The Agent Skills Directory

[COMMAND_EXECUTION]: The Python scripts scripts/get_model_info.py and scripts/get_semantic_model_ai_metadata.py invoke external command-line utilities such as fab and az using the subprocess module. These tools are utilized to interact with Power BI and Fabric service APIs. While the scripts use list-based argument passing to mitigate standard shell injection, they provide a mechanism for the agent to execute system-level commands based on model identifiers.
[INDIRECT_PROMPT_INJECTION]: The skill focuses on managing 'AI instructions' and 'AI schemas' stored within semantic models (e.g., instructions.md). These natural language components steer AI features like Copilot. Because the skill ingests these instructions for model review and auditing, it provides a surface for indirect prompt injection where malicious instructions in a model definition could influence the agent's output.
Ingestion points: Semantic model definition files such as Copilot/Instructions/instructions.md and culture linguistic metadata.
Boundary markers: None; natural language instructions are processed as raw content for review.
Capability inventory: Model mutation via the te CLI and deployment via the fab CLI.
Sanitization: Scripts decode and parse the metadata structures but do not sanitize the natural language content of the instructions sent to the agent.
[EXTERNAL_DOWNLOADS]: As noted in references/documentation-and-bpa.md, the Best Practice Analyzer (te bpa) used by the skill can be configured to fetch ruleset definitions from external URLs.
[DYNAMIC_EXECUTION]: The skill's core workflow relies on te script to execute C# code against the Tabular Object Model. Additionally, interactive C# scripts like scripts/edit-ai-instructions-interactive.csx use reflection (e.g., Assembly.Load, Activator.CreateInstance) to dynamically load and invoke UI components.

semantic-model