CHANGELOG.md

Changelog

All notable changes to OpenReels will be documented in this file.

[0.18.0] - 2026-04-10

Added

• Creative direction file (--direction <file>): attach a free-form markdown brief describing visual style, script notes, music mood, and scene ideas. The AI reads it like a human editor would, constraining the DirectorScore while preserving creative judgment. Works across CLI, API (direction field), and web UI (textarea in Advanced panel).
• Score replay (--score <path>): reuse a previous score.json to skip research, director, and critic stages. Re-renders with different provider config without regenerating the creative plan. Saves $0.50-2.00 and 10-30 seconds per replay.
• Replay-aware cost estimation: cost estimates during replay accurately omit skipped LLM calls (research, creative director, critic), so users see the true execution cost.
• Direction provenance in log.json: direction text, replay flag, and source score path are recorded in the run log for auditability.
• Example creative brief: examples/direction-brief.md demonstrates the direction file format with a deep-sea exploration theme.

For contributors

• PipelineOptions has new optional fields: direction?: string and replayScore?: DirectorScore.
• estimateCost() accepts an optional { replay?: boolean } options object as the last parameter.
• CreateJobBody (server) and CreateJobRequest (web) support direction and score fields.

[0.17.0] - 2026-04-10

Added

• Dedicated video motion prompter: new prompts/video-prompter.md with structured cinematography language. Video prompts now follow a professional anatomy: shot type, subject, action, camera movement, lighting, and style. Temporal progression, single-action constraints, and physics/realism guidance produce more natural motion.
• Negative prompts for video generation: both Veo and Kling providers now receive anti-artifact guidance. Default negatives (blur, flickering, morphing, unnatural physics) are combined with each archetype's existing antiArtifactGuidance for style-aware quality control.
• Video prompt observability: motionPrompt and negativePrompt fields are now logged in VideoResolution metadata (visible in log.json), making video quality issues debuggable without re-running the pipeline.
• Kling 2.6 Pro upgrade: Fal/Kling video provider upgraded from v2.1 standard to v2.6 Pro. Better motion physics and realism at the same $0.07/s price.
• Kling cfg_scale parameter: prompt adherence control (0.5 default) now passed to the Kling API.

Fixed

• Veo API compatibility: removed personGeneration, enhancePrompt, generateAudio, and negativePrompt config params that veo-3.1-lite-generate-preview rejects. These exist in the SDK types but are only supported on the full Veo model (Vertex AI). This fixes Forbidden/INVALID_ARGUMENT errors that silently caused all Veo video generation to fail.
• Factory videoModel routing: config.videoModel is now passed to GeminiVideo regardless of primary/secondary position. FalVideo always uses its own default model, preventing Veo model strings from reaching the Fal endpoint.

For contributors

• VideoProvider.generate() opts now include optional negativePrompt field.
• VideoResolution interface has new motionPrompt and negativePrompt fields.
• image-prompter.ts loads video-prompter.md when mode="video" instead of appending to image-prompter.md.

[0.16.0] - 2026-04-09

Added

• Universal LLM provider gateway: use any OpenAI-compatible endpoint as an LLM provider. New --provider openrouter and --provider openai-compatible options with --llm-model and --llm-base-url flags. OpenRouter gives access to 300+ models; openai-compatible supports Ollama, Together, Groq, vLLM, and any endpoint speaking the OpenAI API.
• Pluggable search backends: independently configure web search with --search-provider native|tavily|none. Native uses each provider's built-in search (Anthropic, OpenAI, Gemini). Tavily adds web search to providers that lack it. None disables search and uses the model's training knowledge instead.
• Tavily web search integration: @tavily/ai-sdk powers web research for OpenRouter and custom providers. Auto-detected when TAVILY_API_KEY is set, with graceful degradation when absent.
• Parametric knowledge fallback: when no search tools are available, the research agent uses a single-pass structured output call with a "use your training knowledge" prompt instead of wasting tokens on an empty two-pass search loop.
• Tool-calling error recovery: if a model doesn't support tool calling, the pipeline catches the error and falls back to parametric knowledge with a warning instead of crashing.
• Pass 2 retry with cached search: when structuring search results fails (model can't produce valid JSON schema), the pipeline retries Pass 2 up to 2 times reusing cached search results from Pass 1 instead of re-running Tavily and burning search credits.
• Web UI provider expansion: OpenRouter and Custom (OpenAI-compatible) options in the LLM dropdown with conditional model ID and base URL inputs. Search provider dropdown (native/tavily/none).

