[open-swe] fix: Handle invalid tool calls and prevent AttributeError in extraction

[open-swe]

Fixes #47

Problem

When an LLM returns invalid tool calls (e.g., due to token limits or parsing errors), the system would raise an AttributeError: 'ExtractionState' object has no attribute 'tool_call_id' during validation and retry attempts.

Solution

This PR implements two key fixes:

1. In _ExtendedValidationNode._func(), added logic to:

   • Check for invalid tool calls in additional_kwargs
   • Create appropriate error ToolMessages for invalid tool calls
   • Ensure graceful handling when tool_calls is empty

2. In validate_or_repatch(), updated routing logic to:

   • Only route to 'patch' when state contains tool_call_id
   • Prevent AttributeError by routing to 'end' for states without tool_call_id

Changes

• Modified trustcall/_base.py to handle edge cases in tool call processing
• Added comprehensive test case in tests/unit_tests/test_extraction.py to validate the fix
• Ensured code passes linting and formatting checks

Test Coverage

• Added async test test_invalid_tool_call_handling() that simulates LLM responses with empty tool_calls and invalid tool call information

Impact

• Prevents AttributeError during extraction and validation
• Gracefully handles scenarios with malformed tool calls
• Improves robustness of the extraction system