Architecture

Last reviewed: 2026-10-10 · Last updated: 2026-10-10 · Change: trivial (status line, revision history and roadmap link added; content checked against the crate list) · Target: Adobe Premiere Pro 2026 (26.5.2)

Kinema is a Cargo workspace of small crates with strictly enforced layering. The engine is headless: every feature can be reached without a window, and the egui UI is one client among the CLI, the JSON control channel and the MCP server.

Design principles:

1. Engine-first. Project-changing actions go through Session::execute(id, params).
2. Everything is a command. Stable id, label, menu path, shortcut, parameter doc, enabled() predicate with a human-readable reason, and run().
3. Exact time. Integer ticks, rational frame rates. No f64 seconds in edit math.
4. Copy-on-write snapshots. The project is an Arc<Project>. Undo is a stack of snapshots, and background readers (playback, export) hold a snapshot without locking.
5. CPU reference, GPU fast path. The CPU compositor is the oracle. The GPU path is tested against it.
6. Pure Rust, clean-room. Codecs and containers are written from public specifications (see AGENTS.md).

1. Layers

 L6  apps/kinema · apps/kinema-cli · apps/kinema-web
 L5  ui-egui · automation · platform
 L4  engine
 L3  render · gpu · export · golden (test-only)
 L2  edit · codecs · interchange · captions · speech · tts · tts-text
 L1  frame · media · project · audio-dsp · text
 L0  foundation: time · geom · color · bitstream · testkit (dev-dependency only)
     codecs/containers: isobmff · matroska · mxf · cfb · mpegts · ogg · h264 · h264enc · hevc · vp9 · av1 · mpeg2v · prores · dnx · apv · dng · aac · ac3 · mpa · opus

Crates are named kinema-<dir> (crates/time is kinema-time). The apps are kinema and kinema-cli.

What cargo xtask layers enforces

The table of layers lives in xtask/src/main.rs (LAYERS). It also reserves names for planned crates. The check reads cargo metadata and looks at normal and build dependencies (dev-dependencies are exempt):

cargo xtask wasm runs cargo check --target wasm32-unknown-unknown on every L0–L4 crate, the egui UI and the web app, so everything up to the engine stays web-portable and the web app builds (web.md). unsafe_code = "forbid" applies workspace-wide, except in platform (deny, allowed only on its FFI modules; ADR 0001).

2. Time base

All time is kinema_time::Tick(i64) at TICKS_PER_SECOND = 254_016_000_000.

That number divides evenly into the frame duration of every broadcast rate (23.976, 24, 25, 29.97, 30, 48, 50, 59.94, 60, 120…) and the sample duration of every common audio rate (8 kHz to 192 kHz, including the 44.1 kHz family). So frame and sample positions are exact integers, edits never drift at 29.97, and audio and video line up to the sample.

Commands take time as time (ticks), frame, seconds or timecode; the engine converts once at the boundary.

3. Data model (kinema-project)

Project
├─ root: Bin                         tree of bins
├─ items: map ItemId → ProjectItem   flat
│    kind: Media(MediaClip) | Sequence(Sequence) | Subclip{..} | AdjustmentLayer{..} | Graphic{..}
└─ next_id

Sequence
├─ settings: frame rate, size, sample rate, …
├─ video_tracks / audio_tracks: Vec<Track>
├─ markers, mark_in / mark_out
Track
├─ locked, sync lock, targeting, mute/solo/visibility
├─ items: Vec<TrackItem>             sorted, never overlapping
├─ transitions: Vec<Transition>
└─ audio: volume_db, pan, effects (mixer inserts), mixer: MixerStrip
     (automation mode + lanes, sends, output, record arm, solo safe, input map)
Sequence (audio) ─ submix_tracks: Vec<Track>, master_volume_db / master_effects / master_mixer
TrackItem (a clip instance)
├─ item: ItemId, start (timeline ticks), source_in (media ticks), duration, speed
├─ link group, label, enabled
└─ effects: Vec<EffectInstance>      intrinsic Motion/Opacity/Volume… first, then standard effects
EffectInstance
├─ effect id, enabled, params: id → constant value or keyframe track
└─ masks: Vec<Mask>                  path (keyframable Bézier), feather, opacity, expansion, inverted, mode

• Graphic clips (graphics.md) reference a Graphic canvas item; their text and shape layers are hidden graphic_text / graphic_shape effect instances on the track item.
• Everything is plain serde data. Sequence::check() validates the invariants (no overlaps, unique ids), and the engine runs it after every sequence edit.
• Timeline positions are sequence ticks; source_in and keyframes are in media time, so trims and splits never move keyframes.
• Effect definitions are data. project::effect::effect_defs() lists every video effect, audio effect and transition with its parameter schema. The Effects panel tree, the Effect Controls rows and the parameter docs agents see are all generated from it.
• Project files (.fcproj) are the project serialised as JSON. Saves are atomic: write a temporary sibling file, then rename it over the target.

4. Command system (kinema-engine)

pub struct CommandSpec {
    pub id: &'static str,                 // "sequence.addEdit"
    pub label: &'static str,              // "Add Edit"
    pub menu: &'static [&'static str],    // ["Sequence"]; empty = not in menus
    pub shortcut: Option<&'static str>,   // "Cmd+K" (Cmd = ⌘ on macOS, Ctrl elsewhere)
    pub params: &'static str,             // r#"{"time":ticks?}"#, shown to agents
    pub enabled: fn(&Session) -> Result<(), String>,  // Err carries the reason
    pub run: fn(&mut Session, &Value) -> Result<Value>,
    pub journal: bool,                    // false for read-only queries
}

• All commands are in crates/engine/src/commands.rs (cmd! for actions, query! for read-only queries such as project.inspect, sequence.inspect, effects.list, jobs.list). Ids follow the menu structure: file.*, edit.*, clip.*, sequence.*, markers.*, timeline.*, effects.*…
• Session::execute(id, params) finds the spec, checks enabled, runs it and appends it to the journal.
• Undo. Edits go through Session::edit(label, |project, state| …) or Session::edit_sequence(label, |seq, ctx, state| …). These clone the project, apply the closure, and on success push the old Arc<Project> with the label onto the undo stack (200 entries). On error nothing changes. edit.undo and edit.redo swap snapshots. Thanks to structural sharing a snapshot costs little.
• Editor state (EditorState: active sequence, playheads, selection, targeting, edit points…) is serde, so agents can read it with state.inspect.
• Events (ProjectChanged, Toast, OpenSequence, OpenSource) are drained by frontends each frame.
• Event log (Session::log, engine::panels): every failed top-level command (an error; a disabled one is a warning), messages and error toasts, auto-save errors, and background jobs starting, finishing, failing or being cancelled. Repeats of the newest entry bump its count. The Events panel shows it; events.list {level?, since?} and events.clear reach it headless.
• UI-only commands (tools, playback, zoom, panels, workspaces) live in crates/ui-egui/src/menus.rs (UI_COMMANDS). The menu bar is built from the engine registry plus this table, and menus::invoke is the single entry point for menus, shortcuts and the control channel.

