Video Factory

Plugin marketplaces

This plugin belongs to AIGC content creation.

Category	Marketplace	Purpose
Full-stack development	Full Stack Plugins	Architecture and UI design, code understanding, quality checks, code review, workflow governance, and server operations
AIGC content creation	Full AIGC Plugins	Image, video, audio, music, 3D, and multimodal content creation

Analyze shots, plan an edit, approve it by hash, render locally with FFmpeg, and hand back a receipt for every output file — no cloud service and no API key.

English | 简体中文 · Install · Quick start · Skills · Troubleshooting

Positioning

video-factory turns existing footage into an approved, verifiable edit. The host agent analyzes shots, proposes a closed EditDecision, quotes the stage, waits for your approval, renders through local FFmpeg, and records media receipts for every output.

The repository keeps the canonical video spelling across repository, plugin, package, CLI, and skill identifiers. There is no cloud service, no API key, and no vendor upload.

Property	Value
Plugin ID	video-factory
Version	0.5.1+codex.20261018
Maturity	Local composition is stable; the recorded runtime acceptance covers the 0.1.0 line
Runtime	Node.js 18+, FFmpeg and ffprobe on PATH
MCP server	None; the plugin exposes a skill-driven CLI instead

The 0.5.0 line has no fresh runtime acceptance record in this repository. Treat recorded media evidence as line-specific and re-verify before claiming a release.

Installation

Prerequisites

• Node.js 18 or newer on PATH.
• FFmpeg and ffprobe on PATH; required for every probe and render.
• Chrome or Chromium only for the enhanced synchronized review video.
• Your footage as local regular files beneath --input-root; the plugin never fetches a URL.

From the plugin marketplace

codex plugin marketplace add partme-ai/full-aigc-plugins
codex plugin add video-factory@full-aigc-plugins

From the repository-local marketplace

codex plugin marketplace add full-aigc-plugins/video-factory-plugin --ref v0.5.1
codex plugin add video-factory@partme-ai-video-factory

Restart Codex or the ChatGPT desktop app and open a new task so the skills load. There is no build step and no npm install: the CLI has zero runtime dependencies.

For development, clone the repository and run bin/video-factory probe; it reports Node, FFmpeg, and ffprobe availability plus whether the enhanced review path is usable. codex plugin list should show video-factory@partme-ai-video-factory installed, enabled.

Quick start

1. Probe the footage, then let the host annotate the shot analysis. Finalize it with the measured track and keyframes:

bin/video-factory analyze source.mp4 --out reelbench-analysis
bin/video-factory analyze-finalize reelbench-analysis/shots.json \
  --track reelbench-analysis/track.json --frames reelbench-analysis/frames

2. Validate the edit decision, quote the rough cut, and approve that exact revision:

bin/video-factory validate-plan video-plan.json
bin/video-factory quote video-plan.json --stage rough
bin/video-factory run video-plan.json --stage rough --approval rough-approval.json

3. Build the synchronized review, quote and approve the final cut, then record your acceptance:

bin/video-factory review-sync output/rough.mp4 reelbench-analysis/shots.json
bin/video-factory quote video-plan.json --stage final
bin/video-factory run video-plan.json --stage final --approval final-approval.json
bin/video-factory accept final-job.json --decision approved --note "played and accepted"

An approval binds the stage, plan hash, edit hash, round, and quote revision. Any change to an asset, an edit, or an output specification invalidates the old approval. The rough cut and the final cut are approved separately, and a rejected stage returns to planning.

What you can build

Workflow	Built-in path	Verifiable output
Break footage into shots	analyze, analyze-finalize, video-shots	shots.json, track, keyframes, and a shot report
Assemble a rough cut	validate-plan → quote → run	Resumable H.264/AAC MP4 with per-segment receipts
Review picture against shot data	review-sync	Internal review MP4 with synchronized shot data
Finish with audio, subtitles, and titles	final stage run	Multi-track final MP4 plus embedded subtitle track
Judge and accept	evaluate, accept	Three-state media gates and a recorded human decision
Resume an interrupted render	recover	Pending segments resumed without re-rendering completed ones
Slice a talking-head recording	episode-slice, episode-roughcut	Rough-cut MP4, aligned SRT, and cutlist.json

