v0.4.0 — contact sheets, purpose-aware manifests, knowledge base

What's new

• --grid — packs consecutive keyframes into 3×3 contact sheets so an LLM reads ordered sequences instead of isolated stills (a direct answer to the HN feedback that "keyframes are not video")
• --why — records the purpose of your watch into MANIFEST.txt so the model knows what to look for
• --kb — saves the digested summary as markdown into a local knowledge-base folder
• Honest privacy wording: frames are extracted locally; whatever you feed your LLM goes to that LLM's provider

Install

pip install claude-real-video
# or the LLM-agnostic alias:
pip install llm-real-video

Runs 100% locally (ffmpeg + faster-whisper). MIT licensed.