Merge pull request #961 from SuanmoSuanyangTechnology/fix/wxy_031

fix(api): fix API Key rate limiting and terminal user quota checks

api/app/controllers/api_key_controller.py

@@ -167,6 +167,8 @@ def update_api_key(
 
         return success(data=api_key_schema.ApiKey.model_validate(api_key), msg="API Key 更新成功")
 
+    except BusinessException:
+        raise
     except Exception as e:
         logger.error(f"未知错误: {str(e)}", extra={
             "api_key_id": str(api_key_id),

api/app/controllers/public_share_controller.py

@@ -219,9 +219,20 @@ def list_conversations(
     end_user_repo = EndUserRepository(db)
     app_service = AppService(db)
     app = app_service._get_app_or_404(share.app_id)
+    workspace_id = app.workspace_id
+
+    # 仅在新建终端用户时检查配额
+    existing_end_user = end_user_repo.get_end_user_by_other_id(workspace_id=workspace_id, other_id=other_id)
+    if existing_end_user is None:
+        from app.core.quota_manager import _check_quota
+        from app.models.workspace_model import Workspace
+        ws = db.query(Workspace).filter(Workspace.id == workspace_id).first()
+        if ws:
+            _check_quota(db, ws.tenant_id, "end_user_quota", "end_user")
+
     new_end_user = end_user_repo.get_or_create_end_user(
         app_id=share.app_id,
-        workspace_id=app.workspace_id,
+        workspace_id=workspace_id,
         other_id=other_id
     )
     logger.debug(new_end_user.id)
@@ -309,7 +320,6 @@ def get_conversation(
     "/chat",
     summary="发送消息（支持流式和非流式）"
 )
-@check_end_user_quota
 async def chat(
         payload: conversation_schema.ChatRequest,
         share_data: ShareTokenData = Depends(get_share_user_id),
@@ -350,6 +360,18 @@ async def chat(
         app_service = AppService(db)
         app = app_service._get_app_or_404(share.app_id)
         workspace_id = app.workspace_id
+
+        # 仅在新建终端用户时检查配额，已有用户复用不受限制
+        existing_end_user = end_user_repo.get_end_user_by_other_id(workspace_id=workspace_id, other_id=other_id)
+        logger.info(f"终端用户配额检查: workspace_id={workspace_id}, other_id={other_id}, existing={existing_end_user is not None}")
+        if existing_end_user is None:
+            from app.core.quota_manager import _check_quota
+            from app.models.workspace_model import Workspace
+            ws = db.query(Workspace).filter(Workspace.id == workspace_id).first()
+            if ws:
+                logger.info(f"新终端用户，执行配额检查: tenant_id={ws.tenant_id}")
+                _check_quota(db, ws.tenant_id, "end_user_quota", "end_user")
+
         new_end_user = end_user_repo.get_or_create_end_user(
             app_id=share.app_id,
             workspace_id=workspace_id,

api/app/repositories/end_user_repository.py

@@ -66,6 +66,17 @@ class EndUserRepository:
             db_logger.error(f"查询宿主 {end_user_id} 时出错: {str(e)}")
             raise
 
+    def get_end_user_by_other_id(self, workspace_id: uuid.UUID, other_id: str) -> Optional["EndUser"]:
+        """按 workspace_id + other_id 查找终端用户，不存在返回 None"""
+        return (
+            self.db.query(EndUser)
+            .filter(
+                EndUser.workspace_id == workspace_id,
+                EndUser.other_id == other_id
+            )
+            .first()
+        )
+
     def get_or_create_end_user(
         self,
         app_id: uuid.UUID,

api/app/services/api_key_service.py

@@ -51,7 +51,7 @@ class ApiKeyService:
             if existing:
                 raise BusinessException(f"API Key 名称 {data.name} 已存在", BizCode.API_KEY_DUPLICATE_NAME)
 
-            # 若 rate_limit 超过租户套餐的 api_ops_rate_limit，自动截断到套餐上限
+            # 若 rate_limit 超过租户套餐的 api_ops_rate_limit，直接报错
             from app.models.workspace_model import Workspace
             from app.core.quota_manager import get_api_ops_rate_limit
 
@@ -59,7 +59,10 @@ class ApiKeyService:
             if workspace:
                 tenant_api_ops_limit = get_api_ops_rate_limit(db, workspace.tenant_id)
                 if tenant_api_ops_limit and data.rate_limit > tenant_api_ops_limit:
-                    data.rate_limit = tenant_api_ops_limit
+                    raise BusinessException(
+                        f"API Key QPS 不能超过套餐上限 {tenant_api_ops_limit}",
+                        BizCode.BAD_REQUEST
+                    )
 
             # 生成 API Key
             api_key = generate_api_key(data.type)
@@ -162,7 +165,7 @@ class ApiKeyService:
             if existing:
                 raise BusinessException(f"API Key 名称 {data.name} 已存在", BizCode.API_KEY_DUPLICATE_NAME)
 
-        # 若 rate_limit 超过租户套餐的 api_ops_rate_limit，自动截断到套餐上限
+        # 若 rate_limit 超过租户套餐的 api_ops_rate_limit，直接报错
         if data.rate_limit is not None:
             from app.models.workspace_model import Workspace
             from app.core.quota_manager import get_api_ops_rate_limit
@@ -171,7 +174,10 @@ class ApiKeyService:
             if workspace:
                 tenant_api_ops_limit = get_api_ops_rate_limit(db, workspace.tenant_id)
                 if tenant_api_ops_limit and data.rate_limit > tenant_api_ops_limit:
-                    data.rate_limit = tenant_api_ops_limit
+                    raise BusinessException(
+                        f"API Key QPS 不能超过套餐上限 {tenant_api_ops_limit}",
+                        BizCode.BAD_REQUEST
+                    )
 
         update_data = data.model_dump(exclude_unset=True)
         ApiKeyRepository.update(db, api_key_id, update_data)