Mutation Proof

Coverage says the line ran. Mutation says the test would have noticed. Only the second is quality.

Level 1 — Targeted proof (always, per test, seconds)

For each new or suspect test:

1. Identify the exact production line the test exists to defend.
2. Apply one mutation to it:

3. Run only that test. Record the result.
4. Red -> restore the code, the test is proven, note the mutation in the commit message. Green -> the test is a false positive. Fix the assertion, do not fix the mutation.
5. Restore the file and re-run to confirm you are back to a clean tree (git diff --exit-code).

Never leave a mutation behind. Verify with git status before finishing — a forgotten mutation shipped to production is the one way this skill can hurt you.

Level 2 — Tool-driven mutation testing (per module, minutes)

Scope it to P0 code only. Whole-repo mutation runs are slow, noisy, and get switched off within a week. Set a threshold on the P0 module and let the rest report without gating:

// stryker.conf.json
{ "mutate": ["src/checkout/**/*.ts"], "thresholds": { "high": 90, "low": 80, "break": 75 } }

Reading the survivors

Each surviving mutant is a sentence: "we can change this behavior and nobody complains". Triage into three buckets:

• Real gap -> write the missing assertion.
• Equivalent mutant (behaviorally identical, e.g. a swapped order in a commutative op) -> annotate and exclude with a comment naming why.
• Dead code -> the mutation survives because nothing depends on the line. Delete the line.

Interaction with other skills

unit-test-authoring, api-contract-testing and e2e-authoring all end here. A test that has not been through Level 1 does not count as done, and the orchestrator will not report it as coverage.

Output

PROVEN  tests/checkout.spec.ts:42  <- mutated charge.ts:88 (`>=` -> `>`)  -> 1 failed  ✓
UNPROVEN tests/cart.spec.ts:10     <- mutated cart.ts:31 (return 0)       -> 0 failed  ✗ false positive
Tree restored: git diff --exit-code -> clean