Skip to content

CHANGELOG.md

Latest commit

 

History

History
1197 lines (915 loc) · 142 KB

CHANGELOG.md

File metadata and controls

1197 lines (915 loc) · 142 KB

Changelog

[1.32.0] - 2026-05-09

Added — Light Theme & Appearance Toggle (CEP Panel)

  • extension/com.opencut.panel/client/index.html: Added FOUC-prevention inline script in <head> that reads opencut_settings.theme from localStorage and applies html.theme-light before first paint. Added Appearance select (Auto / Dark / Light) to Settings > Preferences.
  • extension/com.opencut.panel/client/main.js: Added _applyTheme(pref) function that toggles html.theme-light based on explicit preference or OS prefers-color-scheme. Wired settingsTheme change event for instant switching. Added theme field to saveLocalSettings/loadLocalSettings. Added matchMedia listener to re-apply when OS theme changes in Auto mode.
  • extension/com.opencut.panel/client/style.css: Added comprehensive html.theme-light CSS block — 200+ lines of token overrides and component-specific rules covering: backgrounds, text, borders, sidebar, nav tabs, cards, workspace stage, buttons, form inputs, sub-tabs, empty states, toasts, alert/processing banners, results, job history, command palette, wizard overlay, wizard card, onboarding overlay, scrollbars, and focus rings. Light palette uses professional off-white surfaces (#f0f3f8 void → #ffffff cards), deep blue accent (#1e4fd8) with WCAG AA contrast, and light-calibrated shadows.

Changed — Premium UX/UI Polish (CEP Panel)

  • extension/com.opencut.panel/client/style.css: Full premium design-quality pass on the CEP panel stylesheet.
    • Replaced neon synthwave color system with a professional studio dark theme: accent blue #4d7fff, muted violet secondary, calmer semantic success/warning/error tokens, desaturated background layers.
    • Removed all decorative CSS animations: mesh background, scrolling grid, header gradient bar, logo shimmer, card border rotation, button shine sweeps, and neon button underglow.
    • Static subtle depth gradients and dot grid replace animated backgrounds.
    • Typography cleaned throughout: removed excessive letter-spacing on titles and labels, reduced font-weight on card headings, removed forced uppercase transforms.
    • Eliminated all remaining hardcoded rgba(0,255,255,...) and rgba(255,0,85,...) values — all colors now flow from CSS custom properties.
    • Updated CSS variable fallback strings to match new design tokens.
    • Fixed --text-muted typo (#6868888 → #686888).
    • .btn-primary: removed shine sweep (::before) and neon underglow (::after).
    • .btn-outline: replaced hardcoded neon rgba with design-system variables.
    • .filler-checks, .stem-check checked states: replaced neon glow box-shadow with clean variable-driven shadow.
    • .custom-dropdown-trigger:focus: replaced hardcoded rgba focus ring with var(--cyan-subtle).
    • .drop-zone.drag-over, .drop-zone.drag-active: use var(--cyan-subtle) background.
    • Wizard box-shadow: removed extraneous neon ambient glow.
    • .server-status-banner: replaced hardcoded error-red rgba with var(--error-dim) / var(--red-glow).
  • extension/com.opencut.panel/CSXS/manifest.xml: Updated ExtensionBundleVersion and extension Version to 1.31.0.

[1.30.0] - 2026-05-04

Added — Wave M (Dubbing Pipeline, Sports Highlights, EchoMimic, MCP Server)

  • core/dub_pipeline.py (POST /video/dub): Full automatic dubbing pipeline — delegates to the existing auto_dub_pipeline.auto_dub() engine. Supports 27 target languages, Whisper transcription, deep-translation, edge-tts/ElevenLabs/OpenAI voice synthesis, optional voice cloning, lip-sync, and music preservation. Route is @async_job; registered in _ALLOWED_QUEUE_ENDPOINTS.
  • core/highlights_sports.py (POST /video/highlights/sports): Optical-flow motion scoring (cv2.calcOpticalFlowFarneback at 2fps, 320×180 resize) + ffmpeg astats audio-energy scoring with genre-specific weighting (sports/concert/reaction/gaming/news). Returns non-overlapping top-N highlight windows sorted by start time. Route is @async_job; registered in _ALLOWED_QUEUE_ENDPOINTS.
  • core/lipsync_echomimic.py (POST /video/lipsync/echomimic): Two-backend EchoMimic lipsync — tries echomimic PyPI package first, falls back to diffusers + torch with BadToBest/EchoMimic HuggingFace weights (~2 GB auto-download to ~/.opencut/models/echomimic/). Override model directory via OPENCUT_ECHOMIMIC_MODEL_DIR env var.
  • mcp_server.py: MCP sidecar with 27 tools covering all major OpenCut capabilities. Added main() CLI entry point, --http mode (JSON-RPC 2.0 on port 5681), --port, --list-tools, and --log-level flags. Registered as opencut-mcp-server console script.
  • pyproject.toml: Added mcp optional dependency group (mcp>=1.0). Registered opencut-mcp-server console script.

Changed

  • routes/wave_k_routes.py: Wired route_dub_pipeline() and route_highlights_sports() — replaced _stub_501 with full @async_job implementations.
  • routes/jobs_routes.py: Added /video/dub and /video/highlights/sports to _ALLOWED_QUEUE_ENDPOINTS.
  • mcp_server.py: Added 4 Wave M tools (opencut_dub_video, opencut_sports_highlights, opencut_lipsync_echomimic, opencut_chat_edit) to MCP_TOOLS and _TOOL_ROUTES. Added audio_path to required path validation set.

[1.29.0] - 2026-04-23

Added — Wave L (Premium TTS, Smart Upscaling, AI Face Tools)

  • core/tts_elevenlabs.py (POST /audio/tts/elevenlabs, GET /audio/tts/elevenlabs/voices, GET /audio/tts/elevenlabs/info): ElevenLabs cloud TTS backend — 3,000+ voices across 32 languages. Gated by OPENCUT_ELEVENLABS_API_KEY environment variable; returns 503 with install/setup hint if absent. Supports all four synthesis models (multilingual v2, turbo v2.5, turbo v2, monolingual v1), voice stability, similarity boost, style exaggeration, and speaker boost. Voice can be specified by name or voice_id. Route is @async_job and registered in _ALLOWED_QUEUE_ENDPOINTS.
  • core/upscale_hub.py (POST /video/upscale/smart, GET /video/upscale/smart/info): Smart upscaling dispatcher that auto-selects the best available backend — FlashVSR (streaming SR) → Real-ESRGAN → Video2x → Lanczos (always available). Content hints (auto, fast, quality, anime, face, film) route to appropriate backends and models. Falls back to Lanczos on any backend failure. Route is @async_job. /info endpoint returns live backend availability and auto-selection table.
  • core/face_reshape.py (POST /video/face/reshape, GET /video/face/reshape/info): Liquify-style face warping using MediaPipe FaceMesh (468 landmarks). Supports five operations: slim_face, enlarge_eyes, shrink_nose, raise_cheekbones, smooth_jaw. Builds a smooth per-pixel displacement field from landmark deltas with Gaussian blending, applied via cv2.remap. Strength parameter 0–1. Route is @async_job.
  • core/skin_retouch.py (POST /video/face/retouch, GET /video/face/retouch/info): Automated skin retouching and blemish removal. Bilateral mode applies frequency-separation smoothing (edge-preserving bilateral filter + skin HSV mask + blend). gfpgan mode runs GFPGAN deep face restoration when available; falls back to bilateral if GFPGAN absent. Optional radiance parameter applies a subtle brightness boost to skin areas. Route is @async_job.
  • routes/wave_l_routes.py: New Flask blueprint wave_l_bp with 11 new routes.
  • checks.py: Added check_elevenlabs_available(), check_upscale_hub_available(), check_face_reshape_available(), check_skin_retouch_available(), and check_wave_l() summary function.

Changed

  • routes/__init__.py: Registered wave_l_bp — brings total blueprint count to 99.
  • routes/jobs_routes.py: Added 4 Wave L @async_job routes to _ALLOWED_QUEUE_ENDPOINTS.

[1.28.2] - 2026-04-22

Fixed

  • jobs_routes.py: Added 9 Wave K async routes to _ALLOWED_QUEUE_ENDPOINTS — they were implemented in v1.28.0 but never registered, making them unreachable via the job queue. Routes added: /audio/watermark/embed, /audio/watermark/detect, /settings/brand-kit/preview, /video/reframe/batch, /audio/censor/profanity, /audio/spectral-match, /video/lottie/render, /search/ai, /search/ai/index.
  • wave_k_routes.py: Changed validate_path → validate_filepath on all 16 remaining file-input parameters in Tier 2 synchronous routes (singing_vevo2, lipsync_echomimic, style_tokenflow, track_cutie, track_deva, flow_searaft, restore_diffbir, stabilize_gyroflow, deblur_motion, depth_depthpro, depth_flow, audio_reactive_fx, cinefocus, screenplay_parse, slate_id). Ensures file existence is validated before handing paths to heavy AI modules, giving callers a specific FILE_NOT_FOUND error instead of an opaque internal failure.
  • server.py: Removed duplicate comment block at the PID re-export section.

[1.28.1] - 2026-04-21

Fixed — Wave K Bugfix Pass (Phase 3/4 production hardening)

  • All 6 Wave K core modules: replaced bare "ffmpeg" / "ffprobe" string literals with get_ffmpeg_path() / get_ffprobe_path() from opencut.helpers. Fixes silent failure in PyInstaller-bundled distributions where FFmpeg is not on PATH.
  • All 6 Wave K core modules: wrapped subprocess.run(..., check=True) in try/except CalledProcessError and re-raised as RuntimeError with the FFmpeg stderr included. Debugging FFmpeg failures now gives actionable error messages.
  • clip_rating.py: Added threading.Lock() around all _load_db() / _save_db() cycles in rate(), tag(), untag(), search(), get() to prevent concurrent-request data loss. _save_db() now uses an atomic write (os.replace) to prevent JSON corruption on crash.
  • profanity_censor.py: Fixed critical bleep-mode logic bug — original audio was mixed with the bleep instead of being silenced during profanity windows, leaving the original speech fully audible. Bleep mode now correctly silences the original and gates a 1 kHz tone to those windows only. Added _get_audio_duration() ffprobe helper to generate a full-length bleep track.
  • spectral_match.py: Added -t 60 duration cap to the audio decode step, preventing OOM on long files. Analysis uses the first 60 seconds as a representative spectral sample.
  • lottie_import.py: Fixed silent blank-frame rendering — the PIL fallback created empty transparent frames with no warning. Now raises RuntimeError requiring Cairo. Fixed bg_color being accepted but not applied — compositing now uses filter_complex color + overlay.
  • semantic_search.py: Fixed build_index() which stored only timestamps — it now extracts the first frame with FFmpeg and computes a real CLIP embedding stored in the index. search() now uses pre-built embeddings instead of re-encoding every file per query. Fixed double CLIP import (clip_mod / clip_mod2). Added threading.Lock() for model load and index writes. _save_index() is now atomic via os.replace.
  • subtitle_qa.py: Fixed VTT regex that rejected HH:MM:SS.mmm timestamps (only MM:SS.mmm matched). Added file existence check at top of validate(). Renamed loop variable l to line. Added missing import os.
  • slate_id.py: Added _florence_cache dict to cache the Florence-2 processor and model across calls, avoiding a multi-GB reload on every invocation.
  • wave_k_routes.py: Added validate_filepath to security imports. Changed validate_path → validate_filepath on all file-input route parameters (batch_reframe, subtitle_qa, spectral_match × 3, lottie_info) to enforce file existence before handing paths to modules. Simplified Tier 3 stubs route_gen_video_wan_vace and route_highlights_sports to return _stub_501() directly, consistent with all other Tier 3 routes.

[1.28.0] - 2026-04-21

Added — Wave K Completeness Pass (40 features, 59 routes)

Wave K is the result of a four-angle competitive research pass spanning Kdenlive, DaVinci Resolve 19, CapCut Pro, Runway Gen-3, Pika 2.0, ElevenLabs, Adobe Firefly Video, Kling 2.0, and a broad sweep of 2024–2025 open-source AI papers. Features are tiered: Tier 1 ships working, Tier 2 ships as 503 stubs pending optional dependency installs, Tier 3 ships as 501 stubs on the roadmap for v1.29.0.

Tier 1 — Fully Implemented (8 modules, ~24 routes)

  • K1.1 AudioSeal Watermarking (audio_watermark.py) — AI-inaudible perceptual watermark embed and detect via Facebook Research AudioSeal; wired as a post-export background job so export latency is unaffected. Routes: POST /audio/watermark/embed, POST /audio/watermark/detect, GET /audio/watermark/info.
  • K1.2 Brand Kit (brand_kit.py, existed) — Brand-consistent lower-thirds, intro/outro overlays, and logo-burn. Routes now wired in wave_k_routes.py.
  • K1.3 Podcast Suite (podcast_suite.py, existed) — Multi-speaker transcription, chapter markers, audiogram export. Routes wired.
  • K1.4 Batch Reframe (batch_reframe.py) — AI-crop a folder of clips to any target AR in one call. Routes: POST /batch/reframe, GET /batch/reframe/status.
  • K1.5 Clip Rating (clip_rating.py) — Heuristic quality score (motion blur, low-light, handheld shake) so editors can bin rejects before cut. Routes: POST /clip/rating, GET /clip/rating/info.
  • K1.6 Subtitle QA (subtitle_qa.py) — Lint SRT/VTT for overlaps, encoding errors, and reading-speed violations. Routes: POST /subtitle/qa, GET /subtitle/qa/rules.
  • K1.7 Profanity Censor (profanity_censor.py) — Bleep or silence detected profanity with configurable word list and mode (bleep, silence, mute-word). Routes: POST /audio/censor/profanity, GET /audio/censor/profanity/wordlist.
  • K1.8 Spectral Match (spectral_match.py) — Match the spectral profile of one audio file to a reference using FFmpeg amultiply/EQ. Routes: POST /audio/spectral/match, GET /audio/spectral/info.
  • K1.9 Lottie Import (lottie_import.py) — Render Lottie JSON animations to transparent video via lottie-python or Puppeteer fallback. Routes: POST /lottie/render, GET /lottie/info.
  • K1.10 Semantic Search (semantic_search.py) — CLIP/SentenceTransformers frame-level index so editors can search footage by description. Routes: POST /search/ai, POST /search/ai/index, GET /search/ai/index/status.

Tier 2 — 503 Stubs, install-ready (19 modules, ~38 routes)

  • K2.1 GPT-SoVITS TTS (tts_gptsovits.py) — voice-cloned TTS via local GPT-SoVITS server on port 9880.
  • K2.2 Amphion TTS (tts_amphion.py) — neural TTS with style control.
  • K2.3 Vevo2 Singing (singing_vevo2.py) — AI singing voice synthesis.
  • K2.4 CosyVoice2 TTS (tts_cosyvoice2.py) — multilingual zero-shot voice cloning.
  • K2.5 EchoMimic LipSync (lipsync_echomimic.py) — audio-driven talking-head animation.
  • K2.6 TokenFlow Style Transfer (style_tokenflow.py) — temporally-consistent video style transfer.
  • K2.7 CUTIE Tracking (track_cutie.py) — high-accuracy object segmentation tracking.
  • K2.8 DEVA Tracking (track_deva.py) — open-world decoupled video segmentation.
  • K2.9 SEA-RAFT Optical Flow (flow_searaft.py) — state-of-the-art dense optical flow.
  • K2.10 DiffBIR Restoration (restore_diffbir.py) — face + general image/video restoration.
  • K2.11 Gyroflow Stabilization (stabilize_gyroflow.py) — gyroscope-data lens/stabilization via gyroflow binary.
  • K2.12 Motion Deblur (deblur_motion.py) — blind motion deblur via NAFNet or DeblurGAN-v2.
  • K2.13 Depth Pro (depth_depthpro.py) — Apple Depth Pro metric monocular depth estimation.
  • K2.14 DepthFlow (depth_flow.py) — 2.5D parallax video from still images using ModernGL.
  • K2.15 AudioGen SFX (sfx_audiogen.py) — text-to-sound-effects via Meta AudioGen (CC-BY-NC; weights not auto-downloaded).
  • K2.16 Open-Sora Generation (gen_video_opensora.py) — open-source text-to-video via Open-Sora.
  • K2.17 LTX-Video v2 A+V (gen_video_ltx.py) — joint audio+video generation; new route /generate/ltx/v2 preserves backward compat with existing /generate/ltx.
  • K2.18 Audio-Reactive FX (audio_reactive_fx.py) — beat-synced visual FX driven by audio energy bands.
  • K2.19 CineFocus Rack Focus (cinefocus.py) — AI-driven simulated rack focus using depth maps.

