ta-sandbox

Allowlisted command execution for Trusted Autonomy.

Provides a controlled execution environment where agents can run pre-approved commands (search, format, test) without access to a real shell or host filesystem outside the staging workspace.

Architecture

The sandbox has three enforcement layers:

1. Command allowlist: Only pre-approved binaries can execute
2. CWD enforcement: All execution is confined to the staging workspace
3. Network policy: Per-domain allow/deny for outbound connections

Usage

use ta_sandbox::{SandboxConfig, SandboxRunner, CommandPolicy};

let config = SandboxConfig::default();
let mut runner = SandboxRunner::new(config, "/path/to/workspace");

// Execute an allowed command
let result = runner.execute("rg", &["TODO", "src/"]).unwrap();