Accurate Testing

Accurate Testing

The Deletion Test (Golden Rule)

Before considering any test complete, apply this mental check:

If the production code this test targets were deleted entirely, would this test still pass?

If yes, the test is worthless. It tests framework plumbing, not production logic.

Core Workflow

1. Identify the production code under test — find the exact function, class, or endpoint
2. Read it — understand its real logic, branches, edge cases, and dependencies
3. Apply the import rule — the test file MUST import from production code
4. Choose the right mock boundary — mock I/O at the edges, never mock the thing being tested
5. Write assertions against real behavior — assert on return values, state changes, side effects that matter
6. Run the deletion test mentally — would deleting the production function break this test?

The Five Laws of Accurate Tests

Law 1: Import Production Code

Every test file must import the actual production function/class it claims to test.

# WRONG — tests LangGraph, not your app
from langgraph.graph import StateGraph
graph = StateGraph(MessagesState)
graph.add_node("echo", lambda s: {"messages": [AIMessage(content="Echo")]})

# RIGHT — tests your app
from app.agents.core.graph_builder.build_graph import build_comms_agent
graph = build_comms_agent(checkpointer=MemorySaver())

Law 2: Mock at the Boundary, Not the Core

Mock external I/O (network, database, filesystem). Never mock the logic under test.

# WRONG — mocks the function being tested, tests nothing
with patch("app.services.chat_service.run_chat_stream") as mock:
    mock.return_value = "response"
    result = run_chat_stream(msg)  # just calls the mock

# RIGHT — mocks the dependency, tests the real function
with patch("app.services.chat_service.llm_client.invoke") as mock_llm:
    mock_llm.return_value = AIMessage(content="hello")
    result = run_chat_stream(msg)  # runs real logic, fake LLM

Law 2b: Patch Module Singletons, Not Individual Functions

When production code uses a module-level singleton (a shared client, cache, or connection object), patch the singleton's attribute directly. This ensures all production code that touches the singleton — including code several layers deep — uses the real test resource without any function-level patching.

# Production: redis_cache = RedisCache()  (module singleton)
# StreamManager uses redis_cache.redis internally

# WRONG — patches one function, misses all others that use redis_cache
with patch("app.core.stream_manager.StreamManager.publish_chunk") as mock:
    ...  # other methods still use the broken/missing redis_cache.redis

# RIGHT — patch the singleton attribute; all production code sees real Redis
from app.db.redis import redis_cache

@pytest.fixture
async def real_redis(monkeypatch):
    client = Redis.from_url("redis://localhost:6379", decode_responses=True)
    await client.ping()
    monkeypatch.setattr(redis_cache, "redis", client)   # one patch, everything works
    yield client
    await client.flushdb()
    await client.aclose()

async def test_stream_publishes(real_redis):
    await StreamManager.start_stream("s1", "conv1", "user1")
    await StreamManager.publish_chunk("s1", "data: hello\n\n")
    chunks = []
    async for chunk in StreamManager.subscribe_stream("s1"):
        chunks.append(chunk)
        break
    assert chunks[0] == "data: hello\n\n"

Law 3: Assert on Production Behavior

Assert on what the production code actually does — return values, state mutations, raised exceptions, emitted events.

# WRONG — asserts mock was called (tests your test setup)
mock_service.process.assert_called_once_with(data)

# RIGHT — asserts the actual outcome
result = process_email(raw_email)
assert result.subject == "Re: Meeting"
assert result.is_read is False
assert len(result.attachments) == 2

Law 4: Cover Real Branches

Read the production code. Find the if/elif/else, try/except, and early returns. Write a test for each path.

# Production code has: if user.is_premium: ... else: ...
# Test BOTH paths
def test_premium_user_gets_extended_features(): ...
def test_free_user_gets_basic_features(): ...

Law 5: Test Error Paths, Not Just Happy Paths

Production bugs cluster in error handling. Test what happens when dependencies fail.

def test_handles_api_timeout():
    with patch("app.tools.gmail.client.send") as mock:
        mock.side_effect = httpx.TimeoutException("timeout")
        result = send_email(to="x@y.com", body="hi")
        assert result.error == "Failed to send: timeout"

Anti-Pattern Detection

When writing or reviewing tests, check for these red flags. For detailed examples and fixes, see references/anti-patterns.md.

Mock Hierarchy (What to Mock Where)

Language-Specific Guidance

• Python (pytest): See references/pytest-patterns.md for fixture design, parametrize patterns, and conftest hierarchy
• TypeScript (vitest/jest): See references/vitest-patterns.md for module mocking, type-safe mocks, and async patterns

Pre-Commit Checklist

Before finalizing any test:

• Test imports the production function/class directly
• Removing the production code would break this test
• Assertions check return values or state, not just mock calls
• Each branch in production code has a corresponding test case
• Error/exception paths are tested
• Mock count is proportional to external dependencies, not internal logic
• Test name describes the behavior being verified, not the implementation