saltbo
2.5 KiB
Test pyramid and development selection
Test a behavior at the lowest layer that proves it through a stable boundary. Add a higher-layer case only for risk owned by that boundary.
| Layer | Owns |
|---|---|
| Domain unit | deterministic rules, transitions, predicates, boundary values |
| Application unit | use-case orchestration, port failures, retries, races |
| Component unit | rendered states, local interaction, accessibility contracts |
| Adapter integration | D1 constraints and external request/response translation |
| HTTP integration | auth, tenant/scope denial, validation, headers, Problem Details |
| HTTP contract | OpenAPI declaration and representative runtime agreement |
| Toolbox acceptance | discovery, generated commands, DPoP, deployed resource access |
| Browser E2E | a small set of critical human journeys |
New unit and component tests live in tests/unit/. Real boundaries live in
tests/integration/; schema agreement lives in tests/contract/; deployed
Toolbox journeys live in tests/acceptance/; browser journeys live in
tests/e2e/. Root-level tests are legacy and are migrated when their owning
feature changes.
Before running tests during development:
- name the observable behavior at risk;
- select the lowest owning layer;
- run the exact named cases that prove it;
- after a failure, rerun that case and directly affected cases only;
- run a complete file only when every case in it is relevant.
For each changed behavior, select the applicable case families rather than copying a fixed matrix into every layer:
- successful behavior and stable output;
- invalid input and business rejection;
- authentication, scope, actor, and cross-tenant denial at a protected edge;
- external dependency failure and unknown outcome at application or adapter boundaries;
- retry, idempotency, races, and partial-effect recovery for writes that cross a boundary;
- schema constraints and forward application when storage changes;
- loading, empty, error, and primary interaction states for UI changes;
- one end-to-end journey only when multiple deployed boundaries create risk that lower layers cannot prove.
Not every change needs every family. When a relevant boundary cannot yet prove an applicable family, record the missing coverage and reason instead of hiding the gap or adding a weaker duplicate test.
Tests are deterministic and do not use fixed sleeps, retry-to-green, or shared
mutable fixtures. See
ADR 0010 and
spec/README.md.