The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to use the Salesforce CLI (sf), jq, and python3 to retrieve, patch, and deploy metadata. It also includes instructions for cleanup using rm -rf on a temporary directory created via mktemp -d. These operations are well-scoped to the task of Salesforce metadata management.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection because it ingests and processes external Salesforce metadata (XML files) that could potentially contain malicious content.
Ingestion points: Metadata files are retrieved from a Salesforce organization via the sf project retrieve start command into the local workspace (SKILL.md).
Boundary markers: The instructions mandate the use of a specific inspection script (inspect-dataspace-scopes.sh) to parse the XML, which helps prevent the agent from directly interpreting the XML content as natural language instructions.
Capability inventory: The skill allows for metadata deployment (sf project deploy start), runtime API mutations (REST POST/DELETE via sf org api rest), and file system deletions (rm -rf on the temporary work directory).
Sanitization: The provided parsing script (scripts/inspect-dataspace-scopes.sh) utilizes the Python xml.etree.ElementTree library, which parses the XML structure as data rather than natural language, mitigating the risk of instructions embedded in the data influencing the agent's logic flow.

platform-dataspace-access-configure