WLV-ZEDD/dsh-chatgpt-web

English | 中文

Zero-cost conversational AI model provider for DeepSeek Harness powered by free ChatGPT Web (GPT-6 Luna).

---

Table of Contents

• Overview
• Requirements
• Quick Start
• Diagnostics & Health Check
• Troubleshooting
• Notes & Limitations
• Related Documentation

---

Overview

dsh-chatgpt-web turns your local browser session of ChatGPT into a seamless, $0.00 API-free conversational model provider directly inside DeepSeek Harness (DSH).

It connects via headless or visible Chrome/Chromium browser automation to chatgpt.com, streaming real-time Markdown responses, code solutions, explanations, and reasoning back to your DSH chats without consuming API credits.

Why Pure Chat?

Small conversational web models (like GPT-6 Luna) excel at explanations, dialogue, Q&A, brainstorming, code snippet generation, and side-assistant tasks. By running in Pure Chat Mode, the bridge eliminates prompt overhead, tool hallucinations, and syntax errors of autonomous multi-tool execution loops, providing a fast, rock-solid, zero-cost LLM provider.

Key Features

• 100% Free ($0.00 Cost): Uses your existing free ChatGPT Web session. No OpenAI API keys or credit cards needed.
• Cordis Plugin-First Lifecycle: Managed by DSH via ctx.effect. DeepSeek Harness starts the background sidecar automatically on launch and shuts it down on exit.
• Dynamic Model Auto-Detection: Automatically detects your ChatGPT tier:
  • Free Accounts: Defaults to chatgpt-web/luna (gpt-5-6-luna) and chatgpt-web/think.
  • Plus / Team Accounts: Automatically detects and exposes GPT-6 (chatgpt-web/gpt-6-light, medium, high) and GPT-5.6 Sol (chatgpt-web/light, medium, high), plus Pro tiers when available.
• Full Streaming Markdown & Code Blocks: Delivers tokens in real time directly to the DSH Web UI or CLI.
• Portable, Validated Artifacts: The published tarball is validated by CI (clean install, runtime smoke test, dependency resolution, no build-environment path contamination) before it can reach the registries.

---

Requirements

• Runtime: Node.js >=22.19.0 or Bun >=1.3.0 (both are supported; the CLI runs on either).
• DeepSeek Harness: >=0.2.0-rc.2.
• Browser: Chrome or Chromium on your machine (used for the one-time sign-in and for serving).

---

Quick Start

1. Installation

Install the plugin directly from this GitHub repository:

# Via npm registry (recommended):
dsh plugin --profile <profile> add @wlv-zedd/dsh-chatgpt-web

# Or directly from GitHub:
dsh plugin --profile <profile> add github:WLV-ZEDD/dsh-chatgpt-web

This registers the plugin in the profile's package.json bundles, so its Cordis entries are composed automatically — no manual insert is needed.

2. One-Time Browser Sign-In

Authenticate your ChatGPT account once. The plugin binary lives in the profile's node_modules/.bin:

~/.dsh/profiles/<profile>/node_modules/.bin/dsh-chatgpt-web setup --dsh

A dedicated Chrome/Chromium window will open. Log into your OpenAI / ChatGPT account. Once the composer is visible and the ChatGPT session is authenticated, setup captures the session and writes the config and browser state to the plugin's storage directory (default ~/.dsh/storages/chatgpt-web/). A guest composer alone is not treated as a successful login.

--dsh leaves Codex configuration and OS services untouched; DSH owns the sidecar lifecycle. Existing Codex or DEV storage cannot be repurposed by this command: use a separate DSH_CHATGPT_WEB_HOME if needed. Before rerunning setup, stop or disable the sidecar so its port is free. Without --dsh, the legacy Codex setup path remains unchanged.

If your browser is not at the default path, pass it explicitly:

~/.dsh/profiles/<profile>/node_modules/.bin/dsh-chatgpt-web setup --dsh --chrome /path/to/chrome

Using a snap-confined browser (e.g. snap Chromium)

Snap-confined browsers can only write to non-hidden paths in $HOME. The plugin's default storage directory (~/.dsh/storages/chatgpt-web/) is hidden, so a snap browser cannot create its profile lock there and setup fails with:

Failed to create .../login-profile-XXXX/SingletonLock: Permission denied (13)

The fix is to point the plugin's storage home at a non-hidden directory using the DSH_CHATGPT_WEB_HOME environment variable.