Tier 3 — 501 Stubs, v1.29.0 roadmap (8 modules, ~16 routes)

  • K3.1 Dub Pipeline (dub_pipeline.py) — multi-language dub orchestration (transcribe → translate → TTS → lipsync).
  • K3.2 Trailer Generator (trailer_gen.py) — highlights-driven auto-trailer with title card injection.
  • K3.3 Screenplay Parser (screenplay_parser.py) — FDX/Fountain parse → shot list; partial real implementation via fountain library.
  • K3.4 Face Age Transform (face_age_transform.py) — age/de-age face via SAM2 + inpaint.
  • K3.5 Slate Identify (slate_id.py) — OCR-based clapperboard reader; partial real implementation via Tesseract.
  • K3.6 Video Outpainting (outpaint_video.py) — extend frame borders via diffusion outpaint.
  • K3.7 WAN/VACE Generation (gen_video_wan_vace.py) — Wan2.1 + VACE video generation and editing.
  • K3.8 Sports Highlights (highlights_sports.py) — action-density + audio-peak highlight detection.

Infrastructure

  • opencut/routes/wave_k_routes.py — new blueprint wave_k with 59 handlers (0 route collisions on startup).
  • opencut/checks.py — 35 new check_X_available() functions appended.
  • opencut/routes/__init__.py — wave_k_bp registered; blueprint count now 98.
  • Route collision guard confirmed clean at startup: 1,334 total routes, 0 duplicate method+path pairs.

[1.25.1] - 2026-04-20

Fixed

Two user-reported bugs in the CEP panel: imported media not detected, and processed output landing next to the source clip instead of in the project folder.

  • CEP project-item walker — host/index.jsx _walkProjectItems() now uses getMediaPath() as the primary discriminator instead of an item.type === 2 + !item.getMediaPath heuristic that broke silently on Premiere 25.x. The old heuristic misclassified bins that expose getMediaPath as a function (most versions do), so nested media was never traversed. New logic: if getMediaPath() returns a non-empty string, record it; else if the item has children, walk it regardless of reported type.
  • Project-folder detection fallback — host/index.jsx getProjectMedia() returns a best-guess project folder even when app.project.path is empty (i.e. the .prproj hasn't been saved yet). Priority chain: saved project path → directory of the first imported media → scratch disk path → empty. A new projectFolderSource field in the response tells the panel which fallback fired. Stops processed files from landing in Downloads/ / network shares for unsaved projects.
  • Panel outputDir preference honoured — client/main.js introduces _recomputeEffectiveOutputDir() that layers the user's Settings → Output directory preference on top of the JSX-detected project folder. Previously the preference was saved to localStorage but ignored at job time; every output_dir: projectFolder call site now receives the resolved priority chain (user pref → detected project folder → empty). Recompute fires on settings save and on every media-list refresh.
  • Route ownership cleanup — removed ambiguous duplicate Flask registrations that let the wrong handlers win for /plugins/install, /resolve/import, and /api/macro/play. Local plugin-source install, Resolve import, and macro dry-run now resolve to the intended synchronous handlers again; the async marketplace and developer-only variants were moved onto explicit non-colliding endpoints.
  • Second collision sweep — removed four more overlapping registrations discovered in the full route table: /api/scripting/execute, /audio/me-mix, /audio/dialogue-premix, and /timeline/diff. The stable public endpoints now have a single owner, while the shadowed variants moved onto explicit paths (/api/workflow/scripting/execute, /audio/me-mix/basic, /audio/dialogue-premix/basic, /timeline/otio-diff) instead of silently competing at startup.
  • Runtime route-collision guard — opencut/routes/__init__.py now exposes collision detection helpers used by server.py to fail fast during create_app() if any later blueprint or plugin registration reintroduces a duplicate method/path pair. This turns a previously silent startup-order bug class into an immediate, diagnosable error.
  • Deterministic blueprint registration — opencut/routes/__init__.py no longer relies on locals() reflection to discover blueprints. Built-in route registration now comes from an explicit ordered tuple, which makes startup easier to audit, keeps registration intent readable, and gives tests a stable helper (get_core_blueprints()) to validate directly.
  • Offline-safe webhook URL validation — opencut/core/url_safety.py no longer performs DNS lookups for ordinary public hostnames. Literal localhost/private/reserved IP targets are still blocked, but valid public domains now pass deterministically in offline/test environments instead of failing with hostname does not resolve.
  • Resolve import input hardening — system.py now accepts either a single filepath or a validated paths list on /resolve/import, and missing/invalid path payloads return 400s consistently instead of falling through to the wrong integration route and surfacing as 500s.
  • CEP panel premium polish — extension/com.opencut.panel/client/main.js, index.html, and style.css now preserve per-tab scroll position, move focus correctly from the skip link into the main workspace, calm alert/progress copy, rename history surfaces to activity where that better matches the UX, and add a final shell-state refinement pass so banners, cards, results, and nav states feel more consistent and readable.
  • CEP accessibility hardening — the panel now traps focus inside the active command palette or frame-preview overlay, restores focus to the right launcher on close, keeps custom dropdown keyboard flow anchored to the trigger, stops dropdown wheel scrolling from leaking into the page, keeps warnings on screen until dismissed, and raises disabled/placeholder/focus contrast on key interactive surfaces.
  • CEP workflow refinement — overflowed sub-tab strips now advertise when more tools are off-screen and respond to vertical wheel scrolling as horizontal navigation, while high-risk numeric fields clamp themselves on blur and keep highlight/short-duration min/max pairs internally consistent before a run starts.
  • CEP packaging cleanup — vite.config.mjs now treats CSInterface.js and main.js as explicit classic CEP runtime assets during Vite builds, so client/dist/ contains the real CEP-ready script files without the earlier classic-script build warnings.
  • Panel dead-code reduction — client/main.js dropped unused helpers and stale state, fixed the WebSocket completion fallback to re-enter polling correctly, and reduced the panel lint backlog from 52 warnings to 25 without changing intended behavior.
  • Frontend build dependency restored — extension/com.opencut.panel/package.json now declares terser so the existing Vite production build can honor minify: "terser" without failing on a clean install.

Gotchas

  • JSX walker is getMediaPath-first — sequences and offline media both return empty getMediaPath(); the difference is sequences have 0 children. Both are now skipped cleanly. Bins with type reported as undefined / 0 / 3 on newer Premiere builds are now walked via the "has children" fallback.
  • projectFolder is now the effective output dir, not the raw detected value. The JSX-detected folder lives on the panel side in _detectedProjectFolder. If something needs the raw value, read that — don't read projectFolder.

[1.25.0] - 2026-04-19

Added — Wave H: Commercial Parity & Content-Creator Polish

16 new core modules + 1 new blueprint (wave_h_bp) + 34 new routes (1,241 → 1,275). Extends the OSS survey with commercial-product patterns (Opus Clip, Descript, CapCut, ScreenStudio) and post-April-2026 AI projects (FlashVSR, ROSE, Sammie-Roto-2, OmniVoice, ReEzSynth, VidMuse, VideoAgent, ViMax, Hailuo 2.3, Seedance 2.0, GaussianHeadTalk, FantasyTalking2).

Tier 1 — Content-creator polish (fully working):

  • Virality scoring — new core/virality_score.py + POST /analyze/virality, POST /analyze/virality/rank. Multimodal 0–100 heuristic score blending audio-energy peaks (existing silence.py), transcript hook density (LLM when available, keyword lexicon fallback), and visual salience (FFmpeg signalstats). ViralityResult + ViralitySignals subscriptable for jsonify. Stable-sorted ranking. Opus Clip inspiration.
  • Cursor-zoom sidecar parsing — extended core/cursor_zoom.py + POST /video/cursor-zoom/resolve. Parses ScreenStudio / Screen.Studio / OBS click-log sidecars, falls back to inline events, falls back to the existing frame-diff detector. All coordinates clamped to [0, width] × [0, height].
  • Changelog feed — new core/changelog_feed.py + GET /system/changelog/latest, GET /system/changelog/unseen, POST /system/changelog/mark-seen. Fetches GitHub releases, 15 min cache, graceful offline fallback, persists last-seen tag in ~/.opencut/changelog_seen.json.
  • Issue report bundle — new core/issue_report.py + GET|POST /system/issue-report/bundle. Scrubs all HOME paths to ~ before including crash.log + log tail. Returns a pre-filled GitHub issue URL. 60 KB body cap.
  • Demo bundle — new core/demo_bundle.py + GET /system/demo/list, GET /system/demo/sample, POST /system/demo/download. Serves opencut/data/demo/sample.mp4 when the installer ships one; dev installs can pull it from a GitHub release asset.
  • Gist preset sync — new core/gist_sync.py + POST /settings/gist/push, POST /settings/gist/pull, GET /settings/gist/info. Pure-stdlib urllib; refuses anonymous secret gists; rejects files > 2 MB; whitelists .json/.jsonl/.txt/.md extensions.
  • Onboarding state — new core/onboarding.py + GET|POST /settings/onboarding. Per-profile {seen, step, updated_at} persisted in ~/.opencut/onboarding.json. Delete the file to re-trigger the tour.

Tier 2 — AI model stubs (503 MISSING_DEPENDENCY):

All six ship as check_X_available()-gated stubs returning 503 with install hints, matching the v1.18–1.20 pattern. Full wiring deferred to v1.26.0+ once each upstream pins a stable Python entry point.

  • FlashVSR — streaming diffusion VSR (CVPR'26). core/upscale_flashvsr.py + POST /video/upscale/flashvsr, GET /video/upscale/flashvsr/info.
  • ROSE — video inpainting that preserves shadows/reflections. core/inpaint_rose.py + POST /video/inpaint/rose.
  • Sammie-Roto-2 — AI rotoscoping with VideoMaMa segmentation + in/out markers. core/matte_sammie.py + POST /video/matte/sammie.
  • OmniVoice — zero-shot TTS with 600+ languages. core/tts_omnivoice.py + POST /audio/tts/omnivoice, GET /audio/tts/omnivoice/models.
  • ReEzSynth — flicker-free Ebsynth successor. core/style_reezsynth.py + POST /video/style/reezsynth.
  • VidMuse — video-to-music generation (CVPR'25). core/music_vidmuse.py + POST /audio/music/vidmuse.

Tier 2 — Infra stubs:

  • QE reflection probe — wave_h_routes.py + GET|POST /system/qe-reflect. Panel POSTs the output of ocQeReflect() (host JSX); server caches it in ~/.opencut/qe_reflect.json for future API discovery. Host-side JSX wiring lands in v1.26.0 (see ROADMAP-NEXT Wave H2.8).

Tier 3 — Strategic stubs (501 ROUTE_STUBBED):

All routes return 501 with install hints — promoted to Tier 2 once usage signal confirms demand or upstream licences clarify.

  • VideoAgent + ViMax — core/video_agent.py + POST /agent/search-footage, POST /agent/storyboard. HKUDS agentic pipelines.
  • Hailuo 2.3 / Seedance 2.0 — core/gen_video_cloud.py + POST /generate/cloud/submit, GET /generate/cloud/status/<id>, GET /generate/cloud/backends. Cloud gen-video backends.
  • GaussianHeadTalk + FantasyTalking2 — core/lipsync_advanced.py + POST /lipsync/gaussian, POST /lipsync/fantasy2, GET /lipsync/advanced/backends. Wobble-free lip-sync alternatives to LatentSync/MuseTalk.

Infrastructure

  • 16 new check_*_available() entries in opencut/checks.py (Wave H block).
  • 3 new async routes added to _ALLOWED_QUEUE_ENDPOINTS: /analyze/virality, /analyze/virality/rank, /video/cursor-zoom/resolve.
  • New blueprint wave_h_bp registered (continues wave_a/b/c/d/e/f/g/h naming).
  • New error code ROUTE_STUBBED returned at HTTP 501 for Tier 3 routes. Frontend treats 501 as "coming soon" (greyed-out with tooltip), never as a failed call.
  • ROADMAP-NEXT.md extended with a new Wave H section documenting every item and its upstream source.

Frontend wiring

  • client/main.js — ~330-line WaveH module appended at the end of the main IIFE. Staggered startup probes at 1.2/1.8/2.4/3.0 s to avoid racing the first health check. Exposes window.OpenCutWaveH for the command palette.
  • client/index.html — "Panel Polish" card injected into the Settings tab with 7 data-i18n labelled buttons: Try demo, Restart tour, Gist push, Gist pull, Virality, Cursor zoom, Send log.
  • client/style.css — .oc-wave-h-* + .oc-onboarding-* rules, dark-mode-first, with a 520 px narrow-panel media query.
  • client/locales/en.json — 10 new keys under wave_h.*.
  • host/index.jsx — three new ES3-safe functions: ocQeReflect() (QE API reflection probe), _ocDispatchEvent() (CSXS event dispatcher), ocEmitPingEvent() (panel-callable ack). All CSXS events use the com.opencut.<event> namespace.

Gotchas

  • Virality score is heuristic — no ML model. The absolute number is not comparable across video types; use for relative ranking only.
  • Cursor-zoom: never trust client-supplied sidecar coordinates — the clamp to [0, width] × [0, height] is non-negotiable.
  • Changelog feed never raises. Network / JSON errors return {releases: [], source: "fallback", note: "..."} so the panel keeps working offline.
  • Issue report bundle scrubs HOME paths but does NOT redact API keys / credentials that may leak into logs. Callers must still prompt the user to review the body before opening the GitHub URL.
  • Gist sync refuses anonymous secret gists — anonymous gists are always public. Secret gists require GITHUB_TOKEN env.
  • Tier 3 routes always return 501 ROUTE_STUBBED, even with optional deps installed. These are scaffolding for future releases, not "features waiting for a pip install".

[1.24.0] - 2026-04-17

Added — second wide-net pass

Five infrastructure modules + three governance files + one run_ffmpeg ergonomics extension. Closes the remaining wide-net findings from the v1.14.0 strategic-gap audit. 6 new routes (1,235 → 1,241).

Subprocess tracking for cancel — helpers.run_ffmpeg now accepts an optional job_id=... keyword. When supplied, the spawned Popen is registered with the job subsystem via _register_job_process(job_id, proc) so POST /cancel/<job_id> actually terminates the child. Previous behaviour used blocking subprocess.run() — cancel marked the job "cancelled" but FFmpeg ran to completion, pinning a worker slot indefinitely. The non-job_id call path is unchanged, so all ~158 existing callers keep working; new / refactored callers opt in by passing job_id=job_id. Closes the v1.14.0 audit finding "untracked subprocesses can orphan on cancel".

Disk-space monitor — new core/disk_monitor.py + GET /system/disk, POST /system/disk/preflight, POST /system/disk/track. Tracks tempfile.gettempdir() + ~/.opencut by default; ops add more via register_tracked_path() or the /system/disk/track route. Warn / critical thresholds via OPENCUT_DISK_WARN_MB (2048) / OPENCUT_DISK_CRITICAL_MB (500). Optional background probe thread via OPENCUT_DISK_MONITOR_INTERVAL (default 0 = disabled). preflight(path, required_mb) gates a long render — fails open when the probe itself fails so disk-probe bugs don't cascade into job failures. Started from server.create_app() alongside the temp-cleanup boot.

Request-ID correlation — new core/request_correlation.py + GET /system/request-correlation. Installs Flask before_request / after_request / teardown_request hooks: generates r-{uuid[:16]}, stamps g.request_id, echoes on the response X-Request-ID header (RFC-standard), and attaches a logging.Filter so every log record under logger("opencut") carries the ID. Never trusts client-supplied X-Request-ID — the incoming value is sanitised into g.client_request_id for forensic value but the canonical ID is always server-generated. Prevents log-injection + server-log poisoning.

Deprecation registry — new core/deprecation.py + GET /system/deprecations. @deprecated_route(remove_in="2.0.0", replacement="/new/path", sunset_date="2026-10-01", reason="…") decorator stamps metadata on the view, emits RFC 8594 Deprecation: true + Sunset: ... + Link: <...>; rel="successor-version" response headers, and logs a one-time INFO record the first time each deprecated route is hit per process (including the caller's User-Agent). enrich_openapi_spec(spec, app) helper upgrades the existing OpenAPI generator output with deprecated: true + an x-opencut-deprecation extension block so clients see the warning at schema-generation time.

Governance docs:

  • SECURITY.md — vulnerability reporting process, supported-versions table (1.24.x active / 1.23.x previous / 1.22.x critical-only / ≤1.21 EOL), disclosure SLA (Critical 72h, High 7d, Medium 30d, Low 90d), known-safe-by-design surface, hardening recommendations.
  • CONTRIBUTING.md — 100-line repo tour + common-pattern reference card. Covers async route recipe, optional-dep guard, dataclass result shape, new run_ffmpeg(job_id=...) pattern, rate-limit categories, deprecation decorator.
  • scripts/sbom.py — pure-stdlib CycloneDX 1.5 SBOM generator. Walks pyproject.toml + requirements.txt, emits dist/opencut-sbom.cyclonedx.{json,xml}. First run catalogued 15 required + 71 optional deps. No cyclonedx-bom pip dep required.

