SDK

SDK model

Paperplanes exposes one primary handle: Sandbox.

import { Sandbox, Image } from "@paperplanes/sdk";

const image = await Image.build({
  base: "paperplanes://images/ubuntu-chrome",
  setup: ["npm install", "npm run build"],
  name: "chrome-dev",
});

const sandbox = await Sandbox.create({ image: image.digest });
const checkpoint = await sandbox.checkpoint("before-risky-change");
const branches = await sandbox.fork({ count: 4 });

RL surface

The RL interface is on the same handle:

sandbox.reset()
sandbox.step(action)
sandbox.fork(n)
sandbox.as_env()

No separate Gym-first class is required. as_env() is only an adapter for existing trainers.