Fix task_id/read_pdf bugs, upgrade model, tighten answer formatting for GAIA scoring

- agent.run() was called with the raw question, never full_prompt, so
the agent never saw task_id and could not call fetch_task_file for
any GAIA question with an attached file.
- read_pdf was instantiated but commented out of the tools list, so
the agent had no way to actually read PDFs despite instructions
telling it to.
- Swap Qwen2.5-Coder-32B-Instruct for Qwen3-235B-A22B-Instruct-2507
(via novita) with temperature=0.3 for stronger general reasoning
alongside code generation.
- Add explicit answer-formatting guidance matching GAIA's quasi-exact
match grader (bare numbers, minimal strings, comma-separated lists).

app.py

@@ -156,22 +156,24 @@ class BasicAgent:
         read_pdf = ReadPDFTool()
         analyze_image = AnalyzeImageTool()
         transcribe_audio = TranscribeAudioTool()
-        #read_pdf = ReadPDFTool()
         self.agent = CodeAgent(
             tools=[
                 DuckDuckGoSearchTool(),
                 VisitWebpageTool(),
                 PythonInterpreterTool(),
                 FinalAnswerTool(),
-                #pdf_tool,
                 fetch_task_file,
-                #read_pdf,
+                read_pdf,
                 read_spreadsheet,
                 get_youtube_transcript,
                 analyze_image,
                 transcribe_audio,
                 ],
-            model= InferenceClientModel("Qwen/Qwen2.5-Coder-32B-Instruct") ,
+            model= InferenceClientModel(
+                "Qwen/Qwen3-235B-A22B-Instruct-2507",
+                provider="novita",
+                temperature=0.3,
+            ),
             additional_authorized_imports=["pandas", "requests", "re"],
             instructions = ("You are an advanced CodeAgent that will show your capabilities to work in the real world by being tested in GAIA, the agent testing platform. If the question includes a task_id and mentions a file, call fetch_task_file first; route YouTube URLs to get_youtube_transcript, other URLs to visit_webpage, PDFs to read_pdf, and spreadsheets to read_spreadsheet, using web_search only when no URL or file is given,then respond with only the exact final answer value, no explanation, no prefix."
                            "Use the Thought Action observation to produce high quality results and only answer when you are sure you have performed the necessary steps for the task and question"
@@ -180,7 +182,12 @@ class BasicAgent:
                            "Use the PythonInterpreterTool for code interpretation in python"
                            "NEVER invent, guess, or simulate data you have not actually retrieved. If a "
                            "file cannot be fetched or a page cannot be read, say so explicitly rather "
-                           "than fabricating plausible-looking data or answers."),
+                           "than fabricating plausible-looking data or answers. "
+                           "The grader does an exact string match after light normalization, so format "
+                           "the final answer exactly as the question asks: a bare number with no commas, "
+                           "units, or currency symbols unless explicitly requested; as few words as "
+                           "possible for a string answer, with no articles or explanatory text; and a "
+                           "comma-separated list (no surrounding brackets) if multiple items are asked for."),
             max_steps=20,
         )
 #answering questions 
@@ -190,8 +197,8 @@ class BasicAgent:
         if task_id:
             full_prompt = f"{question}\n\n(task_id for this question: {task_id})"
         try:
-            # Pass the incoming question to the agent
-            answer = self.agent.run(question)
+            # Pass the full prompt (including task_id) so the agent knows to fetch attached files
+            answer = self.agent.run(full_prompt)
             return str(answer).strip()
         except Exception as e:
                 import traceback