Infrastructure

  • 3 new check_*_available() entries in opencut/checks.py: disk_monitor, request_correlation, deprecation_registry.
  • New blueprint wave_g_bp registered (continues wave_a/b/c/d/e/f/g naming).
  • server.create_app() now calls install_middleware() for request-ID wiring + disk_monitor.start_background() for optional background monitoring.

Gotchas

  • run_ffmpeg default path is unchanged — no job_id means the fast subprocess.run path runs exactly as it did pre-v1.24. Only callers that opt in (job_id=job_id) get cancel support. New code should always pass job_id when in an @async_job context.
  • Request-ID middleware regenerates even when the client sends one — this is intentional (prevents log-injection). The client-supplied value is preserved in g.client_request_id for investigation but never echoed back as-is. If downstream systems need client-controlled correlation, use a separate header (e.g. X-Client-Correlation-ID) that we log but don't adopt as the server ID.
  • Disk-monitor preflight fails open — when the probe itself errors (permissions, odd filesystem), it returns ok=True + a diagnostic note. Callers needing guaranteed refusal should check note and bail explicitly. Fail-open matches the existing check_disk_space() semantics.
  • Deprecation registry is keyed by func.__qualname__ at decoration time. When the OpenAPI generator is enriched later via enrich_openapi_spec(spec, app), it walks __wrapped__ to recover the metadata. Keep decorator order @bp.route → @deprecated_route → @require_csrf → @async_job so the walk finds the info.
  • SBOM scope — the generator reads declared deps from pyproject.toml + requirements.txt, not installed ones. Use cyclonedx-py / syft against a populated venv when you need the installed-packages SBOM (e.g. for CVE scanning the actually-running set).
  • X-Request-ID charset is strictly alphanumeric + - _ : ., cap 64 chars. Clients that want richer IDs should derive them from UUID4 hex + hyphens.

[1.23.0] - 2026-04-17

Added — wide-net infrastructure pass

Four cross-cutting modules + one blueprint. No new features — closes long-standing gaps flagged in the v1.14.0 strategic audit and the original ROADMAP.md Wave 3A. 7 new routes, 1,228 → 1,235.

OpenAPI 3.1 spec generator — new core/openapi_spec.py + GET /api/openapi.json, GET /api/docs, GET /api/routes. Walks Flask's url_map, introspects view-function docstrings, detects @require_csrf by walking the wrapper chain, converts Flask <int:foo> path converters to OpenAPI {foo} parameters, and emits a valid 3.1 spec (1,208 paths / 97 tags on a stock install). Swagger UI hosted inline via CDN-loaded script — no extra dep. Per-endpoint schema_hints escape hatch lets callers attach real request/response schemas without rewriting the generator. Closes Wave 3C "FastAPI migration is deferred" with a zero-migration alternative.

GPU-exclusive semaphore — new core/gpu_semaphore.py + GET /system/gpu-semaphore. Ships the minimal version of the Wave 3A P0: a process-wide threading.Semaphore (default MAX_CONCURRENT_GPU_JOBS=3, env-tunable) that routes wrap via @gpu_exclusive on their inner worker body. acquire() honours a configurable OPENCUT_GPU_ACQUIRE_TIMEOUT (default 0 = non-blocking 429). Status endpoint reports active / available / rejected_total / acquired_total for ops dashboards. Upgrade path stays open — the decorator boundary is the same whether the implementation becomes an in-process semaphore or a multi-process worker pool later.

Rate-limit categories — new core/rate_limit_categories.py + GET /system/rate-limits. Four categories (gpu_heavy, cpu_heavy, io_bound, light) with env-tunable ceilings. Each category gets a threading.Semaphore, an acquire / release API, a @rate_limit_category("gpu_heavy") decorator, and a category_of(func) helper so introspection / OpenAPI hints can surface the category on the UI. Closes the v1.14.0 audit finding "4 % of async routes are rate-limited" by reducing the per-route cost of opt-in to a single decorator.

Temp-file startup + periodic sweep — new core/temp_cleanup.py + POST /system/temp-cleanup/sweep, GET /system/temp-cleanup/status. Runs on create_app(): walks tempfile.gettempdir() for opencut_* + opencut-* entries older than a configurable TTL (default 1 h) and removes them. Daemon thread re-sweeps every OPENCUT_TEMP_CLEANUP_INTERVAL seconds (default 3600, 0 disables). Symlink-out-of-tempdir defence via realpath() + commonpath() check. Logs removed N files / N dirs / N.N MB summary per sweep. On the dev machine, the startup sweep of a long-running box reclaimed 4.6 MB across 45 files + 397 dirs.

Infrastructure wiring

  • server.py::create_app now calls temp_cleanup.run_startup_sweep() + start_background_sweep(). Any failure logs at warning and the app continues.
  • 4 new check_*_available() entries in opencut/checks.py (all stdlib-only, so all return True unless the module import itself fails).

Gotchas

  • OpenAPI spec is cached module-level — the first request walks 1,200+ routes; subsequent requests serve from _OPENAPI_CACHE. ?refresh=1 bypasses the cache. Spec drift between deploys is a concern for deploy-time validation, not runtime.
  • @gpu_exclusive goes on the inner worker body, not the Flask route — the @async_job decorator already handles queue back-pressure for the HTTP layer. Putting @gpu_exclusive on the route would mean the HTTP caller waits; putting it on the inner body means the worker thread waits, which is where the actual GPU contention happens.
  • Rate-limit categories coexist with the existing security.rate_limit() / require_rate_limit() — the new category API is for category-scoped work (gpu_heavy, cpu_heavy, io_bound, light). The legacy rate_limit("model_install") API is still the right choice for single-op gates like pip-install serialisation.
  • temp_cleanup prefix match is strict basename-only — we never walk into arbitrary user files. Operators can extend the prefix list via OPENCUT_TEMP_CLEANUP_PREFIXES (comma-separated) but should only add prefixes they own.
  • _route_uses_csrf walks __wrapped__ — add your own decorators above @require_csrf in the stack order @wave_X_bp.route → @require_csrf → @async_job to keep introspection accurate. The OpenAPI spec's CSRFToken security block is derived from this walk.
  • Startup temp sweep is best-effort — logs a warning on failure; never blocks app startup. A corrupt tempfile permission won't prevent the server from coming up.

[1.22.0] - 2026-04-17

Added — Shaka Packager / OBS bridge / RunPod / Plausible telemetry

Four new core modules, 11 new routes (1,216 → 1,228). All opt-in via env + optional deps, with graceful degradation.

B5.2 Shaka Packager HLS/DASH/CENC — new core/shaka_pkg.py + POST /delivery/shaka/package, GET /delivery/shaka/info. Wraps the standalone packager binary (BSD-3, Google). Stream-descriptor builder emits one entry per rendition (video/audio/text), with safe basename sanitisation so user stream labels can't smuggle filter characters. Optional CENC DRM with AES-128 raw-key schema for Widevine / PlayReady / FairPlay / Common System. LL-HLS gate validates protocol=hls. Resolution cache (_AVAILABILITY_CACHE) probes PATH candidates once per process. 2 h default subprocess timeout.

E6 OBS WebSocket v5 bridge — new core/obs_bridge.py + POST /integration/obs/status, POST /integration/obs/switch-scene, POST /integration/obs/recording, POST /integration/obs/screenshot. Implements the v5 JSON-over-WebSocket protocol with SHA-256 challenge auth inline — no dedicated obs-websocket-py dependency. Uses the existing websockets pip package (already checked via check_websocket_available). Blocking sync API with a single background event-loop thread per client. ObsClient is a context-manager (with block). High-level helpers: ping(), status(), switch_scene(), recording(), take_screenshot(). Unlocks the Gaming vertical (Twitch VOD clipping), live-production cue sheets, tutorial-capture auto-split.

E9 RunPod serverless render — new core/runpod_render.py + POST /cloud/runpod/submit, GET /cloud/runpod/status/<endpoint>/<job>, POST /cloud/runpod/cancel, GET /cloud/runpod/info. Pure-urllib transport (no hard pip dep); runpod SDK is an optional ergonomic enhancement. Submit / status / wait / cancel entry points with exponential-ish backoff in wait(). _ENDPOINT_ID_RE prevents URL-injection via malformed endpoint IDs. API key from RUNPOD_API_KEY env or explicit argument — never logged. Sync mode via /runsync for small jobs; default /run returns a job_id to poll.

D5.3 Plausible telemetry — new core/telemetry_plausible.py + POST /telemetry/plausible/track, GET /telemetry/plausible/info. Fire-and-forget event emitter to a self-hosted Plausible instance (AGPL-3). Single background worker thread reads from a bounded queue (_MAX_QUEUED=500, drops oldest on overflow). _scrub_props() hardens user-supplied props: drops _-prefixed keys, caps string lengths to 120 chars, caps dict size to 30 entries, rejects non-alphanumeric key chars. Opt-in via PLAUSIBLE_HOST + PLAUSIBLE_DOMAIN env vars — absent env vars = complete no-op, zero overhead.

Infrastructure

  • 5 new check_*_available() entries in opencut/checks.py: shaka, obs_bridge, runpod (always True, stdlib), runpod_api_key_set, plausible_configured.
  • 1 new route in _ALLOWED_QUEUE_ENDPOINTS: /delivery/shaka/package.
  • New blueprint wave_e_bp registered (v1.22 continues the wave_a/b/c/d/e sequence).

Gotchas

  • Shaka extra_args is power-user escape hatch — args are passed verbatim to the packager binary without validation. Don't expose this via untrusted surfaces. The stream-descriptor builder does sanitise basenames to reject filter-chain metacharacters.
  • OBS ObsClient is not thread-safe — one client per thread; sharing across threads will drop responses because _pending dispatch is lookup-by-request-id, not broadcast. The high-level helpers (status, switch_scene etc.) create a fresh client per call, which is the right pattern for low-volume use.
  • OBS background thread uses daemon=True — clean process exit won't block on a hung OBS connection. If you need guaranteed flush, call client.close() explicitly.
  • RunPod API key env fallback — the submit / status_of / cancel / wait functions all honour RUNPOD_API_KEY when api_key=None. Be careful not to log request bodies that might be echoed back by a misbehaving handler — the SDK fallback path returns raw output verbatim.
  • RunPod wait backoff caps at 30s — poll interval grows ×1.25 every 3 polls. For long-running jobs this is fine; for very short jobs the client may observe COMPLETED on the first poll regardless.
  • Plausible worker is daemon=True, survives the event loop — shutdown() is provided for clean test teardown but normal daemon-thread behaviour is correct for production. queue_depth() lets ops spot backlog.
  • Plausible drops oldest on overflow — _MAX_QUEUED=500. Telemetry is best-effort, so a long Plausible outage doesn't pin memory. If operators care about durability of events, Plausible isn't the right sink — use a proper queue + OpenTelemetry stack.

[1.21.0] - 2026-04-17

Added — Wave B delivery + Wave E voice grammar + D4.2 fuzz harness

Four new core modules, one fuzz-test harness, 9 new routes. 1,207 → 1,216 total.

B5.1 VVC / H.266 export — new core/vvc_export.py + POST /video/encode/vvc, GET /video/encode/vvc/info. Wraps FFmpeg's libvvenc (Fraunhofer HHI, BSD-3). Three presets (faster/balanced/archive), QP override, optional raw .266 / .vvc elementary-stream output for HLS packagers. Availability cached — module only probes FFmpeg's encoder list once per process.

B5.4 SRT streaming — new core/srt_streaming.py + POST /video/stream/srt/start, POST /video/stream/srt/stop, GET /video/stream/srt/info. Spawns FFmpeg → srt:// push (caller) or listen (listener). URL builder validates host/port/passphrase/stream-id, and reserved param names (mode, latency, pkt_size, passphrase, streamid) are immune to extra_params overrides. The 1.5 s probe window on start catches early subprocess exits so clients learn about URL/auth failures before polling. stop_stream() does SIGTERM → wait → SIGKILL / Windows taskkill /F fallback.

B2.2 colour-science scopes — new core/color_scopes_pro.py + POST /video/scopes/pro, GET /video/scopes/pro/info. Mathematically-proper scopes via the colour-science pip package (BSD-3): mean CIE 1931 xy chromaticity, mean CIELUV (u*, v*), and gamut coverage fractions for Rec.709 / DCI-P3 / Rec.2020 computed with a vectorised barycentric test. Optional matplotlib PNG plots (chromaticity horseshoe + LUV vectorscope) — stats return even when matplotlib is absent.

E4 voice-command grammar — new core/voice_command_grammar.py + POST /voice/grammar/parse, GET /voice/grammar/catalogue. Deterministic grammar (no LLM) for hands-free timeline editing. Verbs: cut / trim / mark / slip / nudge / speed / slow / undo / redo / seek / mute / unmute / ripple / zoom_in / zoom_out. Frames + seconds + minutes + beats all normalise to seconds given sequence fps + bpm. Number-words 0..99 parsed inline. parse() is advertised as never-raising — fuzzed to verify. Low-confidence parses emit fallback_route="/nlp/command" for broader LLM-backed dispatch.

D4.2 Atheris fuzz harness — new tests/fuzz/test_parser_fuzz.py. Five fuzz targets: SRT time parser, SRT file parser, .cube LUT parser, voice-grammar parser, event_moments._find_spikes. Opt-in via RUN_FUZZ=1; default pytest runs skip the infinite-loop libFuzzer entry points. Pytest-smoke variant runs a fixed deterministic payload suite per target so CI catches "import-broke-the-fuzz-target" regressions without needing Atheris at ordinary CI time.

Infrastructure

  • 6 new check_*_available() entries in opencut/checks.py: vvc, srt, colour_science, voice_grammar (always True), atheris, plus the deprecation of _AVAILABILITY_CACHE dicts to per-module.
  • 3 new routes in _ALLOWED_QUEUE_ENDPOINTS: /video/encode/vvc, /video/stream/srt/start, /video/scopes/pro.
  • New blueprint wave_d_bp registered (naming: v1.18/19/20 = wave_a/b/c; v1.21 = wave_d).

Gotchas

  • _AVAILABILITY_CACHE is per-module — both vvc_export and srt_streaming memoise their FFmpeg-capability probe. Cache lives for the lifetime of the Python process. A rebuild that flips libvvenc on/off requires a server restart before the cached False becomes True.
  • SRT stop is platform-split — POSIX uses signal.SIGKILL after the grace window; Windows falls back to taskkill /F /PID. Don't remove the Windows branch — Windows Python lacks SIGKILL entirely.
  • Voice-grammar is intentionally narrow — the grammar is a whitelist, not a classifier. Utterances it doesn't match emit fallback_route="/nlp/command" so the broader LLM-backed parser in core/nlp_command.py picks up stragglers. Don't grow the grammar to catch everything — that's what the fallback is for.
  • colour-science sample-count cap is 120 — probing 120 frames at 256² is fast (seconds). Larger values are silently clamped. The xy / LUV means are computed on concatenated pixel arrays across all samples, not per-frame averages, so uniform distributions don't wash out interesting chromaticity.
  • VVC encode timeout is 4 hours — reflects real-world archival encode times at preset=archive on ≥30-minute sources. Lower it for faster preset users only if monitored.
  • Fuzz tests skip by default — RUN_FUZZ=1 pytest tests/fuzz/ enables the deterministic smoke suite. libFuzzer-style real fuzzing uses python -m tests.fuzz.test_parser_fuzz <target> -max_total_time=N.

[1.20.0] - 2026-04-17

Added — Wave D remaining items (D1.1, D3.1, D4.1, D5.2, D6.2)

Three new modules, two LLM-enrichment extensions of existing modules, one server-level observability hook. 1,202 → 1,207 total routes (+5).

D3.1 Semantic OTIO timeline diff — new opencut/export/otio_diff.py + POST /timeline/diff. Loads two timelines through OpenTimelineIO's adapter framework (supports .otio, .otioz, FCP XML, EDL, AAF when the optional adapter is installed) and emits a structured OtioDiffResult with track-level, clip-level, and marker-level diffs. Clip matching by (name, source_url) tuple with order-preserving pair-up; different (start, duration) → retimed, different track index → moved. Includes format_diff_text() for CLI / log output.

D4.1 Objective quality metrics — new opencut/core/quality_metrics.py. VMAF / SSIM / PSNR comparison of a distorted file against a reference via FFmpeg's built-in filters + libvmaf JSON log. Three entry points: measure_vmaf() / measure_ssim() / measure_psnr() for single metrics, compare_videos() for a full QualityReport, and batch_compare() for CI golden-regression suites. Routes: POST /video/quality/compare, POST /video/quality/batch-compare, GET /video/quality/backends. Handles libvmaf absence gracefully — SSIM/PSNR still run; VMAF emits a clear notes entry. Threshold-based passes gate for CI regression budgets.

