Changelog

All notable changes to ReelForge are documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Unreleased

1.0.0 - 2026-11-01

First public release, under the Apache License 2.0 (documentation and examples under CC BY 4.0).

Added

• Visual editor (reelforge edit): canvas with zoom/pan, hover and selection, drag and resize with smart guides, nudging and in-place text editing; layers panel with drag-to-reorder and rename; component library with drag-and-drop; schema-driven inspector for every property (scrubbable numbers, theme color swatches, word-anchored times, lists, effects, states, YAML view); multi-track timeline (scenes, script cues, layer tracks, waveform, SFX) with magnetic snapping to spoken words; undo/redo across episode.yaml and cues.json; ⌘K command palette, context menus, a shortcut sheet, resizable panels and a first-run tour. Edits keep the comments and formatting of episode.yaml.
• Project dashboard (reelforge studio): create English or Arabic episodes, upload recordings (WAV, MP3, M4A, OGG, FLAC…), open the editor, validate and render with live progress, play and download.
• Cues mode: mark sentence starts with Space at 0.5×–1× speed, nudge by 50/10 ms, re-mark any line, word mode, snap to onset, undo, autosave; or drag cues on the timeline.
• Automatic cues (reelforge align, Auto-mark from voice in Cues mode, Auto cues on the dashboard, MCP align_cues): every line and word aligned to the recording. Works without any model (pause detection and dynamic programming over the known script, about 0.05 s per minute of voice). Optional speech models (reelforge align --setup downloads them once into the user cache, about 180 MB; alignment runs offline afterwards): Whisper word timings matched to the script by fuzzy DTW, and Silero VAD for speech versus music and noise. Measured on recorded speech with known timings: every line within 100 ms (mean 15 ms), also under background music. Hand-marked cues are never replaced without confirmation; in the editor the change is one undo step.
• Smart insert: suggestions per line in the editor's Insert panel and MCP suggest_scene. A keyword dictionary (English and Arabic) picks the component, fills its props from the line and makes it enter on the word that names it. Optional local LLM (through Ollama, qwen2.5:1.5b by default) with output constrained to the component schemas.
• Review mode: true 1080×1920 preview synced to the audio, hot reload on save, platform safe-zone, stage and element-box overlays, every check with click-to-select, and an error banner that keeps the last good frame.
• Agent integration (optional): reelforge mcp, a Model Context Protocol server (create, read, edit, validate, preview frames as images, align, render) plus instruction files (AGENTS.md, CLAUDE.md, GEMINI.md, …) and ready-made MCP configurations.
• Examples: the 9-episode "System Design" series in English, the same series in Arabic, and a 45-second product intro.
• 21 animated components: title, label, code, node, arrow, chip, stamp, table, lanes, checklist, counter, meter, timeline, grid, bucket, phone, quote, particles, trap, cta, icon.
• Motion: 17 enter and 9 exit animations (including reveal, spring, drop, blur and wipe), per-element easing (smooth, snappy, gentle, linear), six scene transitions (crossfade, cut, slide-up, zoom, push, blur) and slow camera moves (drift, push-in, pull-out). Everything renders with sub-pixel precision, so slow moves stay smooth; frame-by-frame motion tests guard against jumps and stutter. The inspector previews any animation with one click.
• Captions in English and Arabic: word-by-word karaoke highlight, correct RTL/LTR mixing, الـ term kept together, punctuation on the correct side.
• Audio: procedural sound effects, and voice mastering to −14 LUFS with ducking and a limiter.
• Voice noise cleanup: audio.denoise: true, render --denoise or a checkbox in the render dialogs. RNNoise (bundled 290 KB model, FFmpeg arnndn) runs before loudness mastering; about 15 dB less noise between sentences in tests, with the voice level unchanged.
• Rendering: deterministic parallel frame capture; an H.264 master (CRF 18) and a size-capped share file, plus SRT captions, a contact sheet and publish.md. Every output is verified (codec, profile, pixel format, frame rate, AAC 48 kHz, faststart, size).
• Validation: 14 rules with file:line:col locations and fix hints, including V-14 pacing (lines spoken faster than 3.5 words/s, a hook that starts after 3 s, silences of 1.5 s or more).
• Themes and presets in editable YAML; reelforge theme new --accent #hex.
• CLI: studio, new (--lang), edit, cue, dev, align, mcp, validate, render (--all), snap (--cover), doctor, schema, theme new.
• One-command installers for Windows, macOS and Linux.

Fixed

• Renders failed the output check with pixel format yuvj420p on FFmpeg 7+: frames are now converted to limited-range BT.709 yuv420p.
• The dashboard and editor are responsive (side panels become drawers on tablets and phones).