Skip to content

v0.1.0

v0.1.0
@mathaiostop23 mathaiostop23 tagged this 05 Sep 18:41 
A native code-graph engine that gives a coding agent the relevant code instead
of making it grep for it.

Measured on 963 real issues across three languages, against the tool this
project names as its competitor and against grep:

  SWE-bench Verified (Python)   81.8%  at 15,742 tokens
  Multi-SWE-bench (Rust)        74.2%  at 12,646
  Multi-SWE-bench (TypeScript)  42.6%  at  9,149
  codegraph explore             37.1% / 12.4% / 17.0%
  keyword top-10                50.6%  at 282,720

Every figure rebuilds from a script in bench/. What is not measured is said in
the README rather than left out: whether a patch built on this context passes
the repository's own tests has never been run with a model behind it.
Assets 2
Loading