jev-effort
A research proxy that lets a small decision model choose Claude Code's reasoning effort on every step, measured to cut a benchmark's cost by up to 55% without breaking the prompt cache.
jev-effort is a research proxy that lets TypeSafe's small decision model, Jev, choose Claude Code's reasoning effort on every agent step, on the premise that most steps are routine and can run at low effort while deep reasoning is reserved for hard ones. It relies on Anthropic's per-message effort beta, which lets an effort-only system message change the reasoning level mid-session while leaving the cached prompt prefix intact, so changing effort no longer resets the conversation cache the way it used to.
Measured on a 24-session hidden-test coding benchmark, choosing effort at the high level cut cost by about 1% with thinking down 46% and every test still passing; at the max level it cut cost by about 55% (21% excluding one outlier task), with thinking down 98.5%, wall time down 58%, and every test still passing. In a 470-step, 92-minute real session at max, the estimated saving was 4-9%, and 411 checked steps showed a 99.1% cache hit rate with zero unexpected cache misses.
It is an unofficial research project, not affiliated with Anthropic or TypeSafe, written in JavaScript.