ADR-015: Testing Strategy

Status

Proposed — Pending engineering team review

Context

H7 falsified: test coverage across all Gen 2 services is near-zero. Most services have 2-3 test files with minimal assertions. There are no CI test gates — builds pass regardless of test results. This is the single biggest risk to any migration or consolidation effort.

Current State

Impact

• Every service touched during consolidation (ADR-001) is a regression gamble
• Cannot validate database migration (ADR-005) without integration tests
• Cannot verify BPM migration (ADR-013) without workflow tests
• Cannot validate API compatibility after service merges
• Agentic coding amplifies this: agents can generate code fast, but without tests, they generate regressions equally fast

Decision

Adopt a tiered testing strategy prioritized by blast radius, using Testcontainers for integration tests and BDD/Gherkin for acceptance criteria. Leverage agentic coding to generate test suites from existing code and API contracts.

Testing Pyramid

Priority Order (by blast radius)

Testcontainers Strategy

// Shared test infrastructure — PostgreSQL + RabbitMQ + Redis
@Testcontainers
@SpringBootTest
class PaymentServiceIntegrationTest {
    @Container
    static PostgreSQLContainer<?> postgres = new PostgreSQLContainer<>("postgres:16");
    @Container
    static RabbitMQContainer rabbit = new RabbitMQContainer("rabbitmq:3.12-management");
    @Container
    static GenericContainer<?> redis = new GenericContainer<>("redis:7-alpine");
}

Each service’s integration tests spin up real PostgreSQL, RabbitMQ, and Redis containers. Flyway migrations run against the test database. No mocking of infrastructure — tests validate actual behavior.

Agentic Test Generation Strategy

Agentic coding changes the test generation calculus:

1. Agent reads existing service code — GraphQL schemas, entity models, message handlers, business logic
2. Agent generates test scaffolding — unit tests for all public methods, integration tests for all repository/message handler methods
3. Agent generates BDD scenarios — from GraphQL schema + business rules → Given/When/Then
4. Human reviews — validates test assertions match actual business requirements
5. Agent iterates — fixes failing tests, adds edge cases, improves coverage

This is the highest-ROI use of agentic coding: generating comprehensive test suites from existing code is mechanical work that agents excel at.

CI Gate Enforcement

Coverage Tooling

• JaCoCo for Java code coverage (already in Maven ecosystem)
• SonarQube or Codecov for coverage reporting and PR checks
• GitHub Actions reusable workflow for test execution and coverage enforcement

Hypothesis Background

Primary: A tiered testing strategy prioritized by blast radius, combined with agentic test generation, can achieve 80% coverage on critical services before migration begins.

• Evidence: Near-zero test coverage confirmed across all services (L2 — H7 falsified)
• Evidence: All services use consistent patterns (core-lib, GraphQL, RabbitMQ) making test generation systematic (L1 — H13)
• Evidence: Testcontainers is the de facto standard for Spring Boot integration testing (L1)
• Evidence: Agentic coding can generate test scaffolding from existing code (L1 — demonstrated in other projects, not tested on this codebase)

Alternative 1: Write tests only during migration (test as you touch). - Partially accepted: this is the long-tail strategy for P2/P3 services. But P0 services (payment, purchase-workflow) need tests BEFORE migration.

Alternative 2: E2E tests only (skip unit/integration). - Rejected: E2E tests are slow, flaky, and don’t pinpoint failures. The pyramid exists for a reason.

Falsifiability Criteria

• If agentic test generation produces >30% false-positive tests (tests that pass but don’t actually validate behavior) → manual test writing required for critical paths
• If Testcontainers startup time exceeds 60s per test class → evaluate shared container strategy or TestNG parallel execution
• If 80% coverage target on payment services takes >2 sprints → reduce target to 60% and focus on critical path coverage only
• If CI gate enforcement blocks >50% of PRs in the first month → start with warning-only mode and gradually enforce

Evidence Quality

Overall: L0 (WLNK capped by agentic test generation effectiveness L0)

Bounded Validity

• Scope: All Gen 2 backend services. Frontend testing strategy is separate (Playwright for E2E, Jest/Vitest for unit).
• Expiry: Re-evaluate coverage targets after 6 months of data on actual defect rates.
• Review trigger: If agentic test generation proves ineffective. If Testcontainers adds unacceptable CI time.
• Monitoring: Track coverage % per service, CI pipeline duration, defect escape rate (bugs found in staging/production).

Consequences

Positive: - Safety net for all migration and consolidation work - CI gates prevent regression from shipping - BDD scenarios create living documentation of business rules - Agentic test generation leverages AI for highest-ROI mechanical work - Testcontainers validates real infrastructure behavior (not mocks)

Negative: - Significant upfront investment before migration work begins - CI pipeline time increases (Testcontainers startup) - Coverage targets may feel blocking initially - Test maintenance overhead as services evolve

Decision date: 2026-02-01 Review by: 2026-08-01