Changed

• Cost estimation gracefully handles unknown providers (shows $0.00 with a note) instead of crashing.

For contributors

• BaseLLM now accepts optional searchTools via constructor for search tool injection. Existing providers are unaffected.
• Worker key resolution uses a lookup map instead of hardcoded if/else chain.

[0.15.0] - 2026-04-09

Added

• Director-Critic quality gate: the pipeline now evaluates and revises the DirectorScore before any expensive generation begins. A Critic agent scores the plan (1-10) and if below 7, provides revision instructions. The Creative Director revises the plan and re-submits. Up to 2 revision rounds, tracking the highest-scoring version. Revision costs ~$0.02-0.13 in LLM calls vs $0.50-2.00+ for regenerating the full pipeline.
• Revision cost tracking: cost estimates now include a "Revise" line item showing actual revision LLM cost. The CostBreakdown interface includes revisionCost and revisionRounds fields.
• Worker revision history: meta.json now includes a revisionHistory array with per-round critic scores, enabling future UI display of the revision journey.

Changed

• The Creative Director's system prompt loading is now a shared loadDirectorSystemPrompt() helper, used by both initial generation and revision.
• The step-6 critic no longer displays "Revision support coming in a future release" since revision is now active.
• Cost estimation uses the final revised DirectorScore for accurate scene counts when visual types change during revision.
• Dry-run mode now runs after the revision loop, showing the revised plan instead of the first draft.

[0.14.0] - 2026-04-09

Added

• Spring-animated captions: words now smoothly scale up when spoken, with per-style physics (punchy, elegant, snappy, smooth). Captions feel alive instead of toggling on/off.
• Three-state word rendering: each word transitions through unspoken (dim), active (bright, animated), and spoken (settled). The currently-spoken word is always visually distinct.
• BoxHighlight caption style (7th style): a spring-animated background rectangle tracks the active word across the chunk.
• KaraokeSweep gradient wipe: an accent-colored fill sweeps left-to-right behind each word as it's spoken.
• ColorHighlight spotlight effect: only the active word gets a background highlight, creating a bouncing spotlight.
• Visual snapshot script for QA testing all 7 caption styles (npx tsx src/remotion/captions/visual-snapshot.ts).

Changed

• Caption accent colors now use the archetype's colorPalette.accent instead of hardcoded green/red.
• Chunk sizing is archetype-aware: fast archetypes show 6 words, cinematic show 4 words (was fixed at 5).
• Chunk linger time is archetype-aware: fast=150ms, cinematic=500ms (was fixed at 300ms).
• Chunk transitions advance immediately when the next chunk's first word starts speaking, preventing words from being hidden during the linger window.
• Infographic archetype now uses box_highlight caption style.
• Whisper alignment uses character-proportional duration estimation (~60ms/char, 100ms floor) instead of a flat 200ms for missed words.

[0.13.3] - 2026-04-07

Changed

• Consolidated 75+ hardcoded hex colors into 7 semantic CSS variables (surface-inset, surface-sunken, text-subtle, text-faint, status-info, status-success, status-warning) plus existing theme tokens for consistent theming across all web UI components.
• Added shadow token system (shadow-glow-sm/md/lg) for consistent glow effects.
• Added skeleton loaders for Gallery and Job pages, replacing generic spinners with layout-aware loading states.
• Buttons now show tactile hover (brightness-110) and press (scale-0.97) feedback.
• Page headings use font-semibold tracking-tight instead of font-bold for refined typography hierarchy.
• Cards differentiated by visual weight: primary content cards get shadow-sm, inline status cards use lighter borders, failed panel gets destructive glow.
• Unified archetype color maps into single ARCHETYPE_THEME config per archetype (badge classes + accent hex).
• Converted inline styles (letterSpacing, minHeight, gridTemplateColumns) to Tailwind classes.
• Added will-change hints to animated timeline connectors and spinning indicators.

