Tip: jev-use publishes rerunnable benchmarks for routing non-text agent steps to a decision model

GitHub·high signalXBlueskyLinkedInCopy link

jev-use (created 2026-09-19) is a Claude Code / Codex / pi plugin that hands steps needing no text output to TypeSafe's Jev judgment model, installed with `npx -y jev-use install`. Its README carries measured runs rather than claims: a browser directions task at 10 decisions with p50 274 ms alongside 4 LLM writes, a PreToolUse shell gate denying dangerous commands in ~230 ms at zero LLM tokens, context compaction judging 200 messages in 7 calls to take a window from 94% to 44% with 3/3 recall, and a Pong latency test at 86 Jev decisions in 20 s versus 6 for Haiku and 3 for Gemini called normally (3x after enum-constraining both). Every demo ships as a script in bench/examples with methodology and variance in bench/RESULTS.md.