Graft builds structural codebase graphs for agents

Vector search is fundamentally the wrong primitive for codebase navigation, and deterministic AST graphs prove that structure beats semantic similarity. AST-level parsing extracts exact callers, dependencies, and type hierarchies locally without burning tokens on embedding models or requiring vector databases. Graph-first orientation replaces blind grep loops, cutting agent latency by 60% and reducing tool calls by 46% in benchmark runs. Committing plain-English structural nodes into Git allows agents like Claude Code, Cursor, and Codex to ingest context via files or MCP. Handling highly dynamic codebases and runtime metaprogramming remains an inherent challenge where static syntax trees cannot trace dynamic calls.