Skip to content

CHANGELOG.md

Latest commit

 

History

History
840 lines (692 loc) · 44.7 KB

CHANGELOG.md

File metadata and controls

840 lines (692 loc) · 44.7 KB

Changelog

Notable changes from the June 2026 hardening/audit pass. The authoritative record also lives in the git commit messages.

[Unreleased]

Added — 537 more coverage expansion tests (658 total, 30+ modules)

  • New test files: test_coverage_infra.py (169 tests covering errors, schemas, workers, config, helpers, checks), test_coverage_core_logic.py (137 tests covering timecode_utils, metadata_tools, safe_zones, hdr_tools, dead_time, stream_chapters, timelapse, highlight_detect, soft_subtitles, adr_cueing), test_coverage_core_extra.py (159 tests covering analytics, selects_bin, show_notes, ffmpeg_builder, ab_av1, ab_compare, cinemagraph, surround_mix, gpu_dashboard, display_calibration, apple_silicon, amd_gpu), test_coverage_schemas_mcp.py (72 tests covering schemas, MCP tool catalog validation, error factories, config, helpers, checks, workers). CI coverage floor raised from 55% to 58%.

Added — Centralized config schema registry with versioned migration

  • New register_config_schema() and read_user_file_versioned() in opencut/user_data.py. Config files can declare a schema version and register migration functions. read_user_file_versioned() auto- migrates old configs on read, stamps _schema_version, and persists the upgraded data. Unknown keys are preserved, not stripped.

Added — 121 coverage expansion tests across 10 untested core modules

  • New tests/test_coverage_expansion.py adds focused unit tests for: url_ingest (16), multimodal_index (7), edl_aaf (14), media_relink (9), auto_update (14), structured_ingest (12), batch_conform (5), project_organizer (15), storage_tiering (10), render_queue (9). All tests are fast (<1s total), require no external services, and run without optional ML/GPU dependencies. CI coverage floor raised to 55%.

Changed — Docker compose hardened + MCP sidecar profile

  • Added CPU/memory resource limits (CPU: 4c/4G, GPU: 8c/16G), JSON-file log rotation (50-100MB × 3-5 files), and an opt-in MCP sidecar profile (docker compose --profile mcp up). GPU service now reads NVIDIA_VISIBLE_DEVICES from environment instead of hardcoding all.

Changed — Default Whisper model upgraded to large-v3-turbo

  • CaptionConfig.model default changed from base to large-v3-turbo. Within 1-2% WER of large-v3 at 5.4x speed. Existing users keep their configured model.

Added — CLI config export/import/validate commands

  • opencut config export [-o backup.json] dumps all ~/.opencut/*.json settings to a single JSON bundle. opencut config import backup.json restores them. opencut config validate checks all config files for JSON parse errors (exits non-zero on failure).

Added — 12 new animated caption presets (19 → 31)

  • Neon Glow, Slide Up, Word Pop, Emoji React, Speaker Colored, Retro VHS, Podcast, Gaming, Luxury, TikTok Bold, Handwritten, and RTL Arabic. Caption style count now exceeds 30, matching paid competitor expectations.

Added — Smart pause detection for silence removal

  • New smart_pause=true option on /silence route and detect_speech() that preserves dramatic pauses (0.8-3.0s silences preceded by sentence-ending speech) instead of cutting them. When a transcript is available, uses sentence-final punctuation as context; without transcript, falls back to a duration heuristic that preserves short mid-speech pauses.
  • New filter_smart_pauses() function in opencut/core/silence.py with 2 unit tests.

Added — MCP tool catalog expanded from 39 to 86 curated tools

  • Added 47 new curated MCP tools covering video editing (trim, merge, concat, reframe, stabilize, speed, PiP, blend), video effects (chromakey, LUT, FX, denoise, interpolation, transitions), captions (burn-in, translate, styled, animated, karaoke, SRT import), audio (effects, duck, isolate, normalize, filler removal, beat markers, enhance), timeline/workflow (run, presets, batch rename, smart bins, beat cut, export), deliverables (VFX sheet, ADR list, music cue sheet), search (NLP command, URL ingest), and system (info, GPU, dependencies, feature state, social upload).
  • Agent clients (Claude, Cursor, etc.) now discover 86 OpenCut capabilities vs 39 previously. Extended tools (1,482 route-level) remain opt-in.

Security — Werkzeug floor raised to >=3.1.6 (CVE-2026-27199)

  • Bumped Werkzeug floor from >=3.1.5 to >=3.1.6 in both pyproject.toml and requirements.txt. CVE-2026-27199 is a safe_join() Windows device-name bypass that causes hanging file reads in send_from_directory.

Fixed — Route manifest + README badge drift

  • Regenerated route_manifest.json and feature_readiness.json to reflect recent url_ingest, multimodal_index, and AutoShot additions. README badges and inline counts updated: 1,545 shipped routes, 6 stubs.

Changed — Feature-state now reflects route readiness

  • Generated feature-readiness records now cross-reference the route manifest's readiness field instead of defaulting to "available". Routes classified as dependency-gated produce "missing_dependency" features; stub routes produce "stub" features. Panels disable buttons for these features before the user clicks them (was: all generated features showed as available regardless of route status).

