@yourgpt/llm-sdk 2.5.0 → 2.5.1-beta.1

package/dist/adapters/index.mjs

@@ -112,6 +112,190 @@ function normalizeObjectJsonSchema(schema) {
   }
   return normalized;
 }
+function isOpenAIReasoningModel(modelId) {
+  if (!modelId) return false;
+  return /^(o1|o3|o4|gpt-5)/i.test(modelId);
+}
+function buildOpenAITokenParams(modelId, maxTokens, temperature) {
+  if (isOpenAIReasoningModel(modelId)) {
+    return { max_completion_tokens: maxTokens };
+  }
+  return { max_tokens: maxTokens, temperature };
+}
+function stripSchemaKeys(schema, keysToDrop, options = {}) {
+  if (Array.isArray(schema)) {
+    return schema.map((item) => stripSchemaKeys(item, keysToDrop, options));
+  }
+  if (!schema || typeof schema !== "object") return schema;
+  const out = {};
+  for (const [key, value] of Object.entries(
+    schema
+  )) {
+    if (keysToDrop.has(key)) continue;
+    const renamed = options.renameKeys?.[key] ?? key;
+    out[renamed] = stripSchemaKeys(value, keysToDrop, options);
+  }
+  if (options.forceAdditionalPropertiesFalse && out.type === "object") {
+    out.additionalProperties = false;
+  }
+  return out;
+}
+function toOpenAIResponseFormat(rf) {
+  if (!rf) return void 0;
+  if (rf.type === "json_object") return { type: "json_object" };
+  return {
+    type: "json_schema",
+    json_schema: {
+      name: rf.json_schema.name,
+      schema: normalizeObjectJsonSchema(rf.json_schema.schema),
+      strict: rf.json_schema.strict ?? true
+    }
+  };
+}
+function toOpenAIResponsesTextFormat(rf) {
+  if (!rf || rf.type !== "json_schema") return void 0;
+  return {
+    type: "json_schema",
+    name: rf.json_schema.name,
+    schema: normalizeObjectJsonSchema(rf.json_schema.schema),
+    strict: rf.json_schema.strict ?? true
+  };
+}
+var ANTHROPIC_UNSUPPORTED_KEYS = /* @__PURE__ */ new Set([
+  "minimum",
+  "maximum",
+  "exclusiveMinimum",
+  "exclusiveMaximum",
+  "multipleOf",
+  "minLength",
+  "maxLength",
+  "minItems",
+  "maxItems",
+  "minProperties",
+  "maxProperties",
+  "pattern",
+  "$schema"
+]);
+function toAnthropicOutputConfig(rf) {
+  if (!rf || rf.type !== "json_schema") return void 0;
+  const schema = stripSchemaKeys(
+    rf.json_schema.schema,
+    ANTHROPIC_UNSUPPORTED_KEYS,
+    {
+      forceAdditionalPropertiesFalse: true,
+      renameKeys: { oneOf: "anyOf" }
+    }
+  );
+  return {
+    format: {
+      type: "json_schema",
+      schema
+    }
+  };
+}
+var GEMINI_UNSUPPORTED_KEYS = /* @__PURE__ */ new Set([
+  "oneOf",
+  "anyOf",
+  "$ref",
+  "$defs",
+  "definitions",
+  "pattern",
+  "$schema",
+  "additionalProperties"
+]);
+function toGeminiSchema(rf) {
+  if (!rf || rf.type !== "json_schema") return void 0;
+  return stripSchemaKeys(
+    rf.json_schema.schema,
+    GEMINI_UNSUPPORTED_KEYS
+  );
+}
+function toOllamaFormat(rf) {
+  if (!rf) return void 0;
+  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",
@@ -393,6 +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 "${this.provider}" does not support MCP servers or per-request reasoning effort. Use OpenAI or Anthropic for these features.`
+        );
+      }
+      return true;
+    }
     return request.providerToolOptions?.openai?.nativeToolSearch?.enabled === true && request.providerToolOptions.openai.nativeToolSearch.useResponsesApi !== false && Array.isArray(request.toolDefinitions) && request.toolDefinitions.length > 0;
   }
   buildResponsesInput(request) {
@@ -453,7 +645,7 @@ var OpenAIAdapter = class _OpenAIAdapter {
       strict: true,
       defer_loading: tool.deferLoading === true
     }));
-    return [{ type: "tool_search" }, ...nativeTools];
+    return nativeTools.length > 0 ? [{ type: "tool_search" }, ...nativeTools] : [];
   }
   parseResponsesResult(response) {
     const content = typeof response?.output_text === "string" ? response.output_text : "";
@@ -482,15 +674,33 @@ var OpenAIAdapter = class _OpenAIAdapter {
   async completeWithResponses(request) {
     const client = await this.getClient();
     const openaiToolOptions = request.providerToolOptions?.openai;
+    const responsesTextFormat = toOpenAIResponsesTextFormat(
+      request.config?.responseFormat
+    );
+    const mcpTools = toOpenAIResponsesMcpTools(request.config?.mcpServers);
+    const modelId = request.config?.model || this.model;
