Gameable Engine

Give an NPC a face ​

Goal ​

The spawn-character / set-expression / look-at commands a game already emits stop being recorded no-ops and start moving a real GNM splat head: the NPC smiles when the dialogue system says smile, and turns to look at whoever it is talking to. Nothing in the guest changes — this is entirely a host decision.

Files you will edit ​

  • src/main.ts
  • src/assets.json

Steps ​

  1. Bake a rig pack and serve it. A GNM head is a .aosrig pack: a neutral mesh, an expression basis, a skeleton and the skinning. It is licensed source data, so no template ships one — bake it and drop it in public/, which Vite copies verbatim:

    sh
    npm run gen:pack -w examples/character-showcase
mkdir -p templates/third-person/public/generated
cp examples/character-showcase/public/generated/myra_head.e64.aosrig \
   templates/third-person/public/generated/

    gen-gnm-pack.mjs writes two packs. The e64 one keeps the model's own 64-coefficient reduced view and is 7.8 MB against the full pack's 42 MB; Preview a rig without decoders has what the truncation costs in millimetres. Add public/generated/ to .gitignore: a 7.8 MB binary does not belong in a scaffold.

  2. Point the manifest entry at it. A character entry's rig block says which backend drives the head and, for gnm, which file holds it. pack is resolved like any other src — absolute from the site root here, or relative to the entry's own directory. In src/assets.json:

    json
    {
  "id": "char.guide",
  "type": "character",
  "src": "characters/guide/scene.json",
  "tags": ["character"],
  "rig": { "backend": "gnm", "pack": "/generated/myra_head.e64.aosrig" }
}

  3. Give the adapter a character bridge. The rig stack — the backend, the animator, the animated splat — is the optional half of @aosengine/wasm-host, so a game with no characters never downloads it. Import it dynamically and hand it to createEngineAdapter. In src/main.ts, where the adapter is built:

    ts
    import type { CharacterBridge } from '@aosengine/wasm-host/characters';

const { createCharacterBridge } = await import('@aosengine/wasm-host/characters');
const characters: CharacterBridge = createCharacterBridge({
  engine,
  renderer: engine.renderer,
  scene: engine.scene,
  // An entity's transform is its physics centre; this is where its neck is.
  headOffset: [0, 0.78, 0],
});

const adapter = createEngineAdapter(engine, { modules, characters });

    That is the whole wiring. The bridge resolves the bundle, fetches the pack, builds an AnimatedSplat and a GnmRigBackend under the entity's own object, keeps an Animator per character, and drops the placeholder capsule the moment the head is on screen. On the WebGL fallback it warns once and leaves every capsule where it was, so a player without WebGPU still gets a playable game rather than an exception.

Verify ​

sh
npm run dev -w templates/third-person

Walk up to the guide. The placeholder capsule is gone and in its place, at head height, is a head-shaped cloud of points tinted by skinning joint. Two things about that are deliberate: it is a head rig, so there is no body to draw — GNM bakes a head and a neck, not a skeleton with limbs — and it is the debug renderer, one gaussian per rig vertex, because the geometry and appearance decoders for GNM topology do not exist yet. Press E: the face moves on the lines the script marks smile, and the head turns towards the hero and eases back when the conversation ends.

Nothing on screen means the pack did not load. The bridge never throws — look in the console for the GNM rig for "char.guide" … did not load, which carries the URL it asked for and the reason. A warning about no GNM rig pack instead means step 2 did not take effect.

The template ships this as ?gnm=1 rather than as the default, so npm run dev -w templates/third-person -- --open '/?gnm=1' is the same thing without editing anything, and tests/e2e/third-person.spec.ts asserts it — skipping itself when there is no pack on the machine.

See also ​