Secure Jupyter Sandbox (CONCEPT:AU-ECO.messaging.native-backend-abstraction)

The Secure Sandbox Engine provides a constrained, isolated environment for dynamic code generation and execution, specifically tailored for backtesting and quantitative trading logic.

Overview

Running dynamically generated quantitative strategies natively poses immense risk. agent-utilities mitigates this by integrating a JupyterKernelAdapter wrapped inside a SandboxExecutor. Before any generated code is executed, it undergoes rigorous mathematical and structural validation.

Key Benefits

• Programmatic Safety: By validating the Abstract Syntax Tree (AST), the sandbox blocks dangerous OS-level calls (e.g., os.system).
• Mathematical Invariants: This repository enforce State Machine Invariants (MCS Ch 6) to guarantee that quantitative loops stop and do not exceed pre-defined execution budgets.
• Full Reproducibility: The sandbox can be instantiated identically across test, staging, and production environments.

Implementation Details

Located in agent_utilities.tools.sandbox_executor, the engine uses standard Jupyter kernel protocols (jupyter-client) behind the scenes while front-loading the security checks.

Workflow

1. The orchestrator receives a generated strategy from a specialist agent.
2. The orchestrator forwards the strategy to the SandboxExecutor.
3. The executor runs _validate_invariants(code) against the AST.
4. If the code passes validation, it is forwarded to the isolated JupyterKernelAdapter.
5. The result (or a runtime error) is returned structured as JSON.

Example

from agent_utilities import SandboxExecutor

sandbox = SandboxExecutor()

# Safe execution
result = sandbox.run_safe("import pandas as pd; df = pd.DataFrame()")
print(result["status"]) # ok

# Unsafe execution (Blocked by Invariants)
unsafe_result = sandbox.run_safe("import os; os.system('rm -rf /')")
print(unsafe_result["error"]) # InvariantViolation...