Worked on the BerriAI/litellm repository to enhance the accuracy of Ollama model response handling by addressing a critical bug related to completion reasons. Focused on backend and API development using Python, the work ensured that when responses were truncated due to max_tokens, the actual done_reason of 'length' was correctly propagated as the finish_reason across both streaming and non-streaming paths. This replaced previous hardcoded values, allowing downstream adapters to accurately interpret truncation events. The solution included updating transformation logic and unit tests, and involved cross-team collaboration to maintain consistent behavior and reliable response interpretation throughout the system.