core

작성자: vercel

스키마 정의, 카탈로그 생성, json-render 스펙을 위한 AI 프롬프트 생성. defineSchema를 사용하여 타입이 지정된 스펙과 카탈로그로 스키마를 정의한 후, 컴포넌트 이름을 해당 props 및 설명에 매핑하는 카탈로그를 생성합니다. 선택적 사용자 정의 규칙을 사용하여 카탈로그에서 AI 시스템 프롬프트를 생성하며, createSpecStreamCompiler를 통해 JSONL 패치를 사용한 스펙 스트리밍을 지원합니다. 동적 prop 표현식은 상태 바인딩($state, $bindState), 조건문($cond), 템플릿($template) 등을 가능하게 합니다.

npx skills add https://github.com/vercel-labs/json-render --skill core

@json-render/core

Core package for schema definition, catalog creation, and spec streaming.

Key Concepts

  • Schema: Defines the structure of specs and catalogs (use defineSchema)
  • Catalog: Maps component/action names to their definitions (use defineCatalog)
  • Spec: JSON output from AI that conforms to the schema
  • SpecStream: JSONL streaming format for progressive spec building

Experimental Decision-Model Composition

For decision-model composition, import experimental_composeSpec and experimental_createEvaluator from @json-render/core. These APIs are available since v0.21.0; pin exact versions and keep json-render packages on the same version. Experimental exports and Experimental_ types can change in any release.

  • Run the Gateway evaluator server-side with { model: "typesafe-ai/jev", apiKey: process.env.AI_GATEWAY_API_KEY! }. A plain model identifier is required; Jev is the current example; do not import a provider constructor.
  • Call experimental_composeSpec({ catalog, candidates, prompt, evaluate, initialState, signal }). It is an async generator; stream step.spec snapshots to your existing renderer and inspect complete.stopReason (finish, limit, unavailable). Errors and cancellation throw; retain the last snapshot as partial UI.
  • New trees default to strategy: "batch": one evaluation selects root/membership, then a second arranges the selected elements when needed. The first snapshot contains selected content in catalog order under the root's default/first slot. Resource variants share one exclusive question; repeated counts include the root. Root selection takes precedence over conflicting speculative membership for that recipe/resource. Equal sibling positions retain catalog order. Combined layouts are validated before publication; cycles or excessive depth throw. maxElements caps batched creation (default 32). Limit-truncated selections or a missing required layout call return limit. Use strategy: "sequential" for legacy next/parent adapters or sequential creation. Edits stay sequential.
  • Batched trace steps use choice: "select" | "layout" and an answers record. Count each trace as one evaluation, including its tokens and latency once. Custom evaluators must answer every offered question; names/choices are opaque and include root/select_*, then parent_*/order_* for batches.
  • For follow-up edits, pass the selected version as initialSpec. It is cloned and validated; the evaluator may add, replace, remove non-root subtrees, or move/reorder them. Unchanged IDs, bindings, and state are preserved. Optional elementDescriptions shares identifying descriptions without exposing raw props/state. initialState overrides the seed state. Seeds must be valid trees within the catalog, expression subset, and depth limit. Matching recipes consume usage/resource limits; removals/replacements release them. Replacements/moves use two evaluations (select target, then recipe/destination), each counted against the budget. Treat operation and position keys as opaque.
  • Supply atomic candidates with { id, description, element: { type, props, on?, visible? }, root?, maxUses?, resource? }. Catalog alone is insufficient: the app must supply values and binding recipes. Jev chooses elements and parent slots, never free-form text or code. It never executes actions.
  • Candidates are configured component instances, not page templates. Build them from current app records/operations or bind props to initialState; offer explicit alternatives for chart types, field configurations, and layout variants. The model chooses grouping and order within those options. Name required sections in prompts; structural validity does not imply semantic completeness.
  • V1 supports flat Spec catalogs, named slots, literals, $state, $bindState, and state visibility. No prebuilt children, repeat/watch, computed/template/conditional props, or custom directives. Success/error callbacks must reference allowed actions. Events must be declared in the component catalog.
  • Props and action params are validated against initial state without applying schema transforms/defaults. Supply valid initial values and validate/authorize action calls at runtime. Built-ins without parameter schemas get name validation only.
  • root defaults true, maxUses defaults one, shared resource values make alternatives mutually exclusive. Defaults: 32 evaluations (terminal calls included; no extra finish call for batches), depth eight, 10-second Gateway timeout per call. Supply an overall abort signal.
  • Candidate descriptions, prompt, instructions, topology, and explicit context are sent to the evaluator. Initial state and raw props/binding values are not sent automatically.
  • For custom providers implement Experimental_CompositionEvaluator: accept { state, questions, signal }, return { answers: { [question]: { choice, confidence? } }, usage?: { inputTokens? } }. Only return offered criteria keys.

See packages/core/README.md and /docs/jev for app integration and installation instructions. The web playground is an example consumer, not a dependency of the API.

