langchain-ai/langgraph
1
0
Fork 0
mirror of https://github.com/langchain-ai/langgraph.git synced 2026-07-15 12:35:43 -04:00
Code Issues 314 Packages Projects Releases 17 Wiki Activity

calling model_dump() on pydantic state variables drops tool_calls in AIMessage #1115

New Issue
Closed
opened 2026-02-20 17:43:08 -05:00 by yindo · 7 comments
yindo commented 2026-02-20 17:43:08 -05:00
Owner
Copy Link
Copy Source

Originally created by @Thommy257 on GitHub (Jan 12, 2026).

Checked other resources

  • This is a bug, not a usage question. For questions, please use the LangChain Forum (https://forum.langchain.com/).
  • 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 langchain_core.messages import AIMessage, BaseMessage, ToolCall
from pydantic import BaseModel

class State(BaseModel):
    messages: list[BaseMessage]

msg = AIMessage(content='foo',
                tool_calls=[
                    ToolCall(
                        id='bar',
                        args={'baz': 'qux'},
                        name='quux'
                    ),
                ]
            )
state = State(messages=[msg])

print(msg.model_dump())
print(state.model_dump())

Error Message and Stack Trace (if applicable)

# output
{'content': 'foo', 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'ai', 'name': None, 'id': None, 'tool_calls': [{'name': 'quux', 'args': {'baz': 'qux'}, 'id': 'bar', 'type': 'tool_call'}], 'invalid_tool_calls': [], 'usage_metadata': None}
{'messages': [{'content': 'foo', 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'ai', 'name': None, 'id': None}]}

Description

When using a Pydantic BaseModel as the state schema, tool calls are dropped from AIMessage objects when the state is serialized via model_dump().

Expected Behavior

Both msg.model_dump() and state.model_dump() should include the tool_calls in the serialized output.

Actual Behavior

  • msg.model_dump() correctly includes tool_calls.
  • state.model_dump() drops the tool_calls from the message.

Impact

This causes a KeyError when processing states that have been serialized using Pydantic, as the tool's call information is lost. See example below:

import operator
from typing import Annotated

from dotenv import load_dotenv
from langchain_core.messages import AIMessage, BaseMessage, ToolCall
from langchain_core.prompts.chat import ChatPromptTemplate
from langchain_openai import ChatOpenAI
from langgraph.graph import END, StateGraph
from langgraph.prebuilt import ToolNode
from langchain_core.tools import tool

from pydantic import BaseModel

load_dotenv()


class State(BaseModel):
    messages: Annotated[list[BaseMessage], operator.add]


@tool
def foo(bar: str) -> str:
    """A tool"""
    return f'The argument is {bar}'

prompt = ChatPromptTemplate.from_messages([
    ('system', 'Return "YES" if the tool output is in english, "NO" if not.'),
    ('placeholder', '{messages}')])
llm = ChatOpenAI(name='llm', model='gpt-5-mini', reasoning_effort='low')
agent = prompt | llm

def agent_node(state: State) -> State:
    msg = agent.invoke(state.model_dump())
    return State(messages=[msg])

graph = StateGraph(State)
graph.add_node('tools', ToolNode([foo]))
graph.add_node('agent', agent_node)
graph.set_entry_point('tools')
graph.add_edge('tools', 'agent')
graph.set_finish_point('agent')

compiled_agent = graph.compile()

state = State(messages=[AIMessage(
            content='The word to check is "cat"',
            tool_calls=[
                ToolCall(
                    id='baz',
                    args={'bar': 'cat'},
                    name='foo'
                ),
            ]
        )])

compiled_agent.invoke(state)

--> Triggers:

KeyError: 'tool_call_id'
During task with name 'agent' and id '29960e58-b80d-2b18-c5a2-855168b09a0a'

System Info

System Information

OS: Darwin
OS Version: Darwin Kernel Version 25.2.0: Tue Nov 18 21:07:05 PST 2025; root:xnu-12377.61.12~1/RELEASE_ARM64_T6020
Python Version: 3.12.11 (main, Jul 31 2025, 09:52:35) [Clang 17.0.0 (clang-1700.0.13.5)]

Package Information

langchain_core: 1.2.6
langchain: 1.2.3
langchain_community: 0.4.1
langsmith: 0.6.2
langchain_classic: 1.0.1
langchain_openai: 1.1.7
langchain_text_splitters: 1.1.0
langgraph_sdk: 0.3.1

