The /openai-whisper result can land as a dict or a JSON string depending on how the worker stored it. Normalise to a dict with `duration` and `seg

MuAPI result shapes vary by endpoint — try common keys.

gpt-5-4 sometimes wraps JSON in markdown fences — strip and parse.

Return the .srt cache path for a media file.

Drop a highlight if it overlaps >50% with a higher-scoring one already kept.

Main entry point — returns {highlights: [...]} sorted by score. `llm_fn` swaps the underlying LLM. Defaults to MuAPI gpt-5-mini; local mode p

ffmpeg -ss start -to end → re-encoded mp4 with audio.

Return a cached download path if we already have this YouTube id.

MuAPI wraps results inconsistently across endpoints. Hunt for the verbose_json blob (which has `segments` + `duration`).

Best-effort extraction of a YouTube video id from a URL.

Map our '720' / '1080' shorthand to a yt-dlp format selector.

Parse '9:16' → 9/16, '1:1' → 1.0.

Crop the cut clip to the target aspect ratio, tracking faces if possible.

Return a local filesystem path if the input already points at one.

Normalize model output into the expected shape; skip invalid entries.

Gemini backend used by --mode local when LLM_PROVIDER=gemini.

OpenAI Chat Completions backend used by --mode local.

Submit one autocrop job and return the URL of the rendered short.

Cut + reframe one highlight, returning the local mp4 path.

Crop every highlight, attaching the resulting URL back onto the dict.

Hand a YouTube URL to MuAPI; return a hosted mp4 URL we can read from.

Download a remote URL or return a local file path unchanged.

GET the latest result for a request_id; retry on transient timeouts.

Run the full pipeline and return a structured result. Args: youtube_url: source URL. num_clips: how many shorts to render.

Block until the prediction is done; return the final payload.

POST to /api/v1/{endpoint} and return the request_id; retry transient errors.

Run MuAPI /openai-whisper on a hosted media URL. Returns {duration: float, segments: [{start, end, text}, ...]} so it slots straight into the

Run faster-whisper on a local file path, caching the result as .srt.

Dispatch to the configured local LLM provider.

Default LLM backend: MuAPI gpt-5-mini.

Submit then poll. Returns the final result payload.