Added — Import-by-link footage ingest

  • New url_ingest module (opencut/core/url_ingest.py): fetches video from a URL to ~/.opencut/media_ingest/ so existing routes consume it as a local file. Uses yt-dlp when installed (YouTube, Vimeo, Zoom Clips, Medal, etc.); falls back to plain HTTP download for direct-link URLs. SHA-256 content- addressed caching avoids re-downloads.
  • New routes: POST /search/ingest (async job), GET /search/ingest/cache, DELETE /search/ingest/cache.

Added — Local multimodal footage search (OCR + audio event index)

  • Footage index schema upgraded to v2: ocr_text and audio_tags columns added to the footage table, FTS5 index rebuilt to cover all four fields (file_path, transcript, ocr_text, audio_tags). Existing v1 databases auto-migrate on first access.
  • New multimodal_index module (opencut/core/multimodal_index.py): extracts on-screen text from key frames via pytesseract or easyocr, and classifies audio events (speech, music, silence, tempo) via librosa spectral heuristics with a basic fallback.
  • New POST /search/multimodal-index route: indexes files with transcript + OCR text + audio event tags in a single pass. OCR and audio tags are opt-in (default on) via ocr and audio_tags body params.
  • FTS5 search now matches across transcript, on-screen text, and audio event tags — a query for "music" or "logo" hits audio events and OCR text respectively, not just spoken words.
  • check_ocr_available() added to checks.py.

Added — AutoShot scene-detection engine

  • Added AutoShot as a selectable scene-detection engine with priority 85 (above TransNetV2's 70). detect_scenes_auto() now tries AutoShot first for gradual-transition-aware detection, falling back to TransNetV2 → PySceneDetect → FFmpeg threshold. check_autoshot_available(), engine registry entry, and 5 new tests added.

Added — Model privacy/license posture in feature-state

  • Feature-state entries now expose privacy (local-only / cloud), license (SPDX token), and advisory_notes from the model-card registry. The CEP panel annotates gated buttons with data-feature-privacy/license attributes, privacy chip badges ("Local" / "Cloud"), and advisory notes in tooltips.

Changed — Collapsed 4 duplicate 501 stubs onto canonical pipelines

  • /video/trailer/generate now delegates to trailer_gen.generate().
  • /video/outpaint now delegates to frame_extension.outpaint_aspect_ratio().
  • /agent/search-footage now delegates to semantic_video_search.semantic_search().
  • /agent/storyboard now delegates to ai_storyboard.generate_storyboard_from_script().
  • Shipped route count rises from 1,536 to 1,541. Users no longer hit 501 for capabilities that already have working local implementations.

Changed — Default depth backend upgraded to Depth Anything 3 Small

  • New depth_anything_3 module + a depth-anything/DA3-SMALL (Apache-2.0, single-transformer SOTA, ~+25% geometry) default for CineFocus bokeh/parallax, with Depth-Anything-V2-Small retained as the automatic fallback. CineFocus now routes both depth-load sites through a central _load_depth_backend() that resolves DA3 first and degrades to DA2 on any load failure.
  • New depth_anything_3 engine-registry entry (priority 85 > DA2's 80) + check_depth_anything_3_available(), a da3-small model entry, and tests/test_depth_engines.py.

Changed — Re-aimed the IC-Light relight engine at v1 (Apache-2.0)

  • relight_iclight.py was titled "IC-Light V2" but IC-Light v2 is non-commercial and its full weights were never publicly released — un-shippable under MIT. Re-targeted it at IC-Light v1 (lllyasviel/IC-Light, Apache-2.0, real public weights). Removed the v2 framing from docstrings/hints, added MODEL_ID/NAME/ LICENSE, check_iclight_available(), an ic-light-v1 model entry, and a new relight engine-registry domain. The /video/relight/iclight route now reports the v1 Apache-2.0 licence. Registry test suite added.

Changed — Default speaker diarization upgraded to pyannote community-1

  • diarize() now defaults to pyannote/speaker-diarization-community-1 (CC-BY-4.0, "always freely accessible", lower DER) and automatically falls back to legacy speaker-diarization-3.1 when the community model can't be loaded (terms not accepted, offline cache miss). The model is configurable via DiarizeConfig.model.
  • New diarization engine-registry domain: pyannote_community1 (default, priority 80) > pyannote_legacy (60, fallback) > sortformer (50, optional, NVIDIA NeMo-gated). Added check_diarization_available() / check_sortformer_available(), a pyannote-community-1 model entry, and tests/test_diarization_engines.py.

Added — Normalized 0–100 virality score on highlights/clips

  • EngagementScore.compute_virality() maps the existing engagement dimensions (hook / emotional / pacing / quotability) to a single deterministic, hook-forward 0–100 virality score — the headline number Opus Clip paywalls — so creators can sort and threshold clips at a glance. It is a heuristic, not a performance guarantee, and is distinct from the editing-relevance overall composite.
  • Surfaced as a top-level virality field (plus inside the engagement block) on POST /video/highlights and the magic-clips/shorts response. Deterministic unit tests added.

