Architectural Constraints Skill

Purpose

Establishes execution safety and sandboxing rules for LLM agents. Prevents malicious command injections, infinite code execution loops, unmitigated token spending, data leakage (PII/tokens), network exfiltration, and file corruption during crash scenarios. This skill governs how agent runtimes interact with system memory, processes, external networks, and files.

Core Principles

1. Zero-Trust Execution: Run all user-generated or LLM-suggested scripts inside ephemeral, restricted namespace sandboxes (e.g. docker mounts or process namespaces).
2. Defensive Resource Budgeting: Throttle CPU, memory, and network rates. If an agent loops, immediately terminate the process using SIGKILL limits.
3. Data Anonymization at Source: Scrub PII (Credit Cards, SSNs, Emails) and API credentials from logging systems before writing to disk or sending over HTTP.
4. Hermetic Outbound Networking: Exclude direct internet connectivity. Restrict egress via verified gateway proxy endpoints.
5. Transactional Checkpoint States: Ensure writing checkpoints (e.g., progress.txt) is atomic, utilizing file locks and backup recovery patterns.

Agent Protocol