The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted inputs generated by other agents, creating an exposure to indirect prompt injection where malicious content in an agent's output could attempt to influence the LLM judge's ranking or justification.
Ingestion points: Reads the output of git diff and agent result posts from .agenthub/board/results/agent-{i}-result.md as described in SKILL.md.
Boundary markers: Absent; there are no instructions or delimiters specified to partition untrusted agent data from the judge instructions.
Capability inventory: Executes local Python scripts (result_ranker.py and session_manager.py) to manage sessions and handle ranking.
Sanitization: Absent; no validation or filtering of the agent-generated text or diffs is mentioned before processing.
[COMMAND_EXECUTION]: The skill invokes local command-line operations to execute its ranking logic and update session state. It also accommodates a configurable evaluation command (--eval-cmd "{eval_cmd}") designed to execute within each agent's worktree to obtain metric data.

eval