Fixes#5808
- Add supports_assistant_prefill() to BaseLLM, LLM (litellm), and
AnthropicCompletion
- Claude 4.6+ models reject assistant-role messages as the last message
in a conversation; split observation into a user-role message instead
- Drop temperature parameter for Claude 4.6+ (not supported)
- Update handle_max_iterations_exceeded to use user-role message for
no-prefill models
- Use getattr() for backward compatibility with custom LLM adapters
- Add 24 tests covering prefill detection, temperature dropping, message
splitting, and backward compatibility
Co-Authored-By: João <joao@crewai.com>
449bb0d61e