The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/mcp_validator.py uses importlib.util to dynamically load and execute the validate_metadata_operation.py module from a file path constructed at runtime using the __file__ attribute.
[COMMAND_EXECUTION]: In its cli execution mode, the skill runs shell commands through the Salesforce CLI (sf) to perform bulk metadata retrieval and data queries (e.g., sf data query -q "...").
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted metadata and query results from the target Salesforce org, which creates a surface for indirect prompt injection. Maliciously crafted metadata (such as object descriptions) could influence agent behavior during operations like metadata creation or Apex execution.
Ingestion points: Org metadata and field structures are retrieved through tools like sobject_describe and tooling_api_query as defined in SKILL.md.
Boundary markers: There are no explicit instructions for the agent to use delimiters or ignore embedded instructions within the ingested Salesforce data.
Capability inventory: The skill can perform sensitive operations including metadata_create, metadata_update, and potentially apex_execute (mapped in references/execution-modes.md).
Sanitization: The skill includes a MetadataOperationValidator to check schema and naming conventions, but it does not sanitize natural language fields for potential injection patterns.

sf-metadata