Jev Labs

Open /lab to explore 169 local examples: 110 catalog examples, 56 workspace scenarios, and the original Composer, Decision, and Settings patterns. Use the sidebar search, the mobile Browse Labs panel, or the global command palette. Every example has a stable /lab#example-id link.

Use a Lab inside this workspace:

import { TypedPreview } from "@/components/labs/typed-preview"
import { labById } from "@/lib/lab-registry"

export function Example() {
  return <TypedPreview entry={labById("social-feature-question")!} />
}

These are app imports. @workspace/ui is a private workspace package, and neither the Labs nor the custom Jev primitives have a public npm or shadcn registry installation.

Inspect and reuse an example

1. Select a Lab and edit its text or JSON input.
2. Choose illustrative output values, including Unknown / unscored.
3. Run the local preview. Inspect the typed output and the request shape.
4. Open Source to inspect the actual renderer, helper, shared components, or example data. Switching tabs preserves the preview state.
5. Open Install for workspace setup and Import for the selected Lab's import example.

A/B examples preserve the upstream mutation path and value. Compare A/B changes only that field. Both outputs use your fixture settings; they are not two model predictions. Removing the comparison field produces an error instead of inventing a replacement.

Editing input or fixture values invalidates the previous result. Failure simulation clears results and produces a visible error. Reference notes apply to the original input and are teaching aids, not execution evidence.

Reusable Jev components

import { JevContract } from "@workspace/ui/components/jev-contract"
import { JevResult } from "@workspace/ui/components/jev-result"

JevContract displays choice, noul, and score questions, instructions, and criteria. JevResult displays typed values, unknowns, empty/error states, and explicit response provenance. Their source, previews, and import examples are also available in the main component library.

Keep provider credentials and request handling on your application's server. These components do not create a provider connection or authorize downstream actions.

Coverage and provenance

The input snapshot comes from TypeSafeAI/typesafe-playground at 9d3d990. It includes all 110 entries in web/catalog.json and covers all 19 workspace routes at that revision.

Upstream workspace	Local Labs
Example builder	All 110 original IDs, states, question sets, 41 A/B variants, and available reference notes
Workflow chat	Six starting cases across five playbooks, with rules and follow-up evidence
Tool router and LangChain	Read settings, change production, and request a secret in each workspace
SMT solver	All five named constraint cases and their reference outcomes
Clean-room rebuild	Searchable catalog, contacts CRUD, and support ticket interactions
Meme lab	All four named text/context examples
Jev Chat	All 12 starter prompts across Story studio, Meet Jev, Support desk, and Your notes
Conversation lab	Default transcript and both legacy follow-up/chatter samples
Ask gate	Original context/docs, the single-question example, and all six incoming sample questions
Document extraction	Original synthetic invoice with exact-source candidate inspection
PR review	Original authentication-bypass diff and repository rules
AST governance	Original organization-membership diff and caller/test manifest
Vector reranker	All 200 generated sample candidates with their original vector scores
YouTube extract	A synthetic caption fixture with original-text evidence and relevance controls
Browser agent	Both PC-build and flight task presets with mock policy checkpoints
Chess, MicroDuck, JevDoom	Bounded action-selection components and one-observation illustrations for each game

The canonical catalog content is preserved. Workspace Labs adapt the upstream interaction patterns into reusable component previews; they do not duplicate every upstream engine, benchmark, backend, or integration. Z3 does not run here. No browser automation, OCR, LangChain adapter, provider call, real game engine, purchase, review posting, or business action runs from these previews.

The snapshot and upstream MIT notice live in apps/web/lib/playground/. Credit remains with TypeSafe AI Playground contributors and @nickthompson480's original playground.

Refresh the snapshot

The sync script reads immutable Git objects without importing or executing code from the sibling repository. It does not modify that checkout or consume its uncommitted edits.

node scripts/sync-playground.mjs ../typesafe-playground

To adopt a newer upstream revision, update the pinned revision deliberately, inspect the changed catalog and scenario sets, and add any missing workspace coverage before running verification. provenance.json records every canonical ID and workspace route.

Verify changes

pnpm lint
pnpm typecheck
pnpm build
pnpm test:e2e

For a direct web-workspace run, use pnpm --filter web test:e2e --workers=3. playground-coverage.spec.ts compares the registry to the pinned inventory and exercises every catalog example, A/B variant, and workspace scenario. jev-labs.spec.ts checks navigation, source/import state, error handling, local-only execution, policy gates, ranking, action freshness, and rebuild interactions. Mobile coverage includes the searchable Lab drawer and Preview/Source/Import overflow checks.

Inspect light/dark themes, RTL, narrow screens, and keyboard focus as well. Chromium automation does not establish human screen-reader acceptance or coverage of every browser engine.