fix(llm): handle tool execution exceptions in LiteLLMProvider

[m7mdhka]

Summary

Fixes a critical bug where exceptions raised during tool execution in LiteLLMProvider.complete_with_tools() would crash the entire LLM generation loop, causing loss of conversation context and preventing graceful error recovery.

Problem

The complete_with_tools() method only caught JSONDecodeError when parsing tool arguments, but did not handle exceptions raised during actual tool execution. When a tool raised an exception:

• ❌ The entire LLM generation process crashed
• ❌ Conversation context was lost
• ❌ The agent could not recover or retry
• ⚠️ Security concern: Uncaught exceptions could leak sensitive stack trace information

Solution

• Wrapped tool_executor(tool_use) calls in comprehensive try/except blocks
• Convert tool exceptions into structured ToolResult errors with error details
• Normalize unexpected return types (dict/str) from custom executors into ToolResult
• Added error logging for tool failures
• Tool failures are now communicated back to the LLM, allowing it to:
  • Retry with different parameters
  • Use alternative tools
  • Inform the user about the failure

Changes

• core/framework/llm/litellm.py: Added exception handling and result normalization in complete_with_tools()
• core/tests/test_litellm_provider.py: Added regression test test_complete_with_tools_tool_exception_is_captured
• CHANGELOG.md: Documented the fix

Testing

All 23 tests pass, including the new regression test that verifies:

• Tool exceptions are caught (no crash)
• Errors are converted to ToolResult with is_error=True
• The LLM loop continues and receives the error message
• Conversation context is preserved

uv run pytest tests/test_litellm_provider.py -v

23 passed

[m7mdhka]

#801