Code Interpreter

Tuochat can optionally execute assistant-generated JavaScript or Lua in a tightly limited sandbox. This is not automatic agentic tool use: the user must explicitly approve each execution, and the code runs without filesystem, shell, timer, or network access.

Where it works today

The sandboxed interpreter is integrated into interactive sessions: the terminal REPL (tuochat repl) and the GUI when the Code Int toggle is on. After a response is displayed, tuochat looks for the first fenced code block tagged js, javascript, or lua.

If a matching block is found, tuochat prompts:

Execute javascript code in sandbox? [y/N]
Attach output to conversation? [Y/n]

The second prompt queues the sandbox output as an attachment for your next message.

Installation

Install the bundled interpreter extra:

pip install "tuochat[code-interpreters]"

Or install only the runtime you want:

pip install "tuochat[js-miniracer]"   # JavaScript via mini-racer
pip install "tuochat[js-dukpy]"       # JavaScript via dukpy
pip install "tuochat[lua]"            # Lua via lupa

If no compatible runtime is installed, tuochat still works normally; the sandbox run will just fail with an import error instead of executing.

Using it with the bundled skill

Tuochat ships with a bundled code-interpreter skill:

/skill code-interpreter

That skill tells Duo to respond with short sandbox-friendly js or lua code blocks and embeds the current sandbox rules into the prompt.

Sandbox rules

Supported languages

  • js / javascript
  • lua

Hard limits

  • Timeout: 500 ms
  • Memory limit: 64 MB
  • Max code size: 64 KB
  • Max stdout capture: 200 lines / 64 KB
  • Max emitted result size: 256 KB JSON

Environment restrictions

  • No filesystem access
  • No shell access
  • No network access
  • No imports or external modules
  • Structured JSON-like input/output only

The code must call emit(value) exactly once to produce a final result. Intermediate logging is captured through console.log(...) in JavaScript and print(...) in Lua.

What this is for

The sandbox is useful for small computations, data transformations, and demonstrations where executing a tiny program is clearer than asking the model to reason step by step in prose.

It is not a general-purpose Python execution environment, and it does not let the assistant access your workstation, repository, or network.