fix: set StatusCode.ERROR on tool spans when tool result status is error

[nicolas-bezdolya]

Problem

When a tool raises an exception, Strands catches it in _stream() and converts it to a ToolResult dict with status='error'. The original exception object is lost before end_tool_call_span is called, so _end_span() always falls into the else branch and marks the span as StatusCode.OK.

The gen_ai.tool.status: "error" attribute IS correctly set, but the span StatusCode remains OK. This is critical because observability backends like Langfuse rely on StatusCode.ERROR to raise alarms/statistics on failed tool calls.

Solution

In end_tool_call_span(), before calling _end_span(), check if tool_result["status"] == "error" and synthesize an Exception from the error content. This ensures the span correctly gets StatusCode.ERROR and record_exception() is called.

Testing

• All 76 tracer tests pass ✅
• All 30 executor tests pass ✅

Fixes #2016

[nicolas-bezdolya]

Fixed.

[github-actions]

Issue: Missing test coverage for this new error-handling logic.

Suggestion: Add unit tests to verify:

1. end_tool_call_span sets StatusCode.ERROR when tool_result["status"] == "error"
2. The error message is correctly extracted from content[0]["text"]
3. The default message "Tool call failed" is used when content is empty

Example test structure:

def test_end_tool_call_span_with_error_status(mock_span):
    """Test that tool spans get StatusCode.ERROR when tool result status is error."""
    tracer = Tracer()
    tool_result = {
        "status": "error",
        "content": [{"text": "Error: ValueError - something failed"}],
        "toolUseId": "test-id"
    }
    
    tracer.end_tool_call_span(mock_span, tool_result)
    
    mock_span.set_status.assert_called_once_with(
        StatusCode.ERROR, 
        "Error: ValueError - something failed"
    )
    mock_span.record_exception.assert_called_once()


def test_end_tool_call_span_with_error_status_empty_content(mock_span):
    """Test default error message when content is empty."""
    tracer = Tracer()
    tool_result = {"status": "error", "content": [], "toolUseId": "test-id"}
    
    tracer.end_tool_call_span(mock_span, tool_result)
    
    mock_span.set_status.assert_called_once_with(StatusCode.ERROR, "Tool call failed")

[github-actions]

Assessment: Request Changes

This PR correctly addresses the issue where tool spans receive StatusCode.OK even when the tool fails. The fix properly synthesizes an exception from the tool_result["status"] == "error" condition so that _end_span sets the appropriate error status.

Review Details

• Testing: The PR is missing unit tests for this new error-synthesis logic. Tests should verify the StatusCode.ERROR is set and exception is recorded when tool result status is "error".
• Code Quality: The implementation is clean, minimal, and follows existing patterns.
• Documentation: Not required for this bug fix.

The fix itself looks correct - please add tests to verify the behavior before merging.

[mattdai01]

instead of fixing it here, shouldn't we fix it downstream so that the exception is propagated up the chain? Otherwise, we are doing exception -> text in the tool, and then text -> exception here.