Testing

Gate every agent/LLM release on two checks: a quality eval gate and a safety gate, both run in CI on every model and prompt change.

Prioritize unit tests over integration tests. Test state transitions, edge cases, serialization round-trips. Every im...

Verify that independently deployed services agree on the shape of their interactions before they ship.

Prescriptive rules for testing SQLite-backed code: in-memory vs file-based databases, test isolation strategies, migration testing, sync logic testing, and conflict resolution testing.

Use design-time data contexts for visual preview testing — verify UI renders correctly without running the app.

When the unit under test is agent or LLM behavior, build a calibrated eval harness instead of relying on ordinary unit tests.

Flaky tests destroy confidence. Quarantine them immediately — fix or delete, never ignore.

Detect, quarantine with an owner and deadline, fix the root cause, then return flaky tests to the gating suite.

Fuzz untrusted-input parsing surfaces — especially in memory-unsafe code — with coverage-guided fuzzers; adopt elsewhere only on measured need.

Evaluate RAG groundedness with verified citations, calibrated LLM judges, and retrieval metrics, and abstain when context is insufficient.

Run linting as part of your automated test and verification suite — it catches bugs that unit tests miss.

Mutation testing validates that your tests actually catch bugs — not just achieve coverage.

Every generated artifact MUST be verified:

All UI components MUST include preview declarations for rapid visual verification during development. Previews should...

From Kent Beck's Test Desiderata — tests should be:

When to use: parsers, serializers, data transformers, encoders/decoders, validators — anything

Run security scans as part of post-generation verification (agenticdevelopercookbook://guidelines/testing/post-generation-verification). These a...

Keep snapshots small and deterministic, and deliberately review every snapshot diff before accepting it.

**Construct what you need, per test.** Large shared fixture files SHOULD be avoided.

Use [Martin Fowler's taxonomy](https://martinfowler.com/bliki/TestDouble.html):

Projects SHOULD follow the Google SWE Book ratio: **80% unit / 15% integration / 5% E2E**.

The recommended Claude Code testing workflow, combining all tools:

Evaluate agent tool use for selection, argument correctness, ordering, and stopping against a fixed trajectory suite with error-recovery cases.

**Structure — Arrange, Act, Assert (AAA):**