gnome-jarvis/QVAC.md

QVAC integration pin

Pinned against the live QVAC documentation for SDK release v0.19.0 on 2026-09-11. Source of truth: https://docs.qvac.tether.io/reference/api/ and https://docs.qvac.tether.io/reference/release-notes/.

Rules for Jarvis

• Use the function-centric @qvac/sdk API at version 0.19.0.
• Do not use removed startQVACProvider, delegate, or no_mmap APIs.
• Use load_mode for the replacement load behavior described by the release notes.
• Use completion() for streamed LLM work; its canonical surfaces are events and final.
• Use loadModel() / unloadModel() for model lifecycle.
• Use textToSpeech() or textToSpeechStream() for TTS.
• Use transcribe() / transcribeStream() for ASR, ocr() for OCR, translate() for text translation, and assessModelFit() before model fetch.
• Use cancel({ requestId }) for a specific operation and suspend() / resume() for runtime lifecycle.
• QVAC's local HTTP server is qvac serve --openai; it exposes OpenAI-compatible /v1/* at http://127.0.0.1:11434/v1/. qvac serve openai is deprecated.

v0.19 API names used by the capability registry

completion, batchCompletion, embed, ragIngest, ragSearch, ragReindex, ragDeleteEmbeddings, ragListWorkspaces, diffusion, video, audioGen, transcribe, transcribeStream, textToSpeech, textToSpeechStream, translate, ocr, classify, bciTranscribe, vla, worldCreateScene, worldStep, assessModelFit, downloadAsset, cancel, loadModel, unloadModel, suspend, resume, getSystemResources, getLoadedModelInfo, getModelInfo, modelRegistryList, modelRegistrySearch, modelRegistryGetModel, and state are documented public surfaces. Confirm parameter types against the installed .d.ts files before implementing each skill.

The HTTP server mounts /v1/chat/completions, /v1/responses, /v1/completions, /v1/embeddings, files/vector stores, image/video generation, audio transcription/speech/translation, and /qvac/v1/translate.