5. Media, render and playback pipeline

file ──► codecs (MP4/MOV, MKV, audio)        demux + decode, GOP-aware seek
          │   decoder registry: h264, hevc, vp9, av1, prores, dnx, mjpeg (+ any registered first:
          │   platform's VideoToolbox (macOS) or Media Foundation (Windows) H.264 / HEVC, falling back to ours)
          ▼
        media::MediaSource ──► frame cache (byte-budgeted LRU, shared)
          ▼
        render::render_sequence(project, seq, t, scale)        CPU reference
          per track bottom→top: map timeline t → media t (speed), fetch frame,
          standard effects → Motion → Opacity/blend, transitions, composite
          in linear-light premultiplied f32
          │
          └─ render::plan::plan_frame → gpu::GpuCompositor     GPU path
               layers = decoded YUV/RGBA frames + matrix + opacity + blend mode;
               anything the shaders don't cover is pre-rendered on the CPU
          ▼
        ui-egui frames.rs worker pool ──► monitors (program/source), thumbnails, prefetch

• Sources. media::MediaSource yields video_frame(FrameRequest) and audio(start, frames, rate) in media time. Sources are Send + Sync and shared by monitors, thumbnails, playback and export. The engine's MediaPool creates one per project item, lazily, through registered openers (codecs::openers(): MP4/MOV, MKV/WebM, audio files).
• Pixel aspect ratio. Demuxers report it in VideoStreamInfo::par (MP4/MOV pasp, Matroska display size, MXF aspect ratio, MPEG sequence header / VUI); Interpret Footage's Interpretation::par overrides it (clip.interpretFootage {pixelAspect}), and SequenceSettings::par is the sequence's. Both go through project::checked_par (a zero term or more than 8:1 either way = square). render::motion_matrix works in square display units, so a 1440 x 1080 clip with 4:3 pixels fills a 1920 x 1080 square-pixel sequence at 100 % (Scale to Frame Size, Set to Frame Size and Fit / Fill use Project::conformed_source_size); graphics and adjustment layers are in sequence pixels. New Sequence From Clip copies the clip's ratio, and the Source and Program monitors draw the picture at its display aspect. The renderer does not read decoded frames' own VideoFrame::par.
• Offline media and proxies. The pool caches each item's source per reference (path, offline flag), so relinking and undo take effect at once. Media that can't be opened renders the offline slate (render::offline) instead of failing. With proxies enabled, an item with a proxy reads it through ProxySource, which reports the original's size. Export always uses MediaPool::full_res_provider. See project-files.md.
• Seeking. codecs::Mp4Source seeks to the preceding sync sample and decodes forward, caching every frame of the GOP. Sequential playback reuses the decoder. Decoders implement codecs::VideoDecoder. register_video_decoder puts a factory in front of the built-in ones, so a hardware decoder can take precedence. The GOP cache never holds its lock while decoding.
• Hardware decoding. The apps (desktop, CLI / headless MCP, bench) call kinema_platform::register() at startup, which on macOS registers a VideoToolbox factory for avcC / hvcC streams (8 / 10-bit, 4:2:0 and 4:2:2; elsewhere it does nothing), and on Windows a Media Foundation one (below). The factory declines (our decoder is used) when Settings ▸ Playback ▸ Hardware decoding is Off (codecs::hw::set_hardware_decoding, applied by the engine whenever preferences change), for formats it does not take, and when the OS cannot create a hardware session for the stream. Its decoder is a platform::HybridDecoder: VideoToolbox decodes asynchronously (two access units in flight), a reorder buffer of the stream's own depth (max_num_reorder_frames / sps_max_num_reorder_pics, from codecs::hw::NalStreamInfo) restores presentation order, and the decoded biplanar CVPixelBuffer is copied into planar Yuv8 / Yuv16. Colour, pixel aspect, random-access and disposable answers come from the same helpers as the software decoders (video::vui_color, sar_par, h264_disposable, hevc_disposable), so the two are interchangeable (bit-exact in the parity tests). If the hardware fails mid-stream (decode error, session lost to a GPU change or sleep, changed in-band parameter sets), the hybrid builds codecs::software_video_decoder (built-in factories only), replays the samples since the last restart point and continues in software for that instance; it is logged and counted. A source already open keeps its decoder when the setting changes, until it is reopened. perf.stats decode.hardware reports hardware vs software frames, sessions, declines and fallbacks, and the registered backend. On Windows the decoder (platform::media_foundation) is a Direct3D-aware decoder MFT driven at the level of single access units (no Source Reader: the demuxer above already delivers the samples): the avcC / hvcC sample becomes Annex B, the MFT is given the process's Direct3D 11 video device through an IMFDXGIDeviceManager so it decodes with DXVA, and its NV12 (8-bit) / P010 (10-bit) texture is read back through a staging texture (the one GPU to CPU copy) into planar Yuv8 / Yuv16. H.264 Baseline / Main / High, HEVC Main / Main 10, VP9 profiles 0 / 2 and AV1 main (8- and 10-bit), 4:2:0, progressive (VP9 / AV1 samples go in as they are, not as Annex B); everything else, and any stream the GPU's DXVA decoder does not list, is declined. A decoder that would hand back system-memory pictures (Microsoft's decoder MFTs do that when DXVA is not available) fails the stream, so the hybrid continues with our decoder and Windows' software decoding is never used in its place. Zero-copy (the texture straight to wgpu) is not done yet. Decoders run slices on rayon, so an export worker waiting inside a decode can pick up another frame of the same source. A request that finds the shared decoder busy decodes with a private decoder.
