_Flash ? "low" : "minimal";
-      const levelMap: Record<string, string> = {
-        none: minimumLevel,
-        minimal: minimumLevel,
-        low: "low",
-        medium: "medium",
-        high: "high",
-        xhigh: "high",
-        max: "high",
-      };
-      return {
-        thinkingLevel: levelMap[effort] ?? "low",
-      };
-    }
-
-    // Gemini 2.5 — use thinkingBudget
-    if (GoogleThinkingBudgetModels.includes(model)) {
-      // Map effort to token budget
-      const budgetMap: Record<string, number> = {
-        none: 0,
-        minimal: 512,
-        low: 1024,
-        medium: 8192,
-        high: -1, // dynamic
-        xhigh: -1,
-        max: -1,
-      };
-      return {
-        thinkingBudget: budgetMap[effort] ?? 1024,
-      };
-    }
-
-    return undefined;
-  }
-
-  async createChatCompletion(
-    options: CompletionOptions
-  ): Promise<CompletionResponse> {
-    const thinkingConfig = this.buildThinkingConfig(options);
-    const { systemInstruction, contents } = this.transformMessages(
-      options.messages
-    );
-    const tools = this.transformTools(options.tools);
-
-    try {
-      await wait(2000);
-      const response = await this.client.models.generateContent({
-        model: options.model,
-        contents,
-        config: {
-          systemInstruction,
-          thinkingConfig,
-          tools,
-          ...(tools.length && {
-            toolConfig: {
-              functionCallingConfig: {
-                mode: options.tool_choice === "none"
-                  ? FunctionCallingConfigMode.NONE
-                  : FunctionCallingConfigMode.AUTO,
-              },
-            },
-          }),
-          maxOutputTokens: options.max_tokens,
-        },
-      });
-
-      let toolCalls: ToolCall[] = [];
-
-      if (response.functionCalls) {
-        for (const call of response.functionCalls) {
-          toolCalls.push({
-            id:
-              call.id ||
-              `fc_${Date.now()}_${Math.random().toString(36).substr(2, 9)}`,
-            type: "function",
-            function: {
-              name: call.name,
-              arguments: JSON.stringify(call.args || {}),
-            },
-          });
-        }
-      }
-
-      if (response?.promptFeedback?.blockReason) {
-        // lame
-        throw new Error(
-          `Google GenAI blocked the response due to: ${response.promptFeedback.blockReason}`
-        );
-      }
-
-      // Map Google response to generic CompletionResponse
-      const choices: CompletionResponse["choices"] =
-        response.candidates?.map((candidate) => {
-          const message: OutputMessage = {
-            role: candidate.content.role === "model" ? "assistant" : "user",
-            content: "", // Initialize content
-            tool_calls: [...toolCalls], // Initialize tool calls
-          };
-
-          // Collect text and tool_use parts
-          let textContent = "";
-
-          // after the first message uses the top level tool calls we should empty it
-          if (toolCalls.length) {
-            toolCalls = [];
-          }
-
-          if (!candidate?.content?.parts) {
-            console.warn("No content parts in candidate:", candidate);
-            return { message };
-          }
-
-          // Capture Gemini thoughtSignatures so they can be round-tripped on the
-          // next turn. Gemini 3 REQUIRES these to be echoed back on the matching
-          // functionCall part or it returns a 400 "Thought signature is not
-          // valid" on the 2nd+ tool-calling turn. We key them by tool-call id
-          // (and also keep any standalone thought-part signatures in order).
