Subgraph checkpointer=True causes subgraph to be skipped

[shengbo-ma]

Checked other resources

• This is a bug, not a usage question. For questions, please use GitHub Discussions.
• I added a clear and detailed title that summarizes the issue.
• I read what a minimal reproducible example is (https://stackoverflow.com/help/minimal-reproducible-example).
• I included a self-contained, minimal example that demonstrates the issue INCLUDING all the relevant imports. The code run AS IS to reproduce the issue.

Example Code

from typing import Literal

from langchain_core.runnables import RunnableConfig
from langgraph.checkpoint.memory import MemorySaver
from langgraph.graph import END, START, StateGraph
from langgraph.graph.state import Command
from langgraph.types import interrupt
from rich import get_console
from typing_extensions import TypedDict

###############
# Subgraph
###############


class SubGraphState(TypedDict, total=False):
    parent_counter: int
    sub_counter: int


def subgraph_accumulator(state: SubGraphState) -> SubGraphState:
    get_console().print("---subgraph counter node---")
    get_console().print(f"{state = }")
    # ask for human approval
    human_feedback = interrupt("get human feedback")
    print(f"{human_feedback = }")

    # continue counting
    sub_counter = state["sub_counter"] + 1 if "sub_counter" in state else 1
    return {"sub_counter": sub_counter}


sub_graph = (
    StateGraph(SubGraphState)
    .add_node(subgraph_accumulator)
    .add_edge(START, subgraph_accumulator.__name__)
    .add_edge(subgraph_accumulator.__name__, END)
    .compile(
        checkpointer=True,  # BUG: This causes an issue that subgraph nodes are not executed at all after first interruption
    )
)
sub_graph.name = "sub"

###############
# Parent Graph
###############

MAX_ITERATION = 3


class ParentGraphState(TypedDict):
    parent_counter: int


def parent_graph_accumulator(
    state: ParentGraphState,
) -> Command[Literal["sub", "__end__"]]:
    print("---parent counter node---")
    get_console().print(f"{state = }")
    parent_counter = state["parent_counter"] + 1 if "parent_counter" in state else 0

    # goto end when max iteration reaches
    goto = sub_graph.get_name() if parent_counter < MAX_ITERATION else END
    get_console().print(f"going to node {goto}")
    return Command(
        update={
            "parent_counter": parent_counter,
        },
        goto=goto,
    )


parent_agent = (
    StateGraph(ParentGraphState)
    .add_node(parent_graph_accumulator)
    .add_node(sub_graph)
    .add_edge(START, parent_graph_accumulator.__name__)
    .add_edge(sub_graph.get_name(), parent_graph_accumulator.__name__)
    .compile(checkpointer=MemorySaver())
)

# visualize graph
mermaid_graph = parent_agent.get_graph(xray=True).draw_mermaid()
print(mermaid_graph)

###############
# Conversation
###############

config: RunnableConfig = {"configurable": {"thread_id": "42"}, "recursion_limit": MAX_ITERATION+1}

inputs = [
    ParentGraphState(parent_counter=0),
    Command(resume="human feedback 1"),
    Command(resume="human feedback 2"),
]
for input_ in inputs:
    print(f"{input_ = }")
    for event in parent_agent.stream(
        # resume the conversation
        input_,
        config,
        stream_mode="updates",
        subgraphs=True,
    ):
        print("Streaming event ...")
        print(event)

Error Message and Stack Trace (if applicable)

