vss-generate-video-report

vss-generate-video-report produces analysis reports from NVIDIA's Video Search and Summarization (VSS) blueprint. It solves the problem of turning recorded video clips or time-ranged incident data into a consistently structured markdown report, routing every request to the correct backend rather than misusing the VSS agent's generic POST /generate endpoint.

The skill supports two modes. Mode A (Video clip) resolves a clip URL through the video I/O storage service and runs a VLM chat/completions call to caption or summarize a single recorded clip or sensor video. Mode B (Incident range) queries the video-analytics service for an incident list over a time range and renders a narrative incident report. It enforces strict output contracts — exact report titles and required Basic Information rows — verifies the correct deployment profile with lightweight read-only curl probes (VST/VLM reachability for Mode A, VA-MCP for Mode B), and hands off to a deploy skill when a probe fails. It carefully separates the agent-internal clip URL used for in-cluster VLM frame pulls from a rewritten browser-playable URL surfaced in the rendered report, using public host/port environment variables exported by the deploy layer.

It targets operators and engineers running the VSS blueprint who need repeatable report artifacts. Negative triggers steer ad-hoc visual Q&A, semantic archive search, read-only metrics lookup, deploy/profile changes, and real-time alert management to their dedicated skills instead.