diff --git a/agent/deep_assistant.py b/agent/deep_assistant.py
index 0d9dec4..a68f7aa 100644
--- a/agent/deep_assistant.py
+++ b/agent/deep_assistant.py
@@ -2,13 +2,18 @@ import json
 import logging
 import time
 import copy
+from pathlib import Path
 from typing import Any, Dict
 from langchain.chat_models import init_chat_model
-# from deepagents import create_deep_agent
+from deepagents import create_deep_agent
+from deepagents.backends import CompositeBackend
+from deepagents.backends.filesystem import FilesystemBackend
+from deepagents.backends.sandbox import SandboxBackendProtocol
 from deepagents_cli.agent import create_cli_agent
 from langchain.agents import create_agent
 from langchain.agents.middleware import SummarizationMiddleware
 from langchain_mcp_adapters.client import MultiServerMCPClient
+from sympy.printing.cxx import none
 from utils.fastapi_utils import detect_provider
 from .guideline_middleware import GuidelineMiddleware
 from .tool_output_length_middleware import ToolOutputLengthMiddleware
@@ -19,6 +24,14 @@ from agent.prompt_loader import load_system_prompt_async, load_mcp_settings_asyn
 from agent.agent_memory_cache import get_memory_cache_manager
 from .checkpoint_utils import prepare_checkpoint_message
 from langgraph.checkpoint.postgres.aio import AsyncPostgresSaver
+from langgraph.checkpoint.memory import InMemorySaver
+from langchain.tools import BaseTool
+from langchain_core.language_models import BaseChatModel
+from langgraph.pregel import Pregel
+from deepagents_cli.shell import ShellMiddleware
+from deepagents_cli.agent_memory import AgentMemoryMiddleware
+from deepagents_cli.skills import SkillsMiddleware
+from deepagents_cli.config import settings, get_default_coding_instructions
 import os
 
 # 全局 MemorySaver 实例
@@ -148,13 +161,15 @@ async def init_agent(config: AgentConfig):
     checkpointer = None
     create_start = time.time()
     if config.robot_type == "deep_agent":
-        # 使用 DeepAgentX 创建 agent
-        agent, composite_backend = create_cli_agent(
+        # 使用 DeepAgentX 创建 agent，自定义 workspace_root
+        workspace_root = str(Path.home() / ".deepagents" / config.bot_id)
+        agent, composite_backend = create_custom_cli_agent(
             model=llm_instance,
             assistant_id=config.bot_id,
+            system_prompt=system_prompt,
             tools=mcp_tools,
             auto_approve=True,
-            enable_shell=False,
+            workspace_root=workspace_root
         )
     else:
         # 构建中间件列表
@@ -198,4 +213,164 @@ async def init_agent(config: AgentConfig):
             checkpointer=checkpointer
         )
     logger.info(f"create {config.robot_type} elapsed: {time.time() - create_start:.3f}s")
-    return agent, checkpointer
\ No newline at end of file
+    return agent, checkpointer
+
+
+def create_custom_cli_agent(
+    model: str | BaseChatModel,
+    assistant_id: str,
+    *,
+    tools: list[BaseTool] | None = None,
+    sandbox: SandboxBackendProtocol | None = None,
+    sandbox_type: str | None = None,
+    system_prompt: str | None = None,
+    auto_approve: bool = False,
+    enable_memory: bool = True,
+    enable_skills: bool = True,
+    enable_shell: bool = True,
+    workspace_root: str | None = None,
+) -> tuple[Pregel, CompositeBackend]:
+    """Create a CLI-configured agent with custom workspace_root for shell commands.
+
+    This is a custom version of create_cli_agent that allows specifying a custom
+    workspace_root for shell commands instead of using Path.cwd().
+
+    Args:
+        model: LLM model to use (e.g., "anthropic:claude-sonnet-4-5-20250929")
+        assistant_id: Agent identifier for memory/state storage
+        tools: Additional tools to provide to agent (default: empty list)
+        sandbox: Optional sandbox backend for remote execution (e.g., ModalBackend).
+                 If None, uses local filesystem + shell.
+        sandbox_type: Type of sandbox provider ("modal", "runloop", "daytona").
+                     Used for system prompt generation.
+        system_prompt: Override the default system prompt. If None, generates one
+                      based on sandbox_type and assistant_id.
+        auto_approve: If True, automatically approves all tool calls without human
+                     confirmation. Useful for automated workflows.
+        enable_memory: Enable AgentMemoryMiddleware for persistent memory
+        enable_skills: Enable SkillsMiddleware for custom agent skills
+        enable_shell: Enable ShellMiddleware for local shell execution (only in local mode)
+        workspace_root: Working directory for shell commands. If None, uses Path.cwd().
