[BUG] SlidingWindowConversationManager incorrectly returning the error "The tool result was too large" when the window size is less than number of messages.

[joychak]

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

1.5.0

Python Version

3.13.3

Operating System

macOS 15.6

Installation Method

pip

Steps to Reproduce

from strands import Agent, tool
from strands_tools import calculator
from strands.agent.conversation_manager import SlidingWindowConversationManager

conversation_manager = SlidingWindowConversationManager(
    window_size=5,  # Maximum number of message pairs to keep
)

agent = Agent(
    conversation_manager=conversation_manager,
    tools=[calculator],
    messages=[
        {"role": "user", "content": [{"text": "Hello, my name is Strands!"}]},
        {"role": "assistant", "content": [{"text": "Hi there! How can I help you today?"}]}
    ]
)
agent("What's my name?")
print('\n===================================================')

response = agent.tool.calculator(expression="123 * 456")

# Direct tool call without recording
agent.tool.calculator(expression="765 / 987", record_direct_tool_call=False)

[print(msg) for msg in agent.messages]
  

Expected Behavior

We expect the agent to return the last 5 messages including the successful tool result.

Your name is Strands! Nice to meet you, Strands. Is there anything I can help you with today, perhaps some mathematical calculations or problem-solving?
===================================================
'role': 'assistant', 'content': [{'text': 'Your name is Strands! Nice to meet you, Strands. Is there anything I can help you with today?'}]}
{'role': 'user', 'content': [{'text': 'agent.tool.calculator direct tool call.\nInput parameters: {"expression": "123 * 456"}\n'}]}
{'role': 'assistant', 'content': [{'toolUse': {'toolUseId': 'tooluse_calculator_533712149', 'name': 'calculator', 'input': {'expression': '123 * 456'}}}]}
{'role': 'user', 'content': [{'toolResult': {'status': 'success', 'content': [{'text': 'Result: 56088'}], 'toolUseId': 'tooluse_calculator_533712149'}}]}
{'role': 'assistant', 'content': [{'text': 'agent.tool.calculator was called.'}]}

Actual Behavior

It gives a "The tool result was too large" error for the multiplication operation for the calculator tool.

Your name is Strands! Nice to meet you, Strands. Is there anything I can help you with today, perhaps some mathematical calculations or problem-solving?
===================================================
{'role': 'assistant', 'content': [{'text': 'Your name is Strands! Nice to meet you, Strands. Is there anything I can help you with today, perhaps some mathematical calculations or problem-solving?'}]}
{'role': 'user', 'content': [{'text': 'agent.tool.calculator direct tool call.\nInput parameters: {"expression": "123 * 456"}\n'}]}
{'role': 'assistant', 'content': [{'toolUse': {'toolUseId': 'tooluse_calculator_237287153', 'name': 'calculator', 'input': {'expression': '123 * 456'}}}]}
{'role': 'user', 'content': [{'toolResult': {'status': 'error', 'content': [{'text': 'The tool result was too large!'}], 'toolUseId': 'tooluse_calculator_237287153'}}]}
{'role': 'assistant', 'content': [{'text': 'agent.tool.calculator was called.'}]}

Additional Context

If you set the window_size big enough to accomodate all messages-

conversation_manager = SlidingWindowConversationManager(
    window_size=10,  # Maximum number of message pairs to keep
)

Then, it does produce the correct following output -

Your name is Strands! Nice to meet you, Strands. Is there anything I can help you with today, perhaps some mathematical calculations or problem-solving?
===================================================
{'role': 'user', 'content': [{'text': 'Hello, my name is Strands!'}]}
{'role': 'assistant', 'content': [{'text': 'Hi there! How can I help you today?'}]}
{'role': 'user', 'content': [{'text': "What's my name?"}]}
{'role': 'assistant', 'content': [{'text': 'Your name is Strands! Nice to meet you, Strands. Is there anything I can help you with today?'}]}
{'role': 'user', 'content': [{'text': 'agent.tool.calculator direct tool call.\nInput parameters: {"expression": "123 * 456"}\n'}]}
{'role': 'assistant', 'content': [{'toolUse': {'toolUseId': 'tooluse_calculator_533712149', 'name': 'calculator', 'input': {'expression': '123 * 456'}}}]}
{'role': 'user', 'content': [{'toolResult': {'status': 'success', 'content': [{'text': 'Result: 56088'}], 'toolUseId': 'tooluse_calculator_533712149'}}]}
{'role': 'assistant', 'content': [{'text': 'agent.tool.calculator was called.'}]}

Possible Solution

No response

Related Issues

No response

[pgrayy]

Hey @joychak, thank you for sharing. So right now this is expected behavior. Users can toggle it on and off with the should_truncate_results flag (docs). It may be worth revisiting this logic to see if it is actually beneficial.

[giulio-leone]

Fix submitted in #1837.

Root cause: apply_management() delegates to reduce_context() which has a truncation-first strategy (designed for model context overflow recovery). When called for routine window management, this incorrectly truncates tool results inside the kept window.

Fix: apply_management() now calls _slide_window() directly to remove old messages. reduce_context() retains its truncation-first strategy for context overflow recovery.