Coverage Analysis

<objective>
A suite at 90% line coverage with assertion-free tests catches zero bugs — line coverage
proves code ran, not that a regression would be caught. This skill measures the right
things (branch coverage, mutation score, critical-path coverage), gates them in CI with a
ratchet so coverage can only go up, and surfaces gaps by risk instead of chasing a vanity
number. It prevents the classic failure: a green coverage badge over a test suite that
never asserts anything meaningful, while the payment module sits at 30%.
</objective>

Quick Route

Discovery Questions

Check .agents/qa-project-context.md first — if it exists, use it and skip anything already answered there. Then:

1. What test runner and coverage tooling is configured? Check for vitest.config.* (coverage block), jest.config.* (coverageProvider), .nycrc, c8 in scripts, or [tool.coverage] in pyproject.toml. The runner decides the install — Vitest pulls @vitest/coverage-v8, not c8 (see Coverage Tools).
2. What is the current coverage level? Run the existing coverage command and note line, branch, and function percentages. This is the baseline for the ratchet.
3. Is coverage gated in CI? Check GitHub Actions / GitLab CI for --coverage, coverageThreshold, fail_under, or --cov-fail-under. No gate means coverage is decorative.
4. What is the target, and who set it? A target without rationale ("the VP said 80%") leads to gaming. Targets should reflect risk tolerance and codebase maturity, not a round number.

Core Principles

1. Coverage measures breadth, not depth. A line being executed does not mean it is tested correctly. expect(true).toBe(true) executes the function but asserts nothing. Coverage tells you what code ran, not whether the tests would catch a bug — that is what mutation testing measures.

2. Branch coverage matters more than line coverage. A ternary condition ? a : b on one line counts as fully covered in line coverage even if only one branch ran. Line coverage does not guarantee branch coverage. Gate on branches, not just lines:

function discount(price: number, isPremium: boolean): number {
  return isPremium ? price * 0.8 : price;
}

// Line coverage: 100% (the line executed). Branch coverage: 50% (only the true branch ran).
expect(discount(100, true)).toBe(80);

// Fix — assert BOTH paths so branch coverage reaches 100%:
expect(discount(100, true)).toBe(80);
expect(discount(100, false)).toBe(100);

3. Ratchet pattern: never decrease, only increase. Record current coverage as the minimum threshold. Every PR must meet or exceed it. Coverage climbs over time without forcing an artificial target up front.

4. Focus on gaps by risk, not on the number. A project at 85% is not automatically "better" than one at 75%. What matters is whether the untested slice contains payment, auth, or data-integrity logic. Analyze gaps by risk.

5. New code has a higher bar than legacy code. Require 90%+ on new code in PRs even if the project sits at 65%. This stops coverage decay without demanding a rewrite of legacy code.

Coverage Tools

Pick the provider by test runner first, then by how cleanly it maps to your build output.

Two engines underneath:

• V8 coverage — built into the V8 engine, so it does not instrument source: faster, no Babel transform. For Vitest, install @vitest/coverage-v8 (NOT c8 — that is the standalone CLI for non-Vitest runners). For a plain node:test or mocha project, c8 is the CLI wrapper around the same V8 data. Default for new Node/Vitest projects.
• Istanbul — instruments source code; slower but maps more reliably through transpilers and bundlers. Switch to it (@vitest/coverage-istanbul, or nyc) when V8 maps poorly. Symptom that V8 maps poorly: reported uncovered lines land on blank lines, closing braces, or decorators, or whole covered functions show as red — that means the source map is misattributing lines (common with certain TS bundlers / SWC configs). When you see that, flip to the Istanbul provider.

Node baseline: c8 11.x and nyc 18.x are current. c8 11 still supports Node >=12; nyc 18 requires Node 20 || >= 22. If you must stay on Node 18, pin nyc@^17 (c8 11 runs fine on Node 18). New projects should standardize on Node 20+.

See references/tool-config.md for the full provider configs (Vitest coverage block, .nycrc.json, Jest coverageThreshold, pyproject.toml / .coveragerc.toml), install commands, and run invocations.

