diff --git a/agent/deep_assistant.py b/agent/deep_assistant.py
index cc8e27a..99158b5 100644
--- a/agent/deep_assistant.py
+++ b/agent/deep_assistant.py
@@ -45,10 +45,11 @@ from langchain_core.language_models import BaseChatModel
 from langgraph.pregel import Pregel
 # 新版本导入：MemoryMiddleware 和 SkillsMiddleware 已迁移到 deepagents.middleware
 from deepagents.middleware import MemoryMiddleware, SkillsMiddleware
-from langchain.agents.middleware import AgentMiddleware
+from langchain.agents.middleware import AgentMiddleware, ModelRequest, ModelResponse
+from langchain.agents.middleware import HumanInTheLoopMiddleware, InterruptOnConfig, TodoListMiddleware
+from langchain_core.messages import AIMessage, HumanMessage
 from langgraph.types import Checkpointer
 from deepagents_cli.config import settings, get_default_coding_instructions
-from langchain.agents.middleware import HumanInTheLoopMiddleware, InterruptOnConfig, TodoListMiddleware
 from deepagents.middleware.filesystem import FilesystemMiddleware
 from deepagents.middleware.patch_tool_calls import PatchToolCallsMiddleware
 from langchain_anthropic.middleware import AnthropicPromptCachingMiddleware
@@ -63,6 +64,51 @@ from .custom_filesystem_middleware import CustomFilesystemMiddleware
 # _global_checkpointer = MemorySaver()
 
 logger = logging.getLogger('app')
+
+
+class EmptyResponseRetryMiddleware(AgentMiddleware):
+    """当模型返回空内容（无文字、无工具调用）时自动重试"""
+
+    MAX_RETRIES = 5
+
+    def _is_empty_response(self, result: ModelResponse) -> bool:
+        """判断是否为空响应"""
+        if not result.result:
+            return True
+        msg = result.result[0]
+        if not isinstance(msg, AIMessage):
+            return False
+        content = msg.content or ""
+        has_text = bool(content.strip()) if isinstance(content, str) else bool(content)
+        return not has_text and len(msg.tool_calls) == 0
+
+    def wrap_model_call(self, request, handler):
+        result = handler(request)
+        retries = 0
+        while self._is_empty_response(result) and retries < self.MAX_RETRIES:
+            retries += 1
+            logger.warning(f"Empty response detected, retrying ({retries}/{self.MAX_RETRIES})")
+            retry_messages = list(request.messages) + [
+                HumanMessage(content="Please continue your response.")
+            ]
+            request = request.override(messages=retry_messages)
+            result = handler(request)
+        return result
+
+    async def awrap_model_call(self, request, handler):
+        result = await handler(request)
+        retries = 0
+        while self._is_empty_response(result) and retries < self.MAX_RETRIES:
+            retries += 1
+            logger.warning(f"Empty response detected, retrying ({retries}/{self.MAX_RETRIES})")
+            retry_messages = list(request.messages) + [
+                HumanMessage(content="Please continue your response.")
+            ]
+            request = request.override(messages=retry_messages)
+            result = await handler(request)
+        return result
+
+
 # Utility functions
 def read_system_prompt():
     """读取通用的无状态系统prompt"""
@@ -220,6 +266,8 @@ async def init_agent(config: AgentConfig):
 
     # 构建中间件列表
     middleware = []
+    # 添加空响应重试中间件（最先执行，最外层包裹）
+    middleware.append(EmptyResponseRetryMiddleware())
     # 首先添加 ToolUseCleanupMiddleware 来清理孤立的 tool_use
     middleware.append(ToolUseCleanupMiddleware())
     # 添加工具输出长度控制中间件
@@ -407,22 +455,22 @@ def create_custom_cli_agent(
             # Use LocalShellBackend for filesystem + shell execution
             backend = LocalShellBackend(
                 root_dir=workspace_root,
-                virtual_mode=True,
+                virtual_mode=False,
                 inherit_env=True,
                 env=final_shell_env,
             )
         else:
             # No shell access - use plain FilesystemBackend
-            backend = FilesystemBackend(root_dir=workspace_root, virtual_mode=True)
+            backend = FilesystemBackend(root_dir=workspace_root, virtual_mode=False)
 
