Skip to content

v0.92.6-beta

Choose a tag to compare

@Abrechen2 Abrechen2 released this 09 May 20:22
· 1287 commits to master since this release
f826909

What's new since v0.77.0-beta

10 days, 15 minor releases, ~250 new tests. Highlights: Sublarr's first production-ready waveform timing surface, a comprehensive cross-feature security audit, V1 readiness gates (disk safety + unsourceable escalation), the foundational UI-first configuration model, and 60+ audit-driven fixes across cleanup, standalone, and provider plumbing.

Added

  • Aegisub-class waveform editor. The Waveform tab is now a real timing surface with drag-resize regions, keyframe + scene-cut + neighbour-cue snap (priority keyframe > scene > neighbour, default 150/200/80 ms tolerance), Aegisub L/R click-map (set selected cue's start/end), per-region text labels, sticky time-axis ruler, vertical amplitude zoom (1×–5×), pitch-preserving playback rate (0.5×–2×), gap/overlap quality markers (red bars for overlaps, amber for tight <80 ms gaps), audio scrubbing while dragging, keyframe markers, ASS karaoke syllable overlay, opt-in spectrogram, multi-track audio-stream picker. Stacked-Lanes layout (toolbar → active-cue card → wave → synchronized cue-list). Lock-by-default, Undo/Redo/Save toolbar, inline cue-text editing, sync-scroll highlight (in-viewport / active / now-playing bands). Aegisub keymap (Space / S / D / arrow nudge / +/- zoom / ? help). CSP media-src 'self' blob: for the waveform-audio.
  • UI-first configuration (semi-breaking). Settings is now a composite of BootSettings (11 env-loadable fields) + UISettings (200+ DB-only fields, configured through the Settings UI only). Existing call sites work unchanged via __getattr__ forwarder. Loud per-key warnings flag deprecated SUBLARR_<ui_field> env vars at startup. New CI linter tools/lint_no_new_env_fields.py enforces the contract via AST.
  • Disk safety-valve for wanted_search. When /config device crosses wanted_search_disk_pause_pct (default 98%), the search early-returns with paused_reason="disk_critical" instead of querying providers or writing downloads.
  • Auto-mark-as-unsourceable. Wanted items that fail through wanted_search_max_slow_cycles (default 3) move to status='unsourceable' and drop out of the scheduler queue.
  • Subtitle backup management. Dedicated /settings/cleanup/subtitle-backups page with filter / per-row Restore (atomic swap-rename) / Delete / bulk purge + dry-run preview. Backups now land in a hidden .sublarr/backups/ subdir; idempotent migration script in backend/scripts/migrate_subtitle_baks.py.
  • Subtitle modifier badges. HI / FORCED / SDH / CC badges next to language sidecars in series and movie views.
  • Auto-extract keeps only wanted languages. Scheduler-driven embedded-extraction extracts every text-based stream and trashes any sidecar outside the active profile's target_languages.
  • Provider result-rate metric. Two stacked bars on provider tiles — Treffer-Quote (result_rate) above Download-Quote (download_rate) — makes it obvious whether a 0%-download provider is dead or merely outranked.
  • Scheduler UI redesign. Nine jobs grouped into three SettingsSection cards (Scanning / Metadata sync / Maintenance) with status-dot rows + KebabMenu. ~70% less visual noise per row.
  • Anonymous OpenAPI discovery. GET /api/docs (Swagger UI) and GET /api/v1/openapi.json exempt from X-Api-Key. "Try it out" calls still authenticate via Authorize flow.
  • Compact per-pill action menu. Per-language sidecar row collapsed from ~22 controls to [Lang][Health][×][⋯] with all static actions in the dropdown.

Security

  • Cross-feature audit sweep across 23 areas (notifications, backup, webhooks, Whisper, subtitle tools, cleanup, translation, marketplace, blacklist, integrations, profiles, API keys, system logs, auth, health, onboarding, scoring, activity, providers, scheduler, dashboard, library, audit follow-ups).
  • Path traversal closed in 7+ places — extract-tracks (lang into output path), /api/v1/remux/backups/{restore,delete} (is_safe_path args reversed), /tools/video-sync + /tools/auto-sync symlink bypass, /tools/convert, /cleanup/non-target-subs, /wanted/<id>/extract, _resolve_trash_dir traversal.
  • SSRF guard hardening (P1 + R4-01). validate_service_url rejects loopback literals, decimal/octal/hex IPv4 encodings, IPv4-mapped IPv6, the unspecified address, the Docker bridge, and DNS hostnames that resolve to any of these. Applied at PUT /config, /mediaservers/test, /mediaservers/instances, /config/import.
  • XXE / Billion-Laughs (R4-02). anidb_mapper, anidb_sync, providers/podnapisi, standalone/nfo_parser switched from xml.etree.ElementTree to defusedxml.
  • Provider filename sanitisation (P2). werkzeug.secure_filename at the JSON-response boundary in wanted_search/search.py instead of patching each of 19 provider adapters.
  • Prompt-injection guard hardening (P3). Central translation/prompt_safety.py defends Trojan-Source / zero-width / line-separator / Unicode-Tag-block / soft-hyphen / invisible-operator injections across all 7 LLM backends. Batch-geometry cap (MAX_BATCH_LINES=100, MAX_BATCH_CHARS_TOTAL=50000) rejects 100k+ line attacks before any API cost.
  • Subprocess argv injection defence. _try_reflink passes -- before src/dst to cp; run_ffprobe wraps file_path with _safe_arg_path. A leading - in a media file basename cannot be parsed as a flag.
  • Forbidden-root blocklist. System roots (/etc, /proc, /sys, /dev, /run, /boot, /root, /var/log, /var/run, /var/lib, /usr, /lib, /lib64, /sbin, /bin) refused as remux_trash_dir and watched-folder paths.
  • Secret roundtrip masks no longer destroy values. Config update, media-server PUT, webhook secret all detect ***configured*** / ***abcd and re-merge with the stored value.
  • is_safe_path arg-order linter. tools/lint_is_safe_path.py walks every backend Python file and flags reversed-args calls. Wired into CI; the bug class had previously surfaced twice.
  • +95 security regression tests across test_security_prompt, test_security_download, test_security, test_extract_cleanup_audit_2026_05_09.

