How to Audit What Your AI Agent Actually Did After the Session

When you hand off a multi-hour task to an AI coding agent and come back to the results, the right question isn't "did it finish?" — it's "did it stay within scope?" Agents running Claude Code, Codex, or OpenCode regularly do more than instructed: touching files outside the task boundary, introducing abstractions nobody requested, reorganizing directory structures that were working fine. The damage is usually invisible until it's compounding across three or four subsequent sessions. This tutorial