The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The documentation suggests using the aconn notify utility from the github.com/Paldom/agents-connect repository to handle event hooks. This is a vendor-owned resource belonging to the skill author (paldom).
[INDIRECT_PROMPT_INJECTION]: The skill's primary function is to ingest external human input ("verdicts") to guide agent behavior, which creates a potential surface for indirect prompt injection.
Ingestion points: Human responses land in the agent's thread and are processed to extract verdicts.
Boundary markers: The protocol uses explicit ASK: and OPTIONS: markers to structure requests to the human.
Capability inventory: A human verdict can authorize actions, release blocked turns, and influence the agent's next dispatch body.
Sanitization: The system implements an "action hash" that digests the operation and its parameters. It also pins the policy and run versions. If any of these values change while waiting for a human response (indicating state drift), the verdict is invalidated and the agent is forced to re-ask. This prevents the agent from executing actions based on outdated or misaligned approvals.

army-hitl