打开 crewai/flow/flow.py，开头的模块 docstring 就把整个包的分层说清楚了（flow/flow.py:1）：

# flow/flow.py:1（模块 docstring 摘要）
"""Backwards-compatible re-export surface for the Flow framework.

The implementation now lives in three modules, split by concern:

- ``crewai.flow.dsl``            -- 写法层：@start / @listen / @router、or_ / and_
- ``crewai.flow.flow_definition`` -- 契约层：可序列化的 Flow Definition
- ``crewai.flow.runtime``         -- 引擎层：Flow 执行引擎与状态
- ``crewai.experimental.conversational_mixin`` -- 实验性的对话式扩展
"""

你平时写 class MyFlow(Flow) 用的 Flow，其实是三个来源组合出来的（flow/flow.py:33）：

# flow/flow.py:20
from crewai.experimental.conversational_mixin import _ConversationalMixin
from crewai.flow.dsl import and_, listen, or_, router, start
from crewai.flow.runtime import (
    _INITIAL_STATE_CLASS_MARKER,
    Flow as RuntimeFlow,   # ← 引擎层的 Flow 才是真身
    FlowMeta,
    FlowState,
)

T = TypeVar("T", bound=dict[str, Any] | BaseModel)

# flow/flow.py:33
class Flow(_ConversationalMixin, RuntimeFlow[T]):
    """Public Flow class with experimental conversational extension behavior."""

引擎层的 Flow（runtime/__init__.py:428），注意它继承的是 BaseModel：

# runtime/__init__.py:428
class Flow(BaseModel, Generic[T], metaclass=FlowMeta):
    """Base class for all flows.
    type parameter T must be either dict[str, Any] or a subclass of BaseModel."""

    model_config = ConfigDict(
        arbitrary_types_allowed=True,
        ignored_types=(StartMethod, ListenMethod, RouterMethod),  # 装饰器包装不当字段
        revalidate_instances="never",
    )
    __hash__ = object.__hash__

    _flow_definition: ClassVar[FlowDefinition | None] = None   # 懒构建、按类缓存
    entity_type: Literal["flow"] = "flow"
    _state: Any = PrivateAttr(default=None)                     # ← 运行时状态存这

契约是懒构建的——第一次访问才解析（runtime/__init__.py:469）：

# runtime/__init__.py:469
@classmethod
def flow_definition(cls) -> FlowDefinition:
    """Return the static Flow Definition built from this Flow class."""
    flow_definition = cls.__dict__.get("_flow_definition")
    if flow_definition is None:
        flow_definition = build_flow_definition(cls)   # 扫描类里的装饰器方法
        cls._flow_definition = flow_definition
    return flow_definition

答案是自定义元类 FlowMeta（runtime/__init__.py:375），它在类被创建时改写注解：

# runtime/__init__.py:405
for attr_name, attr_value in list(namespace.items()):
    if attr_name in annotations or attr_name.startswith("_"):
        continue
    if attr_name in parent_fields:
        annotations[attr_name] = Any
        ...
        continue
    if callable(attr_value) or isinstance(
        attr_value, (*_skip_types, FlowMethod)      # 方法/装饰器包装 → 不当字段
    ):
        continue
    annotations[attr_name] = ClassVar[type(attr_value)]  # 裸赋值 → 当类变量
namespace["__annotations__"] = annotations

同步入口 kickoff 其实是异步引擎的薄包装（runtime/__init__.py:1920）：

# runtime/__init__.py:1966
async def _run_flow() -> Any:
    return await self.kickoff_async(inputs, input_files, ...)

runtime_scope = crewai_event_bus._enter_runtime_scope()
try:
    try:
        asyncio.get_running_loop()               # 已在事件循环里？
        ctx = contextvars.copy_context()
        with ThreadPoolExecutor(max_workers=1) as pool:
            return pool.submit(ctx.run, asyncio.run, _run_flow()).result()
    except RuntimeError:                          # 没有运行中的循环
        return asyncio.run(_run_flow())           # 直接开一个跑
finally:
    crewai_event_bus._exit_runtime_scope(runtime_scope)

而 kickoff_async 的核心动作，是执行所有 start 方法、随后让事件链自动扩散（runtime/__init__.py:2442 _execute_start_method）：

# runtime/__init__.py:2473（精简）
result, finished_event_id = await self._execute_method(
    start_method_name, enhanced_method)          # ① 跑一个 @start 方法
...
await self._execute_listeners(                   # ② 广播"它完成了"，触发监听者
    start_method_name, result, finished_event_id)

看 crewai/flow/__init__.py 导出的公开面，就能感受到 Flow 是个独立的编排层（flow/__init__.py:25）：

# flow/__init__.py:25（节选 __all__）
__all__ = [
    "Flow", "start", "listen", "router", "and_", "or_",   # 图的写法
    "FlowStructure", "build_flow_structure", "visualize_flow_structure",  # 可视化
    "persist",                                             # 状态持久化（D45）
    "Expression",                                          # CEL 表达式（D46）
    "ConversationalConfig", "ChatState",                   # 对话式（D47）
    "InputProvider", "HumanFeedbackProvider", ...          # 人在环 / 输入
]