@yourgpt/llm-sdk 2.5.1-beta.0 → 2.5.1-beta.1

package/README.md

@@ -114,6 +114,24 @@ const runtime = createRuntime({
 
 When `search.enabled` is on, deferred tools can be discovered through a hidden `search_tools` server tool. Matching tools are loaded into the next loop iteration instead of sending every deferred tool definition up front.
 
+## Structured output, MCP, and reasoning effort
+
+Pass `responseFormat`, `mcpServers`, and `reasoningEffort` on any `generateText()` / `streamText()` / `runtime.chat()` / `runtime.response()` call:
+
+```ts
+const result = await runtime.response({
+  prompt: "Extract FAQs from this conversation.",
+  mcpServers: [{ label: "kb", url: "https://kb.example.com/sse" }],
+  reasoningEffort: "high",
+  responseFormat: {
+    type: "json_schema",
+    json_schema: { name, schema, strict: true },
+  },
+});
+```
+
+OpenAI routes through `/v1/responses` automatically when MCP or reasoning is set; Anthropic uses the `mcp-client-2025-11-20` beta and adaptive thinking on Claude 4.6/4.7. See the [Structured Output guide](https://copilot-sdk.yourgpt.ai/docs/llm-sdk/structured-output) for the full per-provider mapping.
+
 ## Documentation
 
 Visit **[copilot-sdk.yourgpt.ai](https://copilot-sdk.yourgpt.ai)** for full documentation:
@@ -121,7 +139,7 @@ Visit **[copilot-sdk.yourgpt.ai](https://copilot-sdk.yourgpt.ai)** for full docu
 - [All Providers](https://copilot-sdk.yourgpt.ai/docs/providers) - OpenAI, Anthropic, Google, xAI
 - [Server Setup](https://copilot-sdk.yourgpt.ai/docs/server) - Runtime, streaming, tools
 - [Tools](https://copilot-sdk.yourgpt.ai/docs/tools) - Server-side and client-side tools
-- [LLM SDK Reference](https://copilot-sdk.yourgpt.ai/docs/llm-sdk) - streamText, generateText
+- [LLM SDK Reference](https://copilot-sdk.yourgpt.ai/docs/llm-sdk) - streamText, generateText, runtime.response()
 
 ## License
 

package/dist/adapters/index.d.mts

@@ -1,7 +1,7 @@
-import { L as LLMAdapter, C as ChatCompletionRequest, a as CompletionResult } from '../base-tNgbBaSo.mjs';
-export { A as AdapterFactory, l as AnthropicContentBlock, O as OpenAIContentBlock, j as attachmentToAnthropicDocument, i as attachmentToAnthropicImage, k as attachmentToOpenAIImage, f as formatMessages, c as formatMessagesForAnthropic, d as formatMessagesForOpenAI, b as formatTools, h as hasImageAttachments, g as hasMediaAttachments, m as messageToAnthropicContent, e as messageToOpenAIContent } from '../base-tNgbBaSo.mjs';
-import { W as WebSearchConfig, g as StreamEvent } from '../types-BkQCSiIt.mjs';
-import { d as OllamaModelOptions } from '../types-BSSiJW2o.mjs';
+import { L as LLMAdapter, C as ChatCompletionRequest, a as CompletionResult } from '../base-BYQKp9TW.mjs';
+export { A as AdapterFactory, l as AnthropicContentBlock, O as OpenAIContentBlock, j as attachmentToAnthropicDocument, i as attachmentToAnthropicImage, k as attachmentToOpenAIImage, f as formatMessages, c as formatMessagesForAnthropic, d as formatMessagesForOpenAI, b as formatTools, h as hasImageAttachments, g as hasMediaAttachments, m as messageToAnthropicContent, e as messageToOpenAIContent } from '../base-BYQKp9TW.mjs';
+import { W as WebSearchConfig, h as StreamEvent } from '../types-D774b0dg.mjs';
+import { d as OllamaModelOptions } from '../types-BvkiJ1dd.mjs';
 import 'zod';
 
