[Technical Debt] Add integration tests with real LLM calls using VCR.py

Summary

Current Agent 7 tests mock all LLM responses, which means they test mocking infrastructure rather than actual LLM behavior. Need integration tests with real LLM calls to catch prompt regressions.

Current State

All tests use AsyncMock for LLM responses:

mock_llm.ainvoke = AsyncMock(return_value=expected_plan)

This means:

• Tests pass even if prompts are broken
• Cannot detect structured output validation issues
• Cannot verify that prompts actually work with Claude

Problems

1. False confidence: Tests pass but prompts may fail in production
2. Prompt regressions undetected: Changes to prompts not validated
3. No real-world validation: Cannot verify Claude actually returns expected structure
4. Debugging difficulty: Production prompt issues hard to reproduce in tests

Proposed Solution

Use VCR.py (or similar) to record/replay real LLM interactions:

@pytest.mark.vcr()  # Records first run, replays after
async def test_planning_with_real_llm():
    llm = ChatAnthropic(model="claude-3-5-sonnet-20241022")
    
    plan = await generate_execution_plan(
        workflow_type="order_creation",
        trigger_data=real_trigger_data,
        context=real_context,
        available_tools=real_tools,
        llm=llm
    )
    
    # Verify actual LLM response structure
    assert len(plan.tasks) > 0
    assert plan.tasks[0].tool_name.count('.') == 2  # Real validation

Benefits:

• First run makes real API call and records response
• Subsequent runs replay recorded response (fast, no API cost)
• Easy to update recordings when prompts change (--record-mode=all)
• Catches real prompt/schema mismatches

Acceptance Criteria

• [ ] VCR.py (or vcrpy) integrated into test suite
• [ ] Integration test for generate_execution_plan() with real LLM
• [ ] Integration test for infer_modification_intent() with real LLM
• [ ] Integration test for generate_draft_message() with real LLM
• [ ] Integration test for classify_workflow_type() with real LLM
• [ ] Documentation on updating VCR cassettes when prompts change
• [ ] CI pipeline runs with recorded cassettes (no API calls in CI)

Test Strategy

tests/
  llm/
    test_planning.py              # Unit tests (mocked)
    integration/
      cassettes/                   # VCR recordings
      test_planning_integration.py # Integration tests (VCR)

Priority

Medium - Important for confidence in production behavior

Estimated Effort

4-6 hours (setup VCR, write 4-5 integration tests, document process)

Dependencies

• Install vcrpy or pytest-vcr
• Anthropic API key in test environment (for initial recording only)

Related

• Agent 7 (LLM Modules)
• All tests/llm/test_*.py files
• Prompt engineering quality assurance