---

name: Using ask-question CLI description: Send prompts to ChatGPT via browser automation. Use this when you need to query ChatGPT from the command line or integrate with the /ask-question slash command. allowed-tools: [Bash, Read]

Using ask-question CLI

Send prompts to ChatGPT via Playwright browser automation with headless operation.

Architecture

Copyask-question CLI ──► HTTP POST /ask ──► ask-question-server (daemon)
                                              │
                                              ▼
                                        Chromium (headless)
                                        + storageState session
                                              │
                                              ▼
                                        ChatGPT response

Prerequisites

1. Node.js installed
2. Playwright with Chromium (npm install downloads it)
3. npm link run in the chatgpt-relay directory

Setup (One-Time)

Copycd ~/Code/chatgpt-relay
npm install
npm link
ask-question-login  # Opens browser, log into ChatGPT

Starting the Daemon

Copycd ~/Code/chatgpt-relay
npm run server:log

Keep this running in a dedicated terminal or tmux pane. It runs headless (no visible browser).

Why tee to a log file? This allows Claude Code to check logs via tail ~/.local/state/chatgpt-relay/daemon.log even when the daemon runs in a separate terminal.

Expected output:

Copy[ask-question-server] Starting headless browser...
[ask-question-server] Using session: ~/.chatgpt-relay/storage-state.json
[ask-question-server] ChatGPT page opened.
[ask-question-server] Login verified.
[ask-question-server] HTTP server listening on http://127.0.0.1:3033

Usage

Direct prompt

Copyask-question "What is the capital of France?"

From file

Copyask-question -f question.md -o answer.md

Pipe input

Copyecho "Explain async/await in JavaScript" | ask-question

Options

Integration with /ask-question Slash Command

The /ask-question slash command in Claude Code uses this CLI:

1. Claude drafts a Stack Exchange-formatted question
2. Claude invokes: ask-question -f question.md -o answer.md
3. CLI blocks while ChatGPT responds (typically 1-10 minutes)
4. Claude reads the answer file and discusses

Use /ask-question draft topic to skip automation and just draft the question.

Troubleshooting

"Server not running or not responding"

Start the daemon:

Copyask-question-server

"No session found"

Run login helper:

Copyask-question-login

Session expired

ChatGPT sessions expire after a while. Re-run login:

Copyask-question-login

Response timeout

Default is 20 minutes. Adjust if needed:

Copy# Shorter timeout for quick questions
ask-question -t 120000 "What is 2+2?"

# Longer timeout for complex responses
ask-question -t 1800000 "Write a comprehensive analysis..."

Debugging

Request ID Tracing

Each request gets a unique ID (e.g., [a1b2c3d4]) logged on both CLI and server sides:

Copy[ask-question] [a1b2c3d4] Attempt 1/2...
[ask-question-server] [a1b2c3d4] Received (500 chars)
[ask-question-server] [a1b2c3d4] Processing...
[ask-question-server] [a1b2c3d4] Sending response (1200 chars)
[ask-question] [a1b2c3d4] Success

Use this to correlate CLI errors with server logs.

Connection Failures

The CLI automatically retries once on connection errors (fetch failed, ECONNREFUSED, etc.). If you see:

Copy[ask-question] [a1b2c3d4] Error: fetch failed
[ask-question] [a1b2c3d4] Cause: ...
[ask-question] [a1b2c3d4] Cause code: UND_ERR_HEADERS_TIMEOUT

This indicates Undici's internal timeout fired. The CLI will retry automatically.

Environment Variables

Files

Architecture Deep Dive

Components

Copy┌─────────────────────────────────────────────────────────────────────────┐
│  Claude Code                                                            │
│      │                                                                  │
│      ▼                                                                  │
│  /ask-question (slash command)                                          │
│      │                                                                  │
│      ▼                                                                  │
│  CLI (ask-question) ─► HTTP POST /ask ─► ask-question-server (daemon)   │
│                                               │                         │
│                                               ▼                         │
│                                         Chromium (headless)             │
│                                         + storageState session          │
│                                               │                         │
│                                               ▼                         │
│                                         ChatGPT tab                     │
│                                         (DOM automation)                │
└─────────────────────────────────────────────────────────────────────────┘

Design Decisions

HTTP Daemon over WebSocket Connect

• launchServer() doesn't support persistent sessions
• HTTP is curl-debuggable and simpler
• Request queue handles serialization naturally
• Browser runs headless (no focus-stealing)

Headless with StorageState

• ask-question-login: One-time headed browser for manual login, saves cookies
• ask-question-server: Headless browser loads saved session
• Session persists across server restarts

Reliability Features

Fragile Components (Maintenance Required)

These rely on ChatGPT's DOM structure and may break when ChatGPT updates their UI:

• Composer selectors (div[contenteditable], #prompt-textarea)
• Send button selectors (button[data-testid="send-button"])
• Message extraction (copy button, innerText fallback)
• Login detection (chat history panel visibility)
• Stop button lifecycle (generation progress)

Mitigation: Multiple selector fallbacks, semantic selectors where possible, clear error messages.