Skip to main content

​
Context

0022 says where tests live and that a green pnpm verify proves less than it seems; it does not say which kinds of test exist or how much is enough. Tenant isolation is the failure that costs most, and a mock database cannot show it. The repository is public, so GitHub-hosted runners are free.

​
Decision

  1. Vitest covers logic in packages, services, web and mobile.
  2. Maestro covers mobile E2E, on GitHub-hosted runners; a local Maestro run is the fallback. Flows sit with the app.
  3. Playwright covers web. A shared preset runs a smoke suite on every PR and the full suite nightly. Suites live in apps/<app>/e2e/__tests__/ (0022 rule 3).
  4. Service tests use encore test against a real Postgres. Each tenant-owned table has a mandatory isolation test: tenant A cannot read or write tenant B’s rows.
  5. Coverage floors: 80 and 70 for packages and services. Reading them as lines and branches is an assumption (TODO confirm). Apps have no floor.
  6. Generated clients are type-checked; screenshots serve visual checks; no Storybook.
  7. Accessibility and load: axe in the Playwright suite, k6 for load (0070).
  8. No self-hosted runners on the public repository. Heavy agent runs use a personal box, not CI.

​
Consequences

  • Real-Postgres tests are slower than mocks; they are the only evidence for tenancy.
  • A floor can be gamed; review still reads the tests.

​
Enforcement

Planned: coverage thresholds in @stackra/vitest-config, a check that every tenant-owned table has an isolation test, and the Playwright and Maestro workflows. Test placement is already enforced by tools/scripts/validate-repository.mjs (checkTestPlacement).

​
Supersedes and sources

Supersedes: none. Extends 0022. Sources: owner decision in chat, 2026-10-19 (testing and infrastructure notes).