movie-ingest

Use this skill whenever we need cinematic emotional exemplars (respect, humor, regret, anger, rage, etc.) for PersonaPlex.

It enforces a subtitle-first workflow:

1. Search NZBGeek for scene releases that already include .srt/.sup files with stage directions.
2. Map rage/anger windows from the subtitle text or tags (no decoding needed yet).
3. Clip + Transcribe targeted scenes, marry them with the subtitle cues, and emit Persona-ready JSON.

Commands

run.sh search "<title>"

• Calls NZBGeek’s API (requires NZBD_GEEK_API_KEY) and prints size/date/link.
• Use this to locate releases that explicitly mention subs, CC, or SDH.

run.sh scenes find --subtitle <file.srt> (--query "text" | --tag laugh)

• Quick TUI view of subtitle windows that match free text or cue tags (rage, shout, laugh, etc.).
• Pass --video movie.mkv to see ready-to-copy ffmpeg -ss ... -to ... snippets.
• Designed for “project agent wants a specific sequence” moments.

run.sh scenes extract --subtitle <file.srt> --tag rage --video movie.mkv --clip-dir ./clips

• Writes a PersonaPlex-ready manifest that lists each candidate scene, cue tags, inferred emotion, and clip timing.
• Optional --clip-dir automatically runs ffmpeg to produce numbered .mkv clips plus .srt subtitle snippets per match.
• Use this when we need every rage/anger sequence from a film—no guessing or manual scrubbing.

run.sh transcribe movie_clip.mkv --emotion rage --scene "..."

• Extracts mono WAV via ffmpeg, runs Whisper (default medium), and joins the output with the matching .srt.
• Produces *_persona.json containing:
  • Stable video_id, movie title, scene notes, ISO timestamp.
  • Segment-by-segment tags (laugh, rage, whisper_candidate, etc.).
  • Rhythm metrics (WPM, pause count, duration).
• Pipe straight into horus_lore_ingest.py emotion --input <dir> --emotion rage.

Do we still need the movie if we have subtitles?
Yes for PersonaPlex: subtitles help us find the right windows, but Whisper + rhythm features require the actual audio to model tone, pacing, and breath. Use scenes extract --subtitle-only when scouting, then transcribe once the clip is cut.

Environment

| Variable | Purpose | |----------|---------| | NZB_GEEK_API_KEY (or NZBD_GEEK_API_KEY) | Auth for NZBGeek search | | NZB_GEEK_BASE_URL (or NZBD_GEEK_BASE_URL) | Override for private Usenet proxies | | WHISPER_BIN | Path to Whisper CLI (defaults to ~/.local/bin/whisper) | | FFMPEG_BIN | Path to ffmpeg (defaults to /usr/bin/ffmpeg) | | AUDIO_RMS_THRESHOLD | Base RMS threshold for audio intensity tags (default 0.2) | | AUDIO_RMS_WINDOW_SEC | Window size in seconds for RMS scanning (default 0.5) |

Audio intensity tagging automatically installs numpy + soundfile via uv. If either import ever fails, the CLI logs a warning and continues without those tags so ingestion still works.

Workflow Example

# 1. Find a release that ships with SDH subtitles
cd /home/graham/workspace/experiments/pi-mono/.pi/skills/movie-ingest
./run.sh search "There Will Be Blood"

# 2. Extract rage windows from the downloaded subtitles
./run.sh scenes extract \
  --subtitle "/home/graham/Downloads/media/movies/There Will Be Blood (2007)/blood.en.srt" \
  --tag rage \
  --video "/home/graham/Downloads/media/movies/There Will Be Blood (2007)/There.Will.Be.Blood.2007.PROPER.1080p.BluRay.x264-ELBOWDOWN.mkv" \
  --clip-dir "./clips/there_will_be_blood_rage" \
  --output-json "./clips/there_will_be_blood_rage/scenes.json"

# 3. Transcribe one of the clips with subtitles + metadata baked in
./run.sh transcribe ./clips/there_will_be_blood_rage/clip_01.mkv \
  --subtitle ./clips/there_will_be_blood_rage/clip_01.srt \
  --emotion rage \
  --movie-title "There Will Be Blood" \
  --scene "Most explosive confrontation" \
  --characters "Daniel Plainview,Eli Sunday"

Sanity

./sanity.sh   # Verifies Typer CLI + scenes helpers load