process-test-design

Default to integration or e2e tests against the real dependency, and inject collaborators through a port or interface so tests can substitute a fake behind the same contract. Reserve monkeypatching for true external edges you cannot inject, like third-party globals, the clock, or the network.

Cover text inputs (empty, whitespace, emoji, CJK, RTL, zero-width), bytes (control chars, NUL, invalid UTF-8), wrong types, missing or extra fields, numeric edges (0, -1, max, overflow, NaN), size limits, and malformed structure. Write one test per branch the code actually distinguishes, asserting the specific error or normalization behavior.

Flakiness comes from shared mutable state between tests, real sleep calls for timing, and uninjected clocks or randomness. Each test must set up its own fixtures, poll conditions with timeouts instead of sleeping, and pass both alone and under shuffled ordering.

Match the reproduction pressure to the mechanism: races between tasks reproduce under CPU starvation with busy competitors pinned to one core, while races against a timer require slowing the clock the test races, such as a UI framework tick or poll interval. Always reproduce through the real test entry point.

Delete tests that assert nothing, restate the implementation line-by-line, test the language or framework rather than your behavior, or duplicate another test's coverage. Fewer behavior-focused tests beat many brittle ones; coverage percentage is a smell detector, not a goal.

Yes, all principles are language-neutral with a per-language mechanics table covering Python, TypeScript/JS, Go, Rust, and Bash. The table maps each ecosystem's injection seam, self-mock anti-pattern, real dependency option, and test shuffle flag.