From cd220ee0a39a1e8facfade8342ba762ab89c1095 Mon Sep 17 00:00:00 2001 From: Yinghan Ma Date: Wed, 13 Aug 2025 03:34:53 +0000 Subject: [PATCH 1/2] fix failure reason --- eval_protocol/mcp/execution/manager.py | 15 ++++++++++++--- 1 file changed, 12 insertions(+), 3 deletions(-) diff --git a/eval_protocol/mcp/execution/manager.py b/eval_protocol/mcp/execution/manager.py index aec867fc..41412d1f 100644 --- a/eval_protocol/mcp/execution/manager.py +++ b/eval_protocol/mcp/execution/manager.py @@ -14,6 +14,7 @@ from dataclasses import asdict from typing import TYPE_CHECKING, Any, Callable, Dict, List, Optional, Union +import anyio from openai.types import CompletionUsage from vendor.tau2.data_model.message import AssistantMessage, UserMessage @@ -464,11 +465,19 @@ async def _execute_rollout( ) except asyncio.CancelledError: - logger.error(f"🚨 AsyncIO Cancel Error in roll out {rollout_idx}", exc_info=True) failure_reason = "asyncio context cancelled" + logger.error( + f"🚨 Error in rollout {session.dataset_row.id} {rollout_idx}: {failure_reason}", exc_info=True + ) + except (anyio.ClosedResourceError, anyio.BrokenResourceError): + logger.error( + f"🚨 Error in rollout {session.dataset_row.id} {rollout_idx}: {failure_reason}", exc_info=True + ) + failure_reason = "connection/resource error" except Exception as e: - logger.error(f"🚨 Error in rollout {rollout_idx}: {e}", exc_info=True) - failure_reason = str(e) + error_msg = str(e) if str(e) else f"{type(e).__name__}: Connection/resource error" + logger.error(f"🚨 Error in rollout {session.dataset_row.id} {rollout_idx}: {error_msg}", exc_info=True) + failure_reason = error_msg finally: if failure_reason: trajectory.terminated = True From ad8b34f1fc12058301efe0b2b4a5d306f52a3c7a Mon Sep 17 00:00:00 2001 From: Yinghan Ma Date: Wed, 13 Aug 2025 03:37:58 +0000 Subject: [PATCH 2/2] update --- eval_protocol/mcp/execution/manager.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/eval_protocol/mcp/execution/manager.py b/eval_protocol/mcp/execution/manager.py index 41412d1f..5664e5ac 100644 --- a/eval_protocol/mcp/execution/manager.py +++ b/eval_protocol/mcp/execution/manager.py @@ -470,12 +470,12 @@ async def _execute_rollout( f"🚨 Error in rollout {session.dataset_row.id} {rollout_idx}: {failure_reason}", exc_info=True ) except (anyio.ClosedResourceError, anyio.BrokenResourceError): + failure_reason = "anyioconnection/resource error" logger.error( f"🚨 Error in rollout {session.dataset_row.id} {rollout_idx}: {failure_reason}", exc_info=True ) - failure_reason = "connection/resource error" except Exception as e: - error_msg = str(e) if str(e) else f"{type(e).__name__}: Connection/resource error" + error_msg = str(e) if str(e) else f"{type(e).__name__}: Unexpected error" logger.error(f"🚨 Error in rollout {session.dataset_row.id} {rollout_idx}: {error_msg}", exc_info=True) failure_reason = error_msg finally: