# Aperture 0.7.11 development release ## Studio appearance (dev.8) Studio is an optional application-wide theme with light, dark, and device-following modes. The existing theme remains available as Classic and stays the default until each user opts in. Choose **Avatar → Appearance**, save a theme, and use Studio’s header sun/moon control for a quick mode switch. Returning to Classic preserves the saved Studio mode. The new styling is applied to existing screens and controls, including administration, search, dialogs, charts, graphs, and the assistant. All routes and feature permissions remain in place. No media-server or recommendation settings are changed, and no migration is required. Labels are provided in all 14 interface languages. See [Appearance](../features/user-settings/appearance.md) and [implementation/validation notes](../appearance-development.md). Use `ghcr.io/dgruhin-hrizn/aperture:dev` to follow development releases, or `ghcr.io/dgruhin-hrizn/aperture:0.7.11-dev.8` for this version. Existing accounts remain on Classic after updating. Choose Studio under Avatar → Appearance. ## What users will see - AI Picks records when a title was selected for each account. Movies and shows that were recently suggested take a break. After the cooldown, only a limited number of returning titles can appear in a list. - Movie recommendations use ratings on Aperture’s ten-point scale. Themes from TMDb and MDBList can lift a movie when they recur in recent viewing. The movie algorithm can reserve one fifth of its requested picks for highly rated films outside the viewer’s most-watched genres. - On Emby, admins can publish separately named TV picks and can opt into private native playlists. Native playlists reference original Emby items, so a playback check-in can update the original title. A user’s Emby client and any external Wrapped exporter determine how activity is displayed and counted. ## What admins can configure | Setting | Where to find it | Behavior | | --- | --- | --- | | Repeat suggestion cooldown | Settings → AI Recommendations → Algorithm | Separate movie and TV settings, both 180 days by default | | Emby TV recommendation groups | Settings → Setup → Media Server → Libraries → TV Shows | Group one or more enabled source libraries into named outputs | | Media server output | Settings → AI Recommendations → Output | Choose a generated library, the private Emby playlist preview, or Off independently for movies and TV | | Import Media Server | Admin → Jobs → Global | Import users, catalog, and watch history without AI tokens or publication | Save any desired settings, run the relevant recommendation-generation job, and then run the corresponding library sync to publish. Generation can use model tokens. Library sync publishes saved results. Numeric Emby source IDs are now accepted by the source-library toggle. Generated Emby outputs are checked against the viewer’s source access and the exact item paths. Aperture withholds library files if another active non-admin account can access the output. The native playlist preview requires each recipient to sign into Aperture with Emby and checks the playlist’s audience before adding original items. ## Upgrade and validation The versioned development image is `ghcr.io/dgruhin-hrizn/aperture:0.7.11-dev.8`. It is available for Linux AMD64 and ARM64. Use the versioned tag to reproduce this build. Back up PostgreSQL and uploaded artwork before replacing the app container. Keep the existing database and mounts. Migrations 0115–0117 backfill suggestion history from retained completed runs and add storage for groups and managed playlists. No manual data entry is needed for these migrations. The image was tested against an existing 0.7.10 database. The Emby 4.10.2 checks covered private playlist visibility, original item references, cleanup after access revocation, and playback reporting to the original movie. External Wrapped exporters and every Emby client remain outside that validation. The earlier dev.1 build was superseded after live Emby testing found permission and playlist API compatibility issues. ## Release dialogs (dev.3) The updated interface shows one release dialog at the next authenticated visit. Admins see configuration and jobs guidance. Regular users see the changes to recommendations and playback. Dismissal is saved per account and role in the user’s preferences. ## Related documentation - [Recommendation behavior](../features/recommendations.md) - [Algorithm settings](../admin/algorithm-tuning.md) - [Source libraries and TV groups](../admin/libraries.md) - [Publishing modes and Emby access](../admin/output-format.md) - [Jobs and model usage](../admin/jobs-overview.md) ## Pin to Emby Home (dev.4) Open **AI Picks → Emby Home** to pin a private Aperture-managed playlist as a Home section. Unpin removes only the section Aperture created and preserves the playlist. Shelves created directly in Emby stay under Emby’s control. This build adds migration 0118 to record which sections Aperture owns. ## Local model explanations Movie and TV explanation batches now accept fenced JSON and recover from apostrophes escaped by local models such as Qwen. Structurally invalid responses still use the existing fallback. Generated movie and TV library descriptions use the recipient’s effective AI language for the headings above the explanation and overview. Run recommendation generation to refresh explanations and library sync to rewrite existing descriptions. ## Tonight (dev.5) The new **Tonight** page combines live Continue Watching from Emby with up to three saved AI Picks. Open a pick or in-progress title directly in Emby. The page uses existing recommendations and requires no new settings, model calls, or Home screen changes. This release also includes the local-model explanation and translated-heading fixes above. ## Not tonight (dev.6) Hide a Tonight pick for 12 hours and get the next eligible saved recommendation immediately when one is available. The change is scoped to Tonight and uses no model tokens. Migration 0119 stores these temporary dismissals separately from your taste profile and settings. ## Query performance and recommendation corrections (dev.7) Update with `ghcr.io/dgruhin-hrizn/aperture:dev` to receive this development build. The fixed-version image is `ghcr.io/dgruhin-hrizn/aperture:0.7.11-dev.7`. ### Faster queries and sync - Recommendation scoring loads shared preferences once, reuses franchise lookups, and batches candidate embedding reads. Accounts without compatible custom interests skip that vector fetch. - Watched-title expansion reads the eligible history once and finds duplicate library copies through their external IDs. Read-only comparisons returned identical exclusions for 41 accounts. - TV diversity analysis groups watched series before joining their embeddings. The automatic adjustment reads that statistic directly. Browsing groups viewer counts over matching titles. - All-user watch-history jobs share movie and episode identity maps for the duration of the job. User histories and exclusions are handled separately, and new imports get targeted lookups. - Automatic genre and franchise updates use bounded batches, protect manual preferences, and apply explicit resets atomically. Concurrent bulk updates for an account are serialized. ### Recommendations follow saved preferences - Completed plays and explicit favorites consistently supply positive history. Movie history is selected separately from TV episodes. Partial-watch exclusions keep their 5% threshold. - Movie and TV dislike behavior uses the saved setting. TV respects explicit exclusions when rewatching is enabled. Penalize mode lowers disliked titles before diversity selection. - Normalized scoring weights apply diversity once. Custom interests require a compatible embedding model, and TV sorts candidates before choosing the top 100 for interest scoring. - User documentation now describes the actual controls and their effects. Stored profiles follow their configured refresh interval and lock. History-input corrections apply at the next rebuild. ### Validation and upgrade The release includes regression checks for ranking arithmetic and user syncs, plus PostgreSQL integration checks for batching, manual overrides, duplicate inputs, resets, rollback, and concurrency. Read-only Unraid comparisons cover query results and browsing requests. Timings and the limits of recommendation-quality validation are documented in [Architecture](../architecture.md). This build uses the existing database schema and configuration. Keep the current app environment, volume mounts, and database when replacing the image. The deployment checks compare persisted admin and user settings and preserve uploaded artwork through the app-container replacement.