The Agent Skills Directory

[COMMAND_EXECUTION]: The skill performs file deletion and process termination using shell commands that interpolate variables derived from local state files.\n
Evidence: The implementation steps in SKILL.md include instructions like rm -rf .rcs/state/team/{name}/ and tmux kill-session -t rcs-team-{name} where {name} is extracted from a configuration file.\n
Risk: If the team_name value within config.json is manipulated to include path traversal sequences (e.g., ../../), it could lead to the unintended deletion of files outside the designated state directory or the termination of unrelated tmux sessions.\n- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its reliance on untrusted external state data to drive its logic.\n
Ingestion points: The skill reads various JSON files from the .rcs/state/ directory, including config.json, autopilot-state.json, and forge-state.json, to determine mode status and cleanup targets.\n
Boundary markers: There are no explicit delimiters or 'ignore' instructions provided to the agent to distinguish between control logic and data ingested from these state files.\n
Capability inventory: The skill possesses significant local capabilities, including recursive file deletion (rm -rf), inter-process communication via tmux send-keys, process termination, and the ability to modify workspace documentation (AGENTS.md).\n
Sanitization: The skill logic does not explicitly include steps to sanitize or validate keys and values read from state files before they are interpolated into shell commands or documentation updates.

cancel