Testing Checklist

Testing Checklist

A pragmatic checklist for writing and reviewing tests. Read the relevant section for the change at hand rather than the whole file. Tests are the executable spec — optimize them for reading, not for coverage numbers.

Rationalizations (read first)

What to test (in priority order)

• Branch logic — every non-trivial if/switch/early-return path, especially error paths.
• Boundaries — empty input, one item, max size, null/undefined, zero, negative, off-by-one edges.
• Contracts — what callers rely on: return shapes, thrown error types, emitted events, side effects.
• Error handling — the failure path is production code too; assert what happens when the dependency fails.
• Regressions — every fixed bug gets a test that fails without the fix (write it first, watch it fail).

What NOT to test

• Implementation details — private helpers, internal state, call order that a refactor may change. Test through the public surface.
• The framework — React rendering, ORM query building, library internals. Assume they work; test your logic.
• Trivial pass-throughs — getters, re-exports, config objects with no logic.
• Exact copies of the implementation — a test that mirrors the code's algorithm proves nothing; assert on known input → expected output instead.

Shape of the suite

• Pyramid roughly holds: mostly fast unit tests, some integration at real seams (API route + DB, component + store), few end-to-end. Inverted pyramids (E2E-heavy) are slow and flaky — flag them.
• Integration tests sit at real seams — if everything meaningful is mocked, it's a unit test with extra steps and false confidence.
• Each test asserts one behavior; the name states it (rejects expired token, not test auth 2).
• Deterministic: no real time (vi.useFakeTimers), no real network, no shared mutable state between tests, no order dependence.
• No logic in tests — a for/if inside a test is a test that can be wrong. Table-driven cases (it.each) are fine; computation of the expected value is not.

Test quality (DAMP over DRY)

• The failing output alone tells you what broke — descriptive name, explicit expected value, no dig into helpers to understand the setup.
• Setup visible at the test site: prefer inline literals or a plainly-named builder over layers of beforeEach mutation.
• Assert on values, not on "did not throw".
• Mocks are visible in the test that uses them; a mock configured three files away is a trap for the next reader.

Stack specifics (Vitest / React Testing Library)

• Query by role/label/text (getByRole('button', { name: /save/i })) — testid only when no accessible query exists (that gap is often an a11y bug worth fixing instead).
• userEvent over fireEvent — it simulates real interaction sequences (focus, keydown, click).
• Async UI: await findBy… / waitFor with a real assertion inside — never sleep-and-hope.
• Assert what the user sees/experiences, not component internals (state, props of children, hook return values).
• MSW (or fetch-mock at the boundary) for network — don't mock your own data-fetching wrapper; mock the wire.

TDD loop (new behavior)

For new code, not just bug fixes — let the test drive the design, red before green.

1. Agree the seam first. Decide where the test attaches — a public function, an API route, a component's rendered output — and confirm it before writing the test. Testing below an agreed seam couples the suite to internals; testing above it asserts nothing of yours.
2. Red. Write one failing test for the next thin slice of behavior. Run it, watch it fail for the intended reason.
3. Green. Write the least code that makes it pass. No more.
4. Refactor — outside the loop. Clean up with the test green as the safety net. Refactoring is its own step, never folded into red or green.

Slice vertically — a tracer bullet, one thin path through every layer, demoable — not horizontally (a whole layer at once, testable only much later). One slice at a time.

Bug-fix protocol

1. Write the test that reproduces the bug. Run it — it must fail for the reported reason.
2. Fix. Run it — green.
3. Run the surrounding suite — the fix didn't move the bug next door.

If step 1 is impossible (non-deterministic, environment-bound), say so explicitly per the Verification Exit Criterion — never imply tested-and-covered when it isn't.

Delegation

Delegate the scaffolding, keep the coverage judgment. Hand mechanical generation to a cheaper-tier subagent — test boilerplate, fixtures, obvious table-driven cases, a codemod across many test files — returning ready-to-review tests. Keep on the main model the parts that need judgment: choosing the seam, deciding what is worth testing, and confirming the behavior is actually covered. A green suite is evidence, not proof the acceptance criteria are met.