We propose Open-IRT, a vision-language-based test-time adaptation method for single-sample open-set streams, achieving state-of-the-art performance via polarity-aware OOD filtering and intermediate domain learning.