Digarr v1.19.0 is out. Two additions this time, and quite a few fixes around discovery, playlists, and the messages you see when something doesn’t work.
If you haven’t come across it yet, digarr is my self-hosted music discovery app. It uses your listening history to suggest artists and albums, lets you review them, and can send your picks to Lidarr, slskd, or playlists. You bring your own AI provider, including a local model. You don’t need Lidarr to use it.
You can now give several genres equal priority. Say you want more ambient and doom metal in your queue: you can put both in the preferred group, see that group first, or browse it separately. The usual score ordering is still the default. This changes how you browse, not the recommendation scores or auto-approval rules, and it works from explicit genre tags rather than guessing which genres belong together.
Playlists also explain what happened during generation. You can see which artists made it in, which couldn’t be matched, and which were left out because of a size limit, missing provider, or lookup error. Artist counts and track counts are separate, and those local results stay visible even if sending the playlist to a target fails. Searches now check that the returned artist matches before picking tracks and keep trying the configured fallback sources. With no track resolver available, generation no longer makes up track titles.
Most of the fixes are about the information discovery works from. A Spotify top-artist position and a starred artist in Subsonic aren’t the same kind of evidence; digarr now handles them within their own sources, without counting an overlapping artist several times. AI profiles keep the genre context for each seed artist, so separate musical interests don’t have to collapse into one dominant genre. Artists with exactly equal positive taste weights also get a turn as similarity seeds, rather than the same ones always winning the cutoff.
There are some less visible but useful fixes too: name-only lookups check MusicBrainz names and aliases before accepting a match, preview failures keep their explanation after the queue finishes, and empty listening history is distinguished from a failed request.
None of that proves a recommendation will suit you. AI suggestions and artist matches can still be wrong; listening and review are still part of it. The full changelog has the rest, and the README has setup instructions. If a playlist picks the wrong artist or a recommendation looks confused, an issue with what you expected and what happened would help.