Added
-
AI web search can look further back in someone's history (#152).
Until now it only looked at what they watched lately. Settings, Defaults, How much of their history it
searches from adds their long-time favourites and older watches from across the years, as presets from
Recent only (the default, nothing changes) to Deep. A row or one person on a row can set their own, a
person's page has Show this week's mix with a Don't use button on each title, and How we
picked shows what each search was for. With Exa or SearXNG each extra title is one more search; with
the AI's own search there is no extra cost. A title they rated at or below the "didn't like it" line in
any library is never used as a favourite. Adds migration 0112. -
Choose how many seasons of a show to request, on either route (#154).
How much of a show to grab now has a Start from (First or Last seasons) and a How many seasons
count, or Every season, and it works when requests go through Overseerr, Jellyseerr or Seerr as well as
straight to Sonarr. Rows can set their own. If you request through Overseerr, check this setting after
upgrading: a First or Last Season choice made for Sonarr used to be ignored there (every season was
requested), and now applies. Pilot episode and Add unmonitored stay Sonarr-only; through Overseerr a show
set to one of them is not sent, and waits in the inbox with the reason until you pick another choice.
Rows already set to First or Last Season keep taking one season. On both routes, Last seasons skips a
season that has been announced but hasn't started airing, so a request downloads something now.
Adds migration 0113.
Fixed
-
A request that failed before anything was sent no longer blocks that title for good. If TMDB couldn't be
reached to find a show's TheTVDB id, or Sonarr or Radarr failed while looking a title up, the request was
treated as possibly sent and never tried again until you released it on the Requests page. It now waits in
the inbox with the reason and is tried again next run. A send that may really have reached Sonarr or Radarr
is still held back for you to check. A title that would be refused again the next night (Sonarr or Radarr
can't find it, TMDB lists no seasons yet, or a mode Overseerr can't take) waits in the inbox and is only
sent automatically again after a week, or sooner if you change where or how it is requested, so it never
takes an auto-send slot from the titles below it. You can still send it from the inbox at any time. -
A SearXNG search whose engines were all failing no longer counts as "nothing found" (#152). An empty page from blocked or rate-limited engines was cached for a week as a real empty answer; it is now a failed search, retried next run.
-
A local Ollama model that answered AI web search in prose now says why (#152). Ollama's default context window cuts off the start of Shortlist's prompt; the run trace and Settings Test now point to
OLLAMA_CONTEXT_LENGTH=16384. -
AI web search titles now reach rows (#152). The titles the
AI found were ranked below every title TMDB suggested, so they almost never made it into a row: in one
46-person run, 65 of 92 "Picked for You" rows held none. Each is now credited to the watches that led to it,
so it competes with those watches' similar titles (with Exa or SearXNG; the AI's own search is unchanged, as
crediting it made rows less accurate). With SearXNG the AI is asked to name the watched title behind each
proposal, a few extra output tokens. Run details shows how much of a run the AI search supplied. -
Agregarr trailer placeholders no longer count as the real title (#151).
Playing one of Agregarr's stand-in trailers used to count as watching the film, and as finishing a show
that wasn't on the server yet. A title with both a real copy and a trailer could also put the trailer in a
row. Trailer plays are now ignored and rows use the real copy. A title that only has a trailer can
still go in a row, and playing its trailer no longer counts as watching that pick in the reports. -
AI web search no longer describes titles it shouldn't (#152). The list of someone's recent watches sent to the
AI now leaves out titles marked "Don't seed" and titles they rated at or below the "didn't like it" line
(when "Respect Plex ratings" is on). A row pinned to some libraries now describes only what was watched in
them, as its seeds already did. -
Adding a season to a row is quicker and keeps your place. The Add button no longer waits for the whole
rows list to reload, which took seconds while a run was going, and the page no longer jumps up to the
season list afterwards. -
An upgrade no longer stops on one unreadable history record. Upgrading to 1.11 refused to start if a
run, job or activity-log entry held text that was not valid JSON (only possible after a hand edit). It is
now skipped with a warning in the log.