GitHub · by Ellipsis

blink

Codebase search by sending a hundred walkers through the file tree.

Ellipsis built a code search that takes a plain-language question and a directory, then releases an ensemble of walkers that walk the file system looking for the answer. Each walker asks Jev which child to descend into. Where they end up, and how many of them end up there, is the ranking.

There is no index, no embedding store and no chunking step. A run on a small tree with a hundred walkers and a question like “where is authentication handled?” returns a table of files with the share of walkers that arrived at each: in the repository’s own example, 74% at `src/services/auth/login.ts`, 16% at `src/services/auth/session.ts` and 10% at a file in the UI folder. The percentage is a distribution over a population, not a similarity score, so a split result reads as a genuinely ambiguous question rather than a confident wrong answer.

This works because the decision is cheap. Walking a tree means one decision per directory per walker, which is thousands of decisions for a single query — the kind of arithmetic that rules out a generative model and is the whole argument for a System One one. It needs Bun 1.3.14 or newer and a TypeSafe API key, and takes `--recursive` and `--n_walkers`.