         # Set up composite backend with routing (参考新版本实现)
         large_results_backend = FilesystemBackend(
             root_dir=tempfile.mkdtemp(prefix="deepagents_large_results_"),
-            virtual_mode=True,
+            virtual_mode=False,
         )
         conversation_history_backend = FilesystemBackend(
             root_dir=tempfile.mkdtemp(prefix="deepagents_conversation_history_"),
-            virtual_mode=True,
+            virtual_mode=False,
         )
         composite_backend = CompositeBackend(
             default=backend,
@@ -438,7 +486,7 @@ def create_custom_cli_agent(
 
             agent_middleware.append(
                 CustomSkillsMiddleware(
-                    backend=FilesystemBackend(root_dir=workspace_root, virtual_mode=True),
+                    backend=FilesystemBackend(root_dir=workspace_root, virtual_mode=False),
                     sources=skills_sources,
                 )
             )
diff --git a/agent/logging_handler.py b/agent/logging_handler.py
index de0d3db..26aa6b1 100644
--- a/agent/logging_handler.py
+++ b/agent/logging_handler.py
@@ -41,16 +41,20 @@ class LoggingCallbackHandler(BaseCallbackHandler):
         if hasattr(response, 'generations') and response.generations:
             for gen_idx, generation_list in enumerate(response.generations):
                 for msg_idx, generation in enumerate(generation_list):
-                    if hasattr(generation, 'text'):
-                        output_list = generation.text.split("\n")
-                        for i, output in enumerate(output_list):
-                            if output.strip():
-                                self.logger.info(f"{output}")
-                    elif hasattr(generation, 'message'):
-                        output_list = generation.message.split("\n")
-                        for i, output in enumerate(output_list):
-                            if output.strip():
-                                self.logger.info(f"{output}")
+                    # ChatGeneration: 使用 text 属性获取内容
+                    if hasattr(generation, 'text') and generation.text:
+                        for line in generation.text.split("\n"):
+                            if line.strip():
+                                self.logger.info(f"  {line}")
+                    # 如果有 message 属性，输出额外信息（tool_calls 等）
+                    if hasattr(generation, 'message') and generation.message:
+                        msg = generation.message
+                        content = msg.content if hasattr(msg, 'content') else ''
+                        if not content or (isinstance(content, str) and not content.strip()):
+                            self.logger.info(f"  [EMPTY content]")
+                        tool_calls = msg.tool_calls if hasattr(msg, 'tool_calls') else []
+                        if tool_calls:
+                            self.logger.info(f"  [tool_calls: {[tc.get('name', '') for tc in tool_calls]}")
 
     def on_llm_error(
         self, error: Exception, **kwargs: Any
diff --git a/plans/deepagents-upgrade.md b/plans/deepagents-upgrade.md
index af56e9f..fc9a060 100644
--- a/plans/deepagents-upgrade.md
+++ b/plans/deepagents-upgrade.md
@@ -60,7 +60,7 @@ from deepagents.backends import LocalShellBackend
 # 创建 backend，支持自定义环境变量
 backend = LocalShellBackend(
     root_dir=workspace_root,
-    virtual_mode=True,
+    virtual_mode=False,
     env={"ASSISTANT_ID": "xxx", "USER_IDENTIFIER": "yyy"},  # 自定义环境变量
     inherit_env=True,  # 继承父进程环境变量
 )
@@ -92,7 +92,7 @@ backend = LocalShellBackend(
 ```python
 # 当前实现
 composite_backend = CompositeBackend(
-    default=FilesystemBackend(root_dir=workspace_root, virtual_mode=True),
+    default=FilesystemBackend(root_dir=workspace_root, virtual_mode=False),
     routes={},
 )
 ```
@@ -120,14 +120,14 @@ from deepagents.backends import LocalShellBackend
 # 创建带自定义环境变量的 backend
 shell_backend = LocalShellBackend(
     root_dir=workspace_root,
-    virtual_mode=True,
+    virtual_mode=False,
     env=shell_env,
     inherit_env=True,
 )
 
