Modes
How it works
video_evidence reuses Arka’s describe_video pipeline. When ffmpeg and a
vision backend are available, Arka samples frames and uses those observations as
raw evidence. If the vision stack is unavailable, it still creates the artifact
with a clear note so the workflow does not silently hallucinate video details.
The first implementation is file-first and local-safe. Integrations with GitHub,
GitLab, Jira, Jam, or Loom should pass the downloaded recording path or URL into
this skill, then attach the generated Markdown or JSON to the external system.