OpenAI Agents SDK

The OpenAI Agents SDK exposes model text as an async stream through toTextStream(). Pass that stream to SchemaStream.iterate() on the server, render the progressive snapshots, then treat the Agents SDK result as authoritative after the run completes.

import { Agent, run } from "@openai/agents"
import { SchemaStream } from "schema-stream"
import { z } from "zod"

const analysisSchema = z.object({
  summary: z.string(),
  triage: z.object({
    requiresApproval: z.boolean(),
    rationale: z.string()
  }),
  tags: z.array(z.string())
})

const agent = new Agent({
  name: "Analyst",
  model: "gpt-5.6-luna",
  instructions: "Return a structured analysis.",
  outputType: analysisSchema
})

const result = await run(agent, "Assess the release plan.", { stream: true })
const parser = new SchemaStream(analysisSchema, {
  onValueComplete({ path, value }) {
    const isApproval =
      path.length === 2 && path[0] === "triage" && path[1] === "requiresApproval"

    if (isApproval && value === true) {
      notifyReleaseOwner()
    }
  }
})

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

await result.completed
const finalOutput = analysisSchema.parse(result.finalOutput)

onValueComplete runs after the boolean is syntactically complete, so the branch does not inspect a partial value. Keep callback work short and reversible; final model validation can still fail. See Completion events for ordering and container-value semantics.

Server boundary

• Keep OPENAI_API_KEY, the Agents SDK, and Schema Stream in a trusted server process.
• Forward complete snapshots, patches, or completion events to the browser. Do not forward the key or make the browser understand provider event envelopes.
• Wire the request's AbortSignal through the Agents SDK and stop producing snapshots when no authorized consumer remains.
• Use the SDK's finalOutput only after completed resolves, and validate it before durable writes.

Verify the integration

The repository's examples/sdk-mocks.ts runs the real Agents SDK Runner against a deterministic model implementation. It verifies progressive snapshots, a nested completion-driven decision, and final-result equality without credentials:

bun run example:sdk
bun test tests/sdk-runtime.test.ts

The opt-in live lane requires an already injected key and an explicit model:

SCHEMA_STREAM_LIVE_E2E=1 \
SCHEMA_STREAM_LIVE_PROVIDER=agents \
SCHEMA_STREAM_AGENTS_MODEL=gpt-5.6-luna \
bun run test:live

The live suite uses bounded output, timeouts, disabled tracing, varied nested schemas, and sanitized failures. Never place a key in a command, source file, client bundle, or documentation form.