@dbx-tools/model-shared

The pure, browser-safe surface of @dbx-tools/model: the model-class taxonomy, the serving-endpoint descriptor, the model-lookup request / ranked-result contract (zod schemas + inferred types), and the score-driven class classifier.

No node:* imports, no WorkspaceClient, no I/O - safe to import from a client bundle. A frontend uses this to validate a lookup request, type a ranked response, and bucket a /models payload by class; an agent tool can adopt ModelQuerySchema directly as its inputSchema. Live endpoint listing and fuzzy resolution live in @dbx-tools/model.

Install

npm install @dbx-tools/model-shared

Usage

import {
  ModelClass,
  ModelQuerySchema,
  classifyEndpoints,
  type ServingEndpointSummary,
} from "@dbx-tools/model-shared";

// Validate a caller's lookup request (e.g. an agent tool input).
const query = ModelQuerySchema.parse({ search: "claude sonnet", limit: 5 });

// Bucket a /models response into capability bands, best-first per band.
const byClass = classifyEndpoints(endpoints);
const thinking = byClass[ModelClass.ChatThinking];

What's here

• ModelClass - chat-thinking / chat-balanced / chat-fast / embedding, plus its zod schema.
• ServingEndpointSummary - the stable endpoint descriptor (name, task, state, profile scores, classified class, embedding dimension).
• ModelQuery / RankedModel - the lookup request and ranked result schemas.
• classifyEndpoints - buckets endpoints into classes from the Foundation Model API quality/speed/cost scores (relative quantiles, not fixed cut-offs), with a family-name heuristic for unscored chat endpoints.

The offline fallback model list is a server concern and lives in @dbx-tools/model: a browser never talks to Databricks directly, so it consumes the live /models response rather than a baked-in list.