.gitignoreにcontexts/を追加

LMStudioClientとTaskHandlerでの空レスポンスチェックと関数呼び出し処理を改善

Author: notfolder

.gitignore

@@ -1,6 +1,7 @@
 node_modules/
 gitlab/
 .vscode/
+contexts/
 .env
 *.pyc
 *.log

clients/lmstudio_client.py

@@ -69,20 +69,21 @@ def send_user_message(self, message: str) -> None:
             self.chat.add_user_message(message)
 
     def send_function_result(self, name: str, result: object) -> None:
-        """関数の実行結果を送信する(LM Studioでは未対応).
+        """関数の実行結果を送信する.
 
         Args:
             name: 関数名
             result: 実行結果
 
-        Raises:
-            NotImplementedError: LM Studioは関数呼び出しをサポートしていない
-
         """
-        msg = "LMStudio does not support function calls. Use OpenAI compatible call instead."
-        raise NotImplementedError(
-            msg,
-        )
+        if self.message_store:
+            # File-based mode: add result as user message
+            output_message = f"output: {result}"
+            self.message_store.add_message("user", output_message)
+        else:
+            # Legacy mode: LM Studio does not support function calls
+            msg = "LMStudio does not support function calls. Use OpenAI compatible call instead."
+            raise NotImplementedError(msg)
 
     def get_response(self) -> str:
         """LLMからの応答を取得する.
@@ -91,10 +92,16 @@ def get_response(self) -> str:
             LLMからの応答テキスト
 
         """
+        import logging
+        logger = logging.getLogger(__name__)
+        
         if self.message_store:
             # File-based mode: need to manually create chat from messages
             chat = lms.Chat()
             messages = self._load_messages_from_file()
+            
+            logger.debug("メッセージ数: %d", len(messages))
+            
             for msg in messages:
                 role = msg.get("role")
                 content = msg.get("content", "")
@@ -106,6 +113,8 @@ def get_response(self) -> str:
                     chat.add_assistant_response(content)
 
             result = self.model.respond(chat)
+            logger.debug("LMStudio応答型: %s, 内容: %s", type(result), str(result)[:100])
+            
             self.message_store.add_message("assistant", str(result))
         else:
             # Legacy mode

handlers/task_handler.py

@@ -576,15 +576,24 @@ def _process_llm_interaction_with_client(
         # LLMからレスポンスを取得
         resp, functions = llm_client.get_response()
         self.logger.info("LLM応答: %s", resp)
+        
+        # 空レスポンスのチェック
+        if not resp or not resp.strip():
+            self.logger.error("LLMから空の応答が返されました")
+            if count >= MAX_JSON_PARSE_ERRORS:
+                task.comment("LLM応答エラーでスキップ")
+                return True
+            return False
 
         # レスポンスの前処理
         resp_clean = self._process_think_tags(task, resp)
+        self.logger.debug("think処理後のレスポンス: %s", resp_clean)
 
         # JSON応答の解析
         try:
             data = self._extract_json(resp_clean)
         except Exception:
-            self.logger.exception("LLM応答JSONパース失敗")
+            self.logger.exception("LLM応答JSONパース失敗 (応答内容: %s)", resp_clean[:200])
             if count >= MAX_JSON_PARSE_ERRORS:
                 task.comment("LLM応答エラーでスキップ")
                 return True
@@ -627,6 +636,67 @@ def _process_response_data_with_context(
             task.comment(str(data["plan"]))
             llm_client.send_user_message(str(data["plan"]))
 
+        # function_call形式の処理 (OpenAI/LMStudio互換)
+        if "function_call" in data:
+            func_call = data["function_call"]
+            tool_name = func_call["name"]
+            args_str = func_call.get("arguments", "{}")
+            
+            # argumentsが文字列の場合はパース
+            if isinstance(args_str, str):
+                args = json.loads(args_str) if args_str else {}
+            else:
+                args = args_str
+            
+            args = self.sanitize_arguments(args)
+            mcp_server, tool_func = tool_name.split("_", 1)
+            
+            start_time = time.time()
+            try:
+                output = self.mcp_clients[mcp_server].call_tool(tool_func, args)
+                duration_ms = (time.time() - start_time) * 1000
+                
+                # Record tool execution
+                tool_store.add_tool_call(
+                    tool_name=tool_name,
+                    args=args,
+                    result=output,
+                    status="success",
+                    duration_ms=duration_ms,
+                )
+                context_manager.update_statistics(tool_calls=1)
+                
+                # Send result to LLM
+                llm_client.send_function_result(tool_name, output)
+                
+                # Reset error count on success
+                if error_state.get("last_tool") == tool_name:
+                    error_state["tool_error_count"] = 0
+                
+            except Exception as e:
+                duration_ms = (time.time() - start_time) * 1000
+                error_msg = str(e)
+                
+                # Record tool error
+                tool_store.add_tool_call(
+                    tool_name=tool_name,
+                    args=args,
+                    result=None,
+                    status="error",
+                    duration_ms=duration_ms,
+                    error=error_msg,
+                )
+                context_manager.update_statistics(tool_calls=1)
+                
+                self.logger.exception("ツール呼び出しエラー: %s", error_msg)
+                llm_client.send_function_result(tool_name, {"error": error_msg})
+                
+                # Update error state
+                self._update_error_count(tool_name, error_state)
+                if error_state.get("tool_error_count", 0) >= MAX_CONSECUTIVE_TOOL_ERRORS:
+                    task.comment(f"同じツール({tool_name})で3回連続エラーが発生したため処理を中止します。")
+                    return True
+
         if "call_tool" in data:
             # ツール呼び出し処理
             for tool_call in data["call_tool"]: