The problem: You hand off a multi-hour work queue and walk away, and everything downstream of that moment is a gamble. An agent that stops to ask permission on every judgment call burns the whole night waiting on you. An agent that never stops might quietly delete something, push to main, or send a message it shouldn’t have. Neither failure announces itself until morning, and by then it’s too late to redirect.
This is the actual contract, not a hope. It sets decision rules for choosing without you in the loop (reuse existing patterns, prefer reversible changes, validate small then broaden), a short list of true blockers that always stop the work (missing credentials, anything destructive or irreversible, a decision you explicitly reserved), and a place for every non-blocking question to land instead of interrupting you: a QUESTIONS-FOR-YOU.md file, triaged in the morning instead of pinged at 2am. Status lives on a phone-checkable surface with a three-second glance test: a health callout and a status row above the fold, nothing else. And because in-session task state doesn’t survive a context compaction, the results file, not the task tracker, is the ground truth on resume.
Part of Operator Skills, a collection of skills for running a real Claude Code setup.