The Agent Skills Directory

Architectural Guidelines for Lifecycle Management: The document provides detailed instructions for developers on how to extend agent behavior using a contribution model. This approach promotes modularity and separation of concerns by moving side effects out of central orchestration classes.
Implementation of Security Gates: The guide highlights the onIncomingRequest hook, which is specifically designed to perform admission checks and validation. This hook 'fails closed', meaning any error during validation results in the rejection of the request, which is a defensive programming best practice to prevent unauthorized or malformed turns.
Context and Instruction Management: While the skill describes hooks like onOutgoingTurn that can inject instructions or modify message text, these are presented as managed architectural extensions for the host application rather than untrusted user inputs. The documentation encourages using these channels for enrichment while warning about potential costs like prompt cache invalidation.
Standard Dependency Injection and State Handling: The patterns described, such as using an IInstantiationService and contribution-scoped mementos, are standard for large-scale application development and do not introduce unusual security risks.

agent-host-chat-contributions