Changelog

All notable changes to Deep Cuts will be documented here.

[0.1.7] — 2026-06-06

Features

Song Structure Archetypes: Deep Cuts now automatically groups your tracks by structural blueprint — intro, verse, chorus, outro patterns are compared across your whole library, and tracks with the same shape are clustered together. Labels like “I·VPC×2·O” appear on the track detail pane and in the Music Map structure view. Useful for spotting which tracks have a two-drop structure vs. a long build, or finding all your quick-intro tracks for a live set.
Filter by Structural Archetype: Two filter pills now appear under the structure display in the track pane — one to find tracks with the exact same section sequence, one to find all tracks in the same structural family. Great for building sets where the song shapes complement each other.
Faster Analysis on Large Libraries: The structure encoding step is now dramatically faster on big collections — what previously stalled on libraries of 1000+ tracks now runs in a single pass.

Fixes

Misidentified Outros: Quiet breakdowns and drops in the middle of a track were sometimes labelled as “Outro”, which threw off structure matching. The app now correctly identifies those as Bridges or Breakdowns, reserving Outro for the actual end of the track.
Structure Filter Missing Tracks: Clicking the structure family filter pill was silently excluding roughly half the matching tracks. It now reliably returns every track in the cluster.

[0.1.6] — 2026-06-05

Features

Mood Filter Band with Tolerance Slider: Added a configurable tolerance slider to the mood filter, letting users widen or narrow the acceptance band around a target mood value.
Clickable Classifier Values: Genre and vocal classifier values in the track detail pane are now clickable and apply as active filters.
Pipeline Pause/Resume Controls: Added manual pause/resume functionality to the analysis pipeline from the Analysis Panel, along with a “Paused” badge indicator.
Automatic Chat Session Pausing: Automatically pauses the analysis pipeline when entering a chat session to avoid resource conflicts, and resumes it on navigate-away, with toast notifications and inline banner updates.
Playlist Source in Statistics Panel: Added a playlist source selector to the statistics panel for scoped per-playlist stats.
Analyze Button Glow Pulse: The Analyze button in the navbar now pulses with a glow effect while analysis is active.
Native Zoom Controls: Added zoom in/out/reset entries to the View menu using native macOS zoom.
Collapsible Panes & Hover Preview: Extracted a reusable CollapsiblePane component; both sidebars now show a hover preview while collapsed.

Fixes

Mood Filter Tolerance Accuracy: Corrected tolerance boundary calculations so the filter band matches the intended range.
Theme-Aware CSS Variables: Replaced hardcoded dark rgba values in Autocomplete and FilterSidebar with theme-aware CSS variables.
Full-Height Filter Sidebar: Fixed the filter sidebar to fill the available height correctly.

[0.1.5] — 2026-06-05

Features

Custom User Tags & Suppression: Introduced the ability for users to append manual custom tags (defaulting to the "user" source namespace) and suppress automatic tags from the analysis pipeline. User tags and suppressions are persisted across scans via the database and backed up in .dc.json sidecar files.
Visual Distinction for User Tags: Styled custom user tags in the track detail pane with fully filled backgrounds using the prefix theme color, bold dark-contrast text, and a glowing outer border for clear visual distinction from automated hollow tags.
Suppress & Restore Tag Workflows: Added quick suppression for automated tags via right-click (represented as strikethrough/dimmed chips) and a restoration dropdown menu in the track details pane to easily unsuppress them.
Reusable Autocomplete Component: Refactored the local autocompletion implementations in both the filter sidebar and the track detail pane into a unified, reusable TagsAutocomplete.svelte component.
Align Tag Search Styles: Standardized input fields by introducing a borderless mode to Autocomplete.svelte, ensuring the sidebar tag filter input fits cleanly within its parent wrapper without double borders.

Fixes

User Tag Promotion Conflict: Handled SQLite conflicts gracefully when a user adds/promotes an existing tag, ensuring the database record updates without key violations.

