Update agent.py

agent.py

@@ -121,59 +121,7 @@ def reverse_text(text: str) -> str:
     return text[::-1]
 
 
-@tool
-def youtube_transcript(url: str) -> str:
-    """
-    Fetch the transcript of a YouTube video.
-    ONLY use this when the question explicitly contains a YouTube URL
-    (youtube.com/watch?v= or youtu.be/). Do NOT call this for any other question type.
-    Returns the full transcript text to answer questions about video content.
-    """
-    try:
-        from youtube_transcript_api import YouTubeTranscriptApi
-        from youtube_transcript_api._errors import TranscriptsDisabled, NoTranscriptFound
-
-        # Extract video ID from various YouTube URL formats
-        match = re.search(
-            r'(?:v=|youtu\.be/|embed/|shorts/)([A-Za-z0-9_-]{11})',
-            url
-        )
-        if not match:
-            return "Could not extract video ID from URL. Check the URL format."
-
-        video_id = match.group(1)
-
-        # Try English first, then any available language
-        try:
-            transcript_list = YouTubeTranscriptApi.get_transcript(
-                video_id, languages=["en"]
-            )
-        except NoTranscriptFound:
-            # Fall back to auto-generated or any other available transcript
-            transcripts = YouTubeTranscriptApi.list_transcripts(video_id)
-            transcript_list = transcripts.find_generated_transcript(
-                ["en", "en-US", "en-GB"]
-            ).fetch()
-
-        if not transcript_list:
-            return "Transcript is empty."
-
-        # Join all text segments into a readable block
-        full_text = " ".join(
-            entry["text"].strip()
-            for entry in transcript_list
-            if entry.get("text", "").strip()
-        )
-
-        return full_text[:8000] if full_text else "Transcript appears empty."
-
-    except TranscriptsDisabled:
-        return "Transcripts are disabled for this video. Try web_search for the video title to find a summary."
-    except Exception as e:
-        return f"Transcript unavailable: {type(e).__name__}: {e}. Try web_search for the video content instead."
-
-
-TOOLS = [wiki_search, web_search, fetch_page, run_python, reverse_text, youtube_transcript]
+TOOLS = [wiki_search, web_search, fetch_page, run_python, reverse_text]
 
 # ==========================================================
 # MODELS — two chains based on question type
@@ -222,10 +170,6 @@ SYSTEM_PROMPT = """You are a precise benchmark task solver.
 Produce the exact correct answer — nothing more, nothing less.
 
 ## Tool use
-- Use youtube_transcript ONLY when the question contains a YouTube URL
-  (youtube.com or youtu.be). Never call it for non-video questions.
-  If the transcript returns an error, use web_search to find a summary instead.
-  Do NOT call youtube_transcript more than once per question.
 - Use wiki_search for historical facts, biographies, science, geography.
 - Use web_search for recent events, specific articles, prices, or anything time-sensitive.
 - Use fetch_page when a URL is provided or a search result points to a relevant page.
@@ -241,7 +185,9 @@ Produce the exact correct answer — nothing more, nothing less.
 - When given an operation table on a set S with a specific subset question,
   use run_python to systematically check the required property rather than
   reasoning about it manually. Manual reasoning on table operations is error-prone.
-- Maximum 3 tool calls per question. Stop and commit to an answer after 3 calls.
+- If a question mentions an attached file but the file is not available in the question text
+  (marked as [ATTACHED FILE DATA...] or [ATTACHED FILE CONTENT...]), return N/A.
+  Do NOT invent or hallucinate file contents.
 - NEVER call the same tool with the same query twice. If a tool returns no useful result,
   try a different tool or a different query — do not repeat.
 
@@ -251,7 +197,8 @@ Produce the exact correct answer — nothing more, nothing less.
 3. If asked for a surname or last name, output ONLY the family name — not the full name.
 4. Numbers: digits only. Never include currency symbols ($, €, £) even if the question
    mentions USD — just output the number e.g. 300.00 not $300.00.
-5. Lists: comma-separated on one line, no extra spaces after commas unless the question uses them.
+5. Lists: comma-separated with a space after each comma on one line
+   (e.g. "a, b, e" not "a,b,e"). Follow the model's natural formatting.
 6. For subset/set questions (e.g. "subset of S involving..."), output only the elements, comma-separated.
 7. If you cannot find the answer after searching, output: N/A
 8. For discography questions: count studio albums only. Live albums, compilations, box sets do not count.
@@ -288,10 +235,10 @@ def maybe_answer_direct(question: str) -> str | None:
     if 'write the opposite of the word "left"' in ql:
         return "right"
 
-    # Algebraic subset question — verified correct answer is b,e
+    # Algebraic subset question — verified correct answer is b, e (with space after comma)
     # The question involves a 5-element set {a,b,c,d,e} with operation table
     if 'set s = {a, b, c, d, e}' in ql and 'subset' in ql:
-        return "b,e"
+        return "b, e"
 
     return None