feat: improve error handling

debug_llm_call.py

@@ -0,0 +1,42 @@
+"""Standalone diagnostic script to test a single LLM call with resilience.
+Run: python debug_llm_call.py --provider openai --model gpt-4o-mini --message "Test message".
+It will respect environment variables for keys and SSL the same way the graph does.
+"""
+import argparse
+import os
+from tradingagents.default_config import DEFAULT_CONFIG
+from tradingagents.graph.trading_graph import TradingAgentsGraph
+from langchain_core.messages import HumanMessage
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--provider', default=DEFAULT_CONFIG['llm_provider'])
+    parser.add_argument('--model', default=DEFAULT_CONFIG['quick_think_llm'])
+    parser.add_argument('--message', default='Say hello and include a short market summary placeholder.')
+    args = parser.parse_args()
+
+    cfg = DEFAULT_CONFIG.copy()
+    cfg['llm_provider'] = args.provider
+    cfg['quick_think_llm'] = args.model
+    cfg['deep_think_llm'] = args.model
+
+    graph = TradingAgentsGraph(config=cfg)
+    # Build a minimal state for market analyst
+    state = {
+        'trade_date': '2025-09-29',
+        'company_of_interest': 'AAPL',
+        'messages': [HumanMessage(content=args.message)],
+    }
+    market_node = graph.graph_setup.analyst_nodes.get('market')
+    if not market_node:
+        print('Market node not found in graph setup.')
+        return
+    # Directly invoke underlying function if possible
+    result_state = market_node(state)
+    print('Result keys:', list(result_state.keys()))
+    print('Market report snippet:', str(result_state.get('market_report',''))[:500])
+
+
+if __name__ == '__main__':
+    main()

tradingagents/agents/analysts/fundamentals_analyst.py

@@ -1,6 +1,7 @@
 from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
 import time
 import json
+from tradingagents.agents.utils.llm_resilience import invoke_with_retries
 
 
 def create_fundamentals_analyst(llm, toolkit):
@@ -55,13 +56,19 @@ def create_fundamentals_analyst(llm, toolkit):
         prompt = prompt.partial(ticker=ticker)
 
         chain = prompt | llm.bind_tools(tools)
-
+        try:
-        result = chain.invoke(state["messages"])
+            result = invoke_with_retries(chain, state["messages"], toolkit.config)
+        except Exception as e:  # noqa: BLE001
+            class DummyResult:
+                def __init__(self, content):
+                    self.content = content
+                    self.tool_calls = []
+            result = DummyResult(f"Fundamentals analyst failed after retries. Error: {e}")
 
         report = ""
 
-        if len(result.tool_calls) == 0:
+        if getattr(result, 'tool_calls', []) == []:
-            report = result.content
+            report = getattr(result, 'content', '')
 
         return {
             "messages": [result],

tradingagents/agents/analysts/market_analyst.py

@@ -1,6 +1,7 @@
 from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
 import time
 import json
+from tradingagents.agents.utils.llm_resilience import invoke_with_retries
 
 
 def create_market_analyst(llm, toolkit):
@@ -85,12 +86,22 @@ Bullish and Bearish Candlestick Patterns:
 
         chain = prompt | llm.bind_tools(tools)
 
-        result = chain.invoke(state["messages"])
+        # Resilient invocation with retries
+        try:
+            result = invoke_with_retries(chain, state["messages"], toolkit.config)
+        except Exception as e:  # noqa: BLE001
+            # Provide a graceful degraded response so graph can continue / be logged
+            fallback_content = f"Market analyst failed to retrieve a model response after retries. Error: {e}"
+            class DummyResult:
+                def __init__(self, content):
+                    self.content = content
+                    self.tool_calls = []
+            result = DummyResult(fallback_content)
 
         report = ""
 
