Measure and enforce test coverage with Istanbul/nyc, c8, Jest, and Vitest. Covers branch versus line coverage, per-directory thresholds, CI gates, and correctly excluding generated code from reports.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Measure and enforce test coverage with Istanbul/nyc, c8, Jest, and Vitest. Covers branch versus line coverage, per-directory thresholds, CI gates, and correctly excluding generated code from reports.
This skill makes an AI agent configure coverage collection correctly (Istanbul instrumentation or V8 native coverage), set thresholds that fail builds, read coverage reports to find genuinely untested branches, and exclude generated or config code so the numbers mean something. Trigger it when a user asks "what is our coverage", wants a coverage gate in CI, or when a coverage/ directory, --coverage flag, .nycrc, or coverageThreshold appears in the project.
Core Principles
Branch coverage is the number that matters. Line coverage marks a line as hit even if only one of its outcomes ran. A guard clause if (a && b) return x; can be 100 percent line-covered with three of its four branch outcomes untested.
Coverage is a gap detector, not a quality score. 95 percent coverage with assertion-free tests proves nothing. Use coverage to find untested code paths; use mutation testing to check assertion strength.
Thresholds must fail the build. A coverage report nobody reads is decoration. Wire coverageThreshold (Jest), thresholds (Vitest), or --check-coverage (nyc/c8) so CI goes red on regression.
Set thresholds at current reality, then ratchet. Dropping a 90 percent global gate onto a 60 percent codebase makes the team delete the gate. Start at today's number minus 1, raise it as coverage improves.
Measure all files, not just imported ones. By default some tools only report files touched by tests, so a completely untested module is invisible. Enable all: true (nyc), coverage.all (Vitest), or a broad collectCoverageFrom (Jest).
Exclude generated code explicitly. Protobuf stubs, GraphQL codegen, migrations, and *.d.ts inflate or deflate numbers randomly. Exclude them in config, never by sprinkling ignore comments through generated files.
Setup and Patterns
1. Jest: collection, thresholds, and per-directory gates
// shipping.tsexportfunctionshippingCost(country: string, total: number): number {
if (country === 'US' && total > 50) return0;
return country === 'US' ? 5 : 15;
}
// shipping.test.ts -- this single test yields 100% LINE coverageimport { shippingCost } from'./shipping';
it('ships free for large US orders', () => {
expect(shippingCost('US', 100)).toBe(0);
});
// But branch coverage shows: total <= 50 untested, non-US untested,// the $5 domestic path untested. Three real behaviors have no test.
The branch report (text reporter prints % Branch per file, the HTML report highlights yellow I/E markers) is how you find these.
6. Ignoring unreachable code the honest way
// Istanbul-instrumented runners (nyc, babel-plugin-istanbul)/* istanbul ignore next -- @preserve defensive guard, unreachable after zod validation */if (typeof input !== 'string') thrownewTypeError('input must be a string');
// V8-based runners (vitest --coverage.provider=v8, c8)/* v8 ignore next 2 */if (process.platform === 'win32') {
pathSeparator = '\\';
}
Every ignore comment needs a reason suffix; an ignore without a justification is a coverage lie waiting to rot.
Review the HTML report (coverage/index.html) when writing tests for legacy code; the red/yellow highlighting finds untested branches faster than reading source.
Track coverage trends per package in a monorepo rather than one blended global number that averages away problems.
Gate on json-summary totals for speed, keep lcov output for editors and review tools that show per-line coverage in diffs.
Prefer the V8 provider for TypeScript projects; Istanbul instrumentation of transpiled output can misattribute branches to source maps.
Delete dead code instead of ignoring it; an istanbul ignore on reachable code is technical debt with a receipt.
For bug fixes, check the coverage diff: the fixed lines must be covered by the new regression test.
Anti-Patterns
Chasing 100 percent: the last few points usually buy tests for getters and logging branches while real risk lives in untested integration seams.
Writing tests that execute code without asserting anything, purely to move the number. Mutation testing will expose these instantly.
Excluding files from coverage because they are "hard to test" - that is the highest-risk code in the repository.
A global threshold so low (40 percent) that it never fires; gates that cannot fail teach the team to ignore them.
Measuring coverage only on unit tests when most behavior is exercised by integration tests; merge reports (nyc merge, --coverage.reportsDirectory per suite plus lcov merge) before judging.
Letting coverage/ get committed; add it to .gitignore.
When to Trigger This Skill
The user asks "what is our test coverage", "add a coverage gate", "why is coverage dropping", or "enforce 80 percent coverage".
A PR adds --coverage, coverageThreshold, .nycrc, c8, or @vitest/coverage-v8 to the project.
CI needs a quality gate tied to lines/branches/functions, or thresholds need ratcheting after a coverage push.
Generated code (GraphQL codegen, protobuf, ORM migrations) is skewing the numbers and needs principled exclusion.
Pair it with mutation testing when the user suspects high coverage numbers but weak assertions.
