The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill documents a 'ralph-loop' pattern which uses a local file to maintain agent state and control flow across iterations. \n
Ingestion points: The markdown body of .claude/ralph-loop.local.md is extracted and fed directly back into the agent's prompt as the 'reason' for a tool-use decision (references/real-world-examples.md). \n
Boundary markers: The implementation lacks delimiters or system instructions that would help the agent distinguish between its own logic and instructions embedded within the injected file content. \n
Capability inventory: The documented patterns utilize tools like run_command for shell script execution, write_to_file, and inter-process communication via tmux. \n
Sanitization: While the skill provides examples of sanitizing file paths for traversal, it lacks sanitization for the prompt content injected back into the agent's context.\n- [COMMAND_EXECUTION]: The skill documents techniques for automating actions based on configuration, including inter-process communication. \n
Evidence: The multi-agent-swarm example uses tmux send-keys to send strings to other terminal sessions based on values read from a local configuration file (references/real-world-examples.md). If these values (e.g., agent_name) are not properly sanitized, they could be used to inject and execute arbitrary commands in the target tmux session. \n
Evidence: Multiple provided scripts demonstrate parsing and executing logic based on user-editable local files using standard shell utilities like sed, grep, and awk.\n- [DYNAMIC_EXECUTION]: The skill documents how to generate and update executable content at runtime. \n
Evidence: The skill provides patterns for agents and commands to use write_to_file to create and update shell scripts and configuration files that are subsequently executed or parsed by hooks and agents.

plugin-settings