LLM-free execution sandbox component. It is a strict runtime environment built on top of the Python ecosystem.
Relies on Docker sandboxing for isolating generated code (using containerization features to limit CPU/Memory via --network none when enabled). Alternatively uses a subprocess sandbox with a sanitized environment. Included tools consist of standard PyData ecosystems (Pandas, Scikit-learn, Matplotlib).
Executes after the CoderAgent successfully yields a script. Injects target outputs into a local runtime block and evaluates logic, extracting stdout, stderr, and capturing graphical assets resulting from Matplotlib renders. Its output determines whether the pipeline jumps to the Verifier or Router.
Executes generated code in a python:3.11-slim container (when DOCKER_SANDBOX_ENABLED=true) or via a managed async Python subprocess locally.