testing-without-tautologies

Testing Without Tautologies

Core Idea

Tests should fail when protected behavior breaks. A passing test helps only if it can catch a real problem.

Before writing or changing a test, ask: "What production change should make this test fail?" If you cannot answer, redesign the test.

Quality Gate

Before writing the test body, answer these:

• Who uses this? Prefer public APIs, HTTP responses, SSE events, persisted rows, parsed session output, rendered UI, or CLI output. Avoid private state.
• What example proves it? Use concrete inputs and literal expected outputs. For parsers, hand-write the fixture JSONL and the expected sessions/messages; do not run the parser to generate expectations.
• What break would this catch? Name the wrong branch, missing side effect, wrong argument, boundary case, or contract violation.
• Do we own it? Test our choices at framework, database, and filesystem boundaries. Do not re-test documented dependency mechanics.
• Can you state it? Given this setup, when the user/system does X, then Y observable behavior changes. If Y is not assertable, the test is not ready.

Required Checks

Apply these checks to every new or modified test:

1. Assert observable effects

   • Check returned values, rows in the DB, HTTP status and body, emitted SSE events, parsed message content, rendered DOM, or errors.
   • A no-assertion test is acceptable only when the failure mode is the subject, such as "this config rejects invalid input." Prefer explicit assertions anyway.
   • Use require.X for checks that must abort the test (setup errors, nil receivers, length before indexing) and assert.X for independent checks. Never if got != want { t.Fatalf(...) } in new tests.

2. Make fakes and mocks specific

   • Verify arguments, call counts, order, and branches when they are part of the contract.
   • Do not let a fake accept any input when the code must pass one value.