How can we trim messages while using create_supervisor prebuilt in langGraph?

[iamnikhildogra]

I tried using trim messages as a tool in a separate agent and then passing on to the supervisor agent but it didn't seem to work? Can anyone please help?

Code snippet-

self.supervisor = create_supervisor(
agents=[create_query_agent(auth_user_id), refiner_agent, image_gen_agent, preprocessor_agent],
model=ChatOpenAI(model="gpt-4o", openai_api_key=cnf.OPEN_AI_API_KEY),
prompt=(
generate_supervisor_prompt(auth_user_id)
),
)

Here the preprocessor agent includes the tool for trimming messages-

preprocessor_agent = create_react_agent(
model=ChatOpenAI(model="gpt-4o", openai_api_key=cnf.OPEN_AI_API_KEY),
tools=[trim_messages_wrapper],
prompt="You are a preprocessor agent. You are responsible for preprocessing the messages to the last 4000 tokens.",
name="preprocessor"
)

@tool
def trim_messages_wrapper(state: dict):
"""Trim the messages to the last 4000 tokens."""
llm = ChatOpenAI(model="gpt-4o", openai_api_key=cnf.OPEN_AI_API_KEY)
messages = state.get("messages", [])
trimmed_messages = trim_messages(
messages,
strategy="last",
token_counter=llm,
max_tokens=4000, # Adjust this value based on your needs
start_on="human",
end_on=("human", "tool"),
include_system=True
)
state["messages"] = trimmed_messages
return state