Go Test Quality

Tests are production code. They run in CI on every commit, they document behavior, and they're the first thing you read when a function breaks at 3am. Write them with the same care you'd give to code that handles money.

Community default. A project skill or agent-config rule that explicitly supersedes go-test-quality takes precedence over this one.

Detailed reference material, loaded on demand:

• references/helpers-and-fixtures.md — test helpers, factory functions with options, t.Cleanup, golden files, mock implementations.
• references/integration-testing.md — httptest recorder and server, testcontainers, build tags, TestMain, fuzz testing.
• references/concurrency-and-benchmarks.md — synctest, goleak, t.Context(), b.Loop() benchmarks, t.Attr and artifacts.

Read a reference file only when the summary below is not enough.

1. Test Design Philosophy

Test behavior, not implementation

// ✅ Good — tests what the function DOES
func TestTransferFunds_InsufficientBalance(t *testing.T) {
    from := NewAccount("alice", 100)
    to := NewAccount("bob", 0)

    err := TransferFunds(from, to, 150)

    require.ErrorIs(t, err, ErrInsufficientFunds)
    assert.Equal(t, 100, from.Balance(), "sender balance should be unchanged")
    assert.Equal(t, 0, to.Balance(), "receiver balance should be unchanged")
}

// ❌ Bad — tests HOW the function does it
// asserts debit() was called before credit(), rollback() was called,
// internal mutex was locked — breaks on every refactor

One assertion per logical concept

Multiple assert calls are fine when they verify different facets of the SAME behavior (both accounts after a transfer). A test that checks creation AND update AND deletion is three tests pretending to be one.

Name tests like bug reports

When the test fails, the name alone should say what broke:

// ✅ Good — reads like a sentence
func TestOrderService_Cancel_RefundsPartiallyShippedItems(t *testing.T) { ... }
func TestParseConfig_ReturnsErrorOnMissingRequiredField(t *testing.T) { ... }

// ❌ Bad — says nothing useful
func TestCancel(t *testing.T) { ... }
func TestRateLimiter_Success(t *testing.T) { ... }

Layout

• One _test.go per source file (parser.go → parser_test.go), tests in the order of the functions they cover — tools and reviewers navigate by file.
• Public-API tests live in the pkg_test package so they can only use what callers can; white-box tests stay in pkg.
• Integration tests sit behind //go:build integration.

2. Subtests for Organized Scenarios

Use t.Run to group related scenarios under a parent test. Each subtest gets its own setup, its own failure, and its own name in CI output:

func TestUserService_Create(t *testing.T) {
    svc := setupUserService(t)

    t.Run("succeeds with valid input", func(t *testing.T) {
        user, err := svc.Create(ctx, CreateUserInput{Name: "Alice", Email: "alice@example.com"})
        require.NoError(t, err)
        assert.NotEmpty(t, user.ID)
    })

    t.Run("rejects duplicate email", func(t *testing.T) {
        _, _ = svc.Create(ctx, CreateUserInput{Name: "Alice", Email: "taken@example.com"})
        _, err := svc.Create(ctx, CreateUserInput{Name: "Bob", Email: "taken@example.com"})
        require.ErrorIs(t, err, ErrDuplicateEmail)
    })
}

3. Test Helper Rules

1. Always call t.Helper() in test utilities so failures point to the caller, not the helper.
2. Factory functions with functional options for complex test objects — defaults with per-test overrides, never a 15-parameter constructor.
3. Prefer t.Cleanup over defer — it runs even after t.FailNow() and is scoped to the test, not the function.

Full implementations in references/helpers-and-fixtures.md.

4. Choosing the Test Type

5. Mocking Rules

• Interface-based hand-written mocks for small interfaces (≤3 methods): a struct with function fields plus recorded calls.
• Function injection for simple seams (now func() time.Time).
• Do NOT mock: value objects, pure functions, the standard library, or your own code in the same package. Test the real thing.
• If you mock everything, you're testing your mocks, not your code.

