We just reworked the diagram generation logic using tree-sitter AST parsing. It's much faster and should not exceed context size of the best models when handling bigger repositories. Take a look and let me know if it works for you!