时间旅行的 API 都返回 StateSnapshot，定义在 types.py:643-659：

# types.py:643
class StateSnapshot(NamedTuple):
    values: dict[str, Any] | Any        # 这一步的状态值
    next: tuple[str, ...]               # 下一步该跑哪些节点
    config: RunnableConfig              # 这份快照的坐标(含 checkpoint_id)
    metadata: CheckpointMetadata | None # 元数据(step/source/...)
    created_at: str | None              # 创建时间戳
    parent_config: RunnableConfig | None# 父快照坐标(往回走用)
    tasks: tuple[PregelTask, ...]       # 本步的任务(含中断/错误信息)

main.py:1392 起，先看它怎么拿 checkpointer、怎么处理子图：

# main.py:1392
def get_state(self, config, *, subgraphs=False) -> StateSnapshot:
    """Get the current state of the graph."""
    checkpointer = ensure_config(config)[CONF].get(
        CONFIG_KEY_CHECKPOINTER, self.checkpointer)
    if isinstance(checkpointer, BaseCheckpointSaver):
        checkpointer = self._apply_checkpointer_allowlist(checkpointer)
    if not checkpointer:
        raise ValueError("No checkpointer set")        # ① 没存档 = 没历史可看
    if (checkpoint_ns := config[CONF].get(CONFIG_KEY_CHECKPOINT_NS, "")) \
            and CONFIG_KEY_CHECKPOINTER not in config[CONF]:
        recast = recast_checkpoint_ns(checkpoint_ns)    # ② 子图：剥掉 task_id
        for _, pregel in self.get_subgraphs(namespace=recast, recurse=True):
            return pregel.get_state(...)                #    转发给对应子图
        else:
            raise ValueError(f"Subgraph {recast} not found")
    ...  # 否则读本图最新 checkpoint，组装成 StateSnapshot 返回

main.py:1480-1531，核心是调 checkpointer.list()：

# main.py:1480
def get_state_history(self, config, *, filter=None, before=None, limit=None):
    """Get the history of the state of the graph."""
    config = ensure_config(config)
    checkpointer = config[CONF].get(CONFIG_KEY_CHECKPOINTER, self.checkpointer)
    ...
    if not checkpointer:
        raise ValueError("No checkpointer set")
    # (子图转发逻辑，同 get_state) ...
    config = merge_configs(self.config, config, {...})
    # eagerly consume list() to avoid holding up the db cursor
    for checkpoint_tuple in list(                       # ① 一次性取出，别占着游标
        checkpointer.list(config, before=before, limit=limit, filter=filter)
    ):
        yield self._prepare_state_snapshot(             # ② 每份 → StateSnapshot
            checkpoint_tuple.config, checkpoint_tuple)

update_state（main.py:2515，异步版 aupdate_state main.py:2528）不是"覆盖旧档"，而是造一份新 checkpoint，模拟"某个节点执行并写入"。看它执行更新的核心 main.py:2413-2467：

# main.py:2413
for as_node, values, provided_task_id in valid_updates:
    writers = self.nodes[as_node].flat_writers       # ① 拿"这个节点"的写入器
    if not writers:
        raise InvalidUpdateError(f"Node {as_node} has no writers")
    writes: deque[tuple[str, Any]] = deque()
    task = PregelTaskWrites((), as_node, writes, [INTERRUPT])
    task_id = provided_task_id or (... uuid5(UUID(checkpoint["id"]), INTERRUPT))
    run = RunnableSequence(*writers) if len(writers) > 1 else writers[0]
    await run.ainvoke(values, patch_config(config, ...  # ② 用你的 values 跑写入器
        configurable={ CONFIG_KEY_SEND: writes.extend, ... }))
# main.py:2468
apply_writes(checkpoint, channels, run_tasks,          # ③ 把写 apply 到通道
             checkpointer.get_next_version, self.trigger_to_nodes)
# main.py:2485
checkpoint = create_checkpoint(checkpoint, channels, step + 1, ...)  # ④ 造新档
next_config = await checkpointer.aput(checkpoint_config, checkpoint, ...) # ⑤ 存

你可以显式指定 as_node="某节点"；不指定时，框架会猜（main.py:2381-2397）：

# main.py:2381
elif as_node is None:
    last_seen_by_node = sorted(
        (v, n)
        for n, seen in checkpoint["versions_seen"].items()
        if n in self.nodes
        for v in seen.values())
    # if two nodes updated the state at the same time, it's ambiguous
    if last_seen_by_node:
        if len(last_seen_by_node) == 1:
            as_node = last_seen_by_node[0][1]              # 只有一个候选，就是它
        elif last_seen_by_node[-1][0] != last_seen_by_node[-2][0]:
            as_node = last_seen_by_node[-1][1]             # 最后动的那个节点
if as_node is None:
    raise InvalidUpdateError("Ambiguous update, specify as_node")  # 猜不出→报错

回到过去重跑时，一个棘手问题：图里如果有子图，子图也有自己的 checkpoint。重放到过去某点时，子图该读哪份存档？这归 ReplayState 管（_replay.py:14-73）：

# _replay.py:14
class ReplayState:
    """Tracks which subgraphs have already loaded their pre-replay checkpoint."""
    __slots__ = ("checkpoint_id", "_visited_ns")
    def __init__(self, checkpoint_id: str) -> None:
        self.checkpoint_id = checkpoint_id      # 重放的"目标时间点"
        self._visited_ns: set[str] = set()      # 记哪些子图已加载过
    # _replay.py:52
    def get_checkpoint(self, checkpoint_ns, checkpointer, checkpoint_config):
        if self._is_first_visit(checkpoint_ns):             # ① 该子图第一次被访问
            for saved in checkpointer.list(
                checkpoint_config,
                before={"configurable": {"checkpoint_id": self.checkpoint_id}},
                limit=1):
                return saved                                 #    取"重放点之前"的档
            return None
        return checkpointer.get_tuple(checkpoint_config)     # ② 之后按正常最新档

# 1. 读三个 API 源码
sed -n '1480,1531p' libs/langgraph/langgraph/pregel/main.py     # get_state_history
sed -n '2413,2506p' libs/langgraph/langgraph/pregel/main.py     # update_state 核心
sed -n '14,73p'     libs/langgraph/langgraph/_internal/_replay.py
# 2. 亲手时间旅行
python - <<'PY'
from langgraph.graph import StateGraph, START, END
from langgraph.checkpoint.memory import InMemorySaver
from typing_extensions import TypedDict
class S(TypedDict):
    x: int
g=StateGraph(S)
g.add_node("a", lambda s:{"x":s["x"]+1})
g.add_node("b", lambda s:{"x":s["x"]+10})
g.add_edge(START,"a"); g.add_edge("a","b"); g.add_edge("b",END)
app=g.compile(checkpointer=InMemorySaver())
cfg={"configurable":{"thread_id":"t"}}
app.invoke({"x":0}, cfg)
hist=list(app.get_state_history(cfg))
print("历史档数:", len(hist))                       # 若干份(倒序)
for h in hist: print("  step", h.metadata["step"], "x=", h.values.get("x"), "next=", h.next)
# 挑一个 a 刚跑完的历史点，改 x，再从那分叉
target=[h for h in hist if h.next==("b",)][0]
app.update_state(target.config, {"x": 100}, as_node="a")
print("从改后的点继续:", app.invoke(None, {"configurable":{"thread_id":"t"}}))
PY