Added — LatentSync lip-sync as the dub pipeline's opt-in final stage

  • Wired a new lipsync_latentsync engine into a new lip_sync engine-registry domain. The dub pipeline (transcribe→translate→voice-clone→render) can now re-animate the speaker's mouth to the new audio — the visual lip-sync stage Rask/HeyGen/sync.so paywall — via DubConfig.lip_sync_backend="latentsync" and POST /ai/lip-sync (backend param).
  • Opt-in by design (RESEARCH Open Question 3): LatentSync's code is Apache-2.0 but its checkpoint licence is unconfirmed, so the engine is registered below the always-available heuristic and is never auto-selected. The dub stage degrades gracefully: LatentSync → heuristic jaw-overlay → audio-only dub, so it never hard-fails.
  • Added check_latentsync_available(), a latentsync-1.6 model-download entry, and tests/test_lipsync_engines.py (registry, opt-in ordering, availability gating, license hygiene, dub-config plumbing). Full diffusion forward activates once local weights are present.

Added — SeedVR2 upscaling engine (Apache-2.0, one-step diffusion)

  • Wired the previously-dark upscale_seedvr2 stub into the upscaling engine registry as seedvr2, gated by check_seedvr2_available() (diffusers + torch). It is registered at higher priority than Real-ESRGAN, so it auto-selects when its Apache-2.0 SeedVR2-3B weights are installed and falls back to Real-ESRGAN otherwise. POST /video/ai/upscale gained an engine param (auto/seedvr2/ realesrgan) that tries SeedVR2 first and degrades to Real-ESRGAN on any unavailability (recorded in the response notes), never hard-failing.
  • Added a seedvr2-3b model-download catalog entry (Apache-2.0, ~6 GB) and a registry test suite (tests/test_upscaling_engines.py). The heavy diffusion forward activates once the local weights are present; the entry point raises a structured error when absent (matching the NeMo ASR engine contract).

Added — Beat-synced cuts to the active sequence (POST /timeline/beat-cut)

  • New timeline-native beat-cut path: detect beats in a clip's audio and emit beat-aligned clip boundaries as a cut list the panel reviews and ripple-applies to the active Premiere sequence. Unlike /audio/beat-sync (which renders a new montage file), this produces cut points for the sequence already open.
  • core/beat_sync_edit.py::plan_timeline_beat_cuts() — pure planner over detected beats + a clip duration, with mode selection (every beat / 2 beats / bar / 2 bars / accents) and min_cut_duration merging. Returns {start, end, duration, beat_number, strength, label} segments (the shape ocApplySequenceCuts and the cut-review panel already consume).
  • CEP panel: a "Beat-Synced Cuts" card on the Timeline → Beat Markers sub-tab, wired through the shared cut-review preview + ripple-apply pipeline (the same one silence/filler/auto-edit use). Added to the queue allowlist and i18n.
  • Tests: 4 planner unit tests + 3 route smoke tests; i18n/a11y/parity green.

Changed — Brand/namespace disambiguation: distribution token opencut-ppro

  • Resolved the carried brand/namespace question (RESEARCH Open Question 1). The product keeps the name OpenCut, but the package-manager distribution token is now opencut-ppro to stay unambiguous against the unrelated browser editor opencut.app (~48K stars) across PyPI/Homebrew/winget/SEO. The -ppro suffix marks the Premiere Pro integration that uniquely identifies this tool.
  • pyproject.toml [project].name → opencut-ppro; SBOM root component and README pip install examples updated to match. Unchanged: the import package (import opencut), CLI commands (opencut, opencut-server, opencut-mcp-server), and CEP/UXP extension IDs. Decision recorded in the README "Naming & distribution" section. (All candidate names were free on PyPI; the actual publish remains a credential-gated roadmap item.)

Security — Bundled FFmpeg 8.1.1 + June-2026 security-patch provenance

  • Reconciled the bundled-FFmpeg version pin to 8.1.1-essentials_build-www.gyan.dev across AppConstants.cs, OpenCut.iss, and the manifest test (the Inno pin had drifted to 8.1, leaving test_inno_installer_writes_ffmpeg_manifest red).
  • New opencut/core/ffmpeg_provenance.py is the single source of truth for the acceptable bundled build. It parses ffmpeg -version (gyan.dev release, gyan git-master, BtbN git describe, distro n-prefixed) and grades it against two lanes: a release floor (>= 8.1.1) or a git-master snapshot dated >= 2026-06-10 (reference commit b29bdd3715). This asserts a security patch level — the June-2026 FFmpeg zero-days (CVE-2026-6385 + CVE-2026-39210..39218, crafted-media heap/stack overflows) landed as post-release master commits, so a bare 8.1.x tag is insufficient on its own.
  • GET /system/capabilities now carries ffmpeg.security (graded result) and a ffmpeg_below_security_floor finding so a stale bundled binary is visible at runtime.
  • scripts/verify_ffmpeg_provenance.py is a stdlib build/CI gate that fails closed when the bundled binary is below the floor and records ground-truth provenance (version, git commit/snapshot date, lane, CVE list) to a manifest.
  • Installer manifests (installer.json, both WPF + Inno) now record bundled_ffmpeg_security_floor. Provenance requirements documented in docs/RELEASE_PROVENANCE.md.

