ralph
Self-referential persistence loop — keep working until a verifier confirms the task is genuinely complete.
Ralph is a persistence loop that drives a task to full verified completion. Where a one-shot executor does its best and stops, Ralph keeps iterating: it executes work in parallel, runs verification with fresh evidence, and loops back to fix any failures until an architect-level review approves the result. The phrase "the boulder never stops" captures the contract — Ralph does not quit on partial progress.
When to use
- The task requires guaranteed completion, not a best-effort attempt
- You say "ralph", "don't stop", "must complete", "finish this", or "keep going until done"
- Work may need multiple retry rounds and you want automatic persistence across them
- You want parallel execution with mandatory architect sign-off before the loop exits
- You are following up after
$ralplanapproved a plan and want sequential execution pressure
How to invoke
Natural language triggers: "ralph", "don't stop", "must complete", "keep going until done", "finish this".
Explicit slash: $ralph
codex
> ralph implement the caching layer from the approved plancodex
> don't stop until the auth module passes all testsWhat happens
Ralph begins with a pre-context intake: it loads or creates a context snapshot at .omx/context/{slug}-{timestamp}.md so every iteration shares the same grounded brief. It then reviews any in-progress TODO items and picks up where the last iteration left off. Independent tasks are delegated simultaneously to specialist agents at appropriate tiers — simple lookups at LOW, standard work at STANDARD, and complex analysis at THOROUGH — so parallel work runs concurrently rather than sequentially. Long operations like builds and test suites run in the background. Once all tasks are complete, Ralph collects fresh verification evidence: it runs tests, builds, and lint checks, then reads the output rather than assuming success. An architect agent reviews the result at STANDARD tier minimum; for large or security-sensitive changes the tier rises to THOROUGH. If the architect rejects, Ralph fixes the issues and re-verifies. After architect approval, a mandatory deslop pass cleans AI-generated noise from all changed files, followed by a regression re-run to confirm the pass did not break anything. Only then does Ralph exit cleanly with /cancel.
Outputs
- Source code changes implementing the assigned task
.omx/context/{slug}-{timestamp}.md— context snapshot used across all iterations- Fresh test, build, and lint output as verification evidence
- Architect approval on record before exit
- Clean state via
/cancelon successful completion
Related skills
$ralplan— consensus planning to produce a clear brief before ralph starts$autopilot— full pipeline including ralph as the execution phase$team— parallel coordinated workers when a task needs multiple simultaneous lanes