openai-docs

This file is a documentation/example demonstrating how to wire a model-driven 'local shell' tool to execute commands locally. The code has no hidden payload, but it exemplifies a highly dangerous pattern: running untrusted commands with host privileges and forwarding environment and outputs back to a remote model/API. If used as-is in production it enables remote code execution, secret leakage, destructive operations, and remote-driven persistence. Treat the example as unsafe for production; implement strict allowlist/denylist controls, sandboxing (container/jail), least privileges, explicit redaction of secrets, resource/time limits, and human authorization for high-risk commands before any deployment.