dsh-codex-auth

> Run DeepSeek Harness on your ChatGPT (Codex) subscription quota — one-click ChatGPT login in the UI, automatic token refresh, and a live remaining-quota badge.

   

中文版: README.zh.md

Run models in DeepSeek Harness using your ChatGPT (Codex) quota — no OpenAI API key needed. If you have a ChatGPT Plus/Pro/Team/Enterprise subscription, you already have Codex capacity; this plugin lets the Harness use it.

✨ Features

📸 Screenshots

<img width="1028" height="980" alt="image" src="https://github.com/user-attachments/assets/e6876200-7539-42a8-b330-94f46c61316f" />

🚀 Install (recommended)

# 1. Make sure pnpm is available (once)
npm install -g pnpm

# 2. Install the plugin into the web profile (pnpm fetches from GitHub and registers the bundle)
dsh plugin --profile web add github:nzfern/dsh-codex-auth

# 3. Activate the provider and restart
#    Under llm-pi-ai.providers in $DSH_HOME/settings.yaml add:
#      openai-codex:
#        apiKeyEnv: OPENAI_CODEX_ACCESS_TOKEN
#    Then restart dsh web

Or run the one-liner installer (installs pnpm, installs the plugin, writes settings.yaml):

Set-ExecutionPolicy -Scope Process Bypass; iex (irm https://raw.githubusercontent.com/nzfern/dsh-codex-auth/main/install.ps1)

🖥️ Manual install

1. Clone/download this repo into the profile's plugin directory and register it as a bundle. 2. Activate the provider in $DSH_HOME/settings.yaml:

llm-pi-ai:
  providers:
    openai-codex:
      apiKeyEnv: OPENAI_CODEX_ACCESS_TOKEN

3. Restart dsh web.

🔑 Login

Web UI (recommended): after restart, open Settings → General, find "ChatGPT · Codex 额度", click Log in with ChatGPT → follow the inline link to https://auth.openai.com/codex/device, enter the code and authorize with your ChatGPT account → the card flips to "Connected".

CLI:

node "$env:USERPROFILE\.dsh\profiles\web\node_modules\dsh-codex-auth\bin\codex-login.mjs"
node "$env:USERPROFILE\.dsh\profiles\web\node_modules\dsh-codex-auth\bin\codex-status.mjs"

💡 Usage

After logging in, pick an openai-codex model in the model picker:

• gpt-5.4 / gpt-5.4-mini (image input supported)
• gpt-5.5 / gpt-5.6-luna / gpt-5.6-sol / gpt-5.6-terra
• gpt-5.3-codex-spark

The model list comes from pi-ai's catalog and updates with the dependency.

⚙️ How it works

The Harness ships the openai-codex provider inside dsh-llm-pi-ai (pi-ai), which talks to https://chatgpt.com/backend-api/codex/responses with a ChatGPT access token as the bearer credential — but it cannot log in or keep the token alive (access tokens expire in ~10 minutes). This plugin fills in the OAuth lifecycle:

• Device-code login flow (same as the official Codex CLI); the refresh token is stored in the Harness credential store and hot-reloaded by the web service;
• Auto-refresh: on boot and every 30s, exchange the refresh token for a fresh access token;
• Quota: read plan type and window remaining from chatgpt.com/backend-api/wham/usage.

❓ FAQ

• Requires a subscription: Codex needs ChatGPT Plus / Pro / Team / Enterprise (free tier gets little to none).
• Usage limit errors: ChatGPT rate-limits Codex usage; the error tells you when to retry.
• Login invalidated: if the refresh token is revoked (password change, sign-out elsewhere), /codex-status shows logged out; just log in again.
• Badge missing: toggle "在主页面显示用量" off in Settings → General; or you're logged out / the usage endpoint is unavailable.
• Configuration: the codex-auth: section in $DSH_HOME/settings.yaml can tune clientId, credential refs, refresh interval, etc.

📦 Files

• lib/codex.js — OAuth protocol core (device flow, refresh, WHAM usage query)
• lib/index.js — host plugin (token keep-alive, /api/codex.* endpoints, commands)
• lib/client.js — browser plugin (settings login card, floating quota badge)
• bin/codex-login.mjs / bin/codex-status.mjs / bin/codex-logout.mjs — CLI

🤝 Support & Contributing

• Bugs → Issues
• Ideas → Discussions
• Like it? Give it a ⭐ so more people can use their Codex quota!