D5.2 Optional Sentry / GlitchTip observability — new _init_sentry_if_configured() hook in opencut/server.py, called from create_app(). Activated when SENTRY_DSN env var is set AND sentry_sdk is installed — no behaviour change for installs without either. Reads OPENCUT_SENTRY_ENV, OPENCUT_SENTRY_RELEASE, OPENCUT_SENTRY_SAMPLE_RATE. send_default_pii=False so request bodies (which often carry user media paths) don't leak to the error tracker. Reports via GET /observability/status.

D1.1 LLM-enriched audio description — new describe_scene_llm() in opencut/core/audio_description.py. Takes a timestamp + surrounding transcript + optional brightness hint, asks the configured LLM (core/llm.py) to produce a single present-tense AD line of ≤ ~3·duration words. No vision model needed — pure text-to-text. Falls back to the heuristic template on any LLM error so AD rendering never aborts mid-job. Complements the existing template-based describe_visual_content(); callers can pick per invocation.

D6.2 LLM-enriched quiz generation — new generate_quiz_questions_llm() in opencut/core/quiz_overlay.py. When llm_config is provided, produces genuine comprehension questions with shuffled correct_index. Strict JSON output parsed defensively — markdown code fences stripped, malformed entries dropped. Falls back to the existing TF-IDF fill-in-the-blank generator on any failure, so callers pass llm_config unconditionally.

Infrastructure

  • 4 new check_*_available() entries in opencut/checks.py: otio_diff, quality_metrics, vmaf, sentry.
  • 3 new routes added to _ALLOWED_QUEUE_ENDPOINTS so they're queueable via /queue/add.
  • New blueprint wave_c_bp registered in routes/__init__.py (alongside wave_a_bp, wave_b_bp).

Gotchas

  • VMAF absence is non-fatal — a FFmpeg build without --enable-libvmaf returns a QualityReport with vmaf=None and a diagnostic in notes. Callers must check report.vmaf is not None before using the value. Windows Gyan.dev / BtbN builds include libvmaf; many Linux distro packages don't.
  • Sentry init is idempotent — _SENTRY_INITIALISED guards re-entry. create_app() can be called multiple times in tests without creating multiple hubs. Environment variables are read every call, so changing them between create_app() invocations in test fixtures won't take effect until a fresh interpreter.
  • OTIO AAF reads require the adapter even for diff — /timeline/diff on .aaf inputs fails with MISSING_DEPENDENCY unless otio-aaf-adapter is installed. The error surfaces through the job.error event (the route is @async_job), not inline.
  • LLM quiz JSON fence stripping is defensive — the model may wrap output in ```json ... ```, ```...```, or ~~~...~~~. We strip all three before json.loads. A genuinely non-JSON response falls back to TF-IDF — no error surfaces to the caller.
  • describe_scene_llm word cap is approximate — the max_words + 4 tolerance (vs strict max_words) handles short models that over-shoot by one or two words without truncating natural sentence endings. AD burn-in downstream clamps display duration to the gap length, so a slightly long line is still safely bounded.

[1.19.1] - 2026-04-17

Fixed — production hardening audit on v1.17.0–v1.19.0 additions

  • Broken __import__("PIL.Image") hack in core/matte_birefnet.py cutout branch — the previous code relied on Python's __import__ returning the top-level PIL package, which happened to expose .Image only because from PIL import Image had run earlier in the same function. Fragile, linter-hostile, and confusing to read. Replaced the cutout compositing with a clean from PIL import Image + import numpy as np path, extracted into a _compose_output() helper that also owns the alpha / rgba / cutout branching. Removed the _ = np load-bearing noise line that existed solely to suppress an unused-import lint on the broken code path.
  • Apples-to-oranges score comparison in core/event_moments.py::_find_spikes — the "keep the louder of two close picks" branch compared a raw RMS value (e, typically 0..0.5) against a normalised score (moments[-1].score, 0..1) scaled by mean — different units entirely. Result: the replacement branch almost never fired, so min_spacing was ignored for clustered spikes and duplicate moments survived ranking. Fix: track the raw RMS magnitude of the last pick (last_picked_e) separately from the normalised score used in the final EventMoment, and compare in consistent units. Also extracted an _build_spike() helper to eliminate the copy-pasted EventMoment construction.
  • Model-cache race in core/tts_f5.py — _MODEL_CACHE was a plain module-level dict. Two concurrent POST /audio/tts/f5 requests could both observe cached_model is None, both call F5TTS(model_type=...), both write — doubling VRAM for the race window and wasting ~5 s of load time each. Added _MODEL_LOCK (threading.Lock) around reads and writes; extracted _get_or_load_model() so the locking discipline lives in one place. When a different model is requested than the cached one, the previous is released via torch.cuda.empty_cache() before the new one is loaded — prevents two big torch models pinned in VRAM simultaneously.
  • Backend-cache race in core/neural_interp.py — same shape but less severe: _BACKEND_CACHE["checked"] was written before rife_cli / rife_torch, so a concurrent reader could see checked=True + empty values. Added _BACKEND_LOCK and write checked=True last inside the critical section.

Changed

  • AAF + OTIOZ exports are now async (POST /timeline/export/aaf, POST /timeline/export/otioz) — previously blocked the Flask request thread for tens of seconds on 500-segment timelines. Both promoted to @async_job with a shared _export_pre_validate hook so malformed segments lists short-circuit to HTTP 400 synchronously before a worker slot is consumed. Added to _ALLOWED_QUEUE_ENDPOINTS so they're queueable via /queue/add. Response shape now includes framerate + sequence_name echoes for downstream tooling.
  • core/clip_quality.py — removed the stale # noqa: F401 annotation on from PIL import Image in the piq fallback branch. The import is actually used two lines later (Image.open(fpath).convert("RGB")) — the noqa was a leftover from an earlier refactor and misled readers about the import's purpose.

Gotchas

  • F5-TTS model swap frees VRAM — when a caller requests a different model name than the cached one, _get_or_load_model() deletes the previous model + calls torch.cuda.empty_cache() before loading. This means rapidly switching between F5-TTS and E2-TTS causes repeated cold-loads. Stick to one model per deployment where possible.
  • Spike dedup uses raw RMS — _find_spikes compares raw e against last_picked_e, not the normalised .score field. If future callers add additional scoring criteria to EventMoment, remember the replacement check is by loudness, not by the scored axis.
  • Async AAF/OTIOZ error surface — exports now return 202 + job_id instead of an inline 503 when the adapter is missing. Clients must poll /status/{job_id} and inspect error + code fields. The pre_validate hook still short-circuits malformed requests with 400 synchronously.

[1.19.0] - 2026-04-17

Added — Wave A (remaining items) + Wave D2 + Wave D3.2

Seven new core modules + one jobs-layer hook; 10 new routes; 1,192 → 1,202 total routes. All backends optional and gated behind check_X_available(); FFmpeg fallbacks where applicable keep routes functional on minimal installs.

Wave A2.3 — Matting

  • BiRefNet still/keyframe matte — core/matte_birefnet.py, POST /video/matte/birefnet. ONNX Runtime (preferred, no torch) or transformers/HF fallback. Three output modes: alpha (α only), rgba (RGB + α), cutout (premultiplied RGBA). Targets the ZhengPeng7/BiRefNet HF repo by default. ONNX checkpoint via OPENCUT_BIREFNET_ONNX env var.

Wave A3.1 — Advanced Captions

  • Karaoke / kinetic captions — core/captions_karaoke_adv.py, POST /captions/karaoke-adv/render, GET /captions/karaoke-adv/presets. Six presets — fill, bounce, color_wave, typewriter, karaoke_glow, highlight_word — using libass override tags. PyonFX hook available but defers to built-in renderer until PyonFX-specific FX logic is wired in a follow-up. segments_from_whisperx_dicts() adapter accepts WhisperX segment dicts directly.

Wave A4.2 — Encoding

  • SVT-AV1-PSY — core/svtav1_psy.py, POST /video/encode/svtav1-psy, GET /video/encode/svtav1-psy/info. Three perceptually-tuned presets (social, web, archive). Uses the FFmpeg-integrated libsvtav1 path when available; falls back to a 2-stage rawvideo pipe through the standalone SvtAv1EncApp binary. Outputs .mp4 with libopus audio.

Wave D2 — Restoration

  • DDColor B&W colorisation — core/colorize_ddcolor.py, POST /video/restore/colorize. Frame-by-frame colorisation via ONNX Runtime (checkpoint path via OPENCUT_DDCOLOR_ONNX) or ModelScope (damo/cv_ddcolor_image-colorization). Extract → colourise → reassemble with original audio.
  • VRT / RVRT unified restoration — core/restore_vrt.py, POST /video/restore/vrt. Unifies denoise / deblur / super-res / real-world-SR / unified into one pass via sliding-window temporal inference (default window=8). ONNX or basicsr backend. The basicsr path is a forward-compatible stub — the ONNX path is the complete inference loop.
  • Neural deflicker — core/deflicker_neural.py, POST /video/restore/deflicker. 7-frame temporal sliding window via All-In-One-Deflicker ONNX checkpoint, with auto mode that falls back to FFmpeg's deflicker + tmix chain when the neural backend is absent. Audio stream-copied verbatim.
  • Restoration backends report — GET /video/restore/backends enumerates which backends are installed and provides install hints.

Wave D3.2 — Webhook auto-emit

  • Job-complete webhook events — jobs.py::_emit_job_webhook() automatically fires job.complete / job.error / job.cancelled events via core/webhook_system.fire_event() on terminal job status. Dispatched asynchronously through _io_pool so job finalisation is never blocked on outbound HTTP. Best-effort — any webhook failure is logged at warning and swallowed.
  • Webhook event catalogue route — GET /webhooks/events lists auto-emitted event types + payload schema + retry policy (3 attempts, 1 s / 5 s / 15 s backoff).

Infrastructure

  • 6 new check_*_available() entries in opencut/checks.py — birefnet, pyonfx, svtav1_psy, ddcolor, vrt, neural_deflicker.
  • 5 new routes added to _ALLOWED_QUEUE_ENDPOINTS.
  • _emit_job_webhook() hook in jobs._update_job — lazy-imports webhook_system, never blocks.

Gotchas

  • Webhook auto-emit is best-effort — a failing webhook must never block _update_job. The try/except around _emit_job_webhook() is load-bearing. Dispatch runs on the _io_pool worker; raise only inside the worker where it'll be logged.
  • BiRefNet outputs vary by checkpoint — different exported ONNX files emit different head shapes (single tensor, tuple, list). _postprocess_alpha() handles squeeze + sigmoid + clamp defensively; don't trust mask.shape assumptions when swapping checkpoints.
  • PyonFX is advisory — backend="pyonfx" is reported when the library is installed but the renderer still uses the built-in preset path until we wire per-syllable FX. Users with Aegisub-specific expectations should be warned the implementation is pending.
  • SVT-AV1-PSY integrated vs standalone — the integrated libsvtav1 path is ~2× faster than the rawvideo pipe + standalone binary because it skips an intermediate yuv420p10le encode. Prefer an FFmpeg build with libsvtav1 linked against the PSY fork; the standalone path is there as a fallback, not a goal.
  • VRT basicsr backend is a stub — the ONNX path is production-ready; the torch/basicsr path logs a warning and no-ops on frames. Users hitting the stub should either install the ONNX checkpoint or wait for the full torch wiring (Wave E tracking).
  • Deflicker auto-fallback path — when backend="auto" is chosen and the neural path is registered but fails at runtime (e.g. ONNX checkpoint missing but env var set), the code intentionally re-calls deflicker_video(..., backend="ffmpeg") to guarantee a usable output. Do not remove this recursion — early-stage users frequently mis-set OPENCUT_DEFLICKER_ONNX.
  • DDColor ModelScope import cost — modelscope.pipelines.pipeline("image-colorization") is slow (~10 s cold) and downloads weights on first use. Consider pre-warming via a background job on server start for production installs.
  • Job-complete event payload includes result — the full job result dict is included in the webhook payload. If a webhook endpoint sits over an untrusted network, users should register it via a domain filter (the existing _validate_webhook_url layer blocks private / loopback addresses by default).

[1.18.0] - 2026-04-17

Added — Wave A + Wave D (first batch from ROADMAP-NEXT.md)

Ten features built against the established @async_job + subscriptable- dataclass + graceful-degradation pattern. All new routes registered under the new wave_a_bp blueprint. 1,177 → 1,192 total routes (+15).

Wave A1 — Audio

  • F5-TTS zero-shot voice clone — core/tts_f5.py, POST /audio/tts/f5, GET /audio/tts/f5/models. Supports F5-TTS and E2-TTS models. Module-level model cache to avoid re-loading per request. Subscriptable F5Result.
  • WhisperX --diarize exposure — extended CaptionConfig with diarize, hf_token, min_speakers, max_speakers fields; _transcribe_whisperx now runs pyannote 3.x diarisation when diarize=True and an HF token is available (env HF_TOKEN or explicit). Falls back to plain transcription on pipeline failure — never kills a caption job over a diarisation error.
  • BeatNet downbeat detection — core/beats_beatnet.py, POST /audio/beats/beatnet. Supports offline (best) and realtime modes; configurable meter (3/4/6). Approximates a confidence score as the ratio of downbeat labels to expected-per-meter counts.

Wave A2 — Video Intelligence

  • Scene detection auto-dispatcher — core/scene_detect.py::detect_scenes_auto(), POST /video/scenes/auto. TransNetV2 (ML) → PySceneDetect → FFmpeg threshold priority. Keeps the original detect_scenes / detect_scenes_ml / detect_scenes_pyscenedetect entry points callable by name for users who want a specific backend.
  • CLIP-IQA+ clip quality scoring — core/clip_quality.py, POST /video/quality/score, POST /video/quality/rank. Samples frames via FFmpeg, scores each via pyiqa.create_metric("clipiqa+") (or piq.CLIPIQA fallback). INVERTED_AXES frozenset inverts noisy/blurry/dark so the aggregate is monotone-positive. Batch-rank up to 40 clips per call.
  • HSEmotion emotion arc — core/emotion_arc.py, POST /video/emotion/arc. 8-class emotion timeline (Neutral/Happy/Sad/Surprise/Fear/Disgust/Anger/Contempt). Prefers hsemotion_onnx backend (no torch needed), falls back to torch-based hsemotion. Emits mean probabilities, transition count, and an emotional-range score for highlight ranking.

Wave A4 — Encoding

  • ab-av1 VMAF-target encode — core/ab_av1.py, POST /video/encode/vmaf-target, GET /video/encode/vmaf-target/info. Wraps the ab-av1 CLI for "give me VMAF 95" one-shot encoding (SVT-AV1 / libaom-av1 / libx264 / libx265). Parses both the trailing-JSON format and the legacy crf N VMAF M regex output. Hard 3 h subprocess timeout.

Wave A5 — Interchange

  • OTIO AAF export — export/otio_export.py::export_aaf(), POST /timeline/export/aaf. Uses otio-aaf-adapter (Apache-2) to emit Avid-compatible AAF files. Graceful 503 MISSING_DEPENDENCY when the adapter isn't installed.
  • OTIOZ bundle export — export/otio_export.py::export_otioz(), POST /timeline/export/otioz. Zip container with optional embedded media (bundle_media=True). One-file portable project handoff.

Wave D1 — Compliance

  • Broadcast compliance profiles — extended core/caption_compliance.py with EBU-TT-D (EBU Tech 3380), YouTube broadcast-grade, and accessibility-first (CVAA + WCAG 2.2) profiles in addition to the existing netflix / bbc / fcc / youtube sets. New GET /captions/compliance/standards to let panels enumerate them.

Wave D6 — Verticals

  • Event moment finder — core/event_moments.py, POST /events/moments. Wedding / ceremony / live-event highlight detection via audio-energy RMS spike detection (mean + k·σ threshold) with a min_spacing guard to prevent redundant picks. Optional YAMNet stub for AudioSet tagging when tflite_runtime + a local model path are configured via OPENCUT_YAMNET_MODEL.

Changed

  • Nothing user-visible — all Wave A/D additions live behind check_X_available() guards and degrade gracefully when optional dependencies are absent.

Infrastructure

  • 8 new check_*_available() entries in opencut/checks.py: f5_tts, beatnet, clip_iqa, hsemotion, ab_av1, aaf_adapter, event_moments, plus the pre-existing neural_interp / declarative_compose.
  • 8 new routes added to _ALLOWED_QUEUE_ENDPOINTS so they can be queued via /queue/add.
  • captions.CaptionSegment now carries an optional speaker: str field populated by the WhisperX diarisation pipeline when enabled.

