A local code index for coding agents that resolves imports without a language server (Rust, MIT, runs offline)

Last time I posted about this here someone asked the right question: why not just use rust-analyzer, and how do you build a call graph without real semantics? I did not have a good answer then. I do now, and it comes with a concession. The concession first. This is not LSP-precision. A language server has a type checker; for generics, macros, dynamic dispatch, DI containers and monkey-patching it will be right where this is wrong. What you get instead is no build step, no language server, and no per-language daemon, which matters when the repo does not compile because an agent is halfway through editing it. How the resolution actually works, split by tier, because "300+ languages" is a parsing number and not an understanding number: JavaScript and TypeScript go through oxc. Real scope tree, symbol table, import/export resolution, intra- and cross-file, including tsconfig path aliases in a monorepo. Python and Java go through stack-graphs, GitHub's declarative name-binding rules that run against a tree-sitter parse tree. GitHub archived that project in September 2025. It is forked and maintained here, with four upstream rule bugs fixed that used to abort a whole file's resolution silently on things like a typed splat parameter or a chained assignment. Everything else gets intra-file lexical scope binding only. 371 grammars parse. About 100 of those give up symbols and calls. Three get real resolution. The part that matters for agents is that it refuses to over-claim. Asking for a function's callers returns every call site a name scan finds, annotated with which ones resolution could prove, rather than only the proven subset. That is deliberate: an earlier version filtered to proven hits and reported 2 callers on a file that had 172, with no truncation flag. An agent has no instinct to distrust a small number. Runs fully offline, no API keys, no telemetry, index lives in a local cache. Scanning the TypeScript compiler, 81k files, takes about 18 seconds on an M4; git history queries land in tens of microseconds regardless of history depth. Rust, MIT: github.com/Goldziher/basemind Happy to be argued with on the tiering, particularly by anyone who has tried stack-graphs and given up on it. Three other agent projects evaluated it and walked away, so I am aware I may be the one who is wrong.