The Agent Skills Directory

Indirect Prompt Injection Surface: The skill processes potentially untrusted content from pull request descriptions, comments, and CI logs. This data is ingested through context files and could theoretically contain malicious instructions intended to manipulate the agent. The skill mitigates this risk by providing explicit instructions to the agent to disregard any commands found in the analyzed text and by using a deterministic verdict system that limits the impact of LLM reasoning on the final output.
External Resource Retrieval: The scripts fetch visual failure evidence and test metadata from Azure DevOps (dev.azure.com). While downloading external content is a common vector, the skill employs strict URL allowlisting and performs validation on the downloaded files, such as verifying PNG signatures and enforcing size constraints, to ensure data integrity.
Repository Modification via CLI: The skill utilizes the GitHub CLI (gh) and git to post comments and publish visual assets to a dedicated branch. These operations involve writing to the repository, which is a privileged action. Security is maintained by using a dedicated, validated asset branch and by sanitizing all injected content—such as HTML-encoding text and neutralizing GitHub mentions—to prevent formatting-based attacks or unwanted notifications.

review-test-failures