[davior]

Why

Since #24 the assistant can read whole transcripts, but it reads them inside its own conversation. Everything it reads there is sent again on every later round, so weighing several long interviews gets expensive quickly. The model also has to judge passages while it is busy writing a video.

What changes

• find_clips(gam_id, idea, count=3) is a new assistant tool. It hands one video's whole transcript (numbered sentences) and the idea a scene needs to a separate model call with no tools, using the user's own provider. Only the best passages come back to the request, each with:
  • a ready trim for set_visual,
  • what is said,
  • a score from 1 to 10,
  • one sentence on why it fits.
• The picker names line numbers, never times. Times come from the transcript, so a cut can't land on a moment the model made up. Passages naming lines the picker wasn't shown, or longer than 90 s, or scored below 4, are dropped.
• Long videos are read in overlapping windows of about 15k tokens. The best passages across all windows are kept, with overlapping ones removed.
• Cost controls:
  • Picker calls are capped per request with ASSISTANT_MAX_CLIP_PICKS (default 16). Over the cap, the tool tells the model to use read_transcript instead.
  • Results are cached per video and idea within a request.
  • Picker calls are billed to the request like any other model call.
• Prompt workflow: search for candidates, find_clips each, compare across videos, then set_visual with the chosen trim. read_transcript stays for checking context, or as a fallback.
• Shared padding: the padding logic (_padded) is shared by the automatic hit trim and the picked clips.

No GAM changes are needed.

Testing

• pytest -q: 719 passed. New tests cover:
  • ready trims and what the picker is shown,
  • caching,
  • rejecting lines the picker didn't read,
  • windowed reading of long videos,
  • the call cap,
  • missing picker or transcript, and a bad count,
  • de-duplicating overlapping passages,
  • a whole request through the worker, checking the pass runs as its own tool-less conversation and is billed to the request.
• npm test && npm run build: 187 passed and the build succeeds (no frontend changes).

---

Generated by Claude Code