aiexecode 1.0.91 → 1.0.92

package/src/LLMClient/client.js

@@ -0,0 +1,992 @@
+/**
+ * Unified LLM Client with Responses API Interface
+ * Supports OpenAI, Claude, Gemini, and Ollama with OpenAI Responses API-compatible interface
+ */
+
+import fs from 'fs';
+import path from 'path';
+import OpenAI from 'openai';
+import Anthropic from '@anthropic-ai/sdk';
+import { GoogleGenerativeAI } from '@google/generative-ai';
+
+import { convertResponsesRequestToClaudeFormat, convertClaudeResponseToResponsesFormat } from './converters/responses-to-claude.js';
+import { convertResponsesRequestToGeminiFormat, convertGeminiResponseToResponsesFormat } from './converters/responses-to-gemini.js';
+import { convertResponsesRequestToOllamaFormat, convertOllamaResponseToResponsesFormat } from './converters/responses-to-ollama.js';
+import { normalizeInput } from './converters/input-normalizer.js';
+import { normalizeError, createErrorFromResponse } from './errors.js';
+
+export class UnifiedLLMClient {
+  /**
+   * Create a unified LLM client
+   * @param {Object} config - Configuration object
+   * @param {string} config.provider - Provider name: 'openai', 'claude', 'gemini', or 'ollama' (auto-detected from model if not provided)
+   * @param {string} [config.apiKey] - API key for the provider
+   * @param {string} [config.baseUrl] - Base URL (for Ollama)
+   * @param {string} [config.model] - Default model to use
+   * @param {string} [config.logDir] - Directory path for logging request/response payloads (if not provided, logging is disabled)
+   */
+  constructor(config = {}) {
+    this.provider = config.provider;
+    this.apiKey = config.apiKey;
+    this.baseUrl = config.baseUrl;
+    this.defaultModel = config.model;
+    this.logDir = config.logDir;
+    this.explicitProvider = !!config.provider; // Track if provider was explicitly set
+
+    // Auto-detect provider from model if not explicitly provided
+    if (!this.provider && this.defaultModel) {
+      this.provider = this._detectProviderFromModel(this.defaultModel);
+    }
+
+    // Default to openai if still no provider
+    this.provider = this.provider || 'openai';
+
+    // Create log directory if specified and doesn't exist
+    if (this.logDir) {
+      try {
+        if (!fs.existsSync(this.logDir)) {
+          fs.mkdirSync(this.logDir, { recursive: true });
+        }
+      } catch (error) {
+        console.error(`Failed to create log directory: ${error.message}`);
+        this.logDir = null; // Disable logging if directory creation fails
+      }
+    }
+
+    this._initializeClient();
+  }
+
+  /**
+   * Auto-detect provider from model name
+   */
+  _detectProviderFromModel(model) {
+    if (model.startsWith('gpt-') || model.startsWith('o1-') || model.startsWith('o3-') || model.startsWith('text-')) {
+      return 'openai';
+    } else if (model.startsWith('claude-')) {
+      return 'claude';
+    } else if (model.startsWith('gemini-')) {
+      return 'gemini';
+    } else if (model.includes('llama') || model.includes('mistral') || model.includes('codellama')) {
+      return 'ollama';
+    }
+    return null;
+  }
+
+  /**
+   * Check if model is GPT-5, o3, or other reasoning models
+   */
+  _isReasoningModel(model) {
+    return model === 'gpt-5' || model === 'gpt-5-mini' || model === 'gpt-5-nano' ||
+           model === 'o3' || model === 'o3-mini';
+  }
+
+  _initializeClient() {
+    switch (this.provider) {
+      case 'openai':
+        this.client = new OpenAI({ apiKey: this.apiKey });
+        break;
+
+      case 'claude':
+        this.client = new Anthropic({ apiKey: this.apiKey });
+        break;
+
+      case 'gemini':
+        this.client = new GoogleGenerativeAI(this.apiKey);
+        break;
+
+      case 'ollama':
+        // Ollama doesn't have an official SDK, we'll use fetch
+        this.baseUrl = this.baseUrl || 'http://localhost:11434';
+        break;
+
+      default:
+        throw new Error(`Unsupported provider: ${this.provider}`);
+    }
+  }
+
+  /**
+   * Log payload to file
+   * @param {Object} payload - Request or response payload
+   * @param {string} type - 'REQ' or 'RES'
+   * @param {string} providerName - Provider name
+   */
+  _logPayload(payload, type, providerName) {
+    if (!this.logDir) return;
+
+    try {
+      const now = new Date();
+      const year = now.getFullYear();
+      const month = String(now.getMonth() + 1).padStart(2, '0');
+      const day = String(now.getDate()).padStart(2, '0');
+      const hours = String(now.getHours()).padStart(2, '0');
+      const minutes = String(now.getMinutes()).padStart(2, '0');
+      const seconds = String(now.getSeconds()).padStart(2, '0');
+      const milliseconds = String(now.getMilliseconds()).padStart(3, '0');
