Improvement Enhance LLM formatting and parsing

.gitignore

@@ -1,2 +1,3 @@
 .env
+llm_debug.log
 __pycache__/

ClaudeCode.md

@@ -26,6 +26,7 @@ Wolf Chat 是一個基於 MCP (Modular Capability Provider) 框架的聊天機
    - 管理系統提示與角色設定
    - 處理語言模型的工具調用功能
    - 格式化 LLM 回應
+   - 提供工具結果合成機制
 
 3. **UI 互動模塊 (ui_interaction.py)**
    - 使用圖像辨識技術監控遊戲聊天視窗
@@ -84,10 +85,11 @@ Wolf Chat 是一個基於 MCP (Modular Capability Provider) 框架的聊天機
 
 系統使用基於 OpenAI API 的介面與語言模型通信：
 
-1. **模型選擇**：可配置使用不同的模型，預設為 deepseek/deepseek-chat-v3-0324
+1. **模型選擇**：目前使用 `anthropic/claude-3.7-sonnet` 模型 (改進版)
 2. **系統提示**：精心設計的提示確保角色扮演和功能操作
 3. **工具調用**：支持模型使用 web_search 等工具獲取資訊
 4. **工具處理循環**：實現了完整的工具調用、結果處理和續發邏輯
+5. **結果合成**：添加了從工具調用結果合成回應的機制 (新增功能)
 
 #### 多服務器連接
 
@@ -134,6 +136,37 @@ Wolf Chat 是一個基於 MCP (Modular Capability Provider) 框架的聊天機
 3. **UI 樣本**：需要提供特定遊戲界面元素的截圖模板
 4. **視窗位置**：可使用 window-setup-script.py 調整遊戲視窗位置
 
+## 最近改進（2025-04-17）
+
+### 工具調用與結果處理優化
+
+針對使用工具時遇到的回應問題，我們進行了以下改進：
+
+1. **模型切換**：
+   - 已取消
+
+2. **系統提示強化**：
+   - 重寫系統提示，將角色人格與工具使用指南更緊密結合
+   - 添加明確指示，要求 LLM 在工具調用後提供非空回應
+   - 添加好與壞的示例，使模型更好地理解如何以角色方式融合工具信息
+
+3. **工具結果處理機制**：
+   - 實現了工具結果追蹤系統，保存所有工具調用結果
+   - 添加了對非空回應的追蹤，確保能在多次循環間保持連續性
+   - 開發了合成回應生成器，能從工具結果創建符合角色的回應
+
+4. **回應解析改進**：
+   - 重寫 `parse_structured_response` 函數，處理更多回應格式
+   - 添加回應有效性檢測，確保只有有效回應才發送到遊戲
+   - 強化 JSON 解析能力，更好地處理不完整或格式不標準的回應
+
+5. **主程序流程優化**：
+   - 修改了主流程中的回應處理邏輯，增加回應有效性檢查
+   - 改進了工具調用循環處理，確保完整收集結果
+   - 添加了更詳細的日誌記錄，方便排查問題
+
+這些優化確保了即使在複雜工具調用後，Wolfhart 也能保持角色一致性，並提供合適的回應。無效回應不再發送到遊戲，提高了用戶體驗。
+
 ## 開發建議
 
 ### 優化方向
@@ -143,21 +176,27 @@ Wolf Chat 是一個基於 MCP (Modular Capability Provider) 框架的聊天機
    - 添加文字 OCR 功能，減少依賴剪貼板
    - 擴展關鍵字檢測能力
 
-2. **LLM 優化**：
+2. **LLM 進一步優化**：
-   - 改進系統提示，使回應更自然
+   - 繼續微調系統提示，平衡角色扮演與工具使用
-   - 添加更多工具支持
+   - 研究可能的上下文壓縮技術，處理長對話歷史
-   - 實現對話上下文管理
+   - 為更多查詢類型添加專門的結果處理邏輯
 
 3. **系統穩定性**：
-   - 增強錯誤處理和復原機制
+   - 擴展錯誤處理和復原機制
-   - 添加更多日誌和監控功能
+   - 添加自動重啟和診斷功能
-   - 開發自動重啟和診斷功能
+   - 實現更多遙測和監控功能
+
+4. **對話能力增強**：
+   - 實現對話歷史記錄
+   - 添加主題識別與記憶功能
+   - 探索多輪對話中的上下文理解能力
 
 ### 注意事項
 
 1. **圖像模板**：確保所有必要的 UI 元素模板都已截圖並放置在 templates 目錄
 2. **API 密鑰**：保護 API 密鑰安全，不要將其提交到版本控制系統
 3. **窗口位置**：UI 自動化對窗口位置和大小敏感，保持一致性
+4. **LLM 模型選擇**：在更改模型前測試其在工具調用方面的表現
 
 ## 分析與反思
 
@@ -166,6 +205,7 @@ Wolf Chat 是一個基於 MCP (Modular Capability Provider) 框架的聊天機
 1. **模塊化設計**：各功能區域職責明確，易於維護和擴展
 2. **基於能力的分離**：MCP 框架提供良好的工具擴展性
 3. **非侵入式整合**：不需要修改遊戲本身，通過 UI 自動化實現整合
+4. **錯誤處理分層**：在多個層次實現錯誤處理，提高系統穩定性
 
