Markets
Twinrun Runs Both Versions of Changed Code on the Same Inputs Because a Diff Doesn't Say What the Program Now Does
Posted to Show HN 2026-09-05, MIT-licensed, Twinrun identifies modified callables, builds inputs from type annotations and literals scraped out of the test suite, executes old and new in isolated git worktrees, normalizes and filters non-deterministic output, then clusters differences by root cause. It requires no assertions, which is the point of separation from LLM reviewers that only read the diff, static analyzers that see one revision, test generators that need assertions, and API differs that ignore behavior. The stated motivation is the volume of AI-generated pull requests.
↳ Follow the thread