FORGE

FORGE packages a portable Codex + Cowart image-generation workflow.

It includes:

• A modified Cowart plugin with the in-canvas AI image panel, draggable prompt panel, progress feedback, 4K/2K quality options, A-series and aspect-ratio presets, JPG/PNG output choices, image preview, image grouping, reference-image annotation, uploadable reference images, local replacement boxes, and brush thickness controls.
• A Codex skill, cowart-fast-annotated-imagegen, that claims a Cowart request, extracts prompt/annotation data, calls built-in image_gen, post-processes output, inserts the result back into Cowart, and reports a concise result.
• Helper scripts for precise red-marked local text cleanup when image generation would otherwise drift the full layout.
• Installers for Windows PowerShell and macOS/Linux shells.

Install On Another Codex

Clone the repository, then run one installer from the repository root.

Windows PowerShell:

Set-ExecutionPolicy -Scope Process Bypass
.\install.ps1

macOS/Linux:

chmod +x ./install.sh
./install.sh

The installer copies:

• cowart-plugin/ to $CODEX_HOME/plugins/cache/personal/cowart/0.1.2
• skills/cowart-fast-annotated-imagegen/ to $CODEX_HOME/skills/cowart-fast-annotated-imagegen

If $CODEX_HOME is not set, it uses ~/.codex.

Use

1. Open a Codex project.
2. Ask Codex: Open the Cowart canvas for this project.
3. In Cowart, select or add an image.
4. Use the annotation tool to mark a note, reference image, or local replacement box. After drawing a line or box, a dialog lets you add text, choose an existing reference image, or upload a reference image.
5. Open AI 生图, choose mode, quality, size, direction, format, background, and reference options.
6. Click AI 生图.
7. In Codex, manually confirm with a message such as 确认生成.

FORGE intentionally does not run a background automation. Codex only executes when the user explicitly confirms.

What Gets Transplanted

Cowart UI features:

• Prompt panel can be dragged by its header.
• Prompt panel exposes generate, edit selected image, cutout, and three-view modes.
• Quality options include standard, high, 2K, 4K, and draft.
• Size options include holder match, A4, A3, A2, 4:3, 16:9, 21:9, 3:2, 1:1, and custom.
• Export options include PNG and JPG.
• Background/cutout options include keep, white, transparent, and subject/person extraction.
• Annotation tool supports note, reference image, and local replacement box modes.
• After drawing annotations, a dialog supports text notes, existing canvas reference images, and image upload as a reference.
• Brush thickness can be changed from fine to extra thick.
• Selected images can be previewed, grouped, and selected as a group.
• Generation progress displays pending, running, inserting, done, and error states.

Codex execution features:

• Request claiming through /api/generation-request/claim.
• Prompt extraction from the Cowart request file, current selection, annotations, local replacement regions, and reference images.
• Built-in image_gen workflow with post-processing for 2K/4K, JPG/PNG, and transparent outputs.
• Cowart MCP insertion through insert_cowart_image rather than raw tldraw JSON edits.
• Done/error status feedback so the Cowart button does not stay stuck in "running".

Validation

After installing, open Cowart and submit a small test request. Confirm in Codex with 确认生成. A successful run creates a new image shape beside the source image and writes a done request status.