+    const reasoning = isOpenAIReasoningModel(modelId) ? toOpenAIReasoning(request.config?.reasoningEffort) : void 0;
+    if (request.config?.reasoningEffort && !isOpenAIReasoningModel(modelId)) {
+      console.warn(
+        `[llm-sdk] openai/${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);
@@ -612,14 +822,19 @@ var OpenAIAdapter = class _OpenAIAdapter {
           name: openaiToolOptions.toolChoice.name
         }
       } : openaiToolOptions?.toolChoice;
+      const modelIdForPayload = request.config?.model || this.model;
       const payload = {
-        model: request.config?.model || this.model,
+        model: modelIdForPayload,
         messages,
         tools: tools.length > 0 ? tools : void 0,
         tool_choice: tools.length > 0 ? toolChoice : void 0,
         parallel_tool_calls: tools.length > 0 ? openaiToolOptions?.parallelToolCalls : void 0,
-        temperature: request.config?.temperature ?? this.config.temperature,
-        max_tokens: request.config?.maxTokens ?? this.config.maxTokens,
+        ...buildOpenAITokenParams(
+          modelIdForPayload,
+          request.config?.maxTokens ?? this.config.maxTokens,
+          request.config?.temperature ?? this.config.temperature
+        ),
+        response_format: toOpenAIResponseFormat(request.config?.responseFormat),
         stream: true,
         stream_options: { include_usage: true }
       };
@@ -761,14 +976,19 @@ var OpenAIAdapter = class _OpenAIAdapter {
         name: openaiToolOptions.toolChoice.name
       }
     } : openaiToolOptions?.toolChoice;
+    const modelIdForCompletePayload = request.config?.model || this.model;
     const payload = {
-      model: request.config?.model || this.model,
+      model: modelIdForCompletePayload,
       messages,
       tools: tools.length > 0 ? tools : void 0,
       tool_choice: tools.length > 0 ? toolChoice : void 0,
       parallel_tool_calls: tools.length > 0 ? openaiToolOptions?.parallelToolCalls : void 0,
-      temperature: request.config?.temperature ?? this.config.temperature,
-      max_tokens: request.config?.maxTokens ?? this.config.maxTokens,
+      ...buildOpenAITokenParams(
+        modelIdForCompletePayload,
+        request.config?.maxTokens ?? this.config.maxTokens,
+        request.config?.temperature ?? this.config.temperature
+      ),
+      response_format: toOpenAIResponseFormat(request.config?.responseFormat),
       stream: false
     };
     logProviderPayload("openai", "request payload", payload, request.debug);
@@ -1044,7 +1264,9 @@ var AnthropicAdapter = class {
    * Build common request options for both streaming and non-streaming
    */
   buildRequestOptions(request) {
-    const systemMessage = request.systemPrompt || "";
+    const responseFormat = request.config?.responseFormat;
+    const jsonObjectSuffix = responseFormat?.type === "json_object" ? "\n\nRespond with a single JSON object and no other text." : "";
+    const systemMessage = (request.systemPrompt || "") + jsonObjectSuffix;
     let messages;
     if (request.rawMessages && request.rawMessages.length > 0) {
       messages = this.convertToAnthropicMessages(request.rawMessages);
@@ -1125,32 +1347,58 @@ var AnthropicAdapter = class {
     if (serverToolConfiguration) {
       options.server_tool_configuration = serverToolConfiguration;
     }
-    if (this.config.thinking?.type === "enabled") {
+    const modelForThinking = request.config?.model || this.model;
+    const thinkingTranslation = toAnthropicThinking(
+      request.config?.reasoningEffort,
+      modelForThinking
+    );
+    const outputConfig = toAnthropicOutputConfig(responseFormat);
+    if (outputConfig || thinkingTranslation.outputConfigEffort) {
+      options.output_config = {
+        ...outputConfig ?? {},
+        ...thinkingTranslation.outputConfigEffort ? { effort: thinkingTranslation.outputConfigEffort } : {}
+      };
+    }
+    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",
@@ -1185,17 +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 = [];
@@ -1485,12 +1735,14 @@ var OllamaAdapter = class {
       if (this.config.options) {
         Object.assign(ollamaOptions, this.config.options);
       }
+      const ollamaFormat = toOllamaFormat(request.config?.responseFormat);
       const payload = {
         model: request.config?.model || this.model,
