What problem does it solve? Running model-generated Python against documents or repositories you did not author yourself is a security risk: a prompt-injection payload hidden in an ingested PDF can trick the model into writing destructive code. This Skill provides the policy and checklist for sandboxing the Deep Context Platform's RLM kernel so that process separation is never mistaken for a real security boundary. ## Core Features & Use Cases - Three-Tier Sandboxing Model: Tier 1 subprocess isolation for local dev with trusted documents, Tier 2 container isolation (Docker with dropped capabilities, read-only root filesystem) for any third-party input, and Tier 3 syscall-level isolation (gVisor/nsjail or hosted sandboxes like E2B/Modal) for production. - Mandatory Restriction Checklist: Enforces scratch-directory-only filesystem access, network egress allowlists, credential separation (kernel never holds database credentials or API keys), capped REPL output (8,192 characters per turn), and crash-safe lifecycle management. - Use Case: When deploying the RLM orchestrator against a repository of user-uploaded documents, use this Skill to decide that Tier 2 container isolation is required and to verify the kernel accesses the host only through the typed host-request bridge. ## Quick Start Ask the AI to review the RLM kernel's code-execution path against this sandboxing checklist and recommend the correct isolation tier for your deployment.
