Tests must not leak global or shared state across runs, and must not depend on execution order. Before mutating anything shared with the rest of the process — signal listeners (SIGTERM/SIGINT/beforeExit), module-level caches, environment variables, or mocks — snapshot the prior state in beforeEach/setup and restore it in afterEach/finally, even when the test throws. When adding listeners for a test (e.g. process.once('SIGTERM', ...)), remove any pre-existing listeners first so only the test’s own handlers observe the emitted event, then restore the originals afterward.
Mocks must be reset between tests rather than relying on call order: use vi.resetAllMocks()/explicit re-implementation in beforeEach so a suite is correct under randomized test ordering, not just the order it happens to run in locally.
Avoid writing files or other artifacts to the repository checkout during a test run; commit required fixtures instead (e.g. config/test data files under fixtures/) so a test run leaves the working tree clean.
Before relying on an existing shared test harness for isolation (e.g. an in-memory DB reset between files), verify the harness actually provides the isolation needed rather than adding redundant setup/teardown, but err toward explicit local isolation when a test touches global, process-wide state.
Example:
it('cleans up on SIGTERM without ending the process', async () => {
const saved = process.rawListeners('SIGTERM');
process.removeAllListeners('SIGTERM');
try {
// ...test body registering and emitting SIGTERM...
} finally {
process.removeAllListeners('SIGTERM');
saved.forEach((listener) => process.on('SIGTERM', listener));
}
});