-          const reasoningItems: any[] = [];
-
-          candidate?.content?.parts?.forEach((part) => {
-            if ("text" in part && typeof part.text === "string") {
-              textContent += part.text; // Concatenate text parts
-              if ((part as any).thoughtSignature) {
-                reasoningItems.push({
-                  kind: "thoughtSignature",
-                  thoughtSignature: (part as any).thoughtSignature,
-                });
-              }
-            } else if ("functionCall" in part && part.functionCall) {
-              const callId =
-                part.functionCall.id ||
-                `fc_${Date.now()}_${Math.random().toString(36).substr(2, 9)}`;
-              message.tool_calls.push({
-                id: callId,
-                type: "function",
-                function: {
-                  name: part.functionCall.name,
-                  arguments: JSON.stringify(part.functionCall.args || {}),
-                },
-              });
-              if ((part as any).thoughtSignature) {
-                reasoningItems.push({
-                  kind: "functionCall",
-                  callId,
-                  thoughtSignature: (part as any).thoughtSignature,
-                });
-              }
-            }
-          });
-
-          message.content = textContent || null;
-          if (reasoningItems.length > 0) {
-            (message as any)._reasoning_details = {
-              provider: "gemini",
-              items: reasoningItems,
-            };
-          }
-
-          return { message };
-        }) || []; // Handle case with no candidates
-
-      const usage = response.usageMetadata;
-      const usdCost = usage
-        ? this.calculateCost(options.model, usage)
-        : undefined;
-
-      // Map cachedContentTokenCount → prompt_tokens_details.cached_tokens so that
-      // base.ts can read cache hit tokens via usage.prompt_tokens_details?.cached_tokens
-      const cachedTokens = usage?.cachedContentTokenCount ?? 0;
-      const usageWithCache: TokenUsage | undefined = usage
-        ? ({
-            prompt_tokens: usage.promptTokenCount ?? 0,
-            completion_tokens: usage.candidatesTokenCount ?? 0,
-            total_tokens: usage.totalTokenCount,
-            prompt_tokens_details: { cached_tokens: cachedTokens },
-          } as TokenUsage)
-        : undefined;
-
-      return {
-        choices,
-        model: options.model,
-        usage: usageWithCache,
-        usd_cost: usdCost,
-      };
-    } catch (error) {
-      console.error(
-        "Error calling Google GenAI generateContent:",
-        error.message
-      );
-      throw error;
-    }
-  }
-
-  pricesPerMillion() {
-    return GeminiTextPricing;
-  }
-
-  getPricing(
-    model?: string
-  ):
-    | import("./pricing/types").ModelPricing
-    | Record<string, import("./pricing/types").ModelPricing>
-    | undefined {
-    if (model !== undefined) {
-      return GeminiTextPricing[model as keyof typeof GeminiTextPricing] as
-        | import("./pricing/types").ModelPricing
-        | undefined;
-    }
-    return GeminiTextPricing as unknown as Record<
-      string,
-      import("./pricing/types").ModelPricing
-    >;
-  }
-
-  calculateCost(model: string, usage: UsageMetadata): number | undefined {
-    const pricing = this.pricesPerMillion()[model];
-    if (!pricing || !usage) {
-      return 0;
-    }
-
-    let cost = 0;
-
-    // ── Input tokens ──────────────────────────────────────────────────────────
-    // Use per-modality breakdowns when available (inputTokensDetails) so that
-    // audio tokens (which can cost 3–10× more than text) are billed correctly.
-    const inputDetails = (usage as any).inputTokensDetails as
-      | { modality?: string; tokenCount?: number }[]
-      | undefined;
-
-    if (
-      inputDetails &&
-      inputDetails.length > 0 &&
-      pricing.input_audio !== undefined
-    ) {
-      // Modality-aware billing
-      for (const detail of inputDetails) {
-        const tokens = detail.tokenCount || 0;
-        if (!tokens) continue;
-        const modality = (detail.modality || "").toUpperCase();
-        if (modality === "AUDIO") {
-          const rate = pricing.input_audio ?? pricing.input ?? 0;
-          cost += (tokens * rate) / 1e6;
-        } else {
-          // TEXT / IMAGE / VIDEO all use the base `input` rate
-          const isGt200k = tokens > 200_000 && pricing.input_gt_200k;
-          const rate = isGt200k ? pricing.input_gt_200k! : pricing.input ?? 0;
-          cost += (tokens * rate) / 1e6;
-        }
-      }
