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

Latest (unpinned via pip install strands-agents)

Python Version

3.13

Operating System

Amazon Linux (Bedrock AgentCore Runtime container)

Installation Method

pip

Steps to Reproduce

1. Configure a Strands agent with BedrockModel using model ID moonshotai.kimi-k2.5
2. Register tools (e.g., web_search, output_slide)
3. Call agent.stream_async(user_message) and iterate over the stream
4. The model successfully calls web_search and receives results
5. When the model attempts a subsequent tool call (e.g., output_slide), the event loop crashes

from strands import Agent
from strands.models.bedrock import BedrockModel

model = BedrockModel(model_id="moonshotai.kimi-k2.5")
agent = Agent(model=model, tools=[web_search, output_slide])

# This crashes during tool execution
stream = agent.stream_async("Search for X and create a slide")
async for event in stream:
    print(event)

Expected Behavior

The event loop should handle the model's tool call response gracefully, even if the response format differs slightly from Claude models. At minimum, it should raise a descriptive error rather than an unhandled KeyError.

Actual Behavior

A KeyError: 'toolUse' is raised inside the event loop, crashing the entire stream:

Traceback (most recent call last):
  File ".../strands/event_loop/event_loop.py", line 192, in event_loop_cycle
    async for tool_event in tool_events:
        yield tool_event
  File ".../strands/event_loop/event_loop.py", line 535, in _handle_tool_execution
    async for event in events:
        yield event
  File ".../strands/event_loop/event_loop.py", line 269, in recurse_event_loop
    async for event in events:
        yield event
  File ".../strands/event_loop/event_loop.py", line 155, in event_loop_cycle
    async for model_event in model_events:
        if not isinstance(model_event, ModelStopReason):
            yield model_event
  File ".../strands/event_loop/event_loop.py", line 400, in _handle_model_execution
    raise e

KeyError: 'toolUse'

This is then wrapped as EventLoopException: 'toolUse'.

Additional Context

• The issue occurs specifically with Kimi K2.5 (moonshotai.kimi-k2.5) accessed via Amazon Bedrock. Claude models work fine.
• The first tool call (web_search) succeeds and returns results. The crash happens on a subsequent tool call cycle.
• It appears the model returns a content block that Strands expects to contain a toolUse key, but the key is missing or the structure differs from what the event loop parser expects.
• We observed this in production via CloudWatch Logs on Bedrock AgentCore Runtime (OTEL traces confirmed the sequence: successful web_search -> crash on next tool call).
• As a workaround, we wrap the stream iteration in a try/except to catch the EventLoopException and return a user-friendly error.

Possible Solution

The event loop could add a safety check (e.g., content.get("toolUse") instead of content["toolUse"]) or validate the content block structure before accessing the key. This would make Strands more resilient to non-Claude models that may produce slightly different response formats.

Related Issues

• [BUG] LiteLLM Moonshot AI Kimi K2 thinking fails on multi-turn conversations with 'Invalid part type: thinking' #1150 (Kimi K2 Thinking fails on multi-turn conversations via LiteLLM)