The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill architecture involves processing GraphQL schemas and query strings from external subgraphs and local files, which presents a surface for indirect prompt injection.
Ingestion points: The skill utilizes IntrospectAndCompose in gateway/server.ts to poll subgraphs from remote URLs and uses fs.readFileSync in users-subgraph/resolvers.ts to load schema definitions.
Boundary markers: The skill does not explicitly define prompt boundaries or instructions to ignore embedded directives in the provided templates for external data.
Capability inventory: The provided code samples demonstrate capabilities for filesystem access (fs.readFileSync), database querying (db.users.findMany), and network interaction via the Apollo Gateway.
Sanitization: The skill proactively addresses these risks in references/security.md by providing templates for input validation using zod, query depth limiting with graphql-depth-limit, and query complexity analysis to prevent resource exhaustion.

graphql-architect