Hon maguro

hon-maguro

Binary quality gate: a project either meets every fixed bar (HON MAGURO) or it Fails. No partial credit. Bars → STANDARDS.md. Anti-gaming → MEET.md.

bin/maguro is a self-contained Bash CLI (curl + python3 only). Scripts ship next to this skill.

Paths

Skill root (after install): ~/.agents/skills/hon-maguro/

SKILL="$HOME/.agents/skills/hon-maguro"
"$SKILL/bin/maguro" eval  <path>
"$SKILL/bin/maguro" init  <path>
"$SKILL/bin/maguro" links <path>
"$SKILL/bin/maguro" help
"$SKILL/bin/lighthouse-median" <url> [runs]

From a clone of this repo: ./bin/maguro (wrapper) or ./skills/hon-maguro/bin/maguro.

Invoke

bin/maguro eval  <path>   # full eval — writes HON-MAGURO-EVAL-REPORT.md
bin/maguro init  <path>   # drop blank HON-MAGURO-EVAL.md
bin/maguro links <path>   # SCORE3 only (curl, no install)
bin/maguro help

bin/lighthouse-median <url> [runs]   # SCORE1: Performance median of 3 (=100 to pass)

Typical flow:

1. bin/maguro init <path> — blank EVAL form (bar IDs + evidence cells).
2. Fill each evidence cell with a number, count, or binary (never prose).
3. For web UIs: bin/lighthouse-median <url> → paste Performance median into SCORE1.
4. bin/maguro eval <path> — per-bar table + verdict + fail tips.

What eval checks

• Auto: CODE3 (strict + 0 any), REPO1 (secrets), REPO2 (deps ≤20), SCORE3 (links).
• HUMAN (evidence required): SPEED1–5, TEST1, SCORE1, SCORE2, DOCS1, CODE1, CODE2, REPO3, OPS1–6. Confirm against STANDARDS.md.

Reading the output

• Status: PASS / FAIL / HUMAN.
• result: HON MAGURO only when FAIL=0 and HUMAN=0.
• On FAIL: how to pass: tips (also in HON-MAGURO-EVAL-REPORT.md).

SCORE1 — Lighthouse

Bar: Performance median of 3 runs = 100 (N/A if no web UI).

bin/lighthouse-median https://localhost:3000
bin/lighthouse-median https://myapp.com 3

• Shells out to npx lighthouse (no package.json dep); needs Node + Chrome; median via python3.
• Paste e.g. Lighthouse perf median=100 (3 runs) — https://myapp.com into SCORE1 evidence.
• Real product URL only; fix root cause — never suppress audits.

Inspect → fix (capped, max 2 rounds)

Not open-ended. Infinite "until green" invites cheating the meter.

bin/maguro init  <path>   # if needed
bin/maguro eval  <path>   # round 0 baseline
# fix real FAILs (product, not meter)
bin/maguro eval  <path>   # round 1
# optional second fix
bin/maguro eval  <path>   # round 2 — last

Stop when: HON MAGURO · 2 fix cycles done · remaining is HUMAN/measure · only MEET-violating options left.

Allowed fixes (product, not meter)

FAIL	Honest fix
CODE3	"strict": true; replace any / as unknown as with unknown + narrowing
REPO1	Untrack .env; remove secret-shaped strings; rotate if real
REPO2	Drop unused deps; or one-line why for each past 20 in EVAL
SCORE3	Fix broken links (maguro links)
TEST1	Fix failures/flakes; add flow tests — never delete tests to go green
DOCS1	README alone boots a stranger (commands exit 0)
CODE1/2	Glossary↔schema 1:1; one auth path; 0 user-facing stack leaks
SPEED*	Prefetch first; then real perf — measure every product URL
SCORE1	bin/lighthouse-median; fix root cause to median 100
SCORE2	Fix React Doctor findings — don't disable rules
OPS*	Real napkin/prod numbers; prove restore once

Forbidden (MEET)

• Secret eval mode / cherry-picked routes
• Suppress Lighthouse/Doctor/lint/tests to fake 100
• @ts-ignore, blanket any, delete tests, fake N/A
• Invent SPEED/OPS/Lighthouse numbers
• Loop past 2 fix rounds

Report after fix rounds

rounds: <n>/2
result: HON MAGURO | FAIL
fixed: <bar ids>
still open: <FAIL + HUMAN needing measure>
next human step: <one line or none>

Rules (MEET.md)

• Same path users take.
• Fix the product, not the meter.
• Named exception + measured number beats a fake pass.