@tyvm/knowhow 0.0.159 → 0.0.160

package/src/clients/anthropic.ts

@@ -1,772 +0,0 @@
-import Anthropic from "@anthropic-ai/sdk";
-import { wait } from "../utils";
-import { emitResponseMetadata } from "./responseMetadata";
-import { AnthropicTextPricing, AnthropicModels } from "./pricing";
-import { ContextLimits } from "./contextLimits";
-import { ModelModality } from "./types";
-import { Models } from "../types";
-import {
-  GenericClient,
-  CompletionOptions,
-  CompletionResponse,
-  Tool,
-  Message,
-  EmbeddingOptions,
-  EmbeddingResponse,
-  AudioTranscriptionOptions,
-  AudioTranscriptionResponse,
-  AudioGenerationOptions,
-  AudioGenerationResponse,
-  ImageGenerationOptions,
-  ImageGenerationResponse,
-  VideoGenerationOptions,
-  VideoGenerationResponse,
-} from "./types";
-
-type MessageParam = Anthropic.MessageParam;
-type Usage = Anthropic.Usage;
-type ThinkingParam = Anthropic.Messages.ThinkingConfigParam;
-
-const alwaysOnModels: readonly string[] = [
-  AnthropicModels.Sonnet5,
-  AnthropicModels.Fable5,
-  AnthropicModels.Mythos5,
-  AnthropicModels.MythosPreview,
-];
-const optInModels: readonly string[] = [
-  AnthropicModels.Opus4_8,
-  AnthropicModels.Opus4_7,
-  AnthropicModels.Opus4_6,
-  AnthropicModels.Sonnet4_6,
-];
-
-export class GenericAnthropicClient implements GenericClient {
-  private client: Anthropic;
-  private apiKey?: string;
-
-  constructor(apiKey?: string) {
-    this.setKey(apiKey || process.env.ANTHROPIC_API_KEY || "");
-  }
-
-  private emitErrorMetadata(options: CompletionOptions, error: unknown) {
-    if (!error || typeof error !== "object") return;
-    const apiError = error as {
-      status?: number;
-      headers?: Headers | Record<string, unknown>;
-      request_id?: string;
-      requestID?: string;
-      response?: Response;
-    };
-    const statusCode = apiError.status ?? apiError.response?.status;
-    const headers = apiError.headers ?? apiError.response?.headers;
-    if (typeof statusCode === "number" || headers) {
-      emitResponseMetadata(options, {
-        statusCode,
-        headers,
-        requestId: apiError.request_id ?? apiError.requestID,
-      });
-    }
-  }
-
-  setKey(apiKey: string) {
-    this.apiKey = apiKey;
-    this.client = new Anthropic({
-      apiKey: apiKey || process.env.ANTHROPIC_API_KEY,
-      // Do not hide retry responses from the response-scoped observer.
-      maxRetries: 0,
-    });
-  }
-
-  handleToolCaching(tools: Anthropic.Tool[], longTtl = false) {
-    const lastTool = tools[tools.length - 1];
-
-    if (lastTool) {
-      lastTool.cache_control = longTtl
-        ? ({ type: "ephemeral", ttl: "1h" } as any)
-        : { type: "ephemeral" };
-    }
-  }
-
-  /**
-   * Determine the Anthropic `thinking` param from CompletionOptions.
-   * Returns undefined if the model does not support thinking.
-   */
-  getThinkingParam(
-    model: string,
-    reasoning_effort?: CompletionOptions["reasoning_effort"],
-    reasoning_summary?: boolean
-  ): ThinkingParam | undefined {
-    const effortIsNone = reasoning_effort === "none";
-    const withSummary = reasoning_summary
-      ? ({ display: "summarized" } as const)
-      : {};
-
-    if (alwaysOnModels.includes(model)) {
-      return effortIsNone
-        ? { type: "disabled" }
-        : { type: "adaptive", ...withSummary };
-    }
-
-    if (optInModels.includes(model)) {
-      // Only enable when explicitly requested (not "none" and not undefined).
-      if (!effortIsNone && reasoning_effort !== undefined) {
-        return { type: "adaptive", ...withSummary };
-      }
-      return undefined;
-    }
-
-    // Model does not support thinking.
