The two-step locate/expand pattern is smart — most agent loops I've debugged waste tokens because they pull full file contents way too early. Curious how well the tree-sitter symbol extraction holds up on messier codebases with lots of metaprogramming or dynamic imports (thinking Django or SQLAlchemy-style magic). That's usually where static analysis tools start returning noise. The PyTorch benchmark is impressive but it's also a relatively well-structured codebase would love to see numbers on something gnarlier. Have you tested it on monorepos with mixed languages?