Sandbox Architect

Overview

This skill analyzes a codebase to understand its technology stack, architecture, and dependencies, then generates tailored Claude Code Sandbox configurations. The skill uses an interactive question-and-answer approach to achieve 95% confidence before generating the final configuration.

Workflow

Phase 1: Automatic Codebase Analysis

Before asking questions, analyze the codebase automatically:

1. Run the analysis script to detect the stack:

   python3 scripts/analyze_codebase.py
   

2. If the script is unavailable, manually check for:

   • package.json → Node.js/JavaScript
   • requirements.txt, pyproject.toml, setup.py → Python
   • Cargo.toml → Rust
   • go.mod → Go
   • composer.json → PHP
   • Gemfile → Ruby
   • pom.xml, build.gradle → Java
   • *.csproj → .NET
   • Dockerfile, docker-compose.yml → Container-based

3. Identify key patterns:

   • Web frameworks (Next.js, Django, FastAPI, etc.)
   • Dev servers (Vite, webpack-dev-server, etc.)
   • Testing frameworks (Jest, pytest, etc.)
   • Docker/container usage
   • Database connections
   • External API integrations

Phase 2: Interactive Clarification Questions

Ask clarifying questions until 95% confident. Number each question and provide labeled options.

Question Template Format

**[N]/ [Question Title]**

[Brief context about why this matters]

Options:
- **a)** [Option description]
- **b)** [Option description]
- **c)** [Custom input option if applicable]

**Recommendation**: [Your recommended choice] - [Reason why]

Core Questions to Ask

1/ Security Posture

Different projects require different security levels based on their context.

Options:

• a) Learning/Exploration - I'm experimenting and want minimal friction
• b) Standard Development - Balance between security and productivity
• c) High Security - Working with sensitive code, untrusted dependencies, or production systems
• d) CI/CD Pipeline - Automated environment with strict controls

Recommendation: b) Standard Development - Provides good security while maintaining productivity. Switch to c) if handling credentials or untrusted code.

---

2/ Development Server Usage

Dev servers need to bind to localhost ports (especially on macOS).

Options:

• a) Yes - I run dev servers (npm run dev, python manage.py runserver, etc.)
• b) No - This is a library, CLI tool, or backend-only project
• c) Sometimes - I occasionally test with local servers

Recommendation: Based on detected framework. If web framework detected, recommend a).

---

3/ Container/Docker Usage

Docker commands are incompatible with sandbox and need special handling.

Options:

• a) Yes - I use docker, docker-compose, or similar container tools
• b) No - No containerization in this project
• c) Sometimes - Docker is optional for this project

Recommendation: If Dockerfile detected, recommend a) and add to excludedCommands.

---

4/ Git Operations Preference

Git push/commit can be sensitive operations.

Options:

• a) Auto-allow all git operations (sandboxed)
• b) Ask before push/commit (recommended for team projects)
• c) Ask before all git operations (maximum caution)

Recommendation: b) - Allows routine git status/diff/log but prompts before changing history.

---

5/ SSH Agent Access (Linux/macOS)

Required for SSH-based git authentication.

Options:

• a) Yes - I use SSH keys for GitHub/GitLab
• b) No - I use HTTPS with credential helper
• c) Not sure - Check my setup

Recommendation: If .git/config shows SSH remote URLs, recommend a).

---

6/ Sensitive Files to Protect

Beyond the defaults (.env, secrets), identify project-specific sensitive files.

Options:

• a) Default protection only (.env, .env.*, secrets/)
• b) Add custom paths (please specify)
• c) I have credentials in non-standard locations

If b) or c) selected, ask: "Please list the file patterns to protect (e.g., config/production.yml, *.key)"

Recommendation: a) unless detected files like credentials.py, *.pem, *.key.

---

7/ Package Registry Access

Network access needed for installing dependencies.

Options:

• a) Public registries only (npmjs.org, pypi.org, crates.io, etc.)
• b) Private/internal registry (please specify domain)
• c) Air-gapped - No network access needed
• d) Mixed - Both public and private registries

Recommendation: Based on detected package manager and any .npmrc/.pypirc files.

---

8/ Escape Hatch Behavior

When a command fails in sandbox, should Claude retry outside sandbox?

Options:

• a) Yes - Allow unsandboxed retries with prompt (good for discovering requirements)
• b) No - Fail closed, I'll update config manually (maximum security)

Recommendation: a) for new projects (learning phase), b) for established configs.

---

9/ Auto-Allow Sandboxed Commands

Should sandboxed bash commands run without prompting?

Options:

• a) Yes - Auto-allow (recommended for trusted projects)
• b) No - Still prompt for approval (useful for learning what Claude runs)

Recommendation: a) for productivity, b) when learning Claude Code or auditing behavior.

---

10/ Additional Tools/Commands (if applicable)

Based on codebase analysis, ask about specific tools detected:

• Kubernetes (kubectl, helm)
• Database CLIs (psql, mysql, mongosh)
• Cloud CLIs (aws, gcloud, az)
• Build tools (make, gradle, maven)

Phase 3: Configuration Generation

After collecting answers, generate the configuration by:

1. Reading references/stack-templates.md for the base template
2. Customizing based on user answers
3. Adding detected dependencies and domains
4. Presenting the final configuration with explanations

Output Format

// .claude/settings.json
// Generated by Sandbox Architect for [detected stack]
{
    "sandbox": {
        // [Comment explaining each setting]
    },
    "permissions": {
        // [Comment explaining each rule]
    }
}

Phase 4: Verification

After generating the config:

1. Summarize what the configuration does
2. List any trade-offs or limitations
3. Provide next steps for testing
4. Offer to adjust if needed

Key Principles

Ask Questions Progressively

• Start with high-impact questions (security posture, dev server)
• Skip questions that analysis already answered
• Group related questions when possible
• Explain WHY each question matters

Provide Clear Recommendations

Every question must include:

• Numbered options (a, b, c, etc.)
• A clear recommendation with reasoning
• Context-specific advice based on codebase analysis

Reference the Documentation

For complex scenarios, reference references/sandbox-guide.md which contains:

• Complete setting descriptions
• Platform-specific notes (Linux vs macOS)
• Security threat model
• Troubleshooting guidance

Resources

scripts/

• analyze_codebase.py - Automatic stack and dependency detection

references/

• sandbox-guide.md - Complete Claude Code Sandbox documentation
• stack-templates.md - Pre-built configurations for common stacks

Confidence Threshold

Reach 95% confidence before generating configuration. Factors that increase confidence:

• Clear stack identification (+20%)
• Dev server needs known (+15%)
• Docker usage clarified (+10%)
• Git auth method confirmed (+10%)
• Sensitive file paths identified (+15%)
• Network requirements known (+15%)
• Security posture selected (+15%)

If confidence is below 95%, continue asking clarifying questions about uncertain areas.