Auto

Autonomous mode lets proposals that pass a policy reach the skill library without you reading them first. Handle $ARGUMENTS as follows.

With no arguments, or status — run patina pending auto status and report it. Say plainly whether it is on, what the thresholds are, and what is currently on trial. If anything's trial has expired, name it: the curator will archive it on its next pass.

--dry-run — run patina pending auto --dry-run. Nothing is applied. Report both halves and do not bury the second one: what the policy would take, and what it would hold with the reason for each. The reasons are the point. If the run would take a large number at once, say so with the number rather than summarising it as "several".

on — this is the consequential one. Before running it, put the trade in front of the user in your own words:

• What lands without asking: high-confidence proposals with at least the minimum number of lessons behind them, lint-clean, and no name they have rejected before. Everything else still queues.
• What they get instead of the queue: a git commit per write in ~/.claude/skills that git revert undoes, a trial window after which the curator archives anything nothing loaded, and a prompt the first time an auto-approved skill actually runs.
• What it costs: skill descriptions enter the system prompt of every later session whether or not the skill is invoked.
• What it covers: the whole queue, including whatever is already waiting in it, applied on every scheduled background pass — not only proposals filed from here on. If they have a backlog, say how large it is, because turning this on is what decides it.

If they have a queue they have not seen, run --dry-run first and report it before running on. Then run patina pending auto on.

off — run patina pending auto off. Anything already approved stays where it is; trials already running still expire. Say both.

Applying the policy right now (patina pending auto, no flag) — this is only needed to avoid waiting for the next scheduled pass; autonomous mode does it on its own. Show the dry run first if the user has not seen one.