+
+      const timestamp = `${year}${month}${day}-${hours}${minutes}${seconds}-${milliseconds}`;
+      const filename = `${timestamp}-${type}-${providerName}.json`;
+      const filepath = path.join(this.logDir, filename);
+
+      fs.writeFileSync(filepath, JSON.stringify(payload, null, 2), 'utf-8');
+    } catch (error) {
+      console.error(`Failed to log payload: ${error.message}`);
+    }
+  }
+
+  /**
+   * Create a response (Responses API-compatible interface)
+   * @param {Object} request - Responses API format request
+   * @param {Object} options - Additional options
+   * @param {AbortSignal} options.signal - AbortSignal for cancellation
+   * @returns {Promise<Object>|AsyncGenerator} Responses API format response or stream
+   */
+  async response(request, options = {}) {
+    // Set default model if not provided
+    if (!request.model && this.defaultModel) {
+      request.model = this.defaultModel;
+    }
+
+    // Convert instructions field to input array (Method 2)
+    // If instructions is provided, add it as a system role message at the beginning of input array
+    if (request.instructions) {
+      // Ensure input is an array
+      if (!request.input) {
+        request.input = [];
+      } else if (typeof request.input === 'string') {
+        // Convert string input to array format
+        request.input = [
+          {
+            role: 'user',
+            content: [{ type: 'input_text', text: request.input }]
+          }
+        ];
+      }
+
+      // Check if input already has a system message
+      const hasSystemMessage = Array.isArray(request.input) &&
+                                request.input.some(item => item.role === 'system');
+
+      if (!hasSystemMessage) {
+        // Add instructions as system message at the beginning
+        const systemMessage = {
+          role: 'system',
+          content: typeof request.instructions === 'string'
+            ? [{ type: 'input_text', text: request.instructions }]
+            : request.instructions
+        };
+
+        request.input.unshift(systemMessage);
+      }
+
+      // Remove instructions field to avoid duplication
+      delete request.instructions;
+    }
+
+    // Auto-detect provider from model name in request
+    // Only if provider was not explicitly set in constructor
+    let effectiveProvider = this.provider;
+    if (request.model && !this.explicitProvider) {
+      const detectedProvider = this._detectProviderFromModel(request.model);
+      if (detectedProvider) {
+        effectiveProvider = detectedProvider;
+        // Re-initialize client if provider changed
+        if (effectiveProvider !== this.provider) {
+          const oldProvider = this.provider;
+          this.provider = effectiveProvider;
+          this._initializeClient();
+        }
+      }
+    }
+
+    // Log request payload
+    this._logPayload(request, 'REQ', effectiveProvider);
+
+    // Check if streaming is requested
+    const isStreaming = request.stream === true;
+
+    switch (effectiveProvider) {
+      case 'openai':
+        return isStreaming ? this._responseOpenAIStream(request, options) : await this._responseOpenAI(request, options);
+
+      case 'claude':
+        return isStreaming ? this._responseClaudeStream(request, options) : await this._responseClaude(request, options);
+
+      case 'gemini':
+        return isStreaming ? this._responseGeminiStream(request, options) : await this._responseGemini(request, options);
+
+      case 'ollama':
+        return isStreaming ? this._responseOllamaStream(request, options) : await this._responseOllama(request, options);
+
+      default:
+        throw new Error(`Unsupported provider: ${this.provider}`);
+    }
+  }
+
+  async _responseOpenAI(request, options = {}) {
+    try {
+      // Call OpenAI Responses API directly
+      const response = await this._callOpenAIResponsesAPI(request, options);
+      // Log response payload
+      this._logPayload(response, 'RES', 'openai');
+      return response;
+    } catch (error) {
+      throw normalizeError(error, 'openai');
+    }
+  }
+
+  async _callOpenAIResponsesAPI(request, options = {}) {
+    // Convert to proper Responses API format (like ai_request.js)
+    const openAIRequest = { ...request };
+
+    // 1. Normalize input to Responses API format
+    // Convert from string or Chat Completions format to proper input array
+    if (openAIRequest.input) {
+      openAIRequest.input = normalizeInput(openAIRequest.input);
+
+      // OpenAI Responses API uses role-based input with content arrays
+      // No conversion needed - keep the format as-is
+    }
+
+    // 2. Tools conversion
+    // OpenAI uses: { type: 'function', name, description, parameters }
