| name | lightflow-video-highlights |
| description | This skill should be used when the user asks to identify long-video highlights with VideoScore for LightFlow automatic editing. |
| version | 0.3.0 |
LightFlow VideoScore Highlights
Use lightflow.video_highlights before lightflow.video_auto_edit_plan when
working from a long source. Set LIGHTFLOW_VIDEOSCORE_API_URL and the
runtime-only LIGHTFLOW_VIDEOSCORE_EVIDENCE_KEY before running it. Require the
evidence key to contain at least 32 non-whitespace bytes. This workflow fixes
the model to TIGER-Lab/VideoScore-v1.1; do not override it. The workflow is
Rust-native and sends no credentials in workflow inputs or artifacts.
Mount the source media read-only at the same path in the VideoScore service.
The service must implement POST /v1/highlights: accept model,
source_path, brief, window_seconds, and stride_seconds; return
{"segments":[{"start_seconds":number,"end_seconds":number, "aggregate_score":number,"dimensions":object,"reason":string}]}. Scores
are on the VideoScore 1–4 scale.
Each returned candidate includes workflow, source_path, timestamps, score,
the fixed model, reason, and a lowercase-hex HMAC-SHA256 evidence tag. The
clips output wraps these signed candidates directly as {id,path,start,end, highlight}. Pass clips unchanged to lightflow.video_auto_edit_plan as
, or to as , using the same evidence
key. LightFlow rejects altered paths, timestamps, scores, models, reasons, or
tags before media probing.
