Python Testing Pytest Asyncio

Testing

Stack: pytest, pytest-asyncio, python-dotenv. Setup lives in tests/conftest.py.

Rules

• Mark async tests with @pytest.mark.asyncio (or set the asyncio mode in conftest if it's already configured — check there first).
• Async fixtures: @pytest.fixture + async def returning the value, or @pytest_asyncio.fixture for explicit async fixtures.
• Don't hit real providers in CI. Mock provider SDKs (unittest.mock.patch, or pass an injected client to the constructor).
• .env loaded in conftest.py; tests assume env is available, but secrets should not be required for unit tests — use mocks.
• Test the agentic loop end-to-end when touching providers — invoke() with a tool that returns deterministically, assert the final context has the expected ToolCall + ToolResult + assistant message.
• Test schemas with model_validate, not by constructing fields manually — catches Pydantic v2 regressions.
• Async resource cleanup in tests: use async with or try/finally so a failing test doesn't leak subprocesses (MCP stdio especially).

Existing patterns

• tests/test_llm_response.py — provider response shape.
• tests/test_prompts.py — prompt provider behavior.
• tests/test_postgres_binder.py — binder type mapping.

Common mistakes

• Forgetting @pytest.mark.asyncio → test silently doesn't await.
• Hitting bedrock.client(...) in CI → mock it.
• Leaving MCP stdio managers open after a test → flakes from subprocess leaks.
• Asserting on full LLM response strings → brittle; assert on structure (tool calls, message roles).

See also

• [[python-async-discipline]], [[echo-sdk-llm]]

Source: eka-care/echo-sdk — distributed by TomeVault.