Defining a Schema

import { defineSchema } from "@json-render/core";

export const schema = defineSchema((s) => ({
  spec: s.object({
    // Define spec structure
  }),
  catalog: s.object({
    components: s.map({
      props: s.zod(),
      description: s.string(),
    }),
  }),
}), {
  promptTemplate: myPromptTemplate, // Optional custom AI prompt
});

Creating a Catalog

import { defineCatalog } from "@json-render/core";
import { schema } from "./schema";
import { z } from "zod";

export const catalog = defineCatalog(schema, {
  components: {
    Button: {
      props: z.object({
        label: z.string(),
        variant: z.enum(["primary", "secondary"]).nullable(),
      }),
      description: "Clickable button component",
    },
  },
});

Generating AI Prompts

const systemPrompt = catalog.prompt(); // Uses schema's promptTemplate
const systemPrompt = catalog.prompt({ customRules: ["Rule 1", "Rule 2"] });

SpecStream Utilities

For streaming AI responses (JSONL patches):

import { createSpecStreamCompiler } from "@json-render/core";

const compiler = createSpecStreamCompiler<MySpec>();

// Process streaming chunks
const { result, newPatches } = compiler.push(chunk);

// Get final result
const finalSpec = compiler.getResult();

Form Values in Action Handlers

Use findFormValue("email", params, state) to read a direct parameter, a dotted parameter key (such as "form.email"), a matching flat state key, or a slash-delimited path (such as "/form/email") in nested state. Parameter values are literal, so emails and URLs containing dots are preserved. For action bindings that read nested state, use { $state: "/form/email" }; the resolver passes that value to the handler. A bare "email" field name does not search nested state.form.email.

Dynamic Prop Expressions

Any prop value can be a dynamic expression resolved at render time:

  • { "$state": "/state/key" } - reads a value from the state model (one-way read)
  • { "$bindState": "/path" } - two-way binding: reads from state and enables write-back. Use on the natural value prop (value, checked, pressed, etc.) of form components.
  • { "$bindItem": "field" } - two-way binding to a repeat item field. Use inside repeat scopes.
  • { "$cond": <condition>, "$then": <value>, "$else": <value> } - evaluates a visibility condition and picks a branch
  • { "$template": "Hello, ${/user/name}!" } - interpolates ${/path} references with state values
  • { "$computed": "fnName", "args": { "key": <expression> } } - calls a registered function with resolved args

$cond uses the same syntax as visibility conditions ($state, eq, neq, not, arrays for AND). $then and $else can themselves be expressions (recursive).

Components do not use a statePath prop for two-way binding. Instead, use { "$bindState": "/path" } on the natural value prop (e.g. value, checked, pressed).

{
  "color": {
    "$cond": { "$state": "/activeTab", "eq": "home" },
    "$then": "#007AFF",
    "$else": "#8E8E93"
  },
  "label": { "$template": "Welcome, ${/user/name}!" },
  "fullName": {
    "$computed": "fullName",
    "args": {
      "first": { "$state": "/form/firstName" },
      "last": { "$state": "/form/lastName" }
    }
  }
}
import { resolvePropValue, resolveElementProps } from "@json-render/core";

const resolved = resolveElementProps(element.props, { stateModel: myState });

State Watchers

Elements can declare a watch field (top-level, sibling of type/props/children) to trigger actions when state values change:

{
  "type": "Select",
  "props": { "value": { "$bindState": "/form/country" }, "options": ["US", "Canada"] },
  "watch": {
    "/form/country": { "action": "loadCities", "params": { "country": { "$state": "/form/country" } } }
  },
  "children": []
}

Watchers only fire on value changes, not on initial render.

Validation

Built-in validation functions: required, email, url, numeric, minLength, maxLength, min, max, pattern, matches, equalTo, lessThan, greaterThan, requiredIf.

Cross-field validation uses $state expressions in args:

import { check } from "@json-render/core";

check.required("Field is required");
check.matches("/form/password", "Passwords must match");
check.lessThan("/form/endDate", "Must be before end date");
check.greaterThan("/form/startDate", "Must be after start date");
check.requiredIf("/form/enableNotifications", "Required when enabled");

User Prompt Builder

Build structured user prompts with optional spec refinement and state context:

import { buildUserPrompt } from "@json-render/core";

// Fresh generation
buildUserPrompt({ prompt: "create a todo app" });

// Refinement with edit modes (default: patch-only)
buildUserPrompt({ prompt: "add a toggle", currentSpec: spec, editModes: ["patch", "merge"] });

// With runtime state
buildUserPrompt({ prompt: "show data", state: { todos: [] } });

Available edit modes: "patch" (RFC 6902 JSON Patch), "merge" (RFC 7396 Merge Patch), "diff" (unified diff).

