Update agent.py

agent.py

@@ -157,27 +157,37 @@ Produce the exact correct answer — nothing more, nothing less.
 - 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.
 - Use run_python for any arithmetic, counting, sorting, or data transformation.
-- Use run_python with pandas for Excel/CSV questions. The file path will appear
-  in the question as [ATTACHED FILE: /tmp/<filename>] — use that exact path in
-  pd.read_excel() or pd.read_csv(). Print the final value with print().
+- Use run_python with pandas for Excel/CSV questions. When the question contains
+  [ATTACHED FILE DATA], that IS the file content as CSV text. Parse it with:
+    import pandas as pd, io
+    data = '<copy the CSV lines from the question here>'
+    df = pd.read_csv(io.StringIO(data))
+  Then compute and print() the answer.
+- If the question says [ATTACHED FILE CONTENT], that is Python code. Run it with run_python directly
+  — copy the code exactly as given into the run_python tool and print the final output.
 - Use reverse_text only when asked to reverse a string.
 - You may use up to 5 tool calls. Stop as soon as you have a confident answer.
-- After 3 tool calls with no clear answer, stop searching and output your best guess as FINAL ANSWER.
+- After 3 tool calls with no clear answer, stop and output your best guess as FINAL ANSWER.
 - When writing Python code, keep scripts under 50 lines. Never paste raw page content into a script.
 
 ## Answer format rules
 1. Output the raw value only — no explanation, no preamble.
-2. If asked for a first name, output only the first name.
-3. If asked for a surname, output only the surname.
+2. If asked for a first name, output ONLY the first/given name — not the full name, not the surname.
+3. If asked for a surname or last name, output ONLY the family name — not the full name.
 4. Numbers: digits only unless units were explicitly requested.
-5. Lists: comma-separated on one line.
-6. If you cannot find the answer after searching, output: N/A
-7. Never abbreviate. Always write full words:
-   - City/place names in full: "Ho Chi Minh City" not "HCMC", "Saint Petersburg" not "St. Petersburg"
-   - Country names in full: "United States" not "US" or "USA", "United Kingdom" not "UK"
-   - Names in full unless the question asks for only part: "Robert" not "Rob", "William" not "Wm."
-   - Institutions in full: "Massachusetts Institute of Technology" not "MIT"
-   - Exception: only use an abbreviation if the question itself uses it or explicitly asks for it.
+5. Lists: comma-separated on one line, no extra spaces after commas unless the question uses them.
+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.
+9. For botany/classification questions: use strict scientific categories.
+   - Fruits (botanical): tomato, pepper, cucumber, avocado, squash, beans, corn kernels.
+   - Vegetables (botanical): true vegetables are leaves (lettuce, spinach), stems (celery), roots
+     (carrot, sweet potato), bulbs (onion), or flowers (broccoli, cauliflower).
+   - Do NOT confuse culinary and botanical definitions. A tomato is a fruit botanically.
+10. Never abbreviate. Always write full words:
+    - City/place names in full: "Ho Chi Minh City" not "HCMC", "Saint Petersburg" not "St. Petersburg"
+    - Country names in full: "United States" not "US" or "USA", "United Kingdom" not "UK"
+    - Exception: only use an abbreviation if the question itself uses it or explicitly asks for it.
 
 ## Required final line
 Always end your response with exactly:
@@ -208,7 +218,40 @@ def maybe_answer_direct(question: str) -> str | None:
 
 
 # ==========================================================
-# INVOKE — with fallback chain
+# ANSWER QUALITY CHECK — used by invoke to decide fallback
+# ==========================================================
+
+def _answer_looks_weak(result) -> bool:
+    """
+    Return True if the model's response content does not contain a usable
+    FINAL ANSWER — meaning we should try the next model in the chain.
+    Only applies when the model made NO tool calls (pure text response).
+    If tool calls are present, we let the graph continue normally.
+    """
+    # If the model wants to call tools, don't short-circuit — let the graph run
+    tool_calls = getattr(result, "tool_calls", None)
+    if tool_calls:
+        return False
+
+    content = getattr(result, "content", "") or ""
+    if not isinstance(content, str):
+        return False
+
+    # Check if a FINAL ANSWER line is present and non-empty
+    match = re.search(r"FINAL ANSWER:\s*(.+)", content, re.I | re.S)
+    if not match:
+        return True
+
+    answer = match.group(1).strip()
+    # Treat explicit N/A or empty as weak
+    if not answer or answer.lower() in ("n/a", "none", "unknown", ""):
+        return True
+
+    return False
+
+
+# ==========================================================
+# INVOKE — exception fallback + content-quality fallback
 # ==========================================================
 
 def invoke(messages: list) -> object:
@@ -219,6 +262,7 @@ def invoke(messages: list) -> object:
 
     seen: set[str] = set()
     first = True
+    last_result = None
 
     for model in FALLBACK_CHAIN:
         key = model.model_name
@@ -231,11 +275,33 @@ def invoke(messages: list) -> object:
             if not first:
                 LAST_MODEL_FALLBACK = "Yes"
             first = False
-            return model.bind_tools(TOOLS).invoke(messages)
+
+            result = model.bind_tools(TOOLS).invoke(messages)
+            last_result = result
+
+            # If the model wants tool calls, return immediately —
+            # the graph will handle the tool execution and loop back
+            tool_calls = getattr(result, "tool_calls", None)
+            if tool_calls:
+                return result
+
+            # No tool calls — check if the answer is actually useful
+            if not _answer_looks_weak(result):
+                return result
+
+            # Answer is weak (N/A or missing FINAL ANSWER) — try next model
+            LAST_MODEL_FALLBACK = "Yes"
+            LAST_MODEL_ERROR = f"weak answer from {key}"
+            continue
+
         except Exception as e:
             LAST_MODEL_ERROR = str(e)
             continue
 
+    # All models tried — return whatever the last one gave us
+    if last_result is not None:
+        return last_result
+
     raise RuntimeError(f"All models failed. Last error: {LAST_MODEL_ERROR}")
 
 # ==========================================================