@yourgpt/llm-sdk 2.1.3 → 2.1.4-alpha.0

package/dist/adapters/index.js

@@ -1,6 +1,44 @@
 'use strict';
 
 // src/adapters/base.ts
+function stringifyForDebug(value) {
+  return JSON.stringify(
+    value,
+    (_key, currentValue) => {
+      if (typeof currentValue === "bigint") {
+        return currentValue.toString();
+      }
+      if (currentValue instanceof Error) {
+        return {
+          name: currentValue.name,
+          message: currentValue.message,
+          stack: currentValue.stack
+        };
+      }
+      return currentValue;
+    },
+    2
+  );
+}
+function logProviderPayload(provider, label, payload, enabled) {
+  if (!enabled) {
+    return;
+  }
+  if (label.toLowerCase().includes("stream ")) {
+    return;
+  }
+  try {
+    console.log(
+      `[llm-sdk:${provider}] ${label}
+${stringifyForDebug(payload)}`
+    );
+  } catch (error) {
+    console.log(
+      `[llm-sdk:${provider}] ${label} (failed to stringify payload)`,
+      error
+    );
+  }
+}
 function formatMessages(messages, systemPrompt) {
   const formatted = [];
   if (systemPrompt) {
@@ -38,9 +76,44 @@ function parameterToJsonSchema(param) {
         )
       ])
     );
+    schema.additionalProperties = false;
   }
   return schema;
 }
+function normalizeObjectJsonSchema(schema) {
+  if (!schema || typeof schema !== "object") {
+    return {
+      type: "object",
+      properties: {},
+      required: [],
+      additionalProperties: false
+    };
+  }
+  const normalized = { ...schema };
+  const type = normalized.type;
+  if (type === "object") {
+    const properties = normalized.properties && typeof normalized.properties === "object" && !Array.isArray(normalized.properties) ? normalized.properties : {};
+    normalized.properties = Object.fromEntries(
+      Object.entries(properties).map(([key, value]) => [
+        key,
+        normalizeObjectJsonSchema(value)
+      ])
+    );
+    const propertyKeys = Object.keys(properties);
+    const required = Array.isArray(normalized.required) ? normalized.required.filter(
+      (value) => typeof value === "string"
+    ) : [];
+    normalized.required = Array.from(/* @__PURE__ */ new Set([...required, ...propertyKeys]));
+    if (normalized.additionalProperties === void 0) {
+      normalized.additionalProperties = false;
+    }
+  } else if (type === "array" && normalized.items && typeof normalized.items === "object") {
+    normalized.items = normalizeObjectJsonSchema(
+      normalized.items
+    );
+  }
+  return normalized;
+}
 function formatTools(actions) {
   return actions.map((action) => ({
     type: "function",
@@ -55,7 +128,8 @@ function formatTools(actions) {
             parameterToJsonSchema(param)
           ])
         ) : {},
-        required: action.parameters ? Object.entries(action.parameters).filter(([, param]) => param.required).map(([key]) => key) : []
+        required: action.parameters ? Object.entries(action.parameters).filter(([, param]) => param.required).map(([key]) => key) : [],
+        additionalProperties: false
       }
     }
   }));
@@ -311,7 +385,152 @@ var OpenAIAdapter = class {
     }
     return this.client;
   }
+  shouldUseResponsesApi(request) {
+    return request.providerToolOptions?.openai?.nativeToolSearch?.enabled === true && request.providerToolOptions.openai.nativeToolSearch.useResponsesApi !== false && Array.isArray(request.toolDefinitions) && request.toolDefinitions.length > 0;
+  }
+  buildResponsesInput(request) {
+    const sourceMessages = request.rawMessages && request.rawMessages.length > 0 ? request.rawMessages : formatMessagesForOpenAI(request.messages, void 0);
+    const input = [];
+    for (const message of sourceMessages) {
+      if (message.role === "system") {
+        continue;
+      }
+      if (message.role === "assistant") {
+        const content = typeof message.content === "string" ? message.content : Array.isArray(message.content) ? message.content : message.content ? JSON.stringify(message.content) : "";
+        if (content) {
+          input.push({
+            type: "message",
+            role: "assistant",
+            content
+          });
+        }
+        const toolCalls = Array.isArray(message.tool_calls) ? message.tool_calls : [];
+        for (const toolCall of toolCalls) {
+          input.push({
+            type: "function_call",
+            call_id: toolCall.id,
+            name: toolCall.function?.name,
+            arguments: toolCall.function?.arguments ?? "{}"
+          });
+        }
+        continue;
+      }
+      if (message.role === "tool") {
+        input.push({
+          type: "function_call_output",
+          call_id: message.tool_call_id,
