proof

Write a regression test that fails before the fix and passes after, entering at the boundary where the caller observes the behavior. Record the command and output as evidence, and verify the red state as deliberately as the green.

Place tests at seams where data shape or values observably change: extract output, transform output, parse result, or validator output. Assert success and error envelopes at those boundaries rather than writing one unit test per internal helper function.

Skip proof for edits with no behavior surface such as formatting, typos, comments, and renames that tooling already confirms. Also avoid tests that assert framework behavior, literal text in config files, or trivial passthroughs with no behavior.

A green suite only counts if a named check would fail when the specific claim is false. A passing check that never touches the changed boundary is partial evidence, so identify which test breaks if you are wrong before claiming completion.

Treat heavy mocking as a design signal: the boundary is wrong, not the test framework. Simplify the seam by extracting pure transforms and pushing effects to the edges, or hand off to refactoring before adding more mocks.

Treat a flaky test as a bug in the test, the code, or the environment and fix it directly. Do not use sleeps or retries; make the check deterministic by controlling true system edges like clock, network, and filesystem.