• Hardware encoding. On Windows and 64-bit Linux, platform::nvenc::export::factory is registered with kinema_export::register_encoder (register() does this once), in front of the software H.264 encoder. It takes an export only when Export ▸ Hardware encoding (ExportSettings.hardwareEncoding, HardwareEncoding::Auto) is Auto, which is off by default: hardware streams differ from ours, and exports are otherwise byte-identical from run to run. NVENC (platform::nvenc, the driver's nvEncodeAPI64.dll / libnvidia-encode.so.1 loaded at run time) then takes RGBA frames, converted to 4:2:0 by the software encoder's own conversion, and produces the H.264 samples and avcC of an MP4 / MOV. It declines (the software encoder runs, counted in export.hardware.declined) two-pass VBR, HDR, MXF, interlaced output, sizes outside NVENC's limits, and systems without an NVIDIA GPU or driver. A failure during an export ends it with an error: the software encoder cannot take over a hardware stream. The same NVENC session also encodes H.265 (HEVC Main 8-bit 4:2:0 SDR, or Main 10 HDR: PQ / HLG, BT.2020, limited range, from the float R'G'B' pictures through rgbf_to_yuv420_10 into P010 input buffers, with the HDR10 static metadata as SEI on every IDR for PQ; the hvcC is built from the SPS the encoder wrote). An HDR sequence exports as HDR H.265 only where a registered probe says so (kinema_export::register_hdr_probe, hdr_available; NVENC with 10-bit support), otherwise it is tone-mapped to SDR as on macOS: register() hands platform::nvenc::hevc_available to kinema_export::register_format_probe, so the H.265 format exists where an HEVC encoder does, and choosing it is the opt-in (the Hardware encoding toggle governs H.264 only). With no software H.265 encoder, what NVENC declines is counted and ends the export with an error that says why.
• Compositor. render is the reference for monitors, thumbnails and export. render::plan turns a frame into GPU layers, each with its opacity and blend mode. All 27 blend modes run on the GPU (kinema-gpu): Normal and Dissolve with fixed-function "over" blending, the others by copying the accumulator under the layer into a backdrop texture and compositing in the fragment shader with the CPU reference's formulas. The common standard effects run on the GPU too (render::gpufx, gpu::fx): Brightness & Contrast, ProcAmp, Tint, Black & White, Color Balance, Leave Color, Change to Color, Color Pass, Color Replace, Channel Mix, ASC CDL, Gamma Correction, Levels, Extract, Invert, Posterize, Alpha Adjust, Gaussian Blur and Directional Blur (and their legacy aliases), Camera Blur, Sharpen, Unsharp Mask, Crop, Edge Feather, Transform, Horizontal / Vertical Flip, Mirror, Offset, Vignette and Video Limiter. When every enabled effect of a media clip is in that set (unmasked, with finite parameters, and no Transform shrinking the picture below half size, which the CPU pre-filters), the plan hands the GPU the clip's source with the effects' parameters evaluated at that time: the source is drawn into an Rgba32Float working image at the size the CPU decodes it, each effect runs as compute passes with the CPU reference's math (effects, then Motion, then Opacity / blend, as on the CPU), and the result is placed like any layer. Other standard effects, effect and opacity masks, adjustment layers, nested sequences (except a plain one, whose own layers go into the plan: same frame size and colour settings as its parent, nothing on the clip that changes the picture, every layer inside blended normally) and non-dissolve transitions are rendered on the CPU for that layer or frame and handed to the GPU as an image (a layer image keeps its clip's blend mode), so both paths give the same picture. Setting KINEMA_CPU_COMPOSITE=1 forces the CPU path in the desktop app.
• CPU compositor shortcuts. The CPU compositor, which is what export runs, works on premultiplied linear f32 images, 33 MB at 1080p, so what it does not allocate, convert or mix is time saved. Three shortcuts, each of which gives the general path's bits (compared bit for bit, with every blend mode, in render/src/region_tests.rs): the canvas stays unallocated until a layer needs it, and an opaque bottom layer (an 8-bit or 16-bit Y'CbCr frame without an alpha plane, opacity 100 %, Normal) simply becomes the canvas; a plain media frame with an alpha plane (a ProRes 4444 banner) is converted and mixed only inside the rectangle where its alpha is not zero (VideoFrame::alpha_region, to_linear_f32_region, blend::composite_at), the rest being transparent anyway; and working images are recycled through kinema_frame::pool (take_f32_overwritten / recycle_f32: the buffer comes back with its old contents, so there is no zero-fill) by the export pipeline once a frame is converted to 8 bits; a standalone export's pipeline frees the idle float images (pool::trim_f32) when it goes away, so a finished export leaves none behind (parts of a batch, such as render-preview segments, keep them for the next part). A clip with effects, opacity masks, frame blending, colour management, or a picture that is not placed one to one on the output takes the general path.
• Frame scheduling. crates/ui-egui/src/frames.rs runs a small pool of worker threads with prioritised jobs: the frame on screen first, then playback prefetch, then thumbnails. The UI never decodes. It shows the exact frame when it is ready and holds the nearest cached frame meanwhile. Play waits for the first frames (PREROLL_FRAMES, at most 0.5 s) before starting the clock. Every refresh, schedule_playback asks for the next frames and drops (or cancels, through kinema_media::cancel) jobs for frames the playhead has passed; a new on-screen frame replaces the one asked for before (scrubbing); Stop cancels the prefetch. When frames cost more than the workers can render in real time (CPU effects), it starts only frames that can still be on time and spaces them evenly (playback_plan). When decoding is what falls behind, each job tells its sources which frames are already late (kinema_media::cancel::with_catch_up: the playhead's distance, two seconds for a scrubbed-to frame), and the GOP cache skips non-reference pictures of late frames on the way (VideoDecoder::is_disposable, H.264 nal_ref_idc 0 and HEVC sub-layer non-reference pictures; the wanted frame decodes exactly as before). With Settings ▸ Playback ▸ Draft decoding (off by default), frames played at 1/2 resolution or lower are requested as draft frames (FrameKey::draft, kinema_media::cancel::with_draft): the H.264 decoder skips deblocking of non-reference pictures (VideoDecoder::set_draft), the GOP cache serves those draft frames to draft requests only (a paused frame, render or export decodes them again exactly), and draft plans hand the GPU box-decimated Y'CbCr planes at the drawn size. perf.stats reports the counters. GPU plans carry their texels already converted for upload (kinema_gpu::prepare), so the UI thread only copies them.
• Audio clock. The desktop app passes a cpal output (apps/kinema/src/audio.rs) to the UI as AudioOut. While playing, the samples played by the sound card drive the playhead and video follows. Without an audio device, playback falls back to the wall clock. PlaybackMeter counts timeline frames: shown when the exact picture was on screen while due, dropped otherwise (including frames passed over without a refresh, not while the window is hidden). cargo xtask bench-playback measures the whole path headlessly (testing.md §5). Sequence audio goes through the mixer graph (render::mixer, §5.1); clip audio effects run on audio-dsp via render::audio_fx.

