Watch a Google Drive folder for new audio or video recordings and stage a transcript + summary note for each. Polls the folder via the gdrive file.list operation, picks the newest media file (audio/* or video/*), downloads its bytes straight into the warehouse CAS as a Gateway-gated data.file (file.download surfaces result.file_ref, so the bytes never flow through op-step values). A video is run through the local ffmpeg tool first to extract a compact audio track; an audio file goes straight to the local whisper tool, which transcribes it into plain text captured as a second data.file ref; core-ai-summarize reads that transcript via the ref (a document content part, so it needs a document-capable model) and returns a short summary plus key points, steered by the focus variable. The media bytes and the transcript text never flow through op-step values. Preview-first: by default it shows the summary and writes nothing; set write_mode to stage to persist a durable staged note (summary, key points, and pointers to the extracted audio and the transcript files) into data.shared for later review. A per-run context guard skips re-processing the same newest file across polls. This is the Drive-folder source lane for the media-transcribe toolkit; the manual single-file lanes are transcribe-audio-to-notes and transcribe-video-to-notes.