Test Backfill

Test Backfill

Raise coverage without shipping bad tests.

Process

1. Read existing tests. Learn the framework, assertion style, fixture patterns, naming convention. Never introduce a second style.
2. Run coverage. Capture the report. Identify the gaps.
3. Prioritize, in order:
   • Public API surface (exported functions, HTTP endpoints, CLI commands)
   • Error paths (throws, rejections, non-200s, invariant violations)
   • Edge cases (null, empty, boundary, concurrent)
   • Happy path (least valuable — usually already tested)
   • UI details (lowest priority, hardest to maintain)
4. Write tests in batches of 5–10. Run the suite after each batch.
5. If a batch fails, stop and investigate before writing more.
6. Every new function gets at least one failure-path test.
7. Hand off coverage numbers at the end: "coverage now X%, was Y%, Δ = +Z pp".

Quality Bar

A good test:

• Tests behaviour, not implementation.
• Reads like a specification: "when X, the function does Y".
• Has one assertion per logical check (but may have multiple physical assertions).
• Runs in < 100 ms unless it's explicitly an integration test.
• Is deterministic. No wall-clock, no network, no RNG — or inject them.

A bad test:

• Tests internal state / private methods directly.
• Mocks so much of the system under test that it's not really testing anything.
• Depends on test execution order.
• Has "magic" fixture values with no explanation.
• Uses try/catch instead of expect().toThrow() (or equivalent).

Anti-Patterns to Kill

• Tests that always pass (e.g., expect(true).toBe(true) stubs)
• Tests with no assertions (function runs, nothing's checked)
• Tests that catch all exceptions silently
• sleep(N) — use a proper wait-for-condition utility
• Tests that talk to the real network or real database without a @integration tag

Never

• Modify production code to make a test easier, beyond adding a dependency-injection seam.
• Weaken an assertion to green a test.
• Mark a test .skip without a dated TODO and an owner.
• Copy-paste a test and flip one value without asking "is this actually a different case?"
• Inflate coverage with meaningless tests. 78% with real tests > 85% with theater.