[0.1.4] — 2026-06-04

Features

Autotagging & Unified Tags UI: Introduced a system to automatically emit namespace-prefixed tags from multiple analysis stages. Added a unified Tags section in the track detail pane displaying colored chips that filter the library on click. Includes synonym caching, prompt example cleanup, and instrument hallucination guards.
Pipeline Metrics Database & Diagnostics Drawer: Implemented automatic latency and status tracking per track/pass in a local metrics.db. Added a diagnostics drawer under Library Settings featuring average latency cards, a recent failures log, and a D3 Gantt trace view aggregated by pass for large libraries.
Saved Search Smart Auto-Naming: Automatically generates descriptive, human-readable names for saved searches based on active filters, complete with Vitest unit tests.
Analysis Coverage Panel Stats: Added detailed Qwen metadata coverage statistics to the library Analysis Coverage panel.

Fixes

TrackList Pagination Reactivity Loop: Resolved a bug where the TrackList pagination would reset unexpectedly due to a reactivity loop in the selectedTrack state effect.
Duration Formatting in UI: Replaced raw seconds with formatted H:MM:SS duration strings in the track list, and added millisecond formatting for sub-second durations in the metrics inspector.

[0.1.3] — 2026-06-03

Features

Downgrade llama-server for Metal Audio Stability: Reverted the bundled llama-server to version b9432 to resolve a critical upstream Metal regression (in b9433+) that caused Qwen2-Audio prompts to enter an infinite @@@ token loop.
Improved Music Map Projections & Layouts: Added Mood Centroid projection, Genre Wheel circle layout, Mood coloring scheme with auto-color toggling, and generalized similarity queries with customizable sonic/description blend weighting.
Logarithmic Contour Heatmap Overlays: Integrated D3 density contour overlays on the Music Map to visualize density clusters of sonic similarity.
Fuzzy Gaussian BPM Correction & Joint Preprocessing: Implemented fuzzy Gaussian correction passes and joint Key/BPM preprocessing to improve alignment and fallback reliability.
IDF-Scaled Hybrid Search: Implemented term-frequency (IDF) scaling for hybrid text/vector searches, combining CLAP audio similarity and textual descriptions.
Loudest Region & Analysis Window Markers: Added visual markers in the player waveform showing the exact 10-second high-energy window analyzed by Essentia/Qwen.
Non-Music Skip Optimization: Automatically checks if a track is non-music (derived from Essentia genre probabilities) and skips heavy spectrogram and AI passes.
App Log Opener & License Viewer: Added an application log file opener and a view of the bundled AGPL-3.0 license directly in the Settings drawer.
Faceted & Extended Search: Composer field is now fully indexed and matched by text filters. Collapsible legend added to the Music Map with automatic hiding on pan/zoom.
Mood Filtering: Seven mood axes (happy, sad, aggressive, relaxed, party, acoustic, electronic) exposed as dual-handle histogram range sliders in the filter sidebar, backed by Essentia mood scores.
Mood Radar: Spider/radar chart visualisation in the track detail pane showing the full mood profile of the selected track.
Chat Session Persistence: Chat conversations are now saved to the database with full-text search and a session picker, so you can return to previous Q&A sessions for any track.
Statistics Panel — Set Source Selector: The Statistics page now lets you scope comparisons to the full library, the current filter, or a specific folder.
Reset Analysis Pass: A per-track menu in the track detail pane lets you clear and re-run any individual analysis pass (BPM, key, Essentia, Qwen, etc.) without re-scanning the whole library.
Energy-Based Window Selection for Qwen & Essentia: Both the Qwen2-Audio description pass and the Essentia classifier pass now pick the most energetically interesting (loudest) 10-second audio window rather than a fixed offset, improving description and tag quality on tracks with long or quiet intros.
CLAP-Based Qwen Validation & On-Demand Resampling: Automatically validates generated Qwen descriptions against the track’s CLAP audio embedding. If similarity falls below a calibrated threshold (0.28), it triggers an on-demand resampling pass with a lower temperature (0.2) and strict prompts to automatically correct hallucinations and language slippage.
Pass Dependencies & Cascading Resets: Configured the Qwen analysis pass to depend on CLAP in the pipeline database. Resetting a track’s CLAP pass now automatically cascades to clear and re-verify its Qwen descriptions and tags.
Save Filtered Results as Playlist: A new “Save to Playlist” button directly in the filter sidebar allows users to quickly save the current filtered search results into a new or existing playlist.
Clickable Metadata Filters: Artist and Album fields in the track detail pane are now clickable, allowing users to quickly filter the music library by the selected artist or album.
Pipeline Execution Optimization: Reordered analysis phases so that essentia and bpm_refinement run before qwen and description_embed, ensuring that refined BPM, key, and mood statistics are available during description generation.

