You need to define an escape hatch for the LLM otherwise it will loop on increasingly small issues and even start making up issues that don't actually exist. A simple "fix all medium and high severity issues" guidance will prevent low-priority nits from consuming all your tokens.