I kept shipping AI features and watching the same failure mode: agents and humans both work around dead code instead of removing it. We would leave an unused helper “just in case,” add a new route beside an old one, keep a package in the lockfile after the import was gone. On the next task, Claude or Cursor would route around those files again. The dead path stayed in the tree, and the cost showed up as slower reviews, noisier diffs, and more tokens spent understanding code that never ran. I tried the usual tools. Linters catch unused locals. Bundle tree-shakers help at the edge. Full “AI review” products comment on the diff, not the whole graph. Nothing gave me one local command that said: here is the evidence this export is never reached—and nothing put that same answer inside Claude, Cursor, and Codex without three different products. So I built Deadpath: a deterministic dead-code scan first, optional model explain second, and thin plugins so agents call the scanner instead of guessing. No auto-delete. A plan you can turn into a PR. It sits next to the other tools I’m building in public—Cosen for cost/traces/security at runtime, DeadPath for pre-merge review across UX, API, performance, and engineering. Deadpath is the hygiene step: prove what nothing still reaches, then cut on purpose. If you try it, run deadpath scan --mock on your messiest repo and tell me what it gets wrong. False positives are the product; I want those reports.

Deadpath

find the code your agents keep rewriting around

find the code your agents keep rewriting around