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

@tjakoen/grain

v0.1.24

Published

GRAIN — an AI-interaction design system where every surface is operable by both a human and an AI through one shared vocabulary.

Readme

🌾 GRAIN: usage

Made with Claude npm License: Apache 2.0

GRAIN is the AI-interaction design system: a UI where every surface is addressable and operable by both a human and an AI through one shared vocabulary, with the AI's presence shown as a visible signal (grain = AI). It runs on a substrate (BATCH is the reference one) but imports nothing from it.

  • The beliefs behind it: PHILOSOPHY.md
  • Overview / why: GRAIN
  • The full contract (intent envelope, render ops, manifest, the AI-acts protocol, the two write paths): AI-INTERFACE
  • Built on top of GRAIN: MILL, the Markdown→pages CMS (a layer above).
  • Roadmap / planned features: PLAN.md

Docs live at tjakoen.github.io/grain/docs. GRAIN's explanatory docs are canonically homed in the portfolio repo (tjakoen.github.io/docs/grain/) and rendered there through MILL. This repo keeps its components, code, PLAN, and this README; it no longer ships a docs/ folder.

This file is the practical reference: what a host must provide, the markup conventions, and how to wire it.

What GRAIN gives you (the full, tiered list is the source in GRAIN §"What GRAIN gives you"):

  • Hero: one vocabulary, two operators (human + AI, one door → RenderOps) · grade-as-signal (the AI's presence is visible) · self-documenting (a live /catalog).
  • Also: the manifest (a drift-proof machine map the AI reads) · auditable by design (every interaction is a source-tagged Intent → an interaction log is a server-side drop-in) · the design system works with no AI · a client-side door (the same vocabulary runs on static hosts, no server) · re-skin by token override · machine-readable (SEO/AEO) · built on the modern web platform (native View Transitions, <dialog>, <details>, :has(), color-mix(), not framework JS).

0. Two layers: the AI interface is optional

GRAIN is two things, one-directional:

  • The design system (always usable): the b-* atoms, the default theme (styles/variables.css = tokens + @font-face, styles/global.css = base/skin, fonts/ = the Redaction grades), and the grade-as-signal mechanism (styles/grain.css: the --type-font atom, data-grade / .field / data-commit rules, caret, settle). Grade is useful with no AI at all: draft vs. saved, focus/editing, in-transit vs. committed. It depends on nothing in ai/.
  • The AI-interaction layer (opt-in): ai/* (the door, contract, reasoner boundary, manifest, accepts), the dispatcher island scripts/ai-dispatch.js, and its styling ai/ai.css (the "AI is acting" spotlight). This layer uses the design system (sets data-grade by provenance); the design system never reaches back into it.

So you can adopt just the design system in a plain BATCH app: link GRAIN's three stylesheets (variables.css → global.css → grain.css) + the b-* atoms, and skip ai/ entirely. It looks like GRAIN out of the box; override token slots (§4) to re-skin. Add the AI layer later by wiring §5 and bundling ai/ai.css. (This repo's app uses both; a no-AI consumer simply drops grain/ai from its style roots.)


1. The substrate contract: what a host must provide

GRAIN is portable to any host that supplies three things:

  1. An OpChannel: push(session, event, data) (ai/contract.ts). How render ops reach a client. BATCH = SSE; a WebSocket hub would do. GRAIN imports the interface from its own contract, never the implementation. (Future, additive: a durable sibling, a per-actor turn-status store, lets a reconnecting client reflect a still-running turn. It sits beside this port; the component conventions don't change. See https://tjakoen.github.io/grain/docs/ai-interface §5d.)
  2. A renderer that understands the binding vocabulary below (§3). BATCH's composition engine implements it; another substrate must too.
  3. A filesystem (ai/accepts.ts reads component files to harvest the manifest): any JS runtime, not substrate-specific.

Everything else (the scoped write capability, "render a surface to HTML") is injected by the composition root: GRAIN names no concrete dependency.


