StreamVibe studies both what help a video assistant should provide and when it should intervene. The official project lists 183.4 hours of video, 4,100 evaluation samples and 20,546 help points. Each second, the agent sees only past video and chooses silence or a response, while the evaluation harness manages requests, progress and dialogue. A Qwen3.5-9B-based model uses supervised learning and localized reinforcement learning. Implicit intentions and changing task states remain harder than explicit requests, making this a research effort to measure useful assistance alongside unwanted intervention rather than a finished product launch.