Skill: Test Trophy Architecture

Capability Domain: TEST_STRATEGY
Technology Stack: pytest+behave

Overview

Integration-first testing philosophy replacing the traditional test pyramid. The "trophy" shape reflects where we invest most: thick integration tests in the middle, thin unit tests at the base, acceptance tests above, and selective E2E at the top.

The Test Trophy Model

        ╱╲          E2E (selective, slow, staging only)
       ╱────╲        Acceptance Tests (behave, BDD .features)
      ╱════════╲     INTEGRATION TESTS (main bet, no mocking)
     ╱──────────╲    Unit Tests (thin, only where isolation helps)
    ════════════════  Static Analysis (linting, type checks)

Why Not the Pyramid?

1. Unit tests have limited value: Testing model.save() returns OK tells us nothing about whether data persisted.
2. Acceptance tests are expensive: Real browser, slow, flaky.
3. Integration tests are the sweet spot: Real objects, real connections, real data. Fast enough for CI, thorough enough to catch real bugs.

When Unit Tests ARE Valuable

• Custom QuerySet methods with complex filtering logic
• Pure utility/helper functions with interesting edge cases
• Business logic methods that do computation (not just DB operations)

When Unit Tests ARE NOT Valuable

• Default Model CRUD
• Auto-generated admin views
• Framework behavior
• Anything that requires mocking to test in isolation

Agent test lanes (when SAO §17 applies)

When the project has in-app LLM agents, extend the Trophy with four agent lanes — still integration-first, still no domain mocks:

Lane	Trophy layer	CI gate	Mock seam
1 Control-plane proofs	Integration (thick)	PR when AGENTS_ENABLED=true	CAP-004 ScriptedLLM only
2 Deterministic shell	Integration (thick)	PR (normal integration)	None — no LLM
3 Live contract eval	Integration / contract	Nightly / optional CD	Live provider, temp=0
4 Task performance quality	Eval overlay	Promotion band	Live provider + golden oracle

Placement:
- Lanes 1–2 are part of the integration band — same real DB/services discipline as the Trophy main bet.
- Lane 3 is a contract eval overlay — schema and response bounds, not merge-blocking flake.
- Lane 4 is outside PR merge — TASK-* golden tasks; regression vs last certified band.

Do not move agent orchestration proofs to unit tests or Behave AT. User-visible AT stays in the acceptance band; PRF-* control-plane proofs stay in lane 1 pytest (@agent_proof).

See skill Agent Integration Proof Patterns (lane 1–2) and Agent Task Eval (lane 4).