Dsh Bilibili

Key features

• Optional frame vision descriptions: vision-less main models can still "see" frames — send each frame to a local Ollama / llama.cpp backend (Qwen3-VL 2B/4B/8B tiers) or any OpenAI-compatible vision API for a text description
• The report is attached to every extraction result (an Environment check section)
• Bundled defaulttemplates/summary.md — a concise "time-saver" summary (one-sentence takeaway → timestamped key points → worth-watching segments → shareable closing lines)
• Frame selection is not keyword guessing — the main agent uses the analysis prompts injected into the system prompt to read the transcript and decide which moments are incomplete without the picture.
• the report cites images only when needed