SuperAPI AI Directory

threejs-game-orchestrator

General & other

View repository

Updated 10-16 09:28

About this skill

three.js game orchestration

With threejs-game-workflow-control as the sole authority for global state, risk gates, and task scope. This Skill is responsible for breaking down domain work, freezing handoff artifacts, collecting evidence, and returning to the control plane; it does not approve operations, expand scope, or create a second state machine.

Shortest process

The user first followssimplified-workflow.mdUnderstand the six-stage project view: requirements and scope → global baseline → foundational engineering → scene-by-scene production → global integration verification → release; a single scene is understood as scene definition → breakdown and layout confirmation → formal asset acceptance → formal restoration and runtime acceptance. This view does not change G0-G3, V0, V1, V2, V3, V4, A0-A6, F0-F4, or the hard evidence gates.

  1. Read the Work Item, current task description, current baseline, and applicable state gates; when fields are needed, read references such as quality-gates.md、module-decomposition.md、game-implementation.md、delivery-artifacts.md as needed.
  2. First submit the minimal domain proposal and acceptance boundary; after review is completed, the control plane runs check; stop exploration when the existing entry points, call chain, task scope, and risk facts are sufficient.
  3. Before A3 implementation, freeze the implementation package bound to the current task scope, baseline, file ownership, execution units, acceptance commands, and stop conditions; after implementation, record the candidate change audit and submit the current candidate verification evidence.
  4. Domain delivery proceeds only within the current task path; when the plan, path, or resource list changes within the task, update the plan and re-verify the affected parts; when requirements or scope genuinely change, submit a Change Request. All domains, stages, and work items are judged one by one according toactual completion status, and when the current explicit requirements are already met, do not retroactively amend historical contracts due to mid-course adjustments; finally return to the control plane to run run/check, and the control plane advances the status.
  5. After stage acceptance is completed, the execution agent followsstage wrap-up and Git committo automatically maintain the necessary .gitignore and commit the files for this task, record the commit hash, and hand off; parallel tasks are wrapped up uniformly by the main agent, and sub-agents do not each commit to the shared workspace.

Delivery boundary

  • Hand off records and intermediate artifact organization to threejs-game-artifact-cleanup: stage wrap-up only cleans up temporary files that are regenerable within this task and have no references; cross-task historical cleanup must be planned separately within the scope specified by the user, preserving approvals, rework, and the valid evidence dependency chain.
  • Production, architecture, gameplay, visual, asset, audio, numerical, QA, performance, and release work may only propose, review, or modify within the current task.
  • A0-A3 are based on the current user task; local A4 integration with no external side effects may be executed per task. External writes, payments, real devices, destructive or external deletions, and releases require the control plane to establish a precise pending item and wait for explicit approval; this Skill does not perform these actions.
  • V0→V1→V2→V3→V4, the global static baseline, scene breakdown and layout confirmation, and high-fidelity prerequisites continue to use the control plane's non-bypassable evidence gates; domain documents only supplement domain facts.
  • Do not overwrite modifications by parallel agents, and do not automatically roll back the shared workspace; before starting a local verification service, first follow local-service-validation.md to find a reusable healthy instance.

Visuals and scenes

Planning, concept art, and implementation hand off the sameplayer-visible expressionWhen delegating generation tasks, include the visible text range and the constraint against drawing internal notes; when delegating interface implementation, include short labels, value sources, and display conditions; the full internal GDD/brief cannot replace this handoff, nor can it go directly into the frame.

Image delivery followsimage-first confirmation methodexecution: domain agents organize records internally and directly produce images or real previews; the main agent presents candidates and a small number of checkpoints, without using text proposal confirmation as a routine pre-production prerequisite; within an already confirmed target, rework is submitted directly with before-and-after comparison. Real images and asset dependencies are still confirmed according to the corresponding stage.

Effect image restoration and formal Scene integration belong to the current scene Work Item; persistent HUD follows the scene closed loop. Modal/popup and similar dialogs followindependent work item rulescreate a separate DISPLAY_LAYER Work Item, independently submitting host context, component/state, layout contract, and runtime evidence, without affecting acceptance of the lobby or other host scenes.

Scene V2 layout delivery must simultaneously provide a standard layout PNG,layout-nodes.json、layout-decision.json, a self-contained review.html and generation-result.json, and in the independent layout confirmation and scene root plan associate the same batch of references, breakdown, decisions, nodes, PNGs, and actual review pages. Screen UI also followsThreeJS UI node organizationto define the root hierarchy in V1, confirm parent responsibilities and the ui_layout field in V2, and check responsibilities, input, clipping boundaries, and runtime results under viewport changes in the formal V4 scene. Orchestration only referencesoffline layout review artifacts's general rules, and does not handwrite review pages for scenes; page candidate status does not equal human confirmation.

Juiciness and dynamic feedback routing

Juiciness is an independent domain capability invoked on demand: invoke it when the user explicitly raises juiciness, impact feel, or dynamic feedback needs; it can also be referenced on demand when an existing task includes core feedback but the event semantics are still unclear. It follows the current scene or gameplay Work Item, does not expand ordinary gameplay modifications into full-scale polish, and does not add stages, approval gates, or mandatory artifacts. Followworkflow integration instructionsto arrange responsibilities: production determines intent and key events in the existing GDD/contract, gameplay handles rules, collision, state, triggers, and interaction code, architecture maintains cross-module coordinate conventions; UI handles layout, focus presentation, and layout coordinate contracts, and feedback objects only use independent presentation offsets, without overriding baseline layout values, contending for focus, or input handling. Animation and audio follow their respective production and lifecycle rules, and QA verifies event boundaries and hands dynamic display, comfort, and feel to the user for manual acceptance. For complete semantics and type rules, seeJuiciness Design Guide。

status and rework

This domain does not directly migrate global state. When evidence fails, prefer in-place repair, when upstream facts are unchanged revalidate, and only when upstream facts are invalidated, the task scope truly changes, or a hard gate would be bypassed, perform the minimal-scope return; ordinary candidate identity changes do not trigger rollback, and confirmed facts are not repeatedly searched or overturned.

3D runtime, models, and plugins

3D project foundations use the runtime foundation; formal 3D assets are handed to the model and level pipeline, adopting the plugin solutionaccording to the actual gameplay. The real 3D world is delivered by models/meshes, materials, lighting, Camera, levels, and collision data; character skeletal animation uses imported clips and AnimationMixerFixed-viewpoint renders, 2D breakdowns, and HUD layouts cannot replace world models, cameras, collision, or animation data.

New model production is organized within the current scene or asset Work Item according to theBlender Model Production Contractorchestration: V1 generates/freezes references, V2 breaks down consistent multi-views, V3 invokes Blender MCP to control polygon count modeling, create AI textures and usable normals, and conduct manual model review; after acceptance, integrate into the real scene in V4. Image generation/editing is delegated to image_generator; geometric modeling must be delegated to gpt-6.1-sol、max a modeling sub-agent with reasoning intensity, invoking Blender MCP with parameters specified by the contract. Budget, file ownership, and review evidence are incorporated into the existing implementation package; do not create separate chats, claim model completion using reference images, or automatically determine that screenshot/real-device visual acceptance has passed.