 /**

package/dist/adapters/index.d.ts

@@ -1,7 +1,7 @@
-import { L as LLMAdapter, C as ChatCompletionRequest, a as CompletionResult } from '../base-C58Dsr9p.js';
-export { A as AdapterFactory, l as AnthropicContentBlock, O as OpenAIContentBlock, j as attachmentToAnthropicDocument, i as attachmentToAnthropicImage, k as attachmentToOpenAIImage, f as formatMessages, c as formatMessagesForAnthropic, d as formatMessagesForOpenAI, b as formatTools, h as hasImageAttachments, g as hasMediaAttachments, m as messageToAnthropicContent, e as messageToOpenAIContent } from '../base-C58Dsr9p.js';
-import { W as WebSearchConfig, g as StreamEvent } from '../types-BkQCSiIt.js';
-import { d as OllamaModelOptions } from '../types-BQ31QIsA.js';
+import { L as LLMAdapter, C as ChatCompletionRequest, a as CompletionResult } from '../base-Cxq3ni0t.js';
+export { A as AdapterFactory, l as AnthropicContentBlock, O as OpenAIContentBlock, j as attachmentToAnthropicDocument, i as attachmentToAnthropicImage, k as attachmentToOpenAIImage, f as formatMessages, c as formatMessagesForAnthropic, d as formatMessagesForOpenAI, b as formatTools, h as hasImageAttachments, g as hasMediaAttachments, m as messageToAnthropicContent, e as messageToOpenAIContent } from '../base-Cxq3ni0t.js';
+import { W as WebSearchConfig, h as StreamEvent } from '../types-D774b0dg.js';
+import { d as OllamaModelOptions } from '../types-TMilS-Dz.js';
 import 'zod';
 