-    return undefined;
-  }
-
-  /** Map provider-neutral effort names to Anthropic's output_config.effort. */
-  getEffortOutputConfig(
-    model: string,
-    reasoningEffort?: CompletionOptions["reasoning_effort"]
-  ): Anthropic.OutputConfig | undefined {
-    if (
-      reasoningEffort === undefined ||
-      reasoningEffort === "none" ||
-      (!alwaysOnModels.includes(model) && !optInModels.includes(model))
-    ) {
-      return undefined;
-    }
-
-    return {
-      // Anthropic uses `low` for the provider-neutral `minimal` level.
-      effort: reasoningEffort === "minimal" ? "low" : reasoningEffort,
-    };
-  }
-
-  /**
-   * Clean JSON Schema for Anthropic API compatibility.
-   * Removes unsupported fields like additionalProperties, $ref, $defs, positional.
-   */
-  private cleanSchemaForAnthropic(schema: any): any {
-    if (!schema || typeof schema !== "object") {
-      return schema;
-    }
-
-    // Handle arrays
-    if (Array.isArray(schema)) {
-      return schema.map((item) => this.cleanSchemaForAnthropic(item));
-    }
-
-    const cleaned: any = {};
-
-    for (const key in schema) {
-      if (!Object.prototype.hasOwnProperty.call(schema, key)) {
-        continue;
-      }
-
-      // Skip unsupported properties
-      if (
-        key === "additionalProperties" ||
-        key === "$ref" ||
-        key === "$defs" ||
-        key === "positional"
-      ) {
-        continue;
-      }
-
-      const value = schema[key];
-
-      // Recursively clean nested objects
-      if (typeof value === "object" && value !== null) {
-        cleaned[key] = this.cleanSchemaForAnthropic(value);
-      }
-      // Copy primitive values as-is
-      else {
-        cleaned[key] = value;
-      }
-    }
-
-    return cleaned;
-  }
-
-  transformTools(tools?: Tool[], longTtl = false): Anthropic.Tool[] {
-    if (!tools) {
-      return [];
-    }
-    const transformed = tools.map((tool) => ({
-      name: tool.function.name || "",
-      description: tool.function.description || "",
-      input_schema: this.cleanSchemaForAnthropic(
-        tool.function.parameters
-      ) as any,
-    }));
-
-    this.handleToolCaching(transformed, longTtl);
-
-    return transformed;
-  }
-
-  toBlockArray(content: MessageParam["content"]) {
-    if (typeof content === "string") {
-      return [
-        {
-          text: content,
-          type: "text",
-        },
-      ] as Anthropic.TextBlockParam[];
-    }
-    if (Array.isArray(content)) {
-      return content.map((c) => {
-        if (typeof c === "string") {
-          return {
-            text: c,
-            type: "text",
-          } as Anthropic.TextBlockParam;
-        }
-        return c;
-      });
-    }
-    return content;
-  }
-
-  combineMessages(messages: MessageParam[]): MessageParam[] {
-    if (messages.length <= 1) {
-      return messages;
-    }
-    for (let i = 0; i < messages.length; i++) {
-      if (i - 1 >= 0) {
-        const currentMessage = messages[i];
-        const previousMessage = messages[i - 1];
-        if (currentMessage?.role === previousMessage?.role) {
-          previousMessage.content = this.toBlockArray(previousMessage.content);
-          previousMessage.content.push(
-            ...this.toBlockArray(currentMessage.content)
-          );
-          messages.splice(i, 1);
-          i--;
-        }
-      }
-    }
-    return messages;
-  }
-
-  cacheLastContent(message: MessageParam, longTtl = false) {
-    if (Array.isArray(message.content)) {
-      const lastMessage = message.content[message.content.length - 1];
-      if (
-        lastMessage.type !== "thinking" &&
-        lastMessage.type !== "redacted_thinking"
-      ) {
-        lastMessage.cache_control = longTtl
-          ? ({ type: "ephemeral", ttl: "1h" } as any)
-          : { type: "ephemeral" };
-      }
-    }
-  }
-
-  handleClearingCache(messages: MessageParam[]) {
-    for (const message of messages) {
-      if (Array.isArray(message.content)) {
-        for (const content of message.content) {
-          if ("cache_control" in content && content.cache_control) {
-            delete content.cache_control;
-          }
-        }
-      }
-    }
-  }
-
-  handleMessageCaching(groupedMessages: MessageParam[], longTtl = false) {
-    this.handleClearingCache(groupedMessages);
-
-    // find the last two messages and mark them as ephemeral
-    const lastTwoUserMessages = groupedMessages
-      .filter((m) => m.role === "user")
-      .slice(-2);
-
-    for (const m of lastTwoUserMessages) {
-      if (Array.isArray(m.content)) {
-        this.cacheLastContent(m, longTtl);
-      }
-    }
-  }
-
-  tryParse(str: string): any {
-    try {
-      return JSON.parse(str);
-    } catch (e) {
-      console.error("Invalid JSON from tool call", str);
-      return {};
-    }
-  }
-
-  transformMessages(messages: Message[], longTtl = false): MessageParam[] {
-    const toolCalls = messages.flatMap((msg) => msg.tool_calls || []);
-    // Map each tool_call id → the thinking blocks captured on the assistant
-    // message that produced it, so we can re-inject them before the tool_use.
