Sandboxed Python Variants

This is one of the best patterns in Airlock right now.

The goal is to let the agent do cheap local scripting without making you approve every trivial transform, while keeping a stronger path available for real scripts.

The pattern

• python/sandboxed is your low-friction path for JSON, text, and local transforms
• python/full still exists when the agent genuinely needs more power
• python/github is a middle ground when the script needs a narrow network surface

How it works

All three variants use alias_of: exec/run — they're the same underlying tool with different security envelopes. The agent sees them as separate tools with different descriptions and chooses the right one for the task.

Current implementation

Airlock does not expose a separate Python provider. These variants are aliases of exec/run, and sandboxing is applied by enabling the agent sandbox and setting sandbox presets or per-tool overrides. That keeps all command policy, approval, audit, and sandbox behavior on the same execution path.

Full config example

providers:
  exec: builtin

sandbox_presets:
  local_transform:
    filesystem:
      allow_read: ['.']
      allow_write: ['/tmp', '/private/tmp']
      deny_read: ['~/.ssh', '~/.aws', '.env']
      deny_write: ['.']
    network:
      allowed_domains: []

  github_only:
    network:
      allowed_domains:
        - 'github.com'
        - '*.github.com'
        - 'api.github.com'

agents:
  claude-code:
    allow:
      - python/sandboxed
    ask:
      - python/full
      - python/github

    sandbox:
      enabled: true
      presets:
        - local_transform

    tool_overrides:
      python/sandboxed:
        alias_of: exec/run
        description: 'Run Python for local transformations only (no network, writes to /tmp only)'

      python/full:
        alias_of: exec/run
        description: 'Run Python with broader permissions after approval'
        sandbox:
          filesystem:
            allow_write: ['.', '/tmp', '/private/tmp']
            deny_write: []
          network:
            allowed_domains: ['pypi.org', '*.pythonhosted.org']

      python/github:
        alias_of: exec/run
        description: 'Run Python with GitHub-only network access after approval'
        sandbox_presets:
          - github_only

What each variant can do

Capability	python/sandboxed	python/full	python/github
Read project files	Yes	Yes	Yes
Write to /tmp	Yes	Yes	Yes
Write to project dir	No	Yes (after approval)	No
Network access	None	pypi.org only (after approval)	github.com only (after approval)
Requires approval	No	Yes	Yes
Denied paths	~/.ssh, ~/.aws, .env	~/.ssh, ~/.aws, .env	~/.ssh, ~/.aws, .env

When to use which

• python/sandboxed: JSON transforms, CSV parsing, text munging, code generation, local data processing — anything that reads files and writes to /tmp
• python/full: Installing packages, writing generated code to the repo, running scripts that need pip
• python/github: API scripts that fetch from GitHub, release automation, PR data analysis

Approval experience

When the agent calls python/full, the approval notification includes:

APPROVE? [A1B2C3]
Agent:   claude-code
Tool:    python/full
Sandbox: write:.,/tmp | network:pypi.org,*.pythonhosted.org | deny-read:~/.ssh,~/.aws,.env
Command: python3 -c "import requests; ..."

The sandbox context is visible so you know exactly what the script can do before approving.

See examples/sandbox-presets.yaml for the complete reference config.