AI agents are rats in a maze. They reach for what they know, and what they know is slop.
stupify reviews every PR against a corpus of code you actually respect, names the slop, and tells you how to fix it.
git clone https://github.com/Octember/stupify
Runs on your own Codex plan. No API keys, no servers you run. MIT licensed.
Choose the standard: your best code, or your favorite coder. stupify holds every PR to those standards.
A rubric names what counts as slop and keeps the fix small, and it runs on your own Codex plan, so a careful review costs a fraction of the API.
A linter can’t see slop, because nothing is technically wrong. It compiles, passes a skim, and quietly rots. stupify catches it on real PRs, each finding naming the primitive to use instead.
A helper got inlined, so the call site now manually re-remembers a whole construction contract: sorted clips, zeroed counts, active flag. The corpus keeps that as one named unit.
A WriteStream and three boolean flags to do what fluent-ffmpeg’s .output(path) already does, in the same file. Bigger, more ways to get the ordering wrong, for no new behavior.
A phrase list hand-copied into a prompt instead of rendered from the one exported source. It’s already out of sync with the schema and the eval guard. Two copies, one wrong.
Two parallel nullable fields let the schema hold an impossible half-link. One nullable object makes the malformed state impossible to write in the first place.
The whole PR thread is its memory, so once the findings are addressed it posts one line and goes quiet. The opposite of a bot that re-nags on every push.
Clone it, hand it a repo, open a PR.
git clone https://github.com/Octember/stupify