-    } else if ("promptTokenCount" in usage && usage.promptTokenCount) {
-      // Fallback: no modality breakdown available — use total token count
-      const isGt200k =
-        usage.promptTokenCount > 200_000 && pricing.input_gt_200k;
-      const rate = isGt200k ? pricing.input_gt_200k! : pricing.input ?? 0;
-      cost += (usage.promptTokenCount * rate) / 1e6;
-    }
-
-    // ── Output tokens ─────────────────────────────────────────────────────────
-    const outputDetails = (usage as any).outputTokensDetails as
-      | { modality?: string; tokenCount?: number }[]
-      | undefined;
-
-    if (
-      outputDetails &&
-      outputDetails.length > 0 &&
-      pricing.output_audio !== undefined
-    ) {
-      // Modality-aware billing
-      for (const detail of outputDetails) {
-        const tokens = detail.tokenCount || 0;
-        if (!tokens) continue;
-        const modality = (detail.modality || "").toUpperCase();
-        if (modality === "AUDIO") {
-          const rate = pricing.output_audio ?? pricing.output ?? 0;
-          cost += (tokens * rate) / 1e6;
-        } else {
-          const isGt200k = tokens > 200_000 && pricing.output_gt_200k;
-          const rate = isGt200k ? pricing.output_gt_200k! : pricing.output ?? 0;
-          cost += (tokens * rate) / 1e6;
-        }
-      }
-    } else if ("responseTokenCount" in usage && usage.responseTokenCount) {
-      // Fallback: no modality breakdown — use total token count
-      const isGt200k =
-        usage.responseTokenCount > 200_000 && pricing.output_gt_200k;
-      const rate = isGt200k ? pricing.output_gt_200k! : pricing.output ?? 0;
-      cost += (usage.responseTokenCount * rate) / 1e6;
-    }
-
-    // ── Context caching ───────────────────────────────────────────────────────
-    if (
-      "cachedContentTokenCount" in usage &&
-      usage.cachedContentTokenCount &&
-      pricing.context_caching
-    ) {
-      if (
-        usage.cachedContentTokenCount > 200_000 &&
-        pricing.context_caching_gt_200k
-      ) {
-        cost +=
-          (usage.cachedContentTokenCount * pricing.context_caching_gt_200k) /
-          1e6;
-      } else {
-        cost += (usage.cachedContentTokenCount * pricing.context_caching) / 1e6;
-      }
-    }
-
-    return cost;
-  }
-
-  async getModels(modality?: ModelModality): Promise<{ id: string }[]> {
-    if (modality) {
-      const map: Partial<Record<ModelModality, string[]>> = {
-        completion: GoogleReasoningModels,
-        embedding: GoogleEmbeddingModelsList,
-        image: GoogleImageModels,
-        audio: GoogleTTSModels,
-        video: GoogleVideoModels,
-      };
-      return (map[modality] ?? []).map((id) => ({ id }));
-    }
-    // No modality — live API call (backward compat)
-    try {
-      const models = await this.client.models.list();
-      return models.page.map((m) => ({ id: m.name! }));
-    } catch (error) {
-      console.error("Error fetching Google GenAI models:", error);
-      throw error;
-    }
-  }
-
-  async createEmbedding(options: EmbeddingOptions): Promise<EmbeddingResponse> {
-    if (!options.model) {
-      console.warn(
-        "Embedding model not specified, using default 'text-embedding-004'."
-      );
-    }
-
-    try {
-      const googleEmbedding = await this.client.models.embedContent({
-        model: options.model,
-        contents: options.input,
-      });
-
-      // Map Google EmbeddingResponse to generic EmbeddingResponse
-      const data = googleEmbedding.embeddings.map((e, index) => ({
-        object: "embedding", // Hardcode as per OpenAI's object type for embeddings
-        embedding: e.values, // Google's embedding values
-        index, // Use array index
-      }));
-
-      const usage = {
-        promptTokenCount: googleEmbedding.metadata.billableCharacterCount || 0,
-        totalTokenCount: googleEmbedding.metadata.billableCharacterCount || 0,
-      };
-
-      const usdCost = this.calculateCost(options.model, usage);
-
-      return {
-        data,
-        model: options.model,
-        usage: {
-          prompt_tokens: usage.promptTokenCount,
-          total_tokens: usage.totalTokenCount,
-        },
-        usd_cost: usdCost,
-      };
-    } catch (error) {
-      console.error("Error calling Google GenAI embedContent:", error);
-      throw error;
-    }
-  }
-
-  async createAudioTranscription(
-    options: AudioTranscriptionOptions
-  ): Promise<AudioTranscriptionResponse> {
-    throw new Error(
-      "Audio transcription is not yet supported by the Gemini client. Use OpenAI client with Whisper model instead."
