animforge

Problem

Hand-animating advanced 2D (and eventually 3D) work is normally too expensive to do at real quality. animforge's bet: a workstation built for an AI agent and a human to co-edit the same animation, not a GUI with an assistant bolted on.

Approach

The document is one semantic JSON scene graph — nodes, verb tracks, human pins — edited only through a closed, validated patch layer with deterministic solvers and numeric lints, so an agent edit can't silently corrupt state. The stack is TypeScript: a Vitest-tested engine and a PixiJS editor. Build groups G0 through G8 are merged — schema locked, a renderer, a look engine, and the editor's redline loop browser-verified end to end — backed by 8,471 engine tests and 201 editor tests, all green. Milestone M3 shipped the layer where the agent speaks animation in relative time and named easings, not raw keyframe numbers.

Currently building: after 466 issues and 76 completions, an owner decision recalibrated the product at roughly 2% of the Mappa/Toei quality bar it's held to — engine-rich, experience-poor, since whole built subsystems still have no UI. The plan is now six epics, starting with a reconciliation sweep and the stroke-design research gating the art pipeline.

Outcome

It runs locally; the repo is private. No public build yet — closing the experience gap comes before shipping wider.