vss-ask-video

Send a POST request to the VSS agent's /generate endpoint with an input_message instructing it to call the video_understanding tool for a named sensor. The agent runs VLM inference on the clip frames and returns the answer in the response value field.

The video_understanding tool answers ad-hoc visual questions about a recorded clip, such as objects, people, actions, colors, timing, and safety compliance. It is used when no prior summary, database record, or tool output already answers the question.

Yes. The skill requires listing VST sensors via /vst/api/v1/sensor/list before any /generate call. If no sensor matches the video filename stem, the video must be uploaded with a PUT request and an ISO 8601 timestamp first.

The base profile is recommended because it serves the video_understanding tool on the VSS agent; the lvs profile also works. Probe http://HOST_IP:8000/docs first, and deploy a profile with the vss-deploy-profile skill if the probe fails.

Do not use it when a database, MCP result, prior tool output, or existing summary already answers the question. It is reserved for cases requiring fresh model inference on video pixels, unless the user explicitly wants verification against the video.

The /generate endpoint returns reasoning wrapped in <agent-think> tags inside the value field, with no separate clean-answer field. Consumers must strip those blocks with a regex and return only the remaining final-answer text.