| 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
ask-question CLI ──► HTTP POST /ask ──► ask-question-server (daemon)
│
▼
Chromium (headless)
+ storageState session
│
▼
ChatGPT response
Prerequisites
- Node.js installed
- Playwright dependencies installed with
npm install
- Repo
devenv/direnv shell active so local ask-question* commands are on PATH
Setup (One-Time)
cd ~/Code/chatgpt-relay
npm install
ask-question-login
Do not use npm link in the Nix/devenv setup. The repo provides local
devenv wrappers for ask-question, ask-question-server, and
ask-question-login. If the shell is not already activated, run commands as
devenv shell -- ask-question-login.
Starting the Daemon
cd ~/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:
[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
ask-question "What is the capital of France?"
From file
ask-question -f question.md -o answer.md
Pipe input
echo "Explain async/await in JavaScript" | ask-question
Options
| Option | Description |
|---|
-f, --file <path> | Read prompt from file |
-o, --output <path> | Write response to file |
-t, --timeout <ms> | Response timeout (default: 1200000 / 20 min) |
--continue | Continue existing chat (default: start new chat) |
-h, --help | Show help |
Integration with /ask-question Slash Command
The /ask-question slash command in Claude Code uses this CLI:
- Claude drafts a Stack Exchange-formatted question
- Claude invokes:
ask-question -f question.md -o answer.md
- CLI blocks while ChatGPT responds (typically 1-10 minutes)
- 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:
ask-question-server
"No session found"
Run login helper:
ask-question-login
Session expired
ChatGPT sessions expire after a while. Re-run login:
ask-question-login
Response timeout
Default is 20 minutes. Adjust if needed:
ask-question -t 120000 "What is 2+2?"
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:
[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:
[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
| Variable | Default | Description |
|---|
ASK_QUESTION_SERVER_URL | http://127.0.0.1:3033 | Server URL |
ASK_QUESTION_PORT | 3033 | Server port |
ASK_QUESTION_STORAGE_STATE_FILE | ~/.chatgpt-relay/storage-state.json | Session file |
Files
| Path | Purpose |
|---|
~/.chatgpt-relay/storage-state.json | Saved ChatGPT session (cookies/localStorage) |
Architecture Deep Dive
Components
┌─────────────────────────────────────────────────────────────────────────┐
│ 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 cookiesask-question-server: Headless browser loads saved session- Session persists across server restarts
Reliability Features
| Feature | Description |
|---|
| Request ID Tracing | 8-char ID correlates CLI and server logs |
| Automatic Retry | CLI retries once on connection failures |
| Network Failure Detection | Tracks streaming response via response.finished() |
| Hard Reset on Errors | Closes and recreates page on failure |
| Login Verification | Server verifies login state at startup |
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.
