Update app.py

app.py

@@ -16,6 +16,10 @@ from youtube_transcript_api import YouTubeTranscriptApi
 
 
 DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
+ATTACHMENT_MIRROR = (
+    "https://huggingface.co/spaces/bstraehle/gaia/resolve/"
+    "617431ad4021eb2a62ab81488d1ec1cd7d1e3e19/data"
+)
 TEXT_MODELS = (
     "Qwen/Qwen3-235B-A22B-Instruct-2507:cheapest",
     "openai/gpt-oss-120b:cheapest",
@@ -172,6 +176,13 @@ def download_attachment(task_id: str, file_name: str) -> str:
         headers={"User-Agent": USER_AGENT},
         timeout=30,
     )
+    if response.status_code == 404:
+        print(f"Course file endpoint is missing {task_id}; using pinned mirror.")
+        response = requests.get(
+            f"{ATTACHMENT_MIRROR}/{file_name}",
+            headers={"User-Agent": USER_AGENT},
+            timeout=30,
+        )
     response.raise_for_status()
     handle = tempfile.NamedTemporaryFile(delete=False, suffix=suffix)
     handle.write(response.content)
@@ -232,7 +243,7 @@ def extract_attachment(path: str, file_name: str, question: str, token: str) ->
 
 
 def should_search(question: str, has_attachment: bool) -> bool:
-    if has_attachment or "youtube.com" in question or "youtu.be" in question:
+    if has_attachment:
         return False
     reversed_question = question[::-1].lower()
     if "understand this sentence" in reversed_question:
@@ -319,16 +330,37 @@ def run_and_submit_all(
     oauth_token: gr.OAuthToken | None,
 ):
     """Fetch all questions, solve them, and submit exact-match answers."""
-    if profile is None or oauth_token is None:
+    if profile is None:
         return (
-            "Please sign in with Hugging Face and authorize inference access first.",
+            "Please sign in with Hugging Face first. No answers were submitted.",
             None,
         )
 
     username = profile.username
     space_id = os.getenv("SPACE_ID")
     agent_code = f"https://huggingface.co/spaces/{space_id}/tree/main"
-    agent = ResearchAgent(oauth_token.token)
+    provider_token = os.getenv("HF_TOKEN")
+    if not provider_token and oauth_token is not None:
+        provider_token = oauth_token.token
+    if not provider_token:
+        return (
+            "Missing inference credential. Add a private Space secret named HF_TOKEN "
+            "with 'Make calls to Inference Providers' permission. No answers were submitted.",
+            None,
+        )
+
+    agent = ResearchAgent(provider_token)
+    try:
+        access_check = agent.ask_model("Reply with exactly: OK")
+        if access_check.upper() != "OK":
+            raise RuntimeError(f"Unexpected preflight response: {access_check!r}")
+    except Exception as exc:
+        return (
+            "Inference preflight failed, so evaluation was stopped before any submission. "
+            "Create a fine-grained Hugging Face token with 'Make calls to Inference "
+            f"Providers' permission and save it as the HF_TOKEN Space secret. Detail: {exc}",
+            None,
+        )
 
     try:
         response = requests.get(f"{DEFAULT_API_URL}/questions", timeout=20)
@@ -346,9 +378,23 @@ def run_and_submit_all(
         print(f"Solving {index}/{len(questions)}: {task_id}")
         try:
             answer = agent.solve(question, task_id, file_name)
+            if not answer.strip():
+                raise RuntimeError("The model returned an empty answer")
         except Exception as exc:
             print(f"Agent error on {task_id}: {exc}")
-            answer = ""
+            results_log.append(
+                {
+                    "Task ID": task_id,
+                    "Question": question,
+                    "Submitted Answer": f"ERROR (not submitted): {exc}",
+                }
+            )
+            return (
+                f"Stopped safely at question {index}/{len(questions)}. "
+                "No leaderboard submission was made because at least one answer failed. "
+                f"Task {task_id}: {exc}",
+                pd.DataFrame(results_log),
+            )
         answers_payload.append({"task_id": task_id, "submitted_answer": answer})
         results_log.append(
             {
@@ -358,6 +404,14 @@ def run_and_submit_all(
             }
         )
 
+    if len(answers_payload) != len(questions) or any(
+        not item["submitted_answer"].strip() for item in answers_payload
+    ):
+        return (
+            "Safety check stopped submission because the answer set was incomplete.",
+            pd.DataFrame(results_log),
+        )
+
     submission = {
         "username": username.strip(),
         "agent_code": agent_code,
@@ -390,9 +444,10 @@ with gr.Blocks() as demo:
 This agent searches the web, reads benchmark attachments, transcribes audio,
 and uses a large reasoning model to produce exact-match answers.
 
-1. Sign in with Hugging Face and authorize inference access.
-2. Click **Run Evaluation & Submit All Answers** once.
-3. Keep this tab open while the 20 questions are processed.
+1. Add a private Space secret named `HF_TOKEN` with Inference Providers permission.
+2. Sign in with Hugging Face.
+3. Click **Run Evaluation & Submit All Answers** once.
+4. Keep this tab open while the 20 questions are processed.
 """
     )
     gr.LoginButton()