Gotchas

  • F5-TTS model cache stays alive across requests — _MODEL_CACHE["model"] pins the loaded model to free subsequent requests from the ~5 s load penalty. Call tts_f5.clear_model_cache() from a GPU-pressure code path if VRAM becomes tight.
  • WhisperX diarisation auto-fallback — the try/except Exception wrapper around the pyannote branch is intentional. If pyannote 3.x / the HF token / the model download fails for any reason, the caption job completes without speaker labels rather than erroring out. Any diarisation-specific failure mode must be surfaced via logger.warning, never re-raised.
  • BeatNet wav decode — we intentionally decode to a temp 22 050 Hz mono WAV because BeatNet's pretrained models expect that rate. The temp file is cleaned in finally.
  • ab-av1 output parsing — we check for a JSON summary line first (current versions), then fall back to the crf N VMAF M regex for older builds. Some versions emit neither — in that case the result has achieved_vmaf=0, final_crf=0 and only the output path is reliable.
  • OTIOZ bundle_media pulls media into the zip — relies on OTIO's file_bundle_utils.MediaReferencePolicy. The policy enum name differs across OTIO versions (AllMissingReferences on 0.17, ALL_MISSING_REFERENCES on 0.15). Write the adapter import inside the function so import-time failures don't gate module import.
  • Scene-detect auto-dispatcher is stateless — detect_scenes_auto imports TransNetV2/PySceneDetect fresh per call. Cheap in practice (imports are cached by Python) but avoid calling it in a tight inner loop; use detect_scenes_ml directly if you already know TransNetV2 is installed.
  • Event-moments heuristic fallback — when mode="yamnet" is requested but tflite_runtime / TF / a valid model file are unavailable, the function silently downgrades to mode="heuristic" and records that in result.mode. Never raise in that case — YAMNet is a nice-to-have enrichment, not a requirement.

[1.17.0] - 2026-04-17

Added — cross-project research pass

Two feature surfaces borrowed from OSS editors that were missing from the codebase after a survey of LosslessCut, auto-editor, editly, Shotcut/MLT, Olive, OpenShot, Kdenlive, OpenTimelineIO, WhisperX, and PyAV. (Most research recommendations were already implemented: faster-whisper + whisperx, OTIO export, DeepFilterNet, EBU R128 two-pass loudnorm, and GOP-aware smart_render are all present. These two were the remaining concrete gaps.)

  • Neural frame interpolation (core/neural_interp.py, routes/enhancement_routes.py) — RIFE-NCNN-Vulkan CLI backend with torch-RIFE hook and FFmpeg minterpolate fallback. Single interpolate(video_path, target_fps, backend="auto", ...) entry point with auto backend selection, graceful degradation, and a notes field describing fallbacks. Routes: GET /video/interpolate/backends (list installed/available), POST /video/interpolate/neural (async, @async_job("neural_interp")). Doubles via RIFE up to 3 passes (8×), then decimates to the exact target_fps. Audio stream-copied from the source to avoid resampling drift. Source for the external binary: https://github.com/nihui/rife-ncnn-vulkan.
  • Declarative JSON video composition (core/declarative_compose.py, routes/enhancement_routes.py) — editly-inspired (https://github.com/mifi/editly). Single JSON spec → finished video, no timeline reasoning required by callers. Supports video / image / title / color clip types, ~18 xfade transitions (fade, dissolve, wipe*, slide*, circle*, smooth*, pixelize, radial), optional ducked background audio, burnt captions (top/bottom/center) via drawtext, and per-clip in-points. Every intermediate clip is normalised to the target width×height×fps + yuv420p + AAC 48 k stereo so filter_complex concat/xfade works without "resolution mismatch" errors. Routes: GET /compose/schema (enums + example spec), POST /compose/validate (sync validation without rendering), POST /compose/render (async, @async_job("compose", filepath_required=False)). Max 200 clips per spec. ComposeResult dataclass is subscriptable so Flask jsonify accepts it directly.
  • New checks — check_neural_interp_available() (always True — FFmpeg fallback), check_rife_cli_available() (rife-ncnn-vulkan on PATH), check_declarative_compose_available() (FFmpeg on PATH). Added to opencut/checks.py.
  • Queue allowlist — /video/interpolate/neural and /compose/render added to _ALLOWED_QUEUE_ENDPOINTS in jobs_routes.py so they can be queued via /queue/add.
  • Blueprint — enhancement_bp registered in routes/__init__.py. 5 new routes bring the total from 1,152 → 1,177.

Design notes

  • Both modules follow the established dataclass-with-__getitem__ pattern (InterpResult, ComposeResult) so routes can return them directly to Flask's jsonify without a to_dict() conversion — matches the MEMixResult / PremixResult convention from v1.16.2.
  • Neural interpolation refuses to run when target_fps <= source fps and redirects the caller to the existing /video/framerate-convert route for downsampling.
  • declarative_compose._escape_text scrubs drawtext meta-characters (:, ;, \, ') and control bytes, and caps text at 500 chars — same pattern as the title/overlay route hardening in v1.3.1 batch 10.

[1.16.3] - 2026-04-16

Fixed

  • Demucs separation could hang the worker thread forever — routes/audio.py::audio_separate consumed process.stdout line-by-line then called process.wait() with no timeout. A wedged Demucs (stdout closed but child not reaping) blocked the whole job slot indefinitely. Bounded the final wait to 1800 s with a force-kill + RuntimeError on overrun so the slot is released and the user gets an actionable error instead of a permanent "running" job.
  • UXP _doFetch had no request timeout — every backend call (get / post / del) used fetch(url, opts) with no AbortController. A hung backend pinned the calling button into "Working…" forever. Added a 120 s AbortController.signal on every request, cleared in finally. Also wrapped resp.json() in try/catch so a server that lies about Content-Type: application/json and returns garbage doesn't blow past the status-code check.
  • ocAddNativeCaptionTrack produced malformed SRT for multi-line captions — caption text was written verbatim with tempFile.writeln(seg.text || ""). SRT uses a blank line as the cue terminator, so any caption containing \n\n (very common in Whisper output for sentence breaks) split a single cue into two malformed cues. Premiere either dropped the half-cues or imported them with mis-aligned timecodes. Now collapses \r\n/\r to \n, replaces consecutive newlines with a single \n, trims trailing whitespace, and substitutes a single space when the body is empty (empty cue body is also invalid SRT).

Hardened — encoding consistency

  • open() text-mode without encoding= — added encoding="utf-8" to JSON metadata reads/writes in core/model_manager.py, core/character_consistency.py, core/motion_transfer.py, core/proxy_gen.py, and core/proxy_swap.py. Without an explicit encoding, the Python default on Windows is the system codepage (cp1252), which raises UnicodeDecodeError whenever a profile/proxy mapping mentions a path or speaker name with non-ASCII characters. Concat-list files were intentionally left alone — FFmpeg reads them in the system locale, so forcing UTF-8 there would corrupt non-ASCII paths the writer thinks were saved correctly.

Audit notes (no fix needed)

  • subprocess.run / _sp.run calls — verified all 100+ call sites across the codebase carry an explicit timeout=. Multi-line calls had previously evaded a single-line grep; a syntax-aware sweep confirmed full coverage.
  • urlopen() — verified all call sites carry a timeout= argument.
  • process.wait() on Popen — only one site (routes/audio.py Demucs) lacked a timeout; fixed above.
  • WorkerPool shutdown / cancel paths — all already use set_running_or_notify_cancel() and BaseException to guarantee futures resolve even on SystemExit/KeyboardInterrupt.
  • errors.safe_error() exception classification — verified the order is correct (more specific phrases checked before broader ones); GPU OOM is matched before the generic FFmpeg branch so an "FFmpeg out of memory" RuntimeError reports GPU_OUT_OF_MEMORY (the more actionable code).
  • job_store.save_job ON CONFLICT … DO UPDATE — result_json / error / progress use excluded.X (not COALESCE), so a status-only update wipes the prior value. This is intentional: callers always pass the full job dict from the in-memory jobs registry, so excluded is never partial.

[1.16.2] - 2026-04-16

Fixed — broken production routes (entire endpoints were 500ing on every call)

  • /api/audio/stem-remix was non-functional — routes/audio_expansion_routes.py::stem_remix imports mix_stems and calls remix_stems(stem_paths=LIST, effects_config=, output=, on_progress=) and reads result.stems_processed / result.effects_applied. None of those names existed in core/stem_remix.py. Added the full effect-config API surface to the module: SUPPORTED_EFFECTS dict (10 named effects), StemEffect dataclass, _effect_to_filter(), _build_pan_filter(), apply_stem_effect(), mix_stems(). Made remix_stems() polymorphic so it routes a list-based stem_paths + effects_config call to a new _remix_stems_with_effects() helper while preserving the original dict-based preset path. Added stems_processed / effects_applied fields to RemixResult.
  • /video/data-animation/* was non-functional — routes/color_mam_routes.py calls create_data_animation, render_bar_chart, render_counter from core/data_animation.py; module only had the lower-level render_data_animation. Added a high-level chart API: DataTemplate, DataAnimationResult dataclasses, _load_data_source(), _extract_labels_values(), render_bar_chart(), render_counter(), create_data_animation(). Hoisted run_ffmpeg/get_ffmpeg_path to module-level imports so route/test patching actually intercepts them.
  • /video/shape-animation/* was non-functional — same pattern as data-animation: routes call animate_shape_morph, animate_stroke_draw, animate_fill_transition from core/shape_animation.py; module only had render_shape_animation. Added ShapeDefinition, ShapeAnimationResult dataclasses, _parse_shape(), the three animate-* entry points, plus a _normalise_color() helper that converts 0xFF0000 style hex into the #FF0000 shape FFmpeg accepts. Module-level run_ffmpeg import for testability.
  • /audio/me-mix returned a non-JSON-serialisable dataclass — routes/audio_production_routes.py::route_me_mix does return result straight into Flask's auto-jsonify, but core/me_mix.py::generate_me_mix returned the legacy MEMixResult dataclass which isn't JSON-friendly. Also the legacy _level_match() post-step crashed on every "subtract" call because of the os.replace(tmp_out, filepath) race when the input doesn't actually exist. Rewrote generate_me_mix(): default method="subtract" (was "auto"), added a fast _me_subtract() path that does a single FFmpeg pan=stereo|c0=c0-c1|c1=c1-c0 + loudnorm invocation and returns a hybrid MEMixResult (now subscriptable so result["method_used"] and result.method_used both work). Legacy paths (stem_separation, track_mute, spectral) still route to the original implementation.
  • /audio/dialogue-premix returned dataclass to Flask jsonify — core/dialogue_premix.py::premix_dialogue had target_lufs default of None (used the preset's broadcast value, not what the route passed), and the result was a PremixResult dataclass. Added premix_multi_speaker() as a public alias of the internal helper, defaulted target_lufs=-23.0, and made PremixResult subscriptable so dict and attribute access both work. Wrapped the progress callback so internal 1-arg calls and external 2-arg (pct, msg) calls compose without a TypeError.
  • CLI opencut nlp was completely broken — imported from non-existent opencut.core.nlp (real module is nlp_command), called parse_nlp_command (real name is parse_command), and accessed parsed.route / parsed.params / parsed.execute(file) on a dict that has no such attributes/methods. Rewrote the command to use the correct import + parse_command() dict result, and replaced the hallucinated .execute() call with an actual urllib POST to the matched route (with CSRF refresh). On a missing backend the command now exits non-zero with a helpful "start the server first" hint.
  • CLI chapters / repeat-detect / search index all imported a non-existent transcribe_audio — all three commands do from .core.captions import transcribe_audio, but the captions module only exposes transcribe(filepath, config=...). Added transcribe_audio(filepath, model, language, timeout) to core/captions.py as a thin wrapper that builds a CaptionConfig and projects the TranscriptionResult down to a flat list of segment dicts so callers don't need the full result object.

Added — in-memory subtitle/multilang API

  • MultilangProject, create_multilang_project(), update_language_text(), export_language(), sync_timing_change() in core/multilang_subtitle.py. The persistence-based create_project / add_language / export_srt API is preserved; the new functions are a smaller in-memory surface used by the v1.15.0 editing-workflows tests and by callers that don't need disk-backed project IDs. sync_timing_change supports shift / update / insert / delete operations and mirrors all timing edits across every language array so they never drift out of sync.

Fixed — robustness

  • room_tone.analyze_room_tone() propagated librosa load failures — when librosa was installed, a missing/corrupt input file raised FileNotFoundError from librosa.load() instead of degrading to the FFmpeg-based fallback that the function was specifically designed to provide. Wrapped the whole librosa block in a broad except Exception (still re-raised for ImportError short-circuit) and logged at info level, so callers always get either a librosa-grade or an FFmpeg-grade envelope.
  • CEP recent-files store crashed if localStorage was tampered — getRecentFiles() returned whatever JSON.parse produced, so a corrupted opencut_recent_files value (null, "foo", an object) made populateRecentFiles and addRecentFile crash on .length / .filter / .unshift. Now defensively coerces to a clean Array of {path, name} entries, filtering out anything missing a string path. addRecentFile reuses the same parser instead of re-parsing raw localStorage.

Result counts

  • 7,689 tests passing, 0 failing, 3 skipped (was 7,629 / 60 / 3 before this batch). Net +60 tests fixed across 5 production-route bug clusters and 1 environment-fragility test.

[1.16.1] - 2026-04-16

Fixed

  • outpaint_aspect_ratio silently ran on missing files — core/frame_extension.py::outpaint_aspect_ratio() (and the related extend_frame_spatial/extend_frame_temporal helpers) called get_video_info() first, which logs a warning and returns 1920x1080 defaults on a missing file, then walked the rest of the pipeline producing garbage output paths. Added an explicit os.path.isfile() check that raises FileNotFoundError upfront for all three entry points.
  • Command palette recents sorted by insertion order, not recency — core/command_palette.py::record_feature_use() stamped each entry with time.time(), which on Windows advances in ~15.6 ms ticks. Two consecutive record_feature_use() calls received identical timestamps and the stable sort kept the earlier entry on top, so the most-recently-used feature was hidden behind older ones. Now bumps the timestamp to max(existing) + 1us whenever the wall clock has not advanced past the previous entry — guarantees monotonic ordering without breaking the JSON schema.
  • Step duration always reported 0.0 in autonomous agent — core/autonomous_agent.py::execute_step() measured elapsed time with time.time(), which has ~15.6 ms resolution on Windows. Mocked or fast-completing steps recorded duration_seconds=0.0, breaking downstream timing analytics. Switched to time.perf_counter() with a 1e-6 floor. Same fix applied to execute_plan().
  • Model quantization elapsed always 0.0 for fast paths — core/model_quantization.py::quantize_model() used time.time() then round(elapsed, 2), so any sub-10 ms run rounded to 0.0. Switched to perf_counter, raised reporting precision to 3 decimals, and applied a 1 ms floor so callers can distinguish "ran" from "no-op".
  • get_video_info() raised on ffprobe timeout/missing — helpers.py::get_video_info() called subprocess.run(..., timeout=30) without catching TimeoutExpired, FileNotFoundError, or OSError. A hung or missing ffprobe surfaced as a 500 instead of degrading to safe defaults like every other failure path. Now catches the subprocess exceptions and logs a warning before returning defaults.

[1.16.0] - 2026-04-15

Security & Hardening

  • Path traversal prevention — validate_output_path() added to all 157+ output_path and 100+ output_dir parameters across 43+ route files.
  • FFmpeg concat demux injection — Newline/carriage return stripping in filenames for video_core.py concat operations.
  • SSRF prevention — LLM base_url validation in captions routes blocks non-HTTP schemes.
  • Information disclosure — Error responses no longer leak raw exception messages or detail fields.
  • Docker hardening — Non-root user added to container image.
  • Windows reserved names — validate_path() now blocks CON, PRN, AUX, NUL, etc.
  • Atomic file writes — Plugin marketplace manifest and scripting console history use tempfile + os.replace.
  • Sandbox escape prevention — Scripting console blocked __import__, exec, eval, compile, open, os, sys, subprocess in AST.
  • .gitignore hardened — Added .env.*, *.key, *.pem, credentials*.json.

[1.15.0] - 2026-04-14

Added

  • 5 new feature categories (78-82) — 25 new core modules with routes and comprehensive tests.
  • AI Voice & Speech (Cat. 78): Descript-style transcript-based video editing with undo stack and EDL export, AI eye contact correction via MediaPipe gaze estimation, voice overdub (replace spoken words with TTS-cloned corrections), audio-driven lip sync with Wav2Lip fallback, voice-to-voice conversion with target profiles.
  • Motion Design & Animation (Cat. 79): Kinetic typography engine (12 presets, 7 easing functions), data-driven animation from CSV/JSON (6 chart types), vector shape animation with SVG path morphing, sandboxed expression scripting engine (30+ math functions), configurable particle system (8 presets, 5 emitters, physics simulation).
  • Professional Subtitling (Cat. 80): Shot-change-aware subtitle timing (Netflix/BBC/FCC profiles), multi-language simultaneous editing with shared timing, broadcast caption export (CEA-608/708, EBU-TT, TTML, IMSC1, WebVTT), SDH/HoH auto-formatting (speaker IDs, sound events, music notation), dynamic subtitle positioning to avoid visual obstructions.
  • Developer & Scripting Platform (Cat. 81): Sandboxed Python scripting console with curated OpenCut namespace, macro recording and playback with variable substitution, FFmpeg filter chain builder (20 filter types, cycle detection), event-driven webhook system (5 event types, retry with backoff), batch scripting engine with glob patterns and dry-run.
  • Audio Post-Production (Cat. 82): ADR cueing system with cue sheet export (CSV/JSON) and guide audio, M&E mix export (3 methods: stem separation, track mute, spectral), automated dialogue premix (per-speaker EQ/compression/de-ess, 5 presets), stereo-to-surround upmix (5.1/7.1, 4 modes, VBAP panning), foley cueing with SFX auto-placement (8 categories).
  • 5 new route blueprints — voice_speech_bp, motion_design_bp, subtitle_pro_bp, dev_scripting_bp, audio_post_bp (64 new routes).
  • 5 new test files — 626 new tests covering all v1.15.0 features.
  • 1,152 total API routes (up from 1,088 in v1.14.0).
  • 7,551 total tests across 92 test files (up from 6,925 across 87).
  • 424 core modules (up from 408), 88 route blueprints (up from 83).