6. Parallelism and Coverage

func TestSlugify(t *testing.T) {
    t.Parallel() // safe: pure function, no shared state
    // ...
}

Do NOT use t.Parallel() when tests share mutable state, databases, files, or process-level state — t.Setenv and t.Chdir panic in a parallel test for exactly that reason.

go test -race -coverprofile=coverage.out -coverpkg=./... ./...
go tool cover -html=coverage.out   # read the red lines, not the percentage

Coverage finds untested code; it does not measure whether the assertions are any good. Use it to locate gaps, and let CI enforce a floor that only ratchets up (the go-ci skill owns the gate). Without -coverpkg, a package exercised only by another package's tests reports 0%.

7. Testing Concurrent Code

• Never synchronise a test with time.Sleep — it is slow when it works and flaky when it does not.
• Go 1.25+: run time-dependent code inside synctest.Test. The bubble's fake clock advances instantly once every goroutine is blocked, and synctest.Wait() replaces "sleep until the background goroutine gets there". On Go 1.24 the API is synctest.Run behind GOEXPERIMENT=synctest. Go 1.27 adds synctest.Sleep(d) (advance and settle in one call) and httptest.NewTestServer(t, h), an in-memory server that works inside the bubble and closes itself.
• Any package that starts goroutines verifies none leak: goleak.VerifyTestMain(m) in TestMain, or defer goleak.VerifyNone(t). -race finds data races; it does not find a goroutine that never exits.

8. Test Context, Benchmarks, and Artifacts

• t.Context() (Go 1.24+) instead of context.Background() — it is cancelled before t.Cleanup runs, so blocked work unwinds when the test ends.
• Benchmarks use for b.Loop() (Go 1.24+): setup runs once, results stay alive without a sink. Add b.ReportAllocs() and compare runs with benchstat.
• t.Attr (Go 1.25+) emits key/value pairs CI can parse; t.ArtifactDir() with go test -artifacts (Go 1.26+) keeps failing golden output or captured bodies outside the repository.

Examples for both sections: references/concurrency-and-benchmarks.md.

9. Assertion Library Traps

• Build assert.New(t) / require.New(t) inside each t.Run. An instance created from the parent t reports a subtest's failure against the parent, and the failing subtest itself prints PASS.
• require, t.Fatal and t.FailNow only on the test goroutine. From a spawned goroutine they cannot stop the test. Send results back on a channel and assert after receiving. go vet's testinggoroutine and golangci-lint's testifylint catch most cases.
• require for preconditions, assert for independent checks — a failed precondition should stop the test before it nil-panics.
• assert.Eventually hides timing bugs in polling; prefer synctest.

Anti-Patterns

• 🔴 Test with no assertions — always passes, proves nothing
• 🔴 time.Sleep for synchronization — use testing/synctest, channels, or polling
• 🔴 Test depends on execution order — each test must stand alone
• 🔴 Mocking everything — you end up testing your mocks, not your code
• 🟡 Test names like Test1, TestSuccess — name the scenario
• 🟡 Reaching into private fields — test through the public API
• 🟡 No edge cases: empty, nil, zero, max values, unicode
• 🟡 Giant shared setup — each test should set up only what it needs
• 🟡 context.Background() in a test — use t.Context()
• 🟡 for range b.N on Go 1.24+ — use for b.Loop()
• 🟢 Fuzz anything that takes untrusted input
• 🟢 Golden files for complex output comparisons

Verification Checklist

1. Every test has meaningful assertions (no empty test bodies)
2. Test names describe the scenario, not the method
3. t.Helper() called in every test utility function
4. t.Cleanup() used for resource teardown
5. t.Parallel() used where safe, avoided where not
6. Integration tests guarded with testing.Short() or build tags
7. Mocks are minimal — only mock external dependencies
8. No time.Sleep for synchronization anywhere in the suite
9. Packages that start goroutines verify with goleak
10. Benchmarks use for b.Loop() and call b.ReportAllocs()
11. Edge cases covered: empty, nil, zero, boundary values
12. go test -race ./... passes
13. Coverage is used to find gaps; CI enforces a ratcheting floor, not a target
14. testify instances are built per subtest; no require off the test goroutine