AgentSkillsCN

code-runner

在沙盒环境中执行代码片段。当用户请求运行代码、测试脚本，或在将某项解决方案固化为一项技能之前需要对其进行验证时，可使用此功能。

SKILL.md
--- frontmatter
name: code-runner
description: Execute code snippets in a sandboxed environment. Use when the user asks to run code, test a script, or when you need to validate a solution before crystallizing it as a skill.
metadata:
  openpanda:
    tags: ["code", "execution", "sandbox", "programming"]
    permissions: ["sandbox:exec"]

Code Runner

Execute code in a secure sandboxed environment (Docker or subprocess).

Usage

Run Python:

code
run_code(language="python", code="print('Hello')", timeout=10000)

Run JavaScript:

code
run_code(language="javascript", code="console.log(42)", timeout=5000)

Run shell script:

code
run_code(language="bash", code="curl -s https://api.example.com | jq .", timeout=15000)

Supported Languages

  • •Python 3.x
  • •JavaScript (Node.js)
  • •Bash / Shell
  • •TypeScript (via tsx)

Parameters

  • •language (required): Programming language
  • •code (required): Code to execute
  • •timeout (default: 10000): Max execution time in ms
  • •installDeps (optional): Dependencies to install before execution (e.g., ["requests", "pandas"])

Safety

  • •All code runs in an isolated sandbox
  • •Network access is restricted by default
  • •File system access limited to sandbox working directory
  • •Memory capped at 512MB, CPU time limited
  • •NEVER run on the host machine

For Skill Crystallization

When using code-runner to validate a solution for crystallization:

  1. •Run the code at least once successfully
  2. •Capture stdout, stderr, and exit code
  3. •If it fails, iterate (search docs -> fix -> retry, max 3x)
  4. •On success, the Learner agent can crystallize it into a permanent skill