The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The agent is designed to process tasks (vimarshas) and messages from an external inbox and channel, which constitutes an ingestion surface for untrusted data.\n
Ingestion points: Data enters the agent's context through the repository's vimarsha inbox and the live communication channel (SKILL.md).\n
Boundary markers: The instructions mandate that agents 'grasp' the neighborhood of a task and use 'reality-audit' to verify results independently of the implementer's narrative.\n
Capability inventory: The agent can perform git operations (commit, push, merge), modify internal graph nodes via nks_* tools, and interact with socket-based channels.\n
Sanitization: The skill requires 'Answered when' criteria for tasks and mandates that destructive actions or transcendent will require explicit human sanction.\n- [PERSISTENCE_MECHANISMS]: The skill implements an autonomous 'watch' cycle designed to maintain agent activity over time.\n
The agent is instructed to use webhooks and recurring polling to trigger its work tacts.\n
It maintains a live socket connection to ensure responsiveness to incoming messages.\n
These persistence mechanisms are the core intended functionality of the skill and do not serve a malicious purpose.\n- [DATA_EXPOSURE_AND_EXFILTRATION]: The skill includes guidelines for managing socket URLs, which function as credentials.\n
It warns that whoever holds the socket can read messages addressed to the agent and provides advice on secure connection management and rotation.

autonomous