haiku-skills-sandbox

Docker sandbox skill for haiku.skills. Executes Python code in an isolated Docker container with pre-installed data science packages and filesystem access.

Prerequisites

Build the Docker image (once):

docker build -t haiku-skills-sandbox:latest skills/sandbox/haiku_skills_sandbox/

Usage

Via entry point discovery

HAIKU_SKILLS_SANDBOX_WORKSPACE=/path/to/data haiku-skills chat

Programmatic

from pathlib import Path
from haiku_skills_sandbox import create_skill

skill = create_skill(
    workspace=Path("/path/to/data"),  # mounted at /workspace in the container
    idle_timeout=1800,                # stop container after 30min idle (default: 1h)
    image="my-custom-image:latest",   # custom Docker image (default: haiku-skills-sandbox:latest)
)

Configuration

Priority: create_skill() argument > environment variable > default.

Container lifecycle

• Containers start lazily on the first tool call
• Session binding via SandboxState.session_id — the same AG-UI thread reuses the same container
• Idle containers are stopped automatically (checked on each tool call)
• All containers are stopped on process exit via atexit
• When workspace is mounted, files persist on the host — restarting a container loses nothing

Pre-installed packages

The haiku-skills-sandbox:latest image includes: pandas, numpy, scipy, matplotlib.

Security considerations

• Network access: Containers have full network access (Docker bridge networking) by default.
• Command execution: ConsoleToolset is configured with require_execute_approval=False — the LLM can run arbitrary commands inside the container without user confirmation. Docker provides the isolation boundary.
• Workspace access: When a workspace is mounted, the container has full read/write access to that host directory. The LLM can read, modify, or delete any file in the mounted workspace.
• Container user: The container runs as root. Docker container isolation is the security boundary, not OS-level user separation.