Fix: patch App.__init__ to inject routes into the actual served app

app.py

@@ -27,8 +27,6 @@ if hf_token:
     login(token=hf_token)
 
 
-# ZeroGPU Spaces require at least one @spaces.GPU function to be defined
-# during startup.  This minimal touch satisfies that requirement.
 @spaces.GPU(duration=1)
 def _gpu_startup_touch():
     return "ok"
@@ -145,7 +143,7 @@ API_RESOURCES = {
             "extract_top": "integer (0-5, default 0) - fetch full page content for top N results",
             "region": "string (default 'wt-wt') - e.g. 'us-en', 'gb-en', 'de-de'",
         },
-        "output": "JSON with title, url, content per result. News results include date and source. Extracted results include extracted_content.",
+        "output": "JSON with title, url, content per result.",
     },
 }
 
@@ -206,28 +204,13 @@ try:
                 hits = list(ddgs.news(query, max_results=max_results, region=region))
                 results = []
                 for r in hits:
-                    entry = {
-                        "title": r.get("title", ""),
-                        "url": r.get("url", ""),
-                        "content": r.get("body", ""),
-                        "source": r.get("source", ""),
-                        "date": r.get("date", ""),
-                    }
+                    entry = {"title": r.get("title", ""), "url": r.get("url", ""), "content": r.get("body", ""), "source": r.get("source", ""), "date": r.get("date", "")}
                     if r.get("image"):
                         entry["image"] = r["image"]
                     results.append(entry)
             elif backend == "images":
                 hits = list(ddgs.images(query, max_results=max_results, region=region))
-                results = []
-                for r in hits:
-                    results.append({
-                        "title": r.get("title", ""),
-                        "url": r.get("image", ""),
-                        "source": r.get("source", ""),
-                        "thumbnail": r.get("thumbnail", ""),
-                        "width": r.get("width"),
-                        "height": r.get("height"),
-                    })
+                results = [{"title": r.get("title", ""), "url": r.get("image", ""), "source": r.get("source", ""), "thumbnail": r.get("thumbnail", ""), "width": r.get("width"), "height": r.get("height")} for r in hits]
             else:
                 hits = list(ddgs.text(query, max_results=max_results, region=region))
                 results = [{"title": r.get("title", ""), "url": r.get("href", ""), "content": r.get("body", "")} for r in hits]
@@ -282,23 +265,14 @@ def _ensure_llama_cpp():
     src_dir = os.path.join(_BENCHMARK_CACHE, "llama_src")
     if os.path.isdir(src_dir):
         shutil.rmtree(src_dir, ignore_errors=True)
-    subprocess.check_call(
-        ["git", "clone", "--depth=1", "https://github.com/ggml-org/llama.cpp.git", src_dir],
-        stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL,
-    )
+    subprocess.check_call(["git", "clone", "--depth=1", "https://github.com/ggml-org/llama.cpp.git", src_dir], stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
     _log("Benchmark: building llama.cpp (this takes 2-5 minutes)...")
     n_threads = _parse_cpu_count() or 8
     build_dir = os.path.join(_BENCHMARK_CACHE, "llama_build")
     os.makedirs(build_dir, exist_ok=True)
     os.makedirs(bin_dir, exist_ok=True)
-    subprocess.check_call(
-        ["cmake", "-S", src_dir, "-B", build_dir, "-DLLAMA_NATIVE=ON", "-DBUILD_SHARED_LIBS=OFF"],
-        stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL,
-    )
-    subprocess.check_call(
-        ["cmake", "--build", build_dir, "-j", str(n_threads), "--target", "llama-cli"],
-        stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL,
-    )
+    subprocess.check_call(["cmake", "-S", src_dir, "-B", build_dir, "-DLLAMA_NATIVE=ON", "-DBUILD_SHARED_LIBS=OFF"], stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
+    subprocess.check_call(["cmake", "--build", build_dir, "-j", str(n_threads), "--target", "llama-cli"], stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
     import glob as _glob
     candidates = _glob.glob(os.path.join(build_dir, "**", "llama-cli"), recursive=True)
     if candidates:
@@ -427,78 +401,84 @@ with gr.Blocks(title="Respite API") as demo:
 
 
 # ---------------------------------------------------------------------------
-# Inject routes via monkey-patch on Gradio's App.create_app.
-#
-# Gradio's Blocks.launch() calls App.create_app() to build a NEW FastAPI app,
-# discarding any routes added to demo.app earlier.  We wrap that method so our
-# routes get added to the app that is actually served.
+# Route handlers
 # ---------------------------------------------------------------------------
 
-from gradio.routes import App as GradioApp
 
-_orig_create_app = GradioApp.create_app
+def _resources_handler():
+    return JSONResponse({"resources": API_RESOURCES})
 
 
-def _patched_create_app(blocks, **kwargs):
-    fa_app = _orig_create_app(blocks, **kwargs)
+def _specs_handler():
+    return JSONResponse({**API_SPECS, "runtime": _get_runtime_specs()})
 
-    @fa_app.get("/respite/resources")
-    def _resources():
-        return JSONResponse({"resources": API_RESOURCES})
 
-    @fa_app.get("/respite/specs")
-    def _specs():
-        return JSONResponse({**API_SPECS, "runtime": _get_runtime_specs()})
+def _search_handler(q: str = "", backend: str = "text", max_results: int = 10, extract_top: int = 0, region: str = "wt-wt"):
+    if not q:
+        return JSONResponse({"error": "Missing query parameter 'q'"}, status_code=400)
+    if backend not in ("text", "news", "images"):
+        return JSONResponse({"error": f"Invalid backend '{backend}'. Use 'text', 'news', or 'images'."}, status_code=400)
+    try:
+        results = _search_web(q, backend=backend, max_results=max_results, extract_top=extract_top, region=region)
+        return JSONResponse(results)
+    except Exception as e:
+        return JSONResponse({"error": str(e)}, status_code=502)
 
