The Agent Skills Directory

[COMMAND_EXECUTION]: The skill promotes the 'File System Agent Pattern' in SKILL.md and references/architectural_reduction.md, which involves creating an ExecuteCommand tool that allows an agent to run arbitrary bash commands (e.g., grep, cat, find, ls) for data exploration. Although sandboxing is recommended, granting an agent shell access is a high-risk capability.\n- [REMOTE_CODE_EXECUTION]: The documentation includes implementation patterns for executing code within a sandboxed environment (references/architectural_reduction.md). If the sandbox is improperly configured or bypassed, this enables arbitrary code execution by the agent.\n- [DYNAMIC_EXECUTION]: The skill describes a pattern where an agent explores a data layer by dynamically executing system commands to read and process documentation and configuration files.\n- [INDIRECT_PROMPT_INJECTION]: The skill architecture is designed to ingest and process untrusted external data (such as YAML, Markdown, and JSON files in a data layer) while possessing powerful command execution and SQL tools.\n
Ingestion points: Semantic layer files located at /data/entities/.yaml, /data/measures/.yaml, and /docs/*.md as described in references/architectural_reduction.md.\n
Boundary markers: Relies on tool descriptions acting as contracts and prompt instructions to define agent behavior.\n
Capability inventory: Includes bash command execution (ExecuteCommand) and SQL query execution (ExecuteSQL) as documented in references/architectural_reduction.md.\n
Sanitization: Explicitly mentions the requirement for a sandboxed environment (e.g., Vercel Sandbox) to limit agent access, though the efficacy depends on the implementation.\n- [METADATA_POISONING]: There is an inconsistency between the skill's author metadata in SKILL.md ('Agent Skills for Context Engineering Contributors') and the system-provided author context ('muratcankoylan').

tool-design