/cowatch — live event co-watching

Servy watches a live event WITH Gera in near-real-time. A background capturer (scripts/live_capture.py) turns the stream's audio into a rolling JSONL transcript via local whisper; this skill reads new transcript past a cursor and either reacts when something notable lands (co-watch) or answers on demand (catch-up). Near-real-time means ~1.5-2.5 min behind live — chunk length plus whisper time. Full design: docs/superpowers/specs/2026-06-02-cowatch-live-event-design.md.

Scope (v1)

• Source: web only (public YouTube/stream URLs via yt-dlp). system audio and mic are designed but NOT built yet — if Gera asks for those, say so and point to the spec.
• Engine: whisper.cpp (whisper-cli), models at ~/.local/whisper.cpp/models/ (base.en default, small.en for more accuracy / slower).

Hard rules

• Never fabricate. Only react to text that is actually in the transcript file. Auto-captions garble names/numbers — flag uncertainty, do not "correct" to what you assume.
• SoftServe boundary: v1 is web/public only, so it is unrestricted. (When system/mic land, confirm it is not a client/Internal+ meeting before capturing — see spec §3.)
• External publishing stays gated. If Gera wants a LinkedIn post from the event, draft it and show him first (voice rules); never auto-publish.

Step 0 — args

Parse <url> [--mode cowatch|catchup] [--model base.en] [--ingest] [--interval 60]. Default --mode cowatch, --model base.en, --interval 60, ingest OFF.

Step 1 — verify the stream + launch the capturer

1. Confirm it is a real live stream and a legit channel: yt-dlp --no-warnings --print "%(live_status)s|%(channel)s|%(channel_is_verified)s|%(title)s" <url>
   • If live_status is not is_live/post_live, tell Gera it is not live and stop.
   • Surface the channel + verified flag so Gera knows the source is trustworthy.
2. Derive a session id: <today>-<slug-of-title> (use live_capture.session_id / slugify).
3. Launch the capturer in the background (Bash run_in_background: true): python3 scripts/live_capture.py --source web --url "<url>" --session "<sid>" --model <model>
4. Wait until the first record appears in knowledge/live/<sid>.jsonl (capturer healthy). If the state file flips to status: error or no record appears, fail fast and report.

Step 2a — co-watch mode (it talks)

Drive an interval loop (use the /loop skill at --interval seconds, or self-pace). Each tick:

1. Read cursor: read_cursor(knowledge/live/<sid>.cursor).
2. recs, new_cursor = read_new_records(knowledge/live/<sid>.jsonl, cursor).
3. Judge notability through Gera's lens — react only if a record contains: an announcement, a product/feature/version name, a number/benchmark, a demo, or anything touching his stack (LangChain/LangGraph, Claude Code, Cursor, agents, Azure AI) or Jumpstart/R&D. If notable, post a short reaction with a why-it-matters-to-Gera line. If not, stay silent.
4. write_cursor(...) to the new cursor.
5. Max-quiet: if 3 ticks pass with no reaction, drop one light check-in so you never vanish.
6. Read knowledge/live/<sid>.state.json; when status == ended, leave the loop → Step 3.

Read the JSONL/state via small python3 -c "import live_capture as lc; ..." calls (cheap), not by loading the whole file into context.

Step 2b — catch-up mode (it waits)

No loop. Tell Gera capturing is live ("ask me anything — 'what did they just say about X', or 'summarize the last 10 min'"). On each question, read new records / the buffer and answer from the transcript only. Notice status == ended opportunistically and offer Step 3.

Step 3 — end of event

When the capturer reports status: ended:

• If --ingest: run the wiki-ingest tail (the nate-watch pattern) — write a digest page in references/, update references/index.md, append references/log.md, register the raw transcript in knowledge/README.md, run the build-a-capability gate. Then offer a LinkedIn draft.
• If not: save the transcript, tell Gera where it lives (knowledge/live/<sid>.jsonl), and offer ingestion as a follow-up. The transcript is gitignored (knowledge/*), so it stays local.

Files

• Engine: scripts/live_capture.py (tested: python3 scripts/tests/test_live_capture.py).
• Output: knowledge/live/<sid>.jsonl (transcript), .state.json (capturer state), .cursor (skill read-cursor), .segments/ (temp WAVs, auto-cleaned).
• To stop early: terminate the background capturer process; it flips status: ended on SIGTERM.