Test Coverage

Test Coverage

Audit gaps, write targeted tests, enforce thresholds - across any ecosystem.

Mental Model

The testing pyramid encodes an economic truth: each tier tests what only it can test.

Coverage is a regression gate, not a quality metric. High coverage with bad tests is worse than moderate coverage with good tests. The goal is: new code cannot silently skip tests.

Exclusions are architecture, not exceptions. Every exclusion documents a deliberate decision about where code is tested. An exclusion at one tier should have coverage at another.

Decision Tree

Start here. Follow the branch that matches the current state.

Is there any coverage tooling configured?
├── No → Bootstrap (below)
└── Yes
    ├── Coverage below target? → Audit & Improve (below)
    ├── Coverage adequate but not enforced? → Enforce (below)
    ├── Coverage enforced, writing new code? → Write Tests for New Code (below)
    └── Coverage high but tests feel weak / want to raise quality? → Test Quality: Beyond Coverage (below)

Bootstrap: Setting Up Coverage from Scratch

1. Detect the ecosystem

Check for project markers: package.json, go.mod, Cargo.toml, pyproject.toml, setup.py, *.csproj. See ecosystem patterns for tool recommendations per language. The examples below are TypeScript/Python-flavoured; for Go or Rust follow ecosystem-go.md / ecosystem-rust.md instead (ecosystem-python.md has the full Python setup).

2. Create tiered configs

Each test tier gets its own configuration file with targeted include/exclude patterns. This prevents slow integration tests from blocking fast unit test feedback.

Key principles:

• Each tier has a separate include pattern matching only its source files
• Each tier has a separate coverage output directory (avoids conflicts)
• CI vs local reporter selection: text-summary locally, full HTML/JSON/LCOV in CI

TypeScript/Vitest example structure:

vitest.unit.config.mts    → tests/unit/**/*.unit.spec.ts    → coverage/unit/
vitest.int.config.mts     → tests/int/**/*.int.spec.ts      → coverage/int/
vitest.browser.config.mts → tests/components/**/*.spec.tsx   → coverage/components/

Python example:

pytest -m unit --cov --cov-report=html:coverage/unit
pytest -m integration --cov --cov-report=html:coverage/int

3. Set initial thresholds

Run coverage once, note the baseline. Set thresholds at the current level - this prevents regression while you improve.

# Example: start where you are
thresholds: { lines: 72 }  # measured baseline

Then ratchet up as you add tests. Never ratchet down. See enforcement for the full ratcheting strategy.

4. Add coverage scripts

Create per-tier scripts in your project manifest:

{
  "test:unit": "vitest run --config ./vitest.unit.config.mts",
  "test:unit:coverage": "vitest run --coverage --config ./vitest.unit.config.mts",
  "test:int": "vitest run --config ./vitest.int.config.mts",
  "test:int:coverage": "vitest run --coverage --config ./vitest.int.config.mts",
  "test:components": "vitest run --coverage --config ./vitest.browser.config.mts",
  "test:e2e": "playwright test",
  "test": "pnpm test:unit && pnpm test:int && pnpm test:components && pnpm test:e2e"
}

Audit & Improve: Closing Coverage Gaps

Phase 1: Audit

Run coverage for each tier and examine the output.

# Run with coverage, examine the HTML report or text output
<runner> --coverage

Identify three categories:

• Untested files - no coverage at all (highest priority)
• Untested branches - code paths never exercised
• Untested functions - declared but never called in tests

Phase 2: Classify each gap

For every uncovered file or function, ask:

Phase 3: Prioritise

Triage order (highest value first):

1. Domain logic and business rules (unit)
2. Access control and authorisation (integration)
3. Data validation and input parsing (unit)
4. API endpoints and mutations (integration)
5. UI components with conditional logic (component)
6. Async/server-rendered components (E2E)
7. Configuration and wiring (tested implicitly by higher tiers)

Phase 4: Write tests

For each gap, follow the appropriate tier's patterns. Test expected behaviour through the public API, not implementation details.