-    const thinkingByCallId: { [id: string]: any[] } = {};
-    for (const m of messages) {
-      if (m.role !== "assistant" || !m.tool_calls?.length) continue;
-      const details = (m as any)._reasoning_details;
-      if (details?.provider === "anthropic" && Array.isArray(details.items)) {
-        // Attach to the FIRST tool_call of this assistant turn (thinking blocks
-        // must precede the first tool_use in the turn).
-        thinkingByCallId[m.tool_calls[0].id] = details.items;
-      }
-    }
-    const claudeMessages: MessageParam[] = messages
-      .filter((msg) => msg.role !== "system")
-      .filter((msg) => msg.content || msg.role === "tool")
-      .map((msg) => {
-        if (msg.role === "tool") {
-          const toolCall = toolCalls.find((tc) => tc.id === msg.tool_call_id);
-          const toolMessages = [] as MessageParam[];
-          if (!toolCall) {
-            console.log(
-              "Tool call not found for message",
-              JSON.stringify(msg, null, 2)
-            );
-          } else {
-            // Re-inject any captured thinking / redacted_thinking blocks
-            // (verbatim, with signature/data) BEFORE the tool_use block, as
-            // Anthropic requires for reasoning continuity across tool turns.
-            const thinkingBlocks = thinkingByCallId[msg.tool_call_id] ?? [];
-            toolMessages.push({
-              role: "assistant",
-              content: [
-                ...thinkingBlocks,
-                {
-                  type: "tool_use",
-                  id: msg.tool_call_id,
-                  name: toolCall.function.name,
-                  input: this.tryParse(toolCall.function.arguments),
-                },
-              ],
-            });
-          }
-
-          // Convert tool message content to appropriate format
-          let toolResultContent:
-            | string
-            | (Anthropic.TextBlockParam | Anthropic.ImageBlockParam)[];
-
-          if (typeof msg.content === "string") {
-            toolResultContent = msg.content;
-          } else if (Array.isArray(msg.content)) {
-            // Transform image_url format to Anthropic's image format
-            toolResultContent = msg.content.map(
-              (item): Anthropic.TextBlockParam | Anthropic.ImageBlockParam => {
-                if (item.type === "image_url") {
-                  const url = item.image_url.url;
-                  const isDataUrl = url.startsWith("data:");
-                  const isHttpUrl = url.startsWith("http");
-                  if (isHttpUrl) {
-                    return {
-                      type: "image" as const,
-                      source: {
-                        type: "url" as const,
-                        url,
-                      },
-                    } as Anthropic.ImageBlockParam;
-                  } else {
-                    const base64Data = isDataUrl ? url.split(",")[1] : url;
-                    const mediaType = isDataUrl
-                      ? url.match(/data:([^;]+);/)?.[1] || "image/jpeg"
-                      : "image/jpeg";
-                    return {
-                      type: "image" as const,
-                      source: {
-                        type: "base64" as const,
-                        media_type: mediaType as any,
-                        data: base64Data,
-                      },
-                    };
-                  }
-                } else if (item.type === "text") {
-                  return { type: "text" as const, text: item.text };
-                }
-                // Fallback for unknown types
-                return { type: "text" as const, text: String(item) };
-              }
-            ) as (Anthropic.TextBlockParam | Anthropic.ImageBlockParam)[];
-          } else {
-            toolResultContent = String(msg.content);
-          }
-
-          toolMessages.push({
-            role: "user",
-            content: [
-              {
-                type: "tool_result",
-                content: toolResultContent,
-                tool_use_id: msg.tool_call_id,
-              },
-            ],
-          });
-
-          return toolMessages;
-        }
-
-        return {
-          content: this.transformContent(msg),
-          role: msg.role === "system" ? "assistant" : msg.role,
-        };
-      })
-      .flat();
-
-    const groupedMessages = this.combineMessages(claudeMessages);
-
-    this.handleMessageCaching(groupedMessages, longTtl);
-
-    return groupedMessages;
-  }
-
-  transformContent(message: Message) {
-    if (typeof message.content === "string") {
-      return message.content;
-    }
-
-    const transformContextElement = (
-      e: Message["content"]["0"]
-    ): Anthropic.ContentBlockParam => {
-      if (typeof e === "object" && e.type === "text") {
-        return { type: "text", text: e.text };
-      }
-      if (typeof e === "object" && e.type === "image_url") {
-        const isUrl = e.image_url.url.startsWith("http");
-        if (isUrl) {
-          return {
-            type: "image",
-            source: {
-              type: "url" as const,
-              url: e.image_url.url,
-            },
-          } as Anthropic.ContentBlockParam;
-        } else {
-          // Data URL or raw base64. Anthropic's base64 source expects the
-          // media_type separately and the data WITHOUT the
-          // "data:<mime>;base64," prefix. Parse it out here.
