[BUG] RecursionError + "input length and `max_tokens` exceed context limit: 142422 + 64000 > 204658"

[gargshilpi22]

Checks

• I have updated to the lastest minor and patch version of Strands
• I have checked the documentation and this is not expected behavior
• I have searched ./issues and there are no duplicates of my issue

Strands Version

0.1.0

Python Version

3.10

Operating System

Amazon Linux 2

Installation Method

pip

Steps to Reproduce

RecursionError: maximum recursion depth exceeded

Error Description

When running the Strands agent, I'm encountering a context window overflow error followed by a recursion error.

Error Details

The primary error is a ContextWindowOverflowException with the following details:

input length and `max_tokens` exceed context limit: 142422 + 64000 > 204658

This indicates that the total tokens (input + max_tokens) exceeds the model's context window limit of 204,658 tokens.

Full Error Trace

Traceback (most recent call last):
  # ... [previous trace details]
  
strands.types.exceptions.ContextWindowOverflowException: An error occurred (validationException) when calling the ConverseStream operation: The model returned the following errors: input length and `max_tokens` exceed context limit: 142422 + 64000 > 204658, decrease input length or `max_tokens` and try again

This is followed by a recursion error while handling the initial exception:

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "XXX/.venv/lib/python3.10/site-packages/strands/event_loop/event_loop.py", line 220, in event_loop_cycle
    return _handle_tool_execution(
  File "XXX/.venv/lib/python3.10/site-packages/strands/event_loop/event_loop.py", line 428, in _handle_tool_execution
    return recurse_event_loop(
  File "XXX/.venv/lib/python3.10/site-packages/strands/event_loop/event_loop.py", line 305, in recurse_event_loop
    ) = event_loop_cycle(**kwargs)
  File "XXX/.venv/lib/python3.10/site-packages/strands/event_loop/event_loop.py", line 163, in event_loop_cycle
    return handle_input_too_long_error(
  File "XXX/.venv/lib/python3.10/site-packages/strands/event_loop/error_handler.py", line 108, in handle_input_too_long_error

  File "XXX/.venv/lib/python3.10/site-packages/botocore/validate.py", line 213, in _validate
    getattr(self, f'_validate_{shape.type_name}')(
  File "XXX/.venv/lib/python3.10/site-packages/botocore/validate.py", line 54, in _on_passes_type_check
    return func(self, param, shape, errors, name)
  File "XXX/.venv/lib/python3.10/site-packages/botocore/validate.py", line 282, in _validate_structure
    self._validate(
  File "XXX/.venv/lib/python3.10/site-packages/botocore/validate.py", line 209, in _validate
    special_validator = self._check_special_validation_cases(shape)
  File "XXX/.venv/lib/python3.10/site-packages/botocore/validate.py", line 203, in _check_special_validation_cases
    if is_json_value_header(shape):
RecursionError: maximum recursion depth exceeded

Environment Details

• Python version: 3.10
• Package: Strands
• Running in virtual environment

Question

Has anyone encountered this issue and found a solution? I'm particularly interested in best practices for handling large inputs with the Strands agent.

---

Please let me know if you need any additional information or clarification.

Expected Behavior

Recursion length should be handled.

Actual Behavior

Failed due to - RecursionError: maximum recursion depth exceeded

Additional Context

No response

Possible Solution

No response

Related Issues

No response

[Unshure]

Hi @gargshilpi22, thanks for reporting this bug! Taking a look into it!

[Unshure]

I saw the pull request that you created for this issue, and I wanted to point you to the logic in our Agent class where we are calling reduce_context on the conversation_manager: https://github.com/strands-agents/sdk-python/blob/main/src/strands/agent/agent.py#L506-L509

The reduce_context function is called whenever a model throws a ContextWindowOverflowException, and this should ideally trim down the conversation to avoid this issue. It seems like there might be a bug in this logic. Can you share the state of the agent when you run into this issue? The system_prompt, messages, and tools will help me investigate and reproduce this issue myself.

[ForeverZyh]

Hi @Unshure , to my understanding, the issue is in event_loop_cycle, the ContextWindowOverflowException cannot be handled by handle_input_too_long_error, but was recursively tried in event_loop_cycle. So that the ContextWindowOverflowException will never thrown away from event_loop_cycle, therefore the reduce_context you pointed will not be called because RecursionError will be thrown instead of ContextWindowOverflowException.

[Unshure]

Ahh, thanks for the clarification! Yes, this is definitely a bug in our handle_input_too_long_error method.

I think what is happening is that during a long conversation, the agent's conversation fills up to the point where its right on the edge of the model's conversation context limit. At this point, a toolUse is requested by the model, and the toolResult is added the conversation history which pushes it over the limit. The handle_input_too_long_error tries to truncate the toolResult content, but even just the metadata of the toolUse and toolResult were enough to keep it over the limit. So then the event loop gets stuck trying to truncate toolResult content that is already truncated, and always getting ContextWindowOverflowExceptions.

I just added a unit test to my fork to replicate the issue: Unshure@a0d8a42

As a fix for this issue, we plan to:

1. Move the toolResult truncation logic out of the event_loop_cycle function, and instead have it live in the SlidingWindowConversationManager.
2. Make this truncation logic toggle-able (to address this issue: [BUG] Tool response truncation occurs silently #167)
3. Don't try to truncate when the toolResult text is already The tool result was too large!

[poshinchen]

Hi, we have merged the fix and will release it in the next PyPI release.