Human version

Replace only the opening of a selected video. The rest comes from the original, including its ending and burned captions. A supplied hook uses free local assembly; short text can become a free animated card. Generated footage or voice is optional and requires a separate estimate and approval.

The output is a complete MP4 and an edit manifest identifying the exact original, cut, duration change and preservation checks. It stays needs review until the finished video has been watched and its sound checked.

---

Agent version

Purpose

Perform one hard-cut opening replacement with immutable source identity and decoded retained-body verification. There is no fixed three-second boundary, mandatory creator, new music, body regeneration, whole-video caption pass or automatic final selection. Read the editing guide.

Inputs

• Python 3.10+, working ffmpeg/ffprobe with libx264 and AAC. Test the binaries themselves before any paid step. Supplied-clip execution uses only Python stdlib.
• Config JSON following config.example.json. Paths resolve relative to that config, so a fetched package works outside the author's checkout.
• source.path: exact selected original, never a template sample. Optional source.sha256 refuses changed bytes; project_id/render_id are opaque caller identifiers preserved in the manifest. Compute and freeze the hash at intake.
• hook_end_sec: positive opening boundary leaving a body. If speech crosses it, choose a measured word/silence boundary. No silent correction of that boundary.
• hook.path: supplied video, any dimensions/frame rate, normalized to the original. Or hook.text, duration_sec, font_path and optional #RRGGBB colors. Text needs Pillow (part of the video toolchain) and an explicitly supplied licensed font.
• Optional words:[{start,end,word}] measured on this source, or word_times JSON containing source_sha256 and words. A cut through a measured word fails.
• Optional captions:{path,output_path}: original separate SRT or timed JSON cues. Complete retained cues shift by the measured duration delta. A cue crossing the cut requires measured words, so removed opening copy does not linger over the body.
• output:{path,manifest_path?}: a new .mp4 path. Existing outputs and every input are protected from overwrite. The default manifest is <output>.manifest.json.

The original must be CFR, unrotated square-pixel 8-bit yuv420p with zero video start time. Other sources fail before assembly. If normalization is necessary, make it a separately approved source version and freeze that version; do not quietly remaster the selected original. Audio is optional. An original audio/container tail more than 2ms beyond the last video frame also blocks; make a reviewed source version with a final-frame hold before editing. The renderer never trims that narration/music tail and calls it preserved.

Workflow

1. Watch the exact original and confirm source identity, hook boundary and selected replacement. Keep the original file read-only. Ask only for missing choices.
2. Probe source/hook, validate SHA256, measured words and captions before encoding. Before creating any new hook, run python3 scripts/replace_hook.py --preflight --config /path/to/source-config.json. This needs only source/cut and optional words/captions, verifies runtime encoders and supported source shape, and writes no media or paid job. Generated hooks are separate work: read the chosen generation capability, quote only the required new opening, obtain approval, then bind its finished clip here.
3. Run python3 scripts/replace_hook.py --config /path/to/config.json. The script normalizes only the hook geometry/fps; joins original body at normal speed; uses browser-compatible High-profile libx264 CRF 12; retains original body audio with AAC 256k; shifts separate captions; and checks source integrity before publishing files.
4. Examine the manifest. All four required preservation booleans must be true. Every retained decoded video frame is compared with PSNR ≥45dB and identical frame count/order. Audio is compared at identical sample positions with normalized RMS error ≤0.08 and correlation ≥0.99 (silence uses absolute RMS error ≤0.0001). These are encode tolerances, not a listening verdict.
5. Watch the actual complete output, especially the join and final frame; listen to the hook/body join, narration/music and ending. Check the new hook copy and brand facts. Fix only the opening and re-run to a new path. Keep the result awaiting review until the review report names this output hash and source. Only after actual watch/audio/caption review, record review:{status:'passed',source_sha256:<original hash>,output_sha256:<finished hash>,checked_at:<ISO datetime>} alongside the passing quality report. Never recycle another output's review.

Free demonstration

From a fresh package folder:

python3 scripts/make_demo.py --out-dir demo
python3 scripts/replace_hook.py --config demo/config.json
python3 -m unittest discover -s tests -v

The neutral fixture replaces 2.4s with 1.5s and retains the original body/end. --font /path/to/licensed.ttf adds burned captions to the fixture. Set HOOK_TEST_FONT=/path/to/licensed.ttf to test the optional kinetic text route too. The demo makes no network requests or paid calls. It is a timing fixture, not a customer ad or catalog upload. High-quality retained-body encoding can produce large files.

For a local pack, use python3 scripts/replace_pack.py --config pack.json with the same source, hook_end_sec, optional words/captions, output_dir and variants:[{label,hook:{path}}]. The wrapper freezes one source hash and writes one complete MP4/manifest per label plus pack-manifest.json. Successful siblings remain available if one fails; retry the failed target with the single renderer.

Output

• Complete MP4; original untouched.
• Version 1 manifest with source.{sha256,project_id,render_id}, hook_end_sec, body_start_sec, replacement_duration_sec, new_body_start_sec, duration_delta_sec, output.{path,sha256,duration_sec}, caption artifacts and verification.{source_unchanged,body_video_preserved,body_audio_preserved,caption_timing_preserved}.
• Separate shifted captions if supplied. Existing burned captions stay in the original body frames; no whole-ad reburning or redesign.
• status/review.status: needs_review. No project storage behavior is embedded in this generic renderer; the caller saves candidates and performs final review.

Quality Checks

• Frozen source/hash matches the selected version; old source and hook bytes remain intact.
• No body frame is missing, duplicated, reordered, graded or cropped.
• Original body audio stays at normal speed/gain, with no crossfade or added bed.
• Original ending survives; shifted captions follow the measured duration delta.
• Subframe boundaries preserve the requested cut; the manifest discloses the first selected original video frame and its offset (less than one source frame).
• Invalid inputs fail clearly; no paid work occurs inside this capability.

Failure Modes

Problem	Action
Wrong source hash or stale word file	Refetch the selected version and remeasure; never switch versions silently.
Cut clips speech or a separate caption	Choose a safe boundary; supply measured words for a partial caption.
VFR, HDR/10-bit, rotated or nonzero-start original	Create an explicitly reviewed normalized source version first.
Original audio outlasts video	Make a reviewed source version with a final-frame hold, then freeze that version; never drop the audio tail.
Missing/corrupt media or LFS pointer	Resolve real media before rendering or generation.
Existing output path	Use a fresh version path; never overwrite the original.
Local ffmpeg has no drawtext	Text uses Pillow PNG + fade/slide overlay; drawtext is not required.
Retained-body verification fails	Candidate is not published; inspect the failed video/audio metric and original remains available.
Needed generated creator/voice unavailable	Keep supplied footage/text options; show missing provider and new estimate before a paid retry.

Related

Used by the create-hook-variant-pack recipe. Check each candidate with the watch skill. This renderer's own scripts handle the source preflight, the cut and the caption shift.