+
+    Returns:
+        2-tuple of (agent_graph, composite_backend)
+        - agent_graph: Configured LangGraph Pregel instance ready for execution
+        - composite_backend: CompositeBackend for file operations
+    """
+    if tools is None:
+        tools = []
+
+    # Setup agent directory for persistent memory (if enabled)
+    if enable_memory or enable_skills:
+        agent_dir = settings.ensure_agent_dir(assistant_id)
+        agent_md = agent_dir / "agent.md"
+        if not agent_md.exists():
+            source_content = get_default_coding_instructions()
+            agent_md.write_text(source_content)
+
+    # Skills directories (if enabled)
+    skills_dir = none
+    if enable_skills:
+        skills_dir = settings.ensure_user_skills_dir(assistant_id)
+
+    # Build middleware stack based on enabled features
+    agent_middleware = []
+
+    # CONDITIONAL SETUP: Local vs Remote Sandbox
+    if sandbox is None:
+        # ========== LOCAL MODE ==========
+        composite_backend = CompositeBackend(
+            default=FilesystemBackend(root_dir=workspace_root),  # Current working directory
+            routes={},  # No virtualization - use real paths
+        )
+
+        # Add memory middleware
+        if enable_memory:
+            agent_middleware.append(
+                AgentMemoryMiddleware(settings=settings, assistant_id=assistant_id)
+            )
+
+        # Add skills middleware
+        if enable_skills:
+            agent_middleware.append(
+                SkillsMiddleware(
+                    skills_dir=skills_dir,
+                    assistant_id=assistant_id
+                )
+            )
+
+        # Add shell middleware (only in local mode)
+        if enable_shell:
+            # Create environment for shell commands
+            # Restore user's original LANGSMITH_PROJECT so their code traces separately
+            shell_env = os.environ.copy()
+            if settings.user_langchain_project:
+                shell_env["LANGSMITH_PROJECT"] = settings.user_langchain_project
+
+            # Use custom workspace_root if provided, otherwise use current directory
+            shell_workspace = workspace_root if workspace_root is not None else str(Path.cwd())
+
+            agent_middleware.append(
+                ShellMiddleware(
+                    workspace_root=shell_workspace,
+                    env=shell_env,
+                )
+            )
+    else:
+        # ========== REMOTE SANDBOX MODE ==========
+        composite_backend = CompositeBackend(
+            default=sandbox,  # Remote sandbox (ModalBackend, etc.)
+            routes={},  # No virtualization
+        )
+
+        # Add memory middleware
+        if enable_memory:
+            agent_middleware.append(
+                AgentMemoryMiddleware(settings=settings, assistant_id=assistant_id)
+            )
+
+        # Add skills middleware
+        if enable_skills:
+            agent_middleware.append(
+                SkillsMiddleware(
+                    skills_dir=skills_dir,
+                    assistant_id=assistant_id,
+                    project_skills_dir=project_skills_dir,
+                )
+            )
+
+        # Note: Shell middleware not used in sandbox mode
+        # File operations and execute tool are provided by the sandbox backend
+
+    # Get or use custom system prompt
+    if system_prompt is None:
+        # Import get_system_prompt from deepagents_cli.agent
+        from deepagents_cli.agent import get_system_prompt as _get_system_prompt
+        system_prompt = _get_system_prompt(assistant_id=assistant_id, sandbox_type=sandbox_type)
+
+    # Import InterruptOnConfig
+    from langchain.agents.middleware import InterruptOnConfig
+
+    # Configure interrupt_on based on auto_approve setting
+    if auto_approve:
+        # No interrupts - all tools run automatically
+        interrupt_on = {}
+    else:
+        # Full HITL for destructive operations - import from deepagents_cli.agent
+        from deepagents_cli.agent import _add_interrupt_on
+        interrupt_on = _add_interrupt_on()
+
+    # Import config
+    from deepagents_cli.config import config
+
+    # Create the agent
+    agent = create_deep_agent(
+        model=model,
+        system_prompt=system_prompt,
+        tools=tools,
+        backend=composite_backend,
+        middleware=agent_middleware,
+        interrupt_on=interrupt_on,
+        checkpointer=InMemorySaver(),
+    ).with_config(config)
+    return agent, composite_backend
\ No newline at end of file
diff --git a/prompt/system_prompt_deep_agent.md b/prompt/system_prompt_deep_agent.md
index dcade57..08fd4b7 100644
--- a/prompt/system_prompt_deep_agent.md
+++ b/prompt/system_prompt_deep_agent.md
@@ -1,5 +1,4 @@
 <env>
-<env>
 Working directory: {agent_dir_path}
 Current User: {user_identifier}
 Current Time: {datetime}