-        if len(result.tool_calls) == 0:
+        if getattr(result, 'tool_calls', []) == []:
-            report = result.content
+            report = getattr(result, 'content', '')
        
         return {
             "messages": [result],

tradingagents/agents/analysts/news_analyst.py

@@ -1,6 +1,7 @@
 from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
 import time
 import json
+from tradingagents.agents.utils.llm_resilience import invoke_with_retries
 
 
 def create_news_analyst(llm, toolkit):
@@ -45,12 +46,19 @@ def create_news_analyst(llm, toolkit):
         prompt = prompt.partial(ticker=ticker)
 
         chain = prompt | llm.bind_tools(tools)
-        result = chain.invoke(state["messages"])
+        try:
+            result = invoke_with_retries(chain, state["messages"], toolkit.config)
+        except Exception as e:  # noqa: BLE001
+            class DummyResult:
+                def __init__(self, content):
+                    self.content = content
+                    self.tool_calls = []
+            result = DummyResult(f"News analyst failed after retries. Error: {e}")
 
         report = ""
 
-        if len(result.tool_calls) == 0:
+        if getattr(result, 'tool_calls', []) == []:
-            report = result.content
+            report = getattr(result, 'content', '')
 
         return {
             "messages": [result],

tradingagents/agents/analysts/social_media_analyst.py

@@ -1,6 +1,7 @@
 from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
 import time
 import json
+from tradingagents.agents.utils.llm_resilience import invoke_with_retries
 
 
 def create_social_media_analyst(llm, toolkit):
@@ -44,13 +45,19 @@ def create_social_media_analyst(llm, toolkit):
         prompt = prompt.partial(ticker=ticker)
 
         chain = prompt | llm.bind_tools(tools)
-
+        try:
-        result = chain.invoke(state["messages"])
+            result = invoke_with_retries(chain, state["messages"], toolkit.config)
+        except Exception as e:  # noqa: BLE001
+            class DummyResult:
+                def __init__(self, content):
+                    self.content = content
+                    self.tool_calls = []
+            result = DummyResult(f"Social media analyst failed after retries. Error: {e}")
 
         report = ""
 
-        if len(result.tool_calls) == 0:
+        if getattr(result, 'tool_calls', []) == []:
-            report = result.content
+            report = getattr(result, 'content', '')
 
         return {
             "messages": [result],

tradingagents/agents/utils/llm_resilience.py

@@ -0,0 +1,46 @@
+import time
+import json
+import logging
+from typing import Any, Callable, Dict
+from json import JSONDecodeError
+
+logger = logging.getLogger(__name__)
+
+
+def invoke_with_retries(chain: Any, messages: Any, config: Dict[str, Any]):
+    """Invoke a langchain chain with retries and detailed logging.
+
+    Handles transient HTTP issues and JSON decode errors coming from provider SDKs.
+    """
+    max_retries = config.get("llm_max_retries", 3)
+    backoff = config.get("llm_retry_backoff", 2.0)
+
+    last_err = None
+    for attempt in range(1, max_retries + 1):
+        try:
+            result = chain.invoke(messages)
+            return result
+        except JSONDecodeError as e:
+            last_err = e
+            logger.warning(
+                "JSONDecodeError on attempt %s/%s: %s", attempt, max_retries, e
+            )
+        except Exception as e:  # noqa: BLE001
+            # Capture common transient network / HTTP errors keywords
+            transient = any(
+                kw in str(e).lower() for kw in [
+                    "timeout", "temporarily", "rate limit", "connection reset", "503", "502", "jsondecodeerror"
+                ]
+            )
+            last_err = e
+            logger.warning(
+                "LLM invocation error (transient=%s) attempt %s/%s: %s", transient, attempt, max_retries, e
+            )
+            if not transient and not isinstance(e, JSONDecodeError):
+                # Non transient -> abort early
+                break
+        # Exponential backoff
+        sleep_for = backoff ** (attempt - 1)
+        time.sleep(sleep_for)
+    # All attempts failed
+    raise last_err  # propagate last error

tradingagents/default_config.py

@@ -33,4 +33,8 @@ DEFAULT_CONFIG = {
     # Proxy settings (if needed)
     "http_proxy": os.getenv("HTTP_PROXY"),
     "https_proxy": os.getenv("HTTPS_PROXY"),
+    # LLM resilience settings
+    "llm_max_retries": int(os.getenv("LLM_MAX_RETRIES", "3")),
+    "llm_retry_backoff": float(os.getenv("LLM_RETRY_BACKOFF", "2")),  # seconds exponential base
+    "debug_http": os.getenv("DEBUG_HTTP", "false").lower() in ("1", "true", "yes"),
 }