Jess spotted that dates looked like file dates rather than library-add dates.
He is right, and MediaShelf was not the culprit: it reproduces Plex's addedAt
exactly (verified 500/500 identical to the second). Plex's own field is what
follows the file — replace or re-encode one and Date Added resets while the
item, its ratingKey and its watch history all survive.
Measured on the live library, comparing addedAt against lastViewedAt where both
exist: 55 of 509 movies (10.8%) and 306 of 1,393 TV Show Archive items (22.0%)
were watched BEFORE they were "added" — 19% overall. 2001: A Space Odyssey
reports added 2026-07-31, last watched 2017-08-26.
That is not cosmetic. pre_history is derived from added_at, so an old item whose
file was replaced looks post-coverage and gets promoted into the CONFIDENT
reclaim pool, which is the one pool meant to be trustworthy.
A completed play proves the item already existed, so added_at is now
MIN(provider_added_at, first_watched_at). Plex's raw value is kept in
provider_added_at, added_at_source records which applied, and the item drawer
explains the substitution instead of quietly disagreeing with Plex. Unwatched
items keep Plex's value since nothing contradicts it. first_seen_at is also
recorded now and is authoritative for anything added from here on.
Plex's API has no better field; the true insert time is only in Plex's own
metadata_items.created_at on Loki.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GVbG48GAXfCZatcmX123Ra
The application the design describes: Flask + SQLite, Plex for library data,
Tautulli for watch history, report-only.
Structure follows the design's seams. providers/ splits MediaProvider from
HistoryProvider, because on this network library data and watch data live on
different machines and Jellyfin later will have no Tautulli equivalent.
scoring.py implements the reclaim score twice - as a SQL expression for the
live grid (weights change on every slider drag, so storing it would mean
rewriting thousands of rows per drag) and in Python for CSV export and tests,
with a property test over 500 generated rows asserting the two agree.
rules.py compiles saved views to parameterized SQL through a field/operator
whitelist; nothing user-supplied is ever interpolated.
Three properties are enforced by test rather than asserted in prose:
- Ingest is idempotent. Three consecutive full scans leave every count and
every byte total unchanged. A scanner that double-counts produces a report
that looks plausible and is wrong.
- Keep marks survive Plex reassigning every rating key in the library. They
are keyed on content GUID, scoped per library so the Movies and 4K Movies
copies of the same film mark independently.
- Every config variable the app reads is declared in docker-compose.yml, so
a variable set in Portainer can never silently do nothing.
Also found and fixed while verifying against a fake Plex+Tautulli pair:
executescript() commits the pending transaction, so migrations needed their
BEGIN/COMMIT inside the script; replaceChildren() renders null as the literal
text "null"; a hash-only URL change does not reload the document, so deep
links needed a hashchange listener; and SQLite ROUND rounds half away from
zero where Python rounds half to even.
73 tests, no live server required.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GVbG48GAXfCZatcmX123Ra
