If the current repo has its own rules/skills covering this topic (check .claude/rules/ and repo CLAUDE.md), those take precedence — apply this skill only where they're silent.

Fixing Flaky Tests

Target symptom: Tests pass when run alone, fail when run with other tests.

Diagnose first

Test passes alone, fails with others?
    │
    ├─ Same error every time → Shared state
    │   └─ Database, globals, files, singletons
    │
    ├─ Random/timing failures → Race condition
    │   └─ See async waiting patterns in `writing-tests` skill
    │
    └─ Resource errors (port, file lock) → Resource conflict
        └─ Need unique resources per test/worker

Quick diagnosis:

1. Run failing test 10x alone - does it always pass?
2. Run failing test 10x with the suite - same error or different?
3. Check error message - mentions port/file/connection?

Shared state (deterministic failures)

Tests pollute state that other tests depend on. Fix by isolating state per test.

Database isolation (most common):

# Python: Savepoint rollback - each test gets rolled back
@pytest.fixture
async def db_session(db_engine):
    async with db_engine.connect() as conn:
        await conn.begin()
        await conn.begin_nested()  # Savepoint
        # ... yield session ...
        await conn.rollback()  # All changes vanish

// Jest: Reset mocks between tests
beforeEach(() => {
  jest.clearAllMocks()
  jest.resetModules()  // Clear module cache before test
})

afterEach(() => {
  jest.restoreAllMocks()  // Restore spied functions
})

See language-specific references for complete patterns.

Race conditions (random failures)

Tests don't wait for async operations to complete.

See the writing-tests skill for async waiting patterns:

• Framework-specific waiting (Testing Library findBy, Playwright auto-wait)
• Custom polling helpers
• When arbitrary timeouts are acceptable

Quick summary: Wait for conditions, not time:

// Bad
await sleep(500)

// Good
await waitFor(() => expect(result).toBe('done'))

Resource conflicts (port/file errors)

Multiple tests or workers compete for same resource.

Worker-specific resources:

# Python pytest-xdist: unique DB per worker
@pytest.fixture(scope="session")
def database_url(worker_id):
    if worker_id == "master":
        return "postgresql://localhost/test"
    return f"postgresql://localhost/test_{worker_id}"

// Jest/Node: dynamic port allocation
const server = app.listen(0)  // OS assigns available port
const port = server.address().port

File conflicts:

import tempfile

@pytest.fixture
def temp_dir():
    with tempfile.TemporaryDirectory() as d:
        yield d

Language-specific isolation patterns

Verification

After fixing, verify the fix worked:

# Run the specific test many times
pytest tests/test_flaky.py -x --count=20

# Run with parallelism
pytest -n auto

# Jest equivalent
jest --runInBand  # First verify serial works
jest              # Then verify parallel works