The Agent Skills Directory

[COMMAND_EXECUTION]: The skill performs multiple shell commands using git and python, utilizing string interpolation for parameters like {base_branch}, {winner}, {session-id}, and {agent-id}. If these variables contain unsanitized input from previous agent evaluation steps or user arguments, it could lead to command injection.
Evidence: git checkout {base_branch}
Evidence: git merge --no-ff hub/{session-id}/{winner}/attempt-1
Evidence: python {skill_path}/scripts/session_manager.py --cleanup {session-id}
[INDIRECT_PROMPT_INJECTION]: The skill processes session metadata (winner, session ID, task description) to automate the merge process, which presents a surface for indirect prompt injection.
Ingestion points: Session ID, winner name, and task description are retrieved from the results of a previous evaluation step (/hub:eval) or user input.
Boundary markers: None. The variables are directly interpolated into bash commands and markdown reports without delimiters or sanitization instructions.
Capability inventory: The skill can perform git checkouts, merges, tagging, and branch deletions. It also executes a local Python script (session_manager.py) and writes summary files to the .agenthub/ directory.
Sanitization: None. The skill templates assume the interpolated values are well-formed and safe for use in shell commands and markdown.

merge