Update app.py

app.py

@@ -14,7 +14,6 @@ from smolagents import (
     DuckDuckGoSearchTool,
     LiteLLMModel,
     Tool,
-    VisitWebpageTool,
     WikipediaSearchTool,
 )
 
@@ -25,6 +24,7 @@ HTTP_TIMEOUT = 45
 MAX_EXTRACTED_CHARS = 35_000
 DEFAULT_HF_MODEL = "huggingface/openai/gpt-oss-120b"
 DEFAULT_GEMINI_MODEL = "gemini/gemini-3.5-flash"
+TASK_FILE_CACHE = {}
 
 
 def clean_filename(value: str) -> str:
@@ -123,6 +123,31 @@ def download_gaia_attachment(task_id: str) -> tuple[bytes, str]:
     )
 
 
+def get_task_file_name(task_id: str) -> str:
+    """Retorna o nome oficial do anexo ou uma string vazia."""
+    task_id = str(task_id or "").strip()
+    if not task_id:
+        return ""
+    if task_id in TASK_FILE_CACHE:
+        return TASK_FILE_CACHE[task_id]
+
+    try:
+        response = requests.get(
+            f"{DEFAULT_API_URL}/questions", timeout=HTTP_TIMEOUT
+        )
+        response.raise_for_status()
+        for item in response.json():
+            item_task_id = str(item.get("task_id", "")).strip()
+            TASK_FILE_CACHE[item_task_id] = str(
+                item.get("file_name") or ""
+            ).strip()
+    except Exception as exc:
+        print(f"Could not load task attachment metadata: {exc}")
+        return ""
+
+    return TASK_FILE_CACHE.get(task_id, "")
+
+
 def extract_attachment_text(data: bytes, filename: str) -> str:
     """Extrai conteúdo legível dos formatos mais comuns do GAIA."""
     suffix = Path(filename).suffix.lower()
@@ -218,6 +243,74 @@ def extract_attachment_text(data: bytes, filename: str) -> str:
     return text
 
 
+class OpenWebPageTool(Tool):
+    name = "visit_webpage"
+    description = (
+        "Opens and reads one exact HTTP/HTTPS page. Use it after web_search to "
+        "verify article text, tables, archives, papers, and linked sources. "
+        "It automatically retries blocked HTML pages through a text mirror. "
+        "It is not a binary-file downloader."
+    )
+    inputs = {
+        "url": {
+            "type": "string",
+            "description": "The complete HTTP or HTTPS URL to open.",
+        }
+    }
+    output_type = "string"
+
+    def forward(self, url: str) -> str:
+        from markdownify import markdownify
+
+        url = str(url or "").strip()
+        if not re.match(r"^https?://", url, flags=re.I):
+            return "Invalid URL: visit_webpage requires a full HTTP/HTTPS URL."
+
+        headers = {
+            "User-Agent": (
+                "Mozilla/5.0 (compatible; GAIA-Course-Agent/1.0; "
+                "+https://huggingface.co/learn/agents-course)"
+            )
+        }
+        errors = []
+        targets = [url]
+        if "r.jina.ai/http" not in url:
+            without_scheme = re.sub(r"^https?://", "", url)
+            targets.append(f"https://r.jina.ai/http://{without_scheme}")
+
+        for target in targets:
+            try:
+                response = requests.get(
+                    target,
+                    headers=headers,
+                    timeout=HTTP_TIMEOUT,
+                    allow_redirects=True,
+                )
+                response.raise_for_status()
+                content_type = response.headers.get("content-type", "").lower()
+                if (
+                    "text/" not in content_type
+                    and "json" not in content_type
+                    and "xml" not in content_type
+                ):
+                    return (
+                        f"The URL returned binary content ({content_type}). "
+                        "Use inspect_gaia_attachment for official GAIA files."
+                    )
+
+                text = response.text
+                if "html" in content_type:
+                    text = markdownify(text)
+                text = text.strip()
+                if len(text) > 30_000:
+                    text = text[:30_000] + "\n[page truncated]"
+                return text or "The page was retrieved but contained no text."
+            except Exception as exc:
+                errors.append(f"{target}: {exc}")
+
+        return "Error fetching the webpage: " + " | ".join(errors)
+
+
 class InspectGaiaAttachmentTool(Tool):
     name = "inspect_gaia_attachment"
     description = (
@@ -426,12 +519,6 @@ class BasicAgent:
         gemini_api_key = os.getenv("GEMINI_API_KEY")
         configured_model = os.getenv("GAIA_MODEL_ID")
 
