skill-auditor-in-sandbox（沙盒中的技能审核员）

Skill Auditor in Sandbox

Test and audit Claude Code skills in an isolated NovitaClaw (OpenClaw) sandbox before installing them locally. The skill launches a sandbox, installs the target skill, runs a security scan, and generates a structured risk report.

Quick Reference

Prerequisites

• NOVITA_API_KEY must be set. If not, ask the user to register at https://novita.ai/?utm_source=OpenClaw&utm_medium=Dev_Rel&utm_campaign=Claw-30 and get an API key from https://novita.ai/settings/key-management, then:

export NOVITA_API_KEY=<key>

• novitaclaw CLI must be installed. If not found:

curl -fsSL https://novitaclaw.novita.ai/install.sh | bash

• novita-sandbox npm package must be available. If not:

npm install novita-sandbox

Usage

You are given a skill name (or identifier) as $ARGUMENTS. Your job is to launch a sandbox, install the skill, run a security audit, and generate a report.

Step 1: Launch Sandbox

novitaclaw launch --json

Parse the JSON output and extract sandbox_id and webui. Save these for the report.

If launch fails, check error_code and remediation fields:

• MISSING_API_KEY → ask user for API key
• SANDBOX_TIMEOUT → retry with --timeout 300

Step 2: Install Skill

Run the install script from the project root:

SANDBOX_ID=<sandbox_id> SKILL_NAME="$ARGUMENTS" node scripts/install-skill.mjs

The script outputs JSON: { success, method, skillDir, files, error? }.

• If success is false, show the error and stop.
• Note the method used (clawhub / git-github / git-clawhub) for the report.

Step 3: Security Audit

Run the audit script:

SANDBOX_ID=<sandbox_id> SKILL_NAME="$ARGUMENTS" node scripts/audit-skill.mjs

The script outputs JSON:

• suspicious[] — lines matching risky code patterns (dynamic execution, shell spawning, encoding, etc.)
• urls[] — all URL references found in skill files
• externalPaths[] — references to paths outside the skill directory (system dirs, dotfiles, temp dirs)
• dependencies — contents of requirements.txt or package.json if present
• fileContents[] — full contents of all text files for manual review

Step 4: Assess Risk

Based on audit results, assign a risk level:

Step 5: Generate Report

Output a structured report:

## Skill Installation Report

**Skill:** <skill-name>
**Sandbox ID:** <sandbox_id>
**Web UI:** <webui_url>
**Timestamp:** <current time>

### Installation Status
- **Result:** SUCCESS / FAILED
- **Method:** <clawhub / git-github / git-clawhub>
- **Files Installed:** <count> files

### Installed Files
<table of files and their purpose>

### Security Analysis
- **Risk Level:** LOW / MEDIUM / HIGH / CRITICAL

### Suspicious Patterns Found
| File | Line | Pattern | Severity |
|------|------|---------|----------|
(or "None found")

### URL References
| File | URL | Context |
|------|-----|---------|
(list all URLs and whether they look legitimate)

### External Path References
(list any, or "None found")

### Dependencies
(list any, or "No external dependencies")

### Recommendations
- <recommendation based on findings>

### Sandbox Management
- To access: <webui_url>
- To pause (save costs): `novitaclaw pause <sandbox_id>`
- To stop (permanent): `novitaclaw stop <sandbox_id>`

After generating the report, automatically pause the sandbox to save costs:

novitaclaw pause <sandbox_id> --json

Then inform the user that the sandbox has been paused and can be resumed or stopped:

• To resume: novitaclaw resume <sandbox_id>
• To stop (permanent): novitaclaw stop <sandbox_id>

What Gets Scanned

Important Notes

• Always use --json flag with novitaclaw commands.
• The sandbox auto-terminates based on keep_alive. Suggest pause to save costs.
• Prefer pause over stop — stop is irreversible. Confirm before stopping.

Attribution

• Source: https://github.com/freecodewu/skill-auditor-in-sandbox
• Powered by NovitaClaw sandbox infrastructure