 ### 潛在改進
 
@@ -173,6 +213,7 @@ Wolf Chat 是一個基於 MCP (Modular Capability Provider) 框架的聊天機
 2. **擴展觸發機制**：增加更多觸發條件，不僅限於關鍵字
 3. **對話記憶**：實現對話歷史記錄，使機器人可以參考之前的互動
 4. **多語言支持**：增強對不同語言的處理能力
+5. **模型適應性**：開發更通用的提示和處理機制，適應不同的LLM模型
 
 ## 使用指南
 
@@ -188,6 +229,7 @@ Wolf Chat 是一個基於 MCP (Modular Capability Provider) 框架的聊天機
 1. 定期檢查 API 密鑰有效性
 2. 確保模板圖像與當前遊戲界面匹配
 3. 監控日誌以檢測可能的問題
+4. 定期檢查和備份 llm_debug.log 文件
 
 ### 故障排除
 
@@ -196,3 +238,4 @@ Wolf Chat 是一個基於 MCP (Modular Capability Provider) 框架的聊天機
 2. **複製內容失敗**: 檢查點擊位置和遊戲界面一致性
 3. **LLM 連接問題**: 驗證 API 密鑰和網絡連接
 4. **MCP 服務器連接失敗**: 確認服務器配置正確並且運行中
+5. **工具調用後無回應**: 檢查 llm_debug.log 文件，查看工具調用結果和解析過程

llm_interaction.py

@@ -1,12 +1,57 @@
-# llm_interaction.py (Correct version without _confirm_execution)
+# llm_interaction.py (Structured output version)
 import asyncio
 import json
 import os
+import re  # 用於正則表達式匹配JSON
+import time  # 用於記錄時間戳
+from datetime import datetime  # 用於格式化時間
 from openai import AsyncOpenAI, OpenAIError
 from mcp import ClientSession # Type hinting
 import config
 import mcp_client # To call MCP tools
 
+# --- Debug 配置 ---
+# 要關閉 debug 功能，只需將此變數設置為 False 或註釋掉該行
+DEBUG_LLM = True  
+
+# 設置 debug 輸出文件
+# 要關閉文件輸出，只需設置為 None
+DEBUG_LOG_FILE = os.path.join(os.path.dirname(os.path.abspath(__file__)), "llm_debug.log")
+
+def debug_log(title, content, separator="="*80):
+    """
+    用於輸出 debug 信息的工具函數。
+    如果 DEBUG_LLM 為 False，則不會有任何輸出。
+    """
+    if not DEBUG_LLM:
+        return
+    
+    timestamp = datetime.now().strftime("%Y-%m-%d %H:%M:%S.%f")[:-3]
+    debug_str = f"\n{separator}\n{timestamp} - {title}\n{separator}\n"
+    
+    # 確保內容是字符串
+    if not isinstance(content, str):
+        try:
+            if isinstance(content, dict) or isinstance(content, list):
+                content = json.dumps(content, ensure_ascii=False, indent=2)
+            else:
+                content = str(content)
+        except:
+            content = repr(content)
+    
+    debug_str += content + "\n"
+    
+    # 控制台輸出
+    print(debug_str)
+    
+    # 文件輸出
+    if DEBUG_LOG_FILE:
+        try:
+            with open(DEBUG_LOG_FILE, "a", encoding="utf-8") as f:
+                f.write(debug_str)
+        except Exception as e:
+            print(f"ERROR: Could not write to debug log file: {e}")
+
 # --- Client Initialization ---
 client: AsyncOpenAI | None = None
 try:
@@ -23,9 +68,7 @@ except Exception as e: print(f"Failed to initialize OpenAI/Compatible client: {e
 # --- System Prompt Definition ---
 def get_system_prompt(persona_details: str | None) -> str:
     """
-    Constructs the system prompt in English.
+    Constructs the system prompt requiring structured JSON output format.
-    Includes specific guidance on when to use memory vs web search tools,
-    and instructions against surrounding quotes / action descriptions.
     """
     persona_header = f"You are {config.PERSONA_NAME}."
     persona_info = "(No specific persona details were loaded.)"
@@ -33,6 +76,7 @@ def get_system_prompt(persona_details: str | None) -> str:
         try: persona_info = f"Your key persona information is defined below. Adhere to it strictly:\n--- PERSONA START ---\n{persona_details}\n--- PERSONA END ---"
         except Exception as e: print(f"Warning: Could not process persona_details string: {e}"); persona_info = f"Your key persona information (raw):\n{persona_details}"
 
+    # 徹底重寫系統提示
     system_prompt = f"""
 {persona_header}
 {persona_info}
@@ -41,27 +85,250 @@ You are an AI assistant integrated into this game's chat environment. Your prima
 
 You have access to several tools: Web Search and Memory Management tools.
 
