OpenAI Agents SDK imageGenerationTool action edit and Sume

In openai-agents-js v0.18.0, imageGenerationTool() accepts action: 'generate' | 'edit' | 'auto', and leaving it unset keeps the provider default. That option belongs to the hosted tool. If you want the edit done by Sume, give the agent a function tool that calls Sume's image route with prompt and image_urls.

Vendor facts are from the openai-agents-js release notes; Sume facts are from Image generation and edit, read 2026-09-30.

What changed in the Agents SDK?

The v0.18.0 notes (2026-09-10) say the option is forwarded through streaming and non-streaming Responses requests. The Python release v0.22.2 (2026-09-09) is listed as "support current image generation tool options". The notes say nothing more about how the hosted tool receives the images to edit, so check the SDK reference before relying on a particular input shape.

What does Sume's image edit take?

Sume uses the same image route for generation and edit. The shape of the request decides which one you get.

How do I expose the edit as a function tool?

Wrap the route in a function tool. The docs example sends an Idempotency-Key on the create call and says to reuse it only for the same operation and payload when retrying, so let the caller pass a stable key.

import { tool } from "@openai/agents";
import { z } from "zod";

export const editImage = tool({
  name: "edit_image",
  description: "Edit images at public HTTPS URLs with Sume.",
  parameters: z.object({
    prompt: z.string(),
    image_urls: z.array(z.string().url()).min(1).max(10),
    key: z.string(),
  }),
  async execute({ prompt, image_urls, key }) {
    const res = await fetch("https://api.sume.com/v1/image-1.0/generate", {
      method: "POST",
      headers: {
        Authorization: "Bearer " + process.env.SUME_API_KEY,
        "Content-Type": "application/json",
        "Idempotency-Key": key,
      },
      body: JSON.stringify({ prompt, image_urls }),
    });
    return await res.text();
  },
});

What comes back, and what should I store?

The call returns a job, not a finished image. Poll the job and read its result, as in Jobs and results. Sume mirrors outputs into Sume-owned media URLs, and the docs say integrations should store the Sume URL, not raw provider URLs. Sources must be public HTTPS URLs; localhost, private-network and non-HTTPS URLs are rejected before submission.

Edits cost money, so consider an approval step; see human-in-the-loop for paid tools. For video from the same SDK, see MCP video generation.