meeting-actions

If no job_id was passed as an argument: call list_jobs() to find the right recording, or process_media(path) if the user gave a file path — then follow the workflow below with that job_id.

You are taking minutes from a recorded meeting processed by talkthrough as job $ARGUMENTS. Audio-only jobs are expected here — frame tools are unavailable for them, and that is fine.

Method

1. get_transcript(job_id="$ARGUMENTS", format="segments") — walk the whole meeting (paginate via next_start_ms when truncated).
2. Multi-person meeting without speaker labels on segments? Re-run process_media(path=, diarize=true, num_speakers=) — it adds S1/S2/… labels to the existing job without re-transcribing, and minutes with owners need them.
3. Attendees are listed above and the file is not processed yet (or you are re-running process_media anyway)? Pass vocabulary="<the attendees' names>" in that call — names survive transcription instead of degrading into look-alike words, and owner attribution depends on them.
4. When segments carry speaker labels, map each label to a person before writing minutes. Evidence: self-introductions ("hi, this is Vera"), vocatives ("thanks, Tom"), the attendees list above — and, on video jobs, the SCREEN. The screen check is MANDATORY on video jobs, not optional: for EVERY label you map, call get_frames(job_id="$ARGUMENTS", at_ms=<that label's longest_turn_at_ms from the roster>) and read the meeting-app name plates, the recording's title card (who started/organized it), and the active-speaker highlight BEFORE asserting the mapping — vocative evidence alone mis-attributes (a remark spoken TO Tom is routinely tagged as Tom speaking). Name spellings: STT homophones lie (spoken "profit" vs on-screen "Prophet") — trust OCR/frames over the transcript for names. State the mapping first (e.g. S1 = Vera, S2 = Tom, S3 = unidentified) — never guess beyond the evidence. If the header carries speaker_names_pending_review, those names came from an older roster and are NOT active identities: use their stored context anchors to re-check each one. Confirm/replace only labels that are still current; remove stale labels with an explicit null patch.
5. Persist every defensible mapping with label_speakers(job_id="$ARGUMENTS", labels={"S1":""}, evidence={"S1":"<intro, frame, or attendee proof>"}). Never save OCR name_candidates directly: they are raw hints that may be UI text, a job title, or somebody else's name.
6. Collect: action items (who committed to what), decisions (what was agreed), open questions (raised but unresolved). Keep exact quotes and t_ms for each.
7. search(job_id="$ARGUMENTS", query="") to trace scattered follow-ups on one topic before summarizing it.
8. If the job has video (a screen-share was recorded), attach visual evidence to items that reference the screen via get_moment(job_id="$ARGUMENTS", start_ms=..., end_ms=...).