The Agent Skills Directory

Command Execution: The skill defines a helper function q() to interact with the database using the sqlcmd utility. This allows the agent to execute diagnostic SQL queries directly against the target server as part of its intended administrative functionality.
Privilege Management: The skill provides instructions for granting VIEW DATABASE STATE or adding users to the ##MS_ServerStateReader## server role. While these actions involve granting elevated monitoring permissions, they are presented as a necessary step to resolve visibility issues where a login cannot see other sessions, which is standard for database monitoring tasks.
Indirect Prompt Injection: The skill reads data from system views and Extended Events XML reports, which represents a potential surface for indirect prompt injection if the database content is untrusted. However, the risk is limited to the scope of database diagnostics.
Ingestion points: Database results from queries against sys.dm_exec_requests, sys.dm_xe_database_sessions, and XML deadlock reports (SKILL.md).
Boundary markers: The skill does not implement specific delimiters or "ignore embedded instructions" warnings for the database output.
Capability inventory: The skill uses sqlcmd to execute queries but does not perform file system writes or general network exfiltration based on the ingested data.
Sanitization: The skill processes XML results using standard SQL .nodes() methods without specific sanitization of the string content.
Credential Handling: The skill acknowledges the sensitivity of database passwords by explicitly stating that SQL_PASSWORD should be sourced from a secret store. This follows security best practices for avoiding credential exposure in shell environments.

diagnose-blocking-and-deadlocks