Agent playbook

Any agent with a shell. Do not watch the Space in a browser to listen and type.

1. Take a URL. Accept x.com/i/spaces/<id>, x.com/<user>/status/<id>, or twitter.com equivalents.
2. yt-dlp -f "bestaudio/best" --hls-use-mpegts -o replay.%(ext)s "$URL". Do not remux. If yt-dlp finds no media, stop — the Space was not recorded or the replay expired.
3. Parse Hydra ID3 from the raw download (space_tape/hydra.py). This is who was unmuted, ~1 Hz, with handles.
4. Convert a copy for ASR: ffmpeg -i replay.m4a -ac 1 -ar 16000 space.wav.
5. Run Hugging Face Whisper with return_timestamps=True (word-level if you want karaoke).
6. Merge: each Whisper cue gets the Hydra speaker covering its midpoint.
7. Write cues.json, transcript.md, audio.mp3. Always include the original X URL.
8. Do not watch the Space in a browser to "listen and type." Whisper is the ear. Hydra is the speaker list.

space-tape transcribe "$URL" -o ./out

When an agent does need a computer

Only if step 1 or 2 fails.

1. Open the Space replay on x.com.
2. Play at 1x (unmute UI is not reliable at 16x).
3. Screenshot or DOM-watch the speaker list / pulsing avatar whenever it changes.
4. Log { t: audio.currentTime, handles: [...] }.
5. Merge that log the same way you merge Hydra regions.

Treat this as speakers.json from a different sensor. Same merge code.

Critical

• Never transcode before parsing ID3.
• Host is not in HydraParticipants. Host is slot 0 of HydraAudioLevel.
• Replays expire. Download is the archive.
• Public recorded Spaces only. Keep the link to X.