Proper way of managing large context window for ComputerTool

[huntersgordon]

I am enjoying using ComputerTool, but want to create a persistent agent that lasts for many generations, for example, max_turns > 10000

I run into the basic error:
Error getting response: Error code: 400 - {'error': {'message': 'This request exceeds the context size limit. Please reduce the size of the prompt and try again.', 'type': 'invalid_request_error', 'param': None, 'code': None}}

• Are there builtin functions that can help me prune this?

• I understand this comes from the openAI core API, but can someone please direct me to where I can manage this? What are some good approaches to managing context/input as the size increases for a persistent agent?

[rm-openai]

There are a few possibilities here, and I am not sure which one you are hitting. Would you be able to share some sample code?

I am basically wondering if its:

1. Too many history items being sent to the model (in which case you can just trim the input between calls to Runner.run()
2. Too much data in the request (which could be fixed by using response or reference IDs)

I'm guessing its the former, but let me know!

[huntersgordon]

@rm-openai

sure, I just have my basic Agent instantiation here, followed by an arbitrarily large max_turns

agent = Agent(
              ...,
                tools=[ComputerTool(computer)],
                model="computer-use-preview",
                model_settings=ModelSettings(truncation="auto"),
            )

await Runner.run(agent, prompt, max_turns=10000, hooks=hooks)

Now, I made a modification to run.pyto truncate the input, along the lines of:

run.py -> _run_single_turn()

//if context too long:

pruned_input = input[len(input) // 2:]
print(f"Length of input after pruning: {len(pruned_input)}")

# Retry with pruned input
new_response = await cls._get_new_response(
    agent,
    system_prompt,
    pruned_input,
    output_schema,
    handoffs,
    context_wrapper,
    run_config,
)

and get:

Error getting response: Error code: 400 - {'error': {'message': "Item 'fc_67d21f40a2e88191b3b06fea5e92c66a07505fdb833b0cf2' of type 'function_call' was provided without its required 'reasoning' item: 'rs_67d21f3ce3fc8191a98c966d5161440407505fdb833b0cf2'.", 'type': 'invalid_request_error', 'param': 'input', 'code': None}}

Thanks for your feedback.

[rm-openai]

@huntersgordon do you happen to have a request ID so we can debug this? #114 adds the request ID to the error log

(specifically want to debug the original issue with This request exceeds the context size limit)

[huntersgordon]

openai.BadRequestError: Error code: 400 - {'error': {'message': 'This request exceeds the context size limit. Please reduce the size of the prompt and try again.', 'type': 'invalid_request_error', 'param': None, 'code': None}}

Hi @rm-openai , not seeing any request ID here. What should it be under?

[rm-openai]

@huntersgordon It should have printed a bunch of text including a stack trace - at the very top of the stack trace it should say something like

Error getting response: <error> (request_id: ...)

(I'm assuming you have cloned the repo and hence have #114 included)

[rm-openai]

If you havent cloned, just uninstall and reinstall the SDK with

pip install git+ssh://git@github.com/openai/openai-agents-python.git

[github-actions]

This issue is stale because it has been open for 7 days with no activity.

[maininformer]

@rm-openai

sure, I just have my basic Agent instantiation here, followed by an arbitrarily large max_turns

agent = Agent(
              ...,
                tools=[ComputerTool(computer)],
                model="computer-use-preview",
                model_settings=ModelSettings(truncation="auto"),
            )

await Runner.run(agent, prompt, max_turns=10000, hooks=hooks)

Now, I made a modification to run.pyto truncate the input, along the lines of:

run.py -> _run_single_turn()

//if context too long:

pruned_input = input[len(input) // 2:]
print(f"Length of input after pruning: {len(pruned_input)}")

# Retry with pruned input
new_response = await cls._get_new_response(
    agent,
    system_prompt,
    pruned_input,
    output_schema,
    handoffs,
    context_wrapper,
    run_config,
)

and get:

Error getting response: Error code: 400 - {'error': {'message': "Item 'fc_67d21f40a2e88191b3b06fea5e92c66a07505fdb833b0cf2' of type 'function_call' was provided without its required 'reasoning' item: 'rs_67d21f3ce3fc8191a98c966d5161440407505fdb833b0cf2'.", 'type': 'invalid_request_error', 'param': 'input', 'code': None}}

Thanks for your feedback.

Having the same issue as this second one here