+          output: typeof message.content === "string" ? message.content : JSON.stringify(message.content ?? null)
+        });
+        continue;
+      }
+      input.push({
+        type: "message",
+        role: message.role === "developer" ? "developer" : "user",
+        content: typeof message.content === "string" ? message.content : Array.isArray(message.content) ? message.content : JSON.stringify(message.content ?? "")
+      });
+    }
+    return input;
+  }
+  buildResponsesTools(tools) {
+    const nativeTools = tools.filter((tool) => tool.available !== false).map((tool) => ({
+      type: "function",
+      name: tool.name,
+      description: tool.description,
+      parameters: normalizeObjectJsonSchema(
+        tool.inputSchema ?? {
+          type: "object",
+          properties: {},
+          required: []
+        }
+      ),
+      strict: true,
+      defer_loading: tool.deferLoading === true
+    }));
+    return [{ type: "tool_search" }, ...nativeTools];
+  }
+  parseResponsesResult(response) {
+    const content = typeof response?.output_text === "string" ? response.output_text : "";
+    const toolCalls = Array.isArray(response?.output) ? response.output.filter((item) => item?.type === "function_call").map((item) => ({
+      id: item.call_id ?? item.id ?? generateToolCallId(),
+      name: item.name,
+      args: (() => {
+        try {
+          return JSON.parse(item.arguments ?? "{}");
+        } catch {
+          return {};
+        }
+      })()
+    })) : [];
+    return {
+      content,
+      toolCalls,
+      usage: response?.usage ? {
+        promptTokens: response.usage.input_tokens ?? 0,
+        completionTokens: response.usage.output_tokens ?? 0,
+        totalTokens: response.usage.total_tokens ?? (response.usage.input_tokens ?? 0) + (response.usage.output_tokens ?? 0)
+      } : void 0,
+      rawResponse: response
+    };
+  }
+  async completeWithResponses(request) {
+    const client = await this.getClient();
+    const openaiToolOptions = request.providerToolOptions?.openai;
+    const payload = {
+      model: request.config?.model || this.model,
+      instructions: request.systemPrompt,
+      input: this.buildResponsesInput(request),
+      tools: this.buildResponsesTools(request.toolDefinitions ?? []),
+      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,
+      stream: false
+    };
+    logProviderPayload("openai", "request payload", payload, request.debug);
+    const response = await client.responses.create(payload);
+    logProviderPayload("openai", "response payload", response, request.debug);
+    return this.parseResponsesResult(response);
+  }
   async *stream(request) {
+    if (this.shouldUseResponsesApi(request)) {
+      const messageId2 = generateMessageId();
+      yield { type: "message:start", id: messageId2 };
+      try {
+        const result = await this.completeWithResponses(request);
+        if (result.content) {
+          yield { type: "message:delta", content: result.content };
+        }
+        for (const toolCall of result.toolCalls) {
+          yield {
+            type: "action:start",
+            id: toolCall.id,
+            name: toolCall.name
+          };
+          yield {
+            type: "action:args",
+            id: toolCall.id,
+            args: JSON.stringify(toolCall.args)
+          };
+        }
+        yield { type: "message:end" };
+        yield {
+          type: "done",
+          usage: result.usage ? {
+            prompt_tokens: result.usage.promptTokens,
+            completion_tokens: result.usage.completionTokens,
+            total_tokens: result.usage.totalTokens
+          } : void 0
+        };
+        return;
+      } catch (error) {
+        yield {
+          type: "error",
+          message: error instanceof Error ? error.message : "Unknown error",
+          code: "OPENAI_RESPONSES_ERROR"
+        };
+        return;
+      }
+    }
     const client = await this.getClient();
     let messages;
     if (request.rawMessages && request.rawMessages.length > 0) {
@@ -376,20 +595,32 @@ var OpenAIAdapter = class {
     const messageId = generateMessageId();
     yield { type: "message:start", id: messageId };
     try {
-      const stream = await client.chat.completions.create({
+      const openaiToolOptions = request.providerToolOptions?.openai;
+      const toolChoice = openaiToolOptions?.toolChoice && typeof openaiToolOptions.toolChoice === "object" ? {
+        type: "function",
+        function: {
+          name: openaiToolOptions.toolChoice.name
+        }
+      } : openaiToolOptions?.toolChoice;
+      const payload = {
         model: request.config?.model || this.model,
         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,
         stream: true,