 /**

package/dist/adapters/index.js

@@ -217,6 +217,87 @@ function toOllamaFormat(rf) {
   if (rf.type === "json_object") return "json";
   return rf.json_schema.schema;
 }
+function toOpenAIResponsesMcpTools(mcpServers) {
+  if (!mcpServers || mcpServers.length === 0) return [];
+  return mcpServers.map((mcp) => ({
+    type: "mcp",
+    server_label: mcp.label,
+    server_url: mcp.url,
+    ...mcp.headers ? { headers: mcp.headers } : {},
+    ...mcp.allowedTools ? { allowed_tools: mcp.allowedTools } : {},
+    require_approval: mcp.requireApproval ?? "never"
+  }));
+}
+function toAnthropicMcp(mcpServers) {
+  if (!mcpServers || mcpServers.length === 0) {
+    return { mcpServers: [], tools: [], betas: [] };
+  }
+  const serverEntries = [];
+  const toolEntries = [];
+  for (const mcp of mcpServers) {
+    const authHeader = mcp.headers?.Authorization ?? mcp.headers?.authorization;
+    const token = authHeader?.replace(/^Bearer\s+/i, "");
+    serverEntries.push({
+      type: "url",
+      url: mcp.url,
+      name: mcp.label,
+      ...token ? { authorization_token: token } : {}
+    });
+    if (mcp.allowedTools && mcp.allowedTools.length > 0) {
+      toolEntries.push({
+        type: "mcp_toolset",
+        mcp_server_name: mcp.label,
+        configs: Object.fromEntries(
+          mcp.allowedTools.map((toolName) => [toolName, {}])
+        )
+      });
+    }
+  }
+  return {
+    mcpServers: serverEntries,
+    tools: toolEntries,
+    betas: ["mcp-client-2025-11-20"]
+  };
+}
+function isStringEffort(effort) {
+  return typeof effort === "string" && (effort === "minimal" || effort === "low" || effort === "medium" || effort === "high");
+}
+function toOpenAIReasoning(effort) {
+  if (!effort) return void 0;
+  if (typeof effort === "object" && "raw" in effort) return effort.raw;
+  if (typeof effort === "object" && "budgetTokens" in effort) {
+    const budget = effort.budgetTokens;
+    const mapped = budget >= 16e3 ? "high" : budget >= 8e3 ? "medium" : "low";
+    return { effort: mapped, summary: "auto" };
+  }
+  if (isStringEffort(effort)) {
+    return { effort, summary: "auto" };
+  }
+  return void 0;
+}
+var ANTHROPIC_ADAPTIVE_MODELS = /(claude-opus-4-7|claude-opus-4-6|claude-sonnet-4-6)/i;
+function toAnthropicThinking(effort, modelId) {
+  if (!effort) return {};
+  if (typeof effort === "object" && "raw" in effort) {
+    return { thinking: effort.raw };
+  }
+  const isAdaptive = !!modelId && ANTHROPIC_ADAPTIVE_MODELS.test(modelId);
+  if (typeof effort === "object" && "budgetTokens" in effort) {
+    return {
+      thinking: { type: "enabled", budget_tokens: effort.budgetTokens }
+    };
+  }
+  if (!isStringEffort(effort)) return {};
+  if (isAdaptive) {
+    const mapped = effort === "minimal" ? "low" : effort;
+    return {
+      thinking: { type: "adaptive" },
+      outputConfigEffort: mapped
+    };
+  }
+  const budget = effort === "high" ? 16e3 : effort === "medium" ? 8e3 : effort === "low" ? 4e3 : 2048;
+  return { thinking: { type: "enabled", budget_tokens: budget } };
+}
 function formatTools(actions) {
   return actions.map((action) => ({
     type: "function",
@@ -498,6 +579,14 @@ var OpenAIAdapter = class _OpenAIAdapter {
     return this.client;
   }
   shouldUseResponsesApi(request) {
+    if (request.config?.mcpServers && request.config.mcpServers.length > 0 || request.config?.reasoningEffort !== void 0) {
+      if (this.provider !== "openai" && this.provider !== "azure") {
+        throw new Error(
+          `[llm-sdk] Provider "${this.provider}" does not support MCP servers or per-request reasoning effort. Use OpenAI or Anthropic for these features.`
+        );
+      }
+      return true;
+    }
     return request.providerToolOptions?.openai?.nativeToolSearch?.enabled === true && request.providerToolOptions.openai.nativeToolSearch.useResponsesApi !== false && Array.isArray(request.toolDefinitions) && request.toolDefinitions.length > 0;
   }
   buildResponsesInput(request) {
@@ -558,7 +647,7 @@ var OpenAIAdapter = class _OpenAIAdapter {
       strict: true,
       defer_loading: tool.deferLoading === true
     }));
-    return [{ type: "tool_search" }, ...nativeTools];
+    return nativeTools.length > 0 ? [{ type: "tool_search" }, ...nativeTools] : [];
   }
   parseResponsesResult(response) {
     const content = typeof response?.output_text === "string" ? response.output_text : "";
@@ -590,16 +679,30 @@ var OpenAIAdapter = class _OpenAIAdapter {
     const responsesTextFormat = toOpenAIResponsesTextFormat(
       request.config?.responseFormat
     );
+    const mcpTools = toOpenAIResponsesMcpTools(request.config?.mcpServers);
+    const modelId = request.config?.model || this.model;
+    const reasoning = isOpenAIReasoningModel(modelId) ? toOpenAIReasoning(request.config?.reasoningEffort) : void 0;