         messages,
         tools,
         stream: true,
-        options: ollamaOptions
+        options: ollamaOptions,
+        ...ollamaFormat !== void 0 ? { format: ollamaFormat } : {}
       };
       logProviderPayload("ollama", "request payload", payload, request.debug);
       const response = await fetch(`${this.baseUrl}/api/chat`, {
@@ -1781,6 +2033,12 @@ var GoogleAdapter = class {
     }
     const messageId = generateMessageId();
     yield { type: "message:start", id: messageId };
+    const responseFormat = request.config?.responseFormat;
+    const geminiSchema = toGeminiSchema(responseFormat);
+    const responseFormatGenConfig = responseFormat ? {
+      responseMimeType: "application/json",
+      ...geminiSchema ? { responseJsonSchema: geminiSchema } : {}
+    } : {};
     try {
       logProviderPayload(
         "google",
@@ -1792,7 +2050,8 @@ var GoogleAdapter = class {
           tools: toolsArray.length > 0 ? toolsArray : void 0,
           generationConfig: {
             temperature: request.config?.temperature ?? this.config.temperature,
-            maxOutputTokens: request.config?.maxTokens ?? this.config.maxTokens
+            maxOutputTokens: request.config?.maxTokens ?? this.config.maxTokens,
+            ...responseFormatGenConfig
           },
           messageParts: mergedContents[mergedContents.length - 1]?.parts
         },
@@ -1805,7 +2064,8 @@ var GoogleAdapter = class {
         tools: toolsArray.length > 0 ? toolsArray : void 0,
         generationConfig: {
           temperature: request.config?.temperature ?? this.config.temperature,
-          maxOutputTokens: request.config?.maxTokens ?? this.config.maxTokens
+          maxOutputTokens: request.config?.maxTokens ?? this.config.maxTokens,
+          ...responseFormatGenConfig
         }
       });
       const lastMessage = mergedContents[mergedContents.length - 1];
@@ -1972,6 +2232,12 @@ var GoogleAdapter = class {
       }
     }
     const tools = formatToolsForGemini(request.actions);
+    const responseFormat = request.config?.responseFormat;
+    const geminiSchema = toGeminiSchema(responseFormat);
+    const responseFormatGenConfig = responseFormat ? {
+      responseMimeType: "application/json",
+      ...geminiSchema ? { responseJsonSchema: geminiSchema } : {}
+    } : {};
     const payload = {
       model: modelId,
       history: mergedContents.slice(0, -1),
@@ -1979,7 +2245,8 @@ var GoogleAdapter = class {
       tools: tools ? [tools] : void 0,
       generationConfig: {
         temperature: request.config?.temperature ?? this.config.temperature,
-        maxOutputTokens: request.config?.maxTokens ?? this.config.maxTokens
+        maxOutputTokens: request.config?.maxTokens ?? this.config.maxTokens,
+        ...responseFormatGenConfig
       },
       messageParts: mergedContents[mergedContents.length - 1]?.parts
     };
@@ -1990,7 +2257,8 @@ var GoogleAdapter = class {
       tools: tools ? [tools] : void 0,
       generationConfig: {
         temperature: request.config?.temperature ?? this.config.temperature,
-        maxOutputTokens: request.config?.maxTokens ?? this.config.maxTokens
+        maxOutputTokens: request.config?.maxTokens ?? this.config.maxTokens,
+        ...responseFormatGenConfig
       }
     });
     const lastMessage = mergedContents[mergedContents.length - 1];
@@ -2129,6 +2397,7 @@ var AzureAdapter = class {
         tools,
         temperature: request.config?.temperature ?? this.config.temperature,
         max_tokens: request.config?.maxTokens ?? this.config.maxTokens,
+        response_format: toOpenAIResponseFormat(request.config?.responseFormat),
         stream: true
       };
       logProviderPayload("azure", "request payload", payload, request.debug);
@@ -2228,7 +2497,8 @@ var AzureAdapter = class {
       messages,
       tools,
       temperature: request.config?.temperature ?? this.config.temperature,
-      max_tokens: request.config?.maxTokens ?? this.config.maxTokens
+      max_tokens: request.config?.maxTokens ?? this.config.maxTokens,
+      response_format: toOpenAIResponseFormat(request.config?.responseFormat)
     };
     logProviderPayload("azure", "request payload", payload, request.debug);
     const response = await client.chat.completions.create(payload);