About Ralph

ralph runs a sequence of prompt files in a loop, re-running each with a fresh agent context until the work reports DONE. That’s the whole tool.

It owns the loop and the budget; your prompts own the work. Each prompt ends its turn with one status, and that status is the entire contract between you and ralph:

Status
CONTINUE	Run the same prompt again, fresh context.
NEXT	Advance to the next prompt (wrapping past the last back to the first).
DONE	Stop; exit 0.

Anything else is a bad turn, retried up to --max-retries.

ralph is a primitive building block. It knows nothing about your task, your language, or what “done” means — the prompts decide all of it. Fresh context every turn means nothing survives in memory, so all state lives in files you choose. Swap the prompts and the same binary drives an entirely different job.