Merging coverage across test types

Unit, integration, and E2E runs each produce partial coverage. Combine them so a line covered only by an integration test isn't reported as a gap:

• Vitest — run as multiple projects/configs and let Vitest merge, or merge coverage-final.json outputs.
• nyc — nyc merge .nyc_output merged.json && nyc report -t merged combines .json files from separate runs.
• coverage.py — coverage combine after running each suite with coverage run -p.

Merge first, gate on the merged total. Don't gate each suite's coverage in isolation.

Coverage Report Types

Gap Analysis

Coverage reports show which lines and branches did not run. Not all gaps are equal — prioritize by risk.

Step 1: Generate the report.

npm run test:coverage
# Open coverage/index.html in a browser

Step 2: Sort files by uncovered lines. Parse coverage-summary.json, sort files by (total - covered) descending, focus on the top 20. A small script that reads the JSON summary and outputs file / line% / branch% / uncovered-count makes this repeatable.

Step 3: Map gaps to risk.

Include branch coverage in the sort, not just lines — a file at 100% line / 50% branch hides untested paths a line-only sort would rank as "done."

Coverage as CI Gate

Threshold Configuration

Set a global threshold as the project minimum, then layer per-directory thresholds stricter for critical code (payments, auth) than for utilities. Vitest uses glob keys under thresholds (e.g. "src/payments/**": { lines: 95, branches: 90 }); Jest uses path keys under coverageThreshold. Both support per-path overrides.

See references/ci-gating.md for the global, per-directory, and Jest per-file threshold config.

Ratchet Pattern

Never let coverage decrease. Record the current level as the minimum and floor it upward when coverage improves. A ratchet script reads coverage-summary.json, compares each metric against a committed .coverage-ratchet.json, fails the build on any regression, and updates the baseline when coverage improves.

Commit .coverage-ratchet.json (e.g. { "lines": 82, "branches": 78, ... }). In CI, run the ratchet script after tests. On main-branch merges, auto-commit the updated ratchet file if coverage improved.

See references/ci-gating.md for the full coverage-ratchet.ts script.

PR Diff Coverage Gate

Require new code in a PR to meet a higher threshold (e.g. 90%) than the project baseline. In CI, use git diff --name-only origin/main...HEAD to identify changed files, then check their coverage from coverage-summary.json. Fail the pipeline if changed-file coverage falls below the threshold. This stops decay without rewriting legacy code.

Surface the diff to reviewers with davelosert/vitest-coverage-report-action@v2 (reads the JSON summary) or marocchino/sticky-pull-request-comment@v2 with a script that filters to changed files. See references/ci-gating.md for the full PR workflow.

Hosted alternatives: Codecov, Coveralls, and Trunk Coverage ship first-class differential PR coverage with merge-blocking gates and inline annotations. Most teams prefer these over hand-rolled diff scripts — pick one if you don't already have a coverage host. Codecov + GitHub: codecov/codecov-action@v5 reads lcov.info and posts a PR diff comment automatically.

Mutation Testing

Mutation testing measures assertion quality, not just code execution. It mutates your source (flips a > to >=, deletes a line) and checks whether a test fails. A surviving mutant means a real bug your tests would miss. With Stryker JS v9.6+ and Vitest 4.1+ the cost is low enough to run on PR-changed files; mutmut 3.x covers Python.

Targeting

Mutation testing is expensive on whole codebases — run it incrementally. Stryker's incremental: true (JSON cache) plus --mutate scoped to the git diff re-mutates only touched files; mutmut similarly mutates per path. Restrict to:

• Pure business logic (validators, calculators, transformers)
• Critical paths (payment, auth, data integrity)
• Code with high line coverage but suspect assertions (branch coverage > 90% but few assertion variants)

Skip UI rendering, glue code, and generated code.

See references/mutation-testing.md for the Stryker config (stryker.config.json, incremental, run on changed files only) and the mutmut invocation.