+    // Not: { type: 'custom', name, description, input_schema }
+    if (openAIRequest.tools && Array.isArray(openAIRequest.tools)) {
+      openAIRequest.tools = openAIRequest.tools.map(tool => {
+        let convertedTool;
+
+        if (tool.type === 'custom' && tool.input_schema) {
+          // Convert custom format to OpenAI Responses API format
+          convertedTool = {
+            type: 'function',
+            name: tool.name,
+            description: tool.description || `Tool: ${tool.name}`,
+            parameters: tool.input_schema
+          };
+        } else if (tool.type === 'function' && tool.function) {
+          // Convert Chat Completions format to Responses API format
+          convertedTool = {
+            type: 'function',
+            name: tool.function.name,
+            description: tool.function.description || `Function: ${tool.function.name}`,
+            parameters: tool.function.parameters
+          };
+        } else {
+          // Already in correct format or pass through
+          convertedTool = { ...tool };
+        }
+
+        // Remove 'strict' field to prevent OpenAI SDK from auto-modifying 'required' fields
+        // This ensures all providers receive the same tool schema
+        delete convertedTool.strict;
+
+        return convertedTool;
+      });
+    }
+
+    // Add reasoning configuration for reasoning models (like ai_request.js)
+    // Check if model supports reasoning
+    const reasoningModels = ['gpt-5', 'gpt-5-mini', 'gpt-5-nano', 'o3', 'o3-mini'];
+    const currentModel = openAIRequest.model || 'gpt-4o-mini';
+
+    if (reasoningModels.some(m => currentModel.startsWith(m))) {
+      // Add reasoning configuration if not already present
+      if (!openAIRequest.reasoning) {
+        openAIRequest.reasoning = {
+          effort: 'medium',
+          summary: 'auto'
+        };
+      }
+    }
+
+    // Log raw request payload before API call
+    this._logPayload(openAIRequest, 'REQ-OPENAI-RAW', 'openai');
+
+    // Use OpenAI SDK to call Responses API (like ai_request.js)
+    const createOptions = {};
+    if (options.signal) {
+      createOptions.signal = options.signal;
+    }
+    const data = await this.client.responses.create(openAIRequest, createOptions);
+
+    // Log raw OpenAI API response before normalization
+    this._logPayload(data, 'RES-OPENAI-RAW', 'openai');
+
+    // Ensure all required fields are present and properly formatted
+    const normalizedResponse = {
+      id: data.id,
+      object: data.object || 'response',
+      created_at: data.created_at || Math.floor(Date.now() / 1000),
+      status: data.status || 'completed',
+      background: data.background !== undefined ? data.background : false,
+      billing: data.billing || { payer: 'developer' },
+      error: data.error || null,
+      incomplete_details: data.incomplete_details || null,
+      instructions: data.instructions || openAIRequest.instructions || null,
+      max_output_tokens: data.max_output_tokens || openAIRequest.max_output_tokens || null,
+      max_tool_calls: data.max_tool_calls || null,
+      model: data.model,
+      output: data.output || [],
+      parallel_tool_calls: data.parallel_tool_calls !== undefined ? data.parallel_tool_calls : true,
+      previous_response_id: data.previous_response_id || null,
+      prompt_cache_key: data.prompt_cache_key || null,
+      prompt_cache_retention: data.prompt_cache_retention || null,
+      reasoning: data.reasoning || { effort: openAIRequest.reasoning?.effort || null, summary: openAIRequest.reasoning?.summary || null },
+      safety_identifier: data.safety_identifier || null,
+      service_tier: data.service_tier || 'default',
+      store: data.store !== undefined ? data.store : (openAIRequest.store !== undefined ? openAIRequest.store : true),
+      temperature: data.temperature !== undefined ? data.temperature : (openAIRequest.temperature !== undefined ? openAIRequest.temperature : 1),
+      text: data.text || { format: { type: 'text' }, verbosity: 'medium' },
+      tool_choice: data.tool_choice || openAIRequest.tool_choice || 'auto',
+      tools: data.tools || openAIRequest.tools || [],
+      top_logprobs: data.top_logprobs !== undefined ? data.top_logprobs : 0,
+      top_p: data.top_p !== undefined ? data.top_p : (openAIRequest.top_p !== undefined ? openAIRequest.top_p : 1),
+      truncation: data.truncation || 'disabled',
+      usage: data.usage || {
+        input_tokens: 0,
+        input_tokens_details: { cached_tokens: 0 },
+        output_tokens: 0,
+        output_tokens_details: { reasoning_tokens: 0 },
+        total_tokens: 0
+      },
+      user: data.user || null,