Optional packages not installed

langserve

Other Dependencies

aiohttp: 3.13.3
dataclasses-json: 0.6.7
httpx: 0.28.1
httpx-sse: 0.4.3
jsonpatch: 1.33
langgraph: 1.0.5
numpy: 2.3.5
openai: 2.14.0
opentelemetry-api: 1.39.1
opentelemetry-exporter-otlp-proto-http: 1.39.1
opentelemetry-sdk: 1.39.1
orjson: 3.11.5
packaging: 25.0
pydantic: 2.12.5
pydantic-settings: 2.12.0
pytest: 9.0.2
PyYAML: 6.0.3
pyyaml: 6.0.3
requests: 2.32.5
requests-toolbelt: 1.0.0
sqlalchemy: 2.0.45
SQLAlchemy: 2.0.45
tenacity: 9.1.2
tiktoken: 0.12.0
typing-extensions: 4.15.0
uuid-utils: 0.13.0
zstandard: 0.25.0

Originally created by @Thommy257 on GitHub (Jan 12, 2026). ### Checked other resources - [x] This is a bug, not a usage question. For questions, please use the LangChain Forum (https://forum.langchain.com/). - [x] I added a clear and detailed title that summarizes the issue. - [x] I read what a minimal reproducible example is (https://stackoverflow.com/help/minimal-reproducible-example). - [x] 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 ```python from langchain_core.messages import AIMessage, BaseMessage, ToolCall from pydantic import BaseModel class State(BaseModel): messages: list[BaseMessage] msg = AIMessage(content='foo', tool_calls=[ ToolCall( id='bar', args={'baz': 'qux'}, name='quux' ), ] ) state = State(messages=[msg]) print(msg.model_dump()) print(state.model_dump()) ``` ### Error Message and Stack Trace (if applicable) ```shell # output {'content': 'foo', 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'ai', 'name': None, 'id': None, 'tool_calls': [{'name': 'quux', 'args': {'baz': 'qux'}, 'id': 'bar', 'type': 'tool_call'}], 'invalid_tool_calls': [], 'usage_metadata': None} {'messages': [{'content': 'foo', 'additional_kwargs': {}, 'response_metadata': {}, 'type': 'ai', 'name': None, 'id': None}]} ``` ### Description When using a Pydantic `BaseModel` as the state schema, tool calls are dropped from `AIMessage` objects when the state is serialized via `model_dump()`. ### Expected Behavior Both `msg.model_dump()` and `state.model_dump()` should include the tool_calls in the serialized output. ### Actual Behavior - `msg.model_dump()` correctly includes tool_calls. - `state.model_dump()` drops the tool_calls from the message. ### Impact This causes a KeyError when processing states that have been serialized using Pydantic, as the tool's call information is lost. See example below: ```python import operator from typing import Annotated from dotenv import load_dotenv from langchain_core.messages import AIMessage, BaseMessage, ToolCall from langchain_core.prompts.chat import ChatPromptTemplate from langchain_openai import ChatOpenAI from langgraph.graph import END, StateGraph from langgraph.prebuilt import ToolNode from langchain_core.tools import tool from pydantic import BaseModel load_dotenv() class State(BaseModel): messages: Annotated[list[BaseMessage], operator.add] @tool def foo(bar: str) -> str: """A tool""" return f'The argument is {bar}' prompt = ChatPromptTemplate.from_messages([ ('system', 'Return "YES" if the tool output is in english, "NO" if not.'), ('placeholder', '{messages}')]) llm = ChatOpenAI(name='llm', model='gpt-5-mini', reasoning_effort='low') agent = prompt | llm def agent_node(state: State) -> State: msg = agent.invoke(state.model_dump()) return State(messages=[msg]) graph = StateGraph(State) graph.add_node('tools', ToolNode([foo])) graph.add_node('agent', agent_node) graph.set_entry_point('tools') graph.add_edge('tools', 'agent') graph.set_finish_point('agent') compiled_agent = graph.compile() state = State(messages=[AIMessage( content='The word to check is "cat"', tool_calls=[ ToolCall( id='baz', args={'bar': 'cat'}, name='foo' ), ] )]) compiled_agent.invoke(state) ``` --> Triggers: ``` KeyError: 'tool_call_id' During task with name 'agent' and id '29960e58-b80d-2b18-c5a2-855168b09a0a' ``` ### System Info System Information ------------------ > OS: Darwin > OS Version: Darwin Kernel Version 25.2.0: Tue Nov 18 21:07:05 PST 2025; root:xnu-12377.61.12~1/RELEASE_ARM64_T6020 > Python Version: 3.12.11 (main, Jul 31 2025, 09:52:35) [Clang 17.0.0 (clang-1700.0.13.5)] Package Information ------------------- > langchain_core: 1.2.6 > langchain: 1.2.3 > langchain_community: 0.4.1 > langsmith: 0.6.2 > langchain_classic: 1.0.1 > langchain_openai: 1.1.7 > langchain_text_splitters: 1.1.0 > langgraph_sdk: 0.3.1 Optional packages not installed ------------------------------- > langserve Other Dependencies ------------------ > aiohttp: 3.13.3 > dataclasses-json: 0.6.7 > httpx: 0.28.1 > httpx-sse: 0.4.3 > jsonpatch: 1.33 > langgraph: 1.0.5 > numpy: 2.3.5 > openai: 2.14.0 > opentelemetry-api: 1.39.1 > opentelemetry-exporter-otlp-proto-http: 1.39.1 > opentelemetry-sdk: 1.39.1 > orjson: 3.11.5 > packaging: 25.0 > pydantic: 2.12.5 > pydantic-settings: 2.12.0 > pytest: 9.0.2 > PyYAML: 6.0.3 > pyyaml: 6.0.3 > requests: 2.32.5 > requests-toolbelt: 1.0.0 > sqlalchemy: 2.0.45 > SQLAlchemy: 2.0.45 > tenacity: 9.1.2 > tiktoken: 0.12.0 > typing-extensions: 4.15.0 > uuid-utils: 0.13.0 > zstandard: 0.25.0
yindo added the bugpending labels 2026-02-20 17:43:08 -05:00
yindo closed this issue 2026-02-20 17:43:08 -05:00
yindo commented 2026-02-20 17:43:08 -05:00
Author
Owner
Copy Link
Copy Source

