Robo Refine and Polish

A long roborev refine loop without bookkeeping is how you end up "fixing" the same finding three times, missing a regression you introduced two iterations ago, and silently re-litigating a deliberate design decision because the next agent run can't see your reasoning. This skill is the bookkeeping.

/roborev-refine runs the review-fix-rereview cycle. This skill keeps the ledger that turns the cycle into a converging process.

Use it when:

• The loop is expected to take more than ~2 iterations.
• Two or more reviewer agents are involved (e.g. claude-code + codex).
• The user has asked for convergence ("loop until clean", "address every finding", "max 10 iterations"), or for the loop to be tracked privately.
• A finding has come back after a fix and you need to decide whether to re-fix, defend, or escalate.