ui/src/gradioApi.js

// Gradio Server-mode client — used ONLY on the Hugging Face ZeroGPU Space.
//
// ZeroGPU attributes GPU quota via the HF iframe auth headers, which a plain `fetch`
// does NOT forward — so the GPU narration endpoints (overview/advice/chat/project_*)
// must be called through @gradio/client, which forwards them. The deterministic
// engine endpoints stay on plain `fetch` (see useAnalysis.js).
//
// Transport is chosen at runtime by `apiMode()`: "local" (the `./her` REST server,
// the default) or "space" (this gradio client). useApi() sets it from /api/health.
import { Client } from "@gradio/client";
import { clientId } from "./client.js";

let _mode = "local"; // "local" = REST /api/* · "space" = gradio @app.api endpoints
export function setApiMode(m) { _mode = m === "space" ? "space" : "local"; }
export function apiMode() { return _mode; }

let _clientP = null;
function getClient() {
  if (!_clientP) _clientP = Client.connect(window.location.origin);
  return _clientP;
}

// Call a Gradio @app.api endpoint by name. A single-output endpoint comes back as
// res.data === [value]; unwrap to the value (our endpoints each return one object).
// The per-browser `client` token is added to every payload so the GPU routes scope to
// the caller's namespace (same isolation as the REST X-Her-Client header).
async function call(name, payload) {
  const gc = await getClient();
  const res = await gc.predict("/" + name, { ...(payload || {}), client: clientId() });
  const d = res && res.data;
  return Array.isArray(d) ? d[0] : d;
}

export const callOverview = (path) => call("overview", { path: path || "" });
export const callAdvice = (path) => call("advice", { path: path || "" });
export const callChat = (question, path) => call("chat", { question, path: path || "" });
export const callProjectChat = (question, cwd) => call("project_chat", { question, cwd });
export const callProjectNarrative = (cwd) => call("project_narrative", { cwd });