better-test

Apply the six-condition value gate: the test needs a traceable source, a meaningful failure mode, independent expected values, a stable observation seam, resistance to refactoring, and a single owner. Tests failing any condition are rewritten or deleted with documented evidence.

Judge tests individually rather than by file, reading each test with its production owner, callers, and git history. Issue KEEP, REWRITE, or DELETE verdicts with evidence, remove test-only production seams and dead fixtures, then run the narrowest affected verification.

Mock only at system boundaries such as external APIs, databases, time, and randomness. Never mock your own classes or internal collaborators. Design mockable boundaries with dependency injection and SDK-style interfaces where each function is independently mockable.

Key anti-patterns include tautological tests that recompute expectations like the implementation, change detectors asserting call counts or snapshots, duplicated scenarios across layers, hollow tests without assertions, and treating appearance like layout classes as behavior.

Seeing a test fail red proves it actually exercises the intended behavior rather than only validating mocks. For tests added to existing behavior, temporarily break the implementation, confirm the failure, then restore precisely and verify green.

Use manual evidence gathering in order of closeness to real user paths: browser inspection, curl or CLI calls, asserted logging, or desktop automation. Only record test debt for gaps that pass the value gate and deserve long-term protection.