Fixed

  • Cleanup data-loss class — recoverable trash by default. Every executor (language_filter, format_upgrade, orphan_files, orphan_db) and POST /cleanup/orphaned/delete move files to .sublarr/trash/<date>/ instead of hard-deleting. Hard-delete available via explicit permanent_delete=true opt-in.
  • Mount-blip mass-delete defence. Every cleanup + standalone executor os.stats media_path first and aborts on OSError. NFS reconnects, Unraid share-reloads, USB disconnects no longer trigger a wholesale wipe.
  • Empty keep_languages aborts language_filter. Previously a misconfigured empty list deleted every sidecar.
  • format_upgrade resolves ISO language aliases (ger ↔ de, deu ↔ de, eng ↔ en). Modifier-aware (forced, sdh, hi, cc, signs are part of the bucket key). Verified live on prod by trashing 18 inferior peers (Akame, GARO, Watari-kun, Okami-san).
  • Atomic ffmpeg extraction + repair writes. Both now write to a same-directory tempfile and os.replace into place only on success.
  • legacy_frozen regression in wanted-search. Items hitting wanted_max_search_attempts while wanted_auto_translate was disabled were left with failure_kind=NULL and dropped from _filter_eligible forever. Now funnels through record_search_outcome. Migration a7f2e1c9d3b4 reapplies the resurrection UPDATE to ~2005 already-stuck rows.
  • Sonarr/Radarr quality-upgrade duplicates in wanted_items. When Sonarr swapped a file (WEBRip → WEBDL), upsert_wanted_item left the old row orphaned with error="File not found on disk". Upsert now sweeps shadowed siblings transactionally — same logical entity, same (target_language, subtitle_type), different file_path whose file no longer exists. Mount-blip-safe.
  • Standalone hardening (12 audit fixes). Per-folder rescan was silently broken since the manager refactor; mass-delete on transient mount-loss; orphan wanted_items after stale-series cleanup; watcher overwriting NFO-sourced metadata; resolver cache was dead code (every scan re-hit TMDB/TVDB/AniList); filename-stub cache poisoning; _scan_lock process-wide ineffective; upsert race on UNIQUE; TMDB 429/5xx surrender; /refresh-metadata blocking Flask workers; symlink cycles; bounded watcher stability-check pool.
  • Phantom subtitle tracks in Plex/Emby/Jellyfin. Repair backups named <base>.<lang>.bak.<ext> were parsed as ISO-639-2 "Bashkir" (bak). Backups now in .sublarr/backups/.
  • Episode-vs-pill action split. Per-episode ⋯ no longer duplicates sidecar-bound actions; per-pill ⋯ owns Vorschau/Timing/Auto-Sync/Video-Sync/Health.
  • Bare fetch() bypassed X-Api-Key interceptor. 8 helpers in api/system/tools.ts returned 401 with API-key auth enabled. All routed through shared api axios instance.
  • Atomic swap-rename for /tools/process/undo. 3-step swap with rollback so a second undo is a free re-undo.
  • Auto-translate worker app context — previously raised RuntimeError: Working outside of application context which the catch-all dropped silently.
  • Per-route limits + clamps on amplification routes. Audio / video / OCR / NFO / spell endpoints get rate caps + bounded parameter ranges. Webhook + scoring + log limits added.
  • Provider settings health rail dominated by disabled providers — only active providers' real conditions appear.
  • Stale auto_disabled in /providers/health — status mixin now sources from cooldown-aware enriched mirror.
  • /tools/convert UTF-8 decode failure on legacy SRTs — chardet detection + latin-1 lenient fallback.
  • /api/v1/search?limit=abc 500 → graceful fallback.
  • +19 more audit micro-fixes in 0.77.2 across notifications, backup, Whisper, marketplace, blacklist, integrations export, profiles, API keys, system logs, auth, health, onboarding, scoring, activity, scheduler, dashboard, library — see CHANGELOG.md [0.77.2-beta].

Changed

  • Translation backend recommendations. Dropped the broken anime-translator-v6 preset; UI suggests qwen2.5:14b / llama3.1:8b / cloud backends as general-purpose options.
  • Single source-of-truth filename parser (subtitle_filename.py) replaces per-call regexes in dedup_engine and scan_subtitle_sidecars. Correctly handles nested modifiers like ep.en.hi.bak.srt.

Tests

  • ~250 new tests landed across the cycle. Backend suite: 4900+ tests, full sweep green; +95 security regression tests; +49 cleanup + standalone audit cases; +67 waveform unit tests; +19 format-upgrade alias cases; +7 wanted-upsert prune cases.

Deployed image: ghcr.io/abrechen2/sublarr:0.92.6-beta
Docs: https://sublarr.de/docs/