+    if (request.config?.reasoningEffort && !isOpenAIReasoningModel(modelId)) {
+      console.warn(
+        `[llm-sdk] openai/${modelId} is not a reasoning model; \`reasoningEffort\` is ignored. Use o1/o3/o4/gpt-5.x for reasoning.`
+      );
+    }
+    const functionTools = this.buildResponsesTools(
+      request.toolDefinitions ?? []
+    );
+    const tools = [...functionTools, ...mcpTools];
     const payload = {
       model: request.config?.model || this.model,
       instructions: request.systemPrompt,
       input: this.buildResponsesInput(request),
-      tools: this.buildResponsesTools(request.toolDefinitions ?? []),
+      tools: tools.length > 0 ? tools : void 0,
       tool_choice: openaiToolOptions?.toolChoice === "required" ? "required" : openaiToolOptions?.toolChoice === "auto" ? "auto" : void 0,
       parallel_tool_calls: openaiToolOptions?.parallelToolCalls,
       temperature: request.config?.temperature ?? this.config.temperature,
       max_output_tokens: request.config?.maxTokens ?? this.config.maxTokens,
       ...responsesTextFormat ? { text: { format: responsesTextFormat } } : {},
+      ...reasoning ? { reasoning } : {},
+      store: false,
       stream: false
     };
     logProviderPayload("openai", "request payload", payload, request.debug);
@@ -1246,36 +1349,58 @@ var AnthropicAdapter = class {
     if (serverToolConfiguration) {
       options.server_tool_configuration = serverToolConfiguration;
     }
+    const modelForThinking = request.config?.model || this.model;
+    const thinkingTranslation = toAnthropicThinking(
+      request.config?.reasoningEffort,
+      modelForThinking
+    );
     const outputConfig = toAnthropicOutputConfig(responseFormat);
-    if (outputConfig) {
-      options.output_config = outputConfig;
+    if (outputConfig || thinkingTranslation.outputConfigEffort) {
+      options.output_config = {
+        ...outputConfig ?? {},
+        ...thinkingTranslation.outputConfigEffort ? { effort: thinkingTranslation.outputConfigEffort } : {}
+      };
     }
-    if (this.config.thinking?.type === "enabled") {
+    if (thinkingTranslation.thinking) {
+      options.thinking = thinkingTranslation.thinking;
+    } else if (this.config.thinking?.type === "enabled") {
       options.thinking = {
         type: "enabled",
         budget_tokens: this.config.thinking.budgetTokens || 1e4
       };
     }
-    return { options, messages };
+    const mcp = toAnthropicMcp(request.config?.mcpServers);
+    const betas = [];
+    if (mcp.mcpServers.length > 0) {
+      options.mcp_servers = mcp.mcpServers;
+      betas.push(...mcp.betas);
+      if (mcp.tools.length > 0) {
+        const existingTools = Array.isArray(options.tools) ? options.tools : [];
+        options.tools = [...existingTools, ...mcp.tools];
+      }
+    }
+    return { options, messages, betas };
   }
   /**
    * Non-streaming completion (for debugging/comparison with original studio-ai)
    */
   async complete(request) {
     const client = await this.getClient();
-    const { options } = this.buildRequestOptions(request);
+    const { options, betas } = this.buildRequestOptions(request);
     const nonStreamingOptions = {
       ...options,
       stream: false
     };
     try {
+      const finalOptions = betas.length > 0 ? { ...nonStreamingOptions, betas } : nonStreamingOptions;
+      const messagesApi = betas.length > 0 ? client.beta.messages : client.messages;
       logProviderPayload(
         "anthropic",
         "request payload",
-        nonStreamingOptions,
+        finalOptions,
         request.debug
       );
-      const response = await client.messages.create(nonStreamingOptions);
+      const response = await messagesApi.create(finalOptions);
       logProviderPayload(
         "anthropic",
         "response payload",
@@ -1310,17 +1435,19 @@ var AnthropicAdapter = class {
   }
   async *stream(request) {
     const client = await this.getClient();
-    const { options } = this.buildRequestOptions(request);
+    const { options, betas } = this.buildRequestOptions(request);
     const messageId = generateMessageId();
     yield { type: "message:start", id: messageId };
     try {
+      const finalOptions = betas.length > 0 ? { ...options, betas } : options;
+      const streamApi = betas.length > 0 ? client.beta.messages : client.messages;
       logProviderPayload(
         "anthropic",
         "request payload",
-        options,
+        finalOptions,
         request.debug
       );
-      const stream = await client.messages.stream(options);
+      const stream = await streamApi.stream(finalOptions);
       let currentToolUse = null;
       let isInThinkingBlock = false;
       const collectedCitations = [];