Unit tests: Pure input → output. No database, no network, no filesystem.

describe('slugify', () => {
  it('converts spaces to hyphens', () => {
    expect(slugify('hello world')).toBe('hello-world')
  })
  it('handles empty string', () => {
    expect(slugify('')).toBe('')
  })
})

Integration tests: Real database, real service boundaries, no mocks for things you own.

it('enforces access control on draft posts', async () => {
  const result = await payload.find({
    collection: 'posts',
    where: { _status: { equals: 'draft' } },
    overrideAccess: false,
    user: anonymousUser,
  })
  expect(result.docs).toHaveLength(0)
})

Component tests: Real browser, real DOM queries (accessibility-first via testing-library).

it('renders film title and year', () => {
  render(<FilmCard film={mockFilm} />)
  expect(screen.getByText('Film Title')).toBeInTheDocument()
  expect(screen.getByText('2024')).toBeInTheDocument()
})

E2E tests: Full user flows, real navigation, real network.

test('user can submit a form', async ({ page }) => {
  await page.goto('/submit')
  await page.fill('[name="title"]', 'My Film')
  await page.click('button[type="submit"]')
  await expect(page).toHaveURL(/\/confirmation/)
})

See ecosystem patterns for language-specific runner syntax and config examples.

Enforce: Wiring Coverage into Hooks and CI

Pre-commit (composes with hk)

If using the hk skill, add coverage test steps to hk.pkl:

["test-unit"] {
  check = "pnpm test:unit:coverage"
}
["test-int"] {
  check = "pnpm test:int:coverage"
  depends = List("test-unit")
}

Key principles:

• Coverage thresholds live in the test config, not in hook config
• E2E tests are too slow for pre-commit - run in CI or manually
• Order tiers by speed: unit first (fastest fail), then integration, then components
• Quieting is wrapper-level: run the hook with hk run <hook> -q (hk ≥ 1.51.0, owned by the hk skill), so passing tests produce no output

CI

Run all tiers with coverage in CI. Upload per-tier reports separately for visibility.

- name: Unit tests
  run: pnpm test:unit:coverage
- name: Integration tests
  run: pnpm test:int:coverage
- name: E2E tests
  run: pnpm test:e2e

Ratcheting

For projects not yet at target:

1. Measure current coverage
2. Set threshold at current level
3. After each improvement, bump the threshold
4. Never lower it

See enforcement for detailed CI patterns, PR checks, and ratcheting workflow.

Write Tests for New Code

When adding features to a codebase with established coverage:

1. Identify the tier: What kind of code are you writing? Match to the classification table above
2. Write tests first (TDD): Test the expected behaviour before implementing
3. Run coverage locally: --coverage for the relevant tier
4. Handle exclusions: If code genuinely cannot be tested at this tier, document why and ensure coverage exists at another tier
5. Verify thresholds pass: Pre-commit hooks catch regressions, but check early

Cross-tier exclusion pattern

Every exclusion at one tier names the tier that provides coverage:

// Unit config excludes:
// Cross-tier: Service layer - requires database runtime - tested via integration tests
"src/domain/**/service.ts",

// Integration config excludes:
// Cross-tier: React components - requires browser context - tested via component + E2E tests
"src/components/**",

See coverage exclusions for the full exclusion taxonomy and documentation format.

Test Quality: Beyond Coverage

Coverage is necessary but not sufficient. It tells you which lines ran, never whether a test would fail when the code is wrong. A suite can hit 100% and assert nothing. This section is what to reach for once coverage is high and you need evidence the tests actually catch bugs.

Why a coverage number is not a quality target

• Coverage is weakly correlated with effectiveness. Controlling for test-suite size, coverage is a poor predictor of fault-detection ability (Inozemtseva & Holmes, Coverage Is Not Strongly Correlated With Test Suite Effectiveness, ICSE 2014). High coverage is consistent with a near-useless suite.
• Goodhart's law. "When a measure becomes a target, it ceases to be a good measure." Mandating a coverage percentage incentivises assertion-free tests, deleted edge cases, and excluded hard files - activity that raises the number while lowering quality. (See Seemann, Code coverage is a useless target measure.)
• 100% is a smell, not a goal. Use coverage to find untested code, not as a KPI. Well-tested code tends to land in the 80-90s naturally; a hard 100% mandate signals gaming (Fowler, TestCoverage).

