dsh-llm-chatgpt-web

English · 简体中文

---

English

A ChatGPT Web model provider for DeepSeek Harness. It adds a model route named chatgpt-web and answers model calls by driving your own logged-in chatgpt.com session in a local Chrome — so inference is paid from the ChatGPT web quota, not from the OpenAI API and not from your Codex quota.

• ✅ Spends the ChatGPT web plan quota (the same pool you use chatting on chatgpt.com)
• ❌ Does not use ~/.codex/auth.json, Codex OAuth, or Codex quota
• ❌ No API key, no proxy server, no account sharing

Why a browser and not plain HTTP

ChatGPT's web backend (/backend-api/f/conversation) requires a per-request Sentinel/PoW token plus Cloudflare clearance. Those cannot be forged reliably outside a real browser session. So this plugin drives a real Chromium-family browser over the DevTools Protocol (zero dependencies — Node's built-in WebSocket) and reads the answer from the page.

Requirements

• A Chromium-family browser (Chrome, Edge, Chromium)
• DeepSeek Harness 0.2.x
• A ChatGPT account, signed in once inside the plugin's own browser profile

Install

dsh plugin --profile <name> add /absolute/path/to/dsh-llm-chatgpt-web
# or, inside the Harness: plugin_manager install_bundle <absolute path>

The bundle inserts one row:

- id: llm-chatgpt-web
  name: dsh-llm-chatgpt-web
  config: {}

First run: sign in once

1. Pick ChatGPT Web in the model selector and send anything.
2. The plugin opens a dedicated Chrome window on the ChatGPT login page and reports that web sign-in is still missing.
3. Sign in inside that window. The session is stored in the plugin's own profile (~/.dsh/chatgpt-web/chrome-profile), not in your daily browser.
4. Send again.

The dedicated browser starts on demand in the background (macOS: open -g -j) and is re-launched automatically if you close it, without stealing focus from the window you are working in. The one exception is a sign-in prompt: the plugin then moves the window on screen and activates it, because you cannot type into a login page you cannot see.

Configuration

Key	Default	Meaning
provider	chatgpt-web	Route id used by the model selector
displayName	ChatGPT 网页版	Label shown in DSH
transport	web	fake = offline self-test stream
chromePath	Google Chrome (macOS path)	Browser executable
userDataDir	~/.dsh/chatgpt-web/chrome-profile	Dedicated browser profile (login lives here)
debuggingPort	9333	CDP port for the dedicated browser
windowMode	hidden	visible / hidden (window moved off-screen, launched in the background) / headless (blocked by ChatGPT — not recommended)
temporaryChat	true	Use a temporary chat (not written to your ChatGPT history)
conversationMode	continue	continue = keep one temporary chat per DSH session and send only deltas; fresh = new chat each call
reasoningEffort	high	Aligns the web "thinking effort" slider before every call (low / medium / high)
webModel	empty	Optionally force a web model such as GPT-6; empty keeps the page's current pick
maxPromptChars	60000	Prompt budget in characters
emulateTools	true	Emulate function calling with a text protocol (see below)
responseTimeoutMs	600000	Max wait for one reply

Tool calling

The web UI has no function-calling API. The plugin asks the model for a fenced JSON block:

{"name": "bash", "arguments": {"command": "ls"}}

and translates it into a real DSH tool-call; the tool result is pasted back into the next turn. The prompt explicitly forbids the model's own container/code-interpreter from being used, and a reply that claims it ran something without a tool call is rejected and retried.

Known limitations

• One conversation per DSH session: each session gets its own browser tab, holding one temporary chat for its whole lifetime. Later calls continue in that tab and send only the delta (new user turns / tool results), so the web page itself keeps the context. Different sessions are isolated from each other. Set conversationMode: fresh to always start over.
• Serial: one browser tab, so model calls queue.
• Every turn re-sends the task + tools + recent history (temporary chat means no server-side carry-over). Each step takes roughly 10–40 s with reasoningEffort=high.
• Text only: images are not forwarded.
• Depends on chatgpt.com DOM: when OpenAI ships a redesign, selectors may need updating. Failures save a screenshot to ~/.dsh/chatgpt-web/logs/screenshots/.
• Not a security boundary: the plugin runs with your local user permissions and your own ChatGPT session. Review the source before installing.

Troubleshooting

• Logs: ~/.dsh/chatgpt-web/logs/chatgpt-web-YYYY-MM-DD.log
• Last prompt sent: ~/.dsh/chatgpt-web/logs/debug/last-prompt.txt
• Unit tests: node scripts/test-transcript.mjs

License

MIT.

---

简体中文

把 ChatGPT 网页端 接成 DeepSeek Harness 的模型路由（路由 id：chatgpt-web）。选它之后，DSH 每次模型调用都会用你本机已登录的 chatgpt.com 会话、在本地 Chrome 里完成，因此消耗的是 ChatGPT 网页额度——不是 OpenAI API 额度，也不是 Codex 额度。

• ✅ 花 ChatGPT 网页版套餐额度（和在网页上聊天同一个池子）
• ❌ 不读 ~/.codex/auth.json、不走 Codex OAuth、不花 Codex 额度
• ❌ 不需要 API key、不需要中转服务器、不共享账号

为什么必须是浏览器

ChatGPT 网页后端（/backend-api/f/conversation）要逐请求变化的 Sentinel/PoW 令牌和 Cloudflare 放行，离开真实浏览器会话伪造不出来。所以插件用 DevTools 协议驱动一个真实 Chromium 浏览器（零依赖，只用 Node 内置 WebSocket），从页面里读回复。

安装

dsh plugin --profile <name> add /absolute/path/to/dsh-llm-chatgpt-web

首次使用：登录一次

1. 模型选择器里选 ChatGPT 网页版，随便发一句。
2. 插件会打开一个专用 Chrome 窗口，停在登录页。
3. 在那个窗口里登录（账号密码或 Google 都行），登录态存在插件自己的 profile 里。
4. 回来再发一次。

浏览器是后台按需启动的（macOS 用 open -g -j），不会抢你正在用的窗口的焦点；你手动关掉它也会自动重新拉起。唯一例外是需要登录：这时插件会把窗口挪回屏幕内并激活，否则你看不见登录页。

已知限制

• 每个 DSH 会话一条临时对话：会话独占一个标签页，整个会话期间都用这一条对话；后续只发增量（新的用户消息 / 工具结果），上下文由网页端维护，不同会话互不串味。conversationMode: fresh 可恢复每轮新建。
• 单浏览器标签，调用串行排队。
• 每轮都要重新贴「任务 + 工具 + 最近记录」（用临时对话，不污染历史），reasoningEffort=high 时一步通常 10–40 秒。
• 暂不支持图片输入。
• 依赖 chatgpt.com 的 DOM，改版可能要跟进选择器；失败会自动截图到 ~/.dsh/chatgpt-web/logs/screenshots/。
• 不是安全沙箱：插件用你的本地用户权限和你自己的 ChatGPT 会话，装前请读源码。

许可

MIT。