回顾 D41 的公开 Flow 组合（flow/flow.py:33）：

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

而 crewai/flow/__init__.py 导出的对话式相关符号（flow/__init__.py:7）：

# flow/__init__.py:7
from crewai.flow.conversation import (
    ChatState,               # 推荐的对话状态结构
    ConversationalConfig,    # 类级默认（意图、退出词等）
    ConversationalInputs,    # kickoff 约定的输入键
)

对话状态的推荐形状（flow/conversation.py:51）：

# flow/conversation.py:51
class ChatState(BaseModel):
    """Recommended persisted state shape for multi-turn flows."""
    id: str = Field(default_factory=lambda: str(uuid4()))   # 会话 id（= 持久化键）
    messages: list[LLMMessage] = Field(default_factory=list) # ★多轮消息历史
    last_user_message: str | None = None                     # 最近一句用户输入
    last_intent: str | None = None                           # 最近一次意图分类结果
    session_ready: bool = False                              # 会话是否已初始化

对话入参可能来自不同地方，先归一（flow/conversation.py:71）：

# flow/conversation.py:71
def normalize_kickoff_inputs(inputs, *, user_message=None, session_id=None):
    if inputs is None and user_message is None and session_id is None:
        return None                          # ★都没给 → 返回 None（保持无状态语义，不硬造空 dict）
    merged: dict[str, Any] = dict(inputs or {})
    if session_id is not None:
        merged["id"] = session_id            # 会话 id → 复用 D45 的恢复机制（inputs["id"]）
    if user_message is not None:
        merged["user_message"] = user_message
    return merged

追加一条消息，兼容 dict / BaseModel / 无状态三种情况（flow/conversation.py:116）：

# flow/conversation.py:116
def append_message(flow, role, content, **extra) -> None:
    message: LLMMessage = {"role": role, "content": content}
    for key, value in extra.items():
        if key in ("tool_call_id", "name", "tool_calls", "files"):
            message[key] = value
    state = getattr(flow, "_state", None)
    if state is not None:
        if isinstance(state, dict):
            messages = state.get("messages")
            if isinstance(messages, list):
                messages.append(message); return          # ① dict 状态有 messages → 写进去
        elif isinstance(state, BaseModel) and hasattr(state, "messages"):
            ...
            state.messages.append(message); return        # ② 模型状态有 messages → 写进去
    # ③ 状态里没 messages → 写进 flow 的后备缓冲
    if not hasattr(flow, "_conversation_messages"):
        object.__setattr__(flow, "_conversation_messages", [])
    flow._conversation_messages.append(message)

收到用户消息，可选地做意图分类（flow/conversation.py:160）：

# flow/conversation.py:160
def receive_user_message(flow, text, *, outcomes=None, llm=None) -> str:
    append_message(flow, "user", text)                # 先记进历史
    set_state_field(flow, "last_user_message", text)
    if outcomes and llm is not None:
        intent = flow.classify_intent(                 # ★让 LLM 从候选 outcomes 里挑一个意图
            text, outcomes, llm=llm, context=get_conversation_messages(flow))
        set_state_field(flow, "last_intent", intent)   # 写进 state.last_intent
        return intent
    return text

整轮的准备工作串起来（flow/conversation.py:184）：

# flow/conversation.py:184
def prepare_conversational_turn(flow, *, user_message=None, intents=None, intent_llm=None, config=None):
    if user_message is None:                           # 没直接给 → 从 state 里捞
        ...
    text = _coerce_user_message_text(user_message)
    if not text.strip():
        return
    set_state_field(flow, "last_intent", None)         # ★每轮重新分类，不复用上轮意图
    resolved_intents = intents or (config.default_intents if config else None)
    resolved_llm = intent_llm or (config.intent_llm if config else None)
    if resolved_intents:
        if resolved_llm is None:
            raise ValueError("intent_llm is required when intents are provided")  # 要分类就得给 LLM
        receive_user_message(flow, text, outcomes=resolved_intents, llm=resolved_llm)
    else:
        receive_user_message(flow, text)               # 不分类，纯记录

对话的类级默认配置（flow/conversation.py:37）：

# flow/conversation.py:37
@dataclass
class ConversationalConfig:
    """Optional class-level defaults for conversational flows."""
    default_intents: Sequence[str] | None = None       # 默认候选意图
    intent_llm: str | None = None                       # 默认分类用的 LLM
    interactive_prompt: str = "You: "                   # 交互模式提示符
    interactive_timeout: float | None = None
    exit_commands: Sequence[str] = field(default_factory=lambda: _EXIT_COMMANDS_DEFAULT)  # 退出词
    defer_trace_finalization: bool = True               # ★对话默认推迟 trace 收尾