fix: 彻底修复 RAG 无限循环问题

关键修复：
1. rag_nodes.py: history 里记录的 action 改成大写 RETRIEVE_RAG
2. react_nodes.py: route_by_reasoning 优先检查 DIRECT_RESPONSE
3. 优化路由判断顺序，避免不必要的循环

backend/app/main_graph/nodes/rag_nodes.py

@@ -234,7 +234,7 @@ async def rag_retrieve_node(state: MainGraphState, config: Optional[Dict[str, An
             # 关键修复：把 rag_retrieve 加到 reasoning_history 里，让下次推理知道
             state.reasoning_history.append({
                 "step": state.reasoning_step,
-                "action": "rag_retrieve",
+                "action": "RETRIEVE_RAG",  # 大写，和推理结果保持一致
                 "confidence": 1.0,
                 "reasoning": "RAG 检索完成",
                 "timestamp": datetime.now().isoformat()

backend/app/main_graph/nodes/react_nodes.py

@@ -359,30 +359,36 @@ def route_by_reasoning(state: MainGraphState) -> str:
             return "rag_retrieve"
         return "react_reason"
     
-    # 关键修复：检查是否已经执行过子图，如果是，直接去 llm_call
+    # ========== 关键修复：优先检查当前推理结果 ==========
+    reasoning_result: Optional[ReasoningResult] = state.debug_info.get("reasoning_result")
+    if reasoning_result and reasoning_result.action == ReasoningAction.DIRECT_RESPONSE:
+        info(f"[route_by_reasoning] 当前推理结果=DIRECT_RESPONSE，直接去 llm_call")
+        return "llm_call"
+    
+    # ========== 然后检查历史和状态 ==========
     previous_actions = [h.get("action") for h in state.reasoning_history]
+    
+    # 检查是否已经执行过子图
     if "subgraph_completed" in previous_actions or state.final_result:
         return "llm_call"
     
-    # 关键修复：检测 RAG 重复循环 - 如果发现"RETRIEVE_RAG"出现超过1次，直接去 LLM
+    # 检测 RAG 重复循环 - 如果发现"RETRIEVE_RAG"出现超过1次，直接去 LLM
     rag_count = previous_actions.count("RETRIEVE_RAG")
     if rag_count >= 2:
         info(f"[route_by_reasoning] 检测到 RAG 重复循环（{rag_count}次），直接去 llm_call")
         return "llm_call"
     
-    # 关键修复：如果已经有 rag_docs 或 rag_context，说明已经检索过了，直接去 LLM
+    # 如果已经有 rag_docs 或 rag_context，说明已经检索过了，直接去 LLM
     if (state.rag_docs and len(state.rag_docs) > 0) or (state.rag_context and len(state.rag_context) > 0):
         info(f"[route_by_reasoning] 检测到已存在 RAG 检索结果，直接去 llm_call")
         return "llm_call"
     
-    # 关键修复：限制最多 3 次推理，避免无限循环
+    # 限制最多 3 次推理，避免无限循环
     if len(previous_actions) >= 3:
         info(f"[route_by_reasoning] 已达到最大推理次数 ({len(previous_actions)})，直接去 llm_call")
         return "llm_call"
     
-    # 获取推理结果
+    # ========== 最后处理其他推理结果 ==========
-    reasoning_result: Optional[ReasoningResult] = state.debug_info.get("reasoning_result")
-    
     if not reasoning_result:
         return "llm_call"