Added — NVIDIA NeMo ASR engines (Parakeet TDT 0.6B v3 + Canary 1B Flash)

  • Wired the previously-dark asr_parakeet (streaming) and asr_canary (batch) stubs into the transcription engine registry as parakeet_tdt and canary_1b_flash. Both are gated by a new check_nemo_asr_available() check (accepts the nemo / nemo_toolkit[asr] install), retain faster-whisper and CrisperWhisper as fallbacks, and ship parakeet-tdt-0.6b-v3 / canary-1b-flash model-catalog entries for on-demand download. Engines report unavailable (never crash) when NeMo is absent.

Changed — Route manifest tags readiness; advertised count excludes stubs

  • dump_route_manifest now classifies every route as implemented / dependency-gated / stub via static handler inspection, and records readiness_counts plus a shipped_route_count (total minus the 10 HTTP-501 strategic stubs). Manifest schema bumped to version 3.
  • New GET /system/route-readiness endpoint surfaces the shipped count and the explicit stub / dependency-gated rule lists so panels and tooling stop presenting 501 stubs as shipped capabilities. (Panels already gate data-feature-id actions via F100 feature-state.js; the Tier-3 stub actions are not surfaced as panel buttons, so none are presented as shipped.)
  • README route badge and route-count claims now advertise the shipped count (1,536) instead of the raw total; scripts/sync_badges.py binds the badge to shipped_route_count.

Changed — Structured error responses across video/audio routes

  • Migrated the remaining 35 raw jsonify({"error": ...}) responses in routes/video_editing.py, routes/video_core.py, and routes/audio.py to the structured error_response() helper. Every error now returns a machine-readable code, a suggestion, and request-id correlation, and is logged through the typed-error path — matching the rest of the API surface the frontend already consumes.

Security — AST whitelist for the expression sandbox

  • Replaced the deny-list AST check in expression_engine.py with a closed node whitelist: only arithmetic/boolean/comparison nodes, calls to known sandbox functions, and ternaries are permitted. Every ast.Attribute node is now rejected outright, so dunder-walk escapes (().__class__.__subclasses__) and obfuscated variants cannot be expressed at all. Call targets are constrained to the sandbox's callable table; unknown calls (breakpoint(), eval()) are rejected at validation time.

Security — AST-level dunder/builtin guard for the scripting console

  • Added an obfuscation-proof structural check to scripting_console.py that parses the script and rejects any dunder attribute access (__base__, __flags__, __delattr__, ... — including escape vectors absent from the hand-maintained substring list) and any reference to a blocked builtin (getattr, eval, globals, ...), regardless of source formatting. The lowercased substring scan is retained as a first-pass net.

Security — Transitive web-dependency floor pins (RA-23)

  • Floored Werkzeug (>=3.1.5) and Jinja2 (>=3.1.6) in core pyproject.toml dependencies — they ship transitively with flask and are always installed, so a clean resolver (no requirements-lock.txt) can no longer pull Werkzeug CVE-2026-21860 / CVE-2025-66221 (safe_join Windows device-name DoS) or Jinja2 CVE-2025-27516 (sandbox breakout).
  • Floored requests>=2.33.0 and urllib3>=2.6.3 in the standard and all extras, where they enter via faster-whisper's huggingface-hub fetch path — closing urllib3 CVE-2026-21441 (decompression-bomb DoS) and requests CVE-2026-25645 in the release-audited lane.
  • Synced requirements.txt flask/waitress floors to the pyproject values and added the new security pins; tests/test_dependency_surface.py now asserts the transitive floors and the resolver lane matches the lockfile.

Added — CJK, Bengali, and RTL caption font-fallback

  • Added script-aware font fallback chains to styled_captions.py for CJK (Noto Sans CJK, MS Gothic, SimSun, Malgun, Meiryo, Yu Gothic), Indic (Noto Sans Devanagari/Bengali, Mangal, Vrinda, Nirmala), and RTL (Noto Sans Arabic/Hebrew, Tahoma) scripts in both Pillow and Skia renderers.
  • _load_font() and _load_skia_typeface() now accept text_hint to detect caption text scripts and try script-capable fonts before falling back to Western fonts.
  • Added get_caption_font_info(text, style_name) for the UI to surface which fallback font was selected or warn when no capable font is available for the detected scripts.
  • Added Bengali fixture (bengali_indic) to caption Unicode validation (F223), bringing complex-script fixture count from 5 to 6.
  • Added rtl_arabic burn-in style preset with right-aligned bottom positioning (ASS alignment 6) for natural RTL reading direction.

Added — D3D12VA and Vulkan HW encoder detection

  • Added D3D12VA encoders (h264_d3d12va, hevc_d3d12va, av1_d3d12va) to hw_accel.py — vendor-agnostic Windows hardware encoding for non-NVIDIA users when FFmpeg 8.x is available.
  • Added Vulkan Video encoders (h264_vulkan, hevc_vulkan, av1_vulkan) for cross-platform GPU encoding via Vulkan Video API.
  • Updated priority order: nvenc > qsv > amf > d3d12va > videotoolbox > vulkan.
  • Added quality presets and test-encode options for both new HW types.
  • Added 5 new tests for D3D12VA/Vulkan parsing, priority, and verification.
  • Wired D3D12VA and Vulkan into /hw/encode route validation and all HW export preset descriptions. Installer version strings updated for FFmpeg 8.1.

Fixed — Engineering audit (batch 4): multicam kwarg mismatch

  • video_editing.py: Fixed speaker_map= → speaker_to_track= in generate_multicam_cuts() call. The route passed a keyword argument that didn't match the function signature, causing a TypeError at runtime on every multicam cut request.

