@soullink-emotion/planner-openai

OpenAI-compatible reaction, reflection, and Live2D speaking-motion planners.

The package does not read environment variables. Server applications must pass credentials explicitly through OpenAIClientOptions.

SoullinkSpeakingMotionPlanner accepts (clientOrOpenAIOptions, generationConfig). The generation config exposes mode, fixedFrameCount, frameIntervalSec, minFrameCount, maxFrameCount, and twoStage.

Use mode: "fixed-parallel" (alias fixed) to plan concurrently with TTS. Use mode: "duration" after TTS to derive the frame count from ceil(durationSec / frameIntervalSec). Request-level frameCount and frameIntervalSec override configured values.

When credentials, parameters, or LLM output are unavailable, speaking motion returns provider: "vad-facs" and an empty parameterPlan. The runtime should continue displaying its VAD/FACS expression; no local parameter animation is generated.

Browser clients should use createSpeakingMotionApiClient to call a trusted backend instead of holding an LLM key. The HTTP client deliberately strips the openAI field before sending requests.

Only mouth/jaw opening parameters are reserved for LipSync. Mouth form, smile, pucker, lip shape, phoneme shape, and other non-opening mouth details remain eligible for generated keyframes.

For Node-side .model3.json/CDI3 scanning and soullink.profile.json generation, use the companion @soullink-emotion/profile-generator package.