input_ = {'parent_counter': 0}
---parent counter node---
state = {'parent_counter': 0}
going to node sub
Streaming event ...
((), {'parent_graph_accumulator': {'parent_counter': 1}})
---subgraph counter node---
state = {'parent_counter': 1}
Streaming event ...
((), {'__interrupt__': (Interrupt(value='get human feedback', resumable=True, ns=['sub', 'subgraph_accumulator:f187d019-da4b-d432-bcd2-cea142aa7e35'], when='during'),)})
input_ = Command(resume='human feedback 1')
---subgraph counter node---
state = {'parent_counter': 1}
human_feedback = 'human feedback 1'
Streaming event ...
(('sub',), {'subgraph_accumulator': {'sub_counter': 1}})
Streaming event ...
((), {'sub': {'parent_counter': 1}})
---parent counter node---
state = {'parent_counter': 1}
going to node sub
Streaming event ...
((), {'parent_graph_accumulator': {'parent_counter': 2}})
Streaming event ...
((), {'sub': {'parent_counter': 1}})  <------- BUG: should be subgraph execution like (('sub',), {'subgraph_accumulator': {...}})
---parent counter node---
state = {'parent_counter': 1}
going to node sub
Streaming event ...
((), {'parent_graph_accumulator': {'parent_counter': 2}})
Streaming event ...
((), {'sub': {'parent_counter': 1}})
---parent counter node---
state = {'parent_counter': 1}
going to node sub
Streaming event ...
((), {'parent_graph_accumulator': {'parent_counter': 2}})
Traceback (most recent call last):
  File "/home/linux/arcgis-ai-assistants/python/arcgis-assistant/.tmp/subgraph_state_lose/loop_subgraph_with_interrupt.py", line 99, in <module>
    for event in parent_agent.stream(
  File "/home/linux/miniconda3/envs/test/lib/python3.11/site-packages/langgraph/pregel/__init__.py", line 1690, in stream
    raise GraphRecursionError(msg)
langgraph.errors.GraphRecursionError: Recursion limit of 4 reached without hitting a stop condition. You can increase the limit by setting the `recursion_limit` config key.
For troubleshooting, visit: https://python.langchain.com/docs/troubleshooting/errors/GRAPH_RECURSION_LIMIT

Description

I encountered an issue when building a multi-agent graph for multi-turn conversations, where a subgraph has a human feedback node. It interrupts and takes human feedback.

Here is an example graph to reproduce the issue

• A parent graph has a loop, calling a sub graph until the parent counter reaches a pre-defined limit.
• Subgraph node interrupts and take human feedback.
• The sub graph should remember its state from previous run (checkpointer=True).

Expected Behavior

• graph should interrupt twice, and resume with human inputs
• subgraph should should persist its state on each run (since checkpointer=True)

Actual Behavior

• The first interrupt and resume is as expected
• The second interrupt never happens. The parent graph never executes sub graph counter node after resuming the first interrupt. The sub graph nodes output the same parent counter (=1) repeatedly, leading to recursion limit error since parent counter does not increase.

Observation
If removing checkpointer=True, the graph executes as expected, i.e. the parent counter increases correctly. No bug. (In this cause, ff course, the sub graph states from previous run is not persisted )
It seems in subgraph assigning checkpointer=True and calling interrupt conflicts in some way.

%%{init: {'flowchart': {'curve': 'linear'}}}%%
graph TD;
        __start__([<p>__start__</p>]):::first
        parent_graph_accumulator(parent_graph_accumulator)
        sub(sub)
        __end__([<p>__end__</p>]):::last
        __start__ --> parent_graph_accumulator;
        sub --> parent_graph_accumulator;
        parent_graph_accumulator -.-> sub;
        parent_graph_accumulator -.-> __end__;
        classDef default fill:#f2f0ff,line-height:1.2
        classDef first fill-opacity:0
        classDef last fill:#bfb6fc

Loading

LangGraph Version
0.2.67

System Info

System Information

OS: Linux
OS Version: #135~20.04.1-Ubuntu SMP Mon Oct 7 13:56:22 UTC 2024
Python Version: 3.11.11 (main, Dec 11 2024, 16:28:39) [GCC 11.2.0]

Package Information

langchain_core: 0.3.31
langsmith: 0.3.1
langgraph_sdk: 0.1.51

Optional packages not installed

langserve

Other Dependencies

httpx: 0.28.1
jsonpatch: 1.33
langsmith-pyo3: Installed. No version info available.
orjson: 3.10.15
packaging: 24.2
pydantic: 2.10.6
pytest: Installed. No version info available.
PyYAML: 6.0.2
requests: 2.32.3
requests-toolbelt: 1.0.0
rich: 13.9.4
tenacity: 9.0.0
typing-extensions: 4.12.2
zstandard: 0.23.0

[shengbo-ma]

Since checkpointer=True is not yet officially documented and supported, not sure if I use it in a wrong way.
Could you please take a look? @vbarda @eyurtsev
Appreciate your efforts