5.1 Audio mixer

clip: gain → clip effects → Volume / Channel Volume / Panner (clip keyframes, media time)
      → audio transitions → summed per track                           render::audio::track_input
track / submix strip:  input map, mono fold → pre-fader inserts → pre-fader sends → mute
      → fader (volume) → meter → post-fader inserts → post-fader sends → pan / balance → output
Mix:  bus sum → pre-fader inserts → fader → meter → post-fader inserts → out    render::mixer

• Model (project::mixer). Every audio track, submix and the Mix has a MixerStrip. Static values stay in Track::volume_db, pan, muted and the send/effect parameters; automation is keyframes in sequence ticks: lanes volume, pan, mute (hold), send.<i>.level in MixerStrip::lanes, and insert parameters (fx.<slot>.<param>) in the effect's own keyframes. Up to 5 inserts (EffectInstance::post_fader picks the side) and 5 sends per strip. Submixes feed the Mix or a submix after them (no feedback). All fields have serde defaults, so older projects load unchanged.
• Graph (render::mixer::mix_graph). Lanes are evaluated per sample; effect parameters update on an absolute 64-sample grid, so the output does not depend on how callers cut the timeline into requests (export batches and device callbacks give identical samples). Inserts that report latency delay their strip; each route into a bus gets a compensation delay and the graph is read ahead by its total latency. Graph state (DSP, delay lines) is cached per structure and continued by sequential readers; other requests start fresh with a pre-roll (effect tails, ≤ 3 s). Tracks run in parallel (rayon), buses in order. Mono tracks pan with the −3 dB constant-power law; stereo tracks and sends use balance. Solo keeps soloed and solo-safe strips plus everything feeding them or fed by them. 24 tracks × (EQ + Dynamics + Studio Reverb) + a compressed submix renders at ~6× realtime on one core (release).
• Automation modes (Premiere semantics). Off ignores lanes; Read plays them; Latch records from the first touch and holds the last value until playback stops; Touch records while held and ramps back to the existing automation over the automatch time (Preferences ▸ Audio, 1 s); Write records every control from playback start (then switches to Touch unless "Switch to Touch after Write" is off).
• Recording (engine::mixer). Playback start runs mixer.recordStart, stop runs mixer.recordStop. Fader and knob drags send mixer.touch (value, playhead) while held and mixer.release when let go. Held values go to render::mixer::LiveMix, which the playing mix reads, so moves are heard at once. At stop each gesture stream is thinned (linear keyframe thinning, optional minimum time interval) and written over its time range as one undo step, with boundary keyframes that keep the automation outside the range unchanged.
• Live state. PreviewStore::live (LiveMix) also carries per-strip meter peaks posted by the mix (one per channel; Track Mixer and Audio Meters read them), the voice-over input level (Meter Input(s) Only) and the newest project snapshot, which the audio callback uses, so edits made during playback are heard.
• Audio Clip Mixer automation. Each audio track has a Clip Mixer mode (clipMixer.setMode, session state EditorState::clip_mixer_modes, Read by default). During a pass the Clip Mixer's fader and pan knob are gestures on the track's clip.volume / clip.pan lanes (clipMixer.touch / clipMixer.release): the clip under the playhead plays the held value at once (render::audio::track_input_live), and Latch / Touch / Write follow the track rules. At mixer.recordStop each stream is thinned and written as clip keyframes (Volume level, Panner balance, in media time) into every clip it covers, with boundary keyframes that keep the clip's values outside the range; one undo step for the whole pass.
• Track Mixer view. The panel menu (mixer.menu) has Show/Hide Tracks (UiState::mixer_hidden) and Meter Input(s) Only (UiState::mixer_meter_input_only: record-armed tracks meter the voice-over input). The transport's record button runs voiceover.recordToggle.

5.1.1 Multichannel and 5.1

5.1 track (6 ch: L R C LFE Ls Rs) ─┐                         ┌─► 5.1 Mix ─► 6-ch export / 5.1 device
stereo / mono track ──5.1 panner───┼─► 5.1 submix / 5.1 Mix ─┤
5.1 strip ──BS.775 downmix─────────┴─► stereo bus            └─► stereo device: 5.1 Mixdown Type fold

• Model. Track::channels (Standard = Stereo, Mono, 5.1, Adaptive) and SequenceSettings::audio_master (the Mix: Stereo, Mono, 5.1, Adaptive) existed already; file.newSequence {mix, trackType} and sequence.settings {mix} set them, mixer.setStrip {channels} / mixer.addSubmix {channels} per strip. The 5.1 panner is four lanes in MixerStrip::lanes (pan51.x, pan51.y −100…100, pan51.center 0…100 %, pan51.lfe dB): a lane without keyframes stores the static value, so no new project fields were needed and the puck is automatable like any lane.
• Maths (audio_dsp::channels). 5.1 order is L, R, C, LFE, Ls, Rs everywhere inside Kinema. Downmix is ITU-R BS.775 (Lo = L + k·C + k·Ls, Ro = R + k·C + k·Rs, mono M = k·L + k·R + C + ½·Ls + ½·Rs, k = 1/√2, LFE omitted); upmix puts mono on C and stereo on L/R. The panner splits a point source front/rear and left/right with the sine/cosine constant-power law and gives the centre speaker center·(1 − |x|) of the front power (Σg² = 1). Stereo sources sit at the puck ± 1 in x; 5.1 sources keep their layout moved by the puck's offset from front-centre (the default puck is the identity). LFE gets the source LFE (5.1) or the mean of the source channels (mono/stereo, default −∞) times the LFE level.
• Graph. Buses are 2 or 6 channels wide (render::mixer::width_of); inserts are built for the strip's width; routes convert between widths (5.1 panner into 5.1 buses, BS.775 into stereo buses); meters have one value per channel. mix_graph returns the Mix's width; audio::mix_sequence always returns stereo, mix_sequence_layout any layout and mixdown. Clips on 5.1 tracks play the source's first six channels (or the six picked with Modify ▸ Audio Channels), mono/stereo sources upmixed; clip effects run on all six channels.
• Playback. On a device with ≥ 6 output channels a 5.1 Mix plays as six channels; on a stereo device it is folded with Preferences ▸ Audio ▸ 5.1 Mixdown Type (Front Only, Front + Rear Surround = BS.775, Front + LFE, Front + Rear Surround + LFE; PreviewStore::mix_layout).
• Export. Audio Channels Mono / Stereo / 5.1: WAV writes WAVE_FORMAT_EXTENSIBLE with channel mask 0x3F; QuickTime PCM (ProRes, DNxHR, MJPEG) adds a chan atom (kAudioChannelLayoutTag_MPEG_5_1_A); AAC uses channel configuration 6 (our encoder already supported 1–8 channels; the exporter reorders to AAC's C, L, R, Ls, Rs, LFE). A stereo export of a 5.1 Mix is the BS.775 downmix.