-    );
-  }
-
-  async createAudioGeneration(
-    options: AudioGenerationOptions
-  ): Promise<AudioGenerationResponse> {
-    try {
-      const response = await this.client.models.generateContent({
-        model: options.model,
-        contents: [
-          {
-            role: "user",
-            parts: [{ text: options.input }],
-          },
-        ],
-        config: {
-          responseModalities: ["AUDIO"],
-          speechConfig: {
-            voiceConfig: {
-              prebuiltVoiceConfig: {
-                voiceName: options.voice || "Puck",
-              },
-            },
-          },
-        },
-      });
-
-      // Extract audio data from the response
-      // Gemini returns inline audio data in the response parts
-      const audioPart = response.candidates?.[0]?.content?.parts?.find(
-        (part: any) => part.inlineData?.mimeType?.startsWith("audio/")
-      );
-
-      if (!audioPart || !audioPart.inlineData) {
-        throw new Error("No audio data returned from Gemini TTS");
-      }
-
-      // Convert base64 to buffer
-      const rawBuffer = Buffer.from(audioPart.inlineData.data, "base64");
-      const mimeType = audioPart.inlineData.mimeType || "audio/wav";
-
-      // Gemini returns raw PCM (audio/L16) - convert to WAV format for playability
-      let audioBuffer: Buffer = rawBuffer;
-      if (mimeType.includes("L16") || mimeType.includes("pcm")) {
-        // Parse sample rate from mime type e.g. "audio/L16;codec=pcm;rate=24000"
-        const rateMatch = mimeType.match(/rate=(\d+)/);
-        const sampleRate = rateMatch ? parseInt(rateMatch[1], 10) : 24000;
-        audioBuffer = pcmToWav(rawBuffer, sampleRate);
-      }
-
-      return {
-        audio: audioBuffer,
-        format: "audio/wav",
-      };
-    } catch (error) {
-      console.error("Error calling Gemini TTS:", error);
-      throw error;
-    }
-  }
-
-  async createImageGeneration(
-    options: ImageGenerationOptions
-  ): Promise<ImageGenerationResponse> {
-    try {
-      // Check if using Imagen 3 model or Gemini Flash inline generation
-      const isImagen3 = options.model?.includes("imagen");
-
-      if (isImagen3) {
-        // Imagen 3 uses the generateImages endpoint
-        const response = await this.client.models.generateImages({
-          model: options.model,
-          prompt: options.prompt,
-          config: {
-            numberOfImages: options.n || 1,
-          },
-        });
-
-        // Convert response to ImageGenerationResponse format
-        const generatedImages = response.generatedImages || [];
-        const images = generatedImages.map((img) => ({
-          // imageBytes is already a base64-encoded string from the API
-          // Don't re-encode it, just use it directly
-          b64_json: img.image?.imageBytes ? img.image.imageBytes : "",
-          revised_prompt: options.prompt,
-        }));
-
-        return {
-          created: Math.floor(Date.now() / 1000),
-          data: images,
-          usd_cost: 0.03 * images.length,
-        };
-      } else {
-        // Use Gemini Flash inline image generation (e.g., gemini-2.0-flash-preview-image-generation)
-        const response = await this.client.models.generateContent({
-          model: options.model,
-          contents: [
-            {
-              role: "user",
-              parts: [{ text: options.prompt }],
-            },
-          ],
-          config: {
-            responseModalities: ["IMAGE", "TEXT"],
-          },
-        });
-
-        // Extract image data from the response
-        const imageParts =
-          response.candidates?.[0]?.content?.parts?.filter((part: any) =>
-            part.inlineData?.mimeType?.startsWith("image/")
-          ) || [];
-
-        if (imageParts.length === 0) {
-          throw new Error("No image data returned from Gemini");
-        }
-
-        const images = imageParts.map((part: any) => ({
-          b64_json: part.inlineData.data,
-          revised_prompt: options.prompt,
-        }));
-
-        const usageMetadata = response.usageMetadata;
-        const usdCost = usageMetadata
-          ? this.calculateCost(options.model, usageMetadata)
-          : undefined;
-
-        return {
-          created: Math.floor(Date.now() / 1000),
-          data: images,
-          usd_cost: usdCost,
-        };
-      }
-    } catch (error) {
-      console.error("Error calling Gemini image generation:", error);
-      throw error;
-    }
-  }
-
-  async createVideoGeneration(
-    options: VideoGenerationOptions
-  ): Promise<VideoGenerationResponse> {
-    try {
-      // Submit the video generation job – do NOT poll here.
