Ralphi

galmuri:ralphi — quality-loop discipline

Etymology — Ralph Wiggum (The Simpsons): persistent retry until it works. The -i suffix marks the inspection family.

The lean discipline-gate version. Bounds token spend per cycle and forbids philosophy regressions from sneaking in between cycles. If you need the full file-type-aware inspection skill (criteria-prd / criteria-skill / criteria-script / criteria-code + Necessity Check), use harnish:ralphi.

If invoked bare (no cycle goal), ask: "What is this cycle's narrow goal? e.g. migrate 3 callers off legacy adapter, remove 5 dead exports — name something verifiable in one cycle."

Cycle phases (budget pattern)

#	Phase	Budget	What runs
1	Setup	≤ 500 tokens	State the cycle goal in one sentence; list affected files via rg -l (not rg); confirm baseline branch + SHA. No file reads.
2	Worker Dispatch	≤ 4 parallel sub-agents, each ≤ 2000 tokens out	Each worker emits {file, change, 3-line summary} — no full-file contents in worker output.
3	Verifier	≤ 1000 tokens	One verifier sub-agent re-reads ONLY changed lines via rg -A 3 -B 3 'pattern' <file>; never full files. Apply the Necessity Check (below) per diff hunk.
4	Philosophy Sweep	≤ 200 tokens	rg sweep across the new diff for ROP / version-name violations (below). Halt if any hit.
5	Commit	n/a	One commit per cycle. Never batch multiple cycles into a single commit.
6	Handoff	≤ 300 tokens	5-line summary: what changed, what verified, next cycle's narrow goal. Then advance counter or pause.

If any phase exceeds its budget, write tmp/ralphi-state-<n>.json (cycle counter, goal, pending files, last verifier output) and pause. Next invocation resumes from that state.

Necessity Check (Socratic 4 questions)

Apply to each diff hunk in the Verifier phase. Lifted from the original harnish:ralphi — the only inspection method that gates on purpose before correctness.

1. Why is this here? — (purpose)
2. What is it composed of? — (composition, scaffold for Q1/Q4)
3. What is it really, stripped of framing? — (truth, scaffold for Q1/Q4)
4. What happens if it's removed? — (necessity)

Verdict. Q1 and Q4 are the only gates. Q2 and Q3 are scaffolding to reach honest answers. Items failing Q1 or Q4 → [warning] unjustified existence (kept only by convention / inheritance / authority).

Fix discipline. Subtraction is the default. For [critical/warning], minimal additive patches are allowed (null check, missing import, type fix, error guard). Do NOT introduce new abstractions, modules, or features. If a fix would require new abstraction/module/feature → tag [unfixed] structural change required and hand off to the user.

Philosophy auto-check (between cycles)

Run between cycles, before advancing the counter. Halt the loop and surface to the user on any hit.

# ROP — defensive timeouts / silent catches added by this cycle
git diff HEAD~1 | rg -n 'withTimeout|timeout:\s*\d+|catch\s*\([^)]*\)\s*\{[^}]*//' \
  && echo "VIOLATION: possible ROP breach in last cycle"

# SSOT — hardcoded version-specific field names
git diff HEAD~1 | rg -n 'v\d+\.\w+|schema_v\d' \
  && echo "VIOLATION: possible hardcoded version"

Baseline note: HEAD~1 assumes the previous cycle already committed. If running against a feature branch with multiple cycles, prefer git diff $(git merge-base HEAD main)..HEAD for the full branch range.

Halt conditions

• Philosophy sweep hit → halt, surface, wait for user judgement
• Worker / verifier exceeds its phase budget → write state, pause
• Necessity Check tags [unfixed] structural change required → halt the autonomous loop, hand off to user
• Two consecutive cycles fail the verifier on the same hunk → halt (convergence failure)

What this skill does NOT do

• ✗ Type detection + criteria-file matching (that's harnish:ralphi)
• ✗ Run tests by itself — assumes the caller's loop drives the runner
• ✗ Auto-fix philosophy violations — surfaces them for user judgement
• ✗ Replace galmuri:polish (completion-claim gate) or galmuri:self-audit (single-session audit)