2. GRAIN's own conventions (read by the dispatcher + harvest)

Put these on your components/markup; GRAIN's island + door act on them:

| Attribute | On | Means | |---|---|---| | data-surface="kind:id" | any region | its address: what render ops target (built with surface(kind, id)) | | data-action="verb" | a control | clicking it (or Enter in an input) becomes POST /intent with that verb | | data-target="surface" | a control | the surface the action affects (defaults to the nearest data-surface) | | data-kind + data-accepts="v1 v2" | a component root | harvested into the AI manifest (what verbs this kind accepts) | | data-grade="grain\|smooth\|accent" · .is-ai | any ancestor | provenance: grain = AI, smooth = human (distributed via --type-font) | | data-commit="pending" | any ancestor | in-transit / not-yet-committed → reads grain |

The verbs themselves live in the closed registry (ai/contract.ts: ActionName / SurfaceKind / ACTIONS): the single source of truth.

State invariant (what keeps it from getting fidgety). UI state is declarative attributes components wear: the dispatcher acts on attributes, never component names, so any component participates with no bespoke JS. Transient state (data-commit="pending" / ai-spotlit) follows one rule: every optimistic "pending" is cleared by a server-sent terminal op: a committed op, a flash (rollback), a remove, or a type {done}. Backstops cover the rest: a held control is also released when the AI turn ends (spotlight active:false), re-triggering the same surface clears the old holder first, and a 20s timeout releases a stuck spotlight. Net: the server is the source of truth; the DOM is derived, and never strands a "working" state.


3. The binding vocabulary GRAIN's components are authored in

These are the host-renderer conventions (BATCH defines them in ../batch/render/render.ts). A substrate must interpret them to render GRAIN's b-* atoms. This is the one coupling that lives in markup rather than code.

| Form | Means | |---|---| | data-field="path" | set this element's text from a data path (HTML-escaped); "." = the scope itself | | data-bind-<attr>="path" | set attribute <attr> from a data path; empty → omit; URL attrs are scheme-guarded | | <slot-tag prop-as="default" prop-attr-X="prop" prop-text="prop"> | the polymorphic element: become as; set attrs/text from config props | | each="path" | repeat the component once per array item (item = the scope) | | data="path" | scope a child component to a sub-slice | | <b-button> … (hyphenated tag) | a component: expanded server-side; may self-close |

Example (GRAIN's text atom):

<slot-tag prop-as="span" prop-attr-class="class" data-field=".">Example text</slot-tag>

4. The token slots (GRAIN's default theme: override to re-skin)

GRAIN ships a default theme in styles/variables.css (the Sourdough look: monochrome paper/ink + self-hosted Redaction grades). The mechanism (styles/grain.css) and atoms read these token slots, so a consumer re-skins by overriding the slots in its own sheet (linked after GRAIN's), no component changes:

  • fonts: --type-font (the inherited switch), --font-grain, --font-smooth, --font-accent
  • ink/paper: --ink, --paper, --color-fg, --color-muted, --line-soft
  • scale: --space-1..8, --text-sm, --border, --radius-sm, --radius-md
  • AI spotlight: --ai-veil (the "AI is acting" backdrop, three on-theme picks: color-mix(in srgb, var(--ink) 22%, transparent) dim · color-mix(in srgb, var(--paper) 70%, transparent) wash · transparent lift) and --ai-focus-move (how slowly the spotlight glides between surfaces). Both set once in the theme so every page behaves identically.

5. Wiring it (composition root)

import { createInteractionLayer } from "grain/ai/interaction-layer.ts";
import { createAccepts } from "grain/ai/accepts.ts";

const layer = createInteractionLayer({
  reasoner,                       // your Model-backed reasoner (or the stub)
  stream,                         // an OpChannel (e.g. BATCH's createStream())
  archiveItem,                    // app-scoped write capabilities the reasoner may use…
  renderSurface,                  // …and "render this surface to committed HTML"
});
// mount: POST /intent → layer.handleIntent ; GET /stream → the channel ;
//        GET /ai/manifest → buildManifest(screen, targets) ;
// include the island: <script src="/scripts/ai-dispatch.js" defer></script>
// link GRAIN's stylesheets in <head>: variables.css → global.css → grain.css, then your
//   component bundle; include grain/ai/ai.css (bundled or linked) for the AI spotlight.

Rough edge (honest): the action vocabulary values and the reasoner's tools (archiveItem …) are currently entangled with GRAIN rather than fully app-injected. Pushing the vocabulary down to the consumer is a known next step (see https://tjakoen.github.io/grain/docs/ai-interface §1b note). Until then, edit ai/contract.ts to add verbs.