-      // Use getVideoStatus() to poll and downloadFile() to fetch the result.
-      const operation = await this.client.models.generateVideos({
-        model: options.model,
-        prompt: options.prompt,
-        config: {
-          numberOfVideos: options.n || 1,
-          ...(options.duration && {
-            durationSeconds: Math.max(6, options.duration),
-          }),
-          ...(options.resolution && { resolution: options.resolution }),
-          ...(options.aspect_ratio && { aspectRatio: options.aspect_ratio }),
-        },
-      });
-
-      // Calculate estimated cost using model-specific per-second rate
-      const duration = options.duration || 5; // Default 5 seconds
-      const pricingEntry = GeminiTextPricing[options.model];
-      const ratePerSec = pricingEntry?.video_generation ?? 0.4; // default to $0.40/sec (Veo 3 rate)
-      const usdCost = (options.n || 1) * duration * ratePerSec;
-
-      // Return the operation name as jobId so callers can use getVideoStatus / downloadVideo
-      return {
-        created: Math.floor(Date.now() / 1000),
-        data: [],
-        jobId: operation.name,
-        usd_cost: usdCost,
-      };
-    } catch (error) {
-      console.error("Error calling Gemini video generation:", error);
-      throw error;
-    }
-  }
-
-  async getVideoStatus(
-    options: VideoStatusOptions
-  ): Promise<VideoStatusResponse> {
-    try {
-      const operation = await this.client.operations.getVideosOperation({
-        operation: { name: options.jobId } as any,
-      });
-
-      if (operation.error) {
-        return {
-          jobId: options.jobId,
-          status: "failed",
-          error: JSON.stringify(operation.error),
-        };
-      }
-
-      if (!operation.done) {
-        return {
-          jobId: options.jobId,
-          status: "in_progress",
-        };
-      }
-
-      // Completed – extract file URIs
-      const generatedVideos = operation.response?.generatedVideos || [];
-      const data = generatedVideos.map((vid) => {
-        const videoBytes: string | undefined = vid.video?.videoBytes;
-        const uri: string | undefined = vid.video?.uri;
-        return {
-          b64_json: videoBytes || undefined,
-          url: uri || undefined,
-          fileUri: uri || undefined,
-        };
-      });
-
-      return {
-        jobId: options.jobId,
-        status: "completed",
-        data,
-      };
-    } catch (error) {
-      console.error("Error checking Gemini video status:", error);
-      throw error;
-    }
-  }
-
-  /**
-   * Download a video (or any file) via the Google GenAI Files API.
-   * Pass either `fileId` (the files/* name) or `uri` (the full URI).
-   */
-  async downloadVideo(
-    options: FileDownloadOptions
-  ): Promise<FileDownloadResponse> {
-    return this.downloadFile(options);
-  }
-
-  /**
-   * Upload a file to the Google GenAI Files API.
