fixed tool calling bug; added no_think param and current_time to researcher; updated design diagram; no output guard for researcher

agents.py

@@ -1,6 +1,8 @@
+from typing import List
 from args import Args
 from itf_agent import IAgent
 from toolbox import Toolbox
+import datetime
 
 
 class Manager(IAgent):
@@ -48,6 +50,12 @@ class Researcher(IAgent):
             ]
         super().__init__("05_researcher.txt", Args.PRIMARY_AGENT_PRESET, tools)
 
+    def query(self, messages: List[str]) -> str:
+        last_message = messages[-1]
+        current_time = datetime.datetime.now().isoformat()
+        messages[-1] = f"Current time: {current_time}\n" + last_message
+        return super().query(messages)
+
 
 class Reasoner(IAgent):
     """

app.py

@@ -4,12 +4,7 @@ import requests
 import pandas as pd
 
 from alfred import Alfred
-
-
-# --- Constants ---
-DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
-MOCK_SUBMISSION = True
-QUESTIONS_LIMIT = 3  # Use 0 for no limit !
+from args import Args
 
 
 class Application:
@@ -32,7 +27,7 @@ class Application:
     @staticmethod
     def _get_runtime_and_repo_urls():
         """Determine HF Space Runtime URL and Repo URL"""
-        api_url = DEFAULT_API_URL
+        api_url = Args.AppParams.DEFAULT_API_URL
         questions_url = f"{api_url}/questions"
         submit_url = f"{api_url}/submit"
         return questions_url, submit_url
@@ -88,8 +83,9 @@ class Application:
             - results_log (list): Logs with "Task ID", "Question", and "Submitted Answer".
             - answers_payload (list): Payload with "task_id" and "submitted_answer".
         """
-        if QUESTIONS_LIMIT > 0:
-            questions_data = questions_data[:QUESTIONS_LIMIT]
+        questions_limit = Args.AppParams.QUESTIONS_LIMIT
+        if questions_limit > 0:
+            questions_data = questions_data[:questions_limit]
 
         results_log = []
         answers_payload = []
@@ -128,7 +124,7 @@ class Application:
         """
         print(f"Submitting {len(answers_payload)} answers to: {self.submit_url}")
         try:
-            if MOCK_SUBMISSION:
+            if Args.AppParams.MOCK_SUBMISSION:
                 app_username = self.username
                 class MockResponse:
                     status_code = 200

args.py

@@ -83,7 +83,7 @@ class Args:
     primary_llm_interface=LLMInterface.OPENAI
     # secondary_llm_interface=LLMInterface.HUGGINGFACE
     vlm_interface=LLMInterface.OPENAI
-    primary_model="groot" if TEST_MODE else "qwen2.5-qwq-35b-eureka-cubed-abliterated-uncensored"
+    primary_model="groot" if TEST_MODE else "qwen/qwen3-30b-a3b"
     secondary_model="groot" if TEST_MODE else "qwen2.5-7b-instruct-1m"
     vision_model="groot" if TEST_MODE else "qwen/qwen2.5-vl-7b"
     api_base="http://127.0.0.1:1234/v1"  # LM Studio local endpoint
@@ -102,6 +102,13 @@ class Args:
         vlm_interface, vision_model,
         temperature = None, max_tokens = 2048, repeat_penalty = None
         )
+
+    class AppParams:
+        # --- Constants ---
+        DEFAULT_API_URL = "https://agents-course-unit4-scoring.hf.space"
+        MOCK_SUBMISSION = True
+        QUESTIONS_LIMIT = 3  # Use 0 for no limit !
+
     class AlfredParams:
         # Maximum number of interactions between Manager and Solver
         MAX_INTERACTIONS = 6
@@ -109,3 +116,6 @@ class Args:
         MAX_SOLVING_EFFORT = 6
         # Verification happening every few messages to see whether the manager agent got stuck
         AUDIT_INTERVAL = 3
+
+    class MiscParams:
+        NO_THINK = True

design.puml

@@ -8,31 +8,31 @@ node START TERMINAL_NODE_COLOR[
   START
 ]
 
-node manager IMPLEMENTED_NODE_COLOR[
+node manager TESTED_NODE_COLOR[
   manager
 ]
 
-node final_answer IMPLEMENTED_NODE_COLOR[
+node final_answer TESTED_NODE_COLOR[
   final_answer
 ]
 
-node auditor IMPLEMENTED_NODE_COLOR[
+node auditor TESTED_NODE_COLOR[
   auditor
 ]
 
-node solver IMPLEMENTED_NODE_COLOR[
+node solver TESTED_NODE_COLOR[
   solver
 ]
 
-node researcher IMPLEMENTED_NODE_COLOR[
+node researcher TESTED_NODE_COLOR[
   researcher
 ]
 
-node reasoner IMPLEMENTED_NODE_COLOR[
+node reasoner TESTED_NODE_COLOR[
   reasoner
 ]
 
-node viewer NOT_IMPLEMENTED_NODE_COLOR[
+node viewer IMPLEMENTED_NODE_COLOR[
   viewer
 ]
 

design.yaml

@@ -8,34 +8,34 @@ nodes:
 - name: manager
   connections: [solver, auditor, final_answer]
   description: Orchestrates the workflow by delegating tasks to specialized nodes and integrating their outputs
-  status: IMPLEMENTED
+  status: TESTED
 
 - name: final_answer
   connections: [END]
   description: Formats and delivers the final response to the user
-  status: IMPLEMENTED
+  status: TESTED
 
 - name: auditor
   connections: [manager]
   description: Reviews manager's outputs for accuracy, safety, and quality
-  status: IMPLEMENTED
+  status: TESTED
 
