@soullink-emotion/profile-generator

Node.js SDK for building and validating soullink.profile.json from Live2D .model3.json, .cdi3.json, .exp3.json, and .motion3.json files.

It provides deterministic Cubism-ID and CDI-name mapping first, with optional OpenAI-compatible refinement. LLM output is validated against real CDI parameter IDs before it can be persisted.

import { Live2DProfileAutoGenerator } from "@soullink-emotion/profile-generator";
import { OpenAICompatibleClient } from "@soullink-emotion/planner-openai";

const generator = new Live2DProfileAutoGenerator({
  modelsRoot: "/srv/soullink/models",
  modelsBaseUrl: "https://assets.example.com/models",
  client: new OpenAICompatibleClient({
    apiKey: process.env.OPENAI_API_KEY,
    baseURL: process.env.OPENAI_BASE_URL,
    model: process.env.OPENAI_MODEL
  }),
  useConfiguredOpenAI: true
});

const result = await generator.ensure({
  modelDir: "my-model",
  force: false
});

The package never reads environment variables and never exposes an HTTP server. The host owns authentication, upload policy, modelsRoot, and API keys. It is a Node-only package because profile discovery and persistence use node:fs.

Set useConfiguredOpenAI: false (the default) for deterministic heuristic-only generation. A request-level openAI.apiKey still explicitly enables refinement. Use saveCalibratedProfile() to validate and persist manually edited mappings.

Generated profiles may include privateEmotionMap. These declarative rules map arbitrary model-specific parameters to dominant emotions and/or VAD ranges, with explicit active/neutral values, intensity, priority, and exclusive groups. The generator recognizes high-confidence labels such as 困惑, 生气, 星星, face shadows, and surprise effects; optional LLM refinement handles ambiguous labels. Every target is filtered against the actual CDI3 parameter set.