Agent Skills: Writing Tests

Writing Tests

Core Philosophy: Test user-observable behavior with real dependencies. Tests should survive refactoring when behavior is unchanged.

Iron Laws:

Testing Trophy Model

Write tests in this priority order:

1. Integration Tests (PRIMARY) - Multiple units with real dependencies
2. E2E Tests (SECONDARY) - Complete workflows across the stack
3. Unit Tests (RARE) - Pure functions only (no dependencies)

Default to integration tests. Only drop to unit tests for pure utility functions.

Pre-Test Workflow

BEFORE writing any tests, copy this checklist and track your progress:

Test Writing Progress:
- [ ] Step 1: Review project standards (check existing tests)
- [ ] Step 2: Understand behavior (what should it do? what can fail?)
- [ ] Step 3: Choose test type (Integration/E2E/Unit)
- [ ] Step 4: Identify dependencies (real vs mocked)
- [ ] Step 5: Write failing test first (TDD)
- [ ] Step 6: Implement minimal code to pass
- [ ] Step 7: Verify coverage (happy path, errors, edge cases)

Before writing any tests:

1. Review project standards - Check existing test files, testing docs, or project conventions
2. Understand behavior - What should this do? What can go wrong?
3. Choose test type - Integration (default), E2E (critical workflows), or Unit (pure functions)
4. Identify dependencies - What needs to be real vs mocked?

Test Type Decision

Is this a complete user workflow?
  → YES: E2E test

Is this a pure function (no side effects/dependencies)?
  → YES: Unit test

Everything else:
  → Integration test (with real dependencies)

Mocking Guidelines

Default: Don't mock. Use real dependencies.

Only Mock These

• External HTTP/API calls
• Time-dependent operations (timers, dates)
• Randomness (random numbers, UUIDs)
• File system I/O
• Third-party services (payments, analytics, email)
• Network boundaries

Never Mock These

• Internal modules/packages
• Database queries (use test database)
• Business logic
• Data transformations
• Your own code calling your own code

Why: Mocking internal dependencies creates brittle tests that break during refactoring.

Before Mocking, Ask:

1. "What side effects does this method have?"
2. "Does my test depend on those side effects?"
3. If yes → Mock at lower level (the slow/external operation, not the method test needs)
4. Unsure? → Run with real implementation first, observe what's needed, THEN add minimal mocking

Mock Red Flags

• "I'll mock this to be safe"
• "This might be slow, better mock it"
• Can't explain why mock is needed
• Mock setup longer than test logic
• Test fails when removing mock

Integration Test Pattern

describe("Feature Name", () => {
  setup(initialState)

  test("should produce expected output when action is performed", () => {
    // Arrange: Set up preconditions
    // Act: Perform the action being tested
    // Assert: Verify observable output
  })
})

Key principles:

• Use real state/data, not mocks
• Assert on outputs users/callers can observe
• Test the behavior, not the implementation

For language-specific patterns, see the Language-Specific Patterns section.

Async Waiting Patterns

When tests involve async operations, avoid arbitrary timeouts:

// BAD: Guessing at timing
sleep(500)
assert result == expected

// GOOD: Wait for the actual condition
wait_for(lambda: result == expected)

When to use condition-based waiting:

• Tests use sleep, setTimeout, or arbitrary delays
• Tests are flaky (pass locally, fail in CI)
• Tests timeout when run in parallel
• Waiting for async operations to complete

Delegate to skill: When you encounter these patterns, invoke Skill(ce:condition-based-waiting) for detailed guidance on implementing proper condition polling and fixing flaky tests.

Assertion Strategy

The Golden Rule of Assertions: A test must fail if, and only if, the intention behind the system is not met.

This rule is bidirectional. A test is broken when it:

1. Doesn't fail when the intention is actually broken
2. Does fail when the intention is perfectly fine (implementation detail changed, external service unreachable, etc.)

Before merging any test, ask: "When will this test fail?" If the answer includes anything other than "when the behavior this test describes is broken," the test needs work.

Assert on observable outputs, not internal state

| Context | Assert On | Avoid | | ------- | ----------------------------------------------------- | ------------------------------------- | | UI | Visible text, accessibility roles, user-visible state | CSS classes, internal state, test IDs | | API | Response body, status code, headers | Internal DB state directly | | CLI | stdout/stderr, exit code | Internal variables | | Library | Return values, documented side effects | Private methods, internal state |

Why: Tests that assert on implementation details break when you refactor, even if behavior is unchanged.

