Handle bundled ReadableStream responses in ModelResult

[w0nche0l]

Summary

• the agent package was failing when deployed to a production OpenNext Cloudflare Worker deployment (even though it worked fine in local workerd workflow)
• Removes the brittle constructor.name === "EventStream" check.
• Adds a direct value instanceof ReadableStream inheritance check.

Verified Repro

I verified the naming check with a fresh minimal OpenNext Cloudflare Worker project using:

• next@16.2.4
• @opennextjs/cloudflare@1.19.5
• wrangler@4.87.0
• @openrouter/agent@0.4.0

Minimal route:

import { OpenRouter, stepCountIs } from "@openrouter/agent";

export const runtime = "nodejs";

export async function GET() {
  const openrouter = new OpenRouter({
    apiKey: process.env.OPENROUTER_API_KEY,
  });

  const result = openrouter.callModel({
    model: process.env.OPENROUTER_MODEL || "google/gemini-3-flash-preview",
    input: "Reply with exactly: ok",
    instructions: "Return a very short answer.",
    stopWhen: stepCountIs(1),
  });

  const [text, response] = await Promise.all([
    result.getText(),
    result.getResponse(),
  ]);

  return Response.json({
    text,
    object: response.object,
  });
}

Deploying that project to Cloudflare Workers with real OPENROUTER_API_KEY and OPENROUTER_MODEL secrets returned:

HTTP/2 500

wrangler tail showed:

Error: Unexpected response type from API

Expected result:

{
  "text": "ok",
  "object": "response"
}

With this diff as a local patch, it succeeded.

[robert-j-y]

Thanks for the fix. I’d recommend closing this one in favor of #34, since the two PRs are overlapping fixes to the same stream-detection bug rather than independent changes.

This PR adds ReadableStream detection but keeps the old toReadableStream() fallback. #34 checks for getReader(), which better matches the actual runtime contract used downstream by ReusableReadableStream, and it also updates the non-streaming response guard consistently.

If there’s anything from this PR you want to preserve, I’d fold it into #34; otherwise #34 should be the base to move forward with after it adds the patch changeset and regression tests.

[w0nche0l]

sounds good! primarily submitted both because i couldn't tell if the getReader check was intentional. We probably do want to remove the name check either way in #34 though. closing this out!