Fixed — Engineering audit (batch 3): SQLite connection safety

  • job_store.py: Wrapped PRAGMA setup in try/except to close the connection on failure instead of leaking it (lines 134-140).
  • job_store.py: Added PRAGMA wal_checkpoint(TRUNCATE) before closing connections in close_all_connections() to prevent orphaned WAL/SHM files on Windows shutdown.

Fixed — Engineering audit (batch 2): VideoCapture validation

  • segment_sam2.py: Added cap.isOpened() check after cv2.VideoCapture() to fail early with a clear error instead of silently producing corrupt output.
  • multicam_visual.py: Added cap.isOpened() check in analyze_frames().

Fixed — Engineering audit: correctness, safety, resource leaks

  • segment_sam3.py: Fixed unreliable frame_count scope check ("frame_count" in dir() → initialized at top), added VideoCapture.isOpened() / VideoWriter.isOpened() validation, moved GPU cleanup (del predictor, state + empty_cache()) into a finally block to guarantee cleanup on exception paths.
  • semantic_video_search.py: Fixed unused fps variable in _extract_key_frames(), fixed progress callback always reporting 0% (now computes actual percentage), fixed temp directory leak in build_clip_index() and semantic_search() (added finally: shutil.rmtree), removed dead family variable from _compute_frame_embeddings.
  • object_intel_routes.py: Fixed 4 instances of validate_filepath() misused on output directories (should be validate_path() — files vs. directories use different validation functions).

Added — Versioned visual-search engine registry

  • Added SEARCH_ENGINES registry to semantic_video_search.py with 4 engines: CLIP ViT-B/32 (default), CLIP ViT-L/14, SigLIP ViT-B/16, and SigLIP 2 ViT-B/16.
  • Cache keys now include engine ID and schema version — different models use separate cache directories, preventing embedding dimension mismatches.
  • build_clip_index() and semantic_search() accept an engine parameter.
  • Model loading supports both CLIP (CLIPModel/CLIPProcessor) and SigLIP (AutoModel/AutoProcessor) families.
  • Added list_search_engines() function and GET /search/semantic/engines route for UI engine discovery.
  • Routes /search/semantic and /search/semantic/index now accept engine parameter to select among registered models.

Added — Multicam visual speech cues (audio+visual mode)

  • Added core/multicam_visual.py with MediaPipe Face Mesh lip movement scoring and shot type classification (wide/medium/close by face-to-frame ratio) for multicam cut refinement.
  • /video/multicam-cuts now accepts mode: "audio+visual" to combine diarization with visual lip movement analysis. Cross-mic bleed is resolved by checking which on-camera face is actually speaking.
  • Each cut in visual mode gains visual_confidence, shot_type, and optional visual_override annotations.
  • Added check_visual_multicam_available() to checks.py.

Added — SAM 3 text-prompted object removal engine

  • Added core/segment_sam3.py with text-prompted object segmentation and tracking via SAM 3, plus segment_video_auto() with auto-fallback to CLIP+SAM2 when SAM3 is not installed.
  • Added POST /video/object-remove/sam3 route accepting text queries (e.g., "the watermark in the corner") with SAM2 fallback.
  • Added GET /video/object-remove/engines listing available segmentation backends with capability flags (text_prompts, click_prompts, box_prompts).
  • Updated get_removal_capabilities() to include sam3 availability.
  • Added check_sam3_available() to checks.py.
  • Added /video/object-remove/sam3 to queue allowlist.

Changed — Premiere 26 positioning audit

  • Updated README tagline to lead with silence-cut-to-timeline, stem separation, voice cloning, animated captions, local LLM, and social export.
  • Added "What OpenCut adds beyond Premiere 26" comparison table contrasting 10 OpenCut-unique capabilities against Adobe's native feature set.
  • Added Descript to cost comparison table with stem separation and voice clone columns across all competitors.
  • Documented Premiere 26.x manifest compatibility in UXP panel section (CEP [13.0,99.9], UXP minVersion 25.6).

Changed — onnxruntime floor raise

  • Raised onnxruntime floor from >=1.25 to >=1.26 in both [ai] and [ai-gpu] extras (pyproject.toml). 1.26.0 hardens multiple OOB/overflow scenarios and replaces unrestricted Python setattr config with an allowlist.
  • Documented the floor raise rationale in docs/PYTHON_ADVISORIES.md.

Fixed — UXP i18n

  • Added Spanish diacritics to 400 keys in UXP es.json (507 accent/tilde characters introduced; zero existed before).
  • Replaced hardcoded English "-- Select a clip --" in main.js with i18n t() call and added corresponding en.json/es.json keys.
  • Added scripts/lint_locales.py — locale lint checking key uniqueness, placeholder parity between en/es, missing keys, and diacritics regression. Supports --check flag for CI enforcement.

Fixed — CEP style.css consolidation

  • Merged duplicate @media (prefers-reduced-motion: reduce) blocks into one.
  • Completed light theme variable block with missing accent gradients, mesh background effects, and glow-opacity tokens (8 new variables).
  • Retokenized 14 hardcoded hex color literals in light theme overrides to use CSS custom properties (--text-primary, --text-secondary, --bg-surface, --bg-raised).