5.1.2 Voice-over recording

Voice-over recording (engine::voiceover, audio.voiceover.*). The record point R is the playhead, or the In point when In/Out are set (punch-in; punch-out at Out). Playback starts the pre-roll before it (C = max(0, R − pre-roll)) and the input is captured from C. When the UI's audio clock really starts, audio.voiceover.sync restarts the capture there. audio.voiceover.stop keeps the audio from R to min(stop, Out), writes a mono 32-bit float WAV <Name> <n>.wav (Scratch Disks ▸ Captured, else next to the project, else the data or temporary directory), imports it and overwrites it onto the record track at R as one undo step. The record track is the given one, else the record-armed one, else the first targeted one. Input goes through the AudioInput trait: cpal in the desktop app (apps/kinema/src/audio_in.rs, on its own thread), and SyntheticInput headless, which produces exactly the samples the timeline asks for so recordings land sample-accurately. Voice-Over Record Settings (Source, Input channel, Name, Countdown Sound Cues, Pre-/Post-roll) are preferences (voiceOver). The track header's microphone records or stops, and a right-click opens the settings dialog. Countdown beeps (1 kHz, 100 ms, each whole second of pre-roll and at R) are mixed into playback, an overlay counts down, and playback stops at Out + post-roll. A take records anywhere, also on the empty timeline past the end of the sequence: starting it never returns the playhead to the beginning, and playback keeps going past the end until the take is stopped.

5.1.2a Text to Speech (narrations)

Narrations (engine::narration, tts.*; voices in kinema-tts). The Text to Speech panel (Window ▸ Text to Speech) writes a script with pause markers ([pause 1s]), a voice, a vocal pitch and a pace; tts.create synthesizes it, writes Narration <n>.wav (mono 32-bit float, 24 kHz, where voice-over recordings go), imports it and places it at the playhead on the first free targeted audio track, else the first free audio track, else a new one, as one undo step. The settings are kept per generated item in Project::narrations (schema v14). Selecting a narration clip loads it into the panel; tts.edit writes a new file (beside the previous one) and swaps the clip to it, keeping the clip's start, duration and speed: longer speech is kept whole in the file but plays only to the clip's end; shorter speech is padded with silence to the clip's length. The old item and narration stay, so undo is exact. tts.preview synthesizes without changing the project and the UI plays Session::tts_preview on the audio output. Built-in voices are an original formant synthesizer (robotic, no download). The natural voices (feature neural-voices of the desktop app, off by default like whisper) are Kokoro-82M on candle (kinema-tts) fed by kinema-tts-text (normalizer, CMUdict + letter-to-sound); their 336 MB package is downloaded once after the user confirms (tts.downloadVoices, a job). The UI synthesizes with tts.render (a job filling Session::tts_cache) and then runs create / edit / preview, which use the cache.

5.1.3 Remix

clip media (source In … +duration) ─► audio_dsp::remix::analyze: spectral-flux onsets → tempo (autocorrelation, 50–200 BPM)
  → DP beat tracker → per-beat chroma + mel cepstra → beat self-similarity
  ─► plan(target, Segments, Variations): intro … joints at beat boundaries … outro, length within one beat
  ─► hidden clip effect `remix` (target, sliders, original duration, plan in ticks) ─► render::remix::read (20 ms equal-power crossfades)

Clip ▸ Remix ▸ Enable Remix / Remix Properties… / Revert Remix (clip.remix.*; agents can use clip.remix {duration|seconds…}) and the Remix tool in the Rate Stretch group (drag a music clip's Out edge) retime a music clip to a target duration. The analysis finds the beat and compares every pair of beats by harmony (chroma) and timbre (cepstra). Each joint leaves the music at one beat boundary and continues at the beat boundary whose four beats on either side sound most alike. The intro and outro are kept, and the result lands within one beat of the target. Segments (0–100) sets the number of joints (1–3, more for long extensions). Variations (0–100) sets how far a joint may move from its evenly spaced position (±1–8 beats). The state is a hidden effect instance with no definition, so panels, the effect chain and Paste Attributes ignore it. The plan is stored in ticks, so playback and export read it without re-analysing. A remix never moves other clips: one that would overlap the next clip is refused, and Revert restores the original duration. Results are deterministic and the analysis is cached per media range.