[0.13.2] - 2026-04-07

Fixed

• Gemini structured output compatibility: removed minItems/maxItems array constraints from LLM generation schemas (DirectorScoreRaw, QueryReformSchema) that Gemini's API rejects. Scene count and query count are guided by prompt instructions and validated post-generation.
• Web UI storyboard infinite request loop: removed imgError from useEffect dependency array in StoryboardScene to prevent failed thumbnail loads from triggering an endless retry cycle that could exhaust backend memory.

[0.13.1] - 2026-04-06

Fixed

• Storyboard preview now shows correct thumbnails when stock images/videos fail verification and fall back to AI-generated images. Previously, the URL pointed to the non-existent stock filename, resulting in broken previews.
• Scene type badge updates from "ST"/"SV" to "AI" with a fallback indicator when stock assets fall back to AI generation.
• Video play controls correctly hidden for scenes that fell back from stock video to AI image.
• Eliminated premature 404 requests for scene assets during pipeline execution. Thumbnails now load only after the visuals stage completes.
• Web UI now handles stock_fallback and video_fallback SSE events (previously silently dropped).

Added

• README screenshots: home page, live pipeline, completed job, and gallery views.
• README archetype comparison: same topic rendered in four different visual styles.

[0.13.0] - 2026-04-05

Added

• Documentation site built with Fumadocs (Next.js), deployed to GitHub Pages via static export at openreels.tsensei.dev/docs. 47 MDX pages covering getting started, providers, archetypes, pipeline architecture, configuration, REST API, deployment, and contributing.
• Landing page redirects to /docs for clean entry point.
• Archetype gallery with all 14 visual styles, each with visual parameters, style guides, and sample commands derived from archetype JSON configs at build time.
• Prebuild sync script (docs/scripts/sync-archetypes.ts) copies archetype JSON configs into the docs build with schema validation, ensuring zero drift between pipeline and documentation.
• GitHub Actions workflow (docs.yml) auto-deploys docs on push to main when docs/ or src/config/archetypes/ files change.
• 213 archetype validation tests (required fields, color format, value ranges, sync script output).

Changed

• Converted to pnpm workspace with pnpm-workspace.yaml listing web and docs workspaces. Consolidated web/pnpm-lock.yaml into the root lock file.
• Updated Dockerfile to use workspace-aware install (pnpm install --filter=!docs) so Docker builds exclude the docs workspace.

[0.12.1] - 2026-04-05

Fixed

• Lyria music generation now detects safety-filtered responses that return empty content instead of throwing. Previously, the Google GenAI API could return finishReason: "SAFETY" or promptFeedback.blockReason with empty parts, silently bypassing the sanitize-and-retry logic. The error message now includes the actual rejection reason for debugging.
• isSafetyFilterError now matches BLOCKLIST and PROHIBITED_CONTENT rejection types, ensuring all content-safety related Lyria failures trigger the prompt sanitization retry.

[0.12.0] - 2026-04-05

Added

• Archetype-driven scene pacing tiers. Each of the 14 visual archetypes now has a scenePacing field (fast, moderate, or cinematic) that controls how many scenes the creative director generates and how words are distributed across them.
  • Fast (infographic, bold_illustration, comic_book): 8-12 scenes, 8-12 words per scene, punchy cuts.
  • Moderate (warm_editorial, editorial_caricature, anime_illustration, vintage_snapshot, surreal_dreamscape, gothic_fantasy): 7-10 scenes, 10-16 words per scene.
  • Cinematic (cinematic_documentary, moody_cinematic, studio_realism, warm_narrative, pastoral_watercolor): 5-8 scenes, 15-22 words per scene, deliberate pacing.
• --pacing fast|moderate|cinematic CLI flag to override the archetype's default pacing tier. Explicit pacing always wins, even if it contradicts the archetype.
• Web UI pacing dropdown (Auto Pace / Fast / Moderate / Cinematic) in the controls row alongside style and platform selectors.
• Critic agent now evaluates pacing against tier-specific thresholds instead of a fixed "5-7 scenes" standard. Derives the tier from the score's archetype or from the explicit --pacing override.
• Creative director unit tests (new file: creative-director.test.ts) covering all three pacing injection paths, word budget math, and fallback prompt behavior.

