stepbook

About

stepbook runs a long agent task as a series of short steps, each a fresh Claude Code, Codex or OpenCode session that reads a small validated saved state instead of the whole conversation. Prompt size therefore depends on the state, not on the number of prior steps, and tasks can be paused, given notes, resumed or handed to another agent. Built-in skills cover checklists and eval loops, and you can convert your own SKILL.md.