+      metadata: data.metadata || {},
+      output_text: data.output_text !== undefined ? data.output_text : ''
+    };
+
+    // Normalize output items
+    if (normalizedResponse.output && Array.isArray(normalizedResponse.output)) {
+      for (const item of normalizedResponse.output) {
+        // Ensure function_call items have proper structure
+        if (item.type === 'function_call') {
+          // Ensure id field exists
+          if (!item.id) {
+            item.id = `fc_${item.call_id || Date.now()}`;
+          }
+          // Ensure status field exists
+          if (!item.status) {
+            item.status = 'completed';
+          }
+          // Keep arguments as string (per spec)
+          if (typeof item.arguments !== 'string' && item.input) {
+            item.arguments = JSON.stringify(item.input);
+          }
+          // Ensure call_id exists
+          if (!item.call_id && item.id) {
+            item.call_id = item.id.replace(/^fc_/, 'call_');
+          }
+        }
+        // Ensure message items have proper structure
+        else if (item.type === 'message') {
+          if (!item.id) {
+            item.id = `msg_${Date.now()}`;
+          }
+          if (!item.status) {
+            item.status = 'completed';
+          }
+        }
+        // Ensure reasoning items have proper structure
+        else if (item.type === 'reasoning') {
+          if (!item.id) {
+            item.id = `rs_${Date.now()}`;
+          }
+        }
+      }
+    }
+
+    // Calculate output_text if not provided
+    if (normalizedResponse.output_text === '' && normalizedResponse.output) {
+      let outputText = '';
+      for (const item of normalizedResponse.output) {
+        if (item.type === 'message' && item.content) {
+          for (const content of item.content) {
+            if (content.type === 'output_text' && content.text) {
+              outputText += content.text;
+            }
+          }
+        }
+      }
+      normalizedResponse.output_text = outputText;
+    }
+
+    // Ensure usage has proper structure
+    if (!normalizedResponse.usage.input_tokens_details) {
+      normalizedResponse.usage.input_tokens_details = { cached_tokens: 0 };
+    }
+    if (!normalizedResponse.usage.output_tokens_details) {
+      normalizedResponse.usage.output_tokens_details = { reasoning_tokens: 0 };
+    }
+
+    return normalizedResponse;
+  }
+
+  async _responseClaude(request, options = {}) {
+    try {
+      const claudeRequest = convertResponsesRequestToClaudeFormat(request);
+      // Log raw request payload before API call
+      this._logPayload(claudeRequest, 'REQ-CLAUDE-RAW', 'claude');
+
+      // Use streaming internally to get the response
+      const streamOptions = {};
+      if (options.signal) {
+        streamOptions.signal = options.signal;
+      }
+      const stream = await this.client.messages.stream(claudeRequest, streamOptions);
+
+      // Accumulate streaming response into a complete response object
+      let messageId = null;
+      let model = request.model || 'claude-sonnet-4-5';
+      let stopReason = null;
+      let usage = { input_tokens: 0, output_tokens: 0 };
+      const contentBlocks = [];
+      let currentTextBlock = null;
+      let currentToolBlock = null;
+
+      for await (const event of stream) {
+        // Handle different stream event types
+        if (event.type === 'message_start') {
+          messageId = event.message.id;
+          model = event.message.model;
+          usage.input_tokens = event.message.usage?.input_tokens || 0;
+        } else if (event.type === 'content_block_start') {
+          if (event.content_block?.type === 'text') {
+            currentTextBlock = { type: 'text', text: '' };
+          } else if (event.content_block?.type === 'tool_use') {
+            currentToolBlock = {
+              type: 'tool_use',
+              id: event.content_block.id,
+              name: event.content_block.name,
+              input: {}
+            };
+          }
+        } else if (event.type === 'content_block_delta') {
+          if (event.delta?.type === 'text_delta') {
+            if (currentTextBlock) {
+              currentTextBlock.text += event.delta.text;
+            }
+          } else if (event.delta?.type === 'input_json_delta') {
+            if (currentToolBlock) {
+              // Accumulate JSON string for tool input
+              if (!currentToolBlock._inputJson) {
+                currentToolBlock._inputJson = '';
+              }
+              currentToolBlock._inputJson += event.delta.partial_json;
+            }
+          }
+        } else if (event.type === 'content_block_stop') {
+          if (currentTextBlock) {