-    @fa_app.get("/respite/search")
-    def _search(q: str = "", backend: str = "text", max_results: int = 10, extract_top: int = 0, region: str = "wt-wt"):
-        if not q:
-            return JSONResponse({"error": "Missing query parameter 'q'"}, status_code=400)
-        if backend not in ("text", "news", "images"):
-            return JSONResponse({"error": f"Invalid backend '{backend}'. Use 'text', 'news', or 'images'."}, status_code=400)
-        try:
-            results = _search_web(q, backend=backend, max_results=max_results, extract_top=extract_top, region=region)
-            return JSONResponse(results)
-        except Exception as e:
-            return JSONResponse({"error": str(e)}, status_code=502)
 
-    @fa_app.get("/respite/benchmark")
-    def _benchmark(force: bool = False, model: str = ""):
-        try:
-            global _BENCHMARK_RESULTS
-            if force:
-                _BENCHMARK_RESULTS = []
-            model_filter = model.strip() if model.strip() else None
-            results = _run_benchmark(model_filter=model_filter)
-            return JSONResponse({
-                "status": "complete",
-                "models": results,
-                "system": {"cpu_cores": _parse_cpu_count(), "ram_bytes": _parse_ram_bytes(), "threads": _parse_cpu_count() or 8},
-            })
-        except Exception as e:
-            return JSONResponse({"error": str(e)}, status_code=500)
+def _benchmark_handler(force: bool = False, model: str = ""):
+    try:
+        global _BENCHMARK_RESULTS
+        if force:
+            _BENCHMARK_RESULTS = []
+        model_filter = model.strip() if model.strip() else None
+        results = _run_benchmark(model_filter=model_filter)
+        return JSONResponse({
+            "status": "complete",
+            "models": results,
+            "system": {"cpu_cores": _parse_cpu_count(), "ram_bytes": _parse_ram_bytes(), "threads": _parse_cpu_count() or 8},
+        })
+    except Exception as e:
+        return JSONResponse({"error": str(e)}, status_code=500)
+
+
+def _benchmark_test_handler():
+    import subprocess as _sp
+    try:
+        llama_cli = _ensure_llama_cpp()
+        r = _sp.run([llama_cli, "--version"], capture_output=True, text=True, timeout=10)
+        binary_ok = r.returncode == 0
+        binary_info = r.stdout.strip()[:200] if binary_ok else r.stderr.strip()[:200]
+        from huggingface_hub import hf_hub_download
+        model_path = hf_hub_download(repo_id="unsloth/gemma-4-E4B-it-GGUF", filename="gemma-4-E4B-it-Q4_K_M.gguf", cache_dir=_BENCHMARK_CACHE)
+        model_ok = os.path.isfile(model_path)
+        model_size = os.path.getsize(model_path) if model_ok else 0
+        return JSONResponse({"binary_path": llama_cli, "binary_exists": os.path.isfile(llama_cli), "binary_ok": binary_ok, "binary_info": binary_info, "model_path": model_path, "model_ok": model_ok, "model_size_gb": round(model_size / 1024**3, 2)})
+    except Exception as e:
+        return JSONResponse({"error": str(e)}, status_code=500)
 
-    @fa_app.get("/respite/benchmark/test")
-    def _benchmark_test():
-        import subprocess as _sp
-        try:
-            llama_cli = _ensure_llama_cpp()
-            r = _sp.run([llama_cli, "--version"], capture_output=True, text=True, timeout=10)
-            binary_ok = r.returncode == 0
-            binary_info = r.stdout.strip()[:200] if binary_ok else r.stderr.strip()[:200]
-            from huggingface_hub import hf_hub_download
-            model_path = hf_hub_download(repo_id="unsloth/gemma-4-E4B-it-GGUF", filename="gemma-4-E4B-it-Q4_K_M.gguf", cache_dir=_BENCHMARK_CACHE)
-            model_ok = os.path.isfile(model_path)
-            model_size = os.path.getsize(model_path) if model_ok else 0
-            return JSONResponse({"binary_path": llama_cli, "binary_exists": os.path.isfile(llama_cli), "binary_ok": binary_ok, "binary_info": binary_info, "model_path": model_path, "model_ok": model_ok, "model_size_gb": round(model_size / 1024**3, 2)})
-        except Exception as e:
-            return JSONResponse({"error": str(e)}, status_code=500)
 
-    _log("Respite API routes injected into Gradio app")
-    return fa_app
+# ---------------------------------------------------------------------------
+# Monkey-patch: intercept Gradio server creation to inject routes.
+# ---------------------------------------------------------------------------
+
+_orig_routes = gr.routes.App.__init__
+
+
+def _patched_init(self, *args, **kwargs):
+    _orig_routes(self, *args, **kwargs)
+    _log("Gradio server __init__ patched - adding routes")
+    try:
+        self.add_api_route("/respite/resources", _resources_handler, methods=["GET"])
+        self.add_api_route("/respite/specs", _specs_handler, methods=["GET"])
+        self.add_api_route("/respite/search", _search_handler, methods=["GET"])
+        self.add_api_route("/respite/benchmark", _benchmark_handler, methods=["GET"])
+        self.add_api_route("/respite/benchmark/test", _benchmark_test_handler, methods=["GET"])
+        _log("Routes added via __init__ patch")
+    except Exception as e:
+        _log(f"Route patching failed: {e}")
 
 
-GradioApp.create_app = _patched_create_app
+gr.routes.App.__init__ = _patched_init
 
 # Disable the Gradio Spaces file watcher thread which crashes because
 # importing `spaces` in a new thread after CUDA init fails.