I've been working on a tool that uses tree-sitter grammars to extract structural entities (functions, classes, methods) from source code, then builds a cross-file dependency graph by resolving references between them.
The core problem: traditional diff tools compare lines, but the meaningful unit of change in code is an entity. When you rename a function, move a method, or reformat a file, line-level diff produces noise. Entity-level diff tells you "this function was modified, this one was added, this one moved."
The interesting technical bits:
- Each language gets a config that maps AST node types to entity types (e.g. function_definition in Python, function_item in Rust, method_declaration in Java). Currently supports 25+ languages through tree-sitter.
- Scope resolution walks the AST to resolve which entity references which other entity, handling class scopes, impl blocks, function parameters, and assignment-based type tracking.…