How the variable works: VAR=value command is standard shell syntax — it sets the environment variable only for that one command and does not persist it. The plugin reads DSH_CHATGPT_WEB_HOME on startup; when set, every file it owns (config, browser profile, storage state) lives under that directory instead of ~/.dsh/storages/chatgpt-web/.

Step by step:

1. Run the one-time setup with the variable applied to the command:

   DSH_CHATGPT_WEB_HOME="$HOME/dsh-chatgpt-web" \
     ~/.dsh/profiles/<profile>/node_modules/.bin/dsh-chatgpt-web setup --dsh --chrome /snap/bin/chromium
   

   The browser window opens; log in to ChatGPT. The session is saved under ~/dsh-chatgpt-web/.

2. Persist the variable for the shell that starts DSH. The background sidecar is started by DeepSeek Harness and inherits DSH's environment — it does not inherit a one-off VAR=value prefix from your terminal. So export the variable in your shell (e.g. ~/.bashrc), then start DSH from that shell:

   # add to ~/.bashrc (or ~/.zshrc):
   export DSH_CHATGPT_WEB_HOME="$HOME/dsh-chatgpt-web"
   
   # reload, then start DSH in that shell:
   source ~/.bashrc
   dsh --profile <profile> web
   

   The sidecar now finds the same config.json and browser state that setup created.

3. Verify with the doctor (run it in the same environment):

   ~/.dsh/profiles/<profile>/node_modules/.bin/dsh-chatgpt-web doctor
   

Important: every command that touches the plugin state — setup, login, doctor, and the DSH process that runs the sidecar — must see the same DSH_CHATGPT_WEB_HOME. If some of them use the hidden default path and others use the non-hidden one, the sidecar will report Configuration is missing.

3. Enable in DeepSeek Harness

Provider configuration lives in the profile's Cordis patch layer (~/.dsh/profiles/<profile>/cordis.patch.yml), not in a global settings file. Add the chatgpt-web provider under the existing providers map of the llm-pi-ai entry:

- id: llm-pi-ai
  name: "@deepseek-ai/dsh-llm-pi-ai"
  config:
    providers:
      # ...your existing providers...
      chatgpt-web:
        displayName: "ChatGPT Web (Free)"
        api: openai-responses
        baseURL: http://127.0.0.1:17841/v1
        headers:
          Authorization: "Bearer chatgpt-web-free"
        streamIdleTimeoutMs: 300000
        models:
          - id: chatgpt-web/luna
            name: "ChatGPT Web — Luna (Free)"
            contextWindow: 1050000
            maxTokens: 32768
            input:
              - text
              - image

Optionally, make it the agent's default model in the agent-default-model entry:

- id: agent-default-model
  name: "@deepseek-ai/dsh-agent-default-model"
  config:
    provider: chatgpt-web
    model: chatgpt-web/luna

Now start DeepSeek Harness with the profile:

dsh --profile <profile> web

DSH will automatically start the background sidecar process, connect to your authenticated ChatGPT session, and accept prompts!

---

Diagnostics & Health Check

Verify your setup at any time with the built-in diagnostic doctor:

~/.dsh/profiles/<profile>/node_modules/.bin/dsh-chatgpt-web doctor

Example healthy output:

✓ Configuration is valid (~/.dsh/storages/chatgpt-web/config.json)
✓ Chrome executable found
✓ ChatGPT login state has authenticated browser evidence
✓ Responses proxy is healthy on 127.0.0.1:17841
Doctor result: ready

---

Troubleshooting

See TROUBLESHOOTING.md for common issues: browser executable paths, snap confinement, missing configuration, and runtime (Node/Bun) notes.

---

Notes & Limitations

1. Unofficial Bridge: Operates via local Playwright browser automation on chatgpt.com. Not affiliated with or endorsed by OpenAI.
2. Single-Session Concurrency: Runs within a single browser tab. Sequential queries and normal DSH agent chats work seamlessly; avoid launching parallel multi-subagent swarms against the same tab simultaneously.
3. Pure Chat Only: This provider generates pure conversational responses, explanations, reasoning, and code blocks. It does not execute local filesystem, terminal, or autonomous tool loops.
4. Standard Free Tier Rate Limits: Subject to standard OpenAI free-tier hourly usage limits.

---

Related Documentation

• CONTRIBUTING.md — development setup and contribution guidelines.
• SECURITY.md — how the plugin handles your session credentials.
• TROUBLESHOOTING.md — diagnosing setup and runtime problems.
• LICENSE — MIT.