test-behavior-not-implementation（测试行为而不是实现）

Testing Trophy

Testing philosophy for React and TypeScript projects. Based on Kent C. Dodds' Testing Trophy and TkDodo's testing principles.

Testing Trophy Priorities

From most to least valuable:

1. Integration tests — test multiple units working together (highest confidence per effort)
2. Unit tests — test individual functions in isolation
3. End-to-end tests — test full user flows (highest confidence, highest cost)
4. Static analysis — TypeScript, ESLint (cheapest, catches typos and type errors)

"The more your tests resemble the way your software is used, the more confidence they can give you." — Kent C. Dodds

Core Principles

1. Test behavior, not implementation

Tests MUST assert on what users see or experience — rendered output, fired events, returned values from public API. If a test would break when you refactor internals without changing behavior, it's testing implementation.

Do NOT assert on:

• Internal state values
• Private function calls
• Component instance methods
• Hook return internals that consumers don't use
• How many times a function was called (unless that IS the behavior)

2. Write fewer, longer tests

One test covering a complete user flow is better than five tests each checking one micro-step. Act like a user — users don't stop after clicking one button.

3. Use accessible queries (correct priority)

getByRole > getByLabelText > getByPlaceholderText > getByText > getByDisplayValue > getByAltText > getByTitle > getByTestId

Only use getByTestId when no accessible query works.

4. Don't test third-party code

If a behavior is owned by a library (React, a UI library, a hook library), don't re-test it through your component. Trust the library's tests.

5. Avoid unnecessary mocking

Mocks create a parallel reality where tests pass but production breaks. Only mock at system boundaries:

• Network requests (fetch, API calls)
• Timers (setTimeout, setInterval)
• Browser APIs not available in test env (IntersectionObserver, ResizeObserver, canvas)

NEVER mock internal modules just to isolate units. Import the real module.

6. One test per behavior

"Clicking submit with valid data shows success" and "clicking submit with valid data calls the API" are testing the same behavior from two angles — merge them. Users experience one thing, test one thing.

TkDodo's Principles

Don't test what the framework already tests

React guarantees that useState works. Don't write tests proving it. Your utility library guarantees hotkey matching — don't re-test that matching through every consumer.

Redundant tests are worse than no tests

Every test is code you maintain. A test that duplicates another test's coverage adds maintenance cost with zero confidence gain. When implementation changes, you fix N tests instead of 1.

Test the contract, not the wiring

A hook's contract: "given these inputs, produce these outputs/effects." How it wires state, refs, and effects internally is not the contract.

Implementation detail detector

If a test name contains "should call" followed by an internal function name, it's probably testing implementation.

Decision Tree

For every test case, ask:

Is this testing behavior a user/consumer would notice?
  |-- Yes -> Is there another test covering the same behavior?
  |     |-- Yes -> Merge or remove the weaker one
  |     |-- No  -> Keep as-is
  |-- No  -> Is it testing an internal/implementation detail?
        |-- Yes -> Remove or rewrite to test the observable behavior instead
        |-- Not sure -> Would this test break on a refactor that doesn't
                        change behavior?
                         |-- Yes -> It's implementation-detail testing. Remove.
                         |-- No  -> Keep

Mocking Discipline

• NEVER use vi.mock() or jest.mock() for internal modules. Import the real module.
• If the real implementation works in the test environment, use it.
• When mocking is justified (e.g., ResizeObserver in jsdom), mock the minimum — don't mock the entire module when you only need one function.
• Legitimate mock targets: fetch/network, fs (filesystem), timers, canvas, IntersectionObserver, ResizeObserver.

Test Structure Rules

• Prefer userEvent over fireEvent — userEvent simulates real browser behavior
• No overengineered setup/teardown — beforeEach should only contain truly shared setup (timer mocking, observer polyfills)
• No unnecessary wrapper components — render components directly unless they require a provider
• Test names describe behavior, not implementation:

- Bad: "should call setState with new value" - Good: "updates the display when user types"

• No snapshot tests unless testing serialization output
• Colocate tests with source: stamp-card.test.tsx next to stamp-card.tsx

Anti-Patterns Checklist

When reviewing or auditing tests, check for:

• Duplicate test cases — two tests covering the same behavior through different assertions
• Re-testing upstream behavior — behavior already tested in its own test file
• Assertions on internal state — hook return values that consumers don't use directly
• Assertions on call count — toHaveBeenCalledTimes(1) when count isn't the contract
• Spying on internal methods — vi.spyOn on functions that aren't public API
• Mocking internal modules — vi.mock() for modules that work in test env
• Overengineered setup — beforeEach doing what each test does anyway
• Missing error/edge cases — empty input, null/undefined, boundary values, cleanup on unmount
• Missing accessibility assertions — ARIA attributes, keyboard navigation, focus management

What NOT to Test

• Pure type files, constants, or re-export index files
• Thin wrappers that delegate to tested functions
• Pure styling wrappers with no logic (CVA variants, styled-components with no conditionals)
• Framework behavior (useState works, useEffect fires on mount)

The goal is fewer, better tests — not more tests.