arthur-ai/arthur-engine/test-integration

Run Integration Tests

Integration tests (test_install.py) build the real SDK wheel, install it into a fresh venv, and run subprocesses that confirm:

1. All expected imports succeed from the installed wheel.
2. A real OTLP span is delivered to a mock HTTP collector.
3. OpenAI instrumentation emits a span that reaches the collector.

The wheel build is delegated to scripts/build_sdk_wheel.sh, which copies the SDK to a git-free temp directory so uv includes the gitignored python/src/arthur_genai_client/.

Prerequisites

• python/src/arthur_genai_client/ must exist. If missing, run the generate-client skill first.
• uv must be on PATH.

Run all integration tests

From the arthur-observability-sdk/python/ directory:

uv run pytest tests -m integration_tests -v

Run a specific integration test

uv run pytest tests/test_install.py::test_imports_in_fresh_venv -v
uv run pytest tests/test_install.py::test_telemetry_reaches_collector -v
uv run pytest tests/test_install.py::test_openai_instrumentation_sends_span -v

Build the wheel manually (without running tests)

./scripts/build_sdk_wheel.sh            # → dist/<wheel>.whl
./scripts/build_sdk_wheel.sh /tmp/out   # → /tmp/out/<wheel>.whl

Notes

• Integration tests are marked with @pytest.mark.integration_tests.
• The wheel build copies the SDK to a temp dir with no .git ancestor so that uv uses filesystem-based file discovery (not git ls-files). This ensures the gitignored python/src/arthur_genai_client/ is included in the wheel.
• These tests are slower than unit tests (~1–2 min) due to the wheel build and venv creation.
• The mock server in the fixture listens on a random free port for both POST /v1/traces (OTLP collector) and POST /v1/chat/completions (fake OpenAI API).