Worked on backend development and API integration for the safety-research/safety-tooling repository, focusing on improving the reliability of the AnthropicChatModel. Addressed a bug by ensuring the max_tokens parameter was persisted across API retries, which involved moving the token pop operation outside the retry loop and reusing the value during subsequent attempts. This approach prevented token resets, resulting in more consistent API behavior and reduced response variability during chat interactions. The solution was implemented in Python and demonstrated careful attention to traceability by referencing the specific commit, supporting future maintenance and contributing to a more stable user experience in chat applications.