| name | ralph |
| description | PRD-loop persistence mode: iterate BUILD until verifier passes or max blocked attempts. Triggers ralph keep going keywords and /ohc-ralph. Deactivate via stop ralph.
|
| allowed-tools | Read Write Bash |
Ralph persistence
Operate with scripts/ralph.js and session state files.
Lifecycle
- Activate via keyword
"ralph", "keep going", or /ohc-ralph. On first run without an active PRD, the script initializes .ohc/state/sessions/<id>/prd.json from the current task narrative.
- Loop — execute BUILD iteration; verifier runs tests; persist
RESULT-style narrative in worker notes where applicable.
- Advance — on success mark story done and move to next
prd.stories[] entry; on repeated block update attempt_count; hard-stop after stopOnBlocked (keyword-map default 3).
- Deactivate — user says
"stop ralph" or "ralph stop" (clears active skill flag in .ohc/state/active-skills.json).
State files
.ohc/state/ralph-state.json — active session id, story index, iteration counters
.ohc/state/sessions/<id>/prd.json — { goal, stories:[{title,status,attempt_count}], ... } per scripts/ralph.js
Guardrails
- Do not claim DONE without verifier-aligned evidence.
- If stories are ambiguous, widen them only after syncing with
writing-plans or user OK.