[1.33.1] — 2026-06-13 — CI hardening, privacy mode, v30 compat

Security — dependency floors

  • Flask floor raised to >=3.1.3 (CVE-2026-27205 Vary:Cookie info disclosure).
  • Waitress floor raised to >=3.0.1 (CVE-2024-49768/49769 request smuggling).
  • PyInstaller pinned to >=6.0 in CI (CVE-2025-59042 local privilege escalation).

Changed — CI

  • Added Python 3.13 test matrix job running core test suite on Ubuntu.
  • Added Python 3.13 classifier to pyproject.toml.

Fixed — UXP i18n

  • Deduplicated 113 duplicate keys in UXP en.json (1497 → 1384 unique keys).
  • formatI18n now uses a function replacement to prevent $& and $' in substitution values from corrupting output.

Added — captions

  • Caption export preflight (POST /captions/export/preflight) checks segment validity, timecode range, QC compliance, and host version before export. Returns a fallback strategy (native, srt_sidecar, or burnin) when native Premiere caption export is risky. Premiere 26.0/26.0.1 flagged based on community reports.

Fixed — CEP panel

  • Wizard and audio-preview dialogs now route through the overlay stack (activateOverlay/deactivateOverlay) for proper focus trap, inert background, and topmost-only Escape dispatch. Removes the waterfall Escape listener that closed every visible surface at once. Audio preview gains aria-modal="true".

Added — auto-editor v30

  • detect_auto_editor_generation() distinguishes native v30+ Nim binary from legacy v29 pip package. Install hints across model cards and feature registry now list both install paths.

Added — privacy

  • Local-only privacy mode via OPENCUT_LOCAL_ONLY=1 env var or /settings/local-only UI toggle. Blocks cloud LLM providers (keeps Ollama), cloud vision APIs, cloud TTS, social uploads, stock search, and telemetry with structured local-alternative error messages.

Fixed — documentation

  • Removed stale ROADMAP-NEXT.md and PROJECT_CONTEXT.md references from CONTRIBUTING.md and docs/ROADMAP.md pointer file.

Changed - CEP/UXP polish

  • CEP first-run, launcher, disabled-action, and settings copy now frames the workflow around clear workspace actions instead of shortcut-first language.
  • CEP and UXP panel surfaces now share tighter 8px geometry, clearer focus and disabled states, calmer empty-state feedback, and a more cohesive blue/teal premium accent system across onboarding, workspace, tabs, cards, and controls.
  • UXP tab navigation now binds before backend discovery so the panel remains navigable while the local backend is offline or still being detected.

Fixed — CEP panel

  • Settings no longer exposes non-English UI language choices before matching locale files ship, preventing a selection that cannot persistently localize the panel.
  • Settings import now sanitizes local panel preferences before writing localStorage, ignoring unknown keys and unsupported preference values while surfacing storage failures as a warning toast.

Changed — audio privacy

  • Standalone TTS, auto-dubbing, and overdub workflows now default to local-first TTS selection; Edge TTS and external API providers run only when explicitly selected.

Fixed — release process

  • Release smoke now catches MCP registry version drift, the committed MCP registry manifest is back on the package version, and stale UXP harness/i18n guardrails have been brought back in sync.
  • README test-count badge now syncs to the live suite count so the release readiness gate stays green after new coverage lands.
  • Release smoke and doc-size drift checks now target only live documentation surfaces (CLAUDE.md and README.md) instead of removed planning files.
  • Version sync now covers the security support table, CEP package-lock root metadata, and the C2PA claim-generator string so release smoke fails when those public version surfaces drift.

Security — data loading

  • Character embeddings and depth-compositor archives now call np.load with pickle explicitly disabled, with regression coverage for both paths.

Security — cache safety

  • Preview-cache metadata is now scoped to the active cache directory and ignores metadata entries whose files resolve outside that directory before eviction, invalidation, or flush can delete them.

[1.33.0] — 2026-06-11 — June hardening & quality pass

Fixed — Docker/runtime

  • Docker HTTP containers now set OPENCUT_ALLOW_REMOTE=1 alongside the intentional 0.0.0.0 bind, preventing startup crash-loops while preserving the non-loopback auth gate.

Documentation

  • README planning links now resolve in a fresh clone and name installer artifacts by the release version pattern instead of a stale filename.

Security — dependency policy

  • Standalone depth/model-loading installs now require transformers>=5.3; the lower Transformers floor is confined to the explicit, documented torch-stack/WhisperX exception and the advisory audit allow-list now covers the current Transformers config-injection CVE only in that opt-in lane.

Security — server boundary

  • Default CORS is now closed instead of allowing null/file://, and /health withholds CSRF bootstrap tokens from those origins even if they are explicitly added back for compatibility.
  • Non-loopback server binds now serve through Waitress instead of Flask's Werkzeug development server, covering the Docker/remote runtime path while keeping local loopback/debug launches on the dev server.
  • MCP HTTP sidecar binds on non-loopback interfaces now require the persistent X-OpenCut-Auth token before serving health, tool-list, or JSON-RPC calls.
  • Opt-in generated MCP route tools now run the same fail-fast path validation as curated tools, including nested query and body path fields.
  • Remote HTTP startup now fails closed when OPENCUT_ALLOW_REMOTE=1 is set but the auth-token middleware cannot be installed.
  • Every mutating Flask request now passes through a global CSRF gate, so new POST/PUT/PATCH/DELETE routes stay protected even if a handler omits the per-route decorator.
  • Kinetic text, data-animation, shape-animation, and IMF export routes now clamp render dimensions and font sizes before dispatching to allocation-heavy render code.
  • The writable ~/.opencut/packages fallback install directory is now appended to sys.path instead of taking priority over stdlib and bundled modules.