[1.14.0] - 2026-04-14

Added

  • 5 new feature categories (73-77) — 25 new core modules with routes and comprehensive tests.
  • AI Collaboration & Review (Cat. 73): Frame-accurate review comments with annotations, frame-by-frame version comparison (SSIM/PSNR), multi-stage approval workflows, immutable edit history audit log, team-shared preset library with merge resolution.
  • Advanced Timeline Automation (Cat. 74): AI rough cut assembly from script/brief (LLM + keyword fallback), multi-track audio auto-mix with ducking profiles, smart trim point detection (3 modes), batch timeline operations (6 op types with chaining), template-based video assembly (4 built-in templates).
  • AI Sound Design & Music (Cat. 75): AI sound design from video analysis (12 SFX categories, PCM synthesis), procedural ambient soundscape generator (7 presets), music mood morphing (6 transforms, keyframeable curves), beat-synced auto-edit (6 cut modes, energy matching), creative stem remix (7 presets including lo-fi/nightcore/slowed reverb).
  • Real-Time AI Preview (Cat. 76): Live AI effect preview at 480p (10 effects, LRU cache), GPU-accelerated preview pipeline with CPU fallback, A/B comparison generator (6 modes, quality metrics), real-time video scopes (waveform/vectorscope/histogram/parade/false_color), disk-persistent preview render cache with TTL.
  • Cloud & Distribution (Cat. 77): Cloud/remote render dispatch with health checks and failover, multi-platform auto-publish (10 platforms), content fingerprinting and duplicate detection (pHash + audio), render farm management with segment dispatch, post-publish distribution analytics.
  • 5 new route blueprints — collab_review_bp, timeline_auto_bp, sound_music_bp, preview_realtime_bp, cloud_distrib_bp (57 new routes).
  • 5 new test files — 647 new tests covering all v1.14.0 features.
  • 1,088 total API routes (up from 1,031 in v1.13.0).
  • 6,925 total tests across 87 test files (up from 6,278 across 82).
  • 408 core modules (up from 385), 83 route blueprints (up from 78).

[1.13.0] - 2026-04-14

Added

  • 5 new feature categories (68-72) — 25 new core modules with routes and comprehensive tests.
  • AI Timeline Intelligence (Cat. 68): Holistic timeline quality analysis, per-segment engagement scoring, AI narrative clip chaining, natural language color grading (32 named looks), end-to-end auto-dubbing pipeline (47 languages).
  • Advanced Object AI (Cat. 69): Text-driven video segmentation ("remove the red car"), physics-aware object removal (shadows + reflections), object tracking with graphic overlays (10 overlay types), CLIP-based semantic video search, multi-subject intelligent reframe.
  • Delivery & Mastering (Cat. 70): DCP (Digital Cinema Package) export, IMF (Interoperable Master Format) packaging, automated delivery validation (Netflix/YouTube/Broadcast/DCP/Apple TV+/Amazon specs), multi-format simultaneous render with priority, delivery spec profile manager.
  • AI Content Generation (Cat. 71): Voice-to-avatar video generation (5 styles), ML thumbnail CTR prediction, AI B-roll generation from text prompts, auto chapter artwork/title cards (5 card styles), animated video intro generation (5 intro styles).
  • Pipeline Intelligence (Cat. 72): Pipeline health monitoring dashboard, cron-like scheduled/recurring jobs, smart content routing (10 content types), accurate processing time estimation, real-time CPU/GPU/memory resource monitoring.
  • 5 new route blueprints — timeline_intel_bp, object_intel_bp, delivery_master_bp, content_gen_bp, pipeline_intel_bp (51 new routes).
  • 5 new test files — 596 new tests covering all v1.13.0 features.
  • 1,031 total API routes (up from 980 in v1.12.0).
  • 6,278 total tests across 82 test files (up from 5,742 across 77).
  • 385 core modules (up from 360), 78 route blueprints (up from 73).

[1.12.0] - 2026-04-14

Added

  • Competitive audit features (N1-N20) — 25 new core modules implementing all features identified in AUDIT.md competitive analysis.
  • UX Intelligence: Command palette with fuzzy search (N1), AI contextual suggestions (UX2), smart defaults engine (UX11), unified preview system (UX5).
  • Enhanced Media: AI speech restoration (N2), one-click enhance pipeline (N9), low-light video enhancement (N13), scene edit detection (N14).
  • Engagement & Content: Engagement/retention prediction (N4), 50+ animated caption styles (N5), AI hook generator (N6), A/B variant generator (N7), Essential Graphics caption output (N11).
  • Next-Gen AI: Video LLM integration (N8), AI music remix/duration fit (N10), audio category tagging (N15), AI color match between shots (N20).
  • Motion & Generation: Generative extend (N3), green-screen-free background replacement (N12), consistent character generation (N16), motion brush/cinemagraph+ (N17).
  • Body & Transfer: Body/pose-driven effects (N18), full-body motion transfer (N19), AI foley sound generation, AI face restoration.
  • 6 new route blueprints — enhanced_media_bp, ux_intel_bp, engagement_content_bp, next_gen_ai_bp, motion_gen_bp, body_transfer_bp (48 new routes).
  • 6 new test files — 615 new tests covering all audit features.
  • 980 total API routes (up from 932 in v1.11.0).
  • 5,742 total tests across 77 test files (up from 5,127 across 71).
  • 360 core modules (up from 335), 73 route blueprints (up from 67).

[1.11.0] - 2026-04-13

Added

  • All 302 features fully implemented — Every feature in features.md now has a working core module, route blueprint, and test coverage.
  • 12 new feature categories (Batch 5) — VR/360 video, camera/lens correction, video repair & generation, privacy/redaction, spectral audio editing, split-screen & multicam, content repurposing, storyboard/pre-production, proxy management, composition intelligence, AI dubbing & localization.
  • 52 new core modules — vr_stabilize, spatial_audio_vr, chromatic_aberration, lens_profile, old_restoration, sdr_to_hdr, framerate_convert, img_to_video, scene_extend, video_condensed, bg_replace_ai, plate_blur, pii_redact, profanity_bleep, doc_redact, audio_anon, spectrogram_edit, spectral_repair, noise_classify, split_screen, reaction_template, multicam_grid, long_to_shorts, video_to_blog, podcast_bundle, content_calendar, shot_list_gen, mood_board, script_to_roughcut, proxy_swap, media_relink, composition_guide, saliency_crop, ai_dubbing, isochronous_translate, multilang_audio, emotion_voice, and enhancements to 15 existing modules.
  • 6 new route blueprints — vr_lens_bp, repair_gen_bp, privacy_spectral_bp, multiview_repurpose_bp, preproduction_proxy_bp, composition_dubbing_bp (52 new routes).
  • 6 new test files — 562 new tests covering all Batch 5 features.
  • 932 total API routes (up from 880 in v1.10.5).
  • 5,127 total tests across 71 test files (up from 867 across 65).
  • 335 core modules (up from 298), 67 route blueprints (up from 61).

Fixed

  • Duplicate blueprint registration in server.py that caused ValueError on app startup.

[1.10.5] - 2026-04-13

Added

  • 302-feature expansion plan (features.md) — 62 categories covering every major video editing domain. 12 new categories added: 360/VR, camera correction, video repair, AI generation, privacy/redaction, spectral audio, split-screen, content repurposing, storyboarding, proxy management, composition intelligence, AI dubbing.
  • ROADMAP v3.0 — 7-wave implementation plan organized by dependency chains (Wave 1: quick wins with 0 new deps through Wave 7: emerging tech). Replaces completed Phases 0-6.
  • 230 new core modules (opencut/core/) — implementations for feature categories including deinterlace, lens correction, room tone, credits generator, retro effects, hardware acceleration, GIF export, ProRes/AV1/DNxHR encoding, color scopes, spectral audio, video repair, proxy generation, split-screen templates, AI dubbing, storyboard generation, pacing analysis, shot classification, and 200+ more.
  • 43 new route blueprints (opencut/routes/) — 600+ new API endpoints covering audio production, video processing, encoding, creative effects, QC checks, gaming, education, documentary, integrations, generative AI, professional workflows, and more.
  • 43 new test files (tests/) — smoke tests for all new route blueprints.
  • 880 total routes (up from 254 in v1.9.26). All registered via __init__.py.

Fixed

  • Ruff lint — 1,207 auto-fixed + 45 manual fixes (F401 noqa for lazy imports, F821 forward refs, E741 ambiguous vars, E701 multi-statement lines) across all new files. Codebase fully lint-clean.

[1.10.4] - 2026-04-12

Added

  • Expandable "Why this suggestion?" on Sequence Assistant recommendation cards. Click to reveal the reasoning behind each suggestion with context-specific explanation.

[1.10.3] - 2026-04-12

Added

  • Journal "Apply to selection" — Apply a journal entry's operation to the currently selected clip with one click from the Operation Journal panel.

[1.10.2] - 2026-04-12

Added

  • Persisted Assistant dismissals — Dismissed Sequence Assistant suggestions are remembered across sessions via localStorage. Prevents the same suggestion from reappearing after panel reload.

[1.10.1] - 2026-04-12

Added

  • Preflight checks for /full and /shorts-pipeline — Validate dependencies, disk space, and clip metadata before starting long multi-step pipelines. Prevents wasted time on doomed jobs.
  • More audio preview buttons — Added waveform preview buttons on additional audio processing sub-tabs.

[1.10.0] - 2026-04-12

Added

  • Sequence Assistant — "What should I edit next?" AI-powered recommendation engine. Analyzes the current clip and project context to suggest relevant operations (denoise, captions, color correction, etc.) with one-click execution. Displays as a dismissible card panel.

[1.9.36] - 2026-04-11

Added

  • Live audio preview for denoise — Preview denoised audio in-panel before applying. Streams a short processed sample for A/B comparison.

[1.9.35] - 2026-04-11

Added

  • Visual LUT grid — Browse available LUTs as a visual grid with thumbnail previews showing each LUT applied to a reference frame. Click to apply.

[1.9.34] - 2026-04-11

Added

  • Transcript-to-timeline scrubber — Click any word in the transcript view to seek to that position in the Premiere timeline. Bidirectional sync between text and playhead.

[1.9.33] - 2026-04-11

Added

  • Preflight checklist — Before starting long pipelines (silence + captions + export), display a checklist of required dependencies, available disk space, and estimated processing time. User confirms or cancels.

[1.9.32] - 2026-04-11

Added

  • Batch mode for Interview Polish — Run the Interview Polish workflow on multiple clips sequentially with a single click. Progress tracking per clip.

[1.9.31] - 2026-04-11

Added

  • CLI opencut polish — Headless interview/podcast cleanup pipeline. Chains denoise + normalize + silence removal + filler removal from the command line.

[1.9.30] - 2026-04-11

Added

  • Replay past job on current clip — From the job history panel, replay any previous operation on the currently selected clip with the same parameters. One-click re-run.

[1.9.29] - 2026-04-10

Added

  • Interview Polish — One-click interview/podcast cleanup pipeline. Chains: denoise + normalize + silence removal + filler removal with optimized defaults. Quick Action button on Cut tab.

[1.9.28] - 2026-04-10

Added

  • Operation Journal — Persistent log of all operations performed on each clip with parameters, timestamps, and results. One-click rollback to undo the last operation.

[1.9.27] - 2026-04-10

Added

  • Session Context — "Where you left off" welcome-back card on panel open. Shows last clip worked on, last operation performed, and quick resume button.

[1.9.26] - 2026-04-09

Fixed

  • Install route rate limit timing — make_install_route() checked rate_limit("model_install") inside the @async_job body (async thread), so concurrent installs returned 200 + job_id then failed asynchronously. Moved check to synchronous Flask handler — now returns 429 RATE_LIMITED immediately with no job spawned.
  • @async_job filepath errors lacked structured codes — Filepath validation failures returned bare {"error": "..."} without code or suggestion fields. Now classifies into INVALID_INPUT vs FILE_NOT_FOUND with actionable suggestions. Applies to all 97 async routes.

[1.9.25] - 2026-04-09

Fixed

  • /system/dependencies cold call 6.5s — 20+ importlib.import_module() probes + subprocess calls ran synchronously. Added 60-second TTL cache. Cold: 6,463ms. Warm: <1ms (~6,000x speedup). ?fresh=1 param bypasses cache.

[1.9.24] - 2026-04-09

Fixed

  • Env var config ignored — OPENCUT_PORT, OPENCUT_HOST, OPENCUT_DEBUG env vars were silently ignored. server.py::main() only read argparse with hardcoded defaults. Now reads env vars first, argparse overrides.

[1.9.23] - 2026-04-09

Fixed

  • UXP accessibility — All 7 tab panels now have aria-labelledby. All 72 buttons have type="button". .hidden utility class unified as global display: none !important.

[1.9.22] - 2026-04-08

Fixed

  • /file preview route allowlist restored — Regression had removed the security allowlist.
  • safe_bool() hardened — Rejects NaN, inf, containers. 15 new call sites across route files.
  • MCP serialization fallback — Added JSON fallback for non-serializable job results.
  • highlights.py None-text guard — Prevents crash on segments with null text.

[1.9.21] - 2026-04-08

Added

  • CEP panel UX polish — Friendlier error messages, clearer button labels, improved alert formatting.

Fixed

  • Boolean coercion hardening — safe_bool() added to all 26 boolean flag sites across 8 route files. Prevents "false" string coercing to True.

[1.9.20] - 2026-04-07

Fixed

  • Gemini API key URL leak — core/llm.py passed API key as query string, leaking through error messages. Moved to x-goog-api-key header. Added _sanitize_url().
  • motion_graphics drawtext colon injection — 3 drawtext builders escaped : inside single-quoted values, producing literal \: in rendered text. Removed wrong escapes.
  • Queue allowlist +19 routes — 19 async routes missing from _ALLOWED_QUEUE_ENDPOINTS (total now 101).
  • CEP SSE/poll cancel race — Stale events after cancel could trigger onJobDone(). Added guard.
  • audio_enhance no-op .cpu() calls — Removed ineffective .cpu() calls before del.
  • animated_captions temp file leak — Unlinked orphan tempfile on VideoWriter failure.

[1.9.19] - 2026-04-06

Changed

  • Comprehensive audit + full modernization pass.

[1.9.18] - 2026-04-05

Fixed (Comprehensive 6-Phase Audit — 103+ issues)

  • P0: loudness_match completely broken — _run_ffmpeg() returned str but callers accessed .returncode/.stderr. Renamed to _run_ffmpeg_raw().
  • P0: UXP CSRF token refresh wrong header — Read X-CSRF-Token instead of X-OpenCut-Token.
  • P0: UXP chat endpoint 404 — Posted to /chat/message but route is /chat.
  • P1: WorkerPool shutdown TypeError — None poison pills not comparable with PriorityQueue items.
  • P1: styled captions progress crash — Missing msg="" default.
  • P1: scene_detect + shorts_pipeline bare ffmpeg — 4 bare "ffmpeg" strings.
  • P1: 3 GPU memory leaks — MusicGen (1.2-13GB), SAM2 (1-4GB), Florence-2 (450MB) models never freed.
  • P2: 37 additional fixes — Route allowlists, float coercion, structured errors, resource cleanup, font thread safety, CSS missing rules.

[1.9.17] - 2026-04-03

Changed

  • README update with comprehensive feature documentation and audit prompt.

[1.9.16] - 2026-04-02

Fixed (Performance Audit & Memory Leak Fixes)

  • /video/auto-zoom broken return — Route returned {} instead of result_dict.
  • Unbounded thread spawning — _persist_job() and _schedule_record_time() spawned raw threads per call. Replaced with bounded _io_pool = ThreadPoolExecutor(2).
  • Rate limit slot leak — make_install_route() leaked permanently when rate_limit() returned False.
  • GPU memory leak in audio_enhance — Now deletes all tensor refs + .cpu() + torch.cuda.empty_cache().
  • Timer thread spam — _schedule_temp_cleanup() replaced with single _cleanup_worker daemon thread + event queue.
  • Haar cascade race condition — Added _CASCADE_LOCK with double-checked locking.
  • SQLite connection leak — Added close_all_connections() shutdown hook.
  • CEP health timer leak — setInterval reassigned without clearInterval.
  • Single theme ("Studio Graphite") — Removed multi-theme system (10 themes). Only root dark theme remains.

