npm package discovery and stats viewer.

Discover Tips

  • General search

    [free text search, go nuts!]

  • Package details

    pkg:[package-name]

  • User packages

    @[username]

Sponsor

Optimize Toolset

I’ve always been into building performant and accessible sites, but lately I’ve been taking it extremely seriously. So much so that I’ve been building a tool to help me optimize and monitor the sites that I build to make sure that I’m making an attempt to offer the best experience to those who visit them. If you’re into performant, accessible and SEO friendly sites, you might like it too! You can check it out at Optimize Toolset.

About

Hi, 👋, I’m Ryan Hefner  and I built this site for me, and you! The goal of this site was to provide an easy way for me to check the stats on my npm packages, both for prioritizing issues and updates, and to give me a little kick in the pants to keep up on stuff.

As I was building it, I realized that I was actually using the tool to build the tool, and figured I might as well put this out there and hopefully others will find it to be a fast and useful way to search and browse npm packages as I have.

If you’re interested in other things I’m working on, follow me on Twitter or check out the open source projects I’ve been publishing on GitHub.

I am also working on a Twitter bot for this site to tweet the most popular, newest, random packages from npm. Please follow that account now and it will start sending out packages soon–ish.

Open Software & Tools

This site wouldn’t be possible without the immense generosity and tireless efforts from the people who make contributions to the world and share their work via open source initiatives. Thank you 🙏

© 2026 – Pkg Stats / Ryan Hefner

@superinstance/quilt-tools

v0.1.0

Published

Ten working Quilt tool prototypes on the vendored @quilt/core dist — plus a springboard lab for JEV/MothQuantum experiments.

Readme

quilt-tools

Ten working tools grown on the Quilt reactive spreadsheet engine, one shared harness, three GAN-bred bloodlines of logic, and a lab where the receipts get to mutate the tools that print them.

For the impatient: everything here runs offline, grades its own homework, and says so out loud. No engine build, no cloud, one dependency.

git clone https://github.com/SuperInstance/quilt-tools && cd quilt-tools
npm install                  # the only dep is yaml; the engine dist is vendored
node tools/fleet-pager.mjs   # expect: 7/7 checks green
npm run check                # syntax-check every tool and experiment

The ten tools

Each tools/*.mjs is self-contained — engine wiring, scenario, and a check harness at the bottom. 75 self-checks across the set; green is the only accepted color. Each tool also has a claim card in cards/ saying what it is, who it's for, and where it came from.

| tool | checks | what it actually does | |---|---|---| | fleet-pager | 7/7 | SRE paging with hysteresis band discipline — no more page-flapping at 3 a.m. | | ledger-seal | 6/6 | tamper-evident append-only witness ledger, fnv1a-chained | | ocean-recall | 7/7 | vector-note recall with honest forget — absence says "forgot," never "found" | | triagedesk | 8/8 | support triage routing | | budget-tide | 8/8 | cash-flow envelopes with dry-envelope fences | | home-ecos | 6/6 | home ecosystem automation | | driftwatch | 7/7 | config drift detection | | approvals | 9/9 | spend approvals with tier-based routing | | habit-atlas | 8/8 | habit streaks with momentum physics | | pipeline-guard | 9/9 | pipeline row validation + dead-letter replay |

These are not demos wearing tool costumes. They are the smallest full behaviors that still count: each one makes decisions, keeps receipts, and can tell you why it decided what it decided.

for f in tools/*.mjs; do node $f >/dev/null 2>&1 \
  && echo "${f##*/}: green" || echo "${f##*/}: FAIL"; done

The harness they stand on

src/toolkit.mjs — sheet(), check(), done(), SysOne heuristics, a deterministic embedder, witness-chain helpers. No laptop paths; the engine dist resolves via vendor/quilt-core/ (the exact upstream provenance is recorded in its own README there — vendored so the tools run hermetically and the lab can mutate a copy without touching upstream). Override with QUILT_DIST=/path/to/dist.

The fleet's honesty idiom for any claim is four verdicts: CONFIRMED (re-executed and it holds) · REFUTED (re-executed and it fails) · SIMULATED (asserted, never executed — a model, not a measurement) · MEASURED (executed, nothing claimed). These tools already book their refusals as receipts; coev's auditor names the other two dodges so no claim goes unlabeled.

The lab — where tools meet live models

experiments/ is the springboard: studies that run the heuristic tools against the fleet's typed live models and book whatever happens, including the refusals. The house rule: witness-booked, and REFUSED rows are kept, never retried away. Write-ups and receipts:

  • experiments/README.md — the live-oracle studies (S1: triagedesk vs live jev — 12/12 live calls, honest score-gap reported).
  • experiments/API_LIMITS_R1.md — what the live models tolerate, measured not guessed.
  • e1-ordinal-not-interval.mjs, e2-receipts-change-credit.mjs — instrument studies: how the receipts themselves bend the results (they do).
  • s2-driftwatch-jev.mjs, s3-quantum-tided-budget.mjs — more live pairings.
  • The fleet's newest instrument: coev — adversarial coevolution with a champion-integrity auditor. The gan-elites were bred by divergence (as different as possible, behaviorally exact); coev breeds by rivalry against a live champion and then audits the champion for hollowness. Sibling instruments, one honesty doctrine.

The referral graph — the mesh answers as a distribution

Casey's 2026-09-25 mandate: "not one thing but a distribution of intelligent referrals." src/referral_graph.mjs is the substrate: ideas are nodes, referrals are hash-chained LINK rows (fnv1a-64, the fleet receipt idiom), every edge carries the receipt that proves it — or the falsification condition that would kill it. The answer to a question is a ranked VIEW over modules, not a single confident pointer.

  • experiments/REFERRAL_GRAPH.md — the constitution: weight law (PENDING speculation weighs 0.05; VERIFIED merged provenance weighs 1.0), booking rules, kill switches.
  • referral_graph.seed.mjs / referral_graph.pins.mjs — the real v1 graph and its pin suite (--live audits every provenance PR against GitHub; offline pins are labeled SKIPPED, never passed silently).

The bloodlines — logic the loom bred

gan-elites/ holds crowned implementations bred by the quilt-loom Divergence Foundry: a GAN that breeds logic as different as possible from every prior bloodline while staying behaviorally exact. Each directory carries a vendored oracle, a frozen 200-probe corpus, a provenance banner (family, hash, novelty, voice), and a self-contained verifier:

node gan-elites/pager-band/verify.mjs      # 860/860 — divergent in form, exact in behavior
node gan-elites/witness-fnv/verify.mjs     # 832/832
node gan-elites/cosine-sparse/verify.mjs   # 828/828

Same animal, three skeletons. If you doubt equivalence, run the verifier — that is what it's for.

The site

site/index.html — a tool-picker UI over the ten. Open it, click around, break it, file the receipt.

Reading beyond this repo

  • The engine: SuperInstance/quilt — the reactive spreadsheet substrate everything here stands on.
  • The judges: jeviter (typed model receipts) and quilt-doctor (four diagnostic lenses over any quilt-shaped project — this repo is a patient).
  • The canon: AI-Writings — the essays, including the ones about why tools must grade themselves.
  • The face: SuperInstance.github.io — where the demos live.

Lab mandate (Phase 2+)

These prototypes plus the vendored engine are the substrate for JEV/MothQuantum experiments — instance-graph mutations as sheet diffs, midden harvesting as cell provenance, ZPP certification runs as witness chains. Work happens on branches; main stays green.


A tool that can't check itself is a rumor with a README. Nothing here is a rumor.