Corrections from your UI

The best corrections come from the people already reviewing the agent’s work, inside the product they already use. <CorrectButton> is a React component that posts to a route on your server; your server calls Overturn with the project key, so the key never reaches the browser and the actor comes from your own session.

Server route

// app/api/overturn/correct/route.ts (Next.js; any Request/Response runtime works)
import { Overturn } from "@overturn-ai/sdk";
const ov = new Overturn();
export const POST = ov.correctionHandler({
  actor: async (req) => (await getSession(req))?.email ?? null,
});

Return a Response from actor to refuse (for example a 401). The handler validates the body, calls ov.correct() and returns the correction and the new case.

Python (FastAPI / Flask):

@app.post("/api/overturn/correct")
def correct(body: dict, user=Depends(current_user)):
    return ov.handle_correction(body, actor=user.email)   # raises CorrectionRequestError -> 400

Component

import { CorrectButton } from "@overturn-ai/sdk/react";

<CorrectButton
  endpoint="/api/overturn/correct"
  runId={run.id}            // the run your agent recorded (ov.startRun(...).id)
  original={run.output}     // prefilled when the reviewer picks Edit
  onDone={refresh}
/>

It opens Approve / Edit / Reject / Escalate with a note, posts, and shows “Correction recorded”. Props: stepId to correct one model step instead of the final answer, kinds to narrow the choices, headers for a CSRF token, className/style to restyle. React ≥ 18, no other dependencies.