Test Environments

<objective>
Staging on SQLite passes tests that break on prod Postgres; a shared staging box becomes a
queue where one broken deploy blocks the whole team; an unmocked Stripe call flakes CI at
random. This skill prevents those by designing environment tiers that mirror production where
it matters, isolate per-PR, and stub external dependencies at the HTTP boundary. It delivers a
working `docker compose up` local/CI stack, a parity checklist, and a stubbing strategy keyed
to dependency type.
</objective>

Discovery Questions

Check .agents/qa-project-context.md first — if it exists, use it and skip anything already answered there. Then:

1. How many environments exist today? Local dev, CI, staging, preview, production? Map what you have before designing what you need.
2. Is the app containerized? Check for Dockerfile, docker-compose.yml, or compose.yaml. If yes, multi-stage targets and compose come for free; if not, that is the first deliverable.
3. How is test data seeded? Manual SQL, migration-based, factory libraries, or production snapshots? This decides whether seed scripts are a quick win or a rewrite.
4. How close is staging to production? Same DB engine, queue, cache, auth provider, orchestration? Each mismatch is a class of bugs staging can never catch.
5. External dependencies: How many third-party APIs does the system call, and are they stubbed in non-prod? Unstubbed third parties are the top source of CI flake.

Core Principles

1. Staging must mirror production where bugs hide. If staging uses SQLite and production uses PostgreSQL, staging tests prove nothing about prod behavior. Match the database engine and version, the queue system, the cache layer, and the auth provider — those are where environment-specific bugs live.

2. Ephemeral environments beat long-lived ones. A shared staging environment becomes a bottleneck where one broken deploy blocks the entire team. Per-PR preview environments give isolation and parallel testing; keep staging only for final pre-release validation.

3. Deterministic seed data, not production copies. Production snapshots carry PII, stale references, and non-reproducible state. Build seed data from factories that generate consistent, valid, minimal datasets. (For factory patterns, see test-data-management.)

4. Stub external dependencies at the boundary, not deep inside. Third-party APIs are unreliable, rate-limited, and expensive. Stub them at the HTTP boundary with MSW or WireMock — never by mocking internal service classes, which hides integration bugs between your own code.

5. Environment config is code. Every environment difference (URLs, flags, credentials, resource limits) must be version-controlled and reviewable. No manual setup that cannot be reproduced from the repo.

Environment Strategy

Environment Tiers

Local Development

Fast feedback, zero shared state. Developers must be able to run the full stack locally in under two minutes:

docker compose -f docker-compose.test.yml up -d
npm run db:seed
npm run dev

Use Docker Compose for infrastructure deps (database, cache, queue) but run the application natively for fast reload. External APIs are stubbed with MSW handlers loaded in dev mode.

CI Environment

Fully containerized, created fresh per pipeline run, destroyed after. The block below is the services: fragment of a job — nest it under jobs.<id>.services alongside runs-on and steps; on its own it is not a valid workflow file.

# .github/workflows/test.yml — fragment: nest under jobs.test.services
services:
  postgres:
    image: postgres:18-alpine
    env:
      POSTGRES_DB: testdb
      POSTGRES_USER: test
      POSTGRES_PASSWORD: test
    ports: ['5432:5432']
    options: >-
      --health-cmd="pg_isready -U test"
      --health-interval=5s
      --health-timeout=3s
      --health-retries=5
  redis:
    image: redis:8-alpine
    ports: ['6379:6379']
    options: >-
      --health-cmd="redis-cli ping"
      --health-interval=5s
      --health-timeout=3s
      --health-retries=5

Docker Compose vs Testcontainers

Two ways to give tests real infrastructure. Pick by where the lifecycle should live:

• Docker Compose — declarative stack you bring up before the suite (docker compose up --wait) and tear down after, usually via a trap-guarded script. Best for local dev, a shared CI stack, and E2E where many tests share one set of services.
• Testcontainers (Node / JVM / Python / Go) — containers spun up from test code and auto-torn-down per suite or per test, with no compose file or trap to maintain. Best for integration tests that need isolated, programmatic infra (a throwaway Postgres per test class). The 2026 default for "ephemeral infra owned by the test," and a strong alternative to hand-rolled compose + trap scripts.

Reach for Compose when humans and many tests share the stack; reach for Testcontainers when each test (or suite) wants its own disposable copy.

