Modes

A gate's mode says what kind of test it is — independent of layer (which says which philosophy it serves). See /testing/pyramid for the layer mapping.

Static layer

• static — file-level analysis. Route coverage, schema-runtime parity, lint-shaped invariants. The bread and butter of contract gates.
• drift — generated-file vs. source comparison. OpenAPI ↔ MCP catalog, TS tokens ↔ CSS, Prisma schema ↔ generated client.

Unit layer

• unit — pure-function tests with specific inputs. Vitest-shaped.
• property — fast-check generative tests. "For all inputs in domain D, predicate P holds."
• fuzz — adversarial / random inputs. Catches what enumerated tests don't.
• mutation — Stryker-shaped. Mutates code, verifies tests catch the mutation. Proves your tests can catch bugs.

Integration layer

• integration — multiple units composed. API + DB, MCP + spec, dispatcher + router.
• contract — consumer-provider contract tests (Pact-shaped).
• visual — pixel regression (Chromatic / Percy).

E2E layer

• e2e — real-browser user flows (Playwright).
• smoke — single-page health checks on a running app.
• synthetic — production canary / probe.

AI / Evals layer

• eval — capability + behavior baselines on AI emissions.
• safety — alignment / refusal / constitutional checks.
• red-team — adversarial prompts (jailbreak, prompt injection, data extraction).
• prompt-regression — captured-prompt replay against current model + prompt.
• hallucination — factuality / citation grounding against authoritative oracles.
• llm-judge — LLM-as-judge for soft outputs (voice, style, tone).
• determinism — same input → same output. For structured-output flows.