5.2 Essential Sound

clip.essential (type + settings)  ──apply()──►  clip effects marked `essential`, clip gain, Volume, Panner
essentialSound.autoMatch   BS.1770 integrated loudness of render::audio::clip_signal (mono clips as one channel) → match gain (clip gain)
essentialSound.generateDucking   trigger clips' summed level (10 ms hops) → activity regions → Volume keyframes

• Model (project::essential). A clip's essential: Option<EssentialSound> holds its audio type (Dialogue / Music / SFX / Ambience) and per-type sections: Loudness (match gain, measured and target LUFS), Repair (Reduce Noise, Reduce Rumble, DeHum 50/60 Hz, DeEss, Reduce Reverb), Clarity (Dynamics, EQ preset + amount, Enhance Speech), Creative (Reverb preset + amount, Stereo Width for Ambience), Ducking (against types, sensitivity, reduce by, fades), Pan, Clip Volume and Mute.
• Effects under the hood. essential::apply(item, old, t) turns the settings into ordinary clip effects (Highpass, DeNoise, DeHummer, DeEsser, DeReverb, Dynamics Processing, Parametric Equalizer, Enhance Speech, Stereo Width, Studio Reverb) flagged EffectInstance::essential, in that order, ahead of the user's own effects. Only parameters whose derived value changed are written (at the playhead), so keyframes added in Effect Controls survive. A section switch bypasses its effects, a slot switch removes its effect, clearing the type removes them all. Auto-match gain, Clip Volume (an offset on the Volume level or on all its keyframes) and Pan are applied as deltas, so clearing restores the clip. Because these are normal effects, playback, the mixer and export need nothing special.
• Loudness. The clip signal (clip gain + effects, before Volume) is measured with the BS.1770 meter; the gain is linear after the effects, so one measurement hits the target exactly. Targets are preferences (audio.dialogueTargetLufs −23, musicTargetLufs −25, sfxTargetLufs −21, ambienceTargetLufs −30).
• Ducking. Sensitivity 0…10 maps to a threshold −20 − 4·s dBFS on the summed trigger signal (50 ms window); regions shorter than 100 ms are dropped and pauses under 250 ms bridged; each region gets a fade-down before it and a fade-up after it (audio_dsp::ducking::duck_keyframes), written as Volume level keyframes that replace earlier ones.
• Presets (our own names and values) per type; user presets are saved in preferences (essentialSound.userPresets). Music remixing to a duration is Clip ▸ Remix (§5.1.3).
• Enhance Speech is a DSP chain (high-pass, de-mud, presence and air EQ, expander, compressor), not a model: no speech-enhancement model with an open licence that we could ship and verify is bundled. DeepFilterNet (MIT/Apache-2.0, Rust inference via tract) is the candidate for a future optional integration behind a trait.

5.3 Masks

Every video effect and the intrinsic Opacity carry EffectInstance::masks (project::mask). A mask is a closed cubic Bézier MaskPath in clip pixels (ellipse = four smooth vertices with circular tangents, 4-point polygon = corner vertices, the pen draws arbitrary vertices), stored as a ParamValue::Path so the ordinary keyframe engine animates it (vertex-wise interpolation; paths with different vertex counts hold). Feather, Opacity and Expansion are ordinary float parameters.

• Coverage (render::mask): the path is flattened (≤ 0.05 working px chord error); the signed distance to the polygon (nonzero winding) plus Expansion goes through a falloff of width max(Feather, 1) centred on the edge (linear = exact box-filtered antialiasing at Feather 0, blending into smoothstep as Feather grows). Masks combine top to bottom with Add / Subtract / Intersect / Lighten / Darken / Difference; Inverted and Opacity apply per mask.
• Semantics. A masked effect is lerp(original, effected, coverage) per premultiplied channel (the effect only applies inside); Opacity masks scale the clip's layer before Motion, so they follow the clip's transform. Adjustment-layer masks are in sequence pixels.
• GPU. kinema-gpu::GpuMask evaluates the same coverage and mix in WGSL (compute), tested to agree with the CPU within 3·10⁻⁶. Layers with masks are CPU-rendered images in frame plans.
• Editing. masks.* commands (add / remove / set / moveVertex / translate / addVertex / removeVertex / toggleVertexSmooth / roundCorners / select / list; masks.add takes a cornerRadius for a rounded rectangle); keyframe commands take "mask": n. The Program monitor overlay drags vertices, Bézier handles, the whole mask and the feather / expansion handles; drags merge into one undo step.
• Tracking (render::track, masks.track): Shi–Tomasi features inside the mask, pyramidal Lucas–Kanade (4 levels, 15×15 window) with a forward–backward check, then RANSAC + least squares for Position / Position & Rotation / Position, Scale & Rotation (2D Procrustes). Each frame's transform moves the path, written as Mask Path keyframes while the background job runs (one undo step per run; jobs.cancel stops and keeps what was tracked). Frames are tracked at ≤ 960 px wide. On synthetic footage with known similarity motion the path stays within 0.3 px over 20 frames.

5.4 Colour management

frame (Y'CbCr/RGB + metadata) ─► source colour space: Interpret Footage override, else VUI/colr/MKV Colour
  ─► decode table (sRGB/BT.709, PQ, HLG scene light, camera log) ─► HLG OOTF ─► 3×3 to working gamut
  ─► BT.2390 tone map (HDR/log into an SDR sequence, Auto Tone Map Media) ─► gamut compression
  ─► effects + compositing in working linear (1.0 = reference white = SDR white = 203 cd/m²)
  ─► monitors: working → SDR BT.709 (tone map from HDR, gamut map from BT.2020)
  ─► HDR export: working → PQ/HLG BT.2020 R'G'B' → Y'CbCr (BT.2020 NCL) + VUI/colr/mdcv/clli/SEI