+            contentBlocks.push(currentTextBlock);
+            currentTextBlock = null;
+          } else if (currentToolBlock) {
+            // Parse accumulated JSON for tool input
+            if (currentToolBlock._inputJson) {
+              try {
+                currentToolBlock.input = JSON.parse(currentToolBlock._inputJson);
+              } catch (e) {
+                currentToolBlock.input = {};
+              }
+              delete currentToolBlock._inputJson;
+            }
+            contentBlocks.push(currentToolBlock);
+            currentToolBlock = null;
+          }
+        } else if (event.type === 'message_delta') {
+          stopReason = event.delta?.stop_reason || stopReason;
+          usage.output_tokens = event.usage?.output_tokens || usage.output_tokens;
+        }
+      }
+
+      // Construct the complete Claude response object
+      const claudeResponse = {
+        id: messageId || `msg_${Date.now()}`,
+        type: 'message',
+        role: 'assistant',
+        content: contentBlocks,
+        model: model,
+        stop_reason: stopReason || 'end_turn',
+        usage: usage
+      };
+
+      // Log raw Claude API response before conversion
+      this._logPayload(claudeResponse, 'RES-CLAUDE-RAW', 'claude');
+      const response = convertClaudeResponseToResponsesFormat(claudeResponse, request.model, request);
+      // Log response payload
+      this._logPayload(response, 'RES', 'claude');
+      return response;
+    } catch (error) {
+      throw normalizeError(error, 'claude');
+    }
+  }
+
+  async _responseGemini(request, options = {}) {
+    try {
+      const geminiRequest = convertResponsesRequestToGeminiFormat(request);
+      // Log raw request payload before API call
+      this._logPayload(geminiRequest, 'REQ-RAW', 'gemini');
+
+      // Create model configuration
+      const modelConfig = { model: request.model || 'gemini-2.5-flash' };
+
+      // Add tools to model config if present
+      if (geminiRequest.tools && geminiRequest.tools.length > 0) {
+        modelConfig.tools = geminiRequest.tools;
+      }
+
+      // Add system instruction to model config if present
+      if (geminiRequest.systemInstruction) {
+        modelConfig.systemInstruction = geminiRequest.systemInstruction;
+      }
+
+      const model = this.client.getGenerativeModel(modelConfig);
+
+      // Prepare generateContent request
+      const generateRequest = {
+        contents: geminiRequest.contents
+      };
+
+      if (geminiRequest.generationConfig) {
+        generateRequest.generationConfig = geminiRequest.generationConfig;
+      }
+
+      if (geminiRequest.toolConfig) {
+        generateRequest.toolConfig = geminiRequest.toolConfig;
+      }
+
+      // Add abort signal if provided
+      if (options.signal) {
+        generateRequest.signal = options.signal;
+      }
+
+      const result = await model.generateContent(generateRequest);
+      const response = await result.response;
+
+      // Log raw Gemini response for debugging
+      this._logPayload({
+        candidates: response.candidates,
+        usageMetadata: response.usageMetadata,
+        promptFeedback: response.promptFeedback
+      }, 'RES-GEMINI-RAW', 'gemini');
+
+      // Convert to Responses API format
+      const convertedResponse = convertGeminiResponseToResponsesFormat({
+        candidates: [
+          {
+            content: response.candidates?.[0]?.content,
+            finishReason: response.candidates?.[0]?.finishReason
+          }
+        ],
+        usageMetadata: response.usageMetadata
+      }, request.model, request);
+      // Log response payload
+      this._logPayload(convertedResponse, 'RES', 'gemini');
+      return convertedResponse;
+    } catch (error) {
+      // Log error for debugging
+      this._logPayload({
+        error_message: error.message,
+        error_stack: error.stack,
+        error_name: error.name,
+        full_error: JSON.parse(JSON.stringify(error, Object.getOwnPropertyNames(error)))
+      }, 'ERROR-GEMINI', 'gemini');
+      throw normalizeError(error, 'gemini');
+    }
+  }
+
+  async _responseOllama(request, options = {}) {
+    try {
+      const { url, request: ollamaRequest } = convertResponsesRequestToOllamaFormat(request, this.baseUrl);
+      // Log raw request payload before API call
+      this._logPayload(ollamaRequest, 'REQ-RAW', 'ollama');
+
+      const fetchOptions = {
+        method: 'POST',
+        headers: {
+          'Content-Type': 'application/json'
+        },
+        body: JSON.stringify(ollamaRequest)
+      };
+
+      // Add abort signal if provided
+      if (options.signal) {
+        fetchOptions.signal = options.signal;
+      }
+
+      const response = await fetch(url, fetchOptions);
