render-edl

render-edl

Purpose

Execute an edl.json (Edit Decision List) into a finished master.mp4. The EDL encodes all creator-grammar decisions — clip order, pacing, transitions, audio mix, and captions — produced by generate-edit-decision-list. This atom is the assembly step: it reads the contract and renders it deterministically without any interactive prompts.

Supports both source modes:

• conform (source.mode = "file") — trims ranges from existing video files.
• generative (source.mode = "generate") — locates pre-rendered scene clips under <project_dir>/clips/<id>.mp4.

Inputs

• edl (required) — path to edl.json. Must conform to the EDL schema (schema version 1.0).
• output (required) — path for the output master.mp4.
• project_dir (optional) — base directory for resolving relative asset paths. Defaults to meta.project_dir from the EDL; falls back to the directory containing edl.json.
• crf (optional, default 18) — x264 quality setting; lower = higher quality.

Workflow

1. Load and validate EDL — parse edl.json, check timeline is non-empty and contiguous (each clip's target_out == next clip's target_in within ±1 frame), confirm exactly one payoff-hold role, and that meta.target_duration is within ±0.5s of the last clip's target_out.
2. Resolve and trim clips — for each timeline[] entry:
   • source.mode = "file": trim file from in to out with ffmpeg (re-encoded, frame-accurate).
   • source.mode = "generate": locate <project_dir>/clips/<id>.mp4; if missing, FAIL listing every missing asset.
3. Apply speed — if speed != 1.0, apply setpts=PTS/<speed> (video) + chained atempo (audio).
4. Apply motion — support zoom-punch (stepped crop/scale to the to zoom factor) and push-in (gradual zoompan from from to to). Null motion is a no-op.
5. Normalise — scale every clip to 1080x1920 (9:16) or the appropriate resolution for meta.aspect at meta.fps.
6. Concatenate — hard-cut concat of all processed clips using ffmpeg concat filter. Non-hard-cut transition_in values are treated as hard cuts in v1; a note is written to manifest.json.
7. Build audio bed — mix audio.music + audio.vo + per-clip SFX (placed at sfx[].at on the master timeline) + apply each audio_beats[] silence-beat by muting music at at for dur seconds. When audio.ducking = "sidechain-20:1", compress music under VO with ffmpeg sidechaincompress.
8. Burn captions — build an ASS subtitle file from all timeline[].captions[] cues using the preset from caption-style-presets.json. static-sentence holds the full cue; karaoke-pop renders per-word highlights by splitting the cue text into equal-time words. Falls back to grammar.caption_preset when a cue has no per-cue preset.
9. Mux and deliver — combine video, audio, and subtitle tracks into output (master.mp4). Write manifest.json to the same directory, including a recomputed pacing_curve.

CLI:

python3 skills/atoms/assembly/render-edl/scripts/render_edl.py \
  --edl <edl.json> --output <master.mp4> [--project-dir <dir>] [--crf 18]

Output

• master.mp4 — finished video at meta.fps, normalised resolution, with audio and burned captions.
• manifest.json (sibling to master.mp4) — run metadata including:
  • skill_name, run_id, edl_path, output, created_at, duration_seconds
  • clips_rendered — per-clip records: id, source_mode, file (resolved path), duration, status
  • transitions_logged — list of non-hard-cut transition_in values treated as hard cuts
  • pacing_curve — recomputed mean_shot_len, cuts_per_10s, payoff_hold_ratio
  • status — pass | fail | needs human review
  • warnings, errors

Quality Checks

• master.mp4 exists and is non-zero; ffprobe reports both video and audio streams.
• Duration of master.mp4 is within ±0.5s of meta.target_duration.
• Resolution matches the expected value for meta.aspect (1080×1920 for 9:16).
• All caption cues are present in the ASS file with correct timing.
• manifest.json exists with status: pass and all clips_rendered paths resolve on disk.
• pacing_curve.payoff_hold_ratio is between 2.0 and 4.0 (warn if outside range, don't fail).

Failure Modes

• EDL validation failure — non-contiguous timeline, missing payoff-hold, or target_duration mismatch; script exits with a clear message listing every violation.
• Missing generate-mode clips — script fails immediately listing every <project_dir>/clips/<id>.mp4 that does not exist. Re-run after generating the missing clips.
• Missing conform-mode source file — script exits with a clear error naming the file.
• ffmpeg not installed — install via brew install ffmpeg or apt install ffmpeg.
• No audio streams — if audio.music, audio.vo, and all SFX are null/absent, the master is produced with a silent audio track; a warning is written to the manifest.
• Caption preset not found — if the cue's preset does not exist in caption-style-presets.json, the fallback preset platform-default is used and a warning is written.
• Speed out of range — speed values < 0.1 or > 8.0 are rejected with a clear error.