sandbox

name	sandbox
display_name	Sandbox Specialist
description	Safe multi-language code execution via alibaba/OpenSandbox
version	0.1.0
source_repo	alibaba/OpenSandbox
license	Apache-2.0
tier	experimental
capabilities	["execute","code_execution","sandbox","multi_language"]
allowed_tools	["execute_code","validate_code","list_runtimes"]
output_formats	["python_api","cli","mcp_server","agent_skill","rest_api"]

Sandbox Specialist

Overview

Wraps alibaba/OpenSandbox to provide isolated, resource-limited execution of arbitrary code snippets inside OSS Agent Lab. Each execution runs in a container-backed sandbox with seccomp syscall filters, memory limits, and configurable timeouts — no persistent side effects leak between runs.

Supported runtimes: Python, JavaScript, TypeScript, Bash, Ruby, Go, Rust, Java, C, C++.

Capabilities

• execute: Run a code snippet and capture stdout, stderr, exit code, and timing.
• code_execution: Alias capability tag for routing from generic "run code" intents.
• sandbox: Enforce isolation policies (seccomp, cgroup, read-only overlay FS).
• multi_language: Dispatch to any of the registered runtime backends.

Tools

Tool	Description	Side Effects
execute_code	Execute a code snippet in the sandbox; returns captured output	Subprocess spawn (sandboxed)
validate_code	Static analysis without execution; returns errors and warnings	None
list_runtimes	Enumerate all registered runtimes with version and availability	None

Usage

Python API

from agents.specialists.sandbox.agent import SandboxSpecialist
from oss_agent_lab.contracts import Intent, Query, SpecialistRequest

specialist = SandboxSpecialist()

request = SpecialistRequest(
    intent=Intent(
        action="execute",
        domain="code",
        confidence=0.95,
        parameters={"code": "print('hello, sandbox!')", "language": "python"},
    ),
    query=Query(user_input="print('hello, sandbox!')"),
    specialist_name="sandbox",
)

result = await specialist.execute(request)
(result.result[])