LangChain 1.0 Tool Runtime bloats trace inputs with chat history

[ichitaka]

Describe the feature or potential improvement

Langchain now recommends the runtime to pass parameters into tools through the runtime context. The problem is, that when I add support for the runtime parameter into my tools, my trace shows the full runtime. Here a simplified input of a tool call in langfuse:

{
    <TOOL CALL PARAMETERS REMOVED>
    'runtime': ToolRuntime(state={'messages': [
        SystemMessage(content='<<SYSTEM PROMPT REMOVED>>', additional_kwargs={}, response_metadata={}, id='a1388593-c796-4273-ba24-78d98cf88ba1'), 
        HumanMessage(content=[{'type': 'text', 'text': 'Hi what documents do I have in my knowledgebase?'}], additional_kwargs={}, response_metadata={}, id='410418ba-9a0e-47c4-9edf-2bbd35625a2a'), 
        AIMessage(content='<<TOOL CALL REMOVED>>', tool_calls='<<REMOVED>>', additional_kwargs={}, response_metadata={'model': 'gpt-4.1-2025-04-14', 'status': 'completed'}, id='resp_0071dd80d00c8a5f006928108bfea0819097569b87cc1f551a')
    ]}), 
    'context': Context(user_id=4), 
    'config': {
        'tags': [], 
        'metadata': {'langgraph_step': 2, 'langgraph_node': 'tools', 'langgraph_triggers': ('__pregel_push',), 'langgraph_path': ('__pregel_push', 0, False), 'langgraph_checkpoint_ns': 'tools:<img width="631" height="214" alt="image" src="https://github.com/user-attachments/assets/2b1b1493-fc84-42cf-9141-43eb90ccc45c" />-227a-8a72-133e-9b2de5691c92', 'checkpoint_ns': 'tools:85c45f42-227a-8a72-133e-9b2de5691c92'}, 
        'callbacks': '<langchain_core.callbacks.manager.CallbackManager object>', 
        'recursion_limit': 10000, 
        'configurable': {
            '__pregel_runtime': 'Runtime(...)', 
            '__pregel_task_id': '85c45f42-227a-8a72-133e-9b2de5691c92', 
            '__pregel_send': '<method>', 
            '__pregel_read': '<function>', 
            '__pregel_checkpointer': None, 
            'checkpoint_map': {'': '1f0cb6df-829f-645e-8001-b15d3a826260'}, 
            'checkpoint_id': None, 
            'checkpoint_ns': 'tools:85c45f42-227a-8a72-133e-9b2de5691c92', 
            '__pregel_scratchpad': 'PregelScratchpad(...)', 
            '__pregel_call': '<function>'
        }
    }, 
    'stream_writer': '<function>', 
    'tool_call_id': 'call_0mq7cJIOTpmBSBDIfH3RjHLi', 
    'store': None
}

I know we can manually reconfigure the input in a trace, but I think we could handle this out of the box. I see little value of having the full conversation history inside an Agent trace for every tool call the agent calls. Its also tedious to always having to filter out parameters for every tool just for langfuse. In the best case the runtime context should be saved on an agent level and the runtime is removed from agent tool calls automatically.

Additional information

No response

[hassiebp]

Thanks for sharing! Do you have a minimum repro example for us to reproduce?

[ichitaka]

from langchain.agents import create_agent
from langchain.tools import tool
from langchain_openai import ChatOpenAI
from langfuse import Langfuse
from langfuse.langchain import CallbackHandler

# Langfuse setup
client = Langfuse(
    secret_key="...",
    public_key="...",
)
callback = CallbackHandler()

# Langchain setup
prompt = "You are an helpful assistant"


@tool
def get_current_weather(runtime):
    """Get the current weather for the users location"""
    return "Its 20 Degrees Celsius"


agent = create_agent(
    model=ChatOpenAI(
        api_key="..."
    ),
    system_prompt=prompt,
    tools=[get_current_weather],
)
agent.invoke(
    {"messages": [("user", "what is the current weather like?")]},
    config={"callbacks": [callback]},
)

