ix

search that agrees with itself.

ast-grep × grep × meaning — glued at one spot, with a confidence score.

You already know ast-grep (structure) and grep (text). ix runs both — plus a meaning search — glues the three at the same place, and gives every hit an R: how sure it is, because they agreed. R = 1 when all three land on the same line; lower when only some do. Confidence, not a black-box cosine.

Try it — drop a repo

constructor v0 of the living-files idea: any data → reduced to one latent → glued by the sheaf → an explorable, R-scored world. Repos first; the machine is the same for the rest. Running live below (first query warms the model, ~20 s).

Get it

For an agent

ix "..." --json returns {path, line, R, agreed} per hit — so an agent acts on confidence (take R ≈ 1, escalate R < 1) instead of guessing on a single cosine.

The idea behind it

ix is the runnable instance of The Matryoshka Sheaf: local readings glue into one global result exactly when they agree, with R = exp(−d_tail/scale) as the gluing obstruction. The matryoshka embeddings nest (a prefix is a smaller embedding), so the first pass is coarse-and-fast and only the survivors unfold to full detail.

📄 Paper + runnable code  ·  🧩 ELI5  ·  🌀 artificial intuition