MDT Consultation

What to check before installing: - Confirm platform capabilities and permissions: does your agent platform support sessions_spawn by agentId and asynchronous completion events? Who can spawn the referenced agents (shield, law, devil, etc.)? Ensure the skill won't unexpectedly gain access to agents you don't want invoked. - Decide whether automatic archiving of MDT reports to memory/mdt-reports/... is acceptable for the kinds of documents you'll submit (contracts, speeches). If not, request a setting to disable or encrypt persisted reports, or require explicit user consent before saving. - Resolve the runtime contradiction: clarify whether the integration should use sessions_yield (blocking wait) or the NO_REPLY/asynchronous pattern. Ambiguity can lead to inconsistent behavior and missed outputs. - Be aware of the 'CEO fallback' behavior that instructs the host agent to 'temporarily fill in' missing agent outputs: this can introduce fabricated or overconfident content. If you need strict provenance, require that missing agent perspectives be marked and not auto-filled. - Test in a sandbox with non-sensitive inputs: verify agent spawning, timeouts, retry behavior, and what gets stored in memory. Confirm logs and stored reports are accessible only to intended principals. If these questions are answered and storage/privilege concerns are addressed, the skill's design is plausible; otherwise treat it as risky and restrict or decline installation.