What will your agent get wrong?
Your agent's training data has a cutoff. Your dependencies do not. Drop a manifest and see every major it was trained before — the deprecated APIs it will reach for, scored, with the exact rules that stop it.
Your manifest
Parsed in your browser. Nothing is uploaded.
Score your repo against what an agent can get wrong from memory.
Every dependency major your agent was trained before is a set of APIs it will reach for anyway. Drop a manifest and see which ones — or paste a CLAUDE.md to check it for leaked keys and prompt injection.