 - name: solver
   connections: [manager, researcher, reasoner, viewer]
   description: Central problem-solving node that coordinates with specialized experts based on task requirements
-  status: IMPLEMENTED
+  status: TESTED
 
 - name: researcher
   connections: [solver]
   description: Retrieves and synthesizes information from various sources to answer knowledge-based questions
-  status: IMPLEMENTED
+  status: TESTED
 
 - name: reasoner
   connections: [solver]
   description: Performs logical reasoning, inference, and step-by-step problem-solving
-  status: IMPLEMENTED
+  status: TESTED
 
 - name: viewer
   connections: [solver]
   description: Processes, analyzes, and generates vision related information
-  status: NOT_IMPLEMENTED
+  status: IMPLEMENTED

graph.py

@@ -107,6 +107,14 @@ class Nodes:
         instruction = "Formulate a definitive final answer in english. Be very concise and use no redundant words !"
         state["messages"].append(instruction)
         response = Agents.manager.query(state["messages"])
+
+        # Post process the response
+        if "FINAL ANSWER:" in response:
+            response = response.split("FINAL ANSWER:", 1)[1]
+        if "</think>" in response:
+            response = response.split("</think>", 1)[1]
+        response = response.strip()
+
         state["final_response"] = response
         return state
 
@@ -140,7 +148,8 @@ class Nodes:
         """
         Retrieves and synthesizes information from various sources to answer knowledge-based questions
         """
-        response = Agents.guard_output(Agents.researcher, state["task_progress"])
+        # We do not use the output guard here as it might halucinate results if there are none.
+        response = Agents.researcher.query(state["task_progress"])
         state["task_progress"].append(response)
         return state
 

itf_agent.py

@@ -14,6 +14,7 @@ class IAgent():
         self.name = self._format_name(sys_prompt_filename)
         self.interface = agent_preset.get_interface()
         self.mock = (agent_preset.get_model_name() == "groot")
+        self.tools = tools  # <-- store tools for tool call execution
 
         # Load the system prompt from a file
         system_prompt_path = os.path.join(os.getcwd(), "system_prompts", sys_prompt_filename)
@@ -75,6 +76,43 @@ class IAgent():
         """
         return self.system_prompt
 
+    def _handle_tool_calls(self, tool_calls):
+        """
+        Executes tool calls and returns their results as a string.
+        """
+        tool_results = []
+        for call in tool_calls:
+            tool_name = None
+            tool_args = {}
+            # Qwen-style: call['function']['name'], call['function']['arguments']
+            if "function" in call:
+                tool_name = call["function"].get("name")
+                import json
+                try:
+                    tool_args = json.loads(call["function"].get("arguments", "{}"))
+                except Exception:
+                    tool_args = {}
+            # OpenAI-style: call['name'], call['args']
+            elif "name" in call and "args" in call:
+                tool_name = call["name"]
+                tool_args = call["args"]
+            tool = next((t for t in self.tools if getattr(t, "name", None) == tool_name), None)
+            if tool is not None:
+                try:
+                    # Handle "__arg1" as positional argument for single-argument tools
+                    if isinstance(tool_args, dict) and len(tool_args) == 1 and "__arg1" in tool_args:
+                        result = tool.func(tool_args["__arg1"])
+                    elif isinstance(tool_args, dict):
+                        result = tool.func(**tool_args)
+                    else:
+                        result = tool.func(tool_args)
+                    tool_results.append(f"[{tool_name}]: {result}")
+                except Exception as e:
+                    tool_results.append(f"[{tool_name} ERROR]: {str(e)}")
+            else:
+                tool_results.append(f"[{tool_name} ERROR]: Tool not found")
+        return "\n".join(tool_results)
+
     def query(self, messages: List[str]) -> str:
         """
         Asynchronously queries the agent with a given question and returns the response.
@@ -98,9 +136,29 @@ class IAgent():
             return response
 
         system_prompt = self.get_system_prompt()
+
+        # Disable thinking block for some models
+        if Args.MiscParams.NO_THINK:
+            messages[-1] += "\n/no_think"
+
         messages_with_roles = self._bake_roles(messages)
         conversation = [SystemMessage(content=system_prompt)] + messages_with_roles
-        response = str(self.model.invoke(conversation).content)
+
+        raw_output = self.model.invoke(conversation)
+
+        # --- Unified output and tool call handling ---
+        response = ""
+
+        # 1. Handle tool calls if present
+        tool_calls = getattr(raw_output, "additional_kwargs", {}).get("tool_calls", None)
+        if tool_calls:
+            response = self._handle_tool_calls(tool_calls)
+        # 2. Otherwise, use standard LLM output if present
+        elif hasattr(raw_output, "content") and raw_output.content:
+            response = str(raw_output.content)
+        # 3. Fallback: string conversion
+        else:
+            response = str(raw_output)
 
         Args.LOGGER.log(logging.INFO, f"\nAgent '{self.name}' produced OUTPUT:\n{response}\n{separator}\n")
         return response

test.py

@@ -152,9 +152,9 @@ class TestAlfredAgent(unittest.TestCase):
         """
         nodes = Nodes()
         test_state: State = {
-            "initial_query": "What is the capital of France?",
-            "messages": ["What is the capital of France?"],
-            "task_progress": ["Research: What is the capital of France?"],
+            "initial_query": "What are the latest news headlines about artificial intelligence published this week?",
+            "messages": ["What are the latest news headlines about artificial intelligence published this week?"],
+            "task_progress": ["What are the latest news headlines about artificial intelligence published this week?"],
             "audit_interval": 2,
             "manager_queries": 1,
             "solver_queries": 0,
@@ -379,4 +379,6 @@ class TestAlfredAgent(unittest.TestCase):
 
 
 if __name__ == "__main__":
+    # test = TestAlfredAgent()
+    # test.test_researcher_node()
     unittest.main()