Optional standalone islands (drop-in, no wiring)

Self-contained <dialog> mechanisms: one <script> plus its stylesheet, driven entirely by data-attributes, no composition-root wiring and no AI layer required:

  • scripts/cmdk.js + styles/cmdk.css: the ⌘K command palette. Fetches /search.json; any [data-cmdk-open] element also opens it.
  • scripts/lightbox.js + styles/lightbox.css: the image viewer. [data-lightbox] marks a trigger (if it's an <a href>, the href is the full image, so it stays a real link with no JS); a [data-lightbox-group] ancestor turns its triggers into one gallery with prev/next + a dot rail. Caption comes from data-lightbox-caption or the child <img> alt.
  • scripts/presentation.js + the presentation organism: a deck. Slides, staged fragments, a fit-to-box pass so a dense slide scales instead of overflowing, an overview grid, present mode, a synced presenter window, and a print sheet. Opt in with data-deck. It owns which slide is up and how far into it you are, and nothing else: an app hangs its own figures off one event, presentation:slide.
  • scripts/drawer.js + the drawer organism: the modal side panel. [data-drawer-open="id"] opens, [data-drawer-close] (the scrim carries it) closes, and the script owns the modal obligations: focus into the panel, Tab trapped, the rest of the page inert, focus returned to the opener. Open/closed is the plain hidden attribute, so a server can ship it either way.

6. Build your own: a theme + components (consumer guide)

GRAIN is meant to be consumed, not forked. Two independent extension points, and they compose: one token override re-skins your components and GRAIN's together. The decider for where something lives is one question: "would another product on GRAIN want this?" Yes → it's reusable design, contribute it to grain/. No → it's yours, keep it in your app. This repo is the worked example: project/ and tjakoen.github.io/ are two consumers doing exactly the below.

A · A theme (re-skin): override token slots, never components

  1. Write a stylesheet that redefines the §4 token slots (--paper, --ink, --font-*, --space-*, --radius-*, --ai-veil, …). Touch no component CSS.
  2. Link it after GRAIN's variables.css in your pages. Every b-* atom re-skins.
/* my-theme.css — linked after grain/styles/variables.css */
:root { --paper:#0f1115; --ink:#e8e6df; --font-grain:"IBM Plex Mono"; --radius-md:12px;
        --ai-veil: color-mix(in srgb, var(--ink) 30%, transparent); }

To make it switchable at runtime (light/dark or multiple flavours), scope the block to an attribute and flip it on <html>: [data-theme="mine"] { … } / [data-color-scheme="dark"] { … }. Wire the two theming scripts once at the composition root: scripts/theme.js (deferred: the declarative controls + localStorage persistence) and scripts/theme-boot.js render-blocking in <head> (pre-sets the saved attributes before first paint; without it every navigation flashes the default theme). Inject them on every page shell (including any page that builds its own <html> outside the page server), or that page silently ignores the saved theme. Grade-as-signal is load-bearing: keep --font-grain visibly distinct from --font-smooth in every theme, and verify it still reads in dark.

B · A component: author in the binding vocabulary

  1. Create your own component-root dir (e.g. myapp/components/). Drop a component: mywidget/mywidget.html + mywidget.css (+ mywidget.md for the catalog). One root class, variants as attributes ([data-variant]), and read token slots, so it inherits any theme + dark mode for free. Author the markup in the §3 vocabulary (data-field / data-bind-* / each / <slot-tag>).
  2. Register the dir at your composition root so the host composes <mywidget> and bundles its CSS: add it to the renderer's component roots + the style bundle + the catalog (this repo: componentRoots / styleRoots in tjakoen.github.io/src/config.ts; a standalone consumer passes its dir to createRenderer / the style bundle / createCatalog). It then appears in /catalog automatically.

C · Make it AI-operable (optional: the payoff)

Add the §2 conventions and the shared dispatcher drives it with no bespoke JS: data-surface="kind:id" (its address), data-kind + data-accepts="verb" (harvested into the manifest), a data-action="verb" trigger, and express AI-mode off data-commit/data-grade. Register the verb in the vocabulary (ai/contract.ts, see the §5 rough-edge note). That's the whole path: a themed, catalogued, AI-operable component without editing GRAIN.


🤖 Built with Claude, on the same buttons you'd click. I called the shots, Claude typed them, and the AI never got a backdoor into the DOM, same door as you, every time. I don't prompt and pray, I prompt and prove. How I actually work with AI, receipts and all →