stream-transcript

Stream transcript

Feed a live meeting transcript into the current session while the meeting is still happening, so the agent holds the conversation as context rather than reading it afterwards.

This is the live, advisory half of meeting capture. The authoritative half is unchanged and still owned by Tyto + analyze-meeting: full-file, diarized, stored in Neotoma, reconciled against the graph.

Relationship to the rest of the pipeline

	Live stream (this skill)	Post-hoc (Tyto → analyze-meeting)
When	While recording	After the file settles
Engine	Whisper only, no diarization	Best available, diarized
Boundaries	Arbitrary N-second cuts	Whole file
Speakers	System track only — not the operator's mic	Both tracks, merged
Neotoma	Nothing durable	Authoritative entities
Purpose	Orient the agent mid-meeting	Metabolize the meeting

Live chunks are provisional by construction. They are cut mid-sentence, they garble non-speech, and they never see the operator's own voice. Treat every live claim as something the authoritative pass will restate better.

Recorder control: detect only

This skill does not start or stop Audio Hijack. Audio Hijack 4 exposes scripted control via App Intents (AH4RunScriptIntent), and this machine already has a Tyto session plus start_tyto_* / stop_tyto_test scripts and Shortcuts — but as of 2026-08-28 all three Shortcuts fail with "An unknown error occurred" and change no recording state, likely because scripted control requires confirmation (confirmAH4RunScript:completion:). Until that is resolved, the operator starts and stops the recorder; this skill detects what is running.

Note: shortcuts run exits 0 even when the shortcut fails. Never trust its exit code — parse its output.

---

/stream-transcript start

1. Detect a growing recording

cd "$ATELES_REPO" && execution/venv/bin/python - <<'PY'
import sys; sys.path.insert(0, "execution/scripts")
import live_transcript_tail as lt
rec = lt.find_growing_recording(lt.DEFAULT_DIR)
print(rec if rec else "NONE")
PY