Sandbox Automation Skill

Goal

Safely automate multi-step tasks in the sandbox using run_sandbox_workflow.

How this maps to DeepAgents

The DeepAgents example uses an execute tool plus filesystem tools. In Linea, use:

• •run_sandbox_workflow for command execution (analogous to execute)
• •Shell commands inside the workflow for file ops (e.g., ls, cat, mkdir -p)
• •Keep one session alive to iterate (reuse sessionId)
• •Delegate execution to the sandbox_runner subagent when the workflow is multi-step or likely to need retries.
  • •The sandbox_runner subagent has BaseSandbox-backed execute and filesystem tools.

Workflow

1) Plan and track

• •Outline the steps and use write_todos to track execution.
• •Identify required tooling (apt, pnpm, pip) and include installation as step 1.
• •Decide on a Docker image if a specific toolchain is needed.

2) Execute in a single session

• •Use run_sandbox_workflow with persistWorkspace enabled.
• •Keep the session alive for iterative loops (pass keepAlive: true).
• •Reuse sessionId on follow-up calls to continue work.
• •For complex workflows, spawn sandbox_runner via task and have it execute steps + report results.

3) Verify and summarize

• •Include a verification step (tests, lint, output checks).
• •Summarize results and next actions.

Runbook Template

steps:
1) Install deps (apt/pnpm/pip)
2) Prepare workspace (clone/copy/setup)
3) Run primary task
4) Verify results

Example Workflow

{
  "goal": "Create and run hello.js",
  "steps": [
    { "name": "init", "command": "mkdir -p app && cd app && printf 'console.log(\"Hello\")\\n' > hello.js" },
    { "name": "run", "command": "cd app && node hello.js" }
  ],
  "persistWorkspace": true,
  "keepAlive": true
}

Notes

• •If a step fails, rerun with the same sessionId and updated commands.
• •Keep commands explicit and deterministic.