训练前先把 crew 切到"训练模式"（crew.py:901）：

# crew.py:901
def _setup_for_training(self, filename: str) -> None:
    """Sets up the crew for training."""
    self._train = True
    for task in self.tasks:
        task.human_input = True          # ★每个任务都强制要人工输入（点评）
    for agent in self.agents:
        agent.allow_delegation = False    # ★训练时禁止委派——要单独评估每个 agent
    CrewTrainingHandler(TRAINING_DATA_FILE).initialize_file()   # 建原始反馈文件
    CrewTrainingHandler(filename).initialize_file()             # 建最终建议文件

train 主入口（crew.py:914，节选头部）：

# crew.py:914
def train(self, n_iterations: int, filename: str, inputs: dict | None = None) -> None:
    """Trains the crew for a given number of iterations."""
    inputs = inputs or {}
    try:
        crewai_event_bus.emit(self, CrewTrainStartedEvent(
            crew_name=self.name, n_iterations=n_iterations, filename=filename, inputs=inputs))
        train_crew = self.copy()                 # ★在副本上训练，不污染原 crew
        train_crew._setup_for_training(filename)
        ...

train 的核心循环（crew.py:932）：

# crew.py:932
            for n_iteration in range(n_iterations):        # ★跑 N 遍
                train_crew._train_iteration = n_iteration
                train_crew.kickoff(inputs=inputs)          # 每遍就是一次普通 kickoff

            training_data = CrewTrainingHandler(TRAINING_DATA_FILE).load()   # 读回攒的原始反馈
            ...

训练的产出在这一步生成（crew.py:938）：

# crew.py:938
            for agent in train_crew.agents:
                if training_data.get(str(agent.id)):
                    result = TaskEvaluator(agent).evaluate_training_data(   # ★LLM 提炼反馈
                        training_data=training_data, agent_id=str(agent.id))
                    CrewTrainingHandler(filename).save_trained_data(        # 存进最终文件
                        agent_id=str(agent.role),
                        trained_data=result.model_dump())
            crewai_event_bus.emit(self, CrewTrainCompletedEvent(
                crew_name=self.name, n_iterations=n_iterations, filename=filename))
    except Exception as e:
        crewai_event_bus.emit(self, CrewTrainFailedEvent(error=str(e), crew_name=self.name))
        self._logger.log("error", f"Training failed: {e}", color="red")
        CrewTrainingHandler(TRAINING_DATA_FILE).clear()   # ★失败清空，别留半截脏数据
        CrewTrainingHandler(filename).clear()
        raise

replay 主流程（crew.py:1944）：

# crew.py:1944
def replay(self, task_id: str, inputs: dict | None = None) -> CrewOutput:
    """Replay the crew execution from a specific task."""
    stored_outputs = self._task_output_handler.load()      # ① 读磁盘上的任务产出存档
    if not stored_outputs:
        raise ValueError(f"Task with id {task_id} not found in the crew's tasks.")

    start_index = self._find_task_index(task_id, stored_outputs)   # ② 找到从哪个任务开始
    if start_index is None:
        raise ValueError(f"Task with id {task_id} not found in the crew's tasks.")

    replay_inputs = inputs if inputs is not None else stored_outputs[start_index]["inputs"]
    self._inputs = replay_inputs
    if replay_inputs:
        self._interpolate_inputs(replay_inputs)             # 用当时/新的 inputs 插值

    if self.process == Process.hierarchical:
        self._create_manager_agent()                        # 层级流程要重新造经理

    for i in range(start_index):                            # ③ 重建断点之前的任务产出
        stored_output = stored_outputs[i]["output"]
        task_output = TaskOutput(
            description=stored_output["description"], agent=stored_output["agent"],
            raw=stored_output["raw"], pydantic=stored_output["pydantic"],
            json_dict=stored_output["json_dict"], output_format=stored_output["output_format"],
            messages=stored_output.get("messages", []))
        self.tasks[i].output = task_output                  # 把存档结果填回任务，不重跑
    self._logging_color = "bold_blue"
    return self._execute_tasks(self.tasks, start_index, True)   # ④ 从 start_index 往后真跑

怎么按 task_id 找到位置（crew.py:1934）：

# crew.py:1934
@staticmethod
def _find_task_index(task_id: str, stored_outputs: list[Any]) -> int | None:
    return next(
        (index for (index, d) in enumerate(stored_outputs)
         if d["task_id"] == str(task_id)),      # 在存档里逐个比对 task_id
        None)                                    # 没找到返回 None

存档的内容（回看 Day 20 的 _store_execution_log，crew.py:1445）：

# crew.py:1445
log = {
    "task": task,
    "output": {"description": ..., "raw": output.raw, "pydantic": output.pydantic,
               "json_dict": output.json_dict, "agent": output.agent, "messages": ...},
    "task_index": task_index,
    "inputs": inputs,                # ★连当时的 inputs 也存了
    "was_replayed": was_replayed,
}
self._task_output_handler.update(task_index, log)   # 写进磁盘存档