Update app.py

app.py

@@ -17,6 +17,7 @@ from tools.calculator import Calculator
 from tools.final_answer import FinalAnswerTool
 from tools.visit_webpage import VisitWebpageTool
 from tools.web_search import DuckDuckGoSearchTool
+from tools.attachment_reader import AttachmentReaderTool, download_attachment
 
 from research_state import ResearchState
 from verifier import verify_research
@@ -40,7 +41,7 @@ MAX_TOKENS = 900
 TEMPERATURE = 0.2
 
 # Number of questions to skip from the start of the list (set to 0 to run all)
-SKIP_FIRST_N = 1
+SKIP_FIRST_N = 0
 
 # Pacing / rate-limit handling
 MIN_SECONDS_BETWEEN_QUESTIONS = 60
@@ -267,6 +268,66 @@ class DailyQuotaExhaustedError(Exception):
     pass
 
 
+class NonAnswerError(Exception):
+    """Raised when the agent returns something that is not a real answer
+    (a refusal, an apology, an error message, empty output, etc.)."""
+    pass
+
+
+# Prefixes that indicate the agent is not giving an answer.
+NON_ANSWER_PREFIXES = (
+    "i am sorry",
+    "i'm sorry",
+    "sorry,",
+    "i cannot",
+    "i can't",
+    "i am unable",
+    "i'm unable",
+    "unable to",
+    "error:",
+    "agent error",
+    "skipped",
+    "traceback",
+    "as an ai",
+)
+
+# Phrases that indicate a failure anywhere in the text.
+NON_ANSWER_PHRASES = (
+    "i cannot directly access",
+    "cannot access local files",
+    "unable to access",
+    "please provide the data",
+    "please provide the file",
+    "i do not have access",
+    "i don't have access",
+    "reached max steps",
+    "max steps reached",
+    "could not find an answer",
+    "i was unable to",
+    "no answer found",
+)
+
+
+def _check_is_real_answer(answer) -> None:
+    """Raise NonAnswerError if `answer` is not a usable answer."""
+    if answer is None:
+        raise NonAnswerError("Agent returned None.")
+
+    text = str(answer).strip()
+    if not text:
+        raise NonAnswerError("Agent returned an empty answer.")
+
+    lowered = text.lower()
+
+    for prefix in NON_ANSWER_PREFIXES:
+        if lowered.startswith(prefix):
+            raise NonAnswerError(f"Answer looks like a refusal/error: {text[:200]}")
+
+    for phrase in NON_ANSWER_PHRASES:
+        if phrase in lowered:
+            raise NonAnswerError(f"Answer looks like a refusal/error: {text[:200]}")
+
+
 def _is_rate_limit_error(exc: Exception) -> bool:
     """
     Detect rate-limit / quota errors across providers, since exact
@@ -317,11 +378,15 @@ def _run_agent_with_backoff(agent_runner, question_text,
     backoff if a transient rate-limit error is hit. If the error looks
     like a daily/per-day quota exhaustion, raises DailyQuotaExhaustedError
     immediately instead of retrying, since backoff won't help.
-    Non-rate-limit errors are raised immediately (handled by the caller).
+    NonAnswerError is re-raised untouched (a refusal that happens to
+    mention "quota" or "429" must not be mistaken for a rate limit).
+    Other errors are raised immediately (handled by the caller).
     """
     for attempt in range(max_retries + 1):
         try:
             return agent_runner(question_text)
+        except NonAnswerError:
+            raise  # not a rate-limit issue, don't inspect or retry
         except Exception as e:
             if _is_daily_quota_exhausted(e):
                 raise DailyQuotaExhaustedError(str(e)) from e
@@ -361,6 +426,8 @@ verify_research_tool = VerifyResearchTool()
 
 final_answer = FinalAnswerTool()
 
+attachment_reader = AttachmentReaderTool()
+
 youtube_video_analysis = YouTubeVideoAnalysisTool(
     model="gemini-2.5-flash"
 )
@@ -371,6 +438,7 @@ tools = [
     calculator,
     verify_research_tool,
     youtube_transcript,
+    attachment_reader,
     final_answer,
     youtube_video_analysis,
 ]
@@ -445,9 +513,11 @@ class ResearchAgentWrapper:
     BasicAgent: agent(question) -> answer string.
     Resets the shared research context for every new question
     so state doesn't leak between questions.
+    Raises NonAnswerError instead of returning anything that
+    isn't a real answer.
     """
 
-    def __init__(self, agent: CodeAgent):
+    def __init__(self, agent):
         self.agent = agent
         print("ResearchAgentWrapper initialized.")
 
@@ -457,7 +527,10 @@ class ResearchAgentWrapper:
         research_context.start(question)
         answer = self.agent.run(question)
         print(f"Agent returning answer (first 50 chars): {str(answer)[:50]}...")
-        return str(answer)
+
+        # Never let a non-answer leave the wrapper.
+        _check_is_real_answer(answer)
+        return str(answer).strip()
 
 
 # =========================================================
@@ -534,22 +607,51 @@ def run_and_submit_all(profile: gr.OAuthProfile | None):
 
         start_time = time.monotonic()
 
+        # Download the attachment (if the task has one) and point the agent at it.
+        # download_attachment never raises; it returns None if there is no file.
+        attachment_path = download_attachment(api_url, task_id)
+        agent_question = question_text
+        if attachment_path:
+            agent_question += (
+                f"\n\nAn attached file for this task is saved locally at: "
+                f"{attachment_path}\nUse the read_attachment tool to read it."
+            )
+
         try:
-            submitted_answer = _run_agent_with_backoff(agent_runner, question_text)
+            submitted_answer = _run_agent_with_backoff(agent_runner, agent_question)
             answers_payload.append({"task_id": task_id, "submitted_answer": submitted_answer})
-            results_log.append({"Task ID": task_id, "Question": question_text, "Submitted Answer": submitted_answer})
+            results_log.append({
+                "Task ID": task_id,
+                "Question": question_text,
+                "Submitted Answer": submitted_answer,
+                "Status": "OK",
+            })
+        except NonAnswerError as e:
+            print(f"No valid answer for task {task_id}, not submitting: {e}")
+            results_log.append({
+                "Task ID": task_id,
+                "Question": question_text,
+                "Submitted Answer": "",
+                "Status": f"NO ANSWER: {e}",
+            })
         except DailyQuotaExhaustedError as e:
             print(f"Daily quota exhausted, stopping run early: {e}")
             results_log.append({
                 "Task ID": task_id,
                 "Question": question_text,
-                "Submitted Answer": f"SKIPPED (daily quota exhausted): {e}",
+                "Submitted Answer": "",
+                "Status": f"SKIPPED (daily quota exhausted): {e}",
             })
             quota_exhausted = True
             break
         except Exception as e:
             print(f"Error running agent on task {task_id}: {e}")
-            results_log.append({"Task ID": task_id, "Question": question_text, "Submitted Answer": f"AGENT ERROR: {e}"})
+            results_log.append({
+                "Task ID": task_id,
+                "Question": question_text,
+                "Submitted Answer": "",
+                "Status": f"AGENT ERROR: {e}",
+            })
 
         # Pace requests: ensure at least MIN_SECONDS_BETWEEN_QUESTIONS
         # seconds between question starts, accounting for however long