The trace for the tool call looks like this:
{'runtime': ToolRuntime(state={'messages': [HumanMessage(content='what is the current weather like?', additional_kwargs={}, response_metadata={}, id='d9174cc3-9c84-4be0-8fc1-65348bff50fc'), AIMessage(content='', additional_kwargs={'refusal': None}, response_metadata={'token_usage': {'completion_tokens': 15, 'prompt_tokens': 61, 'total_tokens': 76, 'completion_tokens_details': {'accepted_prediction_tokens': 0, 'audio_tokens': 0, 'reasoning_tokens': 0, 'rejected_prediction_tokens': 0}, 'prompt_tokens_details': {'audio_tokens': 0, 'cached_tokens': 0}}, 'model_provider': 'openai', 'model_name': 'gpt-3.5-turbo-0125', 'system_fingerprint': None, 'id': 'chatcmpl-CkbeEWFirgeifu2JgG5h3UDTrWk51', 'service_tier': 'default', 'finish_reason': 'tool_calls', 'logprobs': None}, id='lc_run--019aff8a-3fe9-7df2-a702-46e4b29be6ab-0', tool_calls=[{'name': 'get_current_weather', 'args': {'runtime': {}}, 'id': 'call_ldOqOuy7QWKcRcelJ12akwCJ', 'type': 'tool_call'}], usage_metadata={'input_tokens': 61, 'output_tokens': 15, 'total_tokens': 76, 'input_token_details': {'audio': 0, 'cache_read': 0}, 'output_token_details': {'audio': 0, 'reasoning': 0}})]}, context=None, config={'tags': [], 'metadata': {'langgraph_step': 2, 'langgraph_node': 'tools', 'langgraph_triggers': ('__pregel_push',), 'langgraph_path': ('__pregel_push', 0, False), 'langgraph_checkpoint_ns': 'tools:8c13797e-74d0-2e46-5bac-78420c81d1cc', 'checkpoint_ns': 'tools:8c13797e-74d0-2e46-5bac-78420c81d1cc'}, 'callbacks': <langchain_core.callbacks.manager.CallbackManager object at 0x11bcd9fd0>, 'recursion_limit': 10000, 'configurable': {'__pregel_runtime': Runtime(context=None, store=None, stream_writer=<function Pregel.stream.<locals>.stream_writer at 0x11bc3d1c0>, previous=None), '__pregel_task_id': '8c13797e-74d0-2e46-5bac-78420c81d1cc', '__pregel_send': <built-in method extend of collections.deque object at 0x11ba734c0>, '__pregel_read': functools.partial(<function local_read at 0x1197587c0>, PregelScratchpad(step=2, stop=10000, call_counter=<langgraph.pregel._algo.LazyAtomicCounter object at 0x11bc49f60>, interrupt_counter=<langgraph.pregel._algo.LazyAtomicCounter object at 0x11bc4a020>, get_null_resume=<function _scratchpad.<locals>.get_null_resume at 0x11bc3d3a0>, resume=[], subgraph_counter=<langgraph.pregel._algo.LazyAtomicCounter object at 0x11bc49f30>), {'messages': <langgraph.channels.binop.BinaryOperatorAggregate object at 0x11bc6b780>, 'jump_to': <langgraph.channels.ephemeral_value.EphemeralValue object at 0x11bc6b6c0>, 'structured_response': <langgraph.channels.last_value.LastValue object at 0x11bc6b700>, '__start__': <langgraph.channels.ephemeral_value.EphemeralValue object at 0x11bc6b7c0>, '__pregel_tasks': <langgraph.channels.topic.Topic object at 0x11b22ec00>, 'branch:to:model': <langgraph.channels.ephemeral_value.EphemeralValue object at 0x11bc6b680>, 'branch:to:tools': <langgraph.channels.ephemeral_value.EphemeralValue object at 0x11bc6b600>}, {}, PregelTaskWrites(path=('__pregel_push', 0, False), name='tools', writes=deque([]), triggers=('__pregel_push',))), '__pregel_checkpointer': None, 'checkpoint_map': {'': '1f0d4701-c866-6f8e-8001-71df0cd24e00'}, 'checkpoint_id': None, 'checkpoint_ns': 'tools:8c13797e-74d0-2e46-5bac-78420c81d1cc', '__pregel_scratchpad': PregelScratchpad(step=2, stop=10000, call_counter=<langgraph.pregel._algo.LazyAtomicCounter object at 0x11bc49f60>, interrupt_counter=<langgraph.pregel._algo.LazyAtomicCounter object at 0x11bc4a020>, get_null_resume=<function _scratchpad.<locals>.get_null_resume at 0x11bc3d3a0>, resume=[], subgraph_counter=<langgraph.pregel._algo.LazyAtomicCounter object at 0x11bc49f30>), '__pregel_call': functools.partial(<function _call at 0x119772520>, <weakref at 0x11bc88db0; to 'PregelExecutableTask' at 0x11bc75760>, retry_policy=None, futures=<weakref at 0x11bca99e0; to 'FuturesDict' at 0x11bc88be0>, schedule_task=<bound method SyncPregelLoop.accept_push of <langgraph.pregel._loop.SyncPregelLoop object at 0x11bc49c40>>, submit=<weakref at 0x11bc7a2d0; to 'BackgroundExecutor' at 0x11ba58e00>)}}, stream_writer=<function Pregel.stream.<locals>.stream_writer at 0x11bc3d1c0>, tool_call_id='call_ldOqOuy7QWKcRcelJ12akwCJ', store=None)}

I worked my way around it by having my own decorator that basically wraps the function and sets the trace inputs to everything besides runtime, so for me this is a non issue now. But you still might want to think about this, or give recommendations in the docs. Its a lot of unnecessary data sent your way.

[bourne-AJ]

Hey @ichitaka do you have any example code of that decorator? Currently seeing the same problem in our agent setup. Cheers!

[dahnny012]

Just ran into this issue as well.

@ichitaka can you share your code?