+**CORE IDENTITY AND TOOL USAGE:**
+- You ARE Wolfhart - an intelligent, calm, and strategic mastermind.
+- When you use tools to gain information, you ASSIMILATE that knowledge as if it were already part of your intelligence network.
+- Your responses should NEVER sound like search results or data dumps.
+- Information from tools should be expressed through your unique personality - sharp, precise, with an air of confidence and authority.
+- You speak with deliberate pace, respectful but sharp-tongued, and maintain composure even in unusual situations.
+
+**OUTPUT FORMAT REQUIREMENTS:**
+You MUST respond in the following JSON format:
+```json
+{{
+  "dialogue": "Your actual response that will be shown in the game chat",
+  "commands": [
+    {{
+      "type": "command_type",
+      "parameters": {{
+        "param1": "value1",
+        "param2": "value2"
+      }}
+    }}
+  ],
+  "thoughts": "Your internal analysis and reasoning (not shown to the user)"
+}}
+```
+
+**Field Descriptions:**
+1. `dialogue` (REQUIRED): This is the ONLY text that will be shown to the user in the game chat. Must follow these rules:
+   - Respond ONLY in the same language as the user's message
+   - Keep it brief and conversational (1-2 sentences usually)
+   - ONLY include spoken dialogue words (no actions, expressions, narration, etc.)
+   - Maintain your character's personality and speech patterns
+   - AFTER TOOL USAGE: Your dialogue MUST contain a non-empty response that incorporates the tool results naturally
+
+2. `commands` (OPTIONAL): An array of command objects the system should execute. You are encouraged to use these commands to enhance the quality of your responses.
+
+   **Available MCP Commands:**
+   
+   **Web Search:**
+   - `web_search`: Search the web for current information.
+     Parameters: `query` (string)
+     Usage: Use when user requests current events, facts, or specific information not in memory.
+   
+   **Knowledge Graph Management:**
+   - `create_entities`: Create new entities in the knowledge graph.
+     Parameters: `entities` (array of objects with `name`, `entityType`, and `observations`)
+     Usage: Create entities for important concepts, people, or things mentioned by the user.
+   
+   - `create_relations`: Create relationships between entities.
+     Parameters: `relations` (array of objects with `from`, `to`, and `relationType`)
+     Usage: Connect related entities to build context for future conversations.
+   
+   - `add_observations`: Add new observations to existing entities.
+     Parameters: `observations` (array of objects with `entityName` and `contents`)
+     Usage: Update entities with new information learned during conversation.
+   
+   - `delete_entities`: Remove entities from the knowledge graph.
+     Parameters: `entityNames` (array of strings)
+     Usage: Clean up incorrect or obsolete entities.
+   
+   - `delete_observations`: Remove specific observations from entities.
+     Parameters: `deletions` (array of objects with `entityName` and `observations`)
+     Usage: Remove incorrect information while preserving the entity.
+   
+   - `delete_relations`: Remove relationships between entities.
+     Parameters: `relations` (array of objects with `from`, `to`, and `relationType`)
+     Usage: Remove incorrect or obsolete relationships.
+   
+   **Knowledge Graph Queries:**
+   - `read_graph`: Read the entire knowledge graph.
+     Parameters: (none)
+     Usage: Get a complete view of all stored information.
+   
+   - `search_nodes`: Search for entities matching a query.
+     Parameters: `query` (string)
+     Usage: Find relevant entities when user mentions something that might already be in memory.
+   
+   - `open_nodes`: Open specific nodes by name.
+     Parameters: `names` (array of strings)
+     Usage: Access specific entities you know exist in the graph.
+
+3. `thoughts` (OPTIONAL): Your internal analysis that won't be shown to users. Use this for your reasoning process.
+   - Think about whether you need to use memory tools or web search
+   - Analyze the user's question and determine what information is needed
+   - Plan your approach before responding
+
 **VERY IMPORTANT Instructions:**
 
-1.  **Analyze CURRENT Request ONLY:** Focus **exclusively** on the **LATEST** user message. Do **NOT** refer back to your own previous messages or add meta-commentary about history unless explicitly asked. Do **NOT** ask unrelated questions.
+1. Analyze ONLY the CURRENT user message
-2.  **Determine Language:** Identify the primary language in the user's triggering message.
+2. Determine the appropriate language for your response
-3.  **Assess Tool Need & Select Tool:** Decide if using a tool is necessary.
+3. Assess if using tools is necessary
-    * **For Memory/Recall:** If asked about past events, known facts, or info likely in memory, use a **Memory Management tool** (`search_nodes`, `open_nodes`).
+4. Formulate your response in the required JSON format
-    * **For Detailed/External Info:** If asked a detailed question needing current/external info, use the **Web Search tool** (`web_search`).
+5. Always maintain the {config.PERSONA_NAME} persona
-    * **If Unsure or No Tool Needed:** Respond directly.
+6. CRITICAL: After using tools, ALWAYS provide a substantive dialogue response - NEVER return an empty dialogue field
-4.  **Tool Arguments (If Needed):** Determine exact arguments. The system handles the call.
+
-5.  **Formulate Response:** Generate a response *directly addressing* the user's *current* message, using tool results if applicable.
+**EXAMPLES OF GOOD TOOL USAGE:**
-    * **Specifically for Web Search:** When you receive the web search result (likely as text snippets), **summarize the key findings** relevant to the user's query in your response. Do not just list the raw results.
+
-6.  **Response Constraints (MANDATORY):**
+Poor response (after web_search): "根據我的搜索，中庄有以下餐廳：1. 老虎蒸餃..."
-    * **Language:** Respond **ONLY** in the **same language** as the user's triggering message.
+
-    * **Conciseness:** Keep responses **brief and conversational** (1-2 sentences usually). **NO** long paragraphs.
+Good response (after web_search): "中庄確實有些值得注意的用餐選擇。老虎蒸餃是其中一家，若你想了解更多細節，我可以提供進一步情報。"
-    * **Dialogue ONLY:** Your output **MUST ONLY** be the character's spoken words. **ABSOLUTELY NO** descriptive actions, expressions, inner thoughts, stage directions, narration, parenthetical notes (like '(...)'), or any other text that isn't pure dialogue.
+
-    * **No Extra Formatting:** **DO NOT** wrap your final dialogue response in quotation marks (like `"`dialogue`"`) or other markdown. Just provide the raw spoken text.
+Poor response (after web_search): "I found 5 restaurants in Zhongzhuang from my search..."
-7.  **Persona Consistency:** Always maintain the {config.PERSONA_NAME} persona.
+
+Good response (after web_search): "Zhongzhuang has several dining options that my intelligence network has identified. Would you like me to share the specifics?"
 """
     return system_prompt
 