[1.9.15] - 2026-03-27

Release Audit (Batch 47)

Full 3-agent audit (routes, core modules, frontend+infra). Results:

  • Routes: CLEAN. All 17 route files pass security, consistency, and contract checks.
  • Core modules: 1 real bare ffmpeg fixed (audio_enhance.py audio extraction). All other flagged items were either already using run_ffmpeg() (which auto-resolves) or false positives.
  • Frontend/infra: No critical issues. All XSS, thread-safety, and MCP claims verified as false positives.

Fixed

  • audio_enhance.py bare ffmpeg — _extract_audio() used "ffmpeg" directly in subprocess instead of get_ffmpeg_path(). Crashes on bundled installs.
  • 4 test failures — test_score_frame_colorful_image threshold too high (20→5), test_registry_has_builtin_engines expected wrong dict key, test_depth_map_rate_limited + test_emotion_rate_limited couldn't reach rate limit due to filepath validation happening first.

Verified

  • 0 lint warnings
  • All 18 version targets in sync
  • 254 routes, 23 MCP tools, 23 schemas, 16 CLI commands
  • All subprocess calls use get_ffmpeg_path()/get_ffprobe_path() or go through run_ffmpeg()
  • All open() calls have encoding="utf-8" for text mode
  • All on_progress closures have msg="" default

[1.9.14] - 2026-03-27

Fixed (Batch 46 — Final Sweep)

  • 4 more bare "ffmpeg" missed by agents — audio.py (3 in stem separation format conversion + audio extraction) and video_core.py (1 in stabilize audio merge). All replaced with get_ffmpeg_path().
  • MCP server stale docstring — Claimed --sse mode existed but was never implemented. Fixed docstring to accurately describe stdio-only transport.
  • MCP resources/list and prompts/list — Added empty handlers required by MCP spec. Modern MCP clients (Claude Code) expect these methods to exist.

[1.9.13] - 2026-03-27

Fixed (Batch 45 — Complete ffmpeg/ffprobe Path Resolution)

  • 23 bare "ffmpeg" in subprocess calls — 13 in route files (audio.py, system.py, timeline.py, video_core.py x7, video_editing.py x3) and 10 in core modules (audio.py x2, audio_suite.py x2, highlights.py, loudness_match.py x4, scene_detect.py). All replaced with get_ffmpeg_path(). On bundled installs where FFmpeg is not in system PATH, every one of these would crash with "command not found".

[1.9.12] - 2026-03-27

Fixed (Batch 44 — Dependencies & Hardcoded Ports)

  • psutil added to core dependencies — /system/status endpoint used psutil but it wasn't in dependencies, causing ImportError on fresh installs.
  • numpy upper bound removed — <2 cap prevented numpy 2.x which is already used in the dist build. Now >=1.24 with no upper bound.
  • 3 new optional dependency groups — otio (opentimelineio), tts (edge-tts + kokoro), depth (torch + transformers). All were installable but not declared in pyproject.toml.
  • opentimelineio added to [all] extras — OTIO export was available in code but not installed via pip install opencut[all].
  • social_post.py hardcoded port — get_oauth_url() hardcoded localhost:5679 in all 3 OAuth redirect URIs. Now accepts port parameter (default 5679) for correct redirect on non-default ports.

[1.9.11] - 2026-03-27

Fixed (Batch 43 — Encoding, Docker, CI)

  • 13 open() calls missing encoding="utf-8" — 6 core modules (captions, footage_search, lut_library, motion_graphics, social_post, speed_ramp) and 1 route (video_core) used system-locale encoding on Windows, causing UnicodeDecodeError on non-ASCII file content.
  • Docker GPU variant — docker-compose.yml GPU service now usable via docker compose --profile gpu up instead of requiring manual uncommenting. Added WebSocket port 5680 to EXPOSE. Added psutil and python-json-logger to Dockerfile deps. Plugin dir created at build time.
  • CI smoke tests expanded — Added imports for create_app (factory pattern), WorkerPool/JobPriority, MCP_TOOLS (count check), cli.commands (count check), WorkflowResult schema.

[1.9.10] - 2026-03-27

Added

  • CLI: denoise command — Remove background noise from audio/video via FFmpeg filters (afftdn, highpass, gate) with adjustable strength.
  • CLI: scene-detect command — Detect scene boundaries with method selection (ffmpeg/ml/pyscenedetect), threshold control, and optional JSON export. Rich table output.

Fixed (Batch 42 — CLI Crash Bugs)

  • CLI color-match crash — Imported non-existent from .core.color import match_color. Fixed to from .core.color_match import color_match_video.
  • CLI loudness-match crash — Imported non-existent from .core.loudness import normalize_loudness. Fixed to from .core.loudness_match import batch_loudness_match with correct call signature.
  • CLI auto-zoom crash — Imported non-existent from .core.zoom import apply_auto_zoom. Fixed to from .core.auto_zoom import generate_zoom_keyframes with FFmpeg zoompan apply step.
  • CLI deliverables crash — Imported non-existent generate_deliverables() function. Fixed to import individual generators (generate_vfx_sheet, generate_adr_list, etc.) and dispatch based on --type flag. Now properly loads sequence JSON and routes to correct generator.

[1.9.9] - 2026-03-27

Added

  • UXP: 4 new Video features — AI Upscale (Real-ESRGAN 2x/3x/4x), Scene Detection (FFmpeg/TransNetV2/PySceneDetect), Style Transfer (8 preset styles with intensity slider), Shorts Pipeline (one-click short-form clip generation with face tracking + captions).
  • UXP: full JS handlers — runUpscaleUxp(), runSceneDetectUxp(), runStyleTransferUxp(), runShortsPipelineUxp() with proper job polling, error handling, and toast notifications.

Fixed

  • UXP stale version — Settings tab showed hardcoded "1.9.2" instead of actual version. Now dynamic via version sync.

[1.9.8] - 2026-03-27

Added

  • 9 new response schemas — WorkflowResult, ContextAnalysisResult, VideoAIResult, ShortsPipelineResult, DepthMapResult, BrollPlanResult, BatchResult, PluginListResult (22 total schemas, up from 13). All mapped into OpenAPI spec.
  • OpenAPI endpoint coverage +15 — 15 more async routes registered in _JOB_ENDPOINTS and _ENDPOINT_SCHEMAS for typed OpenAPI responses.

Fixed (Batch 40)

  • 3 more bare "ffprobe" in routes/utils — audio.py waveform probe, video_editing.py reframe probe, utils/media.py media probe.
  • 2 broken tests fixed — test_info_with_csrf expected 200 from /info with no filepath (returns 400), test_file_no_path expected 400 but route returns 404.
  • pytest hanging on exit — WorkerPool daemon threads prevented clean exit. Added session-scoped _shutdown_worker_pool fixture in conftest.py.

[1.9.7] - 2026-03-27

Added

  • 5 new MCP tools — opencut_denoise_audio, opencut_upscale, opencut_scene_detect, opencut_depth_map, opencut_shorts_pipeline (23 total, up from 18). AI clients can now drive denoising, upscaling, scene detection, depth mapping, and the full shorts pipeline.
  • WorkerPool priority queue — Replaced FIFO ThreadPoolExecutor with custom PriorityQueue-backed pool. JobPriority.CRITICAL (0) jobs now leapfrog BACKGROUND (200) work when all workers are busy. Equal-priority jobs still run in FIFO order.

Fixed (Batch 39)

  • 3 more bare "ffprobe" in routes/utils — audio.py waveform probe, video_editing.py reframe probe, utils/media.py media probe — all crash on bundled installs where ffprobe isn't in PATH.
  • FFmpeg/ffprobe path warning — get_ffmpeg_path() and get_ffprobe_path() now log WARNING when binary not found in PATH (was silently falling back to bare name, causing confusing subprocess errors later).

[1.9.6] - 2026-03-27

Fixed (Batch 38 — Route Cleanup & Remaining Crash Bugs)

  • 10 remaining _p(pct, msg) closures missing msg="" default — Found in audio.py (3), video_fx.py (5), captions.py (1), workflow.py (1). All core modules call on_progress(pct) with 1 arg, causing TypeError crash on every job using these routes. All now have msg="" default.
  • Redundant filepath re-validation in 7 routes — face_enhance, face_swap, upscale_run, remove_watermark, title_overlay, particle_apply, color_correct, color_convert, color_external_lut all re-read data.get("filepath") and re-validated despite @async_job already doing this. Removed ~80 lines of dead validation code, now using the decorator-provided filepath param directly.
  • Unused validate_filepath import — Removed from video_specialty.py after watermark/title routes stopped re-validating.

[1.9.5] - 2026-03-27

Fixed (Batch 37 — Infrastructure & Hardening)

  • Bare ffprobe in 8 core modules — audio_enhance, color_management, highlights, motion_graphics, scene_detect (2 sites), shorts_pipeline, transitions_3d, and video_ai all used "ffprobe" directly in subprocess commands. On systems where ffprobe is bundled (not in PATH), all of these would crash. Fixed to use get_ffprobe_path().
  • GPU VRAM check can hang — check_vram() in gpu.py called torch.cuda.mem_get_info() with no timeout. On hung NVIDIA drivers, this blocked forever. Added 5-second timeout via ThreadPoolExecutor.
  • safe_error GPU OOM false positives — errors.py pattern "cuda" in lower and "memory" in lower matched unrelated errors like "CUDA device not found in memory". Tightened to exact phrases "cuda out of memory" / "cuda error: out of memory".
  • _unique_output_path 10K stat loop — helpers.py looped up to 9,998 times trying filenames. On full disk or permission errors, wasted thousands of stat() calls. Capped at 100.
  • get_video_info silent defaults — Fallback to 1920x1080@30fps was logged at DEBUG level (invisible). Upgraded all 3 fallback paths to WARNING so users know when probe data is missing.
  • Job cleanup on every creation — _new_job() called _cleanup_old_jobs() synchronously on every job creation. With 1000+ old jobs, blocked the HTTP handler. Removed — periodic cleanup thread (already running every 5 minutes) handles this.
  • proc.poll() crash on stale handles — _cleanup_old_jobs() called .poll() on potentially corrupt Popen objects. Added try/except per handle.
  • async_job future not stored for cancel — Race window between _new_job() returning job_id and storing the future. Cancel requests during that window couldn't find the future. Now stores _thread reference immediately after submit.
  • multicam_xml path encoding — _path_to_url() didn't URI-encode spaces or special characters in filenames. Premiere Pro rejected the XML. Added urllib.parse.quote().
  • main.js parentNode null guard — select.parentNode.insertBefore() could crash if select element had no parent. Added null check.

[1.9.4] - 2026-03-27

Fixed (Batch 36 Audit)

  • P0: face_enhance/face_swap/upscale _p(pct, msg) crash — 3 on_progress closures in video_ai.py missing msg="" default; core modules call with 1 arg → TypeError. Added default.
  • P0: engagement attribute crash — shorts pipeline response accessed c.engagement.hook_strength directly; switched to getattr() with defaults for all 5 engagement fields.
  • P0: broll_plan inconsistent response — empty segments returned 2 keys but success path returned 4; frontend expects all 4. Fixed + added plan is None guard + getattr() for all window fields.
  • Security: style_arbitrary path traversal — /video/style/arbitrary accepted style_image without validate_filepath(); attacker could read arbitrary files. Added validation.
  • Security: plugin install path traversal — /plugins/install accepted arbitrary source directory without validate_path(). Added validation.
  • Security: plugin uninstall symlink escape — /plugins/uninstall didn't verify resolved plugin_dir stays within PLUGINS_DIR. Added os.path.realpath() containment check.
  • Security: depth_effects model_id injection — model_size param interpolated into HuggingFace model ID without validation in apply_bokeh_effect() and apply_parallax_zoom(). Added allowlist in all 3 functions.
  • GPU rate limiting gaps — /video/ai/interpolate and /video/ai/denoise (basicvsr method) bypassed rate_limit("ai_gpu"), allowing concurrent GPU OOM. Added guards.
  • Queue allowlist +9 routes — 9 newer routes missing from _ALLOWED_QUEUE_ENDPOINTS: interpolate, depth/map/bokeh/parallax, broll-plan, remove/watermark, upscale/run, multicam-xml, search/auto-index.
  • title_overlay preset allowlist — Missing lower_third, countdown, kinetic_bounce presets (present in title_render but not title_overlay).
  • Engine registry cache race — _availability_cache reads/writes in get_available_engines() were outside _lock; added lock protection.
  • main.js timer leak — _scanDebounceTimer missing from cleanupTimers(); leaked on panel close.
  • main.js safeFixed — defaultZoomVal slider used raw toFixed() instead of safeFixed() wrapper.
  • FFmpeg stderr truncation UX — run_ffmpeg() truncated stderr silently; now prepends "...[truncated] " marker when truncating.
  • Test fix — test_system_gpu expected gpu_available key but API returns available.

[1.9.3] - 2026-03-27

Added

  • JSON Structured Logging (Phase 0.3) — File handler now outputs JSON via python-json-logger. Each log line includes timestamp, level, module, job_id, message. Console handler stays plain text. Graceful fallback if python-json-logger not installed.
  • CI Coverage Enforcement (Phase 0.1) — .github/workflows/build.yml now runs pytest-cov with --cov-fail-under=50 threshold. CI also triggers on PRs and pushes to main. PyInstaller build skipped for non-release runs.
  • Structured Error Migration (Phase 0.2) — All route error handlers migrated from bare {"error": str(e)} to safe_error() from opencut/errors.py. Zero bare 500 error patterns remain across all 13 route files. Every API error returns {error, code, suggestion}.
  • Smart Tab Reordering (Phase 3.2) — Sub-tabs now physically reorder in the DOM based on contextual relevance scores. Highest-scoring features move to front within each tab group. resetTabOrder() restores original order on clip deselection.
  • Frontend Error Code Mapper (Phase 0.2) — enhanceError() now reads data.code field before regex fallbacks. Code-to-action mapping for GPU_OOM, MISSING_DEPENDENCY, FILE_NOT_FOUND, etc. with navigable settings links.
  • Core Module Unit Tests (Phase 0.1) — 15 additional core modules tested (silence, fillers, scene_detect, auto_edit, highlights, workflow, speed_ramp, audio, face_reframe, chromakey, video_fx, export_presets, diarize, audio_duck, thumbnail).
  • i18n String Extraction (Phase 6.2) — ~200 additional data-i18n attributes added to buttons, headers, labels, and tabs in index.html with corresponding en.json keys.
  • Pre-commit Hooks (Phase 0.1) — .pre-commit-config.yaml now includes ruff lint/format + pytest smoke suite (on pre-push). pre-commit added to dev dependencies.
  • Log Levels Audit (Phase 0.3) — 22 log calls across 10 files corrected: verbose processing steps downgraded INFO→DEBUG, fallback/degraded paths upgraded INFO→WARNING. No secrets found in logging.
  • Core Module Unit Tests Batch 2 (Phase 0.1) — 135 additional tests across 28 previously untested core modules (animated_captions, audio_enhance, audio_pro, audio_suite, caption_burnin, captions_enhanced, color_management, emotion_highlights, face_swap, face_tools, lut_library, motion_graphics, music_ai, music_gen, object_removal, particles, shorts_pipeline, style_transfer, styled_captions, transitions_3d, upscale_pro, video_ai, voice_gen, zoom, broll_insert, broll_generate, multimodal_diarize, social_post).
  • ExtendScript Mock Harness (Phase 0.1) — tests/jsx_mock.js provides fake Premiere Pro DOM (app.project, activeSequence, tracks, markers, ProjectItem). Tests 38 ExtendScript functions including ocApplySequenceCuts, ocBatchRenameProjectItems, ocAddSequenceMarkers, ocAddNativeCaptionTrack. 35 assertions pass under Node.js.

Refactored

  • async_job decorator adoption — All 97 manual _new_job() + threading.Thread + _update_job(status="complete") patterns across 6 route files converted to @async_job decorator. Fixes race condition where cancelled jobs could be overwritten to "complete". Removes ~2,800 lines of boilerplate. Decorator extended with filepath_required and filepath_param parameters.
  • Split video.py into 5 domain blueprints — Monolithic 3636-line video.py split into video_core.py (1395), video_editing.py (750), video_fx.py (687), video_specialty.py (489), video_ai.py (443). No file exceeds 1400 lines. All URL paths unchanged.
  • App factory pattern — server.py now exposes create_app(config) for isolated Flask instances. New opencut/config.py centralizes env var reads into OpenCutConfig dataclass. Tests use independent app instances. Module-level app = create_app() preserved for backward compat.
  • Install route factory — make_install_route() in opencut/jobs.py replaces 6 identical install endpoint handlers (depth, emotion, multimodal-diarize, broll-generate, face, crisper-whisper) with a single factory call per route.
  • Version sync CI enforcement — scripts/sync_version.py now includes package.json in targets and supports --check flag for CI. Added to .github/workflows/build.yml.

