I audited my own AI-generated refactor and found 46 bugs. Here's what that taught me.

Short summary

A developer refactored a 1,920-line CLI file using AI coding agents, shipped it with green tests, then discovered 46 undefined-reference bugs across 7 files that tests never caught. The root cause was mechanical extraction silently dropping imports in code paths only triggered by specific flags or project states. The fix wasn't a smarter model but adding ESLint's no-undef rule and manually verifying every CLI command against a real project.

• •AI-assisted refactor of 1,920-line file shipped with green tests but 46 hidden undefined-reference bugs
• •Tests passed because broken code paths only executed on specific flags/project states
• •Real fix was static analysis (ESLint no-undef) plus manual verification of all 14 CLI commands

Generated with AI, which can make mistakes.