subsvibe

Guide for using the OpenAI Python SDK in SubsVibe client code. Covers synchronous and asynchronous chat completion and audio transcription calls via OpenAI-compatible APIs (Ollama, vLLM, LM Studio, OpenAI), plus structured output with Pydantic, streaming responses, and tool / function calling. Invoke this when implementing transcription workers (`./client/transcribe.py`), LLM context refinement or translation (`./client/llm.py`), or any code that needs to call OpenAI-compatible HTTP endpoints for chat completions or Whisper-compatible transcription using the openai SDK. Also use this when adding structured / JSON-mode outputs with response_format, streaming tokens to the subtitle UI, function calling for glossary lookup or branching logic, debugging API client setup, handling streaming responses, error handling with retries, or configuring base URLs for custom backends.

Guide for navigating the SubsVibe transcription reference implementations. Use this skill whenever working on the transcription server (`./server/server.py`, `./server/model.py`), the client transcription worker (`./client/transcribe.py`), designing API endpoints, handling audio input, WAV encoding, integrating model backends (Faster Whisper or Qwen3-ASR), or any question about how transcription should behave end-to-end. Also trigger when the user asks "how does X work in the reference" or "check the reference for Y".

Reference for py-webrtcvad, the Python wrapper around Google's WebRTC Voice Activity Detector (`import webrtcvad`, the `Vad` class, `is_speech`). Use this whenever working with webrtcvad: splitting PCM audio into speech vs. silence, segmenting recordings, gating a transcription/ASR pipeline on speech, building a frame_generator / vad_collector, or choosing an aggressiveness mode. Trigger it for the strict audio-format constraints (16-bit mono PCM at 8/16/32/48 kHz, frames of exactly 10/20/30 ms), for `is_speech` raising errors or returning garbage, for the sliding-window "triggered" segmenter pattern, and any time someone mentions webrtcvad, WebRTC VAD, or wiseman/py-webrtcvad — even if they don't name the exact API. Note this is the classic energy/GMM WebRTC VAD, NOT a neural model like Silero; pick the right tool when both are in play.