Author a Mesh Model Test Case

Scaffold a new static mesh-model asset-generation test case — a 3D model the model sculpts by compositing a continuous signed-distance field and meshing it with a binary (mc → mc-model, sn → sn-model, or dc → dc-model), one recorded operation at a time. This is the short version; Authoring a Mesh Model Test Case covers it in full, and Manifests is the authoritative schema.

A rigged, animated mesh instead? See Author a Mesh Animation Test Case. Discrete cube voxels rather than a meshed field? See Author a Voxel Model Test Case.

Layout

A version lives at test-cases/<type>/<difficulty>/<slug>/<version>/ and is immutable once runs reference it — revise by adding a new version.

test-cases/<type>/<difficulty>/<slug>/<version>/
  test-case.toml   # manifest: type, asset_kind, [voxel], [tool], [output], the overall domain
  variants/        # one standalone TOML file per variant (listed in `variants`)
  prompt.hbs       # rendered into the harness instruction (NOT seeded)
  description.md   # site blurb (NOT seeded)
  README.md        # human overview (NOT seeded)
  specs/brief.md   # what to sculpt + how the tool behaves — SEEDED

A run seeds only the brief, the binary on PATH, a pre-seeded <binary>.config.json, and a blank preview. There is no target model and no operations schema — the binary’s --help is the contract.

Steps

Pick a catalog slug (e.g. aegis-dc) and the subject to sculpt. It should read clearly at the volume size from silhouette and palette alone and be achievable by compositing CSG primitives (add-*/subtract-*, --blend, mirror).
Pick the algorithm for the surface you want — it fixes the asset_kind and [tool].binary: mc (low-poly faceted), sn (smooth watertight), dc (crisp sharp edges; only dc exposes a per-primitive --sharp/--smooth tag). The kind is a property of the whole version, not a variant axis.
Write specs/brief.md: subject, silhouette, orientation, exact opaque #rrggbb palette, the [voxel] volume, and — factually — which extractor meshes the field. Keep it self-contained; state the extractor’s behaviour, don’t prescribe the look.
Write prompt.hbs using only the documented template variables ({{variant.*}}, {{#each specs}}) — it renders in strict mode — pointing at the binary’s --help and restating that it must render before finishing so the mesh is emitted.
Write test-case.toml per the Voxel cases schema: metadata (name, difficulty, tags — include 3d/mesh and the algorithm), type = "asset-generation", asset_kind ("mc-model"/"sn-model"/"dc-model"), a variants list (a root key, so it precedes the first table; first = default), the [voxel] volume, and the [tool]/[output] tables. [tool].binary is the meshing binary; [tool].preview (e.g. model.png) and [output].actions (the op log) are each a single file — no {part} token. Add the single overall [[domain]] a human rates under; there is no [[review_item]] checklist. The extracted mesh.glb is emitted by core on render, not declared in the manifest.
Reject the wrong tables: a meshed case declares no [canvas], no [model] (a static model has no rig), no [[reference]] (no target), no [build], and no [[check]]. Write description.md and README.md (both non-seeded).

Worked examples: the Aegis Bastion walking fortress, authored once per algorithm (test-cases/asset-generation/medium/aegis-mc/v1.0.0, aegis-sn/v1.0.0, aegis-dc/v1.0.0). Read the one matching your surface alongside the full guide before you start.

Validate

npm run lint:specs   # markdownlint-cli2 + cspell over test-cases/**
tcab prompt --test-case <slug> --version <version> --variant <variant>
tcab seed   --test-case <slug> --version <version> --variant <variant>

Render the prompt and inspect the seeded repository for every variant to confirm the manifest resolves and the seeded set is self-contained. After editing, force-re-ingest so a backend-driven run picks up the change — see Running the services locally.

Next steps

Create a Mesh Model Variant to add a brief variation.
Run a Test Case to exercise it end to end.
Review a Run to assess the result.