tashfeenahmed/scallopbot/run-code

Run Code Skill

Author a small program and run it. This is the "code-execution-as-action" primitive: rather than being limited to a fixed catalog of tools, the agent can write a throwaway script to do exactly what a task needs and execute it.

When to Use

Reach for run_code over bash when the task is really a small program:

• Data transforms: parse JSON/CSV, reshape, aggregate, filter
• Math / computation: anything beyond a one-liner
• Using a library: call a python or node module to do the heavy lifting
• Generating a tool on the fly: write the logic once, run it, read the result

For a single shell command (ls, git status, npm install), use bash.

Input

{
  "language": "python",
  "code": "import json\nprint(json.dumps({'sum': sum(range(10))}))",
  "timeout": 60000
}

Output

{
  "success": true,
  "output": "{\"sum\": 45}\n",
  "error": "",
  "exitCode": 0,
  "language": "python"
}

Notes

• The program runs in the workspace directory; temp source files are written to

the OS temp dir and cleaned up after execution.

• python requires python3 on PATH; if it's missing, the result explains how

to install it. javascript runs under node. bash runs under bash.

• Output is captured up to 1 MB; long-running programs are killed at the timeout.
• Same trust model as the bash skill: it runs on the host unless the operator enabled SANDBOX_MODE, in which case writes outside the workspace (and, if configured, network access) are denied.