Fixed — UXP panel

  • UXP job submission now rejects concurrent backend jobs before they overwrite the shared tracker/SSE stream, and job-action buttons stay locked while a backend job is starting or running.
  • The UXP Refresh button now rescans backend ports 5679-5689, and failed background health checks retry against a newly detected port before staying offline.
  • The UXP live-updates WebSocket now consumes the bridge URL reported by the backend, uses a capped reconnect backoff, and honors manual disconnects so the Stop button actually stops reconnect attempts.
  • UXP job polling now tolerates short /status/<job_id> interruptions before failing a running job, preventing one dropped request from abandoning work.
  • Periodic UXP backend health checks no longer repaint the workspace as connecting/offline before each background probe completes.
  • FCC caption display settings now unwrap BackendClient response payloads before populating token selects or applying preview CSS, making the compliance card functional again.
  • Cancelling an active job now clears any button left in the panel's loading state, so the initiating action is usable again without reloading the panel.

Security — routes layer

  • Closed the unvalidated output_path write sinks. @async_job validates only the primary input filepath; eight route modules forwarded a user-supplied output_path/output_file/db_path straight into an FFmpeg -y or open(..., "w") sink, allowing arbitrary local file overwrite via a crafted request (CSRF-protected, but reachable by any local caller). All secondary write targets now pass validate_output_path/validate_path:
    • audio_production_routes (declip/dehum/decrackle/dewind/dereverb, room-tone generate/fill, M&E mix, fingerprint db_path read+write)
    • integration_routes (adjustment-layers apply, flight-map, jog-mapping save)
    • overlay_routes (safe-zones + the three output_path_override routes)
    • solver_agent_routes (3D overlay), timeline_auto_routes (auto-mix — replaced the brittle startswith/colon check that let absolute paths and ..\ traversal through), cloud_distrib_routes (farm render), editing_workflow_routes (7 export/assemble routes), color_mam_routes (22 output sinks via a shared validated helper).

Fixed — CEP panel (extension/com.opencut.panel)

  • Cut Review panel was completely broken. A for (var t = ...) loop counter in showCutReview() shadowed the t() i18n function in the same scope, so every silence / filler / auto-edit / highlights / repeat-detect completion threw TypeError: t is not a function and the review panel never opened. Renamed the loop variable.
  • A throwing job-done listener could lock the UI forever. The completion dispatch loop ran listeners with no isolation before hideProgress(), so any listener exception left the processing banner and body.job-active pointer lock stuck on. Each listener is now wrapped in try/catch.
  • Escape no longer silently cancels a running job when it was pressed only to dismiss a dropdown, menu, wizard, modal, or popover. Added a single "any dismissible surface open" guard to the cancel-job shortcut.
  • Enter no longer hijacks focused controls. The "Enter runs the primary action" shortcut now bails when focus is on a button/link/tab/dropdown/ menuitem, so keyboard activation of those controls works and an unrelated job isn't launched.
  • ExtendScript string safety: migrated the five remaining PremiereBridge call sites (importFiles, removeSequenceMarkers, unrenameItems, removeImportedSequence, removeImportedItem) from the narrow replace(/'/g,"\\'") escape to escSingleQuote(), which also handles newlines and U+2028/U+2029 line separators in clip/marker names.
  • Accessibility: document.documentElement.lang now tracks the active locale so assistive tech uses correct pronunciation after a language switch.
  • Stale per-run time estimate is cleared at job start instead of lingering from the previous run until the new estimate resolves.

Fixed — FFmpeg integration (backend)

  • Styled-caption overlay encoding now drains FFmpeg stderr in a background thread while raw frames are piped to stdin, preventing stderr backpressure from deadlocking long renders.
  • The Real-ESRGAN upscale tier now resolves and caches official model weights before constructing RealESRGANer, instead of passing model_path=None.
  • Caption burn-in failed on every Windows drive-letter path. The subtitles=/ass= filter value dropped colon escaping, so C:/... was parsed as filename C. Added escape_filter_path() with the verified two-level escaping (drive colons, apostrophes, spaces) and applied it. Verified end-to-end against real ffmpeg.
  • Large edits crashed on Windows with "filename or extension is too long". Silence-removal export and speed_up_silences built -filter_complex as a single argv string; past ~220 segments this exceeds the 32 KB command-line limit (OSError WinError 206). Both now spill large graphs to a temp file and use -filter_complex_script. Verified: 300-segment graph (43 KB) fails inline, succeeds via script file.
  • Merge/concat broke on filenames with apostrophes and corrupted non-ASCII names. The concat-demuxer escaping used \' (rejected by ffmpeg) and the default text codec (cp1252 on Windows). Added write_concat_list() (UTF-8 + the correct '\'' idiom). Verified: apostrophe and Cyrillic/CJK names.
  • All core concat-demuxer writers found by the repo-wide scan now use write_concat_list() or the shared concat-line formatter instead of local platform-codec path writes.
  • New shared escaping helpers in opencut/helpers.py (escape_filter_path, escape_drawtext, write_concat_list) with a string-level regression test (tests/test_ffmpeg_escaping.py) locking the ffmpeg-verified behavior for CI.
  • escape_drawtext documents/requires expansion=none: under default drawtext expansion a literal % cannot be escaped at all ("Stray %") and caption text containing %{...} would be interpreted as an expression.
  • Caption styles, click/keystroke overlays, tickers, quiz cards, telemetry overlays, callouts, audiograms, brand watermarks, end screens, and guest lower-thirds now use the shared escape_drawtext() contract with expansion=none.
  • Audio-only export to .aac and .ogg no longer fails — those containers were handed the mp3 encoder; now mapped to aac / libvorbis.
  • Progress-parsing FFmpeg subprocess now decodes stdout/stderr as UTF-8 with replacement, preventing a UnicodeDecodeError in the stderr drain thread from deadlocking the progress loop on non-ASCII output.