package/dist/adapters/index.mjs

@@ -215,6 +215,87 @@ function toOllamaFormat(rf) {
   if (rf.type === "json_object") return "json";
   return rf.json_schema.schema;
 }
+function toOpenAIResponsesMcpTools(mcpServers) {
+  if (!mcpServers || mcpServers.length === 0) return [];
+  return mcpServers.map((mcp) => ({
+    type: "mcp",
+    server_label: mcp.label,
+    server_url: mcp.url,
+    ...mcp.headers ? { headers: mcp.headers } : {},
+    ...mcp.allowedTools ? { allowed_tools: mcp.allowedTools } : {},
+    require_approval: mcp.requireApproval ?? "never"
+  }));
+}
+function toAnthropicMcp(mcpServers) {
+  if (!mcpServers || mcpServers.length === 0) {
+    return { mcpServers: [], tools: [], betas: [] };
+  }
+  const serverEntries = [];
+  const toolEntries = [];
+  for (const mcp of mcpServers) {
+    const authHeader = mcp.headers?.Authorization ?? mcp.headers?.authorization;
+    const token = authHeader?.replace(/^Bearer\s+/i, "");
+    serverEntries.push({
+      type: "url",
+      url: mcp.url,
+      name: mcp.label,
+      ...token ? { authorization_token: token } : {}
+    });
+    if (mcp.allowedTools && mcp.allowedTools.length > 0) {
+      toolEntries.push({
+        type: "mcp_toolset",
+        mcp_server_name: mcp.label,
+        configs: Object.fromEntries(
+          mcp.allowedTools.map((toolName) => [toolName, {}])
+        )
+      });
+    }
+  }
+  return {
+    mcpServers: serverEntries,
+    tools: toolEntries,
+    betas: ["mcp-client-2025-11-20"]
+  };
+}
+function isStringEffort(effort) {
+  return typeof effort === "string" && (effort === "minimal" || effort === "low" || effort === "medium" || effort === "high");
+}
+function toOpenAIReasoning(effort) {
+  if (!effort) return void 0;
+  if (typeof effort === "object" && "raw" in effort) return effort.raw;
+  if (typeof effort === "object" && "budgetTokens" in effort) {
+    const budget = effort.budgetTokens;
+    const mapped = budget >= 16e3 ? "high" : budget >= 8e3 ? "medium" : "low";
+    return { effort: mapped, summary: "auto" };
+  }
+  if (isStringEffort(effort)) {
+    return { effort, summary: "auto" };
+  }
+  return void 0;
+}
+var ANTHROPIC_ADAPTIVE_MODELS = /(claude-opus-4-7|claude-opus-4-6|claude-sonnet-4-6)/i;
+function toAnthropicThinking(effort, modelId) {
+  if (!effort) return {};
+  if (typeof effort === "object" && "raw" in effort) {
+    return { thinking: effort.raw };
+  }
+  const isAdaptive = !!modelId && ANTHROPIC_ADAPTIVE_MODELS.test(modelId);
+  if (typeof effort === "object" && "budgetTokens" in effort) {
+    return {
+      thinking: { type: "enabled", budget_tokens: effort.budgetTokens }
+    };
+  }
+  if (!isStringEffort(effort)) return {};
+  if (isAdaptive) {
+    const mapped = effort === "minimal" ? "low" : effort;
+    return {
+      thinking: { type: "adaptive" },
+      outputConfigEffort: mapped
+    };
+  }
+  const budget = effort === "high" ? 16e3 : effort === "medium" ? 8e3 : effort === "low" ? 4e3 : 2048;
+  return { thinking: { type: "enabled", budget_tokens: budget } };
+}
 function formatTools(actions) {
   return actions.map((action) => ({
     type: "function",
@@ -496,6 +577,14 @@ var OpenAIAdapter = class _OpenAIAdapter {
     return this.client;
   }
   shouldUseResponsesApi(request) {
+    if (request.config?.mcpServers && request.config.mcpServers.length > 0 || request.config?.reasoningEffort !== void 0) {
+      if (this.provider !== "openai" && this.provider !== "azure") {
+        throw new Error(
+          `[llm-sdk] Provider "${this.provider}" does not support MCP servers or per-request reasoning effort. Use OpenAI or Anthropic for these features.`
+        );
+      }
+      return true;
+    }
     return request.providerToolOptions?.openai?.nativeToolSearch?.enabled === true && request.providerToolOptions.openai.nativeToolSearch.useResponsesApi !== false && Array.isArray(request.toolDefinitions) && request.toolDefinitions.length > 0;
   }
   buildResponsesInput(request) {
@@ -556,7 +645,7 @@ var OpenAIAdapter = class _OpenAIAdapter {
       strict: true,
       defer_loading: tool.deferLoading === true
     }));
-    return [{ type: "tool_search" }, ...nativeTools];
+    return nativeTools.length > 0 ? [{ type: "tool_search" }, ...nativeTools] : [];
   }
   parseResponsesResult(response) {
     const content = typeof response?.output_text === "string" ? response.output_text : "";
@@ -588,16 +677,30 @@ var OpenAIAdapter = class _OpenAIAdapter {
     const responsesTextFormat = toOpenAIResponsesTextFormat(
       request.config?.responseFormat
     );
+    const mcpTools = toOpenAIResponsesMcpTools(request.config?.mcpServers);
+    const modelId = request.config?.model || this.model;
+    const reasoning = isOpenAIReasoningModel(modelId) ? toOpenAIReasoning(request.config?.reasoningEffort) : void 0;