-          const dataUrlMatch = e.image_url.url.match(
-            /^data:([^;]+);base64,(.*)$/s
-          );
-          const mediaType = dataUrlMatch ? dataUrlMatch[1] : "image/jpeg";
-          const data = dataUrlMatch
-            ? dataUrlMatch[2]
-            : e.image_url.url.replace(/^data:[^,]*,/, "");
-          return {
-            type: "image",
-            source: {
-              type: "base64" as const,
-              media_type: mediaType,
-              data,
-            },
-          } as Anthropic.ContentBlockParam;
-        }
-      }
-    };
-
-    if (Array.isArray(message.content)) {
-      return message.content.map((e) => transformContextElement(e));
-    }
-  }
-
-  async createChatCompletion(
-    options: CompletionOptions
-  ): Promise<CompletionResponse> {
-    const longTtl = !!options.long_ttl_cache;
-
-    const thinkingParam = this.getThinkingParam(
-      options.model,
-      options.reasoning_effort,
-      options.reasoning_summary
-    );
-    const effortOutputConfig = this.getEffortOutputConfig(
-      options.model,
-      options.reasoning_effort
-    );
-
-    const systemMessage = options.messages
-      .filter((msg) => msg.role === "system")
-      .map((msg) => msg.content || "")
-      .join("\n");
-
-    const claudeMessages = this.transformMessages(options.messages, longTtl);
-
-    const tools = this.transformTools(options.tools, longTtl);
-    try {
-      const result = await this.client.messages
-        .create(
-          {
-            model: options.model,
-            messages: claudeMessages,
-            system: systemMessage
-              ? [
-                  {
-                    text: systemMessage,
-                    cache_control: longTtl
-                      ? ({ type: "ephemeral", ttl: "1h" } as any)
-                      : { type: "ephemeral" },
-                    type: "text",
-                  },
-                ]
-              : undefined,
-            max_tokens: options.max_tokens || 8000,
-            ...(thinkingParam && { thinking: thinkingParam }),
-            ...(effortOutputConfig && { output_config: effortOutputConfig }),
-            ...(tools.length && {
-              tool_choice: {
-                type: options.tool_choice ?? "auto",
-              },
-              tools,
-            }),
-          },
-          { signal: options.signal }
-        )
-        .withResponse();
-      const { data: response, response: nativeResponse, request_id } = result;
-      emitResponseMetadata(options, {
-        statusCode: nativeResponse.status,
-        headers: nativeResponse.headers,
-        requestId: request_id,
-      });
-
-      if (!response.content || !response.content.length) {
-        console.log("no content in Anthropic response", response);
-      }
-
-      // Capture thinking / redacted_thinking blocks verbatim (with their
-      // signature / data) so they can be round-tripped unmodified on the next
-      // tool-use turn. Anthropic REQUIRES thinking blocks from the assistant
-      // message to be passed back, complete and unmodified, when returning tool
-      // results — otherwise reasoning continuity and prompt-cache hits are lost.
