The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes shell scripts (setup-ralph-loop.sh, stop-hook.sh) to manage the iterative loop state. These scripts employ common system utilities like jq, perl, sed, and awk to parse agent transcripts and maintain a local state file in the .ralph directory. This is standard behavior for managing session state.\n- [INDIRECT_PROMPT_INJECTION]: The skill implement an iterative loop by reading the agent's transcript and re-injecting the original prompt. The implementation is robust as it extracts the original prompt from a local state file rather than taking instructions from the assistant's output, and it uses jq to safely format the response back to the agent platform.\n- [SAFE]: The loop mechanism includes several safety features, including a maximum iteration limit (--max-iterations), a manual cancellation command (/cancel-ralph), and explicit instructions for the agent regarding completion promises. The technique is based on a public development methodology and the code originates from a trusted internal source.

ralph