review-video-pacing-rhythm

name	review-video-pacing-rhythm
description	Detect cut points (ffmpeg scene-detect) and music beats (librosa via the beat-sync-reel skill); flag cuts > 100 ms off the nearest beat and scenes > 4 s with no internal motion change. Emits a pacing-report.md with a waveform overlay PNG.

review-video-pacing-rhythm

Purpose

Off-beat cuts are felt before they're heard — the eye registers the mismatch and the ad feels amateur. This atom is the deterministic check: every cut, every beat, every delta.

It also catches the other pacing failure mode — a scene that runs too long without internal change (static framing, no motion, no caption shift) — which kills retention.

Inputs

• <video_path> (required)
• <music_path> — the music bed audio (default: look in <video_folder>/audio/variants/<active>/music.mp3, fall back to extracting audio from the video)
• <output_dir> — default <video_folder>/review-notes/pacing-<ts>/
• <beat_tolerance_ms> — max acceptable cut-to-beat delta (default 100)
• <scene_drag_threshold_s> — flag scenes longer than this with no motion change (default 4.0)

Workflow

Step 1 — Detect cuts

Use ffmpeg scene-detect filter:

ffmpeg -i <video_path> -vf "select='gt(scene,0.4)',showinfo" -f null - 2>&1 \
  | grep showinfo | extract pts_time

Persist cuts.json = list of cut timestamps in seconds.

Step 2 — Detect beats

Use librosa (reuse the beat-sync-reel skill's setup — do not re-vendor):

import librosa
y, sr = librosa.load(music_path)
tempo, beat_frames = librosa.beat.beat_track(y=y, sr=sr)
beat_times = librosa.frames_to_time(beat_frames, sr=sr)

Persist beats.json = list of beat timestamps in seconds, plus tempo.

Step 3 — Compute deltas

For each cut, find the nearest beat. Record delta_ms = (cut_time - nearest_beat_time) * 1000.

• Cuts within ±beat_tolerance_ms → on-beat
• Cuts outside → off-beat (flagged)

Step 4 — Detect static-stretch

For each scene (cut-to-cut interval longer than scene_drag_threshold_s):

• Sample frames at 4 fps within the scene
• Compute mean absolute difference between consecutive frames
• If mean MAD < 2 (essentially a still frame) → flag as drag

Step 5 — Overlay PNG

Render pacing-overlay.png: x-axis = time, plot:

• Music waveform (decimated)
• Beat tick marks (vertical lines, alpha 0.3)
• Cut tick marks (vertical lines, alpha 0.8, green if on-beat, red if off-beat)
• Drag region rectangles (yellow, alpha 0.2)

Use matplotlib if available; otherwise emit a simpler ASCII summary.

Step 6 — Write report

# Pacing report — <video>

- **Tempo:** 128 BPM
- **Cuts:** 14 (12 on-beat, 2 off-beat)
- **Drag scenes:** 1
- **Overlay:** [pacing-overlay.png](./pacing-overlay.png)

## Off-beat cuts

| Cut | Time | Nearest beat | Delta | Severity |
|---|---|---|---|---|
| 7 | 8.43s | 8.59s | -160 ms | P1 |
| 11 | 15.21s | 15.00s | +210 ms | P1 |

## Drag scenes

| Scene | Range | Duration | Mean frame-diff | Note |
|---|---|---|---|---|
| 6 | 12.0–17.2s | 5.2s | 1.1 | "Static framing on bike for >5s; no caption shift" |

## Status

- ✅ APPROVED if 0 off-beat cuts AND 0 drag scenes
- ⚠ NEEDS REVISION otherwise

Output

• cuts.json, beats.json
• pacing-overlay.png (or pacing-overlay.txt if matplotlib unavailable)
• pacing-report.md
• manifest.json

Quality checks

• Tempo detected and stated; if librosa returns implausible BPM (< 40 or > 220), flag as tempo_uncertain.
• Off-beat cuts each have a delta_ms value; sign of delta tells you whether the cut is early or late.
• Drag scenes are only flagged when duration is over threshold AND mean frame diff is low (no false-positive on long busy scenes).

Failure modes

• No music bed (talking-head ad with no music) — skip beat detection; report only drag scenes. Manifest beat_check: skipped (no_music).
• Music drops mid-ad — beat detection still works but beats stop appearing; cuts past the music drop are unscoreable. Flag in manifest.
• Cuts are intentionally off-beat for effect — possible but rare in social ads. The skill flags; the operator decides.
• librosa not available — abort with pip install librosa instructions. Do not silently fall back to ffmpeg-only beat detection (less accurate).

Example

review-video-pacing-rhythm
  video_path=peloton/ads/video-03-ironman-comeback/edits/master-final.mp4
  music_path=peloton/ads/video-03-ironman-comeback/audio/variants/B_cinematic/music.mp3

Phase 7b (creative) of the new review-video.md.