I was tired of agents ignoring AGENTS.md mid-session, then introducing invalid states (representable) into interfaces and building upon them, until finally being caught only by reviewer agents (only if!) or by manual reviews.
Agents are much more cooperative when their hand is slapped right after a wrong assumption. That assumption doesn't poison the context no more after that. So, Hapsland was born. Behold.
TLDR: edit tools hook + semantic enhancement + decision api; full control over "what goes to Jev" and "what is saved to the disk"