+
+      if (!response.ok) {
+        const error = await createErrorFromResponse(response, 'ollama');
+        throw error;
+      }
+
+      const ollamaResponse = await response.json();
+      // Log Ollama's raw response for debugging
+      this._logPayload(ollamaResponse, 'RES-RAW', 'ollama');
+      const convertedResponse = convertOllamaResponseToResponsesFormat(ollamaResponse, request.model, request);
+      // Log response payload
+      this._logPayload(convertedResponse, 'RES', 'ollama');
+      return convertedResponse;
+    } catch (error) {
+      throw normalizeError(error, 'ollama');
+    }
+  }
+
+  // ============================================
+  // Streaming Methods
+  // ============================================
+
+  /**
+   * OpenAI streaming response
+   * @param {Object} request - Responses API format request
+   * @param {Object} options - Additional options
+   * @returns {AsyncGenerator} Responses API format stream
+   */
+  async *_responseOpenAIStream(request, options = {}) {
+    const chunks = []; // Collect all chunks for logging
+    try {
+      // Prepare streaming request (same as non-streaming)
+      const streamRequest = { ...request, stream: true };
+
+      // 1. Normalize input to Responses API format
+      if (streamRequest.input) {
+        streamRequest.input = normalizeInput(streamRequest.input);
+      }
+
+      // 2. Tools conversion
+      if (streamRequest.tools && Array.isArray(streamRequest.tools)) {
+        streamRequest.tools = streamRequest.tools.map(tool => {
+          let convertedTool;
+
+          if (tool.type === 'custom' && tool.input_schema) {
+            convertedTool = {
+              type: 'function',
+              name: tool.name,
+              description: tool.description || `Tool: ${tool.name}`,
+              parameters: tool.input_schema
+            };
+          } else if (tool.type === 'function' && tool.function) {
+            convertedTool = {
+              type: 'function',
+              name: tool.function.name,
+              description: tool.function.description || `Function: ${tool.function.name}`,
+              parameters: tool.function.parameters
+            };
+          } else {
+            convertedTool = { ...tool };
+          }
+
+          // Remove 'strict' field to prevent OpenAI SDK from auto-modifying 'required' fields
+          delete convertedTool.strict;
+
+          return convertedTool;
+        });
+      }
+
+      // Add reasoning configuration for reasoning models
+      const reasoningModels = ['gpt-5', 'gpt-5-mini', 'gpt-5-nano', 'o3', 'o3-mini'];
+      const currentModel = streamRequest.model || 'gpt-4o-mini';
+
+      if (reasoningModels.some(m => currentModel.startsWith(m))) {
+        if (!streamRequest.reasoning) {
+          streamRequest.reasoning = {
+            effort: 'medium',
+            summary: 'auto'
+          };
+        }
+      }
+
+      // Log raw request payload before API call
+      this._logPayload(streamRequest, 'REQ-RAW', 'openai');
+
+      // Use OpenAI SDK for streaming (like ai_request.js)
+      const createOptions = {};
+      if (options.signal) {
+        createOptions.signal = options.signal;
+      }
+      const stream = await this.client.responses.create(streamRequest, createOptions);
+
+      // Stream the response chunks and convert to consistent format
+      const streamId = `resp_${Date.now()}`;
+      const created = Math.floor(Date.now() / 1000);
+      let currentMessageId = null;
+
+      for await (const chunk of stream) {
+        // Convert OpenAI SDK streaming format to our standard format
+        if (chunk.type === 'response.output_text.delta') {
+          // Text delta
+          const deltaChunk = {
+            id: streamId,
+            object: 'response.delta',
+            created_at: created,
+            model: streamRequest.model,
+            delta: {
+              type: 'output_text',
+              message_id: chunk.item_id,
+              text: chunk.delta
+            }
+          };
+          chunks.push(deltaChunk);
+          yield deltaChunk;
+        } else if (chunk.type === 'response.done' || chunk.type === 'response.completed') {
+          // Stream completed
+          const doneChunk = {
+            id: streamId,
+            object: 'response.done',
+            created_at: created,
+            model: streamRequest.model,
+            status: 'completed'
+          };
+          chunks.push(doneChunk);
+          // Log all chunks
+          this._logPayload(chunks, 'RES', 'openai');
+          yield doneChunk;
+        }
+        // Ignore other chunk types for now (response.created, response.in_progress, etc.)
