Test Strategy

Test Strategy Skill

A test suite is a product: its users are future changes. Optimize for fast, trustworthy feedback — tests that fail when behavior breaks and stay green when only implementation moves. Language-agnostic; Go + TypeScript used as reference stacks.

When to use

Setting up testing for a new project or module, deciding which test type a piece of code needs, improving coverage that exists but proves nothing, or diagnosing a slow/flaky suite.

The test pyramid (normative shape)

Inverted pyramids (many E2E, few unit) produce slow, flaky, undebuggable suites. If a behavior is testable at a lower level, test it there.

What to test where — decision table

Bug fixes always get a regression test that fails before the fix — no exceptions.

Test quality rules (normative)

• Assert behavior, not implementation: test through the public API; never assert "method X was called" when you can assert the observable outcome.
• One behavior per test; the name states scenario + expectation (TestDiscount_ExpiredCoupon_Rejected, it("rejects an expired coupon")).
• Independent and order-free: each test creates its own fixtures; no shared mutable state; parallel-safe (t.Parallel(), default in vitest).
• Deterministic: inject clocks, seed randomness, never sleep to wait — poll or use synchronization.
• Arrange–Act–Assert visible in every test body; extract builders for noisy setup.
• Tests mirror source structure and live next to (Go) or beside (__tests__/*.test.ts) the code they test.

Coverage policy

Coverage is a gap detector, not a goal. Measure it (go test -cover, vitest run --coverage), look at what is uncovered, and decide if it matters. A hard threshold (e.g. 80%) invites assertion-free tests written to pass the gate; instead enforce: critical-path packages reviewed for uncovered branches in [[code-review]].

Anti-patterns

• Testing implementation details — suite breaks on every refactor ([[refactor-safely]] depends on this not happening).
• Mocking everything, including the thing under test's collaborators' collaborators — green tests, untested system.
• Shared mutable fixtures (global test DB rows, package-level vars) — order-dependent flakiness.
• Sleep-based waits and real-clock assertions — flaky by design.
• Assertion-free or snapshot-everything tests that pass no matter what.
• One E2E per acceptance criterion — push them down the pyramid.

Verification (executable)

The strategy is enforced by commands, not intentions: go test -race -cover ./... (Go), vitest run --coverage or jest --coverage (TS) in CI on every PR; integration tests gated behind a tag/profile so the unit loop stays under 30s; flakiness surfaced with go test -count=10 / vitest --retry=0. Full tooling per ecosystem and the flaky-test triage checklist are in reference.md.

Going deeper

• reference.md — per-ecosystem tooling table, suite-layout conventions, flaky-test triage, mutation testing.
• examples.md — table-driven unit test, integration test with testcontainers, contract test, in Go and TypeScript.