Fixed — backend correctness

  • Transcription timeout was a no-op. The ThreadPoolExecutor was used as a context manager, whose exit blocks on shutdown(wait=True) until the worker finishes — so a timed-out transcription still pinned the caller for the full run. Switched to manual non-blocking shutdown with cancel_futures.
  • smart_trim now detects trailing silence (a clip that ends in silence emits an unmatched silence_start), so "trim trailing silence" actually works.
  • Stem recombination uses amix=...:normalize=0; the default normalize=1 scaled each stem by 1/N, making a recombined mix ~12 dB too quiet.
  • loudness_match clamps non-finite (-inf) measurements from silent input (previously produced invalid -Infinity JSON and out-of-range pass-2 args) and anchors on the last loudnorm JSON block.
  • smart_reframe guards against ZeroDivisionError from a zero aspect-ratio component (e.g. "16:0") or a zero source dimension from a probe failure.
  • extract_audio_wav no longer leaks its temp WAV when extraction fails.
  • speed_up_silences now treats MP3/M4A cover art as attached pictures, not real video streams, so audio-only files stay on the audio concat path.
  • Waveform image uses a unique temp name instead of waveform_<pid>.png, which collided across requests in the long-lived server.
  • waveform_timeline uses array.array instead of Python list for PCM samples (~8x less memory for long audio files).
  • smart_trim now surfaces subprocess timeout/errors instead of silently returning an empty result as "no speech detected".
  • Export cancel now calls proc.wait() after killing FFmpeg before unlinking the partial output file, preventing silent failure on Windows.
  • preview_frame checks FFmpeg return code and output file size before returning success.
  • Captions temp WAV cleanup uses _schedule_temp_cleanup when immediate unlink fails from a timed-out Whisper thread on Windows.
  • auth.json now applies a restrictive DACL via icacls on Windows (was POSIX-only chmod).
  • Multiview/repurpose routes now validate secondary file paths (video_paths, content_path, reaction_path) before forwarding to core.
  • Installers prefer requirements.txt/requirements-lock.txt over loose pip specs in fallback paths.
  • CLI loudness-match resolves bare filenames to absolute paths before passing output_dir.
  • CLI auto-zoom uses .get() for width/height to avoid KeyError.
  • Fixed placeholder github.com/opencut URLs in CLI banner and InstallerBuilder.

Fixed — CEP panel

  • Job history dedupe now compares by job_id (or type+createdAt fallback) instead of (type, status).
  • NLP auto-execute raised confidence threshold to 0.8 with a confirm step.
  • body.job-active UI lock now sets inert on main content for keyboard and AT users (was pointer-events-only).
  • showAlert() accepts an explicit tone parameter, bypassing English-only regex inference for non-English locales.
  • Time estimate reads numeric clip duration from state instead of parsing rendered DOM text.
  • Command palette section labels ("Recent", "Favorites", etc.) are now i18n-configurable via sectionLabels on the palette context.
  • Language dropdown reduced to only en (the only shipped locale).

Fixed — UXP panel

  • Await precedence fix: (await getOutPoint?.())?.seconds instead of await getOutPoint?.()?.seconds.
  • addMarkers now awaits getFirstMarkerAtTime so names/colors apply.
  • Enter key handlers guarded with hasActiveJob() to prevent duplicate submissions.
  • 10 handlers changed else to else if (!r.ok) so ok-without-job_id synchronous responses aren't reported as failures.
  • udt-smoke.js gated behind localStorage opencut_debug=1.
  • postMessage targetOrigin restricted from "*" to location.origin.
  • cep_node.exec passthrough blocked in the WebView shim.

Fixed — Installer

  • Inno Setup PlayerDebugMode registry writes use runasoriginaluser via reg.exe so keys land in the invoking user's HKCU, not the admin's.
  • VBS launcher uses WshShell.Environment instead of cmd /c string concatenation, preventing injection from paths containing & or ^.

Changed

  • Background sweep threads (temp cleanup, disk monitor) are skipped during test runs via create_app(testing=True).
  • README: added June 2026 cost comparison table vs CapCut Pro, Submagic, AutoCut, and FireCut.