I’m the maker of Superloopy, a small MIT-licensed workflow layer for Codex and Claude Code. I built it around a problem I kept running into with coding agents: after a long task, the final answer often sounds confident, but the human still has to reconstruct what was actually checked. The pattern I’m trying is an evidence gate before the agent can call work done: - define acceptance criteria up front - route specialized work through skills/subagents when useful - run command-backed che...