Skills

Skill	Purpose
video-factory-use	Routes an editing request to the correct workflow
video-factory-plan	Converts local assets and shot evidence into a validated EditDecision
video-factory-run	Renders an approved stage and records receipts without automatic retry
video-factory-judge	Runs media gates, advisory semantic review, and human acceptance or rework labels
video-factory-recover	Reads the durable ledger and determines the only legal next action
video-factory-harness	Calling spec for the CLI surface and the rough → review → final approval gate
video-episode-slicing	Word-level ASR timeline → cut points → rough cut, SRT, and cutlist
video-shots	Vendored ReelBench shot breakdown with deterministic evidence
video-sync	Vendored annotated review video that pairs picture with shot data

The video-factory-use/plan/run/judge/recover skills are vendored from video-factory-skills v1.0.2 through skills.lock.json. video-shots and video-sync are ReelBench snapshots pinned by Git blob and SHA-256 in upstream/reelbench.lock.json; video-episode-slicing and video-factory-harness are plugin-local. Edit none of these copies in place.

Slash commands

Command	What it does
/video-factory	Total entry point; routes by capability
/video-factory-plan	Builds the timeline and edit decision
/video-factory-run	Renders the approved edit and collects receipts
/video-factory-judge	Reviews finished media and records the verdict
/video-factory-recover	Resumes an interrupted composition job
/video-episode	Slices and rough-cuts a talking-head recording

Configuration

Setting	Where it lives	Notes
Job ledger	<plan>.job.json, or the --ledger path	Plain JSON with atomic per-segment updates
Input root	--input-root	Only regular files beneath it are accepted
Render outputs	The output directory you choose	Never overwritten silently
Review assets	The analysis directory	Keyframes and contact sheets stay local
Credentials	None	The runtime never reads an API key or token

Architecture and security

flowchart LR
    A[Footage] --> B[analyze / probe]
    B --> C[EditDecision]
    C --> D{quote + approval}
    D -->|approved| E[FFmpeg segments]
    D -->|rejected| C
    E --> F[Assemble audio + subtitles]
    F --> G[Media gates]
    E -->|interrupted| H[(Job ledger)]
    H --> E

Component	Owns	Does not own
src/cli.mjs	Command dispatch, exit codes, argument validation	Media processing
src/orchestrator.mjs	Run, approve, and recover sequencing	FFmpeg argv construction
src/ffmpeg-compiler.mjs	Deterministic argv and content-addressed segment keys	Approval decisions
src/job-ledger.mjs, src/approval.mjs	Durable state, resume points, hash-bound approvals	Cost estimation
skills/ (9)	Routing, planning, review, and recovery instructions	Runtime behaviour

The plugin makes no network call for generation; local FFmpeg is the only execution engine. Approvals are single-use and bound to content hashes, so they cannot be replayed against different content. Input containment rejects paths outside --input-root, path traversal, symlink escapes, special files, arbitrary FFmpeg filter graphs, and credential fields. All child processes run with shell: false.

Troubleshooting

Symptom	Check first	Resolution
probe reports a missing tool	FFmpeg and ffprobe on PATH	Install FFmpeg and re-run probe
Plan validation fails	EditDecision schema	Fix the named field; unknown fields are rejected
Run refuses to start	The approval file	Re-quote the stage and approve the new revision
Render stops mid-way	Ledger state	Run recover; completed segments are never re-rendered
Review video is unavailable	Chrome presence	Install Chrome, or skip the enhanced review
Subtitle burn-in is missing	The local FFmpeg build	The verified FFmpeg 9.0.1 lacks libass; the step is recorded as SKIPPED
Semantic consistency is unproven	Gate status	It stays NOT_RUN; Codex or a human reviews it
Verification is missing for your platform	Recorded evidence	Media acceptance was recorded on macOS; re-verify on your platform
A remote generation path was expected	Plugin scope	Native video generation is blocked by design

Upgrade

codex plugin marketplace upgrade partme-ai-video-factory
codex plugin add video-factory@partme-ai-video-factory

Contributing and support

Open functional issues at https://github.com/full-aigc-plugins/video-factory-plugin/issues. Before proposing a change, state the target Node version and whether it alters the approval binding or the ledger format, and include the affected tests.

License

Apache-2.0 — see LICENSE and THIRD_PARTY_NOTICES.md. Vendored ReelBench skills keep their upstream terms.