Changed

• DirectorScore schema max scenes increased from 10 to 16. DirectorScoreRaw (LLM-facing schema) now also has .min(3).max(16) bounds to prevent token waste on overshoot.
• Creative director prompt rewritten with a pacing tier table. When no archetype is specified, the LLM receives the full tier lookup table to self-select after choosing an archetype.
• Critic and playbook pacing rubric updated to reference tier-specific ranges instead of hardcoded "5-7 scenes".
• buildDefaultPrompt() fallback no longer hardcodes "4-7 scenes", instead references archetype's recommended pacing.

[0.11.0] - 2026-04-04

Added

• AI-generated background music via Google Lyria 3 Pro (--music-provider lyria). Every video gets a unique score that follows the emotional arc of the narration, with per-scene instrument direction and intensity mapping. No more random track selection.
• An LLM music prompter agent writes detailed Lyria prompts from the video's mood, emotional arc, and scene narratives, with timestamp sections like [0:00 - 0:12] sparse piano, building tension. Intensity: 3/10.
• --music-provider CLI flag: bundled (default, free) or lyria ($0.08/track).
• --provider google now sets everything to Google in one shot: LLM, images, TTS, video, and music. One API key, full pipeline.
• Music generates in parallel with visuals, so AI scoring adds zero extra wait time.
• If Lyria's safety filter rejects a prompt, the pipeline strips intense adjectives and retries. If that fails too, it falls back to bundled tracks silently.
• Music cost shows in both the pre-run estimate and post-run actual cost breakdown.
• Web UI: music provider dropdown and audio preview player in the completed panel.
• Music generation metadata (the Lyria prompt, provider, response data) saved to log.json and meta.json for prompt iteration and debugging.

Fixed

• --provider google now correctly sets TTS to gemini-tts (was defaulting to elevenlabs, requiring an ElevenLabs key).
• Gemini TTS pricing updated from $0 (preview) to $0.02/1K chars (GA pricing).
• Stock footage verification with Gemini now works correctly (GOOGLE_API_KEY passed to Vercel AI SDK which expects GOOGLE_GENERATIVE_AI_API_KEY).
• Lyria temp files cleaned up after copy to prevent /tmp disk fill on servers.

[0.10.0] - 2026-04-04

Added

• Unified TTS alignment layer. Any TTS provider that returns audio now gets word-level timestamps automatically via Whisper forced alignment. Enables karaoke captions and scene duration calculation for providers without native timestamp support.
• Kokoro TTS provider (--tts-provider kokoro). Local, zero-cost text-to-speech using Kokoro 82M. No API key needed. Runs in a subprocess to avoid ONNX runtime conflicts.
• Gemini TTS provider (--tts-provider gemini-tts). Uses your existing Google API key, no additional setup.
• OpenAI TTS provider (--tts-provider openai-tts). Uses your existing OpenAI API key with gpt-4o-mini-tts.
• --provider local convenience flag that sets TTS to Kokoro for fully local voiceover generation.
• --kokoro-voice flag for selecting Kokoro voice presets (default: af_heart).
• Per-provider TTS pricing in cost estimator. Kokoro shows $0.00, Gemini TTS shows free tier, OpenAI TTS shows token-based estimate.

Changed

• TTS cost estimator uses an explicit pricing map (Record<TTSProviderKey, number>) instead of a ternary fallback. Adding a new provider without pricing now fails at compile time.

[0.9.0] - 2026-04-04

Added

• Google Gemini as third LLM provider. Use --provider gemini to run the full pipeline with Gemini for reasoning, research, and creative direction.
• --provider google convenience flag that sets LLM, image, and video providers to Gemini in one shot. One API key for three capabilities.
• Gemini pricing in cost estimator. Pre-run and actual cost estimates now use correct per-provider pricing instead of always defaulting to Anthropic rates.

