Merge pull request #91 from aguzererler/fix-analyst-tool-loop-18174774985345323969

fix: use run_tool_loop instead of invoke in analyst agents

tests/unit/agents/test_analyst_agents.py

@@ -0,0 +1,73 @@
+from unittest.mock import MagicMock
+import pytest
+from langchain_core.messages import AIMessage, HumanMessage, ToolMessage
+from langchain_core.runnables import Runnable
+from tradingagents.agents.analysts.fundamentals_analyst import create_fundamentals_analyst
+from tradingagents.agents.analysts.market_analyst import create_market_analyst
+from tradingagents.agents.analysts.social_media_analyst import create_social_media_analyst
+from tradingagents.agents.analysts.news_analyst import create_news_analyst
+
+class MockRunnable(Runnable):
+    def __init__(self, invoke_responses):
+        self.invoke_responses = invoke_responses
+        self.call_count = 0
+
+    def invoke(self, input, config=None, **kwargs):
+        response = self.invoke_responses[self.call_count]
+        self.call_count += 1
+        return response
+
+class MockLLM(Runnable):
+    def __init__(self, invoke_responses):
+        self.runnable = MockRunnable(invoke_responses)
+        self.tools_bound = None
+
+    def invoke(self, input, config=None, **kwargs):
+        return self.runnable.invoke(input, config=config, **kwargs)
+
+    def bind_tools(self, tools):
+        self.tools_bound = tools
+        return self.runnable
+
+@pytest.fixture
+def mock_state():
+    return {
+        "messages": [HumanMessage(content="Analyze AAPL.")],
+        "trade_date": "2024-05-15",
+        "company_of_interest": "AAPL",
+    }
+
+@pytest.fixture
+def mock_llm_with_tool_call():
+    # 1. First call: The LLM decides to use a tool
+    tool_call_msg = AIMessage(
+        content="",
+        tool_calls=[
+            {"name": "mock_tool", "args": {"query": "test"}, "id": "call_123"}
+        ]
+    )
+    # 2. Second call: The LLM receives the tool output and writes the report
+    final_report_msg = AIMessage(
+        content="This is the final report after running the tool."
+    )
+    return MockLLM([tool_call_msg, final_report_msg])
+
+def test_fundamentals_analyst_tool_loop(mock_state, mock_llm_with_tool_call):
+    node = create_fundamentals_analyst(mock_llm_with_tool_call)
+    result = node(mock_state)
+    assert "This is the final report after running the tool." in result["fundamentals_report"]
+
+def test_market_analyst_tool_loop(mock_state, mock_llm_with_tool_call):
+    node = create_market_analyst(mock_llm_with_tool_call)
+    result = node(mock_state)
+    assert "This is the final report after running the tool." in result["market_report"]
+
+def test_social_media_analyst_tool_loop(mock_state, mock_llm_with_tool_call):
+    node = create_social_media_analyst(mock_llm_with_tool_call)
+    result = node(mock_state)
+    assert "This is the final report after running the tool." in result["sentiment_report"]
+
+def test_news_analyst_tool_loop(mock_state, mock_llm_with_tool_call):
+    node = create_news_analyst(mock_llm_with_tool_call)
+    result = node(mock_state)
+    assert "This is the final report after running the tool." in result["news_report"]

tradingagents/agents/analysts/fundamentals_analyst.py

@@ -11,6 +11,8 @@ from tradingagents.agents.utils.fundamental_data_tools import (
     get_sector_relative,
 )
 from tradingagents.agents.utils.news_data_tools import get_insider_transactions
+from tradingagents.agents.utils.tool_runner import run_tool_loop
+from tradingagents.agents.utils.agent_utils import build_instrument_context
 from tradingagents.dataflows.config import get_config
 
 
@@ -66,12 +68,9 @@ def create_fundamentals_analyst(llm):
 
         chain = prompt | llm.bind_tools(tools)
 
-        result = chain.invoke(state["messages"])
+        result = run_tool_loop(chain, state["messages"], tools)
 
-        report = ""
+        report = result.content or ""
-
-        if len(result.tool_calls) == 0:
-            report = result.content
 
         return {
             "messages": [result],

tradingagents/agents/analysts/market_analyst.py

@@ -3,6 +3,8 @@ import time
 from tradingagents.agents.utils.core_stock_tools import get_stock_data
 from tradingagents.agents.utils.technical_indicators_tools import get_indicators
 from tradingagents.agents.utils.fundamental_data_tools import get_macro_regime
+from tradingagents.agents.utils.tool_runner import run_tool_loop
+from tradingagents.agents.utils.agent_utils import build_instrument_context
 from tradingagents.dataflows.config import get_config
 
 
@@ -73,16 +75,14 @@ Volume-Based Indicators:
 
         chain = prompt | llm.bind_tools(tools)
 
-        result = chain.invoke(state["messages"])
+        result = run_tool_loop(chain, state["messages"], tools)
 
-        report = ""
+        report = result.content or ""
         macro_regime_report = ""
 
-        if len(result.tool_calls) == 0:
+        # Extract macro regime section if present
-            report = result.content
+        if report and ("Macro Regime Classification" in report or "RISK-ON" in report.upper() or "RISK-OFF" in report.upper() or "TRANSITION" in report.upper()):
-            # Extract macro regime section if present
+            macro_regime_report = report
-            if "Macro Regime Classification" in report or "RISK-ON" in report.upper() or "RISK-OFF" in report.upper() or "TRANSITION" in report.upper():
-                macro_regime_report = report
 
         return {
             "messages": [result],

tradingagents/agents/analysts/news_analyst.py

@@ -1,6 +1,8 @@
 from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
 import json
 from tradingagents.agents.utils.news_data_tools import get_news, get_global_news
+from tradingagents.agents.utils.tool_runner import run_tool_loop
+from tradingagents.agents.utils.agent_utils import build_instrument_context
 from tradingagents.dataflows.config import get_config
 
 
@@ -42,12 +44,9 @@ def create_news_analyst(llm):
         prompt = prompt.partial(instrument_context=instrument_context)
 
         chain = prompt | llm.bind_tools(tools)
-        result = chain.invoke(state["messages"])
+        result = run_tool_loop(chain, state["messages"], tools)
 
-        report = ""
+        report = result.content or ""
-
-        if len(result.tool_calls) == 0:
-            report = result.content
 
         return {
             "messages": [result],

tradingagents/agents/analysts/social_media_analyst.py

@@ -2,6 +2,8 @@ from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
 import time
 import json
 from tradingagents.agents.utils.news_data_tools import get_news
+from tradingagents.agents.utils.tool_runner import run_tool_loop
+from tradingagents.agents.utils.agent_utils import build_instrument_context
 from tradingagents.dataflows.config import get_config
 
 
@@ -43,12 +45,9 @@ def create_social_media_analyst(llm):
 
         chain = prompt | llm.bind_tools(tools)
 
-        result = chain.invoke(state["messages"])
+        result = run_tool_loop(chain, state["messages"], tools)
 
-        report = ""
+        report = result.content or ""
-
-        if len(result.tool_calls) == 0:
-            report = result.content
 
         return {
             "messages": [result],