Spec Validation

Validate spec structure and auto-fix common issues:

import { validateSpec, autoFixSpec } from "@json-render/core";

const { valid, issues } = validateSpec(spec);
// issues include: missing_child, invalid_visible (malformed conditions),
// repeat_without_children, repeat_item_outside_scope, repeat_state_mismatch

const { spec: fixed, fixDetails } = autoFixSpec(spec);
// fixDetails entries are { message, lossy }. Lossless fixes relocate
// misplaced fields; lossy fixes prune dangling children references.
// In a repair loop, withhold lossy fixes until retries are exhausted:
const attempt = autoFixSpec(spec, { lossy: retriesExhausted });

Visibility Conditions

Control element visibility with state-based conditions. VisibilityContext is { stateModel: StateModel }.

import { visibility } from "@json-render/core";

// Syntax
{ "$state": "/path" }                    // truthiness
{ "$state": "/path", "not": true }      // falsy
{ "$state": "/path", "eq": value }      // equality
[ cond1, cond2 ]                         // implicit AND

// Helpers
visibility.when("/path")                 // { $state: "/path" }
visibility.unless("/path")               // { $state: "/path", not: true }
visibility.eq("/path", val)              // { $state: "/path", eq: val }
visibility.and(cond1, cond2)             // { $and: [cond1, cond2] }
visibility.or(cond1, cond2)              // { $or: [cond1, cond2] }
visibility.always                        // true
visibility.never                         // false

Built-in Actions in Schema

Schemas can declare builtInActions -- actions that are always available at runtime and auto-injected into prompts:

const schema = defineSchema(builder, {
  builtInActions: [
    { name: "setState", description: "Update a value in the state model" },
  ],
});

These appear in prompts as [built-in] and don't require handlers in defineRegistry.

StateStore

The StateStore interface allows external state management libraries (Redux, Zustand, XState, etc.) to be plugged into json-render renderers. The createStateStore factory creates a simple in-memory implementation:

import { createStateStore, type StateStore } from "@json-render/core";

const store = createStateStore({ count: 0 });

store.get("/count");         // 0
store.set("/count", 1);      // updates and notifies subscribers
store.update({ "/a": 1, "/b": 2 }); // batch update

store.subscribe(() => {
  console.log(store.getSnapshot()); // { count: 1 }
});

The StateStore interface: get(path), set(path, value), update(updates), getSnapshot(), subscribe(listener).

Key Exports

ExportPurpose
defineSchemaCreate a new schema
defineCatalogCreate a catalog from schema
createStateStoreCreate a framework-agnostic in-memory StateStore
resolvePropValueResolve a single prop expression against data
resolveElementPropsResolve all prop expressions in an element
buildUserPromptBuild user prompts with refinement and state context
buildEditUserPromptBuild user prompt for editing existing specs
buildEditInstructionsGenerate prompt section for available edit modes
isNonEmptySpecCheck if spec has root and at least one element
deepMergeSpecRFC 7396 deep merge (null deletes, arrays replace, objects recurse)
diffToPatchesGenerate RFC 6902 JSON Patch operations from object diff
EditModeType: "patch" | "merge" | "diff"
validateSpecValidate spec structure
autoFixSpecAuto-fix common spec issues; classifies fixes lossy/lossless, { lossy: false } withholds pruning
createSpecStreamCompilerStream JSONL patches into spec
createJsonRenderTransformTransformStream separating text from JSONL in mixed streams
parseSpecStreamLineParse single JSONL line
applySpecStreamPatchApply patch to object
StateStoreInterface for plugging in external state management
ComputedFunctionFunction signature for $computed expressions
checkTypeScript helpers for creating validation checks
BuiltInActionType for built-in action definitions (name + description)
ActionBindingAction binding type (includes preventDefault field)

vercel의 다른 스킬

vercel
vercel
로컬 개발 및 테스트를 위한 Vercel REST API 에뮬레이션입니다. 사용자가 로컬에서 Vercel API 엔드포인트와 상호작용하거나 Vercel 통합을 테스트해야 할 때 사용합니다.
cron-jobs
vercel
Vercel Cron Jobs 구성 및 모범 사례. vercel.json에서 예약된 작업을 추가, 편집 또는 디버깅할 때 사용합니다.
codegen
vercel
json-render을 위한 코드 생성 유틸리티입니다. UI 명세서에서 코드를 생성하거나, 사용자 정의 코드 내보내기를 구축하거나, 명세서를 탐색하거나, props를 직렬화할 때 사용합니다.
enable-shopify-menus
vercel
하드코딩된 내비게이션 및 푸터 메뉴를 Shopify 기반 메뉴로 교체합니다.
next-best-practice
vercel
Next.js 모범 사례 - 파일 규칙, RSC 경계, 데이터 패턴, 비동기 API, 메타데이터, 오류 처리, 라우트 핸들러, 이미지/폰트 최적화,…
benchmark-sandbox
vercel
Vercel Sandbox에서 vercel-plugin eval 시나리오를 로컬 WezTerm 패널 대신 실행합니다. Claude Code와 플러그인이 사전 설치된 임시 마이크로VM을 프로비저닝합니다.
write-guide
vercel
점진적인 예제를 통해 실제 사용 사례를 가르치는 기술 가이드를 제작합니다. 개념은 독자가 필요로 할 때만 소개됩니다.
benchmark-testing
vercel
벤치마크 테스트 프로젝트를 생성하고 실행하여 실제 시나리오에서 vercel-plugin 스킬 인젝션을 테스트합니다. 격리된 디렉토리를 설정하고, 설치하며…