KodaCut

开源、可审计的 AI 口播剪辑 Agent。

KodaCut is a reusable Codex skill for turning raw Chinese talking-head footage into a reviewed A-roll, compact captions, evidence-driven B-roll, designed HUD cards, and a final 1.25x delivery.

It preserves the real camera background, uses compact punctuation-free captions with yellow keyword emphasis, adds evidence-driven B-roll and cyan/yellow HUD cards, and delivers the approved master at 1.25x with pitch-preserved audio.

Why KodaCut

• 先审粗剪，再做包装，避免把时间浪费在错误版本上。
• EDL 使用视频帧与音频采样点作为唯一时间真相，结果可复查。
• 自动防止字幕过度改写、英文拆词和未经确认的超长字幕。
• 保留真人背景，支持证据型 B-roll、重点词高亮和 HUD 信息卡。
• 所有核心门禁都可以离线运行，不绑定某个云端剪辑平台。

Install

Clone or copy this repository to:

%USERPROFILE%\.codex\skills\kodacut

Restart Codex after installation if the skill is not immediately visible.

Use

Invoke the template explicitly:

$kodacut

Then provide a talking-head source video and ask Codex to apply the saved edit template.

Modes

• Full template: rough cut, hash-bound A-roll approval, guarded transcript correction, validated semantic caption lines, keyword emphasis, evidence B-roll, HUD cards, QA, and final 1.25x delivery.
• Retime only: apply the final 1.25x speed-up to an already approved edit without changing visuals or captions.

See SKILL.md and references/style-spec.md for the complete editing contract.

Deterministic workflow gates

The upgrade inspired by yh-chatcut-skills adds two standard-library tools:

• scripts/aroll_approval.py prevents expensive packaging work from starting before the exact A-roll EDL and preview are approved.
• scripts/caption_plan.py rejects over-aggressive transcript correction, non-semantic rewrites, overlong unreviewed captions, and line breaks inside protected English or numeric tokens.

Both tools write atomically and refuse to overwrite an existing artifact unless replacement is explicitly requested. Run their offline checks with:

python scripts/caption_plan.py self-test
python scripts/aroll_approval.py self-test

The JSON contracts are published under schemas/, and GitHub Actions runs the offline gate tests on every push and pull request.

Digital-human roadmap

The repository also contains the development-ready PRD for extending this editing template into a local-first digital-human video agent:

• Koda Digital Human Video Agent PRD

The proposed MVP uses AIGCPanel as the local model gateway, LatentSync 1.5 as the primary avatar engine, MuseTalk 1.5 as the local fallback, Alibaba Cloud EMO as the low-cost cloud fallback, and HeyGen only as the premium-quality fallback. These roadmap capabilities are not yet part of the installed kodacut skill.