Skip to content

sandbox-mode.md

Latest commit

 

History

History
56 lines (41 loc) · 2.17 KB

sandbox-mode.md

File metadata and controls

56 lines (41 loc) · 2.17 KB
Raw

Sandbox Mode

Sandbox mode is a rollout runtime alongside Docker and RJob. Safactory creates a Brainbox Sandbox Instance through the OpenSandbox SDK, runs the agent entrypoint inside it, and deletes the instance after rollout and optional evaluation.

Prerequisites

  1. Create a Sandbox Environment in advance using SandboxAPI.md at the repository root.
  2. Its image must match the agent config env_image.
  3. It must expose the command port, 44772 by default.
  4. Configure data and result storage as Environment volumes; local Docker bind mounts are not translated.
  5. The Sandbox must be able to reach sandbox.gateway_base_url.

Install dependencies and provide credentials:

pip install -r requirements.txt
export OPEN_SANDBOX_API_KEY='<ak>:<sk>'

Copy config.sandbox.example.yaml and set project, environment_id, and a cluster-reachable gateway URL. An agent can override the Environment in its start config:

agent_name: openrt

container:
  workdir: /app
  runner_entrypoint:
    source: ./runner.py
    target: /tmp/safactory-openrt-runner.py
    command: "python /tmp/safactory-openrt-runner.py"

sandbox:
  environment_id: env-openrt
  required_mount_paths: [/app/data, /app/results]

The local runner source is installed after instance allocation. container.mounts remain Docker-only.

Run the flow with:

python launcher.py \
  --mode sandbox \
  --sandbox-config config.sandbox.yaml \
  --agent-config env/openrt/openrt_config.yaml \
  --agent-start-config env/openrt/openrt_start.yaml \
  --gateway-base-url http://YOUR_GATEWAY_HOST:8000/v1/sessions \
  --llm-model YOUR_ROUTE_KEY \
  --pool-size 8

--pool-size controls the number of active instances. The manager fills those leases before workers start, and the Environment capacity must cover that concurrency.

Rule evaluation needs no Sandbox-specific settings. The rule evaluator receives the rollout result and the persisted trajectory after the Gateway session is closed.

Brainbox does not implement pause. Set lifecycle_minutes long enough to cover rollout, telemetry flush, and evaluation. Use cleanup_on_finish: false only for debugging because it preserves quota-consuming instances.