[BUG]: TypeError in Bedrock instrumentation when tool_input is dict instead of JSON string

[johaywood]

Tracer Version(s)

4.3.1

Python Version(s)

3.14

Pip Version(s)

uv

Bug Report

The ddtrace Bedrock instrumentation crashes with a TypeError when processing tool use responses from AWS Bedrock. The issue occurs in ddtrace/llmobs/_integrations/utils.py in the get_final_message_converse_stream_message function.

The function attempts to parse tool_input with json.loads(), but AWS Bedrock sometimes returns tool inputs as already-parsed dict objects rather than JSON strings. This causes the instrumentation to fail with:

TypeError: the JSON object must be str, bytes or bytearray, not dict

Expected Behavior

The instrumentation should handle tool inputs regardless of whether they are:

• JSON strings (need parsing with json.loads())
• Already-parsed dict objects (use directly)

Actual Behavior

The instrumentation crashes with TypeError when tool_input is a dict, breaking LLM observability and potentially disrupting the application flow.

Root Cause

In ddtrace/llmobs/_integrations/utils.py, the get_final_message_converse_stream_message function assumes tool_input is always a JSON string:

# Problematic code (approximate line 1448)
tool_input = tool_block.get("input")
tool_args = json.loads(tool_input)  # Fails when tool_input is already a dict

Proposed Fix

Add type checking before attempting to parse:

tool_input = tool_block.get("input")
tool_args = {}

if isinstance(tool_input, dict):
    # Already parsed - use directly
    tool_args = tool_input
elif isinstance(tool_input, str):
    # JSON string - parse it
    try:
        tool_args = json.loads(tool_input)
    except (json.JSONDecodeError, ValueError):
        # Invalid JSON - wrap in dict
        tool_args = {"input": tool_input}
else:
    # Unexpected type - convert to string
    tool_args = {"input": str(tool_input)}

Workaround

We've implemented a monkey-patch that wraps the problematic function with the type checking above. This is a temporary solution until the upstream fix is available.

Reproduction Code

1. Set up AWS Bedrock with Claude model (Sonnet 4.5)
2. Enable ddtrace LLM observability:

   from ddtrace.llmobs import LLMObs
   
   LLMObs.enable(
       ml_app="my_app",
       api_key=os.getenv("DD_API_KEY"),
       site="datadoghq.com",
       agentless_enabled=True,
       integrations_enabled=True
   )

3. Create an agent with tools using the Bedrock Converse API
4. Invoke the agent with a prompt that triggers tool use
5. Observe the TypeError when the agent attempts to use a tool

Error Logs

Traceback (most recent call last):
  File "/usr/local/lib/python3.14/site-packages/strands/models/bedrock.py", line 770, in _stream
    for chunk in response["stream"]:
  File "/usr/local/lib/python3.14/site-packages/ddtrace/llmobs/_integrations/base_stream_handler.py", line 142, in __iter__
    self._self_handler.process_chunk(chunk, self._self_stream_iter)
  File "/usr/local/lib/python3.14/site-packages/ddtrace/contrib/internal/botocore/services/bedrock.py", line 92, in process_chunk
    stream_processor.send(chunk)
  File "/usr/local/lib/python3.14/site-packages/ddtrace/llmobs/_integrations/bedrock.py", line 325, in _converse_output_stream_processor
    get_final_message_converse_stream_message(current_message, text_content_blocks, tool_content_blocks)
  File "/usr/local/lib/python3.14/site-packages/ddtrace/llmobs/_integrations/utils.py", line 1448, in get_final_message_converse_stream_message
    tool_args = json.loads(tool_input)
                ^^^^^^^^^^^^^^^^^^^^^^
  File "/usr/local/lib/python3.14/json/__init__.py", line 345, in loads
    raise TypeError(f'the JSON object must be str, bytes or bytearray, '
                    f'not {s.__class__.__name__}')
TypeError: the JSON object must be str, bytes or bytearray, not dict

Libraries in Use

No response

Operating System

No response