• Model. SequenceSettings::color (ColorPipeline: working space Rec. 709 / Rec. 2100 PQ / Rec. 2100 HLG, wide gamut, auto tone map) and Interpretation::color_space (per media item; None = from metadata). Commands: sequence.colorSettings, clip.interpretFootage, color.spaces, media.colorInfo.
• Maths in kinema-color (transform, log, spaces; formulas and sources in crates/color/README.md); the renderer side is render::colorman.
• Fast path. A Rec. 709 sequence without wide gamut and media whose metadata says Rec. 709 / sRGB decodes exactly as before, and its layers stay on the GPU. Log, HDR or wide-gamut media, and every layer of an HDR/wide-gamut sequence, are converted on the CPU (the GPU path draws them as pre-rendered images, like layers with effects).
• Outputs. RenderOptions::working_output returns working-space pixels (HDR exports, scopes); otherwise the top-level render is converted for an SDR monitor. The other colour effects work on display-encoded values clamped to 0..1 (they clip HDR highlights above reference white).
• HDR grading (Lumetri). FxCtx::working carries the sequence's working space. In a PQ or HLG sequence Lumetri grades the working space's own signal normalised to HDR White (cd/m², Basic Correction; the curves, wheels, looks and HSL Secondary use the Curves section's HDR Range): kinema_color::GradeSpace (PQ: PQ⁻¹(nits/10000) / PQ⁻¹(white/10000); HLG: the BT.2100 inverse OOTF of a display with that peak, then the OETF). So the sliders and curves span 0 … HDR White like black … white in SDR, exposure is stops of light, and highlights above HDR White pass through (contrast leaves them alone; HDR Specular scales them). Rec. 709 sequences grade exactly as before. HSL Secondary ▸ Refine: Denoise (median of the key, radius 1–3 px) and Blur (Gaussian of the key, σ ≤ 20 px), both scaled with the playback resolution.
• HDR metadata. VideoStreamInfo::hdr (HdrMetadata: ST 2086 mastering luminance, MaxCLL, MaxFALL) is read from mdcv / clli (MP4 / MOV) and Matroska MasteringMetadata / MaxCLL / MaxFALL. PQ media tone mapped into an SDR working space uses it as the BT.2390 source peak (MaxCLL, else the mastering peak; 1000 cd/m² without metadata), so a 4000-nit master keeps its highlight detail instead of clipping at 1000. media.colorInfo reports it (hdrMetadata, toneMapPeakNits). The field is optional and serde-defaulted (no schema change).
• HDR export. H.264 and ProRes exports of a PQ/HLG sequence encode BT.2020 PQ/HLG (ProRes 10-bit; our H.264 encoder is 8-bit, so H.264 HDR is 8-bit) and signal it in the VUI / ProRes frame header, colr, and for PQ mdcv (BT.2020 / D65, 1000 / 0.0001 cd/m²) and clli (MaxCLL/MaxFALL 0 = unknown, they are not measured) plus the matching H.264 SEI. sdr: true exports the tone-mapped SDR picture instead; render previews always do.
• Monitors / display colour management. The monitors are SDR (sRGB-encoded RGBA8 textures): HDR sequences are shown tone mapped. macOS EDR (extended-range CAMetalLayer output) is not wired up: there is no platform crate yet and eframe/wgpu do not expose an EDR surface, so HDR values above SDR white are never sent to the display. The scopes of an HDR sequence show the working-space values (waveform in cd/m² on a PQ scale, BT.2020 vectorscope).
• LUTs. Lumetri Input LUT and Creative Look reference lib:<id> (the project's LUT library, Project::luts, embedded .cube/.3dl text) or builtin:<id> (code-generated camera conversions and looks). kinema-gpu::GpuLut is the WGSL tetrahedral counterpart, tested for parity.

5.4 Multi-camera and synchronisation

clips ──sync (in | out | timecode[±hours] | marker | audio)──► anchors (media time ↔ common instant)
  ├─ clip.synchronize        moves selected timeline clips (link groups together) onto the reference
  ├─ clip.mergeClips         video + ≤16 audio clips → a merged-clip sequence
  └─ clip.createMulticam     cameras → a multi-camera source sequence (one video track per angle)
multi-camera clip = nested source + TrackItem::multicam {enabled, angle}
  render: only the angle's video track · audio: camera 1 | all | the angle (Switch Audio)

• Model (project::multicam). Sequence::multicam (MulticamSource: cameras with their video track, audio tracks, name, shown flag and source item; audio mode) marks a multi-camera source; Sequence::merged a merged clip. TrackItem::multicam (MulticamSel) makes a nested clip a multi-camera clip; any nest can be one (its video tracks are then the angles, Sequence::cameras()). Editing a multi-camera source into a sequence gives an enabled clip on the first angle (Project::make_track_item). Project schema v7.
• Sync (engine::sync). Each method reduces a clip to an anchor (the media time that lines up with the common instant). Audio uses audio_dsp::sync::find_offset: DC removal, windowed-sinc decimation to ≤ 8 kHz, GCC-PHAT-β (β = 0.75) via one packed complex FFT for the coarse lag, then the same at the full rate on the loudest common window (≤ 2.7 s) and parabolic interpolation. Recordings with different gains, microphones (filtered), 0 dB SNR noise or a strong echo are aligned to the sample; two 10-minute recordings take ~2.4 s (release). Clips are placed on frame boundaries with the sub-frame remainder taken from their source In, so video stays on frames while audio keeps sample accuracy.
• Render. render::item_layer renders a multi-camera clip's angle track only (render_seq_tracks); render::audio mixes the nested source with only the audible tracks (Sequence::with_angle_audio). Nested audio now plays at all (it was skipped when the nest had no media source) and is limited to the clip's range. render::multicam::render_grid_page renders one page of the shown angles at the cell scale in parallel (rayon) and tiles them: the Multi-Camera view is one frame job (frames::Target::MulticamGrid(source, layout, page)), prefetched while playing like the program. Only the page's angles are decoded.
• Multi-Camera view settings (EditorState::multicam_view, Program ▸ wrench menu): grid layout (automatic = smallest square up to 4×4, or fixed 2×2 / 3×3 / 4×4) with pages when the angles don't fit (render::multicam::page_layout; multicam.gridLayout, multicam.page, page arrows; keys 1–9 pick cameras on the shown page, multicam.cut {camera} stays absolute); Multi-Camera Selection Top Down (stacked multi-camera clips: topmost instead of lowest); Show Multi-Camera Preview Monitor (off: the grid fills the monitor); Auto-Adjust Multi-Camera Playback Quality (grid_cell_scale: ½ / ¼ of the cell scale while playing); Transmit Multi-Camera View (stored; there is no transmit device yet). multicam.grid returns the page. Edit Cameras… shows a thumbnail per angle (frames::Target::MulticamAngle).
• Editing (edit::multicam, engine::multicam). multicam.switchAngle (click an angle, Ctrl/⌘-click for video only), multicam.selectCamera1…9 (keys 1–9) and cutToCamera1…9 (Ctrl+1–9), Enable/Flatten, Edit Cameras, Audio Follows Video (EditorState). Live switching: playback in the Multi-Camera view runs multicam.recordStart; each key/click is a multicam.cut that is applied at once (the program shows it) by re-applying the whole pass to the project from before the pass (edit_merged), so a pass is one undo step; Stop runs multicam.recordStop, which ends the last angle at the stop point. Through edits inside the recorded range are healed, so pressing the angle already showing adds no edit. Flatten replaces a clip by the clip(s) its angle shows (outer effects carried over; linked pairs stay linked).

