CoreSAFE

The Generative UI framework

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: core
description: Core package for defining schemas, catalogs, and AI prompt generation for json-render. Use when working with @json-render/core, defining schemas, creating catalogs, or building JSON specs for UI/video generation.
---

# @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 unreleased; use a source build until published, then pin exact versions. 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 source-build instructions. The web playground is an example consumer, not a dependency of the API.

## Defining a Schema

```typescript
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

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

export const catalog = defineCatalog(schema, {
  componen

SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.

What the source does

Filesystem

none-observed

Network

none-observed

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (0)

No findings outside the package's declared scope.

Gates applied: no_behavioural_pass.

[![ModelsAtlas audit: Core](https://modelsatlas.com/skills/vercel-labs/core/badge.svg)](https://modelsatlas.com/skills/vercel-labs/core/)

Not your listing, or something out of date? Claim or update your listing.