-        if not gemini_api_key:
-            raise RuntimeError(
-                "O secret GEMINI_API_KEY é obrigatório porque todas as "
-                "respostas passam pela revisão final do Gemini."
-            )
-
         model_id = configured_model or DEFAULT_HF_MODEL
 
         if model_id.startswith("gemini/"):
@@ -466,12 +553,7 @@ class BasicAgent:
             "snippets. Use it to discover candidate sources. It does NOT open "
             "or read the full pages; call visit_webpage on a returned URL."
         )
-        visit_page_tool = VisitWebpageTool(max_output_length=30_000)
-        visit_page_tool.description = (
-            "Opens one exact HTTP/HTTPS URL and returns the readable page "
-            "content. Use it after web_search when facts depend on the page "
-            "itself, a table, article text, or linked source."
-        )
+        visit_page_tool = OpenWebPageTool()
         wikipedia_tool = WikipediaSearchTool(
             user_agent="GAIA-Course-Agent/1.0 (educational project)",
             language="en",
@@ -490,14 +572,15 @@ class BasicAgent:
             InspectGaiaAttachmentTool(),
             YouTubeTranscriptTool(),
             AnalyzeGaiaImageTool(),
-            ConsultGeminiTool(),
         ]
+        if gemini_api_key:
+            agent_tools.append(ConsultGeminiTool())
 
         self.agent = CodeAgent(
             tools=agent_tools,
             model=self.model,
             max_steps=10,
-            planning_interval=4,
+            planning_interval=None,
             additional_authorized_imports=[
                 "collections",
                 "datetime",
@@ -540,8 +623,18 @@ can be opened. Never invent a tool, use subprocess, or use shell commands.
 Do not repeat nearly identical searches; change the source or method.
 
 FINAL RESPONSE POLICY:
-Call final_answer with only the requested value. Never include reasoning,
-explanations, labels, Markdown, citations, or the words "FINAL ANSWER".
+Call final_answer with only the requested value. The final action must always
+be valid executable code inside the required code tags, for example:
+<code>
+final_answer("Claus")
+</code>
+or:
+<code>
+final_answer(5)
+</code>
+Never write the answer as plain text outside a final_answer tool call. Never
+include reasoning, explanations, labels, Markdown, citations, or the words
+"FINAL ANSWER" inside the submitted value.
 - Quantity/count: return only the number, unless units or currency are requested.
 - Person: return only the requested name component.
 - City/country/code: return only that value.
@@ -560,11 +653,23 @@ explanations, labels, Markdown, citations, or the words "FINAL ANSWER".
         if not question:
             raise ValueError("Digite uma pergunta para testar o agente.")
 
-        task_context = (
-            f"GAIA task_id: {task_id}\n\nQuestion: {question}"
-            if task_id
-            else question
-        )
+        if task_id:
+            attachment_name = get_task_file_name(task_id)
+            attachment_context = (
+                f"Official attachment: {attachment_name}. "
+                "Use the appropriate attachment tool."
+                if attachment_name
+                else (
+                    "Official attachment: NONE. Do not call "
+                    "inspect_gaia_attachment or analyze_gaia_image."
+                )
+            )
+            task_context = (
+                f"GAIA task_id: {task_id}\n"
+                f"{attachment_context}\n\nQuestion: {question}"
+            )
+        else:
+            task_context = question
         try:
             result = self.agent.run(task_context, reset=True)
         except Exception as exc:
@@ -653,6 +758,12 @@ explanations, labels, Markdown, citations, or the words "FINAL ANSWER".
         self, question: str, candidate: str, task_id: str | None = None
     ) -> str:
         """Revisa obrigatoriamente conteúdo e formato antes de salvar a resposta."""
+        if not self.gemini_api_key:
+            fallback = self.deterministic_answer_cleanup(candidate)
+            print("Gemini review status: SKIPPED — GEMINI_API_KEY is missing")
+            print(f"Primary answer preserved: {fallback}")
+            return fallback
+
         reviewer_model = os.getenv(
             "GAIA_GEMINI_REVIEW_MODEL", DEFAULT_GEMINI_MODEL
         )
@@ -748,10 +859,17 @@ Do not return JSON. Do not wrap the fields in Markdown or a code block.
             except Exception as exc:
                 last_error = exc
 
-        raise RuntimeError(
-            "A revisão obrigatória do Gemini falhou; a resposta não foi salva. "
-            f"Detalhe: {last_error}"
-        )
+        fallback = self.deterministic_answer_cleanup(candidate)
+        if not fallback:
+            raise RuntimeError(
+                "A revisão do Gemini falhou e a resposta primária estava vazia. "
+                f"Detalhe: {last_error}"
+            )
+
+        print(f"Candidate answer: {fallback}")
+        print("Gemini review status: FAILED — primary answer preserved")
+        print(f"Gemini review error: {last_error}")
+        return fallback
 
 
 def empty_results() -> pd.DataFrame: