[davior]

Why

When the assistant picked clips, all it saw was the search result: one moment per video, the best match for its query words (a single transcriber segment plus a 240-character snippet). Transcriber segments end wherever the speaker pauses, often mid-sentence, so the cuts started and stopped mid-thought. Any trim the assistant wrote itself was a guess, because it had never seen the rest of the transcript.

What changes

• read_transcript(gam_id, start?, end?) is a new assistant tool that calls GAM's existing GET /api/assets/{id}/transcript.
  • It regroups the words into sentences, one line each: [start-end] text.
  • It labels speakers when there is more than one, and treats a segment corrected by hand in GAM as its text.
  • Long transcripts come back in pages of about 30k characters, each ending with the next start to read from.
  • Each transcript is fetched at most once per turn.
• Clip-choosing workflow in the prompt. Search to find candidate videos, read the most promising transcripts (several at once), choose the passage that best demonstrates the idea, then set_visual with a trim taken from the transcript lines.
• The automatic trim from a hit grows to the whole sentences it falls in. Growth is capped at 12 s each way. The usual 0.5 s padding now stops where the neighbouring sentence begins. If no transcript is available, the trim behaves as before.
• more_moments on search hits passes through GAM's other_matches, so the assistant knows the video matched in other places too.

No GAM changes are needed.

Testing

• pytest -q: 712 passed. New tests cover sentence regrouping, speaker labels, paging and caching, untranscribed and unknown assets, hand-edited segments, sentence-snapped trims and more_moments.
• npm test && npm run build: 187 passed and the build succeeds (no frontend changes).

---

Generated by Claude Code