Eidoverse Video — Data skill for Claude Code

Eidoverse Video — prealpha 0.01

**A film studio for AI agents.** Eidoverse is a toolkit an agent uses to make finished videos — world-building, characters, physics, simulation, music, sound, lipsync, cameras, and rendering, all behind one documented contract (`AGENTS.md`) that coding agents read natively. A human and an agent work in it together: the conversation is the writers' room, the agent is the filmmaker, and the toolkit is the studio.

Everything renders through **Deno + WebGPU + three.js/TSL at real-time speeds with minimal CPU** — GPU compute and node materials throughout, no per-frame CPU loops, no baking. Extracted from a production pipeline that has shipped hundreds of videos.

Quickstart

# 1. Install Deno 2.8.1 (pinned) + ffmpeg          → docs/SETUP.md
python eido.py bootstrap      # one-time dependency fetch
python eido.py doctor         # health check
python eido.py render eidoverse/examples/basic_vrm.json   # smoke test

Then open the repo in your agent (Claude Code, codex, opencode). It reads `AGENTS.md` and knows the studio. Working rhythm — single-frame probes, review, iteration: `docs/HARNESS_MODE.md`.

Branches

• main (this branch) — the harness edition: Deno + ffmpeg on your machine, no containers anywhere.
• auto — the containerized edition: a Docker render image and a runner that wraps the toolkit as a subagent for autonomous agentic loops (a parent orchestrator hands in a brief file and gets back a finished mp4, engine mounted read-only).

The studio, room by room

Render engine — `eidoverse/render_scene.mjs`

Scene scripts get the GPU device, injected assets, every helper below as a global, and an ffmpeg NVENC pipe out. Auto-enhance runs on every frame: N8AO ambient occlusion, screen-space reflections, bloom, FXAA. At the end of a render the engine **audits its own output** and names defects: flo