English · 中文

Pull the latest code out of the installed OpenAI Codex macOS desktop app and turn its minified/bundled JavaScript back into readable, well-named source.

The repository ships two agent skills that work together:

Typical flow: refresh ./ref from the app → deobfuscate ./ref into ./restored.

• macOS with the ChatGPT desktop app installed at /Applications/ChatGPT.app (required by the refresh skill — it reads the app's app.asar).
• Node.js (the refresh skill runs node and npx @electron/asar / prettier).
• Bun (the deobfuscate skill's scripts are TypeScript run with bun).
• An agent that can run skills (Claude Code, or any Codex/agent harness that reads .agents/skills). You can also run the bundled scripts by hand — see below.

The skills live under .agents/skills/. Point your agent at this repo, then invoke a skill by name; or run the scripts directly.

Recreates ./ref from the installed ChatGPT app, so the source you analyze always matches the version you have installed. It deletes and replaces ./ref, extracts /Applications/ChatGPT.app/Contents/Resources/app.asar into it, then formats every extracted .js/.css with Prettier (skipping ref/node_modules).

From the repo root, ask the agent:

Use codex-app-ref-refresh to refresh ./ref from the installed ChatGPT app.

# from the repo root (the directory that should own ./ref)
node .agents/skills/codex-app-ref-refresh/scripts/refresh-codex-ref.mjs

Options:

⚠️ Run it from the directory that should own ./ref — the script intentionally wipes ./ref first. It refuses to run unless the target resolves to ./ref under the current working directory.

After it finishes, the bundled web UI lives in ref/webview/ (entry: ref/webview/index.html, chunks: ref/webview/assets/).

Turns the minified, bundled JS extracted into ./ref back into readable code. It's a multi-stage pipeline (deobfuscate → smart-rename → polish → optional typed .tsx rewrite). The agent itself does the semantic renaming; bundled bun scripts handle the mechanical work. Restored files land in ./restored/ with meaningful names, semantic-domain subfolders, provenance headers, and a shared restored/IMPORT_MAP.json. Only finalized files land in ./restored/: the skill stages mechanical batch/script output in a hidden working area (restored/.deobfuscate-javascript/) first, and promotes a file only after it's organized — readable, well-named, clearly structured (and fully typed in deep mode).

After ./ref exists, ask the agent (English or Chinese both trigger it):

Use deobfuscate-javascript to restore ./ref.

or, for example: "deobfuscate ref/webview/assets", "反混淆 ./ref", "看懂这段 JS".

By default it works on the whole import tree: it reads ref/webview/index.html, auto-discovers the app entry, and recursively restores every reachable project-local chunk. Paste a single file instead and it restores just that file.

cd .agents/skills/deobfuscate-javascript
bun install

The individual bun scripts/*.ts tools (detect, extract, smart-rename, polish, quality-gate, …) are documented in the skill's SKILL.md. In normal use you don't call them by hand — the agent orchestrates them.

decode-codex/
├─ .agents/skills/
│  ├─ codex-app-ref-refresh/      # Skill 1: sync ChatGPT.app → ./ref
│  └─ deobfuscate-javascript/     # Skill 2: ./ref → ./restored
├─ ref/                           # extracted ChatGPT app source (generated by Skill 1)
└─ restored/                      # readable, deobfuscated output (generated by Skill 2)

ref/ and restored/ are generated artifacts — regenerate them with the skills rather than editing by hand.

This project is for personal study and interoperability research of software you have installed. The extracted code is © OpenAI; respect the ChatGPT app's license and terms of service. Don't redistribute the extracted or restored code.