Skill: Agent Integration Proof Patterns

Capability Domain: TEST_INTEGRATION
Technology Stack: pytest, CAP-004, artifact-56
Linked Activities: TFK-02, BPE-02, DTA-19

Overview

Patterns for lane 1 (control-plane proofs) and lane 2 (deterministic shell) agent testing. Proves orchestration under ScriptedLLM — not model quality. Generic for any project with SAO §17 and selected PRF-* rows from artifact 56 Part 4.5.

Principles

1. Only CAP-004 mocks the LLM. Real DB, real broker (if used), real tool registry.
2. Assert the control plane, not chat prose. Workflow id, tool names, domain writes, plan/step transitions, blackboard keys.
3. Scripted queue = test oracle for LLM turns. Temperature irrelevant in lane 1.
4. Happy + reject for adverse PRF rows (429, parse fail, HITL block, exhausted queue).

Harness (TFK-02)

See rule do-assert-agent-story and TFK-02 for assert_agent_story and @pytest.mark.agent_proof.

Lane 2 (deterministic shell)

Non-LLM steps are ordinary pytest — no @agent_proof, no ScriptedLLM.

CI

Lane 1 via make test-agent-proof when AGENTS_ENABLED=true. Do not add -m agent_proof to default PR CI until PRF tests exist.

Anti-patterns

FakeLLM silent fallback; @patch on ToolExecutor; asserting assistant wording.

References

Artifact 56 Part 4.5; Rule do-assert-agent-story.