v0.7.0
See the footage: LLM-visible framegrabs
A connected agent can now actually see the media it edits over MCP — three visual tools return real image content blocks an LLM can look at, not just JSON.
New MCP tools
get_framenow returns a real image block (a low-res JPEG drill-in at a timestamp) instead of an unusabledata:URL string.skim_asset— a tiled contact-sheet montage of an asset sampled across a range, plus a text index mapping each cell to its timestamp, for finding the good parts of footage.preview_timeline— the composited cut at a given timeline time: every clip visible at that moment overlaid onto a black canvas, mirroring the export geometry, so the agent can see the edit.
Engine
New frame_jpeg, contact_sheet, and timeline_frame in the always-on CLI engine (driven by the ffmpeg/ffprobe binaries — no dev libraries needed, works in the --no-default-features build). The build_* argument builders are pure and unit-tested.
get_info now points the agent at the visual tools.
Full changelog: v0.6.0...v0.7.0