The Agent Skills Directory

[COMMAND_EXECUTION]: The metric-based evaluation mode constructs a shell command that includes a placeholder {eval_cmd}. This command is passed to a local script result_ranker.py which is intended to execute the specified evaluation command. If the command string is populated from untrusted sources, such as user-defined configuration or external session data, it could lead to arbitrary command execution on the host system.
[PROMPT_INJECTION]: The skill implements an 'LLM Judge' mechanism that ingests untrusted data to perform rankings, creating a surface for indirect prompt injection.
Ingestion points: The skill reads agent-generated markdown files from .agenthub/board/results/ and examines git diff outputs.
Boundary markers: There are no explicit instructions or delimiters defined to prevent the LLM from following instructions embedded within the results being evaluated.
Capability inventory: The skill has the capability to execute shell commands and modify session states via session_manager.py.
Sanitization: No sanitization or validation of the ingested agent results is mentioned, allowing potential malicious content within agent outputs to influence the judge's behavior.

eval