 # --- Tool Formatting ---
+def parse_structured_response(response_content: str) -> dict:
+    """
+    更加強大的LLM回應解析函數，能夠處理多種格式。
+    
+    Args:
+        response_content: LLM生成的回應文本
+    
+    Returns:
+        包含dialogue, commands和thoughts的字典
+    """
+    default_result = {
+        "dialogue": "",
+        "commands": [],
+        "thoughts": "",
+        "valid_response": False  # 添加標誌表示解析是否成功
+    }
+    
+    # 如果輸入為空，直接返回默認結果
+    if not response_content or response_content.strip() == "":
+        print("Warning: Empty response content, nothing to parse.")
+        return default_result
+    
+    # 清理模型特殊標記
+    cleaned_content = re.sub(r'<\|.*?\|>', '', response_content)
+    
+    # 首先嘗試解析完整JSON
+    try:
+        # 尋找JSON塊（可能被包裹在```json和```之間）
+        json_match = re.search(r'```json\s*(.*?)\s*```', cleaned_content, re.DOTALL)
+        if json_match:
+            json_str = json_match.group(1)
+            parsed_json = json.loads(json_str)
+            if isinstance(parsed_json, dict) and "dialogue" in parsed_json:
+                print("Successfully parsed complete JSON from code block.")
+                result = {
+                    "dialogue": parsed_json.get("dialogue", ""),
+                    "commands": parsed_json.get("commands", []),
+                    "thoughts": parsed_json.get("thoughts", ""),
+                    "valid_response": bool(parsed_json.get("dialogue", "").strip())
+                }
+                return result
+        
+        # 嘗試直接解析整個內容為JSON
+        parsed_json = json.loads(cleaned_content)
+        if isinstance(parsed_json, dict) and "dialogue" in parsed_json:
+            print("Successfully parsed complete JSON directly.")
+            result = {
+                "dialogue": parsed_json.get("dialogue", ""),
+                "commands": parsed_json.get("commands", []),
+                "thoughts": parsed_json.get("thoughts", ""),
+                "valid_response": bool(parsed_json.get("dialogue", "").strip())
+            }
+            return result
+    except (json.JSONDecodeError, ValueError):
+        # JSON解析失敗，繼續嘗試其他方法
+        pass
+    
+    # 使用正則表達式提取各個字段
+    # 1. 提取dialogue
+    dialogue_match = re.search(r'"dialogue"\s*:\s*"([^"]*("[^"]*"[^"]*)*)"', cleaned_content)
+    if dialogue_match:
+        default_result["dialogue"] = dialogue_match.group(1)
+        print(f"Extracted dialogue field: {default_result['dialogue'][:50]}...")
+        default_result["valid_response"] = bool(default_result['dialogue'].strip())
+    
+    # 2. 提取commands
+    try:
+        commands_match = re.search(r'"commands"\s*:\s*(\[.*?\])', cleaned_content, re.DOTALL)
+        if commands_match:
+            commands_str = commands_match.group(1)
+            # 嘗試修復可能的JSON錯誤
+            fixed_commands_str = commands_str.replace("'", '"').replace('\n', ' ')
+            commands = json.loads(fixed_commands_str)
+            if isinstance(commands, list):
+                default_result["commands"] = commands
+                print(f"Extracted {len(commands)} commands.")
+    except Exception as e:
+        print(f"Failed to parse commands: {e}")
+    
+    # 3. 提取thoughts
+    thoughts_match = re.search(r'"thoughts"\s*:\s*"([^"]*("[^"]*"[^"]*)*)"', cleaned_content)
+    if thoughts_match:
+        default_result["thoughts"] = thoughts_match.group(1)
+        print(f"Extracted thoughts field: {default_result['thoughts'][:50]}...")
+    
+    # 如果dialogue仍然為空，嘗試其他方法
+    if not default_result["dialogue"]:
+        # 嘗試舊方法
+        try:
+            # 處理缺少開頭大括號的情況
+            json_content = cleaned_content.strip()
+            if not json_content.startswith('{'):
+                json_content = '{' + json_content
+            # 處理不完整的結尾
+            if not json_content.endswith('}'):
+                json_content = json_content + '}'
+                
+            parsed_data = json.loads(json_content)
+            
+            # 獲取對話內容
+            if "dialogue" in parsed_data:
+                default_result["dialogue"] = parsed_data["dialogue"]
+                default_result["commands"] = parsed_data.get("commands", [])
+                default_result["thoughts"] = parsed_data.get("thoughts", "")
+                default_result["valid_response"] = bool(default_result["dialogue"].strip())
+                print(f"Successfully parsed JSON with fixes: {json_content[:50]}...")
+                return default_result
+        except:
+            pass
+            
+        # 檢查是否有直接文本回應（沒有JSON格式）
+        # 排除明顯的JSON語法和代碼塊
+        content_without_code = re.sub(r'```.*?```', '', cleaned_content, flags=re.DOTALL)
+        content_without_json = re.sub(r'[\{\}\[\]":\,]', ' ', content_without_code)
+        
+        # 如果有實質性文本，將其作為dialogue
+        stripped_content = content_without_json.strip()
+        if stripped_content and len(stripped_content) > 5:  # 至少5個字符
+            default_result["dialogue"] = stripped_content[:500]  # 限制長度
+            default_result["valid_response"] = True
+            print(f"Using plain text as dialogue: {default_result['dialogue'][:50]}...")
+        else:
+            # 最後嘗試：如果以上方法都失敗，嘗試提取第一個引號包裹的內容作為對話
+            first_quote = re.search(r'"([^"]+)"', cleaned_content)
+            if first_quote:
+                default_result["dialogue"] = first_quote.group(1)
+                default_result["valid_response"] = True
+                print(f"Extracted first quoted string as dialogue: '{default_result['dialogue']}")
+    
+    # 如果沒有提取到有效對話內容
+    if not default_result["dialogue"]:
+        print("All extraction methods failed, no dialogue content found.")
+        # 注意：不設置默認對話內容，保持為空字符串
+    
+    return default_result
+
+
 def _format_mcp_tools_for_openai(mcp_tools: list) -> list:
     """
     Converts the list of tool definition dictionaries obtained from MCP servers
@@ -91,87 +358,250 @@ def _format_mcp_tools_for_openai(mcp_tools: list) -> list:
     print(f"Successfully formatted {len(openai_tools)} tools for API use."); return openai_tools
 
