Skip to content

provider-cubesandbox.md

Latest commit

 

History

History
47 lines (32 loc) · 3.94 KB

provider-cubesandbox.md

File metadata and controls

47 lines (32 loc) · 3.94 KB
Raw

CubeSandbox provider

CubeSandbox is a catalog Provider consumed by the CLI and Python ProviderSandbox through the local engine. The Node.js server HTTP client is a separate consumer and does not manage cloud sandboxes. 中文.

The authoritative plugin and operation declarations live in awesome-providers/cubesandbox. Install its Git tree with abx provider install alibaba-agent-sandbox/awesome-providers//cubesandbox --allow-unpinned --yes, or use the Python registry installer python -m agent_sandbox.providers install registry:cubesandbox --yes. Pin a reviewed full commit for reproducible use.

Exact interpreter boundary

The adapter implements only code.default-context.execute: persistent variables in a sandbox's default Jupyter context. It sends JSON code, language and optional env_vars to /execute on port 49999 and reads the buffered NDJSON result. Call await sandbox.run_code("x = 42", mode="default-context") explicitly.

Stateless code.execute, explicit context create/execute/list/restart/delete, realtime code.stream and code.results.structured are not declared. Context IDs and callbacks are rejected before code I/O. A buffered result does not promise realtime callbacks, and missing interpreter services do not fall back to a shell.

This mapping is distinct from the removed guessed Connect interpreter RPC. Cube's official Python SDK README documents /execute and persistent variables. This source mapping and loopback tests do not establish a live Cube deployment pass.

Configure and create

Field Environment Default
API origin ABX_CUBESANDBOX_BASE_URL http://127.0.0.1:3000
Optional provider key ABX_CUBESANDBOX_API_KEY Unset
Port-host suffix ABX_CUBESANDBOX_DOMAIN cube.app
Remote node proxy ABX_CUBESANDBOX_PROXY Unset

Provider-scoped keys stay separate from Express credentials. Process and filesystem calls use envd on port 49983; Jupyter uses 49999. Port hosts are {port}-{sandboxID}.{domain}. The adapter's optional proxy controls TCP dialing while retaining the virtual Host name.

from agent_sandbox import ProviderSandbox

sandbox = await ProviderSandbox.create(
    provider="cubesandbox",
    template="YOUR_NATIVE_TEMPLATE_ID",
    provider_options={"base_url": "http://127.0.0.1:3000", "domain": "cube.app"},
    timeout=60,
)
try:
    result = await sandbox.run_code("print(42)", mode="default-context")
finally:
    await sandbox.kill()

The control-plane mapping follows the Cube OpenAPI and official SDK. The plugin maps lifecycle/timeout/pause/resume, exec, files and ports independently. Its manifest is the current operation authority; coarse feature groups do not authorize another workflow. Process management, PTY, file transfer, native template inventory, storage and VPC are not inferred from another supported operation.

Source preparation and evidence

CLI source preparation requires an explicit compatible Dockerfile/profile, Docker, a writable image repository and native Cube registration prerequisites. It preserves FROM, USER, COPY and ENTRYPOINT. It returns the same provider's native managed reference; raw OCI references are not treated as Cube template IDs. Python does not declare template.prepare and consumes an already prepared reference.

Cube remains Beta. Unit/loopback tests exercise mapping and failure behavior; native create requires an actual CubeAPI, valid native template and node/runtime setup. This environment has not completed that native acceptance. New source or binding identities require new scoped evidence; a historical record cannot certify the current workflow.