Changed

• LLM providers refactored with a shared base class. Each provider (Anthropic, OpenAI, Gemini) is now ~25 lines instead of ~105, with shared web search and structured output logic extracted.
• Cost estimates use per-provider pricing. Gemini users see accurate estimates (~30x cheaper than Anthropic) instead of inflated numbers.
• Settings page shows all provider API keys with correct names and labels.

Fixed

• Settings page used wrong env var name GEMINI_API_KEY instead of GOOGLE_API_KEY.
• Settings page was missing OpenAI, Inworld, and Pixabay API key entries.
• Worker verification model key lookup now handles Gemini provider correctly.

[0.8.0] - 2026-04-04

Added

• AI video generation. Scenes marked ai_video get a first-frame AI image, then animate it into a 4-10 second video clip via Google Veo or fal.ai (Kling). Motion is the story... explosions, flowing water, rocket launches come alive.
• Two video providers: Google Veo (--video-provider gemini) and fal.ai Kling (--video-provider fal). Cross-provider fallback tries the secondary provider on failure, then degrades gracefully to a static AI image.
• LLM-generated motion prompts. When generating video, a separate prompt optimization pass produces motion-rich descriptions ("asteroid rotates against star field, debris trail, Earth grows in background") instead of static scene descriptions.
• Per-scene cost breakdown in the confirmation prompt. Each scene now shows its estimated cost (Scene 1: ai_video $0.33, Scene 2: stock $0.00).
• Smart duration matching. Video providers declare their supported durations, and the resolver picks the smallest duration >= the scene's voiceover length. Never loops AI video.
• Video generation telemetry in log.json: timing, provider, success/failure, and fallback chain per scene.
• New CLI flags: --video-provider, --video-model, --no-video.
• Web UI: new "AV" badge for AI video scenes, video asset URLs in scene preview.

Changed

• Creative director prompt is now cost-aware about video. It only suggests ai_video when video providers are available, and recommends 1-3 video scenes per Short where motion adds real value.
• optimizeImagePrompt() now takes an options bag ({ mode, rejectionContext }) instead of a trailing optional parameter. Existing callers updated.
• Cost estimator includes video generation cost in totals and per-scene breakdowns.

Fixed

• Anti-loop guard in score-to-props.ts was ineffective due to mutation-before-read bug. The capped duration is now computed from the original value.
• Video model override (--video-model) no longer leaks to the secondary provider. Only the primary provider receives the user-specified model.
• Web UI SceneCard, scene-assets, and useApi type now handle ai_video visual type (enum completeness fix).
• StockVideoBeat no longer loops AI-generated video clips, preventing visible seam artifacts.

[0.7.0] - 2026-04-03

Added

• Vision-verified stock footage. The pipeline now uses a multimodal LLM to check whether stock footage actually matches what was requested before putting it in the video. No more toy rockets in your Artemis launch video.
• Query reformulation. When stock footage doesn't match, the pipeline automatically rewrites the search query in stock-API-friendly terms (strips proper nouns, uses concrete visual nouns) and retries.
• Multi-provider stock search. Both Pexels and Pixabay are searched when both API keys are configured. Query-first ordering tries the same query on all providers before reformulating.
• AI image fallback with negative examples. When all stock options fail verification, the pipeline falls back to AI image generation, feeding the VLM's rejection reasons into the image prompt for better results.
• Stock verification metadata in log.json. Every stock scene now logs what was tried, what was rejected, why, and whether it fell back to AI.
• New CLI flags: --stock-verify / --no-stock-verify, --stock-confidence, --stock-max-attempts, --verification-model.
• Lazy sequential downloading for stock candidates. Search returns metadata first, assets are downloaded one at a time and verified before downloading the next.
• Cost estimator now shows max additional cost if stock scenes fall back to AI generation.

Changed

• Actual cost report now counts AI images from what was produced on disk (including stock fallbacks), not from what the director originally planned.