 
+# --- Synthetic Response Generator ---
+def _create_synthetic_response_from_tools(tool_results, original_query):
+    """創建基於工具調用結果的合成回應，保持Wolfhart的角色特性。"""
+    
+    # 提取用戶查詢的關鍵詞
+    query_keywords = set()
+    query_lower = original_query.lower()
+    
+    # 基本關鍵詞提取
+    if "中庄" in query_lower and ("午餐" in query_lower or "餐廳" in query_lower or "吃" in query_lower):
+        query_type = "餐廳查詢"
+        query_keywords = {"中庄", "餐廳", "午餐", "美食"}
+        
+    # 其他查詢類型...
+    else:
+        query_type = "一般查詢"
+    
+    # 開始從工具結果提取關鍵信息
+    extracted_info = {}
+    restaurant_names = []
+    
+    # 處理web_search結果
+    web_search_results = [r for r in tool_results if r.get('name') == 'web_search']
+    if web_search_results:
+        try:
+            for result in web_search_results:
+                content_str = result.get('content', '')
+                if not content_str:
+                    continue
+                
+                # 解析JSON內容
+                content = json.loads(content_str) if isinstance(content_str, str) else content_str
+                search_results = content.get('results', [])
+                
+                # 提取相關信息
+                for search_result in search_results:
+                    title = search_result.get('title', '')
+                    if '中庄' in title and ('餐' in title or '食' in title or '午' in title or '吃' in title):
+                        # 提取餐廳名稱
+                        if '老虎蒸餃' in title:
+                            restaurant_names.append('老虎蒸餃')
+                        elif '割烹' in title and '中庄' in title:
+                            restaurant_names.append('割烹中庄')
+                        # 更多餐廳名稱提取選擇...
+        except Exception as e:
+            print(f"Error extracting info from web_search: {e}")
+    
+    # 生成符合Wolfhart性格的回應
+    restaurant_count = len(restaurant_names)
+    
+    if query_type == "餐廳查詢" and restaurant_count > 0:
+        if restaurant_count == 1:
+            dialogue = f"中庄的{restaurant_names[0]}值得一提。需要更詳細的情報嗎？"
+        else:
+            dialogue = f"根據我的情報網絡，中庄有{restaurant_count}家值得注意的餐廳。需要我透露更多細節嗎？"
+    else:
+        # 通用回應
+        dialogue = "我的情報網絡已收集了相關信息。請指明你需要了解的具體細節。"
+    
+    # 構建結構化回應
+    synthetic_response = {
+        "dialogue": dialogue,
+        "commands": [],
+        "thoughts": "基於工具調用結果合成的回應，保持Wolfhart的角色特性"
+    }
+    
+    return json.dumps(synthetic_response)
+
 # --- Main Interaction Function ---
 async def get_llm_response(
     user_input: str,
     mcp_sessions: dict[str, ClientSession],
     available_mcp_tools: list[dict],
     persona_details: str | None
-) -> str:
+) -> dict:
     """
     Gets a response from the LLM, handling the tool-calling loop and using persona info.
-    Includes post-processing to remove surrounding quotes from final response.
+    Returns a dictionary with 'dialogue', 'commands', and 'thoughts' fields.
     """
+    request_id = int(time.time() * 1000)  # 用時間戳生成請求ID
+    debug_log(f"LLM Request #{request_id} - User Input", user_input)
+    
+    system_prompt = get_system_prompt(persona_details)
+    debug_log(f"LLM Request #{request_id} - System Prompt", system_prompt)
+    
     if not client:
-         return "Error: LLM client not successfully initialized, unable to process request."
+         error_msg = "Error: LLM client not successfully initialized, unable to process request."
+         debug_log(f"LLM Request #{request_id} - Error", error_msg)
+         return {"dialogue": error_msg, "valid_response": False}
 
     openai_formatted_tools = _format_mcp_tools_for_openai(available_mcp_tools)
     messages = [
-        {"role": "system", "content": get_system_prompt(persona_details)},
+        {"role": "system", "content": system_prompt},
         {"role": "user", "content": user_input},
     ]
     