@ahmedibrahimX commented on GitHub (Jan 13, 2026):

Hi! I'd like to help with this issue. I've investigated the root cause and have a proposed solution I'd like to run by you before implementing.

Root Cause

The issue occurs during Pydantic serialization. When State.model_dump() is called on a model with list[BaseMessage] type annotations, Pydantic treats this as a union and serializes each message using only the BaseMessage schema, which drops subclass-specific fields like tool_calls.

Why the behavior differs:

  • AIMessage.model_dump() → uses concrete AIMessage schema → all fields preserved
  • State(messages=[msg]).model_dump() → uses base BaseMessage schema → subclass fields lost

Proposed Solution

Modify JsonPlusSerializer in langgraph/checkpoint/serde/jsonplus.py to handle BaseMessage objects specially:

  1. In _msgpack_default(): Check for BaseMessage instances before the general Pydantic model check and serialize using the concrete message class
  2. In _msgpack_ext_hook(): Use LC_REVIVER during deserialization to properly reconstruct message objects with all fields

Benefits:

  • Fixes the issue without requiring changes to user code
  • Preserves backward compatibility
  • Graceful fallback if langchain_core isn't available
  • Works for any State schema with message lists

Would this approach work for you? Happy to share a draft PR or adjust based on your feedback!

@ahmedibrahimX commented on GitHub (Jan 13, 2026): Hi! I'd like to help with this issue. I've investigated the root cause and have a proposed solution I'd like to run by you before implementing. ## Root Cause The issue occurs during Pydantic serialization. When `State.model_dump()` is called on a model with `list[BaseMessage]` type annotations, Pydantic treats this as a union and serializes each message using only the `BaseMessage` schema, which drops subclass-specific fields like `tool_calls`. **Why the behavior differs:** - `AIMessage.model_dump()` → uses concrete `AIMessage` schema → all fields preserved - `State(messages=[msg]).model_dump()` → uses base `BaseMessage` schema → subclass fields lost ## Proposed Solution Modify `JsonPlusSerializer` in `langgraph/checkpoint/serde/jsonplus.py` to handle `BaseMessage` objects specially: 1. **In `_msgpack_default()`**: Check for `BaseMessage` instances before the general Pydantic model check and serialize using the concrete message class 2. **In `_msgpack_ext_hook()`**: Use `LC_REVIVER` during deserialization to properly reconstruct message objects with all fields **Benefits:** - [x] Fixes the issue without requiring changes to user code - [x] Preserves backward compatibility - [x] Graceful fallback if `langchain_core` isn't available - [x] Works for any State schema with message lists Would this approach work for you? Happy to share a draft PR or adjust based on your feedback!
yindo commented 2026-02-20 17:43:08 -05:00
Author
Owner
Copy Link
Copy Source

