InjectedToolArg works with OpenAI gpt-4o-mini but not longer with gpt-4o

[lorerave85]

Checked other resources

• I added a very descriptive title to this question.
• I searched the LangChain documentation with the integrated search.
• I used the GitHub search to find a similar question and didn't find it.

Commit to Help

• I commit to help with one of those options 👆

Example Code

from dotenv import load_dotenv
from langchain import hub
from langchain.agents import AgentExecutor, create_openai_tools_agent
from langchain.tools.render import render_text_description_and_args
from langchain_core.tools import InjectedToolArg, tool
from langchain_openai import ChatOpenAI
from typing_extensions import Annotated
import os
import requests

# load local parameters
load_dotenv()

# method to invoke an external api
def post_to_api(url, payload):
    """
    Makes a POST request to an API with the specified body.

    :param url: The API endpoint.
    :param payload: The payload to send in the request body.
    :return: The API response.
    """
    # HTTP headers (optional, configure if necessary)
    headers = {
        "Content-Type": "application/json"
    }

    try:
        # Make POST request
        response = requests.post(url, json=payload, headers=headers)

        # Check if the request was successful
        response.raise_for_status()
        
        # Returns the response in JSON format
        return response.json()  
    except requests.exceptions.RequestException as e:
        # Error Handling
        print(f"Errore durante la richiesta POST: {e}")
        return None
    
# tool definition
@tool(parse_docstring=True)
def exec_device_user(user_id: Annotated[str, InjectedToolArg]) -> str:
  
    """Exec query on DB to retrive information about user's device

    Args:
        user_id: User's ID.
    """
    
    print(f"my user {user_id}")
    
    api_url = "http://localhost:5000/execute_query"  # Sostituisci con l'URL reale dell'API
    payload = {
        "fields": ["device_integration_type","model_id"],
        "distinct": True,
        "table": "user_metrics",
        "filter": "user_id = %s",
        "params": [user_id]
    }

    return post_to_api(api_url, payload)


# LLM Agent definition
prompt = hub.pull("hwchase17/react")
tools = [exec_device_user]
prompt = prompt.partial(tools=render_text_description_and_args(list(tools)))
prompt = prompt.partial(tool_names=", ".join([t.name for t in tools]))

llm = ChatOpenAI(
    model_name="gpt-4o",
    temperature=os.getenv('OPENAI_TEMPERATURE'),
    max_tokens=os.getenv('OPENAI_MAX_TOKENS'),
)

agent = create_openai_tools_agent(
    llm=llm,
    tools=tools,
    prompt=prompt
)

agent_executor = AgentExecutor(
    agent=agent,
    tools=tools,
    max_iterations=os.getenv('LANGCHAIN_MAX_ITERATIONS'),
)

res = agent_executor.invoke(input={"input":[{"content":"which is my bucket?","type":"human","user_id":"1"}]})

print(res)

Description

for a few weeks I have had a strange behavior of my agent.
I have reported a simplified version that shows the behavior.

my agent executes an API call passing a parameter that arrives in input on the invoke.

if I use the OpenAI gpt-4o-mini model the parameter is read (and printed), while if I use the gpt-4o model (which has always worked) now it shows me a validation error:

pydantic_core._pydantic_core.ValidationError: 1 validation error for exec_device_user
user_id

as if it lost the parameter.

System Info

langchain==0.3.19
langchain-community==0.3.17
langchain-core==0.3.35
langchain-openai==0.3.6

platform mac, Python 3.12.8

[lorerave85]

I add a detail: I'm debugging and I found this behavior:

in the agent.py file in the AgetExecution class there is this method _iter_next_step:

this step:

# Call the LLM to see what to do.
output = self._action_agent.plan(
intermediate_steps,
callbacks=run_manager.get_child() if run_manager else None,
**inputs,
)

in the case of 4o-mini the output is:
[ToolAgentAction(tool='exec_device_user', tool_input={'user_id': '1'},...

in the case of 4o the output is:
[ToolAgentAction(tool='exec_device_user', tool_input={},...