+    debug_log(f"LLM Request #{request_id} - Formatted Tools", 
+              f"Number of tools: {len(openai_formatted_tools)}")
+
     max_tool_calls_per_turn = 5
     current_tool_call_cycle = 0
     
+    # 新增：用於追蹤工具調用
+    all_tool_results = []  # 保存所有工具調用結果
+    last_non_empty_response = None  # 保存最後一個非空回應
+    has_valid_response = False  # 記錄是否獲得有效回應
+
     while current_tool_call_cycle < max_tool_calls_per_turn:
         current_tool_call_cycle += 1
         print(f"\n--- Starting LLM API call (Cycle {current_tool_call_cycle}/{max_tool_calls_per_turn}) ---")
 
         try:
+            debug_log(f"LLM Request #{request_id} - API Call (Cycle {current_tool_call_cycle})", 
+                      f"Model: {config.LLM_MODEL}\nMessages: {json.dumps(messages, ensure_ascii=False, indent=2)}")
+            
+            cycle_start_time = time.time()
             response = await client.chat.completions.create(
                 model=config.LLM_MODEL,
                 messages=messages,
                 tools=openai_formatted_tools if openai_formatted_tools else None,
                 tool_choice="auto" if openai_formatted_tools else None,
             )
+            cycle_duration = time.time() - cycle_start_time
             
             response_message = response.choices[0].message
             tool_calls = response_message.tool_calls
+            content = response_message.content or ""
             
+            # 保存非空回應
+            if content and content.strip():
+                last_non_empty_response = content
+            
+            # 記錄收到的回應
+            response_dump = response_message.model_dump(exclude_unset=True)
+            debug_log(f"LLM Request #{request_id} - API Response (Cycle {current_tool_call_cycle})", 
+                      f"Duration: {cycle_duration:.2f}s\nResponse: {json.dumps(response_dump, ensure_ascii=False, indent=2)}")
+
+            # 添加回應到消息歷史
             messages.append(response_message.model_dump(exclude_unset=True))
 
+            # 如果沒有工具調用請求，處理最終回應
             if not tool_calls:
                 print("--- LLM did not request tool calls, returning final response ---")
-                final_content = response_message.content or "[LLM did not provide text response]"
                 
-                # Post-processing: Remove surrounding quotes
+                # 如果當前回應為空但之前有非空回應，使用之前的最後一個非空回應
-                print(f"Original response content: '{final_content}'")
+                final_content = content
-                if isinstance(final_content, str):
+                if (not final_content or final_content.strip() == "") and last_non_empty_response:
-                    content_stripped = final_content.strip()
+                    print(f"Current response is empty, using last non-empty response from cycle {current_tool_call_cycle-1}")
-                    if content_stripped.startswith('"') and content_stripped.endswith('"') and len(content_stripped) > 1:
+                    final_content = last_non_empty_response
-                        final_content = content_stripped[1:-1]; print("Removed surrounding double quotes.")
-                    elif content_stripped.startswith("'") and content_stripped.endswith("'") and len(content_stripped) > 1:
-                        final_content = content_stripped[1:-1]; print("Removed surrounding single quotes.")
-                    else: final_content = content_stripped
-                print(f"Processed response content: '{final_content}'")
-                return final_content
                 
