You start from a clear ask, after answering only what you alone know.
Claude Code. With moo off, you say the agent box is running out of disk space and ask what can be cleaned safely. Claude reads the dumped df, du and worktree state and ranks what is safe to clean: idle worktrees and build output, about 315 GB. It warns against deleting straight from the stale snapshot and lists a second tier to check first. With moo on, you type /hope:intent and the same words. /reframe comes up, and one question notes the agent runner never removes its worktree, so free space falls about 12 GB a day, and asks which job you want; you pick cleanup plus stopping the growth. /deepen follows. The reply says it will write a sweep script and change the runner, then tables what is safe to clean, what to keep and what to ask about, and the rules: a dry run by default, how you will know it worked, and what it must not touch. It ends with your step: run the sweep on the box, dry run first.