+    if (request.config?.reasoningEffort && !isOpenAIReasoningModel(modelId)) {
+      console.warn(
+        `[llm-sdk] openai/${modelId} is not a reasoning model; \`reasoningEffort\` is ignored. Use o1/o3/o4/gpt-5.x for reasoning.`
+      );
+    }
+    const functionTools = this.buildResponsesTools(
+      request.toolDefinitions ?? []
+    );
+    const tools = [...functionTools, ...mcpTools];
     const payload = {
       model: request.config?.model || this.model,
       instructions: request.systemPrompt,
       input: this.buildResponsesInput(request),
-      tools: this.buildResponsesTools(request.toolDefinitions ?? []),
+      tools: tools.length > 0 ? tools : void 0,
       tool_choice: openaiToolOptions?.toolChoice === "required" ? "required" : openaiToolOptions?.toolChoice === "auto" ? "auto" : void 0,
       parallel_tool_calls: openaiToolOptions?.parallelToolCalls,
       temperature: request.config?.temperature ?? this.config.temperature,
       max_output_tokens: request.config?.maxTokens ?? this.config.maxTokens,
       ...responsesTextFormat ? { text: { format: responsesTextFormat } } : {},
+      ...reasoning ? { reasoning } : {},
+      store: false,
       stream: false
     };
     logProviderPayload("openai", "request payload", payload, request.debug);
@@ -1244,36 +1347,58 @@ var AnthropicAdapter = class {
     if (serverToolConfiguration) {
       options.server_tool_configuration = serverToolConfiguration;
     }
+    const modelForThinking = request.config?.model || this.model;
+    const thinkingTranslation = toAnthropicThinking(
+      request.config?.reasoningEffort,
+      modelForThinking
+    );
     const outputConfig = toAnthropicOutputConfig(responseFormat);
-    if (outputConfig) {
-      options.output_config = outputConfig;
+    if (outputConfig || thinkingTranslation.outputConfigEffort) {
+      options.output_config = {
+        ...outputConfig ?? {},
+        ...thinkingTranslation.outputConfigEffort ? { effort: thinkingTranslation.outputConfigEffort } : {}
+      };
     }
-    if (this.config.thinking?.type === "enabled") {
+    if (thinkingTranslation.thinking) {
+      options.thinking = thinkingTranslation.thinking;
+    } else if (this.config.thinking?.type === "enabled") {
       options.thinking = {
         type: "enabled",
         budget_tokens: this.config.thinking.budgetTokens || 1e4
       };
     }
-    return { options, messages };
+    const mcp = toAnthropicMcp(request.config?.mcpServers);
+    const betas = [];
+    if (mcp.mcpServers.length > 0) {
+      options.mcp_servers = mcp.mcpServers;
+      betas.push(...mcp.betas);
+      if (mcp.tools.length > 0) {
+        const existingTools = Array.isArray(options.tools) ? options.tools : [];
+        options.tools = [...existingTools, ...mcp.tools];
+      }
+    }
+    return { options, messages, betas };
   }
   /**
    * Non-streaming completion (for debugging/comparison with original studio-ai)
    */
   async complete(request) {
     const client = await this.getClient();
-    const { options } = this.buildRequestOptions(request);
+    const { options, betas } = this.buildRequestOptions(request);
     const nonStreamingOptions = {
       ...options,
       stream: false
     };
     try {
+      const finalOptions = betas.length > 0 ? { ...nonStreamingOptions, betas } : nonStreamingOptions;
+      const messagesApi = betas.length > 0 ? client.beta.messages : client.messages;
       logProviderPayload(
         "anthropic",
         "request payload",
-        nonStreamingOptions,
+        finalOptions,
         request.debug
       );
-      const response = await client.messages.create(nonStreamingOptions);
+      const response = await messagesApi.create(finalOptions);
       logProviderPayload(
         "anthropic",
         "response payload",
@@ -1308,17 +1433,19 @@ var AnthropicAdapter = class {
   }
   async *stream(request) {
     const client = await this.getClient();
-    const { options } = this.buildRequestOptions(request);
+    const { options, betas } = this.buildRequestOptions(request);
     const messageId = generateMessageId();
     yield { type: "message:start", id: messageId };
     try {
+      const finalOptions = betas.length > 0 ? { ...options, betas } : options;
+      const streamApi = betas.length > 0 ? client.beta.messages : client.messages;
       logProviderPayload(
         "anthropic",
         "request payload",
-        options,
+        finalOptions,
         request.debug
       );
-      const stream = await client.messages.stream(options);
+      const stream = await streamApi.stream(finalOptions);
       let currentToolUse = null;
       let isInThinkingBlock = false;
       const collectedCitations = [];

