优化测试脚本stage分别输出

backend_service/src/llm/gateway.py

@@ -15,14 +15,17 @@ class LLMGateway:
         self.generator = generator
         self.stage1_enable_thinking = os.getenv("STAGE1_ENABLE_THINKING", "true").lower() in ("1", "true", "yes")
 
-    def classify_scene(self, user_prompt: str) -> str:
+    def classify_scene(self, user_prompt: str, drone_state: str = "on_ground") -> str:
         scene_mode = "scene1"
         try:
+            user_content = user_prompt
+            if drone_state and drone_state in ("on_ground", "in_air"):
+                user_content = f"无人机当前状态：{drone_state}\n\n用户指令：{user_prompt}"
             classifier_resp = self.generator.classifier_client.chat.completions.create(
                 model=self.generator.classifier_model,
                 messages=[
                     {"role": "system", "content": self.generator.scene_classifier_prompt or "你是一个分类器，只输出JSON。"},
-                    {"role": "user", "content": user_prompt},
+                    {"role": "user", "content": user_content},
                 ],
                 temperature=0.0,
                 response_format={"type": "json_object"},

backend_service/src/pipeline/orchestrator.py

@@ -9,8 +9,8 @@ class GenerationOrchestrator:
     def __init__(self, generator: Any):
         self.stages = PipelineStages(generator)
 
-    async def generate(self, user_prompt: str) -> Dict:
-        understanding = self.stages.stage1_task_understanding(user_prompt)
+    async def generate(self, user_prompt: str, drone_state: str = "on_ground") -> Dict:
+        understanding = self.stages.stage1_task_understanding(user_prompt, drone_state=drone_state)
         context = self.stages.stage2_context_binding(user_prompt, understanding)
         
         if understanding.scene_mode == "simple":

backend_service/src/pipeline/stages.py

@@ -71,18 +71,15 @@ class PipelineStages:
     def __init__(self, generator: Any):
         self.generator = generator
 
-    def stage1_task_understanding(self, user_prompt: str) -> TaskUnderstanding:
+    def stage1_task_understanding(self, user_prompt: str, drone_state: str = "on_ground") -> TaskUnderstanding:
         logging.info("========== [Stage 1] Task Understanding ==========")
-        scene_mode = self.generator.llm_gateway.classify_scene(user_prompt)
+        if drone_state not in ("on_ground", "in_air"):
+            drone_state = "on_ground"
+        scene_mode = self.generator.llm_gateway.classify_scene(user_prompt, drone_state=drone_state)
         intent_type = _infer_intent_type(user_prompt, scene_mode)
         risk_flags = _extract_risk_flags(user_prompt)
         requires_relative = "relative_reference_detected" in risk_flags
-        
-        # 简单提取状态
-        drone_state: DroneState = "on_ground"
-        if "在空中" in user_prompt or "已起飞" in user_prompt:
-            drone_state = "in_air"
-            
+
         entities = {"raw_prompt": user_prompt}
         
         logging.info(f"Task Understanding Results: mode={scene_mode}, state={drone_state}, intent={intent_type}, risks={risk_flags}")

backend_service/src/prompts/scene_classifier_prompt.txt

@@ -1,9 +1,12 @@
-你是指令分类器。只输出一个JSON，无其它内容。
-输入：无人机状态{on_ground/in_air}+指令。
+你是指令分类器，仅输出JSON对象，无任何多余内容。
+输入：无人机状态(on_ground/in_air)+指令。
 输出仅三选一：{"mode":"simple"}、{"mode":"scene1"}、{"mode":"scene4"}。
 
