results field in Code Interpreter always returns None

[Ladun]

Please read this first

• Have you read the docs?Agents SDK docs
• Have you searched for related issues? Others may have had similar requests

Question

I'm trying to access the result values from a code interpreter used by an agent.
Using the following code, I attempted to retrieve and log the executed code and its results after running Runner.run.

result = await Runner.run(agent, natural_language_query)

for new_item in result.new_items:
    if (new_item.type == "tool_call_item" 
        and new_item.raw_item.type == "code_interpreter_call"):
        logger.info(f"Code interpreter ======================================= ")
        logger.info(f"code:\n```\n{new_item.raw_item.code}\n```\n")
        logger.info(f"Results: {new_item.raw_item}")
        if new_item.raw_item.results is not None:
            logger.info(f"results:")
            for idx, r in enumerate(new_item.raw_item.results):
                logger.info(f"Result {idx}")
                logger.info(r)

From the traces, I can confirm that the code interpreter runs the code and produces output as expected.
However, the value of new_item.raw_item.results is always None.

I’m not sure if:

• This is the intended behavior,
• This is a bug where the results are not being set, or
• There is a separate step I need to take to retrieve the results properly.

I would appreciate any clarification on this.
Thank you!

[rm-openai]

Can you try this in the OpenAI playground or directly via a curl to openai, and let me know if you still see the same?

[seratch]

Indeed, the tool call item does not contain any values, while the subsequent LLM call properly uses the tool's output.

DEBUG:openai.agents:LLM resp:
[
  {
    "id": "ci_685a62ade044819c859d99ca53c332fe07fbf10a2fb463eb",
    "code": "import math\n\n# Calculate the expression\nresult = math.sqrt(273 * 312821 + 1782)\nresult",
    "results": null,
    "status": "completed",
    "type": "code_interpreter_call",
    "container_id": "cntr_685a62ad87988191834a261b3cba3ac6074320d957a966d3",
    "outputs": null
  },
  {
    "id": "msg_685a62b5a274819c86c0768e689fbb8307fbf10a2fb463eb",
    "content": [
      {
        "annotations": [],
        "text": "The square root of \\(273 \\times 312821 + 1782\\) is approximately \\(9241.32\\).",
        "type": "output_text",
        "logprobs": null
      }
    ],
    "role": "assistant",
    "status": "completed",
    "type": "message"
  }
]

I know this isn’t an ideal workaround, but if you need the tool’s output outside the agent, you might find it helpful to ask the agent to save a file in the container and then fetch the file’s content later on.

import asyncio
from openai import OpenAI
from agents import Agent, CodeInterpreterTool, Runner, trace

# import logging
# logging.basicConfig(level=logging.DEBUG)

async def main():
    agent = Agent(
        name="Code interpreter",
        instructions="You love doing math.",
        tools=[CodeInterpreterTool(tool_config={"type": "code_interpreter", "container": {"type": "auto"}})],
    )
    with trace("Code interpreter example"):
        result = await Runner.run(
            agent,
            "What is the square root of273 * 312821 plus 1782? When you use Python tool, ask the tool to save only its final output in a text file, then return the final outuput.",
        )
        for new_item in result.new_items:
            if (
                new_item.type == "tool_call_item"
                and new_item.raw_item.type == "code_interpreter_call"
            ):
                print(f"code:\n```\n{new_item.raw_item.code}\n```\n")
                openai_client = OpenAI()
                files = openai_client.containers.files.list(
                    container_id=new_item.raw_item.container_id,
                )
                for file in files:
                    file_content = openai_client.containers.files.content.retrieve(
                        container_id=new_item.raw_item.container_id,
                        file_id=file.id,
                    )
                    print(file_content.text)

        print(f"Final output: {result.final_output}")

if __name__ == "__main__":
    asyncio.run(main())

With this prompt, the generated Python code looks like this:

import math

# Given numbers
num1 = 273
num2 = 312821
num3 = 1782

# Calculate the expression
result = math.sqrt(num1 * num2 + num3)

# Save the result to a text file
output_file_path = '/mnt/data/result.txt'
with open(output_file_path, 'w') as file:
    file.write(str(result))

result

[Ladun]

Thank you very much for your prompt response!
I had been considering various ways to utilize the output generated by the agent’s code, and the method you suggested seems to offer a clear and effective solution.

[Ladun]

I’ve confirmed that the method you suggested works well in code — the file is correctly saved and read as expected.
However, when I applied it to my own code, although the code interpreter clearly executed the file-saving logic, the file does not appear in the container’s file list.
For reference, I didn’t let the container be created automatically; instead, I pre-created the container and uploaded a specific file beforehand to proceed.
To clarify, the container is created in advance, and the agent runs with one file already uploaded to it.

Here is the code I used to create the container:

async def _setup_openai_container_and_files(data_filename: str) -> str:
    
    target_container = None    
    raws = client.containers.list(limit=100)
    if raws.data is not None:
        for container in raws.data:
            if container.status == 'running':
                target_container = container
    
    if target_container == None:
        target_container = client.containers.create(
            name="Test container",
            expires_after=ExpiresAfter(
                anchor='last_active_at',
                minutes=20
            )
        )
        
    if target_container == None:
        logger.error("Create container error")
        return None
    
    target_file = None
    raws = client.containers.files.list(target_container.id)
    logger.info(f"Current container file list")
    for file in raws.data:
        logger.info(f"File: {file}")
        file_name = "-".join(file.path.split("-")[1:])
        if file_name in data_filename:
            target_file = file
            
    if target_file is None:
        logger.info(f"Create file '{data_filename}'")
        client.containers.files.create(
            container_id=target_container.id,
            file=open(data_filename, 'rb'),
        )    
      
    print(f"{target_container=}")
    return target_container.id

Here is the code that was executed by the agent:

result = {
    ...
}
json_path = '/mnt/data/result.json'
with open(json_path, 'w') as f:
    json.dump(result, f, indent=4)
json_path