Preview Environments (Per-PR)

Each pull request gets its own isolated environment; reviewers click a link and test the exact changes without interfering with other PRs.

Hosting options (2026), pick by stack:

• Vercel preview deployments — Next.js / static / serverless; per-PR URL automatically.
• Cloudflare Pages preview — git-integrated, generous free tier.
• Render / Railway preview environments — full-stack including databases.
• Northflank, Qovery, Bunnyshell, Uffizzi — full ephemeral-environment platforms (Kubernetes-backed) when previews need the whole stack, not just a frontend.

For each preview, pair the env lifecycle with a database branch (Neon, Supabase, PlanetScale-style): create a branch on PR open, drop it on close. That gives every preview a cheap, instant, isolated DB copy instead of a shared staging DB. (See test-data-management.)

For local-dev parity with CI:

• Devcontainers (.devcontainer/devcontainer.json) — VS Code, Codespaces, JetBrains. The standard for "everyone gets the same Docker-backed dev env."
• Tilt (Tiltfile) — Kubernetes-first local dev with hot reload and multi-service orchestration. Pick when staging itself is K8s.

A frontend preview with E2E against the generated URL is a few lines:

- name: Run E2E against preview
  env:
    BASE_URL: ${{ steps.deploy.outputs.preview-url }}
  run: npx playwright test --project=chromium

A custom Docker preview keyed to a per-PR namespace, auto-torn-down on close:

- name: Deploy preview
  run: |
    NAMESPACE="pr-${{ github.event.number }}"
    docker compose -f docker-compose.preview.yml -p "$NAMESPACE" up -d
    echo "preview-url=https://${NAMESPACE}.preview.example.com" >> "$GITHUB_OUTPUT"

- name: Teardown preview
  if: github.event.action == 'closed'
  run: |
    NAMESPACE="pr-${{ github.event.number }}"
    docker compose -p "$NAMESPACE" down -v

Staging

Long-lived environment that mirrors production infrastructure. Reset weekly or on-demand to prevent drift:

#!/bin/bash
# scripts/reset-staging.sh
set -euo pipefail

echo "Resetting staging database..."
psql "$STAGING_DATABASE_URL" -c "DROP SCHEMA public CASCADE; CREATE SCHEMA public;"

echo "Running migrations..."   # migrations MUST recreate extensions + grants (see caveat below)
npm run db:migrate -- --env staging

echo "Seeding anonymized data..."
npm run db:seed -- --env staging --dataset production-anonymized

echo "Verifying staging health..."
curl -sf https://staging.example.com/health || exit 1
echo "Staging reset complete."

Caveat: DROP SCHEMA public CASCADE also drops the schema's default privileges and any installed extensions (uuid-ossp, pgcrypto, …). Your migration pipeline must recreate them (CREATE EXTENSION IF NOT EXISTS …, re-grant defaults) or the migrate step fails. Don't assume a bare CREATE SCHEMA public restores the prior grants — it does not.

Docker Compose for Testing

A production-quality docker-compose.test.yml spins up the full stack (app, Postgres, Redis, a one-shot seed container, Mailpit) for integration and E2E tests. Two details that matter:

• Health checks gate depends_on. Without a healthcheck + condition: service_healthy, depends_on only waits for the container to start, not for the service to accept connections — tests then race the database and fail with connection errors.
• Seed is a one-shot container, not a long-running service. It uses depends_on: condition: service_completed_successfully, so the app starts only after seeding exits 0. Teams that model seed as a long-running service get a race where the app boots mid-seed.

See references/docker-compose.md for the full docker-compose.test.yml, the trap-guarded integration test runner, the multi-stage Dockerfile (with the production target), and the MinIO block.

Multi-Stage Dockerfile

One base layer installs deps once; development, test, and seed stages reuse it; and a slim production stage runs prod deps only (npm ci --omit=dev) with build artifacts copied from the test stage. The split keeps test dependencies and source out of the shipped image while giving each environment its own entrypoint. Use npm ci --include=dev in base — the modern flag; --production=false is legacy --omit/--include syntax. Full Dockerfile in references/docker-compose.md.

External Dependency Management

Stubbing Strategy by Dependency Type

Avoid: MailHog — unmaintained, last release 2020. Use Mailpit (axllent/mailpit); it is a drop-in on the same ports (1025 SMTP / 8025 UI).