         stream_options: { include_usage: true }
-      });
+      };
+      logProviderPayload("openai", "request payload", payload, request.debug);
+      const stream = await client.chat.completions.create(payload);
       let currentToolCall = null;
       const collectedCitations = [];
       let citationIndex = 0;
       let usage;
       for await (const chunk of stream) {
+        logProviderPayload("openai", "stream chunk", chunk, request.debug);
         if (request.signal?.aborted) {
           break;
         }
@@ -471,6 +702,70 @@ var OpenAIAdapter = class {
       };
     }
   }
+  async complete(request) {
+    if (this.shouldUseResponsesApi(request)) {
+      return this.completeWithResponses(request);
+    }
+    const client = await this.getClient();
+    let messages;
+    if (request.rawMessages && request.rawMessages.length > 0) {
+      messages = request.rawMessages;
+      if (request.systemPrompt && !messages.some((message2) => message2.role === "system")) {
+        messages = [
+          { role: "system", content: request.systemPrompt },
+          ...messages
+        ];
+      }
+    } else {
+      messages = formatMessagesForOpenAI(
+        request.messages,
+        request.systemPrompt
+      );
+    }
+    const tools = request.actions?.length ? formatTools(request.actions) : [];
+    const openaiToolOptions = request.providerToolOptions?.openai;
+    const toolChoice = openaiToolOptions?.toolChoice && typeof openaiToolOptions.toolChoice === "object" ? {
+      type: "function",
+      function: {
+        name: openaiToolOptions.toolChoice.name
+      }
+    } : openaiToolOptions?.toolChoice;
+    const payload = {
+      model: request.config?.model || this.model,
+      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,
+      stream: false
+    };
+    logProviderPayload("openai", "request payload", payload, request.debug);
+    const response = await client.chat.completions.create(payload);
+    logProviderPayload("openai", "response payload", response, request.debug);
+    const choice = response.choices?.[0];
+    const message = choice?.message;
+    return {
+      content: message?.content ?? "",
+      toolCalls: message?.tool_calls?.map((toolCall) => ({
+        id: toolCall.id ?? generateToolCallId(),
+        name: toolCall.function?.name ?? "",
+        args: (() => {
+          try {
+            return JSON.parse(toolCall.function?.arguments ?? "{}");
+          } catch {
+            return {};
+          }
+        })()
+      })) ?? [],
+      usage: response.usage ? {
+        promptTokens: response.usage.prompt_tokens,
+        completionTokens: response.usage.completion_tokens,
+        totalTokens: response.usage.total_tokens
+      } : void 0,
+      rawResponse: response
+    };
+  }
 };
 function extractDomain(url) {
   try {
@@ -693,6 +988,26 @@ var AnthropicAdapter = class {
     }
     return messages;
   }
+  buildNativeSearchTools(tools, variant = "bm25") {
+    const nativeSearchTool = variant === "regex" ? {
+      type: "tool_search_tool_regex_20251119",
+      name: "tool_search_tool_regex"
+    } : {
+      type: "tool_search_tool_bm25_20251119",
+      name: "tool_search_tool_bm25"
+    };
+    const providerTools = tools.filter((tool) => tool.available !== false).map((tool) => ({
+      name: tool.name,
+      description: tool.description,
+      input_schema: tool.inputSchema ?? {
+        type: "object",
+        properties: {},
+        required: []
+      },
+      defer_loading: tool.deferLoading === true
+    }));
+    return [nativeSearchTool, ...providerTools];
+  }
   /**
    * Build common request options for both streaming and non-streaming
    */
@@ -705,7 +1020,11 @@ var AnthropicAdapter = class {
       const formatted = formatMessagesForAnthropic(request.messages, void 0);
       messages = formatted.messages;
     }
-    const tools = request.actions?.map((action) => ({
+    const anthropicNativeSearch = request.providerToolOptions?.anthropic?.nativeToolSearch;
+    const tools = anthropicNativeSearch?.enabled ? this.buildNativeSearchTools(
+      request.toolDefinitions ?? [],
+      anthropicNativeSearch.variant
+    ) : request.actions?.map((action) => ({
       name: action.name,
       description: action.description,
       input_schema: {
@@ -758,6 +1077,19 @@ var AnthropicAdapter = class {
       messages,
       tools: tools.length ? tools : void 0
     };
+    const anthropicToolOptions = request.providerToolOptions?.anthropic;
+    if (tools.length > 0 && anthropicToolOptions) {
+      if (anthropicToolOptions.toolChoice || anthropicToolOptions.disableParallelToolUse !== void 0) {
+        const toolChoice = typeof anthropicToolOptions.toolChoice === "object" ? {
