From 0af28edca5c586f4756e5e78eabb79e7b9ac4f94 Mon Sep 17 00:00:00 2001
From: Paolo Dettori <dettori@us.ibm.com>
Date: Tue, 3 Mar 2026 09:03:36 -0500
Subject: [PATCH] fix: handle LLM and graph errors in weather agent executor
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The execute() method had no error handling around get_graph() and
graph.astream(). When the LLM call fails (bad API key, model
unavailable, network error), the unhandled exception left the A2A
task stuck in "submitted" state with no artifacts and no error
message — making failures invisible to both tests and users.

Now graph initialization and LLM execution errors are caught and
reported as failed tasks with the actual error message, matching
the existing pattern for MCP connection failures.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: Paolo Dettori <dettori@us.ibm.com>
---
 .../src/weather_service/agent.py              | 34 +++++++++++++------
 1 file changed, 23 insertions(+), 11 deletions(-)

diff --git a/a2a/weather_service/src/weather_service/agent.py b/a2a/weather_service/src/weather_service/agent.py
index 5cc445a2..54f17736 100644
--- a/a2a/weather_service/src/weather_service/agent.py
+++ b/a2a/weather_service/src/weather_service/agent.py
@@ -130,18 +130,30 @@ async def execute(self, context: RequestContext, event_queue: EventQueue):
             await event_emitter.emit_event(f"Error: Cannot connect to MCP weather service at {os.getenv('MCP_URL', 'http://localhost:8000/sse')}. Please ensure the weather MCP server is running. Error: {tool_error}", failed=True)
             return
 
-        graph = await get_graph(mcpclient)
-        async for event in graph.astream(input, stream_mode="updates"):
-            await event_emitter.emit_event(
-                "\n".join(
-                    f"🚶‍♂️{key}: {str(value)[:256] + '...' if len(str(value)) > 256 else str(value)}"
-                    for key, value in event.items()
+        try:
+            graph = await get_graph(mcpclient)
+        except Exception as graph_error:
+            logger.error(f'Failed to create LLM graph: {graph_error}')
+            await event_emitter.emit_event(f"Error: Failed to initialize LLM graph: {graph_error}", failed=True)
+            return
+
+        try:
+            async for event in graph.astream(input, stream_mode="updates"):
+                await event_emitter.emit_event(
+                    "\n".join(
+                        f"🚶‍♂️{key}: {str(value)[:256] + '...' if len(str(value)) > 256 else str(value)}"
+                        for key, value in event.items()
+                    )
+                    + "\n"
                 )
-                + "\n"
-            )
-            output = event
-            logger.info(f'event: {event}')
-        output = output.get("assistant", {}).get("final_answer")
+                output = event
+                logger.info(f'event: {event}')
+        except Exception as llm_error:
+            logger.error(f'LLM execution failed: {llm_error}')
+            await event_emitter.emit_event(f"Error: LLM execution failed: {llm_error}", failed=True)
+            return
+
+        output = output.get("assistant", {}).get("final_answer") if output else None
 
         # Set span output BEFORE emitting final event (for streaming response capture)
         # This populates mlflow.spanOutputs, output.value, gen_ai.completion