The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/result_ranker.py utilizes subprocess.run(eval_cmd, shell=True) to execute evaluation benchmarks. The eval_cmd is retrieved from a session-specific config.yaml file populated during initialization. This allows for the execution of arbitrary shell commands in the repository's context.
[REMOTE_CODE_EXECUTION]: By design, the skill allows users to configure evaluation commands (eval_cmd) that are executed during the ranking phase. This provides a surface for remote code execution if the configuration is sourced from or influenced by untrusted external data.
[INDIRECT_PROMPT_INJECTION]: The hub coordinator agent processes markdown files written by sub-agents to the .agenthub/board/ directory. This creates a surface for indirect prompt injection where malicious or compromised sub-agents could embed instructions into their progress or result reports to influence the coordinator's ranking or merge decisions.
Ingestion points: The coordinator agent reads contents from .agenthub/board/results/ and .agenthub/board/progress/ (referenced in agents/hub-coordinator.md).
Boundary markers: The system does not implement explicit boundary markers or "ignore instructions" delimiters when the coordinator reads agent-generated markdown.
Capability inventory: The coordinator has access to git tools and is instructed to execute the result_ranker.py script, which triggers shell commands.
Sanitization: While board_manager.py sanitizes author names for use in filenames, there is no evidence of sanitization or filtering for the body content of the markdown posts.
[DYNAMIC_EXECUTION]: The scripts/dry_run.py script dynamically executes other Python files within the skill package using sys.executable to verify that their --help menus exit correctly, which is a form of internal dynamic code execution.

agenthub