+          type: "tool",
+          name: anthropicToolOptions.toolChoice.name
+        } : anthropicToolOptions.toolChoice ? { type: anthropicToolOptions.toolChoice } : { type: "auto" };
+        if (anthropicToolOptions.disableParallelToolUse !== void 0) {
+          toolChoice.disable_parallel_tool_use = anthropicToolOptions.disableParallelToolUse;
+        }
+        options.tool_choice = toolChoice;
+      }
+    }
     if (serverToolConfiguration) {
       options.server_tool_configuration = serverToolConfiguration;
     }
@@ -780,7 +1112,19 @@ var AnthropicAdapter = class {
       stream: false
     };
     try {
+      logProviderPayload(
+        "anthropic",
+        "request payload",
+        nonStreamingOptions,
+        request.debug
+      );
       const response = await client.messages.create(nonStreamingOptions);
+      logProviderPayload(
+        "anthropic",
+        "response payload",
+        response,
+        request.debug
+      );
       let content = "";
       let thinking = "";
       const toolCalls = [];
@@ -813,6 +1157,12 @@ var AnthropicAdapter = class {
     const messageId = generateMessageId();
     yield { type: "message:start", id: messageId };
     try {
+      logProviderPayload(
+        "anthropic",
+        "request payload",
+        options,
+        request.debug
+      );
       const stream = await client.messages.stream(options);
       let currentToolUse = null;
       let isInThinkingBlock = false;
@@ -820,6 +1170,7 @@ var AnthropicAdapter = class {
       let citationIndex = 0;
       let usage;
       for await (const event of stream) {
+        logProviderPayload("anthropic", "stream event", event, request.debug);
         if (request.signal?.aborted) {
           break;
         }
@@ -1102,18 +1453,20 @@ var OllamaAdapter = class {
       if (this.config.options) {
         Object.assign(ollamaOptions, this.config.options);
       }
+      const payload = {
+        model: request.config?.model || this.model,
+        messages,
+        tools,
+        stream: true,
+        options: ollamaOptions
+      };
+      logProviderPayload("ollama", "request payload", payload, request.debug);
       const response = await fetch(`${this.baseUrl}/api/chat`, {
         method: "POST",
         headers: {
           "Content-Type": "application/json"
         },
-        body: JSON.stringify({
-          model: request.config?.model || this.model,
-          messages,
-          tools,
-          stream: true,
-          options: ollamaOptions
-        }),
+        body: JSON.stringify(payload),
         signal: request.signal
       });
       if (!response.ok) {
@@ -1138,6 +1491,7 @@ var OllamaAdapter = class {
           if (!line.trim()) continue;
           try {
             const chunk = JSON.parse(line);
+            logProviderPayload("ollama", "stream chunk", chunk, request.debug);
             if (chunk.message?.content) {
               yield { type: "message:delta", content: chunk.message.content };
             }
@@ -1391,6 +1745,22 @@ var GoogleAdapter = class {
     const messageId = generateMessageId();
     yield { type: "message:start", id: messageId };
     try {
+      logProviderPayload(
+        "google",
+        "request payload",
+        {
+          model: modelId,
+          history: mergedContents.slice(0, -1),
+          systemInstruction: systemInstruction ? { parts: [{ text: systemInstruction }] } : void 0,
+          tools: toolsArray.length > 0 ? toolsArray : void 0,
+          generationConfig: {
+            temperature: request.config?.temperature ?? this.config.temperature,
+            maxOutputTokens: request.config?.maxTokens ?? this.config.maxTokens
+          },
+          messageParts: mergedContents[mergedContents.length - 1]?.parts
+        },
+        request.debug
+      );
       const chat = model.startChat({
         history: mergedContents.slice(0, -1),
         // All but the last message
@@ -1406,6 +1776,7 @@ var GoogleAdapter = class {
       let currentToolCall = null;
       const collectedCitations = [];
       for await (const chunk of result.stream) {
+        logProviderPayload("google", "stream chunk", chunk, request.debug);
         if (request.signal?.aborted) {
           break;
         }
@@ -1468,6 +1839,12 @@ var GoogleAdapter = class {
       let usage;
       try {
         const response = await result.response;
+        logProviderPayload(
+          "google",
+          "response payload",
+          response,
+          request.debug
+        );
         if (response.usageMetadata) {
           usage = {
             prompt_tokens: response.usageMetadata.promptTokenCount || 0,
@@ -1547,6 +1924,18 @@ var GoogleAdapter = class {
       }
     }