package/dist/{base-tNgbBaSo.d.mts → base-BYQKp9TW.d.mts}

@@ -1,4 +1,4 @@
-import { M as Message, A as ActionDefinition, d as ToolDefinition, R as ResponseFormat, W as WebSearchConfig, P as ProviderToolRuntimeOptions, g as StreamEvent, J as TokenUsage, X as LLMConfig, a6 as MessageAttachment } from './types-BkQCSiIt.mjs';
+import { k as Message, A as ActionDefinition, d as ToolDefinition, R as ResponseFormat, M as McpServerConfig, g as ReasoningEffort, W as WebSearchConfig, P as ProviderToolRuntimeOptions, h as StreamEvent, O as TokenUsage, Z as LLMConfig, a8 as MessageAttachment } from './types-D774b0dg.mjs';
 
 /**
  * Request-level LLM configuration overrides
@@ -8,6 +8,10 @@ interface RequestLLMConfig {
     temperature?: number;
     maxTokens?: number;
     responseFormat?: ResponseFormat;
+    /** MCP servers exposed to the model for this request (provider-translated). */
+    mcpServers?: McpServerConfig[];
+    /** Reasoning/thinking effort knob (provider-translated). */
+    reasoningEffort?: ReasoningEffort;
 }
 /**
  * Chat completion request

package/dist/{base-C58Dsr9p.d.ts → base-Cxq3ni0t.d.ts}

@@ -1,4 +1,4 @@
-import { M as Message, A as ActionDefinition, d as ToolDefinition, R as ResponseFormat, W as WebSearchConfig, P as ProviderToolRuntimeOptions, g as StreamEvent, J as TokenUsage, X as LLMConfig, a6 as MessageAttachment } from './types-BkQCSiIt.js';
+import { k as Message, A as ActionDefinition, d as ToolDefinition, R as ResponseFormat, M as McpServerConfig, g as ReasoningEffort, W as WebSearchConfig, P as ProviderToolRuntimeOptions, h as StreamEvent, O as TokenUsage, Z as LLMConfig, a8 as MessageAttachment } from './types-D774b0dg.js';
 
 /**
  * Request-level LLM configuration overrides
@@ -8,6 +8,10 @@ interface RequestLLMConfig {
     temperature?: number;
     maxTokens?: number;
     responseFormat?: ResponseFormat;
+    /** MCP servers exposed to the model for this request (provider-translated). */
+    mcpServers?: McpServerConfig[];
+    /** Reasoning/thinking effort knob (provider-translated). */
+    reasoningEffort?: ReasoningEffort;
 }
 /**
  * Chat completion request

package/dist/fallback/index.d.mts

@@ -1,7 +1,7 @@
-import { L as LLMAdapter } from '../base-tNgbBaSo.mjs';
-import { F as FallbackChainConfig, c as FallbackFailure, R as RoutingStore } from '../types-B6dhnguR.mjs';
-export { d as FallbackInfo, b as RetryBackoff, e as RetryInfo, a as RoutingStrategy } from '../types-B6dhnguR.mjs';
-import '../types-BkQCSiIt.mjs';
+import { L as LLMAdapter } from '../base-BYQKp9TW.mjs';
+import { F as FallbackChainConfig, c as FallbackFailure, R as RoutingStore } from '../types-mwMhCwOq.mjs';
+export { d as FallbackInfo, b as RetryBackoff, e as RetryInfo, a as RoutingStrategy } from '../types-mwMhCwOq.mjs';
+import '../types-D774b0dg.mjs';
 import 'zod';
 
 /**

package/dist/fallback/index.d.ts

@@ -1,7 +1,7 @@
-import { L as LLMAdapter } from '../base-C58Dsr9p.js';
-import { F as FallbackChainConfig, c as FallbackFailure, R as RoutingStore } from '../types-CCxPmkmK.js';
-export { d as FallbackInfo, b as RetryBackoff, e as RetryInfo, a as RoutingStrategy } from '../types-CCxPmkmK.js';
-import '../types-BkQCSiIt.js';
+import { L as LLMAdapter } from '../base-Cxq3ni0t.js';
+import { F as FallbackChainConfig, c as FallbackFailure, R as RoutingStore } from '../types-ChORafYS.js';
+export { d as FallbackInfo, b as RetryBackoff, e as RetryInfo, a as RoutingStrategy } from '../types-ChORafYS.js';
+import '../types-D774b0dg.js';
 import 'zod';
 
 /**