修复循环推理bug

backend/app/main_graph/nodes/routing.py

@@ -0,0 +1,120 @@
+"""
+路由与初始化模块
+包含状态初始化节点和条件路由函数
+
+三层统一循环防护：
+1. 全局步数硬上限（reasoning_step > max_steps）
+2. 路由模式检测（A→B→A→B 交替循环）
+3. 状态停滞检测（连续相同动作）
+"""
+
+from datetime import datetime
+
+from app.core.intent import get_route_by_reasoning, ReasoningAction
+from app.main_graph.state import MainGraphState
+from app.logger import info
+
+
+# ========== 初始化状态节点 ==========
+def init_state_node(state: MainGraphState) -> MainGraphState:
+    """初始化状态节点：在流程开始时设置初始值"""
+    state.current_phase = "initializing"
+    state.reasoning_step = 0
+    state.start_time = datetime.now().isoformat()
+
+    if not state.user_query and state.messages:
+        last_msg = state.messages[-1]
+        state.user_query = getattr(last_msg, "content", str(last_msg))
+
+    return state
+
+
+# ========== 条件路由函数 ==========
+def route_by_reasoning(state: MainGraphState) -> str:
+    """
+    根据推理结果决定下一步路由，带三层统一循环防护
+
+    核心逻辑：
+    1. DIRECT_RESPONSE → 直接返回 llm_call
+    2. 子图完成/已有结果 → 直接返回 llm_call
+    3. 步数超限 → 直接返回 llm_call
+    4. 其他 → 正常路由
+    """
+    # 获取历史动作
+    previous_actions = [h.get("action") for h in state.reasoning_history]
+
+    info(f"[条件路由] step={state.reasoning_step}, phase={state.current_phase}, history={previous_actions}")
+
+    # ========== 获取推理结果 ==========
+    reasoning_result = state.debug_info.get("reasoning_result")
+    latest_action = reasoning_result.action.name if reasoning_result else None
+
+    # ========== 核心检查：DIRECT_RESPONSE 优先 ==========
+    # 从 reasoning_result 检查（最新）
+    if latest_action == "DIRECT_RESPONSE":
+        info(f"[条件路由] 推理结果为 DIRECT_RESPONSE，直接去 llm_call")
+        return "llm_call"
+
+    # 备用：从历史记录检查
+    if previous_actions and previous_actions[-1] == "DIRECT_RESPONSE":
+        info(f"[条件路由] 历史记录最新动作为 DIRECT_RESPONSE，直接去 llm_call")
+        return "llm_call"
+
+    # ========== 子图完成/已有结果 ==========
+    if "subgraph_completed" in previous_actions or state.final_result:
+        info("[条件路由] 子图已完成或已有结果，直接终止")
+        return "llm_call"
+
+    # ========== 步数超限 ==========
+    if state.reasoning_step > state.max_steps:
+        info(f"[条件路由] 步数超限 ({state.reasoning_step}/{state.max_steps})，强制终止")
+        return "llm_call"
+
+    # ========== 特殊阶段快速通道 ==========
+    if state.current_phase in ("max_steps_exceeded", "finalizing", "done"):
+        return "llm_call"
+    if state.current_phase == "error_handling" or state.current_error:
+        return "handle_error"
+
+    # ========== 无推理结果，默认终止 ==========
+    if not reasoning_result:
+        info("[条件路由] 无推理结果，默认去 llm_call")
+        return "llm_call"
+
+    # ========== 计算目标路由 ==========
+    route = get_route_by_reasoning(reasoning_result)
+
+    route_mapping = {
+        "direct_response": "llm_call",
+        "retrieve_rag": "rag_retrieve",
+        "re_retrieve_rag": "rag_retrieve",
+        "web_search": "web_search",
+        "clarify": "llm_call",
+        "call_tool": "llm_call",
+        "contact": "contact_subgraph",
+        "dictionary": "dictionary_subgraph",
+        "news_analysis": "news_analysis_subgraph",
+    }
+    target = route_mapping.get(route, "llm_call")
+
+    # ========== 循环防护检测 ==========
+    # 1. 路由模式检测（A→B→A→B 交替）
+    if len(previous_actions) >= 4:
+        if (previous_actions[-4] == previous_actions[-2]
+                and previous_actions[-3] == previous_actions[-1]
+                and previous_actions[-2] != previous_actions[-1]):
+            info(f"[条件路由] 检测到路由循环: {previous_actions[-4:]}，强制终止")
+            return "llm_call"
+
+    # 2. 状态停滞检测（连续相同动作）
+    if len(previous_actions) >= 2 and previous_actions[-1] == previous_actions[-2]:
+        info(f"[条件路由] 连续相同动作 '{previous_actions[-1]}'，强制终止")
+        return "llm_call"
+
+    # ========== 智能优化 ==========
+    if target == "rag_retrieve" and (state.rag_docs or state.rag_context):
+        info("[条件路由] RAG 结果已存在，跳过检索")
+        return "llm_call"
+
+    info(f"[条件路由] 动作={latest_action}, 目标={target}")
+    return target