先看底座 BaseMessage（libs/core/langchain_core/messages/base.py:93），一条"发言记录"的全部字段：

# libs/core/langchain_core/messages/base.py:93
class BaseMessage(Serializable):
    """Base abstract message class.
    Messages are the inputs and outputs of a chat model."""

    content: str | list[str | dict[Any, Any]]          # 发言内容：字符串 或 多模态块列表
    additional_kwargs: dict[Any, Any] = Field(default_factory=dict)
    # 厂商私有的附加数据（例如 OpenAI 原始格式的 tool calls）
    response_metadata: dict[Any, Any] = Field(default_factory=dict)
    # 响应元数据：response headers、logprobs、token 数、模型名…
    type: str            # 消息类型标签，每个子类唯一（"human"/"ai"/"tool"…），用于序列化识别
    name: str | None = None    # 可选的人类可读名字
    id: str | None = Field(default=None, coerce_numbers_to_str=True)  # 可选唯一 ID，最好由厂商提供

HumanMessage（libs/core/langchain_core/messages/human.py:9）薄得惊人——它几乎只是"把 type 写死成 human"：

# libs/core/langchain_core/messages/human.py:9
class HumanMessage(BaseMessage):
    """Message from the user."""

    type: Literal["human"] = "human"        # ★唯一的实质区别：发言人标签写死

    def __init__(self, content=None, content_blocks=None, **kwargs):
        """Specify `content` as positional arg or `content_blocks` for typing."""
        if content_blocks is not None:      # 新式多模态写法：显式传内容块列表
            super().__init__(content=cast("list[str | dict[Any, Any]]", content_blocks), **kwargs)
        else:
            super().__init__(content=content, **kwargs)   # 常规写法：HumanMessage("你好")

SystemMessage（messages/system.py:9）结构完全一样，只是 type="system"。真正"重"的是 AIMessage（libs/core/langchain_core/messages/ai.py:160）——模型的回复除了文本还带三样标准化好的东西：

# libs/core/langchain_core/messages/ai.py:160
class AIMessage(BaseMessage):
    """Message from an AI.
    ...consists of both the raw output as returned by the model and
    standardized fields (e.g., tool calls, usage metadata) added by the
    LangChain framework."""

    tool_calls: list[ToolCall] = Field(default_factory=list)          # ① 工具调用请求（标准格式）
    invalid_tool_calls: list[InvalidToolCall] = Field(default_factory=list)  # ② 解析失败的调用
    usage_metadata: UsageMetadata | None = None                       # ③ 跨厂商统一的 token 用量
    type: Literal["ai"] = "ai"

工单本体是个 TypedDict：ToolCall（libs/core/langchain_core/messages/tool.py:206）：

# libs/core/langchain_core/messages/tool.py:206
class ToolCall(TypedDict):
    """Represents an AI's request to call a tool.
    Example:
        {"name": "foo", "args": {"a": 1}, "id": "123"}
    """
    name: str                 # 要调哪个工具
    args: dict[str, Any]      # 调用参数（已经解析成 dict，不是 JSON 字符串！）
    id: str | None            # ★工单号：多个并行调用时靠它对号入座
    type: NotRequired[Literal["tool_call"]]   # 判别标签

工具执行完，结果装进 ToolMessage（libs/core/langchain_core/messages/tool.py:26）送回给模型：

# libs/core/langchain_core/messages/tool.py:26
class ToolMessage(BaseMessage, ToolOutputMixin):
    """Message for passing the result of executing a tool back to a model.

    `tool_call_id` is used to associate the tool call request with the tool
    call response. Useful in situations where a chat model is able to
    request multiple tool calls in parallel."""

    tool_call_id: str                    # ★回填的工单号，必填！
    type: Literal["tool"] = "tool"
    artifact: Any = None                 # 不发给模型的"完整原始产物"

答案是 Chunk 类支持 + 运算：BaseMessageChunk.__add__（libs/core/langchain_core/messages/base.py:412），核心内容拼接靠 merge_content（libs/core/langchain_core/messages/base.py:366）：

# libs/core/langchain_core/messages/base.py:366
def merge_content(first_content, *contents):
    """Merge multiple message contents."""
    merged = "" if first_content is None else first_content
    for content in contents:
        if isinstance(merged, str):
            if isinstance(content, str):
                merged += content                # 字符串 + 字符串：直接拼
            else:
                merged = [merged, *content]      # 字符串 + 列表：升格成列表
        elif isinstance(content, list):
            merged = merge_lists(merged, content)   # 列表 + 列表：按块合并
        elif merged and isinstance(merged[-1], str):
            merged[-1] += content                # 列表末尾是字符串：接着写
        ...
    return merged