Ralph Executor
Stateless multi-session execution that survives any context break
Overview
The execution backbone for complex, multi-phase projects. Each invocation reads state from a PRD file, executes one action, updates state, and signals completion. Zero state held in memory. 42 PRDs, 300+ iterations, zero lost state.
Execution Phases
State Read
Reads current phase, iteration count, and next action from PRD file
Pre-Flight Check
Validates context budget, tool availability, and scope boundary
Execute Action
Performs exactly one action toward the next success criterion
State Write
Updates PRD state section, iteration log, and running learnings
Signal
Outputs CONTINUE, TASK_COMPLETE, or BLOCKED_NEED_HUMAN
Outputs
- Updated PRD state
- Iteration log entry
- Completion signal
Example
A 24-phase SEO campaign spans 3 weeks of sessions. Each iteration reads the PRD, performs one action (write an outline, build a page, deploy), updates state, and signals ready for next. Session breaks, context compaction, agent handoffs: none of them lose state.
Deep Dives
Ready to deploy Ralph Executor?
This agent ships with Knowledge OS. Set up your system in 90 minutes.
Built and maintained by Victor Sowers at STEEPWORKS