Mount API routes on Gradio's app instead of wrapping — fixes HF Spaces

app.py

@@ -518,32 +518,34 @@ def ask_question(question: str) -> dict:
     }
 
 
-if __name__ == "__main__":
-    import fastapi
-    from fastapi.middleware.cors import CORSMiddleware
-
-    WORKSPACE_DIR = load_workspace()
-    demo = build_app()
-
-    # Mount the JSON API on the same FastAPI app
-    fastapi_app = fastapi.FastAPI()
-    fastapi_app.add_middleware(
-        CORSMiddleware,
-        allow_origins=["*"],
-        allow_methods=["POST"],
-        allow_headers=["*"],
-    )
+WORKSPACE_DIR = load_workspace()
+demo = build_app()
+
+# Add API routes to Gradio's underlying FastAPI app
+import fastapi as _fastapi
+from fastapi.middleware.cors import CORSMiddleware
+
+demo.app.add_middleware(
+    CORSMiddleware,
+    allow_origins=["*"],
+    allow_methods=["*"],
+    allow_headers=["*"],
+)
+
 
-    @fastapi_app.post("/api/ask")
-    async def api_ask(request: fastapi.Request):
-        body = await request.json()
-        question = body.get("question", "")
-        if not question:
-            return {"error": "No question provided"}
-        return ask_question(question)
+@demo.app.post("/api/ask")
+async def api_ask(request: _fastapi.Request):
+    body = await request.json()
+    question = body.get("question", "")
+    if not question:
+        return {"error": "No question provided"}
+    return ask_question(question)
 
-    @fastapi_app.get("/api/remaining")
-    async def api_remaining():
-        return {"remaining": _remaining(), "limit": DAILY_LIMIT}
 
-    demo.launch(ssr_mode=False, app=fastapi_app)
+@demo.app.get("/api/remaining")
+async def api_remaining():
+    return {"remaining": _remaining(), "limit": DAILY_LIMIT}
+
+
+if __name__ == "__main__":
+    demo.launch(ssr_mode=False)