A social post points at an agent-training idea where terminal output becomes part of what a CLI agent learns from, not just background noise after a command. The available files only preserve metadata and prior notes, so the linked ECHO paper is the next artifact to inspect before anyone treats the method as validated. Still worth a click if you like arguments about where automation feedback actually lives.