The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to use terminal commands like nc -U and cat to communicate with a local Unix domain socket at ~/.codegrid/socket. This interface is used to control the CodeGrid application via a JSON-RPC API.
[PROMPT_INJECTION]: The skill implements a multi-agent orchestration pattern where the agent reads output from sibling panes using the read_pane tool. This creates an indirect prompt injection surface where instructions embedded in the output of one agent could influence the behavior of the orchestrator.
Ingestion points: Terminal output from sibling agent sessions is ingested via the read_pane MCP tool.
Boundary markers: There are no technical delimiters (such as XML tags or Markdown code blocks) defined in the skill's operational logic to isolate ingested content from instructions, although the documentation advises using natural language scope constraints.
Capability inventory: The agent has the ability to message other agents (message_agent), interact with the local control socket, and potentially execute shell commands in sibling panes.
Sanitization: No explicit sanitization, filtering, or validation of the content retrieved from other agents is implemented within the skill's protocol.

codegrid