Not Mock In Integration Tests

Do not use mocks in integration tests — use real objects, real connections, real or real-like data from fixtures. Think of them as acceptance tests minus the UI.

CAP-004 carve-out (agent integration proofs)

When SAO §17 applies and a test is marked @pytest.mark.agent_proof, only CAP-004 ScriptedLLM may stand in for the LLM provider. Everything else stays real:

• Real database and transactions
• Real tool registry / ToolExecutor wiring
• Real broker and result backend (when CAP-060 in scope)
• Real domain services — no @patch on services the agent invokes

ScriptedLLM replays ordered responses from tests/fixtures/llm_scripts/<prf-id>/. It proves orchestration, not model quality.

FakeLLM ban

Do not use FakeLLM, hand-rolled stub clients, or test doubles that:

• Return empty lists / {} when the script queue is exhausted (hides missing turns)
• Silently swallow parse errors
• Bypass the LLM port protocol (CAP-001)

If the scripted queue runs out, the test must fail loudly. Use CAP-004 ScriptedLLM or a real provider in lane 3/4 eval tests — never a silent fallback mock in PR merge gates.

Lane 2 reminder

Deterministic shell tests (D0 pre-filter, JSON parse, validate-before-LLM) are ordinary integration tests — no LLM mock at all.