Claude Code sometimes hallucinates user messages (lesswrong.com)

wondering if these specific hallucinations end up costing more tokens because of req/resp cycles :/

failproofai Agent Monitor (github.com/exospherehost/failproofai) logs the full pre/post tool-call payload for every session, so you can enforce rules to detect wired patterns and reduce hallucination a lot