| name | ralph |
| description | In-session OMA persistence loop — invoke /oh-my-agy:ralph; keep going until PRD stories pass with verify evidence (CLI optional) |
| argument-hint | <task description> |
ralph (OMA / in-session)
You are already in the agent session
When invoked via /oh-my-agy:ralph or this ralph skill, treat $ARGUMENTS as the task and run the persistence loop HERE until stories pass with verification evidence.
- Do not require
oma ralph -- … / terminal / SID as the default path.
- Outer
oma ralph is an optional managed launch (see Appendix); it is not required to start.
- Canonical slash:
/oh-my-agy:ralph.
Purpose
Self-referential persistence loop until work is actually done. Sibling of OMC/OMX Ralph:
- Inner loop (primary): you keep going in this session until stories pass + verify skill is satisfied
- Outer launch (optional):
oma ralph -- <task> for managed directive + skill protocol
Use when
- User invokes
/oh-my-agy:ralph or says ralph / don't stop / keep going until done / must complete
- Work needs story tracking and anti-premature-stop
Do not use when
- Full product-from-idea pipeline →
/oh-my-agy:autopilot
- Pure parallel fan-out without persistence →
ultrawork
- Read-only research →
search
PRD (required)
Maintain a PRD file (create if missing):
- Preferred:
.agy/ralph/prd.json (workspace)
- Session note: update progress in
.agy/ralph/progress.md
Schema (minimal):
{
"schemaVersion": 1,
"goal": "<task>",
"stories": [
{
"id": "US-001",
"title": "…",
"acceptanceCriteria": ["specific, testable criterion"],
"passes": false
}
]
}
Startup gate: replace any generic criteria (“implementation is complete”) with task-specific criteria before coding.
Steps (in-session)
- Init PRD — refine stories + acceptance criteria from
$ARGUMENTS.
- Pick next story — highest priority with
passes: false.
- Implement — only that story; record files touched.
- Verify story — for each criterion, produce fresh evidence.
- Mark passes — only when all criteria for the story are evidenced.
- Loop until all stories
passes: true.
- Final verify — run
verify skill on the whole change set.
- Complete — summary + evidence; do not stop after first story.
Anti-patterns (forbidden)
- Claiming done because code “looks good”
- Deleting/skipping tests to go green
- Stopping after an intermediate approval without final verify
- Silent scope reduction
- Telling the user to open a terminal and run
oma ralph before working
Stop conditions
- Hard blocker (credentials, ambiguous product decision) → report + wait
- User cancel →
cancel skill
- Same root failure 3 iterations → escalate with diagnosis
Final checklist
Appendix: optional oma CLI
Use only when the user wants a managed outer launch:
oma ralph -- "<task description>"
Managed binding injects skill protocol when configured. The in-session PRD under .agy/ralph/ remains the progress source of truth for this loop.