The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/result_ranker.py executes a user-provided evaluation command using shell=True. While this is a core feature intended for running tests or benchmarks, it allows for the execution of arbitrary shell commands in the agent's environment. Evidence: subprocess.run(eval_cmd, shell=True, ...) in scripts/result_ranker.py.
[PROMPT_INJECTION]: The skill's architecture relies on a coordinator reading and processing result summaries and progress updates written by potentially untrusted subagents to the .agenthub/board/ directory, creating a surface for indirect prompt injection. Ingestion points: Result summaries located in .agenthub/board/results/ and progress updates in .agenthub/board/progress/ (accessed via scripts/board_manager.py and scripts/result_ranker.py). Boundary markers: Absent; data is stored in markdown files without explicit delimiters or structural isolation to prevent subagent content from being interpreted as instructions by the coordinator. Capability inventory: The coordinator possesses the capability to execute shell commands via scripts/result_ranker.py, perform git merges, and manage the filesystem via worktree commands. Sanitization: Absent; the skill does not perform validation or sanitization on the content written by subagents before it is processed by the coordinator.

agenthub