A QUALITY GATE FOR GITHUB ISSUES

refusepr

Coding agents will happily rebuild what you already shipped. This reads the ticket first and says no — before a single token is spent writing code.

INSTALL ON GITHUB SEE IT REFUSE

DUPLICATE OF #412ALREADY FIXED IN a3f9c21NO REPRO STEPS14 FILES / 900 LINES — TOO BIGPR #9 ALREADY DOES THISNO ACCEPTANCE CRITERIASUPERSEDED — 118 DAYS QUIETSAME BUG, DIFFERENT WORDS

WHAT IT ACTUALLY POSTS

A real refusal from a real run. It costs about a tenth of a cent to produce, and it stops a generation run that would have cost a thousand times more.

#847Clear the session cookie on logout

ALREADY DONE

This appears to be done already, so no code was generated.

commit a3f9c21 — “Clear session cookie on logout” — 93% confidence

THE ARITHMETIC

Checking is cheap. Generating is not. Every refusal is a generation run that never had to happen.

$0.001to check a ticketmeasured across live runs

$1.50+to generate a pull requesttypical agent run, low estimate

37%of agent pull requests are rejectedn=9,799 reviewed PRs · arXiv 2605.22534

~1 in 3of those rejections are nothing to do with the code — duplicates, superseded work, wrong contextsame study

Separately, roughly half of patches that pass their benchmark would still not be merged by a maintainer. METR, March 2026

AND WHEN A TICKET IS GOOD

It writes the change and opens a draft pull request — small enough that someone will actually review it. Work that grows past the limit is discarded rather than pushed.

SEE A REAL PR IT OPENED