KarLex AI

Graphicle gives Meta's models the right part of a whole code repository, cited to file and line, and closes the loop from a bug on the screen to a tested fix. Built during the event, in one live run on excalidraw issue #12053 (30 Oct 2026): Muse Spark drives the app with computer use and reproduces the bug; SAM 3.1 outlines what it pointed at and Muse Spark reads it, so Graphicle can map it to the code; Muse Wonder reads a 193,474-token pack of the repository, plans, patches and runs the project's tests; when the first round fails, Graphicle feeds the failing lines and the files they name into the next round, which passes excalidraw's own regression test for the bug; then Muse Spark checks the rebuilt app on screen. Fixed and checked in 5 min 12 s, 12 model calls, $1.64 (cost computed from Meta's reported tokens at list prices). The same loop passed the tests in 3 of 3 live runs with the new fix rounds, and in 0 of 2 with Graphicle 0.1.10. We also measured whether the pack helps a model find the right code, live and on repositories we never tuned on. 90 held-out questions on pytorch, vscode and elasticsearch, Muse Spark 1.3 with 128K tokens of context: the right file was in the top 5 for 71% with Graphicle's pack, 63% with BM25 and 51% with the first 128K tokens of the repository (pack against the first 128K: p = 0.0003; against BM25: p = 0.09, not significant). On Muse Spark Open (open weights, served by Fireworks) at 1M tokens: 66% against 48% for the first 1M tokens (67 paired questions, p = 0.002). A 1M-token pack on the Meta Model API: the first Muse Spark call cost $1.26, the next 19 on the same pack $0.25 to $0.27 each, about 90% served from Meta's prompt cache. Graphicle is open source (Apache-2.0), and the event's work shipped in four releases, 0.1.9 to 0.1.12, with release archives for macOS, Linux and Windows and a Homebrew formula. HACKATHON.md in the repository lists everything built during the event with when it was written, and what existed before it (Graphicle 0.1.8, and first drafts of the pack, `ask --pack`, `plan` and the benchmark harness, written on 26 October before the kickoff and not released).