HTTP Client (HttpSandboxManager)

HttpSandboxManager is the remote counterpart of LocalSandboxManager. It wraps REST calls to the HTTP server. Its API is identical to the local manager — switching between local and remote is a one-line change.

Basic usage

import asyncio
from ms_enclave.sandbox.manager import HttpSandboxManager
from ms_enclave.sandbox.model import DockerSandboxConfig, SandboxType

async def main():
    async with HttpSandboxManager(base_url='http://127.0.0.1:8000') as manager:
        sandbox_id = await manager.create_sandbox(
            SandboxType.DOCKER,
            DockerSandboxConfig(
                image='python:3.11-slim',
                tools_config={'python_executor': {}},
            ),
        )
        result = await manager.execute_tool(
            sandbox_id, 'python_executor',
            {'code': 'import platform; print(platform.node())'},
        )
        print(result.output.strip())
        await manager.delete_sandbox(sandbox_id)

asyncio.run(main())

With auth

If the server has api_key enabled, the client must pass it:

HttpSandboxManager(base_url='http://server:8000', api_key='your-secret-key')

Unified factory: local vs. remote

SandboxManagerFactory switches based on whether base_url is provided:

from ms_enclave.sandbox.manager import SandboxManagerFactory

# Local (dev)
async with SandboxManagerFactory.create_manager() as manager:
    ...

# Remote (production)
async with SandboxManagerFactory.create_manager(base_url='http://server:8000') as manager:
    ...

Business code stays the same; deployment mode can change freely.

Timeouts & retries

HttpSandboxManager accepts a timeout argument (seconds, per HTTP request). For long-running tasks (large image pulls, container cold starts), increase it:

HttpSandboxManager(base_url='...', timeout=120)

If image pulls on the server side are too slow, docker pull the image ahead of time.

Notes

• The client doesn't know whether the server has a pool enabled. To use the pool, call POST /pool/initialize on the server first, then use execute_tool_in_pool.
• Retries are not built in; add your own (e.g. tenacity) for flaky networks.