🤖 AI coding agents have one surprisingly annoying problem: they can build the thing but they can't explain what the hell they did You ask Claude Code / Codex / Cursor to fix something, it works for a minute, and then you get: - auth middleware refactored - session TTL adjusted - fallback branch patched - changes committed Cool. But is it actually finished? Did anything break? Do I need to do something now? Found a small open-source project called Open Steps that tries to solve exactly…