520kiss/wushu

Fork 0

Files

Developer 9175ff9905 Initial commit: Flutter 无书应用项目

2026-03-30 02:35:31 +08:00

6.5 KiB

Raw Blame History

BoxLite Setup

Using planning-with-files inside BoxLite micro-VM sandboxes via ClaudeBox.

Overview

BoxLite is a micro-VM sandbox runtime — hardware-isolated, stateful, embeddable as a library with no daemon and no root required. The analogy their team uses is "SQLite for sandboxing": you import it directly into your application and each box gets its own kernel, not just namespaces or containers.

ClaudeBox is BoxLite's official Claude Code integration layer. It runs the Claude Code CLI inside a BoxLite micro-VM with persistent workspaces, optional skills, security policies, and resource limits. planning-with-files loads into ClaudeBox as a Skill object — the skill's SKILL.md and scripts get injected into the VM filesystem at startup, exactly where Claude Code expects to find them.

Why use this combination:

Run untrusted AI-generated code in hardware isolation without touching your host filesystem
planning files (task_plan.md, findings.md, progress.md) persist across sessions via ClaudeBox's stateful workspaces at ~/.claudebox/sessions/
Snapshot before a risky phase, roll back if needed
All hooks (PreToolUse, PostToolUse, Stop) work because Claude Code runs natively inside the VM

Prerequisites

Python 3.10+
BoxLite runtime (installed automatically as a ClaudeBox dependency)
Docker — used to pull and manage OCI images (the sandbox itself runs as a micro-VM, not a Docker container)
CLAUDE_CODE_OAUTH_TOKEN or ANTHROPIC_API_KEY set

Installation

Install ClaudeBox

pip install claudebox

Set your API credentials

export CLAUDE_CODE_OAUTH_TOKEN=sk-ant-oat01-...
# or
export ANTHROPIC_API_KEY=sk-ant-...

Quick Start

The minimal example — planning-with-files inside a BoxLite VM:

import asyncio
from pathlib import Path
from claudebox import ClaudeBox, Skill

# Load the SKILL.md content from the installed skill
SKILL_MD = Path.home() / ".claude" / "skills" / "planning-with-files" / "SKILL.md"

PLANNING_SKILL = Skill(
    name="planning-with-files",
    description="Manus-style file-based planning with persistent markdown",
    files={
        "/root/.claude/skills/planning-with-files/SKILL.md": SKILL_MD.read_text()
    }
)

async def main():
    async with ClaudeBox(
        session_id="my-project",
        skills=[PLANNING_SKILL]
    ) as box:
        result = await box.code(
            "Plan and implement a user authentication feature for the Express API"
        )
        print(result.response)

asyncio.run(main())

Claude Code finds the skill at /root/.claude/skills/planning-with-files/SKILL.md inside the VM. The three planning files (task_plan.md, findings.md, progress.md) are written to the box's working directory and persist across sessions.

Persistent Sessions

ClaudeBox workspaces survive restarts. Pick up exactly where you left off:

async def main():
    # First session — starts the plan
    async with ClaudeBox(session_id="auth-feature", skills=[PLANNING_SKILL]) as box:
        await box.code("Create task_plan.md for the authentication feature")

    # Later session — same workspace, plan files still there
    async with ClaudeBox.reconnect("auth-feature") as box:
        await box.code("Continue implementing the login endpoint from the plan")

    # Clean up when done
    await ClaudeBox.cleanup_session("auth-feature", remove_workspace=True)

asyncio.run(main())

How Hooks Work Inside the VM

ClaudeBox runs Claude Code CLI natively inside the BoxLite micro-VM. This means planning-with-files hooks execute exactly as they do on your local machine:

Hook	Trigger	What It Does
PreToolUse	Before Write, Edit, Bash, Read, Glob, Grep	Reads first 30 lines of `task_plan.md` — keeps goals in attention
PostToolUse	After Write, Edit	Reminds agent to update plan status after file changes
Stop	When agent finishes	Runs completion check script before stopping

The hook scripts need to be injected alongside SKILL.md. See the full example in examples/boxlite/quickstart.py for how to include them.

Session Recovery

When your context fills up and you run /clear, the skill recovers the previous session automatically on next activation. Inside a ClaudeBox session, run manually:

async with ClaudeBox.reconnect("my-project") as box:
    await box.code(
        "Run session catchup: python3 ~/.claude/skills/planning-with-files/scripts/session-catchup.py $(pwd)"
    )

Snapshots

BoxLite supports checkpointing. Snapshot before a risky implementation phase:

from claudebox import ClaudeBox
from boxlite import Box

async def main():
    async with ClaudeBox(session_id="risky-refactor", skills=[PLANNING_SKILL]) as box:
        # Phase 1 complete — snapshot before the destructive refactor
        await box.box.snapshot("pre-refactor")

        result = await box.code("Refactor the database layer")

        if "error" in result.response.lower():
            await box.box.restore("pre-refactor")
            print("Rolled back to pre-refactor snapshot")

asyncio.run(main())

Troubleshooting

Skill not activating inside the VM?

The skill file must be at /root/.claude/skills/planning-with-files/SKILL.md inside the VM. Verify by injecting a validation step:

result = await box.code("ls ~/.claude/skills/planning-with-files/")
print(result.response)

Hooks not running?

The Stop hook script (check-complete.sh) must also be injected. Add it to the files dict in your Skill object. See examples/boxlite/quickstart.py for the full implementation.

Docker not found?

ClaudeBox uses Docker to pull OCI images for the VM. Install Docker Desktop (macOS/Windows) or Docker Engine (Linux). The sandbox itself does not run as a Docker container — Docker is only used for image management.

BoxLite not available on Windows host?

BoxLite requires Linux KVM (Linux) or Hypervisor.framework (macOS). On Windows, use WSL2 with KVM enabled.

6.5 KiB Raw Blame History