@ahmedibrahimX commented on GitHub (Jan 13, 2026):

I've implemented this solution in PR #6685. The fix adds special handling for BaseMessage objects in both standalone and nested contexts (within Pydantic State objects). All tests passing and ready for review!

@ahmedibrahimX commented on GitHub (Jan 13, 2026): I've implemented this solution in PR #6685. The fix adds special handling for BaseMessage objects in both standalone and nested contexts (within Pydantic State objects). All tests passing and ready for review!
yindo commented 2026-02-20 17:43:08 -05:00
Author
Owner
Copy Link
Copy Source

@ahmedibrahimX commented on GitHub (Jan 14, 2026):

Hi @Thommy257,

I did a deep investigation into this issue and wanted to share what I found.

The Good News

LangGraph’s checkpoint system already handles this correctly, tool_calls are preserved when using MemorySaver and other checkpoint savers. I tested this with State defined as both a TypedDict and a Pydantic BaseModel, and both work as expected.

What’s Causing the Behavior

The issue you’re seeing happens when state.model_dump() is called directly inside the node:

def agent_node(state: State) -> State:
    msg = agent.invoke(state.model_dump())
    return State(messages=[msg])

When Pydantic serializes a model containing list[BaseMessage], it uses the base class schema, which doesn’t include subclass-specific fields like tool_calls. This is expected Pydantic behavior and occurs outside LangGraph’s serialization path.

Recommended Pattern

Instead of calling model_dump() on the full state, let LangGraph pass message objects through directly:

def agent_node(state: State) -> State:
    msg = agent.invoke({"messages": state.messages})
    return State(messages=[msg])

This preserves message types and tool_calls.

Why This Happens

LangGraph’s checkpoint persistence uses JsonPlusSerializer, which has special handling to preserve message subclasses. Calling model_dump() directly skips that mechanism and relies purely on Pydantic’s base-class serialization.

PR Status

I opened PR #6685 initially, but closed it after realizing it addressed a different edge case (nested Pydantic data in custom checkpoint payloads). This issue is best solved through usage rather than a code change.

Hope this helps — happy to clarify or test further.

@ahmedibrahimX commented on GitHub (Jan 14, 2026): Hi @Thommy257, I did a deep investigation into this issue and wanted to share what I found. ### The Good News LangGraph’s checkpoint system already handles this correctly, `tool_calls` are preserved when using MemorySaver and other checkpoint savers. I tested this with State defined as both a TypedDict and a Pydantic BaseModel, and both work as expected. ### What’s Causing the Behavior The issue you’re seeing happens when `state.model_dump()` is called directly inside the node: ```python def agent_node(state: State) -> State: msg = agent.invoke(state.model_dump()) return State(messages=[msg]) ``` When Pydantic serializes a model containing `list[BaseMessage]`, it uses the base class schema, which doesn’t include subclass-specific fields like tool_calls. This is expected Pydantic behavior and occurs outside LangGraph’s serialization path. ### Recommended Pattern Instead of calling `model_dump()` on the full state, let LangGraph pass message objects through directly: ```python def agent_node(state: State) -> State: msg = agent.invoke({"messages": state.messages}) return State(messages=[msg]) ``` This preserves message types and tool_calls. ### Why This Happens LangGraph’s checkpoint persistence uses JsonPlusSerializer, which has special handling to preserve message subclasses. Calling `model_dump()` directly skips that mechanism and relies purely on Pydantic’s base-class serialization. ### PR Status I opened PR #6685 initially, but closed it after realizing it addressed a different edge case (nested Pydantic data in custom checkpoint payloads). This issue is best solved through usage rather than a code change. Hope this helps — happy to clarify or test further.
yindo commented 2026-02-20 17:43:08 -05:00
Author
Owner
Copy Link
Copy Source

@Thommy257 commented on GitHub (Jan 14, 2026):

Hi @ahmedibrahimX,

Thanks for your response. In my specific use case, State has numerous variables that I would potentially like to pass to the agent. I have to admit that I'm a bit lazy here, but keeping track of what variables need to be passed to the ChatPromptTemplate and which not seems a bit tedious.