-   */
-  async uploadFile(options: FileUploadOptions): Promise<FileUploadResponse> {
-    try {
-      const blob = new Blob([new Uint8Array(options.data)], {
-        type: options.mimeType,
-      });
-      const uploadedFile = await this.client.files.upload({
-        file: blob,
-        config: {
-          mimeType: options.mimeType,
-          displayName: options.displayName,
-          name: options.fileName,
-        },
-      });
-
-      return {
-        fileId: uploadedFile.name,
-        uri: uploadedFile.uri,
-        url: uploadedFile.downloadUri || uploadedFile.uri,
-        mimeType: uploadedFile.mimeType,
-        sizeBytes: uploadedFile.sizeBytes
-          ? Number(uploadedFile.sizeBytes)
-          : undefined,
-      };
-    } catch (error) {
-      console.error("Error uploading file to Google GenAI Files API:", error);
-      throw error;
-    }
-  }
-
-  /**
-   * Download a file from the Google GenAI Files API.
-   *
-   * The SDK's `files.download()` writes to disk, so we use a temp file and
-   * read it back as a Buffer. Pass either:
-   *  - `fileId`: the files/* resource name (e.g. "files/abc-123") or a Video uri
-   *  - `uri`: the full Video.uri returned in GeneratedVideo (also accepted as fileId)
-   *
-   * For generated videos the `file` param accepts the Video object directly
-   * (uri + optional mimeType), which the SDK resolves to a download URL.
-   */
-  async downloadFile(
-    options: FileDownloadOptions
-  ): Promise<FileDownloadResponse> {
-    const mimeMap: Record<string, string> = {
-      ".mp4": "video/mp4",
-      ".webm": "video/webm",
-      ".mov": "video/quicktime",
-      ".png": "image/png",
-      ".jpg": "image/jpeg",
-      ".jpeg": "image/jpeg",
-      ".gif": "image/gif",
-      ".wav": "audio/wav",
-      ".mp3": "audio/mpeg",
-    };
-
-    try {
-      // The Google GenAI SDK's files.download() uses an async pipe that is NOT
-      // properly awaited, so we fetch the file directly via HTTP instead.
-      // Build the download URL from the uri/fileId.
-      const rawUri = options.uri || options.fileId || "";
-
-      // If it's already a full https URL, use it directly (append API key).
-      // Otherwise construct the Files API download URL from the resource name.
-      let downloadUrl: string;
-      if (rawUri.startsWith("https://")) {
-        // Append API key if not already present
-        const sep = rawUri.includes("?") ? "&" : "?";
-        downloadUrl = `${rawUri}${sep}key=${this.apiKey}`;
-      } else {
-        // Strip leading "files/" if present to get just the file ID
-        const fileId = rawUri.replace(/^files\//, "");
-        downloadUrl = `https://generativelanguage.googleapis.com/v1beta/files/${fileId}:download?alt=media&key=${this.apiKey}`;
-      }
-
-      const response = await fetch(downloadUrl);
-      if (!response.ok) {
-        throw new Error(
-          `HTTP ${response.status} ${response.statusText} downloading ${downloadUrl}`
-        );
-      }
-
-      const arrayBuffer = await response.arrayBuffer();
-      const data = Buffer.from(arrayBuffer);
-
-      // If caller supplied a filePath, write to it (creating dirs as needed)
-      if (options.filePath) {
-        fsSync.mkdirSync(pathSync.dirname(options.filePath), {
-          recursive: true,
-        });
-        fsSync.writeFileSync(options.filePath, data);
-      }
-
-      // Infer mime type from the URI/fileId first (more reliable), then from the path
-      const sourceForExt =
-        options.uri || options.fileId || options.filePath || "";
-      const ext = pathSync.extname(sourceForExt.split("?")[0]).toLowerCase();
-      const mimeType = mimeMap[ext] || "video/mp4";
-
-      return { data, mimeType };
-    } catch (error) {
-      console.error(
-        "Error downloading file from Google GenAI Files API:",
-        error
-      );
-      throw error;
-    }
-  }
-
-  getContextLimit(
-    model: string
-  ): { contextLimit: number; threshold: number } | undefined {
-    const contextLimit = ContextLimits[model];
-    if (contextLimit === undefined) return undefined;
-    const pricing = GeminiTextPricing[model];
-    // If the model has tiered pricing above 200k tokens, use 200k as the threshold
-    const threshold =
-      pricing && "input_gt_200k" in pricing ? 200_000 : contextLimit;
-    return { contextLimit, threshold };
-  }
-}