For contributors

• StockProvider interface now returns StockCandidate[] (metadata without download) and has a download() method for lazy fetching. Both Pexels and Pixabay providers updated.
• PipelineOptions.stock changed from single StockProvider to StockProvider[] for multi-provider support.
• createProviders() factory constructs both stock providers when both API keys are available.
• Remotion score-to-props mapper detects when stock scenes fell back to AI images and renders them correctly as AIImageBeat instead of StockVideoBeat.

Fixed

• OpenAI structured output schema rejection for the transition field (regression from v0.6.0). Changed .nullish() to .nullable() in both director-score.ts and creative-director.ts.

[0.6.0] - 2026-04-03

Changed

• The pipeline is now a Mastra workflow. Contributors can read the 6-stage flow (research, director, TTS, visuals, assembly, critic) as a declarative .then() chain instead of scrolling through 630 lines of imperative code. Same behavior, much easier to extend.
• LLM providers now use Vercel AI SDK 6 instead of raw Anthropic/OpenAI SDK calls. Structured output, web search, and token tracking all go through a unified generateText() API. Adding new LLM providers is simpler.
• Fixed an OpenAI structured output compatibility issue where the transition field on scenes was rejected by AI SDK 6's stricter schema validation.
• Pipeline utility functions (splitWordsIntoScenes, getVideoDuration, etc.) are now importable from src/pipeline/utils.ts for use outside the orchestrator.

Added

• Mastra workflow framework (@mastra/core) for pipeline orchestration.
• Vercel AI SDK 6 (ai, @ai-sdk/anthropic, @ai-sdk/openai) for unified LLM provider interface.

[0.5.0] - 2026-04-02

Added

• Background music for every generated video. 25 royalty-free tracks across 8 moods (epic cinematic, tense electronic, chill lofi, uplifting pop, mysterious ambient, warm acoustic, dark cinematic, dreamy ethereal), automatically selected from the creative director's mood tag.
• --no-music CLI flag and noMusic API option to generate videos without background music.
• Music manifest integrity check validates all referenced MP3 files exist on startup.
• Music selection progress event in SSE stream shows which track was picked and why.
• Selected music track info written to job meta.json for provenance.

Changed

• MusicMood is now a strict 8-value enum instead of a free string. Creative director prompt updated with exact valid values to prevent LLM hallucination and retry costs.
• MusicTrack.tsx simplified to flat volume (0.15) instead of per-word ducking, which caused audible volume pumping during continuous voiceover.

[0.4.0] - 2026-04-02

Added

• Web UI with live pipeline visualization. Watch AI research, write, speak, and paint in real time across 6 stages. React + Tailwind frontend served from the same Docker image.
• REST API (Fastify) for programmatic job management. Create, list, monitor, cancel, and delete render jobs via /api/v1/ endpoints.
• Server-Sent Events (SSE) streaming for real-time job progress. Each pipeline stage emits typed events as it runs.
• BullMQ job queue backed by Redis. Jobs survive restarts, support cancellation between stages, and auto-prune when MAX_JOBS is set.
• Background worker process runs rendering jobs independently from the CLI, with live progress callbacks.
• Bring-your-own-key (BYOK) support. Submit your own API keys per job without storing them server-side.
• Job gallery page with archetype color palette (14 distinct colors) and time-ago labels.
• Settings page for API key management (stored in browser localStorage).
• Download generated videos, images, and audio directly from the web UI.
• Health check endpoint (/api/v1/health) with Redis and filesystem status.
• Progressive reveal during rendering. Research summary, director score, and critic review appear in real time as each stage completes, so you can watch the AI think.
• Reconnecting clients recover full pipeline state automatically, no missed events.
• Three dedicated job views: running (progressive data reveal), completed (video player, quality review, cost breakdown), and failed (error details with the exact stage that broke).
• Scene cards with visual type badges show what's being generated (AI image, stock photo, stock video, or text card).

Changed

• Docker Compose upgraded from single-service CLI to 3-service architecture (Redis + API + Worker) with shared job volume.
• Dockerfile builds the React frontend during image creation and serves it via Fastify.
• Pipeline timeline shows running substatus (e.g., "Rendering 450 frames") and connector lines between stages.

