skillranker: a 139,000-line Rust project that uses Jev for skill routing
Hands the question of which skill to load next to Jev as a typed Choice decision with a probability, and the author distills five rules for wiring it in, including separating deciding from generating and requiring real abstention.
JEV JUST TURNED THE HARDEST PART OF AI ENGINEERING INTO A MULTIPLE CHOICE TEST a model that cannot write code, cannot write prose, cannot reason - and that is exactly why it wins. six days ago somebody shipped 139,000 lines of Rust that does nothing but ask it ONE question: which skill loads next. 114 stars already. skillranker does not generate anything. it hands your live session to Jev as a typed Choice, gets back one option from a set it defined, plus a probability. the agent still decides what to consult. 5 rules before you wire it in. -> SEPARATE the deciding from the generating. "which skill fits this turn" is multiple choice with a bounded answer set. most loops burn a frontier generative call on it. two logical requests, at most four HTTP attempts, a three second total budget - that is the entire ranking path. -> DEMAND a real none-of-these. both comparison passes carry an explicit abstain option and the thing stays silent when nothing fits. a ranker that always ranks is a ranker that always interrupts. abstention shows up in 36 of its files - it is load bearing, not a flag. -> READ the limitations section before the benchmarks. this repo does that work for you: "these targets are not remote-service guarantees or measured benchmark results." a perf table that disclaims itself is worth more than one that does not. -> KEEP observations apart from judgments. it logs observed load, unobservable, censored, and explicit usefulness judgment as four different things, then says outright that load observations are affected by the suggestion itself. most agent telemetry collapses all of it into "it worked." -> ASSUME retrieval is the ceiling. past 254 skills a lexical prefilter narrows the set before Jev ever sees it, and the repo admits that step misses paraphrases. your ranker is only as good as the candidate list you handed it. third time this year the interesting engineering was in the abstain path and not the ranking would you let something rank your skills if it could not tell you it had nothing
This post quotes another post: x.com/me_barnyx/status/2101976764779999381