Most test suites get feature flag isolation backwards. The instinct is to opt in to a feature when a test needs it. The correct default is to force the feature off for every test and make the feature's own tests opt back in.

Here is the concrete version. I am building a warmup phase into the publishing engine, the part that decides whether a new account needs a slow ramp before full volume. The feature has its own config flag, WARMUP_ENABLED, and it touches state in ways that bleed across tests: writes to a ledger, mutates a counter, gates publish decisions.

The first version of the test suite left WARMUP_ENABLED at its production default, True. Tests that did not care about warmup were running against a warmup enabled engine. Most passed anyway because the warmup state happened to be inert for their inputs. That is luck, not design. Any time the warmup logic grows more aggressive or initial state shifts, those tests start failing for reasons that have nothing to do with what they are actually testing.

The fix is two parts. First, add WARMUP_ENABLED = False to the test config so the module level default is already off. Second, wire up an autouse conftest fixture to make it impossible for a test to accidentally inherit a dirty enabled state:

# conftest.py
import pytest
from x_engine import config

@pytest.fixture(autouse=True)
def disable_warmup(monkeypatch):
 monkeypatch.setattr(config, "WARMUP_ENABLED", False)

autouse=True means every test in the suite gets warmup disabled before it runs. No per test annotation, no risk of forgetting. The warmup tests themselves opt back in explicitly:

# test_warmup.py
def test_warmup_evaluates_correctly(monkeypatch):
 monkeypatch.setattr(config, "WARMUP_ENABLED", True)
 # now the feature is live; test it

The real tradeoff: you are encoding the assumption that off is the right baseline for most tests. If your feature is deeply load bearing and most tests genuinely need it on, flip the logic. But for warmup, a gate that the rest of the engine ignores unless it fires, off is the correct starting point. The tests that care about the feature should say so explicitly. That explicitness is the point.

What I would do differently: wire this fixture up the same day the feature flag is created, not after the suite accumulates tests. I added WARMUP_ENABLED in an earlier commit and let a handful of tests pile up before patching the suite wide behavior. That gap cost me debugging time on a failure that was impossible to explain until I traced it back to warmup state leaking through.

The rule I am now following: every feature flag gets a suite wide opt out fixture on the same commit it gets added to config. Not the day tests start failing.