Respect test boundaries

A test that makes a real network request doesn't just test your function -- it tests DNS resolution, network connectivity, server uptime, and response timing. When any of those fail, your test fails even though your code is fine. That violates the Golden Rule.

Fix: Mock at the boundary of what you don't own or control. The function under test isn't responsible for the server's validity -- it's responsible for making the right request and handling the response correctly. Use API mocking (e.g., MSW, respx, httpmock) to make external interactions fixed, predictable givens.

This is not a contradiction with "default: don't mock." Internal modules stay real. External boundaries (network, third-party services) get mocked so your test only fails when your code's intention is broken.

Test Data Management

Use source constants and fixtures, not hard-coded values:

// Good - References actual constant or fixture
expected_message = APP_MESSAGES.SUCCESS
assert response.message == expected_message

// Bad - Hard-coded, breaks when copy changes
assert response.message == "Action completed successfully!"

Why: When product copy changes, you want one place to update, not every test file.

Anti-Patterns to Avoid

Testing Mock Behavior

// BAD: Testing that the mock was called, not real behavior
mock_service.assert_called_once()

// GOOD: Test the actual outcome
assert user.is_active == True
assert len(sent_emails) == 1

Gate: Before asserting on mock calls, ask "Am I testing real behavior or mock interactions?" If testing mocks → Stop, test the actual outcome instead.

Test-Only Methods in Production

// BAD: destroy() only used in tests - pollutes production code
class Session:
    def destroy(self):  # Only exists for test cleanup
        ...

// GOOD: Test utilities handle cleanup
# In test_utils.py
def cleanup_session(session):
    # Access internals here, not in production code
    ...

Gate: Before adding methods to production code, ask "Is this only for tests?" Yes → Put in test utilities.

Mocking Without Understanding

// BAD: Mock prevents side effect test actually needs
mock(database.save)  # Now duplicate detection won't work!

add_item(item)
add_item(item)  # Should fail as duplicate, but won't

// GOOD: Mock at correct level
mock(external_api.validate)  # Mock slow external call only

add_item(item)  # DB save works, duplicate detected
add_item(item)  # Fails correctly

Incomplete Mocks

// BAD: Partial mock - missing fields downstream code needs
mock_response = {
    status: "success",
    data: {...}
    // Missing: metadata.request_id that downstream code uses
}

// GOOD: Mirror real API completely
mock_response = {
    status: "success",
    data: {...},
    metadata: {request_id: "...", timestamp: ...}
}

Gate: Before creating mocks, check "What does the real thing return?" Include ALL fields.

TDD Prevents Anti-Patterns

1. Write test first → Think about what you're testing (not mocks)
2. Watch it fail → Confirms test tests real behavior
3. Minimal implementation → No test-only methods creep in
4. Real dependencies first → See what test needs before mocking

If testing mock behavior, you violated TDD - you added mocks without watching test fail against real code.

Language-Specific Patterns

For detailed framework and language-specific patterns:

• JavaScript/React: See references/javascript-react.md for React Testing Library queries, Jest/Vitest setup, Playwright E2E, and component testing patterns
• Python: See references/python.md for pytest fixtures, polyfactory, respx mocking, testcontainers, and FastAPI testing
• Go: See references/go.md for table-driven tests, testify/go-cmp assertions, testcontainers-go, and interface fakes

Quality Checklist

Before completing tests, verify:

• [ ] Happy path covered
• [ ] Error conditions handled
• [ ] Edge cases considered
• [ ] Real dependencies used (minimal mocking)
• [ ] Async waiting uses conditions, not arbitrary timeouts
• [ ] Golden Rule: test fails if and only if the intention is broken
• [ ] Tests survive refactoring (no implementation details)
• [ ] No test-only methods added to production code
• [ ] No assertions on mock existence or call counts
• [ ] Test names describe behavior, not implementation

What NOT to Test

• Internal state
• Private methods
• Function call counts
• Implementation details
• Mock existence
• Framework internals

Test behavior users/callers observe, not code structure.

Quick Reference

| Test Type | When | Dependencies | | ----------- | ----------------------- | ---------------------------- | | Integration | Default choice | Real (test DB, real modules) | | E2E | Critical user workflows | Real (full stack) | | Unit | Pure functions only | None |

| Anti-Pattern | Fix | | ------------------------------- | --------------------------------------- | | Testing mock existence | Test actual outcome instead | | Test-only methods in production | Move to test utilities | | Mocking without understanding | Understand dependencies, mock minimally | | Incomplete mocks | Mirror real API completely | | Tests as afterthought | TDD - write tests first | | Arbitrary timeouts/sleeps | Use condition-based waiting |