using-chatgpt-relay

name	using-chatgpt-relay
description	Send ChatGPT queries from any remote NixOS host (devbox or cloudbox) via the ask-question CLI. Use when /ask-question fails with "Server not running" or when setting up chatgpt-relay for the first time on a new host.
allowed-tools	["Bash","Read"]

Using chatgpt-relay

The ask-question CLI runs on a remote NixOS host (devbox or cloudbox) and talks to ask-question-server running on macOS with a real browser. The relay is an SSH reverse-tunnel pattern, not host-specific.

Architecture

┌─────────────────────────────────────────────────────────────────────────┐
│  REMOTE NIXOS HOST (devbox or cloudbox)                                 │
│                                                                         │
│  OpenCode → /ask-question → ask-question CLI                            │
│                                      │                                  │
│                                      │ HTTP POST                        │
│                                      ▼                                  │
│                          localhost:3033 (tunnel)                        │
└─────────────────────────────────────│───────────────────────────────────┘
                                      │ SSH reverse tunnel
                                      │
┌─────────────────────────────────────│───────────────────────────────────┐
│  MACOS (local machine)              ▼                                   │
│                                                                         │
│                          ask-question-server:3033                       │
│                                      │                                  │
│                                      ▼                                  │
│                              Chromium (headless)                        │
│                              + ChatGPT session                          │
└─────────────────────────────────────────────────────────────────────────┘

Setup

1. On macOS: Start the server

cd ~/Code/chatgpt-relay
npm install
ask-question-login          # One-time: log into ChatGPT in browser
ask-question-server         # Keep running

2. On macOS: Create SSH reverse tunnel

When connecting to the remote host, add a reverse tunnel (substitute devbox or cloudbox for <host>):

ssh -R 3033:localhost:3033 <host>

Or add to your SSH config:

Host <host>
    HostName <host-ip>
    User dev
    RemoteForward 3033 localhost:3033

In this repo the persistent tunnel launchd agents (devbox-dev-tunnel, cloudbox-dev-tunnel) already carry port 3033 — see the troubleshooting-nixos-host skill for the full port table.

3. On the remote host: the CLI is already installed

On the NixOS hosts (devbox + cloudbox) the ask-question CLI is installed declaratively as a Nix package (pkgs/ask-question, wired into home.base.nix under isDevbox || isCloudbox). There is no npm install / npm link step, and the repo is not cloned on these hosts — only the client is packaged. The ask-question-server and ask-question-login halves need playwright + a browser and run only on macOS (~/Code/chatgpt-relay).

If ask-question is missing (e.g. on a freshly provisioned host), apply home-manager:

# devbox
nix run home-manager -- switch --flake ~/projects/workstation#dev
# cloudbox
nix run home-manager -- switch --flake ~/projects/workstation#cloudbox

To bump the pinned version, edit pkgs/ask-question/default.nix (rev + src.hash, plus undiciVersion and the node-modules outputHash if the upstream lockfile changed), then re-apply.

4. On the remote host: Verify connection

# Check tunnel is working
curl -s http://localhost:3033/health | jq .
# Expected: {"ok":true}

# Test a query
ask-question "What is 2+2?"

Environment Variables

The CLI uses ASK_QUESTION_SERVER_URL (default: http://127.0.0.1:3033).

With the SSH reverse tunnel, the default works. If using a different setup:

export ASK_QUESTION_SERVER_URL=http://localhost:3033

Research Caveats

DB-layer / schema claims: verify before coding

ChatGPT-relay answers about database constraints, unique indexes, or model fields can be stale or fabricated, especially for fast-moving projects like Letta. A 2026-04-17 planning pass against Letta 0.16 (eternal-machinery bead zvul) asked whether server_name uniqueness had been dropped; ChatGPT answered yes. First live run crashed on a still-enforced uix_name_organization_mcp_server UNIQUE constraint, costing a full session of rework.

Before relying on a schema claim, verify against one of:

1. The live API (e.g. deliberately POST a second instance, inspect the error shape returned by the server).
2. The upstream source tree (e.g. letta/server/db/models.py or equivalent — a quick GitHub search of the upstream repo usually settles it).
3. A second ChatGPT pass that is required to cite a file path or migration in the upstream repo, not just "I recall from the changelog...".

Non-schema behavioral claims (how a tool is called, what shape a JSON response has) can usually be trusted without extra verification because they're immediately falsifiable by writing the code.

Troubleshooting

"Server not running or not responding"

1. Check tunnel on the remote host:

   curl -s http://localhost:3033/health
   

   If no response, the SSH tunnel isn't active.

2. Check server on macOS:

   curl -s http://localhost:3033/health
   

   If no response, start ask-question-server.

3. Reconnect with tunnel:

   # On macOS, reconnect to the remote host with -R flag
   ssh -R 3033:localhost:3033 <host>   # devbox or cloudbox
   

"No session found" or "Login required"

On macOS:

ask-question-login

Tunnel drops on disconnect

The SSH tunnel only exists while your SSH connection is active. If you disconnect, reconnect with the -R flag.

For persistent tunnels, consider:

• Using autossh for automatic reconnection
• Running the tunnel in a separate tmux pane

Response timeout

Default is 20 minutes. For complex questions:

ask-question -t 1800000 -f question.md -o answer.md