Tested a few queries in Docker.

  - "compress video without losing quality" → #1 was jpegrescan (a JPEG tool). ffmpeg is in the index but didn't surface at all.
                                                                                                                
  - "pretty print json" → jq ranked #7. 
                                         
  - "find files by name" → returned duplicate-file finders (rdfind, duff). fd/fzf didn't appear. 
                                                                                                                                               
  - "download youtube video" → yt-dlp not in results at all. 

Hope this helps

Thank you!

The lazy-loading abilities pattern makes a lot of sense for context bloat. But I'm curious about the practical middle ground? have you found that agents reliably know when to call load_ability vs trying to work with what they already have? In my experience the hardest part of tool use isn't having tools available, it's getting the model to pick the right one at the right time. What's been your experience?