     const tools = formatToolsForGemini(request.actions);
+    const payload = {
+      model: modelId,
+      history: mergedContents.slice(0, -1),
+      systemInstruction: systemInstruction ? { parts: [{ text: systemInstruction }] } : void 0,
+      tools: tools ? [tools] : void 0,
+      generationConfig: {
+        temperature: request.config?.temperature ?? this.config.temperature,
+        maxOutputTokens: request.config?.maxTokens ?? this.config.maxTokens
+      },
+      messageParts: mergedContents[mergedContents.length - 1]?.parts
+    };
+    logProviderPayload("google", "request payload", payload, request.debug);
     const chat = model.startChat({
       history: mergedContents.slice(0, -1),
       systemInstruction: systemInstruction ? { parts: [{ text: systemInstruction }] } : void 0,
@@ -1559,6 +1948,7 @@ var GoogleAdapter = class {
     const lastMessage = mergedContents[mergedContents.length - 1];
     const result = await chat.sendMessage(lastMessage.parts);
     const response = result.response;
+    logProviderPayload("google", "response payload", response, request.debug);
     let textContent = "";
     const toolCalls = [];
     const candidate = response.candidates?.[0];
@@ -1663,16 +2053,19 @@ var XAIAdapter = class {
     const messageId = generateMessageId();
     yield { type: "message:start", id: messageId };
     try {
-      const stream = await client.chat.completions.create({
+      const payload = {
         model: request.config?.model || this.model,
         messages,
         tools,
         temperature: request.config?.temperature ?? this.config.temperature,
         max_tokens: request.config?.maxTokens ?? this.config.maxTokens,
         stream: true
-      });
+      };
+      logProviderPayload("xai", "request payload", payload, request.debug);
+      const stream = await client.chat.completions.create(payload);
       let currentToolCall = null;
       for await (const chunk of stream) {
+        logProviderPayload("xai", "stream chunk", chunk, request.debug);
         if (request.signal?.aborted) {
           break;
         }
@@ -1749,13 +2142,16 @@ var XAIAdapter = class {
       );
     }
     const tools = request.actions?.length ? formatTools(request.actions) : void 0;
-    const response = await client.chat.completions.create({
+    const payload = {
       model: request.config?.model || this.model,
       messages,
       tools,
       temperature: request.config?.temperature ?? this.config.temperature,
       max_tokens: request.config?.maxTokens ?? this.config.maxTokens
-    });
+    };
+    logProviderPayload("xai", "request payload", payload, request.debug);
+    const response = await client.chat.completions.create(payload);
+    logProviderPayload("xai", "response payload", response, request.debug);
     const choice = response.choices[0];
     const message = choice?.message;
     const toolCalls = (message?.tool_calls || []).map((tc) => ({
@@ -1851,7 +2247,7 @@ var AzureAdapter = class {
     const messageId = generateMessageId();
     yield { type: "message:start", id: messageId };
     try {
-      const stream = await client.chat.completions.create({
+      const payload = {
         // Azure uses deployment name, not model name
         model: this.config.deploymentName,
         messages,
@@ -1859,9 +2255,12 @@ var AzureAdapter = class {
         temperature: request.config?.temperature ?? this.config.temperature,
         max_tokens: request.config?.maxTokens ?? this.config.maxTokens,
         stream: true
-      });
+      };
+      logProviderPayload("azure", "request payload", payload, request.debug);
+      const stream = await client.chat.completions.create(payload);
       let currentToolCall = null;
       for await (const chunk of stream) {
+        logProviderPayload("azure", "stream chunk", chunk, request.debug);
         if (request.signal?.aborted) {
           break;
         }
@@ -1938,13 +2337,16 @@ var AzureAdapter = class {
       );
     }
     const tools = request.actions?.length ? formatTools(request.actions) : void 0;
-    const response = await client.chat.completions.create({
+    const payload = {
       model: this.config.deploymentName,
       messages,
       tools,
       temperature: request.config?.temperature ?? this.config.temperature,
       max_tokens: request.config?.maxTokens ?? this.config.maxTokens
-    });
+    };
+    logProviderPayload("azure", "request payload", payload, request.debug);
+    const response = await client.chat.completions.create(payload);
+    logProviderPayload("azure", "response payload", response, request.debug);
     const choice = response.choices[0];
     const message = choice?.message;
     const toolCalls = (message?.tool_calls || []).map((tc) => ({