Twinrun

Description

Refactoring code — or migrating it to a new library version, a new framework, or a rewritten module — always carries the same fear: did I just quietly change behavior somewhere I didn't test? Writing new unit tests to catch every possible regression is slow and you can never be sure you covered everything the old code actually did. Twinrun takes a different approach: instead of writing new assertions by hand, it runs both the old and new versions of your code on the same generated inputs and automatically flags anywhere the outputs, exceptions, printed output, or even mutated arguments diverge.

Under the hood, Twinrun checks out both code revisions into isolated git worktrees, generates test inputs automatically from type annotations, existing test-suite literals, and guard constants in the code (with an optional --llm flag to generate more when it runs out of ideas), and executes both versions in separate subprocesses for clean isolation. It compares far more than return values — types, exceptions raised, stdout, argument mutation, and instance state are all checked — then filters out flaky/non-deterministic noise and clusters related findings so a hundred symptoms of one root cause don't look like a hundred separate bugs. An --accept flag lets you formally record a behavior change as intentional rather than a regression. It's free, open source, Python 3.10+, and has zero external dependencies.