The Agent Skills Directory

[PERSISTENCE]: The skill instructs the agent to install specialist sub-skills globally using the npx skills add command with the -g flag. This is explicitly designed to make the code persist across different agent sessions and projects by storing it in global directories such as ~/.agents/skills and ~/.codex/skills.
[PROMPT_INJECTION]: The instructions employ forceful language to override standard agent workflows, telling the agent to 'Act now,' 'do not ask first,' and 'do not stop or report a blocker.' Most notably, it specifically directs the agent to 'request network escalation and retry' if environment security policies block necessary downloads, which is an attempt to influence the agent to bypass platform-level network restrictions.
[EXTERNAL_DOWNLOADS]: The skill performs several installations from external sources, including sub-skills from the microsoft/Dataverse-skills repository, the @microsoft/dataverse CLI, and several Python libraries. These resources originate from well-known services and a trusted organization.
[COMMAND_EXECUTION]: The skill relies extensively on executing shell commands for setup and operation, including package managers (npm, pip, npx), the Dataverse CLI for authentication and data operations, and Python scripts for configuration.
[INDIRECT_PROMPT_INJECTION]: The skill enables the agent to query and process data from external Dataverse environments via SQL and OData, creating a potential attack surface if the retrieved data contains malicious instructions.
Ingestion points: Data retrieved through dataverse data query or Python SDK analytical calls.
Boundary markers: None are specified in the instructions to ensure the agent ignores instructions embedded in the external data.
Capability inventory: The agent possesses the capability to execute shell commands, perform network requests, and write to local configuration files like .env.
Sanitization: No validation or escaping mechanisms are mentioned for handling the contents of external environment data.

dataverse