I solved the issue by adding a shallow_dump method to my State class:

def shallow_dump(self, /) -> dict[str, Any]:
    """Get a shallow dictionary dump of the model, depth 1 only."""
    keys = chain(self.__class__.model_fields.keys(),
                 self.__class__.model_computed_fields.keys())
    return {k: getattr(self, k) for k in keys}
@Thommy257 commented on GitHub (Jan 14, 2026): Hi @ahmedibrahimX, Thanks for your response. In my specific use case, `State` has numerous variables that I would potentially like to pass to the agent. I have to admit that I'm a bit lazy here, but keeping track of what variables need to be passed to the ChatPromptTemplate and which not seems a bit tedious. I solved the issue by adding a `shallow_dump` method to my `State` class: ```python def shallow_dump(self, /) -> dict[str, Any]: """Get a shallow dictionary dump of the model, depth 1 only.""" keys = chain(self.__class__.model_fields.keys(), self.__class__.model_computed_fields.keys()) return {k: getattr(self, k) for k in keys} ```
yindo commented 2026-02-20 17:43:08 -05:00
Author
Owner
Copy Link
Copy Source

@ahmedibrahimX commented on GitHub (Jan 14, 2026):

Hi @Thommy257 , that makes total sense, especially when the state grows and you want a low-friction way to pass context through.

I like this approach! Conceptually, shallow_dump is closer to state projection than serialization, which avoids the Pydantic base-class issue while keeping the ergonomics you want.

As long as the receiving code can handle the actual message objects (rather than needing pre-serialized dictionaries), this feels like a clean and explicit solution. Thanks for sharing, I think others may run into the same tradeoff.

@ahmedibrahimX commented on GitHub (Jan 14, 2026): Hi @Thommy257 , that makes total sense, especially when the state grows and you want a low-friction way to pass context through. I like this approach! Conceptually, `shallow_dump` is closer to state projection than serialization, which avoids the Pydantic base-class issue while keeping the ergonomics you want. As long as the receiving code can handle the actual message objects (rather than needing pre-serialized dictionaries), this feels like a clean and explicit solution. Thanks for sharing, I think others may run into the same tradeoff.
yindo commented 2026-02-20 17:43:08 -05:00
Author
Owner
Copy Link
Copy Source

@Momo-Not-Emo commented on GitHub (Jan 29, 2026):

This could be fix by using SerializeAsAny from pydantic. I have found a similar issue in https://github.com/langchain-ai/langchain/issues/34925

from pydantic import SerializeAsAny

class State(BaseModel):
    messages: list[SerializeAsAny[BaseMessage]]
@Momo-Not-Emo commented on GitHub (Jan 29, 2026): This could be fix by using `SerializeAsAny` from pydantic. I have found a similar issue in https://github.com/langchain-ai/langchain/issues/34925 ```py from pydantic import SerializeAsAny class State(BaseModel): messages: list[SerializeAsAny[BaseMessage]] ```
yindo commented 2026-02-20 17:43:08 -05:00
Author
Owner
Copy Link
Copy Source

@mdrxy commented on GitHub (Jan 30, 2026):

Closing as this is not related to LangGraph. See the issue mentioned by @Momo-Not-Emo

@mdrxy commented on GitHub (Jan 30, 2026): Closing as this is not related to LangGraph. See the issue mentioned by @Momo-Not-Emo
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
agents
awaiting author revision
bug
change
chat models
checkpointer
ci
code quality
code quality
dependencies
documentation
drawing
duplicate
enhancement
good first issue
help wanted
invalid
maintainer
meta
missing information
not maintained by langchain
open-swe-auto
open-swe-max
open-swe-max-auto
pending
performance
pull-request
Mirrored from GitHub Pull Request
 question
requires design
sdk
server
stale
studio
tests
type checking
usage question
v0.5
v1.1
No Label bug pending
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
yindo
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: langchain-ai/langgraph#1115
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?