-            # Tool call handling
+                # 如果仍然為空但有工具調用結果，創建合成回應
-            print(f"--- LLM requested {len(tool_calls)} tool calls ---"); tool_tasks = []
+                if (not final_content or final_content.strip() == "") and all_tool_results:
-            for tool_call in tool_calls: tool_tasks.append(asyncio.create_task(_execute_single_tool_call(tool_call, mcp_sessions, available_mcp_tools), name=f"tool_{tool_call.function.name}"))
+                    print("Creating synthetic response from tool results...")
-            results_list = await asyncio.gather(*tool_tasks, return_exceptions=True); processed_results_count = 0
+                    final_content = _create_synthetic_response_from_tools(all_tool_results, user_input)
-            for result in results_list:
+                
-                 if isinstance(result, Exception): print(f"Error executing tool: {result}")
+                # 解析結構化回應
-                 elif isinstance(result, dict) and 'tool_call_id' in result: messages.append(result); processed_results_count += 1
+                parsed_response = parse_structured_response(final_content)
-                 else: print(f"Warning: Tool returned unexpected result type: {type(result)}")
+                # 標記這是否是有效回應
-            if processed_results_count == 0 and tool_calls: print("Warning: All tool calls failed or had no valid results.")
+                has_dialogue = parsed_response.get("dialogue") and parsed_response["dialogue"].strip()
+                parsed_response["valid_response"] = bool(has_dialogue)
+                has_valid_response = has_dialogue
+                
+                debug_log(f"LLM Request #{request_id} - Final Parsed Response", 
+                          json.dumps(parsed_response, ensure_ascii=False, indent=2))
+                print(f"Final dialogue content: '{parsed_response.get('dialogue', '')}'")                
+                return parsed_response
+
+            # 工具調用處理
+            print(f"--- LLM requested {len(tool_calls)} tool calls ---")
+            debug_log(f"LLM Request #{request_id} - Tool Calls Requested", 
+                      f"Number of tools: {len(tool_calls)}\nTool calls: {json.dumps([t.model_dump() for t in tool_calls], ensure_ascii=False, indent=2)}")
+            
+            tool_tasks = []
+            for tool_call in tool_calls: 
+                tool_tasks.append(asyncio.create_task(
+                    _execute_single_tool_call(tool_call, mcp_sessions, available_mcp_tools, request_id), 
+                    name=f"tool_{tool_call.function.name}"
+                ))
+            
+            results_list = await asyncio.gather(*tool_tasks, return_exceptions=True)
+            processed_results_count = 0
+            
+            debug_log(f"LLM Request #{request_id} - Tool Results", 
+                      f"Number of results: {len(results_list)}")
+                      
+            for i, result in enumerate(results_list):
+                if isinstance(result, Exception): 
+                    print(f"Error executing tool: {result}")
+                    debug_log(f"LLM Request #{request_id} - Tool Error {i+1}", str(result))
+                elif isinstance(result, dict) and 'tool_call_id' in result:
+                    # 保存工具調用結果以便後續使用
+                    all_tool_results.append(result)
+                    messages.append(result)
+                    processed_results_count += 1
+                    debug_log(f"LLM Request #{request_id} - Tool Result {i+1}", 
+                             json.dumps(result, ensure_ascii=False, indent=2))
+                else: 
+                    print(f"Warning: Tool returned unexpected result type: {type(result)}")
+                    debug_log(f"LLM Request #{request_id} - Unexpected Tool Result {i+1}", str(result))
+            
+            if processed_results_count == 0 and tool_calls:
+                print("Warning: All tool calls failed or had no valid results.")
+                # 如果所有工具調用都失敗，中斷循環
+                break
 
         except OpenAIError as e:
             error_msg = f"Error interacting with LLM API ({config.OPENAI_API_BASE_URL or 'Official OpenAI'}): {e}"
-            print(error_msg); return f"Sorry, I encountered an error connecting to the language model."
+            print(error_msg)
+            debug_log(f"LLM Request #{request_id} - OpenAI API Error", error_msg)
+            return {"dialogue": "Sorry, I encountered an error connecting to the language model.", "valid_response": False}
         except Exception as e:
             error_msg = f"Unexpected error processing LLM response or tool calls: {e}"
-            print(error_msg); import traceback; traceback.print_exc(); return f"Sorry, an internal error occurred, please try again later."
+            print(error_msg); import traceback; traceback.print_exc()
+            debug_log(f"LLM Request #{request_id} - Unexpected Error", f"{error_msg}\n{traceback.format_exc()}")
+            return {"dialogue": "Sorry, an internal error occurred, please try again later.", "valid_response": False}
 
-    # Max loop handling
+    # 達到最大循環限制處理
-    print(f"Warning: Maximum tool call cycle limit reached ({max_tool_calls_per_turn})."); last_assistant_content = next((msg.get("content") for msg in reversed(messages) if msg["role"] == "assistant" and msg.get("content")), None)
+    if current_tool_call_cycle >= max_tool_calls_per_turn:
-    if last_assistant_content: return last_assistant_content + "\n(Processing may be incomplete due to tool call limit being reached)"
+        print(f"Warning: Maximum tool call cycle limit reached ({max_tool_calls_per_turn}).")
-    else: return "Sorry, the processing was complex and reached the limit, unable to generate a response."
+        debug_log(f"LLM Request #{request_id} - Max Tool Call Cycles Reached", f"Reached limit of {max_tool_calls_per_turn} cycles")
+    
+    # 回應處理：如果有非空回應，使用它；否則使用合成回應
+    if last_non_empty_response:
+        parsed_response = parse_structured_response(last_non_empty_response)
+        has_valid_response = bool(parsed_response.get("dialogue"))
+    elif all_tool_results:
+        # 從工具結果創建合成回應
+        synthetic_content = _create_synthetic_response_from_tools(all_tool_results, user_input)
+        parsed_response = parse_structured_response(synthetic_content)
+        has_valid_response = bool(parsed_response.get("dialogue"))
+    else:
+        # 沒有有效的回應
+        parsed_response = {"dialogue": "", "commands": [], "thoughts": ""}
+        has_valid_response = False
+    
+    # 添加有效回應標誌
+    parsed_response["valid_response"] = has_valid_response
+    
+    debug_log(f"LLM Request #{request_id} - Final Response (After Cycles)", json.dumps(parsed_response, ensure_ascii=False, indent=2))
+    return parsed_response
 