6. Export jobs (kinema-export)

file.exportMedia {path, preset?, settings?, format?, range?, …}     export.quick {preset?, path?}
  → engine::export_tools: preset ⊕ settings JSON ⊕ flat overrides → ExportSettings; range resolved
  → a Job {id, label, progress, result} on a background thread (export.queue.* runs them in order)
  → export: per output frame render (output rate) → fit into the output frame → overlays → limiter
            → encode in order → mux; audio mixed at the output rate, folded to mono/stereo,
            loudness-normalized (measuring pass + gain + true-peak limiter) per batch
  → jobs.list shows progress; jobs.cancel sets the shared cancel flag

• Reproducible. The same project and settings give the same file on every machine. Frames render in batches of one per core, but the muxer gets them in fixed groups of 16 output frames (INTERLEAVE in crates/export/src/job.rs), each followed by its audio, and H.264 uses one slice per four macroblock rows instead of one per core. MXF files are the exception: their UMIDs and modification date come from the clock, as SMPTE ST 330 / ST 377 expect.
• Settings. ExportSettings (serde, camelCase, every field optional) holds Video (frame size or Match Source, frame rate, Scale to Fit / Fill / Stretch, pixel aspect, field order — progressive only, profile / level, bitrate encoding, target / maximum / adaptive bitrate, keyframe distance, maximum render quality), Audio (codec, sample rate, channels, AAC bitrate, PCM / FLAC sample size: bits 16, 24 or 32 and float for 32-bit float PCM in WAV, AIFF and QuickTime — supports_sample_format says which a codec and container store, and Export mode greys out the rest), Multiplexer (audio channels: mono, stereo or 5.1, §5.1.1), Captions (burn-in or SRT / WebVTT sidecar), Effects (image / name / timecode overlays, video limiter, loudness normalization), Metadata (udta ©nam, ©ART, ©cpy, ©des, ©cmt). settings.summary() / estimate_bytes() feed Export mode's Summary.
• Presets. export::presets::builtin_presets() are our own definitions (Match Source adaptive H.264 at 0.2 / 0.1 / 0.05 bits per pixel, 1080p / 2160p delivery, vertical 1080×1920, ProRes, DNxHR, MXF OP1a (DNxHR HQ / ProRes 422 HQ / H.264) and OP-Atom (DNxHR, Avid), image sequences, GIF, WAV / AIFF). User presets and favourites persist in <data dir>/export-presets.json (export.presets.*).
• Queue. export.queue.add snapshots the project and the resolved settings (several sequences or ranges add several items); export.queue.start encodes ready items one after another as ordinary jobs, advanced by Session::poll_persistence / pump_jobs (or synchronously with wait). Cancel, retry, reorder, remove and clear work per item. The queue is session state, not part of the project file.

Video encoders implement export::VideoEncoder. Codec crates plug in with register_encoder and register_audio_encoder.

Timelines can be exchanged as EDL, FCP7 XML, FCPXML, OTIO, AAF or OMF. file.import detects these formats and merges the result into the project as one undoable step (AAF / OMF embedded audio is written next to the document first). file.import {replace: true} starts from an empty project with the same name (file.newProject) so a headless import→export round-trip contains only the imported timeline. file.exportInterchange, file.exportEdl, file.exportFcpxml, file.exportOtio, file.exportAaf and file.exportOmf write them. file.exportInterchange {format} takes edl, xml (the default), fcpxml, otio, aaf or omf; any other value is a parameter error and nothing is written. For AAF and OMF the engine (engine::aaf_omf) first prepares the audio the document references: it lists the used ranges (interchange::essence::audio_needs), decodes or renders them (clip effects through the export audio pipeline), embeds them or writes WAV / AIFF files, and optionally renders a video mixdown. A nested sequence is exported the way Premiere Pro does it: a composition of its own in AAF (the media of the clips inside it is prepared with the rest), its sound mixed into the document in OMF, a nested sequence in FCP7 XML, FCPXML and OTIO, and one AX event in an EDL (table in the interchange README).

7. Automation surfaces

All of these dispatch the same command ids.

• Automation ids. Every interactive widget calls app.auto.add(id, rect, label) each frame (crates/ui-egui/src/automation.rs). Agents click by id, e.g. tools.Razor, timeline.clip.<id>, effects.item.gaussian_blur, panel.Timeline.

• UI state that is not project data (tool, workspace, dock layout, zoom, scroll, monitor settings) is in crates/ui-egui/src/state.rs as serde structs, so ui.inspect and ui.set can read and write it.

• Panels with engine data (M8.9 / M12.6): Lumetri Scopes (scopes.read returns the same scopes as numbers), Metadata (metadata.get / metadata.set, one undo step per edit, stored in ProjectItem::metadata), Events (events.list / events.clear), Progress (jobs.list / jobs.cancel), Project Notes (project.notes / project.setNotes, plain text in Project::notes, schema v15; one typing session is one undo step). Timecode and Reference Monitor are frontend-only views. Their settings are UiState::panels (ui.set {"panels": {...}}).

• Project panel and Media Browser (M12.7). View settings (List / Icon / Freeform, thumbnail and font size, Preview Area, Hover Scrub), the List view's columns, widths and sort, and ten view presets are preferences (Preferences::project_panel; project.view.*, project.columns.*, project.sort, project.viewPreset.*). Freeform positions, Clip Size, stacks and saved arrangements are item metadata (Freeform Position, Freeform Stack, Freeform Arrangement: <name>; project.freeform.*), so they are undoable and saved with the project without a schema change. The Media Browser lists directories through Services::list_entries / volumes / home_dir (native: std::fs; web: the virtual file table; tests: a fake filesystem): engine::media_browser keeps navigation history and the selection in Session::browser, Favorites / recent directories / file types / columns in Preferences::media_browser (mediaBrowser.*). Bins opened in a tab or a window, the inline rename, dialogs and the hover-scrubbed card are UiState::project_panel; the browser's tree state is UiState::media_browser (ui.set {"projectPanel": …, "mediaBrowser": …}).

Protocol reference: control-protocol.md. Agent guide: agents.md.

8. Not built yet

The layer table reserves names for crates that don't exist yet: riff, mjpeg, keyframe, effects, audio and playback. (platform exists but holds only OS media FFI; other OS integration stays in apps/kinema.) Until they exist, that work lives elsewhere: keyframes and effect definitions in project, effects and the audio mix in render, playback in ui-egui, and OS integration (cpal, rfd, native menus) in apps/kinema. roadmap.md has the milestone status and gaps.md what is missing.

Revision history