ralph

name	ralph
description	Single-owner execute-fix-verify loop for one clear task. Use with /ralph when one agent should keep going until evidence or blocker.

Ralph

Use /ralph when one owner should complete one clear task end-to-end.

When to use

• A plan or concrete task already exists
• The work is a single logical unit (not parallelisable)
• You need persistent execution until done or blocked

Do not use when

• Work has independent parallel lanes — use /team
• 5+ mechanical tasks — use /ultrawork
• No plan exists yet — use /ralplan first

Input

Accept a plan from /ralplan, a ticket, or a concrete task description. If a /ralplan plan exists, use its acceptance criteria as your definition of done.

Steps

1. Register the loop FIRST — before editing any file, run omp ralph start "<task>" --max-iterations 10. This is mandatory even for small tasks: it tracks the loop so omp ralph status/cancel and /team nudges can see it and the iteration cap is enforced. Skipping it leaves the loop invisible to the CLI. State what "done" looks like, then start writing code.
2. Implement one slice at a time, in plan order.
3. Verify after each slice — run tests, lint, type-check. Do not batch verification to the end.
4. Fix any failures immediately before moving to the next slice.
5. Record slice completion — after a verified slice, omp ralph tick may be used for completed-slice bookkeeping only. The agent-stop hook is the sole writer of the loop iteration counter and enforces the cap.
6. Repeat until all slices complete with evidence, or a blocker is hit.
7. Final verification — run the full test suite one last time after all slices.
8. End the loop — run omp ralph cancel once done or blocked, so the tracked state is cleared.

Circuit breaker

If the same issue fails 3 times after 3 different fix attempts, stop. Report:

• What was tried
• Why each attempt failed
• What information is missing

Do not keep trying the same approach. Escalate the blocker.

Scope freeze

If you discover work outside the original plan:

• Note it in your output under "Known gaps"
• Do not chase it — finish the planned work first
• If it blocks the plan, stop and report

Rules

• Never claim "done" without running verification
• Commit working increments — don't batch everything into one commit
• Read test output — don't assume green means pass

Final checklist

Before claiming done:

• Every plan slice implemented and verified individually
• Full test suite passes after all slices
• No lint or type errors introduced
• Evidence of completion attached (test output, build log)

Output

• Done — what was completed
• Evidence — test output, build logs, or behaviour proof
• Known gaps — anything intentionally left or discovered but out of scope