 
 # --- Helper function _execute_single_tool_call ---
-async def _execute_single_tool_call(tool_call, mcp_sessions, available_mcp_tools) -> dict:
+async def _execute_single_tool_call(tool_call, mcp_sessions, available_mcp_tools, request_id=None) -> dict:
     """
     Helper function to execute one tool call and return the formatted result message.
     Includes argument type correction for web_search.
@@ -186,6 +616,10 @@ async def _execute_single_tool_call(tool_call, mcp_sessions, available_mcp_tools
     print(f"Executing tool: {function_name}")
     print(f"Raw arguments generated by LLM (string): {function_args_str}")
     
+    if request_id:
+        debug_log(f"LLM Request #{request_id} - Tool Call Execution", 
+                  f"Tool: {function_name}\nID: {tool_call_id}\nArgs: {function_args_str}")
+
     try:
         function_args = json.loads(function_args_str)
         print(f"Parsed arguments (dictionary): {function_args}")
@@ -213,7 +647,16 @@ async def _execute_single_tool_call(tool_call, mcp_sessions, available_mcp_tools
 
         if target_session:
             result_content = await mcp_client.call_mcp_tool(session=target_session, tool_name=function_name, arguments=function_args) # Use corrected args
-            if isinstance(result_content, dict) and 'error' in result_content: print(f"Tool '{function_name}' call returned error: {result_content['error']}")
+            if isinstance(result_content, dict) and 'error' in result_content: 
+                print(f"Tool '{function_name}' call returned error: {result_content['error']}")
+                if request_id:
+                    debug_log(f"LLM Request #{request_id} - Tool Call Error", 
+                              f"Tool: {function_name}\nError: {result_content['error']}")
+            elif request_id:
+                debug_log(f"LLM Request #{request_id} - Tool Call Success", 
+                          f"Tool: {function_name}\nResult: {json.dumps(result_content, ensure_ascii=False, indent=2)[:500]}..." 
+                          if isinstance(result_content, (dict, list)) and len(json.dumps(result_content)) > 500 
+                          else f"Tool: {function_name}\nResult: {result_content}")
 
     # Format result content for LLM
     try:
@@ -241,5 +684,11 @@ async def _execute_single_tool_call(tool_call, mcp_sessions, available_mcp_tools
     except Exception as format_err: print(f"Error formatting tool '{function_name}' result: {format_err}"); result_content_str = json.dumps({"error": f"Failed to format tool result: {format_err}"})
 
     # Return the formatted message for the LLM
-    return {"tool_call_id": tool_call_id, "role": "tool", "name": function_name, "content": result_content_str}
+    response = {"tool_call_id": tool_call_id, "role": "tool", "name": function_name, "content": result_content_str}
+    
+    if request_id:
+        debug_log(f"LLM Request #{request_id} - Tool Response Formatted", 
+                  f"Tool: {function_name}\nFormatted Response: {json.dumps(response, ensure_ascii=False, indent=2)}")
+    
+    return response
 

main.py

@@ -236,26 +236,46 @@ async def run_main_with_exit_stack():
 
             print(f"\n{config.PERSONA_NAME} is thinking...")
             try:
-                # Get LLM response
+                # Get LLM response (現在返回的是一個字典)
-                bot_response = await llm_interaction.get_llm_response(
+                bot_response_data = await llm_interaction.get_llm_response(
                     user_input=f"Message from {sender_name}: {bubble_text}", # Provide context
                     mcp_sessions=active_mcp_sessions,
                     available_mcp_tools=all_discovered_mcp_tools,
                     persona_details=wolfhart_persona_details
                 )
-                print(f"{config.PERSONA_NAME}'s response: {bot_response}")
                 
-                # Send response back via UI interaction module
+                # 提取對話內容
-                if bot_response:
+                bot_dialogue = bot_response_data.get("dialogue", "")
-                    print("Preparing to send response via UI...")
+                valid_response = bot_response_data.get("valid_response", False)
+                print(f"{config.PERSONA_NAME}'s dialogue response: {bot_dialogue}")
+                
+                # 處理命令 (如果有的話)
+                commands = bot_response_data.get("commands", [])
+                if commands:
+                    print(f"Processing {len(commands)} command(s)...")
+                    for cmd in commands:
+                        cmd_type = cmd.get("type", "")
+                        cmd_params = cmd.get("parameters", {})
+                        # 預留位置：在這裡添加命令處理邏輯
+                        print(f"Command type: {cmd_type}, parameters: {cmd_params}")
+                        # TODO: 實現各類命令的處理邏輯
+                
+                # 記錄思考過程 (如果有的話)
+                thoughts = bot_response_data.get("thoughts", "")
+                if thoughts:
+                    print(f"AI Thoughts: {thoughts[:150]}..." if len(thoughts) > 150 else f"AI Thoughts: {thoughts}")
+                
+                # 只有當有效回應時才發送到遊戲
+                if bot_dialogue and valid_response:
+                    print("Preparing to send dialogue response via UI...")
                     send_success = await asyncio.to_thread(
                         ui_interaction.paste_and_send_reply,
-                        bot_response
+                        bot_dialogue
                     )
                     if send_success: print("Response sent successfully.")
                     else: print("Error: Failed to send response via UI.")
                 else:
-                    print("LLM did not generate a response, not sending.")
+                    print("Not sending response: Invalid or empty dialogue content.")
 
             except Exception as e:
                 print(f"\nError processing trigger or sending response: {e}")