v0.19.0

AutoDJ 0.19.0 is a large simplification release: about 45,000 lines removed, no backward compatibility, and many fixes found by an NVDA release sample and a fresh-clone walkthrough. Full details are in the CHANGELOG.

Before you upgrade

Nothing old is converted automatically. Read the CHANGELOG "Removed" and "Changed" sections, then run uv run autodj doctor after upgrading.

• No web build step. The server serves src/autodj/static directly. You no longer need Node or npm run build to run AutoDJ; delete any leftover src/autodj/static_dist folder.
• Index format. Indexes from 0.18 use manifest schema 2 and are refused with a rebuild message. Rebuild with uv run autodj index --force. The tracks database and vectors themselves did not change, so the rebuild re-embeds your library.
• Model folder. The model is now stored in the Hugging Face cache layout under model_dir, so the first autodj index downloads it once more. You can delete the old model folder afterwards.
• Removed: autodj play (use autodj serve --server-audio), the DJ-mix options on serve (use config.toml or Settings), --export-m3u, --history-file and autodj playlist, Daypart and Mood arc, the Entropy walk pick mode, the preset genres filter, presets.toml (presets live in config.toml), backup --online, and the non-worklet effect fallbacks.
• Stricter config. Unknown keys and sections, and out-of-range values, now stop AutoDJ with an error that names them. Profiles and backups from earlier versions are refused.

Highlights

• HTTPS from config ([server] ssl_certfile and ssl_keyfile), with renewed certificates picked up without a restart, and a guide to using your own domain with a Cloudflare Tunnel and Cloudflare Access.
• Paired devices in Settings: rename and revoke devices, and show the pairing code to pair a new one.
• Picks stay musical: silent tracks are never picked automatically, and sets follow Similarity (optionally anchored to the seed).
• Serato cues are imported from audio file tags, and imported cues now reach indexed and player-analysed tracks.
• Simpler backup and restore, simpler liner uploads that work in group-writable folders, and one beat-grid method on every machine.
• Screen reader and safety fixes: shortcuts do nothing while a dialog is open, Play now works on a fresh page, nothing ever plays without your own Play, every sound follows the page volume, and many announcements and labels were corrected.

Accessibility testing

The release sample is in docs/accessibility-samples/0.19.0.md: NVDA 2026.2 with Chrome (all flows), Firefox and Edge (a subset), on Windows 11 over HTTPS. An automated agent pressed the keys and captured NVDA's speech; no person listened. The defects it found were fixed before this release. Known quirks, both from how NVDA and Chrome handle dialogs: a dialog's name is read twice when it opens, and one stray line is read when a dialog closes.