The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill documentation in SKILL.md describes a 'SCAPI code mode' that enables the execution of arbitrary JavaScript async functions.
Evidence: Instructions specify using 'native Node execution' for snippets and recipes.
Guardrails: The skill mandates using a 'disposable child' to bound runtime and explicitly restricts Node permissions, disabling filesystem, subprocess, worker, and addon grants for the child process.
[INDIRECT_PROMPT_INJECTION]: The skill establishes several surfaces for ingesting untrusted data from external systems which could contain malicious instructions.
Ingestion points: Data is ingested from the Salesforce Commerce API (SCAPI), WebDAV instance files, system logs, and CIP analytics reports.
Capability inventory: The skill includes capabilities to write files via WebDAV, execute code via 'code mode', and perform state-changing API operations.
Sanitization: The instructions emphasize the use of SDK safety rules, resolution enrichment, and 'input_required' elicitation for confirmations, though these are implementation guidelines rather than hard enforcements in the provided text.
[COMMAND_EXECUTION]: The skill provides instructions for tools that perform state-changing and destructive operations on remote environments.
Evidence: references/tool-effects.md defines scapi_execute as a 'destructive' tool that performs create, update, and delete operations via the Commerce API. It also mentions 'cartridge deployment' which overwrites existing files on the instance environment.

mcp-development