For contributors

• Pipeline orchestrator uses a PipelineCallbacks interface with onProgress events, decoupling progress from terminal output.
• CLI entry point uses the shared provider factory and callback system.
• Imports alphabetized across all source files.

[0.3.0] - 2026-04-01

Added

• Docker containerization. Run docker run --env-file .env --shm-size=2gb -v ./output:/output openreels "your topic" on any platform with Docker installed. No Node.js, Chrome, or ffmpeg setup required.
• --yes / -y CLI flag to auto-confirm cost estimation prompt for non-interactive environments (Docker, CI).
• Startup API key validation that checks required keys based on selected providers before any API calls, with clear error table showing missing keys and signup URLs.
• --preview flag now detects non-interactive terminals and skips gracefully instead of hanging.
• GitHub Actions workflow for building and pushing Docker images to GHCR on release.
• docker-compose.yml with pre-configured shared memory, output volume, and stock asset cache.

Changed

• .env.example now includes provider signup URLs and Docker usage instructions, so you know exactly where to get each key.

[0.2.0] - 2026-04-01

Added

• Scene transitions between beats. Videos now use crossfade, slide, wipe, and flip transitions instead of hard cuts. The Creative Director selects transitions per scene based on emotional content, with each archetype providing sensible defaults.
• Remotion TransitionSeries replaces raw Sequence assembly, with overlap-aware duration calculation that prevents voiceover truncation.
• Full test suite for the score-to-props mapper (12 new tests covering transition cascade, duration calculation, and voiceover clamp).

Changed

• DirectorScore schema extended with optional transition field per scene (backward-compatible).
• All 14 archetype configs now include defaultTransition and transitionDurationFrames.
• getTotalDurationInFrames accounts for transition overlaps and extends the last scene to prevent black frames.

[0.1.1] - 2026-04-01

Added

• Inworld TTS as an alternative voice provider (--tts-provider inworld). Native word-level timestamps, MP3 output, 15-language support. Includes response validation and a 2000-character input guard.
• Provider key type aliases (LLMProviderKey, TTSProviderKey, ImageProviderKey) for DRY type usage across CLI, pipeline, and cost estimation.
• Cost estimation support for Inworld TTS pricing.

[0.1.0] - 2026-03-31

Added

• Give it a topic, get a fully rendered YouTube Short. Research, script, voiceover, visuals, captions, and assembly in one command.
• 14 visual archetypes covering factual, narrative, stylized, dramatic, and aesthetic content: editorial-caricature, warm-narrative, studio-realism, infographic, anime-illustration, pastoral-watercolor, comic-book, gothic-fantasy, vintage-snapshot, surreal-dreamscape, warm-editorial, cinematic-documentary, moody-cinematic, bold-illustration.
• 6 animated caption styles: bold-outline, color-highlight, clean, karaoke-sweep, gradient-rise, block-impact. Each renders word-level synchronized captions from TTS timestamps.
• DirectorScore system: the LLM generates a per-scene production plan with variety enforcement (no 3+ consecutive same visual type) and automatic retry on validation failure.
• 5 AI agents working together: research (web-grounded facts), creative director (script + scene planning), image prompter (matches visuals to archetype style), critic (scores output against a rubric), plus a content playbook for style guidance.
• Dual LLM provider support: Anthropic Claude (with native web search) and OpenAI GPT-4.1 (with two-pass web search via gpt-4o-search-preview).
• ElevenLabs TTS with word-level timestamps and character-to-word aggregation for precise caption sync.
• Gemini Flash image generation with archetype style bible context.
• Pexels and Pixabay stock providers with local caching and portrait-orientation preference.
• Remotion-based video renderer with scene beats (AI image, stock image, stock video, text card), music ducking, and configurable platform output (YouTube Shorts default).
• CLI with --archetype override, --provider selection, --dry-run mode, --preview for Remotion Studio, and real-time progress display.
• Cost estimation (pre-run) and actual cost reporting (post-run) for LLM, TTS, and image generation.
• Vitest test suite covering archetype registry, DirectorScore schema validation, caption timing utilities, playbook parsing, and cost estimation.