基类在 libs/core/langchain_core/tools/base.py:427。先看它"是什么"和"带什么"：

# libs/core/langchain_core/tools/base.py:427（字段裁剪汇总）
class BaseTool(RunnableSerializable[str | dict[str, Any] | ToolCall, Any]):
    """Interface LangChain tools must implement."""

    name: str                                   # :468 品名：模型用它点名调用（唯一）
    description: str                            # :471 用途说明：模型靠它决定"何时用这件工具"
    args_schema: ArgsSchema | None = Field(     # :477 插口规格：参数的 Pydantic 模型
        default=None, description="The tool schema.")
    return_direct: bool = False                 # :489 结果直接返回用户,不再交回模型(Day15 会用到)
    response_format: Literal["content", "content_and_artifact"] = "content"  # :541

    @abstractmethod
    def _run(self, *args: Any, **kwargs: Any) -> Any:   # :877 ★子类唯一必须实现的：真干活
        ...

手写 BaseTool 子类太啰嗦，99% 的场景用 @tool。实现在 libs/core/langchain_core/tools/convert.py:77（前面 :18/:32/:48/:63 是四个 @overload 类型重载，真身是第五个）：

# libs/core/langchain_core/tools/convert.py:77
def tool(
    name_or_callable=None,          # 既支持 @tool 也支持 @tool("自定义名")
    runnable=None, *args,
    description=None,               # 不给就用函数 docstring
    return_direct=False,
    args_schema=None,               # 不给就从函数签名推断
    infer_schema=True,
    response_format="content",
    parse_docstring=False, ...
) -> BaseTool | Callable[..., BaseTool]:

它的核心分拣逻辑（convert.py:297-319，裁剪）——判断你给的是同步函数、异步函数还是 Runnable，最后统一送进 StructuredTool.from_function：

# libs/core/langchain_core/tools/convert.py:297（裁剪）
elif inspect.iscoroutinefunction(dec_func):
    coroutine = dec_func            # async def → 装进 coroutine 槽
    func = None
    schema = args_schema
else:
    coroutine = None
    func = dec_func                 # 普通 def → 装进 func 槽
    schema = args_schema

if infer_schema or args_schema is not None:
    return StructuredTool.from_function(     # ★绝大多数工具的最终归宿
        func, coroutine,
        name=tool_name,                      # 默认 = 函数名
        description=tool_description,        # 默认 = docstring
        args_schema=schema,
        infer_schema=infer_schema, ...)

StructuredTool（libs/core/langchain_core/tools/structured.py:40）是 @tool 产物的真身——"函数 + 说明书"的合体：

# libs/core/langchain_core/tools/structured.py:40
class StructuredTool(BaseTool):
    """Tool that can operate on any number of inputs."""

    description: str = ""
    args_schema: Annotated[ArgsSchema, SkipValidation()] = Field(
        ..., description="The tool schema.")     # 这里变成必填（基类里可为 None）
    func: Callable[..., Any] | None = None       # 同步执行体
    coroutine: Callable[..., Awaitable[Any]] | None = None   # 异步执行体

    def _run(self, *args, config, run_manager=None, **kwargs):   # :74
        if self.func:
            ...
            return self.func(*args, **kwargs)    # ★_run 就是转调你的原函数

造它的工厂 from_function（structured.py:133）里，最关键的是"推断 schema"这几行（structured.py:203）：

# libs/core/langchain_core/tools/structured.py:202
name = name or source_function.__name__          # 品名 = 函数名
if args_schema is None and infer_schema:
    args_schema = create_schema_from_function(   # ★读函数签名，现场生成 Pydantic 模型
        name, source_function,
        parse_docstring=parse_docstring,         # True 时还会解析 Google 风格 docstring
        error_on_invalid_docstring=error_on_invalid_docstring,
        filter_args=_filter_schema_args(source_function))
...
if description is None and not parse_docstring:
    description_ = source_function.__doc__ or None    # 说明书 = docstring

执行入口 BaseTool.run（libs/core/langchain_core/tools/base.py:977）——又一次模板方法，骨架如下：

# libs/core/langchain_core/tools/base.py:977（裁剪骨架）
def run(self, tool_input, ..., tool_call_id=None, **kwargs) -> Any:
    callback_manager = CallbackManager.configure(...)        # ① 装回调
    run_manager = callback_manager.on_tool_start(            # ② 记"工具开始"（含入参）
        {"name": self.name, "description": self.description},
        tool_input_str, ..., tool_call_id=tool_call_id)
    try:
        tool_args, tool_kwargs = self._to_args_and_kwargs(   # ③ ★解析+schema 校验入参
            tool_input, tool_call_id)
        response = context.run(self._run, *tool_args, **tool_kwargs)  # ④ ★真干活
        if self.response_format == "content_and_artifact":   # ⑤ 双通道返回的拆包
            content, artifact = response
        ...
    # （异常处理 / ToolException 兜底 / on_tool_end 上报，略）