Provider portability

Schema Stream is provider-agnostic at its input boundary: it consumes JSON text or UTF-8 bytes. A provider integration is portable only when the surrounding SDK reliably exposes the structured response as raw JSON text and separately validates an authoritative final result.

The Vercel AI SDK is the repository's portability boundary. Keep the parser function generic over AI SDK's LanguageModel and configure the provider elsewhere:

import { type LanguageModel, Output, streamText } from "ai"
import { SchemaStream } from "schema-stream"
import { z } from "zod"

const schema = z.object({
  answer: z.string(),
  confidence: z.number(),
  sources: z.array(z.string())
})

async function runStructuredModel({
  model,
  prompt
}: {
  model: LanguageModel
  prompt: string
}): Promise<z.output<typeof schema>> {
  const result = streamText({ model, output: Output.object({ schema }), prompt })
  const parser = new SchemaStream(schema)

  for await (const snapshot of parser.iterate(result.textStream)) {
    renderProgress(snapshot)
  }

  return await result.output
}

The same function can receive an AI SDK model supplied by a gateway or a separately configured AI SDK provider package. Credential loading, model identifiers, and provider options stay outside Schema Stream.

What is tested

The repository does not currently install, execute, or promise direct Anthropic SDK or Google GenAI SDK stream adapters. Do not pass provider-specific SSE events to Schema Stream unless an adapter first extracts only the structured JSON text and has its own fixture and live canary.

Qualify a model before shipping

1. Run the credential-free AI SDK fixtures with the exact package versions used by the application.
2. Add a provider-shaped deterministic fixture that splits JSON across strings, escapes, UTF-8 boundaries, nested arrays, and completion boundaries.
3. Run the opt-in live Gateway lane with the exact model id and several structurally different schemas.
4. Assert meaningful intermediate snapshots, child-before-root completion ordering, and equality between the final snapshot and result.output.
5. Bound output, retries, and timeouts; sanitize failures so prompts, responses, headers, and keys do not enter logs.
6. Keep one scheduled canary for every provider/model combination the application depends on.

Provider behavior and structured-output support can change independently of Schema Stream. An AI SDK type check proves the interface compiles; it does not prove a particular model emits useful or correct incremental JSON. See Integration testing for the existing fixtures and live-test contract.