fix: lazy init TavilySearch to avoid startup crash without API key

- Add get_tavily_search() for lazy initialization (renamed from _get_tavily_search)
- Update deep_agent/graph.py to use lazy initialization
- Add assertion check when DeepAgent loads to ensure search tool is available
- Convert research_sub_agent to function to receive tools dynamically

When TAVILY_API_KEY is not configured, the API server can now start normally.
DeepAgent will show a clear error message when used without the API key.

src/agents/common/tools.py

@@ -4,15 +4,25 @@ from typing import Annotated, Any
 
 from langchain.tools import tool
 from langchain_core.tools import StructuredTool
-from langchain_tavily import TavilySearch
 from langgraph.types import interrupt
 from pydantic import BaseModel, Field
 
 from src import config, graph_base, knowledge_base
 from src.utils import logger
 
-search = TavilySearch()
+# Lazy initialization for TavilySearch (only when TAVILY_API_KEY is available)
-search.metadata = {"name": "Tavily 网页搜索"}
+_tavily_search_instance = None
+
+
+def get_tavily_search():
+    """Get TavilySearch instance lazily, only when API key is available."""
+    global _tavily_search_instance
+    if _tavily_search_instance is None and config.enable_web_search:
+        from langchain_tavily import TavilySearch
+
+        _tavily_search_instance = TavilySearch()
+        _tavily_search_instance.metadata = {"name": "Tavily 网页搜索"}
+    return _tavily_search_instance
 
 
 @tool(name_or_callable="计算器", description="可以对给定的2个数字选择进行 add, subtract, multiply, divide 运算")
@@ -101,7 +111,9 @@ def get_static_tools() -> list:
 
     # 检查是否启用网页搜索
     if config.enable_web_search:
-        static_tools.append(search)
+        tavily_search = get_tavily_search()
+        if tavily_search:
+            static_tools.append(tavily_search)
 
     return static_tools
 

src/agents/deep_agent/graph.py

@@ -8,25 +8,26 @@ from langchain.agents.middleware import ModelRequest, SummarizationMiddleware, T
 
 from src.agents.common import BaseAgent, load_chat_model
 from src.agents.common.middlewares import inject_attachment_context
-from src.agents.common.tools import search
+from src.agents.common.tools import get_tavily_search
 
 from .context import DeepContext
 from .prompts import DEEP_PROMPT
 
-search_tools = [search]
 
+def _get_research_sub_agent(search_tools: list) -> dict:
+    """Get research sub-agent config with search tools."""
+    return {
+        "name": "research-agent",
+        "description": ("利用搜索工具，用于研究更深入的问题。将调研结果写入到主题研究文件中。"),
+        "system_prompt": (
+            "你是一位专注的研究员。你的工作是根据用户的问题进行研究。"
+            "进行彻底的研究，然后用详细的答案回复用户的问题，只有你的最终答案会被传递给用户。"
+            "除了你的最终信息，他们不会知道任何其他事情，所以你的最终报告应该就是你的最终信息！"
+            "将调研结果保存到主题研究文件中 /sub_research/xxx.md 中。"
+        ),
+        "tools": search_tools,
+    }
 
-research_sub_agent = {
-    "name": "research-agent",
-    "description": ("利用搜索工具，用于研究更深入的问题。将调研结果写入到主题研究文件中。"),
-    "system_prompt": (
-        "你是一位专注的研究员。你的工作是根据用户的问题进行研究。"
-        "进行彻底的研究，然后用详细的答案回复用户的问题，只有你的最终答案会被传递给用户。"
-        "除了你的最终信息，他们不会知道任何其他事情，所以你的最终报告应该就是你的最终信息！"
-        "将调研结果保存到主题研究文件中 /sub_research/xxx.md 中。"
-    ),
-    "tools": search_tools,
-}
 
 critique_sub_agent = {
     "name": "critique-agent",
@@ -73,7 +74,16 @@ class DeepAgent(BaseAgent):
 
     async def get_tools(self):
         """返回 Deep Agent 的专用工具"""
-        tools = search_tools
+        tools = []
+        tavily_search = get_tavily_search()
+        if tavily_search:
+            tools.append(tavily_search)
+
+        # Assert that search tool is available for DeepAgent
+        assert tools, (
+            "DeepAgent requires at least one search tool. "
+            "Please configure TAVILY_API_KEY environment variable to enable web search."
+        )
         return tools
 
     async def get_graph(self, **kwargs):
@@ -88,6 +98,9 @@ class DeepAgent(BaseAgent):
         sub_model = load_chat_model(context.subagents_model)
         tools = await self.get_tools()
 
+        # Build subagents with search tools
+        research_sub_agent = _get_research_sub_agent(tools)
+
         # 使用 create_deep_agent 创建深度智能体
         graph = create_agent(
             model=model,