Practical stance for this skill: keep coverage as the regression gate (the Enforce section), and use the techniques below to measure and raise quality. The single most useful quality signal is the oracle gap - high coverage paired with low mutation score flags files where weak tests execute important code (see mutation testing).

Which technique, when

This skill owns the techniques you measure and gate in CI (the left two columns below). The test-design techniques (when/how to write a good test) live in the testing skill - load it for those; this skill only points at them.

The two quality metrics you can gate on each have a dedicated reference here with per-ecosystem tooling and CI/diff enforcement: mutation testing (does the suite detect changes - the strongest quality signal, and the rigorous version of "do my assertions matter") and fuzzing (coverage-guided crash-finding). The oracle gap above is computed from mutation score.

For the design side - property-based testing (invariants over generated inputs, shrinking, stateful/model-based), snapshot/approval pitfalls, differential & metamorphic testing, assertion density, and flaky-test management - see the testing skill (../testing/SKILL.md). That's where "how to write the test" lives; this skill is where "how to measure and enforce it" lives.

Enforcing quality techniques in CI/hooks

These are slower than unit tests - keep them fast and actionable the same way coverage stays fast:

• Diff-scoped: mutate/fuzz only changed code (mutation --in-diff/--incremental/--git-diff-lines; see refs).
• Time-boxed: fuzzing runs a fixed budget (e.g. 5 min smoke test), never open-ended in PR CI.
• Capped & separated: cap mutants per file; run mutation/fuzz as a dedicated CI job after the suite passes, not in the fast pre-commit path. PBT runs inside the unit tier with a bounded example count.
• Thresholds live in the tool, per the Enforce-section boundary - break (Stryker), --min-msi (go-mutesting), exit-code 2 (cargo-mutants), numRuns (fast-check). hk/CI just runs the command and checks the exit code.

Test Organisation Patterns

Directory structure

tests/
  unit/          *.unit.spec.ts       Pure functions, domain logic
  int/           *.int.spec.ts        Database, API, access control
  components/    *.browser.spec.tsx   Rendered UI in real browser
  e2e/           *.e2e.spec.ts        Full user flows
  fixtures/      index.ts             Shared test data factories
  setup/         Per-tier setup files (DB init, browser cleanup)

Naming conventions

Suffix encodes the tier - config include patterns use these suffixes for zero-ambiguity matching:

Test data factories

Use factory functions with auto-incrementing counters for unique identifiers:

let counter = 0
function createTestUser(overrides = {}) {
  counter++
  return {
    email: `test-${counter}@example.com`,
    name: `Test User ${counter}`,
    ...overrides,
  }
}

Counter-based (not random) for deterministic debugging. Reset between test runs if needed.

Mock boundaries

• Do mock: External APIs, third-party SDKs, environment-specific runtimes
• Do not mock: Code you own - test through the public API
• Database: Use a real local database for integration tests (SQLite, test containers)
• Browser: Use a real browser for component tests (Playwright, Vitest browser mode)
• Server-side imports: Stub server-only modules when testing in browser context

Coverage Providers: Quick Reference

Gotchas

References

• Ecosystem Patterns - Index of per-language references:
  • TypeScript/JS | Python | Go | Rust | Merging
• Coverage Exclusions - How to document and justify every exclusion
• Enforcement - Wiring coverage into hk hooks, CI pipelines, and PR checks
• Test quality you measure & gate (beyond line/branch coverage):
  • Mutation Testing - Does the suite detect changes? Per-ecosystem tools, the oracle gap, diff-based CI
  • Fuzzing - Coverage-guided crash-finding (Go native, cargo-fuzz, Atheris, Jazzer, OSS-Fuzz), time-boxed CI
• For test design quality - property-based, snapshot/approval, differential/metamorphic, contract, assertion density, flaky-test management - see the testing skill (../testing/SKILL.md)