+      }
+    } catch (error) {
+      throw normalizeError(error, 'openai');
+    }
+  }
+
+  /**
+   * Claude streaming response
+   * @param {Object} request - Responses API format request
+   * @param {Object} options - Additional options
+   * @returns {AsyncGenerator} Responses API format stream
+   */
+  async *_responseClaudeStream(request, options = {}) {
+    const chunks = []; // Collect all chunks for logging
+    try {
+      const claudeRequest = convertResponsesRequestToClaudeFormat(request);
+      // Log raw request payload before API call
+      this._logPayload(claudeRequest, 'REQ-RAW', 'claude');
+
+      const streamOptions = {};
+      if (options.signal) {
+        streamOptions.signal = options.signal;
+      }
+      const stream = await this.client.messages.stream(claudeRequest, streamOptions);
+
+      const streamId = `resp_${Date.now()}`;
+      const created = Math.floor(Date.now() / 1000);
+      let currentMessageId = `msg_${Date.now()}`;
+
+      for await (const event of stream) {
+        if (event.type === 'content_block_delta' && event.delta?.type === 'text_delta') {
+          // Convert to Responses API streaming format
+          const deltaChunk = {
+            id: streamId,
+            object: 'response.delta',
+            created_at: created,
+            model: request.model || 'claude-sonnet-4-5',
+            delta: {
+              type: 'output_text',
+              message_id: currentMessageId,
+              text: event.delta.text
+            }
+          };
+          chunks.push(deltaChunk);
+          yield deltaChunk;
+        } else if (event.type === 'message_delta' && event.delta?.stop_reason) {
+          // Final chunk
+          const doneChunk = {
+            id: streamId,
+            object: 'response.done',
+            created_at: created,
+            model: request.model || 'claude-sonnet-4-5',
+            status: 'completed'
+          };
+          chunks.push(doneChunk);
+          // Log all chunks
+          this._logPayload(chunks, 'RES', 'claude');
+          yield doneChunk;
+        }
+      }
+    } catch (error) {
+      throw normalizeError(error, 'claude');
+    }
+  }
+
+  /**
+   * Gemini streaming response
+   * @param {Object} request - Responses API format request
+   * @param {Object} options - Additional options
+   * @returns {AsyncGenerator} Responses API format stream
+   */
+  async *_responseGeminiStream(request, options = {}) {
+    const chunks = []; // Collect all chunks for logging
+    try {
+      const geminiRequest = convertResponsesRequestToGeminiFormat(request);
+      // Log raw request payload before API call
+      this._logPayload(geminiRequest, 'REQ-RAW', 'gemini');
+
+      // Create model configuration
+      const modelConfig = { model: request.model || 'gemini-2.5-flash' };
+
+      if (geminiRequest.tools && geminiRequest.tools.length > 0) {
+        modelConfig.tools = geminiRequest.tools;
+      }
+
+      if (geminiRequest.systemInstruction) {
+        modelConfig.systemInstruction = geminiRequest.systemInstruction;
+      }
+
+      const model = this.client.getGenerativeModel(modelConfig);
+
+      const generateRequest = {
+        contents: geminiRequest.contents
+      };
+
+      if (geminiRequest.generationConfig) {
+        generateRequest.generationConfig = geminiRequest.generationConfig;
+      }
+
+      if (geminiRequest.toolConfig) {
+        generateRequest.toolConfig = geminiRequest.toolConfig;
+      }
+
+      // Add abort signal if provided
+      if (options.signal) {
+        generateRequest.signal = options.signal;
+      }
+
+      const result = await model.generateContentStream(generateRequest);
+
+      const streamId = `resp_${Date.now()}`;
+      const created = Math.floor(Date.now() / 1000);
+      const messageId = `msg_${Date.now()}`;
+
+      for await (const chunk of result.stream) {
+        const text = chunk.text();
+
+        if (text) {
