Skip to content

The all-in-one local AI studio for your desktop: chat, image and video generation and a coding agent in one free, open source app. Windows and Linux. No Docker, no terminal, no cloud required.

1.8 KB
Raw Blame

Remote Smoke Harness

Self-contained Node.js harness for reproducing and verifying Remote Access bugs without needing a phone. Talks to the same /remote-api/agent-tool endpoint the mobile page uses.

When to use

Use this whenever a Remote bug is reported (file_list wrong path, agent hanging, shell command failing) — saves a full round-trip through the user's phone.

One-time bootstrap

  1. Build & launch the LU desktop exe:
    npm run tauri:build
    target\release\locally-uncensored.exe
    
  2. Open Remote drawer (right side) → click LAN → click Dispatch on a chat → pick a folder.
  3. The drawer will display:
    • URL: e.g. http://192.168.1.42:8765
    • Code: 6-digit (lasts 5 minutes; refresh if expired).

Running the harness

node scripts/remote-smoke.mjs \
  --base http://192.168.1.42:8765 \
  --code 123456 \
  --chat-id remote-smoke

Output:

  • 12 assertions covering: auth, file_write, file_list, file_read, shell cwd default, error response shape.
  • Exit code 0 = green, 1 = at least one fail.

The harness writes a file at <your-folder>/remote-smoke/hello.txt, lists the folder, reads back, runs pwd (or $pwd.Path on Windows), and asserts none of the paths leak through to ~/agent-workspace/__remote__/.

What to assert next time

Bug 1 (path resolution) — the path is NOT ~/agent-workspace/__remote__ checks fail when the override map isn't honored.

Bug 2 (echo / consecutive errors) — exercise via a multi-step agent prompt that fails 5+ times. The new guard should bail with a clean "too many consecutive tool errors" step instead of leaking the system prompt back to the chat.

Adding more checks

agentTool(base, token, chatId, tool, args) is the wrapper. Stack more calls inside main(). Use check(label, ok, detail) to record an assertion.