 # 或使用 CompositeBackend 路由
 composite_backend = CompositeBackend(
-    default=FilesystemBackend(root_dir=workspace_root, virtual_mode=True),
+    default=FilesystemBackend(root_dir=workspace_root, virtual_mode=False),
     routes={
         "/shell/": shell_backend,  # shell 命令路由
     },
@@ -209,7 +209,7 @@ if enable_shell:
     final_shell_env = shell_env or {}
     shell_backend = LocalShellBackend(
         root_dir=workspace_root,
-        virtual_mode=True,
+        virtual_mode=False,
         env=final_shell_env,
         inherit_env=True,  # 继承 os.environ
     )
diff --git a/routes/chat.py b/routes/chat.py
index 40cf8ee..b7169b4 100644
--- a/routes/chat.py
+++ b/routes/chat.py
@@ -512,7 +512,7 @@ async def chat_completions(request: ChatRequest, authorization: Optional[str] =
         project_dir = create_project_directory(request.dataset_ids, bot_id, request.skills)
 
         # 收集额外参数作为 generate_cfg
-        exclude_fields = {'messages', 'model', 'model_server', 'dataset_ids', 'language', 'tool_response', 'system_prompt', 'mcp_settings' ,'stream', 'robot_type', 'bot_id', 'user_identifier', 'session_id', 'enable_thinking', 'skills', 'enable_memory', 'n', 'shell_env'}
+        exclude_fields = {'messages', 'model', 'model_server', 'dataset_ids', 'language', 'tool_response', 'system_prompt', 'mcp_settings' ,'stream', 'robot_type', 'bot_id', 'user_identifier', 'session_id', 'enable_thinking', 'skills', 'enable_memory', 'n', 'shell_env', 'max_tokens'}
         generate_cfg = {k: v for k, v in request.model_dump().items() if k not in exclude_fields}
         # 处理消息
         messages = process_messages(request.messages, request.language)
@@ -666,7 +666,7 @@ async def chat_warmup_v2(request: ChatRequestV2, authorization: Optional[str] =
         messages = process_messages(empty_messages, request.language or "ja")
 
         # 收集额外参数作为 generate_cfg
-        exclude_fields = {'messages', 'stream', 'tool_response', 'bot_id', 'language', 'user_identifier', 'session_id', 'n', 'model', 'model_server', 'api_key', 'shell_env'}
+        exclude_fields = {'messages', 'stream', 'tool_response', 'bot_id', 'language', 'user_identifier', 'session_id', 'n', 'model', 'model_server', 'api_key', 'shell_env', 'max_tokens'}
         generate_cfg = {k: v for k, v in request.model_dump().items() if k not in exclude_fields}
         # 从请求中提取 model/model_server/api_key，优先级高于 bot_config（排除 "whatever" 和空值）
         req_data = request.model_dump()
@@ -773,7 +773,7 @@ async def chat_completions_v2(request: ChatRequestV2, authorization: Optional[st
         # 处理消息
         messages = process_messages(request.messages, request.language)
         # 收集额外参数作为 generate_cfg
-        exclude_fields = {'messages', 'dataset_ids', 'language', 'tool_response', 'system_prompt', 'mcp_settings', 'stream', 'robot_type', 'bot_id', 'user_identifier', 'session_id', 'enable_thinking', 'skills', 'enable_memory', 'n', 'model', 'model_server', 'api_key', 'shell_env'}
+        exclude_fields = {'messages', 'dataset_ids', 'language', 'tool_response', 'system_prompt', 'mcp_settings', 'stream', 'robot_type', 'bot_id', 'user_identifier', 'session_id', 'enable_thinking', 'skills', 'enable_memory', 'n', 'model', 'model_server', 'api_key', 'shell_env', 'max_tokens'}
         generate_cfg = {k: v for k, v in request.model_dump().items() if k not in exclude_fields}
         # 从请求中提取 model/model_server/api_key，优先级高于 bot_config（排除 "whatever" 和空值）
         req_data = request.model_dump()