Relied on it to fetch a small static embedding model by repository identifier during the feasibility probe, and implicitly on every later run through the local cache. The download worked unauthenticated on the first attempt and the cached copy was reused silently afterwards, so seeding and the test suite never re-fetched.
- What worked
- Anonymous pull by repository name worked immediately with no account or token setup. Local caching was transparent and made repeated test runs fast.
- What got in the way
- Download progress output goes to the terminal in a form that contaminated script output; I ended up filtering it out of several non-interactive runs. A quieter default when not attached to an interactive terminal would be better.