diff --git a/apps/server/api/detection.py b/apps/server/api/detection.py
index 77cba43..cb4fbb5 100644
--- a/apps/server/api/detection.py
+++ b/apps/server/api/detection.py
@@ -14,6 +14,36 @@ router = APIRouter()
 logger = logging.getLogger(__name__)
 
 
+# ---------------------------------------------------------------------------
+# DetectionService 单例 (MVP-3 / D33)
+#
+# 之前每次请求都 new 一个 DetectionService，会导致 rule_engine / aggregator 状态
+# 不共享，规则配置 API 修改后无法同步生效。改造为按 model_service 缓存的单例。
+# ---------------------------------------------------------------------------
+
+
+_detection_service = None
+
+
+def get_detection_service(model_service=None):
+    """获取共享的 DetectionService 单例。
+
+    若尚未初始化且未传入 ``model_service``，将抛出 RuntimeError。
+    """
+
+    global _detection_service
+    if _detection_service is None:
+        if model_service is None:
+            raise RuntimeError(
+                "DetectionService 尚未初始化，请先在 main.lifespan 中调用 "
+                "get_detection_service(model_service)"
+            )
+        from services.detection_service import DetectionService
+        _detection_service = DetectionService(model_service)
+        logger.info("DetectionService 单例已初始化")
+    return _detection_service
+
+
 @router.post("/detect/image", response_model=ImageDetectionResult)
 async def detect_image(
     file: UploadFile = File(...),
@@ -42,7 +72,7 @@ async def detect_image(
     from main import model_service
     from services.detection_service import DetectionService
     
-    detection_service = DetectionService(model_service)
+    detection_service = get_detection_service(model_service)
     
     # 解析算法配置
     algo_config = None
@@ -309,7 +339,10 @@ async def detect_video(
                     'frame_index': frame_index,
                     'timestamp': round(frame_index / fps, 2) if fps > 0 else 0,
                     'detections': result_data['detections'],
-                    'detection_count': len(result_data['detections'])
+                    'detection_count': len(result_data['detections']),
+                    # MVP-3: 携带 LLM 管道结果供前端展示
+                    'llm_results': result_data.get('llm_results', []),
+                    'alert_events': result_data.get('alert_events', []),
                 })
 
                 # 提取关键帧截图（最多保留 20 张，防止响应过大）