threejs-build-code-models

Build Three.js Code Models

Engine activation

Run commands/resolve-engine.sh and commands/resolve-skill.sh threejs-build-code-models before activation, edits or tool launches. Require compatible THREEJS selection. An explicit isolated prototype uses recorded --evidence and validates each --output; CODE producer commands accept --engine-evidence FILE and explicit prototype source paths. Manifests/reports stay in the recorded task. Never switch engines or art pipelines implicitly.

When to use

• The resolved art pipeline is CODE.
• A web, frontend, game, or technical-3D task needs a procedural Three.js model authored in source.
• Optional reference images are authorized for review evidence but must not become runtime dependencies.
• Route characters, creatures, animated actors, enemies, NPCs, avatars, mounts, and other embodied subjects to threejs-build-code-characters; keep props, environments, and generic procedural models in this skill.

Workflow

1. Read STYLE.md, resolve CODE with select-art-pipeline, and read .styles/.pipelines/code/PIPELINE.md.
2. If STYLE.md configures external asset libraries, inspect them first with use-external-asset-library. Copy any authorized template into the repository, but keep it as reference evidence rather than a primary runtime mesh.
3. Inspect the host's existing Three.js version and documented model interface. Do not add or install Three.js. Use the interface recorded by the code-model manifest when the host has no interface.
4. Scaffold a manifest below .hephaera/.cache/tasks/ with the text brief, semantic parts, pivots or anchors, deterministic seed, explicit budgets, cleanup ownership, and any authorized reference-image hashes.
5. Validate the manifest before authoring. Treat references as review inputs only.
6. Work in order: blockout, structure, form, materials, interaction, then optimization. Query the next unlocked pass before each change.
7. For each pass, author source, build or typecheck with the host toolchain, and capture deterministic browser evidence. Non-planar models require front, orbit-left, and orbit-right views.
8. Record exactly one review decision: pass, revise-spec, revise-code, request-input, or stop. Keep evidence hashed and project-confined.
9. Stop after three attempts on one pass or twelve attempts overall. After all passes, run complete validation and representative browser verification, including budgets and disposal.

Output

• Host-native Three.js TypeScript or JavaScript source and its factory export.
• A validated code-model manifest with semantic parts, anchors, budgets, and cleanup ownership.
• Hashed review evidence for every ordered pass.
• Typecheck/build and browser evidence for deterministic output, views, budgets, and disposal.

Guardrails

• Do not download, generate, or load a mesh/model-library asset as the primary implementation.
• Do not add Three.js to Hephaera or silently install it in a host repository.
• Do not use a reference image without a recorded rights basis and applicable consent.
• Do not skip or reorder passes, exceed correction limits, or fabricate evidence.
• Do not silently switch from CODE when implementation fails.

Commands

• ./commands/create-code-model-manifest.sh MODEL_ID --brief TEXT --prompt-override [options]
• ./commands/validate-code-model-manifest.sh [--strict|--complete] MANIFEST.json
• ./commands/next-code-model-pass.sh [--json] MANIFEST.json
• ./commands/record-code-model-review.sh MANIFEST.json --pass PASS --decision DECISION --summary TEXT [--evidence VIEW=FILE ...]