Sandbox

Sandbox Specialist

Overview

Mirrors part of the alibaba/OpenSandbox response contract for OSS Agent Lab tests. It applies simple string validation and returns synthetic execution records.

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

That is a simulated catalogue. The current execute_code function never evaluates the supplied code, starts a process, or invokes OpenSandbox; it returns synthetic output based on string markers.

Prerequisites

• Use Python 3.11+ in a local OSS Agent Lab checkout and run pip install -e ..
• Treat the tool as a validator/contract simulator, not a security boundary or execution engine.
• Read the runtime contract before interpreting availability.

Capabilities

• execute: Return simulated stdout, stderr, exit code, and timing without running a snippet.
• code_execution: Alias capability tag for routing from generic "run code" intents.
• sandbox: Exercise an isolation-shaped response without providing an isolation boundary.
• multi_language: Validate names against a static runtime catalogue.

Tools

Instructions

1. Prefer validate for static keyword warnings; supply a supported language identifier.
2. Use execute only to test the response contract and label its stdout/stderr as simulated.
3. Use list_runtimes as a configured catalogue, not proof that interpreters are installed.
4. Route real execution to a separately configured and independently verified sandbox backend.

Examples

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)
print(result.result["stdout"])

CLI

oss-lab run sandbox "print('hello, sandbox!')"

With language override

request = SpecialistRequest(
    intent=Intent(
        action="execute",
        domain="code",
        confidence=0.95,
        parameters={
            "code": "console.log('hello from node');",
            "language": "javascript",
            "timeout": 10,
        },
    ),
    query=Query(user_input="console.log('hello from node');"),
    specialist_name="sandbox",
)

Validate without executing

request = SpecialistRequest(
    intent=Intent(
        action="validate",
        domain="code",
        confidence=0.9,
        parameters={"code": "import os; os.system('ls')", "language": "python"},
    ),
    query=Query(user_input="import os; os.system('ls')"),
    specialist_name="sandbox",
)

result = await specialist.execute(request)
print(result.result["valid"])       # False (warnings present)
print(result.result["warnings"])    # ['os.system() executes shell commands...']

List available runtimes

request = SpecialistRequest(
    intent=Intent(action="list_runtimes", domain="code", confidence=1.0),
    query=Query(user_input=""),
    specialist_name="sandbox",
)

result = await specialist.execute(request)
for rt in result.result["runtimes"]:
    print(rt["name"], rt["version"], rt["available"])

Output

execute

{
  "stdout": "[sandbox:python] Execution started\n[sandbox:python] 1 line(s) processed\n[sandbox:python] Execution complete",
  "stderr": "",
  "exit_code": 0,
  "execution_time_ms": 12.4,
  "language": "python",
  "execution_id": "3f2504e0-4f89-11d3-9a0c-0305e82c3301"
}

validate

{
  "valid": true,
  "errors": [],
  "warnings": ["eval() executes arbitrary code strings."],
  "language": "python"
}

list_runtimes

{
  "runtimes": [
    {"name": "python", "version": "3.12.3", "available": true},
    {"name": "javascript", "version": "node 22.2.0", "available": true}
  ],
  "total_count": 12,
  "available_count": 10
}

Error Handling

• Reject empty code, unknown language, and timeouts outside 1–300 seconds.
• A valid: true result means no built-in string rule fired; it is not proof of safety or correctness.
• Never execute or recommend executing untrusted code outside an actual isolation boundary.

Resources

Wraps alibaba/OpenSandbox. The current local specialist does not include that runtime. See the runtime contract.