The Agent Skills Directory

Dynamic Execution: The skill includes a script used to calibrate model graders by executing local Python files.
Evidence: finetuning/scripts/calibrate_grader.py uses compile() and exec() to load and run user-supplied Python code.
Context: This is a standard requirement for the Reinforcement Fine-Tuning (RFT) workflow, allowing for custom evaluation logic. The script includes documentation advising users to only run trusted code they have reviewed.
Indirect Prompt Injection: The skill features workflows for harvesting production logs and traces to create and refine evaluation datasets.
Evidence: The eval-datasets sub-skill processes production telemetry from Application Insights in the trace-to-dataset.md workflow.
Context: Ingesting untrusted data from production logs creates a surface for indirect prompt injection. The risk is minimized by the use of structured KQL filters and a mandatory human-in-the-loop review step for all harvested candidates.
Command Execution: The skill frequently interfaces with system tools using shell commands.
Evidence: Scripts like deploy_model.py and check-and-setup-dependencies.sh invoke the az and azd CLI tools via subprocess calls.
Context: These operations are necessary for authentication, infrastructure provisioning, and deployment within the Azure ecosystem.
External Downloads: The skill automates the installation of official extensions to provide additional functionality.
Evidence: scripts/check-and-setup-dependencies.sh and other sub-skills use azd extension install to manage the microsoft.foundry, azure.ai.agents, and azure.ai.toolboxes extensions.
Context: These extensions are provided by the trusted vendor and are part of the core toolchain for Microsoft Foundry development.

microsoft-foundry