Time travel to resumed node

[balex89]

Problem

Having a node that was interrupted with interrupt and then resumed with Command(resume=...), when start with that checkpoint again and try to resume with a different value, the new value won't apply.

Minimal example:

from langgraph.checkpoint.memory import MemorySaver
from langgraph.graph import StateGraph, START
from langgraph.types import Command, interrupt
from typing_extensions import TypedDict

class State(TypedDict):
    pass

graph = StateGraph(State)

def node(_: State):
    data = interrupt(None)
    print(f'{data=}')

# a graph with  1 node
graph.add_node(node)
graph.add_edge(START, 'node')

saver = MemorySaver()
agent = graph.compile(saver)

config = {'configurable': {'thread_id': '1'}}

# 1-st invoke -> interrupt -> resume 
agent.invoke(State(), config)
agent.invoke(Command(resume="foo"), config)

# 2-nd resume from node checkpoint (skip start)
old_config = list(agent.checkpointer.list(config))[-2].config
agent.invoke(Command(resume="bar"), old_config)

Output i expect:

data='foo'
data='bar'

What i get:

data='foo'
data='foo'

Curiously enough the checkpoint after second resume contains following pending_writes:

...
('b570d40d-d313-1b55-76ad-d30ff326af64', '__interrupt__', [Interrupt(value=None, resumable=True, ns=['node:b570d40d-d313-1b55-76ad-d30ff326af64'])])
('00000000-0000-0000-0000-000000000000', '__resume__', 'bar'), 
('b570d40d-d313-1b55-76ad-d30ff326af64', '__resume__', ['foo']),
...

Don't know what that means exactly. But apparently the first __resume__ somehow sticks.

Am i doing it wrong? What are the cleanest way to work it around?
Thanks.