Testing

Testing

A test earns its place by failing when the behavior breaks and passing otherwise. Tests that never fail are noise; tests coupled to implementation break on every refactor. Aim for the first, avoid the second.

What to test (and what not to)

• Test behavior and contracts, not implementation details. Assert on what a function returns / what the user sees, not which private helper it called.
• Prioritize by risk × traffic: money, auth, permissions, data integrity, and anything in lib/auth/ get thorough coverage. Trivial getters and framework glue don't.
• Cover the boundaries where bugs live: empty, one, many; null/undefined; zero/negative/overflow; the error path, not just the happy path.
• Every bug fix ships with a test that fails before the fix and passes after. That test is the proof the bug is gone and the guard against its return.
• Don't chase a coverage number. 100% line coverage with no assertions on behavior is theater. Coverage shows what's untested, not what's tested well.

The test pyramid

• Unit (most): pure logic, one module, no I/O — fast, run constantly.
• Integration (some): modules together, real DB/API routes — catch the wiring bugs units miss.
• E2E (few): full user flows — expensive and slower, reserve for critical paths (signup, checkout, auth).
• Invert it and your suite gets slow and flaky. Push each test to the lowest level that can still catch the bug.

Structure

• Arrange–Act–Assert, visibly separated. One logical behavior per test.
• Name tests by the behavior: returns 401 when the session is expired, not test auth 3. The name should read as a spec line.
• Assert on specific values, not just "truthy" or "no throw". expect(res.status).toBe(403) beats expect(res).toBeDefined().
• Keep tests independent and order-agnostic — no test may depend on another having run. Reset shared state (DB rows, mocks, module state) in beforeEach/afterEach.
• Prefer real objects over mocks. Mock only at the true boundary (network, clock, randomness, third-party SDKs). Over-mocking tests your mocks, not your code.

Databases (Postgres)

• Run integration tests against a real Postgres (a disposable/test database), not an in-memory fake — you want the real query planner, constraints, and types.
• Isolate tests: wrap each in a transaction and roll back, or truncate between tests. Never let one test's rows leak into the next.
• Test the migration and the constraint, not just the query: a NOT NULL/UNIQUE/FK is behavior worth asserting.
• Seed the minimum data the test needs, inline and visible — a giant shared fixture hides what the test actually depends on.

Async, time, and flakiness

• await every promise and assert after it resolves; a floating promise makes a green test that proves nothing.
• Never sleep(n) to wait for async — await the actual signal (the resolved value, the state change, a polled condition). Fixed sleeps are the #1 source of flakiness.
• Freeze time and seed randomness. Tests that read the real clock, timezone, or Math.random() fail intermittently.
• A flaky test is a bug — in the test or the code. Quarantine it loudly, then fix the race; don't paper over it with retries.

Working with the harness

• The Stop gate runs pnpm test before letting a task finish, so a red suite blocks completion — keep it green as you go, not at the end.
• Run the focused test while iterating (pnpm test / -t ""), the full suite before you consider the work done.
• When you change behavior deliberately, update the test in the same change and say why in the diff — don't delete a failing test to get green.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: ksed8
• Source: ksed8/cc-loopkit
• License: MIT

Install and usage instructions live in the source repository linked above.