Discard old thinking steps

If you're building an agent harness, discard all past thinking steps except the most recent. The model summarises its thinking for the user, so those massive thinking steps are just bloating your context. All the major harnesses do this.