Architecture

CineCode is a Rust workspace of seven focused crates. Each owns one stage of the journey from a text project to a finished video.

The crates

Crate	Responsibility
`cinecode-core`	The Scene Document / IR — the data model every other crate reads or writes.
`cinecode-timeline`	Animation tracks: evaluating keyframes and easing per frame.
`cinecode-highlight`	Syntax highlighting via syntect (~212 languages).
`cinecode-sim`	Deterministic simulations: terrain, pathfind, particles, graph, life, boids.
`cinecode-render`	Rendering: a CPU backend (cosmic-text + swash) and a GPU backend (wgpu, WGSL, bloom).
`cinecode-script`	Loading projects: parsing `codescene.toml` (TOML) and `.ccs` scripts (Rhai).
`cinecode-cli`	The `cinecode` command (clap) and the ffmpeg encode pipeline.

Load. cinecode-script reads the manifest and runs each .ccs script (Rhai), with cinecode-highlight colorizing any code. The result is a Scene Document: a JSON-serializable intermediate representation.
Evaluate. For each frame, cinecode-timeline evaluates every animation track and the camera to produce an AnimatedState and a CameraState. Simulations from cinecode-sim contribute their state for that frame.
Draw. cinecode-render projects the camera and rasterizes the state into an RGBA frame — on the CPU, or on the GPU when --gpu is set.
Encode. cinecode-cli streams raw frames into ffmpeg, which encodes H.264 and muxes audio into the final MP4.

The Scene Document is the contract. Anything that can produce one can be rendered; anything that can read one can analyze or transform a film.
Separating evaluate from draw is what makes rendering deterministic and seekable — see The rendering pipeline.
Keeping simulations in their own crate lets them be pure functions of (parameters, frame), so they slot onto the timeline like any other node.

The IR at the centre of the system.

Per-frame evaluation through to ffmpeg.

Why the same project always renders the same pixels.