Security

  • Replaced all __import__() calls with importlib.import_module() / importlib.util.find_spec() in helpers.py, system.py, engine_registry.py, plugins.py

Fixed

  • _safe_error undefined — Fixed _safe_error reference in video.py multicam XML (F821 undefined name bug)
  • safe_int missing import — Added missing import in settings.py log tail endpoint (F821)
  • Timecode watermark plugin — start_tc variable now used in FFmpeg drawtext filter (was extracted but unused)
  • Version sync — All version strings (pyproject.toml, init.py, CEP manifest, UXP manifest, server startup banner, installer AppConstants.cs) now read from single source. Server banner uses __version__ instead of hardcoded string.
  • TooManyJobsError handling — 7 remaining _new_job() calls in system.py, search.py, timeline.py now properly catch TooManyJobsError and return HTTP 429 instead of 500.
  • audio_separate cleanup crash — temp_audio variable moved above try block to prevent NameError in finally clause if _resolve_output_dir() fails.
  • GPUContext memory leak — __exit__ now calls .cpu() on registered models before clearing, actually freeing GPU VRAM instead of only deleting a loop variable.
  • batch_id collision risk — Batch IDs extended from 8 to 16 hex chars, reducing collision probability from 1-in-65K to 1-in-4-billion batches.
  • XSS in engine registry — All API-sourced values (domain, engine name, display_name, quality, speed) now wrapped in esc() before innerHTML insertion.
  • elapsedTimer leak — startJob() now clears any existing elapsed timer before creating a new one, preventing interval accumulation on rapid re-starts.
  • mediaScanTimer leak — Periodic media scan setInterval now assigned to tracked variable and cleared on beforeunload.
  • pollTimer leak — trackJobPoll() now clears any existing poll timer before creating a new one; SSE fallback no longer creates duplicate trackers.
  • SSE parse errors silent — SSE onmessage handler now logs JSON parse failures to console instead of silently swallowing them.
  • Reframe dimension DoS — /video/reframe now enforces min_val=16, max_val=7680 on width/height parameters, preventing absurd allocation requests.
  • ASS subtitle injection — Caption burn-in now strips backslash sequences from source text before ASS formatting, preventing style override injection via subtitle content.
  • 23 import sorting violations fixed (I001)
  • 15 unused imports removed (F401)
  • 4 unused variables removed (F841)
  • Codebase at 0 lint warnings matching CI config

[1.9.1] - 2026-03-26

Added

  • Frontend Context Awareness (Phase 3.2) — CEP panel now calls POST /context/analyze after clip selection. Guidance banner shows clip-specific recommendations. Sub-tabs for high-scoring features get visual highlights. Dismiss button hides banner.
  • Parallel Batch Processing (Phase 5.4) — process_batch_parallel() in batch_process.py uses ThreadPoolExecutor for concurrent multi-clip operations. GPU ops limited to 1 worker, CPU ops scale to core count. Partial failure isolation — one item crash doesn't kill the batch.
  • Clip Notes Plugin (Phase 6.1) — Second example plugin at opencut/data/example_plugins/clip-notes/. SQLite-backed per-clip notes with POST /note, GET /notes, DELETE /note, GET /export (text/CSV). Shipped alongside timecode-watermark as reference plugins.
  • Route Smoke Tests (Phase 0.1) — Comprehensive smoke test suite in tests/test_route_smoke.py covering all 175+ endpoints across 13 blueprints. Verifies routes don't crash with minimal payloads.

Fixed

  • Ruff lint cleanup — Fixed unused imports in new test files (8 auto-fixed F401 violations)

[1.9.0] - 2026-03-26

Added

  • Contextual Awareness (Phase 3.2) — Clip type detection and feature relevance scoring. POST /context/analyze accepts clip metadata and returns scored features, guidance messages, and per-tab relevance scores. 35 features scored across 4 tabs based on clip tags (talking_head, audio_only, long_duration, etc.)
  • Plugin System (Phase 6.1) — Plugin loader discovers/validates/loads plugins from ~/.opencut/plugins/. Each plugin has a plugin.json manifest and optional Flask Blueprint routes registered under /plugins/<name>/. Endpoints: GET /plugins/list, GET /plugins/loaded, POST /plugins/install, POST /plugins/uninstall. Includes example timecode-watermark plugin.
  • Multicam XML Export — Generate Premiere Pro compatible FCP XML from multicam diarization cut data. POST /video/multicam-xml endpoint. CLI multicam command now exports XML instead of showing placeholder message.
  • Background Indexing with SQLite FTS5 (Phase 5.3) — Persistent footage index at ~/.opencut/footage_index.db with full-text search. Incremental re-indexing via mtime comparison. Endpoints: POST /search/auto-index, POST /search/db-search, GET /search/db-stats, POST /search/cleanup.
  • Response Streaming (Phase 5.2) — NDJSON streaming utilities for progressively delivering large result sets. GET /jobs/stream-result/<job_id> streams completed job results (segments, scenes, thumbnails) in batches of 50.
  • Context blueprint (routes/context.py) — 2 new endpoints
  • Plugins blueprint (routes/plugins.py) — 4 new endpoints
  • 4 new search endpoints — auto-index, db-search, db-stats, cleanup

Fixed

  • TooManyJobsError handling — All 96 manual _new_job() calls across video.py (47), audio.py (23), and captions.py (13) now catch TooManyJobsError and return proper 429 responses instead of 500 errors
  • AI GPU rate limiting — Added rate_limit("ai_gpu") to 6 GPU-heavy routes: video AI upscale, background removal, shorts pipeline, and 3 music AI generation routes. Prevents concurrent GPU OOM crashes
  • Settings import validation — /settings/import now validates workflow steps (requires endpoint + label per step) before saving, preventing malformed workflow injection
  • Preview frame bounds — width parameter bounded to 32-3840, detection_skip bounded to 1-30
  • Secure tempfile — Preview frame extraction uses tempfile.mkstemp() instead of predictable path construction

Security

  • 96 routes hardened against job limit bypass (TooManyJobsError → 429)
  • Input bounds added to prevent resource exhaustion via unbounded parameters
  • Workflow import step validation prevents malformed workflow injection
  • Fixed double rate-limit release in demucs/watermark install routes (decorator + explicit call)

[1.5.0] - 2026-03-23

Added

  • Repeated Take Detection — Jaccard similarity sliding-window to find and remove fumbled/repeated takes from transcriptions
  • YouTube Chapter Generation — LLM-powered (Ollama/OpenAI/Anthropic) topic detection and chapter timestamp generation
  • Footage Search — Index media library by spoken content; search by keyword or phrase across all clips
  • Post-Production Deliverables — Generate VFX sheets, ADR lists, music cue sheets, and asset lists as CSV from sequence data
  • Color Match — YCbCr histogram matching to match color profile of one clip to a reference clip
  • Multicam Auto-Switching — Diarization-driven cut generation for podcast multicam editing
  • Auto Zoom Keyframes — Face-detected keyframe generation for dynamic push-in zoom effect
  • Loudness Match — FFmpeg two-pass LUFS normalization to match loudness across multiple clips
  • NLP Command Parser — Natural language command dispatch via keyword matching and LLM
  • Timeline Write-Back — Apply cuts, markers, and keyframes directly to the active Premiere Pro sequence via ExtendScript
  • Beat Markers — Sync detected beat timestamps as sequence markers in Premiere Pro
  • Batch Rename — Rename project panel clips with find/replace pattern support
  • Smart Bins — Auto-sort project items into bins by rule (name contains, type, duration)
  • SRT to Native Captions — Import SRT files as native Premiere Pro caption tracks
  • Batch Export from Markers — Export individual clips defined by sequence markers
  • UXP Panel — New parallel panel for Premiere Pro 25.6+ using modern UXP APIs (com.opencut.uxp)
  • Timeline Tab — New CEP panel tab with all timeline write-back operations
  • Search & NLP Tab — New CEP panel tab with footage search and AI command input
  • Deliverables Section — Post-production document generation in the Export tab

Fixed

  • Improved progress callback support in color match, auto zoom, and loudness match operations
  • LLMConfig now properly instantiated as dataclass object across all routes

v1.2.0 (2026-03-15)

New Features

  • Waveform Preview: Visual waveform display on Silence tab with draggable threshold line synced to slider
  • Side-by-Side Preview: Before/after frame comparison modal for video effects
  • Dependency Health Dashboard: Grid view of all 24 optional dependencies with install status in Settings
  • First-Run Wizard: Animated 3-step onboarding overlay for new users (dismissible, persisted)
  • Output File Browser: Browse recent output files with Import-to-Premiere button
  • Favorites Bar: Pin frequently-used operations as quick-access chips below the nav tabs
  • Batch Multi-Select: Multi-file picker for batch operations (add selected, add all, clear)
  • Parameter Tooltips: Hover tooltips on range sliders explaining what each parameter controls
  • Custom Workflow Builder: Chain multiple operations into named reusable workflows (saved to ~/.opencut/workflows.json)
  • Audio Preview Player: Floating audio player to preview generated audio before importing
  • Settings Import/Export: Export all settings (presets, favorites, workflows) as JSON, import on another machine
  • Right-Click Context Menu: Quick-action context menu on clip selector (silence, transcribe, denoise, etc.)
  • Collapsible Cards: Click card headers to collapse/expand dense form sections
  • Job Time Estimates: Estimated processing time in the progress banner based on historical job data
  • Localization Framework: Language selector in Settings (placeholder for future i18n support)
  • Video Reframe: Resize/crop video for TikTok, Shorts, Reels, Instagram, Square, or custom dimensions with crop/pad/stretch modes
  • Clip Preview Thumbnail: Visual thumbnail + duration/resolution metadata when selecting a clip
  • Command Palette: Ctrl+K fuzzy search across all 28+ operations with keyboard navigation
  • Recent Clips: Dropdown of last 10 used clips, persisted across sessions
  • Trim Tool: Set in/out points to extract a clip portion (stream copy or re-encode)
  • Merge/Concatenate: Join multiple clips into one (fast stream copy or re-encoded for mixed formats)
  • Auto-Crop Detect: Smart reframe anchor using FFmpeg cropdetect for talking-head content
  • Audio Waveform Everywhere: Waveform preview buttons on Denoise and Normalize tabs (reuses Silence tab infrastructure)
  • Per-Operation Presets: Save/load settings per operation (persisted to localStorage)
  • Server Health Monitor: 10-second heartbeat with reconnect banner and toast notification
  • Output Deduplication: Auto-increment suffix prevents overwriting previous outputs

Backend Improvements

  • FFmpeg Progress Parsing: _run_ffmpeg_with_progress() helper parses -progress pipe:1 for real percentage updates
  • Subprocess Kill on Cancel: Job cancellation now terminates the running FFmpeg process via _kill_job_process()
  • Temp File Cleanup: Stale opencut_preview_*.jpg files cleaned up on server startup
  • File Serving Endpoint: GET /file serves local audio/video for preview player (path-restricted)

Backend (14 new endpoints)

  • POST /audio/waveform - Extract waveform peak data via FFmpeg PCM extraction
  • POST /video/preview-frame - Extract single frame as base64 JPEG at timestamp
  • GET /system/dependencies - Check 24 optional deps + FFmpeg install status
  • GET /outputs/recent - List recent output files from completed jobs
  • GET/POST /favorites - Favorites persistence to ~/.opencut/favorites.json
  • GET/POST/DELETE /workflows - Custom workflow CRUD to ~/.opencut/workflows.json
  • GET/POST /settings/export|import - Bundle presets + favorites + workflows
  • POST /system/estimate-time - Historical ratio-based time estimation
  • POST /video/reframe - Resize/crop video for target aspect ratio
  • GET /video/reframe/presets - Available reframe platform presets
  • GET /file - Serve local files for audio/video preview
  • POST /video/merge - Concatenate multiple clips (demux or filter modes)
  • POST /video/trim - Extract clip portion by in/out timecodes

v1.1.0 (2026-03-15)

New Features

  • Preset Save/Load: Save and load all panel settings as named presets (persisted to ~/.opencut/user_presets.json)
  • AI Model Manager: View all downloaded AI models (HuggingFace, Torch, Whisper) with sizes, delete unused models to free space
  • GPU Auto-Detection & Recommendation: Automatically recommend optimal model sizes and settings based on detected GPU VRAM
  • Job Queue System: Queue multiple jobs for sequential processing instead of waiting for each to finish
  • Enhanced Job History: Re-run previous jobs directly from history, toast notifications on completion
  • Keyboard Shortcuts: Enter to run active operation, 1-6 to switch tabs, Escape to cancel
  • Enhanced Drag-and-Drop: Drop files anywhere on the panel, not just the drop zone
  • Toast Notifications: Non-intrusive slide-in notifications for job completion/errors
  • Transcript Search: Search and navigate through transcript segments with highlight
  • Social Platform Export Presets: Quick export presets for YouTube Shorts, TikTok, Instagram Reel/Story/Post, Twitter/X, LinkedIn
  • Premiere Theme Sync: Detect Premiere Pro's UI brightness via CSInterface
  • Universal Auto-Import: Smart ExtendScript import function that routes files to appropriate project bins

Bug Fixes (v1.0.0 audit)

  • Fixed 6 broken install routes using subprocess.run instead of _sp.run
  • Fixed SSE race condition (dict copy + status read now inside job_lock)
  • Fixed VBS launcher path quoting for C:\Program Files\OpenCut
  • Replaced eval() with JSON.parse() in ExtendScript
  • Fixed wrong /cut/silence endpoint in social_ready workflow
  • Fixed jobStarting guard and no_input flag in all applicable routes
  • Fixed temp file leaks in audio_separate and watermark routes
  • Added user-select: text override for CEP Chromium inputs

v1.0.0-beta (2026-02-09)

First public beta release with full feature set.

Core (34 modules, 116 routes)

  • Silence detection and removal with configurable thresholds
  • Filler word detection (um, uh, like, you know, so, actually)
  • Speaker diarization via pyannote.audio
  • Job queue with async processing and progress polling

Captions (5 modules)

  • WhisperX / faster-whisper transcription with word-level timestamps
  • 19 caption styles (YouTube Bold, Neon Pop, Cinematic, Netflix, Sports, etc.)
  • 7 animated caption presets (pop, fade, slide, typewriter, bounce, glow, highlight)
  • Caption burn-in with style selection
  • Translation to 50+ languages

Audio (8 modules)

  • Stem separation via Demucs (vocals, drums, bass, other)
  • AI noise reduction + spectral gating
  • Loudness normalization (LUFS targeting)
  • Beat detection and BPM analysis
  • Pro FX chain: compressor, EQ, de-esser, limiter, reverb, stereo width
  • Text-to-speech with 100+ voices (Edge-TTS / F5-TTS)
  • Procedural SFX generator (tones, sweeps, impacts, noise)
  • AI music generation via MusicGen (AudioCraft)
  • Audio ducking (auto-lower music under dialogue)

Video (12 modules)

  • FFmpeg stabilization (deshake/vidstab)
  • Chromakey with green/blue/red screen + spill suppression
  • Picture-in-picture overlay with position/scale
  • 14 blend modes (multiply, screen, overlay, etc.)
  • 34 xfade transitions (fade, wipe, slide, circle, pixelize, radial, zoom)
  • 7 particle effect presets (confetti, sparkles, snow, rain, fire, smoke, bubbles)
  • 6 animated title presets (fade, slide, typewriter, lower third, countdown, kinetic)
  • Speed ramping with ease curves
  • Scene detection via PySceneDetect
  • Film grain, letterbox, vignette effects
  • 15 built-in cinematic LUTs + external .cube/.3dl support
  • Color correction (exposure, contrast, saturation, temperature, shadows, highlights)
  • Color space conversion (sRGB, Rec.709, Rec.2020, DCI-P3)

AI Tools (6 modules)

  • AI upscaling: Lanczos (fast), Real-ESRGAN (balanced), Video2x (premium)
  • Background removal via rembg (U2-Net)
  • Face enhancement via GFPGAN
  • Face swapping via InsightFace
  • Neural style transfer
  • Object/watermark removal (FFmpeg delogo + LaMA inpainting)
  • Auto-thumbnail extraction with AI scoring

Export & Batch (3 modules)

  • Platform presets: YouTube, TikTok, Instagram, Twitter, LinkedIn, Podcast
  • Batch processing with parallel execution
  • Transcript export (SRT, VTT, plain text, timestamped)

UI

  • 6 dark themes (Cyberpunk, Midnight OLED, Catppuccin, GitHub Dark, Stealth, Ember)
  • 6 main tabs with 24 sub-tabs
  • Persistent processing banner with progress
  • Single-job enforcement with UI lockout
  • Auto-import results back to Premiere Pro timeline