The Agent Skills Directory

[SAFE]: The skill outlines a plan to refactor reviewer transport mechanisms into a centralized Go-based contract. This approach simplifies the architecture and ensures that security enforcement (validation, schema checking, and admission) is handled by native code rather than disparate runtime adapters.\n- [INDIRECT_PROMPT_INJECTION]: The skill manages a potential attack surface by defining how the agent processes untrusted model output. Ingestion points include the ProviderReviewRequest which carries prompts and evidence to models. Boundary markers are explicitly established by defining runtime output as advisory and non-authoritative. Capability inventory includes the invocation of reviewer models via thin adapters. Sanitization is enforced through mandatory strict JSON extraction and schema validation in the native Go backend.\n- [SAFE]: The proposal to remove legacy isolation mechanisms, such as OpenCode specific environment variables and process restarts, is documented as a transition to a more robust validation-based security model. This is a technical design choice consistent with the stated architectural goals.

rdd-advisory-transport