Honey (I Shrunk the AI) by GreenPT

AI coding agents burn most of their tokens on waste: code that didn't need to exist, prose nobody reads, bloated agent-to-agent handoffs. Honey fixes all three reflexively, YAGNI/stdlib-first code, answer-first prose, handoffs halved via compact JSON or ESON, plus ~75% cheaper bulk reads as rendered images (pxpipe). Benchmarked head-to-head vs Caveman & Ponytail on Opus and GPT: −49% output tokens at 98% quality, only variant 100% lossless on handoffs. Free, MIT, ships to ten agents from skills.