The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/result_ranker.py script uses subprocess.run with shell=True to execute evaluation commands (e.g., benchmarks or test suites) defined in the session configuration. While intended for developer workflows, this facilitates the execution of arbitrary shell commands in the coordinator's environment.
[DYNAMIC_EXECUTION]: The evaluation workflow executes code within worktrees where sub-agents have made arbitrary file modifications. This results in the coordinator executing code generated or altered by an AI agent, which is a vector for unintended code execution.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: Sub-agents modify project source files and test scripts in isolated worktrees (SKILL.md).
Boundary markers: The skill includes documentation warning users to review evaluation commands and candidate results before execution, though the /hub:run command automates the entire lifecycle.
Capability inventory: The skill can execute arbitrary shell commands via result_ranker.py, modify the local filesystem, and perform complex Git DAG operations.
Sanitization: No automated sanitization or sandboxing (beyond filesystem isolation) is applied to the agent-generated code before it is executed by the coordinator's evaluation script.

agent-hub