+          const deltaChunk = {
+            id: streamId,
+            object: 'response.delta',
+            created_at: created,
+            model: request.model || 'gemini-2.5-flash',
+            delta: {
+              type: 'output_text',
+              message_id: messageId,
+              text: text
+            }
+          };
+          chunks.push(deltaChunk);
+          yield deltaChunk;
+        }
+      }
+
+      // Final chunk
+      const doneChunk = {
+        id: streamId,
+        object: 'response.done',
+        created_at: created,
+        model: request.model || 'gemini-2.5-flash',
+        status: 'completed'
+      };
+      chunks.push(doneChunk);
+      // Log all chunks
+      this._logPayload(chunks, 'RES', 'gemini');
+      yield doneChunk;
+    } catch (error) {
+      throw normalizeError(error, 'gemini');
+    }
+  }
+
+  /**
+   * Ollama streaming response
+   * @param {Object} request - Responses API format request
+   * @param {Object} options - Additional options
+   * @returns {AsyncGenerator} Responses API format stream
+   */
+  async *_responseOllamaStream(request, options = {}) {
+    const chunks = []; // Collect all chunks for logging
+    try {
+      const { url, request: ollamaRequest } = convertResponsesRequestToOllamaFormat(request, this.baseUrl);
+
+      // Ensure stream is enabled
+      ollamaRequest.stream = true;
+
+      // Log raw request payload before API call
+      this._logPayload(ollamaRequest, 'REQ-RAW', 'ollama');
+
+      const fetchOptions = {
+        method: 'POST',
+        headers: {
+          'Content-Type': 'application/json'
+        },
+        body: JSON.stringify(ollamaRequest)
+      };
+
+      // Add abort signal if provided
+      if (options.signal) {
+        fetchOptions.signal = options.signal;
+      }
+
+      const response = await fetch(url, fetchOptions);
+
+      if (!response.ok) {
+        throw new Error(`Ollama API error: ${response.statusText}`);
+      }
+
+      const streamId = `resp_${Date.now()}`;
+      const created = Math.floor(Date.now() / 1000);
+      const messageId = `msg_${Date.now()}`;
+
+      // Parse streaming response (newline-delimited JSON)
+      const reader = response.body.getReader();
+      const decoder = new TextDecoder();
+      let buffer = '';
+
+      while (true) {
+        const { done, value } = await reader.read();
+
+        if (done) break;
+
+        buffer += decoder.decode(value, { stream: true });
+        const lines = buffer.split('\n');
+        buffer = lines.pop() || '';
+
+        for (const line of lines) {
+          if (line.trim()) {
+            try {
+              const chunk = JSON.parse(line);
+
+              if (chunk.message?.content) {
+                const deltaChunk = {
+                  id: streamId,
+                  object: 'response.delta',
+                  created_at: created,
+                  model: request.model,
+                  delta: {
+                    type: 'output_text',
+                    message_id: messageId,
+                    text: chunk.message.content
+                  }
+                };
+                chunks.push(deltaChunk);
+                yield deltaChunk;
+              }
+
+              if (chunk.done) {
+                const doneChunk = {
+                  id: streamId,
+                  object: 'response.done',
+                  created_at: created,
+                  model: request.model,
+                  status: 'completed'
+                };
+                chunks.push(doneChunk);
+                // Log all chunks
+                this._logPayload(chunks, 'RES', 'ollama');
+                yield doneChunk;
+              }
+            } catch (parseError) {
+              console.error('Error parsing Ollama stream chunk:', parseError);
+            }
+          }
+        }
+      }
+    } catch (error) {
+      throw normalizeError(error, 'ollama');
+    }
+  }
+}