Can propose, review, and modify assets and integration code within the current Work Item, and upon completion return to threejs-game-workflow-control to submit asset identity and verification evidence.
Blender model generation
When the user requests generation of a new model, within the current asset or scene Work Item follow reference image → multi-view breakdown → Blender MCP modeling and AI texturing → human model review to execute. Read Blender model production contract, write the subject matter, style, polygon count, and texture budget into the current implementation package; image generation and editing are handed to image_generator, and actual geometry and material production must be delegated to a modeling sub-agent with gpt-6.1-sol、max reasoning intensity to invoke Blender MCP for execution. When valid references or multiple views are already provided, reuse them and do not regenerate.
This route reuses V1 reference, V2 breakdown, and V3 formal asset production and acceptance, and does not create an independent chat or a second set of stages. Only after human confirmation of the current model and texture candidates can they enter the integration stage as accepted assets; automated static checks do not replace human review. Screenshots, renders, and on-device acceptance are executed according to the user's current explicit authorization.
Model integration and optimization
- V1 determines world scale, axes, pivot, bounding box, animation naming, PBR style, lighting, and 3D budget; the effect image is only a reference for the target viewpoint. The interactive world must deliver model/level data, collision proxies, and observable state, and must not disguise the level with a full-screen texture.
- V2 breaks down characters, environments, interactive objects, effects, and HUD. 3D objects record parent, TRS, units, camera, materials, collision, and visibility; the 2D layout JSON only constrains the HUD and fixed-viewpoint annotations and cannot prove that 3D space is correct.
- V3 asset manifest records item by item the GLB path and optional tracking summary, source/license, triangles, draw calls, material/texture, texture dimensions/VRAM estimate, skeleton/clip, LOD, collision proxy, and decoding dependencies. Texture disk size cannot replace the GPU memory budget, and generated images cannot serve as evidence that a model has been generated.
- Load with GLTFLoader, and register KTX2/meshopt/Draco according to the actual asset extensions; deploy the decoder/transcoder and verify CSP, paths, and failure prompts. Color textures are handled according to color space, while data textures such as normal/roughness are not treated as sRGB; do not blindly override materials already set by the loader.
- Each instance has an AnimationMixer and character state; SkinnedMesh clones use SkeletonUtils.clone, and clip/mixer time is used for presentation, while damage/rewards are driven by gameplay state. On switching, stop the action and uncacheRoot, and release exclusive geometry/material/texture; ImageBitmap and shared textures are managed by asset ownership.
- Test density first, then adopt InstancedMesh, merged static geometry, LOD, frustum culling, texture compression, and shadow limits. Instance picking is bound to instanceId; animation bounds, transparency sorting, and collision data are verified together with the degradation tier. Changing LOD must not change gameplay hit semantics.
V4 provides the same checkpoints for camera, lighting, timing, and game state for manual visual acceptance, and checks movement, collision, animation, scene transitions/re-entry, and failure recovery; items not actually executed are marked as unverified. Screenshots are collected only when the user explicitly requests automated visual verification, and on-device acceptance requires separate explicit authorization. Record the actually obtained renderer.info, load times, and the resource baseline after exit. Asset processing tools are chosen per project; do not automatically upload assets or request external paid generation.
Reference:glTF official loader、Resource disposal、Plugin integration。
Static GLB asset acceptance commands are in validate-model.mjs, and a checklist example is in model-manifest.example.json：node skills/threejs-game-model-pipeline/scripts/validate-model.mjs --repo . --manifest <模型清单.json>。
The production unit for 3D assets uses production_method=authored-model、delivery_kind=gltf-model, bind the model checklist and acceptance report path and submit the general file evidence for the control plane.sceneVisualRouteAnalysis serves only 2D effect-image coverage; GLB does not enter its image-asset/threejs-native branch. Fixed-view screenshots can prove visual results, and the model checklist proves spatial assets; the two are recorded separately.
The summary serves only as optional tracking information, is not required to be provided or compared, and does not block acceptance due to absence, errors, or changes; the model's actual content, path ownership, resource budget, and manual review results must still be checked.