-      const thinkingBlocks = response.content.filter(
-        (c) => c.type === "thinking" || c.type === "redacted_thinking"
-      );
-      const reasoningDetails =
-        thinkingBlocks.length > 0
-          ? { provider: "anthropic", items: thinkingBlocks }
-          : undefined;
-      // Extract a human-readable summary from thinking blocks (when present).
-      const reasoningSummary =
-        thinkingBlocks
-          .map((b) => ("thinking" in b ? (b as any).thinking : ""))
-          .filter(Boolean)
-          .join("\n") || undefined;
-
-      // Determine which resulting choice should carry the thinking blocks. They
-      // are re-injected before the first tool_use on the next turn, so prefer
-      // the first tool_use block; if there are no tool calls, fall back to the
-      // first (text) block. This keeps a single copy that transformMessages can
-      // reliably find via the assistant message's tool_calls.
-      const nonThinking = response.content.filter(
-        (c) => c.type !== "thinking" && c.type !== "redacted_thinking"
-      );
-      const firstToolUseIdx = nonThinking.findIndex(
-        (c) => c.type === "tool_use"
-      );
-      const carryIdx = firstToolUseIdx >= 0 ? firstToolUseIdx : 0;
-
-      return {
-        choices: nonThinking.map((c, idx) => {
-          // Only the FIRST resulting assistant message carries the thinking
-          // blocks (Anthropic requires a single thinking block sequence at the
-          // start of the assistant turn); attaching to every choice would
-          // duplicate them once the messages are re-transformed.
-          const carryReasoning = idx === carryIdx;
-          if (c.type === "tool_use") {
-            return {
-              message: {
-                role: "assistant",
-                content: "",
-                tool_calls: [
-                  {
-                    id: c.id,
-                    type: "function",
-                    function: {
-                      name: c.name,
-                      arguments: JSON.stringify(c.input),
-                    },
-                  },
-                ],
-                ...(carryReasoning &&
-                  reasoningDetails && { _reasoning_details: reasoningDetails }),
-                ...(carryReasoning &&
-                  reasoningSummary && { reasoning_summary: reasoningSummary }),
-              },
-            };
-          } else {
-            return {
-              message: {
-                role: "assistant",
-                content: "text" in c ? c.text : c.type,
-                tool_calls: [],
-                ...(carryReasoning &&
-                  reasoningDetails && { _reasoning_details: reasoningDetails }),
-                ...(carryReasoning &&
-                  reasoningSummary && { reasoning_summary: reasoningSummary }),
-              },
-            };
-          }
-        }),
-
-        model: options.model,
-        usage: response.usage
-          ? {
-              input_tokens: response.usage.input_tokens ?? 0,
-              output_tokens: response.usage.output_tokens ?? 0,
-              prompt_tokens: response.usage.input_tokens ?? 0,
-              completion_tokens: response.usage.output_tokens ?? 0,
-              total_tokens:
-                (response.usage.input_tokens ?? 0) +
-                (response.usage.output_tokens ?? 0),
-              cache_creation_input_tokens:
-                response.usage.cache_creation_input_tokens ?? 0,
-              cache_read_input_tokens:
-                response.usage.cache_read_input_tokens ?? 0,
-              ...((response.usage as any).thinking_input_tokens != null ||
-              (response.usage as any).output_tokens_details?.thinking_tokens !=
-                null
-                ? {
-                    output_tokens_details: {
-                      thinking_tokens:
-                        (response.usage as any).thinking_input_tokens ??
-                        (response.usage as any).output_tokens_details
-                          ?.thinking_tokens ??
-                        0,
-                    },
-                  }
-                : {}),
-            }
-          : undefined,
-        usd_cost: this.calculateCost(options.model, response.usage),
-      };
-    } catch (err) {
-      this.emitErrorMetadata(options, err);
-      if ("headers" in err && err.headers?.["x-should-retry"] === "true") {
-        console.warn("Retrying failed request", err);
-        await wait(2500);
-        return this.createChatCompletion(options);
-      } else {
-        console.error("Error in createChatCompletion", err);
-        throw err;
-      }
-    }
-  }
-
-  pricesPerMillion() {
-    return AnthropicTextPricing;
-  }
-
-  getPricing(
-    model?: string
-  ):
-    | import("./pricing/types").ModelPricing
-    | Record<string, import("./pricing/types").ModelPricing>
-    | undefined {
-    if (model !== undefined) {
-      return AnthropicTextPricing[
-        model as keyof typeof AnthropicTextPricing
-      ] as import("./pricing/types").ModelPricing | undefined;
-    }
-    return AnthropicTextPricing as unknown as Record<
-      string,
-      import("./pricing/types").ModelPricing
-    >;
-  }
-
-  calculateCost(model: string, usage: Usage): number | undefined {
-    const rawP = this.pricesPerMillion()[model];
-    // Fall back to pricing file for unknown/newer models
-    const fallback =
-      AnthropicTextPricing[model as keyof typeof AnthropicTextPricing];
-    // Second fallback: short aliases like "claude-sonnet-4-5" don't have their
-    // own entry but map to a dated version like "claude-sonnet-4-5-20250929".
