shine-sandbox

Execute code safely in an isolated Docker container. Prevents untrusted code from running directly on the host.

<objective> Safely execute code in isolation:

  1. Risk assessment:

    • 🟢 Safe (direct OK): read-only ops, linting, type-checking.
    • 🟡 Moderate (sandbox recommended): package installs, API calls, file I/O.
    • 🔴 High (sandbox REQUIRED): untrusted code, global installs, network access, unknown scripts.

  2. Sandbox selection (Rule #21):

    • Tier 1: docker MCP → microsandbox → Docker CLI via Bash.
    • Tier 2: e2b (ASK first — "This uses E2B cloud sandbox (freemium). Proceed?").
    • No sandbox: REFUSE for 🔴 code. Explain why + give Docker install command.

  3. Execution:

    • Select container image for language (python:3.12, node:22, golang:1.22, etc.).
    • Mount code read-only unless writes required.
    • Resource limits: 2 CPU, 1 GB RAM, 5-min timeout.
    • Capture stdout, stderr, exit code, duration.

  4. Cleanup: Remove container + temporary files.

Delegate to shine-sandbox-runner agent for complex multi-step executions. </objective>

<guardrails> - Never run 🔴-risk code directly on host — always sandbox. - Never expose host network to container without explicit approval. - Never mount home directory or sensitive paths into containers. - Always clean up containers after execution. - If Docker not installed → state gap + give install link, never simulate. </guardrails>