write-tests

Author tests directly from immutable contracts like OpenAPI specs, Protobuf schemas, or feature tickets before writing any implementation code. Then run them against the baseline codebase to confirm they fail deterministically, proving the tests verify the specification rather than the implementation.

Property-based testing verifies mathematical invariants like round-trip serialization and idempotency across thousands of generated inputs instead of hardcoded examples. Use fast-check for TypeScript, Hypothesis for Python, RapidCheck for C++, and proptest for Rust, all with automatic counter-example shrinking.

Inject network faults in tests using Shopify/toxiproxy between the client and a mock service, adding latency, jitter, or connection resets. Then assert the circuit breaker transitions from CLOSED to OPEN to HALF-OPEN and verify fallback handling and single-canary probe recovery.

Line coverage only proves code was executed, not that assertions verify correctness. Mutation testing introduces deliberate faults like inverted conditionals, and critical packages should kill at least 75-80% of mutants using tools like Stryker, mutmut, or cargo-mutants.

Run each subtest inside a GORM transaction and roll it back in cleanup, so no committed data leaks between tests. Combine this with ephemeral Testcontainers PostgreSQL instances and never use global database handles.

No. Tests must run in an air-gapped sandbox with no network access, stubbing HTTP dependencies with MSW v2 or local fixtures. For LLM features, record real responses once as VCR cassettes and replay them deterministically in CI.