-规则：
-1. 指令含“面前”→scene1
-2. 状态=in_air，指令是：飞到某地/飞到某地+方位距离/往某方向飞X米/降落/旋转/悬停 →simple
-3. 状态=on_ground，指令含去/飞到/回到某地 →非simple
-4. 多动作/序列任务→scene4
+规则（必须严格执行）：
+1. 指令包含“面前”→输出{"mode":"scene1"}；
+2. 无人机状态=in_air且指令包含“飞到”→输出{"mode":"simple"}；
+3. 无人机状态=in_air且指令包含“往某方向飞”或“降落”或“旋转”或“悬停”→输出{"mode":"simple"}；
+4. 无人机状态=on_ground且指令包含“起飞”或“旋转”或“悬停”→输出{"mode":"simple"}；
+5. 无人机状态=on_ground且指令包含“飞到”或“去”或“回到”→输出{"mode":"scene4"}；
+6. 指令包含“先”或“再”或“搜索”或“监控”或“拍照”或“返航”或“确认”→输出{"mode":"scene4"}；
+7. 其他所有情况→输出{"mode":"scene4"}。

backend_service/src/py_tree_generator.py

@@ -790,9 +790,89 @@ class PyTreeGenerator:
         except Exception as e:
             logging.warning(f"保存推理链Markdown失败: {e}")
 
-    async def generate(self, user_prompt: str) -> Dict[str, Any]:
-        logging.info(f"接收到用户请求: {user_prompt}")
-        return await self.orchestrator.generate(user_prompt)
+    async def generate(self, user_prompt: str, drone_state: str = "on_ground") -> Dict[str, Any]:
+        logging.info(f"接收到用户请求: {user_prompt}, drone_state={drone_state}")
+        return await self.orchestrator.generate(user_prompt, drone_state=drone_state)
+
+    def run_debug_stage(
+        self, user_prompt: str, drone_state: str = "on_ground", target_stage: int = 1
+    ) -> Dict[str, Any]:
+        """
+        分阶段调试：运行到指定 stage 并返回该 stage 的输出。
+        当 target_stage >= 2 时，响应中附带 upstream（前一 stage 的 output）。
+        当 target_stage >= 3 时，响应中附带 stage1 与 stage2 的 output。
+        以此类推，便于追溯完整流水线。
+        """
+        if target_stage < 1 or target_stage > 6:
+            return {"error": f"target_stage 必须在 1-6 之间，当前为 {target_stage}"}
+
+        upstream: Dict[str, Any] = {}
+
+        understanding = self.orchestrator.stages.stage1_task_understanding(
+            user_prompt, drone_state=drone_state
+        )
+        if target_stage == 1:
+            return {
+                "target_stage": 1,
+                "stage_name": "TaskUnderstanding",
+                "output": understanding.model_dump(),
+                "upstream": upstream,
+            }
+
+        upstream["stage1"] = understanding.model_dump()
+        context = self.orchestrator.stages.stage2_context_binding(user_prompt, understanding)
+        if target_stage == 2:
+            return {
+                "target_stage": 2,
+                "stage_name": "ContextBinding",
+                "output": context.model_dump(),
+                "upstream": upstream,
+            }
+
+        upstream["stage2"] = context.model_dump()
+        draft = self.orchestrator.stages.stage3_macro_planning(
+            user_prompt, understanding, context
+        )
+        if target_stage == 3:
+            return {
+                "target_stage": 3,
+                "stage_name": "BTDraft",
+                "output": draft.model_dump(),
+                "upstream": upstream,
+            }
+
+        upstream["stage3"] = draft.model_dump()
+        resolved_data = self.orchestrator.stages.stage4_middleware_resolution(draft)
+        if target_stage == 4:
+            return {
+                "target_stage": 4,
+                "stage_name": "MiddlewareResolution",
+                "output": {"resolved_data": resolved_data},
+                "upstream": upstream,
+            }
+
+        upstream["stage4"] = {"resolved_data": resolved_data}
+        final_tree = self.orchestrator.stages.stage5_micro_filling(
+            draft, resolved_data, understanding
+        )
+        if target_stage == 5:
+            return {
+                "target_stage": 5,
+                "stage_name": "MicroFilling",
+                "output": {"final_tree": final_tree},
+                "upstream": upstream,
+            }
+
+        upstream["stage5"] = {"final_tree": final_tree}
+        payload = self.orchestrator.stages.stage6_validate_and_postprocess(
+            user_prompt, understanding, context, draft, final_tree
+        )
+        return {
+            "target_stage": 6,
+            "stage_name": "ValidateAndPostprocess",
+            "output": payload,
+            "upstream": upstream,
+        }
 
 # Create a single instance for the application
 py_tree_generator = PyTreeGenerator()