Fixes

Scope Reset Pass: Fixed the reset analysis pass action to correctly scope the database and column resets to the current track only (preventing multi-track reset bugs).
Theme Visibility & Contrast: Improved the visibility and contrast of WaveSurfer waveforms and the Mood Radar spider chart when using the light theme.
Pagination: Replaced load-more behavior with pagination to avoid performance degradation when updating search filters on large libraries.
Bounds-Aware Windows: Fixed bounds-checking for Qwen, Essentia, and CLAP analysis windows on very short audio files.
UI & Theme Enhancements: Integrated theme-aware waveform styling in the chat pane, unified Update/Network cards, and disabled chat input during active pipeline analysis.

[0.1.2]

Changes

Sidecar file writing (.dc.json) is now opt-in (disabled by default); toggle in Settings.
Fixed notarization issues blocking macOS Gatekeeper approval.
App version is now derived automatically from Cargo.toml at compile time.
Updated app icon.

[0.1.1]

Features

AcoustID Metadata Enrichment: Fingerprints tracks against the AcoustID/MusicBrainz database to fill in missing title, artist, album, and year metadata. Includes cover art fetching with local database caching.
Playlists: Save and load named playlists; export any playlist or filtered set as a native M3U file via the macOS file dialog.
Statistics Page: Compare audio feature distributions (BPM, key, loudness, genre, mood) between the full library and the current filter set, with interactive histograms.
WaveSurfer Region Selector in Chat: Draw a region on the waveform to send only a specific segment to the Qwen2-Audio model, reducing context window usage on long tracks.

Fixes

Cap chat audio at 4 minutes to stay within llama-server context budget.
Resolve llama-server audio processing bugs and restore SSE streaming with a waiting indicator.
Smooth-scrolling UI improvements and miscellaneous Svelte 5 / compiler warning fixes.

[0.1.0]

Initial public release.

Features

Recursive library scanner with sidecar (.dc.json) persistence
Audio analysis pipeline: BPM (with correction and refinement), key, loudness (EBU R128), waveform, duration
Essentia ONNX classifier: 400-class Discogs genre, vocal/instrumental detection, seven mood axes
CLAP audio embeddings (512-dim) stored in sqlite-vec for KNN similarity search
Qwen2-Audio-7B local descriptions via llama-server
all-MiniLM-L6-v2 description embeddings for semantic text search
Music Map: UMAP 2D projection of CLAP embeddings, D3 zoom/pan, filter-aware, KNN inspection pane
Full-text and faceted filtering: genre, key, BPM range, vocals, folder, similarity
WaveSurfer.js waveform and spectrogram player with track detail pane
Local Multimodal AI Chat: Interactive Q&A with any track using Qwen2-Audio-7B-Instruct, with real-time SSE token streaming
In-app Model Downloader: Download, verify, and monitor all neural network models directly inside the app with resumable transfers
Startup Update Checks: Automatic update notifications on launch
Bundled llama-server sidecar with relocation-safe shared libraries
Duplicates view: CLAP-based similarity detection with clickable track rows
Sonic Glitch design system with dark, light, and accessible high-contrast themes
macOS sandbox-compatible, fully offline