-    // Find the first pricing entry whose key starts with the short model name.
-    const aliasFallback =
-      !rawP && !fallback
-        ? Object.entries(AnthropicTextPricing).find(([key]) =>
-            key.startsWith(model + "-")
-          )?.[1]
-        : undefined;
-    const p: any = rawP || fallback || aliasFallback || undefined;
-    if (!p) return undefined;
-
-    const inputTokens = usage.input_tokens ?? 0;
-    const cacheWriteTokens = usage.cache_creation_input_tokens ?? 0;
-    const cacheReadTokens = usage.cache_read_input_tokens ?? 0;
-    const outputTokens = usage.output_tokens ?? 0;
-
-    const totalInputTokens = inputTokens + cacheWriteTokens + cacheReadTokens;
-
-    const useLongContextTier = totalInputTokens > 200_000 && !!p.input_gt_200k;
-    const inputRate = useLongContextTier
-      ? (p.input_gt_200k as number)
-      : p.input;
-    const outputRate =
-      useLongContextTier && p.output_gt_200k ? p.output_gt_200k : p.output;
-
-    // Prefer modeling cache pricing as multipliers, but if you keep absolute numbers,
-    // you MUST scale them when usingLongContextTier.
-    //
-    // Anthropic docs describe cache read/write as multipliers of the base input rate. :contentReference[oaicite:7]{index=7}
-    // If your `cache_write` + `cache_hit` are absolute $/MTok at base tier, scale them:
-    const cacheWriteRate = (p.cache_write / p.input) * inputRate; // preserves your multiplier
-    const cacheReadRate = (p.cache_hit / p.input) * inputRate; // preserves your multiplier
-
-    const nonCachedInputCost = (inputTokens * inputRate) / 1e6;
-    const cacheWriteCost = (cacheWriteTokens * cacheWriteRate) / 1e6;
-    const cacheReadCost = (cacheReadTokens * cacheReadRate) / 1e6;
-    const outputCost = (outputTokens * outputRate) / 1e6;
-
-    return nonCachedInputCost + cacheWriteCost + cacheReadCost + outputCost;
-  }
-
-  async getModels(modality?: ModelModality): Promise<{ id: string }[]> {
-    if (modality) {
-      if (modality === "completion") {
-        return Object.values(Models.anthropic).map((id) => ({ id }));
-      }
-      return [];
-    }
-    // No modality — live API call (backward compat)
-    const models = await this.client.models.list();
-    return models.data.map((m) => ({ id: m.id }));
-  }
-
-  async createEmbedding(options: EmbeddingOptions): Promise<EmbeddingResponse> {
-    throw new Error("Provider does not support embeddings");
-  }
-
-  async createAudioTranscription(
-    options: AudioTranscriptionOptions
-  ): Promise<AudioTranscriptionResponse> {
-    throw new Error("Anthropic does not support audio transcription");
-  }
-
-  getContextLimit(
-    model: string
-  ): { contextLimit: number; threshold: number } | undefined {
-    const contextLimit = ContextLimits[model];
-    if (contextLimit === undefined) return undefined;
-    const pricing = AnthropicTextPricing[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 };
-  }
-
-  async createAudioGeneration(
-    options: AudioGenerationOptions
-  ): Promise<AudioGenerationResponse> {
-    throw new Error("Anthropic does not support audio generation");
-  }
-
-  async createImageGeneration(
-    options: ImageGenerationOptions
-  ): Promise<ImageGenerationResponse> {
-    throw new Error("Anthropic does not support image generation");
-  }
-
-  async createVideoGeneration(
-    options: VideoGenerationOptions
-  ): Promise<VideoGenerationResponse> {
-    throw new Error(
-      "Video generation is not supported by the Anthropic provider."
-    );
-  }
-}