diff --git a/scripts/deploy-ubuntu.sh b/scripts/deploy-ubuntu.sh
index e592178..174b352 100755
--- a/scripts/deploy-ubuntu.sh
+++ b/scripts/deploy-ubuntu.sh
@@ -213,6 +213,7 @@ export OPENAI_API_KEY=$(printf "%q" "${api_key}")
 export OPENAI_BASE_URL=$(printf "%q" "${base_url}")
 export OPENAI_MODEL=$(printf "%q" "${model}")
 export OPENAI_TIMEOUT_SECONDS="${OPENAI_TIMEOUT_SECONDS:-3600}"
+export CLAW_MODEL_IDLE_TIMEOUT_SECONDS="${CLAW_MODEL_IDLE_TIMEOUT_SECONDS:-60}"
 
 export CLAW_DEPLOY_INSTANCE=$(printf "%q" "${DEPLOY_INSTANCE}")
 export CLAW_BACKEND_HOST=$(printf "%q" "${backend_host}")
@@ -264,6 +265,7 @@ ensure_env_file() {
 		echo "OPENAI_BASE_URL=${OPENAI_BASE_URL:-}"
 		echo "OPENAI_MODEL=${OPENAI_MODEL:-}"
 		echo "OPENAI_TIMEOUT_SECONDS=${OPENAI_TIMEOUT_SECONDS:-3600}"
+		echo "CLAW_MODEL_IDLE_TIMEOUT_SECONDS=${CLAW_MODEL_IDLE_TIMEOUT_SECONDS:-60}"
 		echo "CLAW_SERVICE_SCOPE=${SERVICE_SCOPE}"
 		echo "CLAW_ENABLE_LINUX_ACCOUNTS=${CLAW_ENABLE_LINUX_ACCOUNTS:-0}"
 		echo "OPENAI_API_KEY=已配置"
diff --git a/scripts/start-backend.sh b/scripts/start-backend.sh
index 28d9fd6..a573e19 100755
--- a/scripts/start-backend.sh
+++ b/scripts/start-backend.sh
@@ -20,7 +20,7 @@ export OPENAI_API_KEY
 export OPENAI_BASE_URL="${OPENAI_BASE_URL:-http://model.mify.ai.srv/v1}"
 export OPENAI_MODEL="${OPENAI_MODEL:-tongyi/deepseek-v4-pro}"
 export OPENAI_TIMEOUT_SECONDS="${OPENAI_TIMEOUT_SECONDS:-3600}"
-export CLAW_MODEL_IDLE_TIMEOUT_SECONDS="${CLAW_MODEL_IDLE_TIMEOUT_SECONDS:-${OPENAI_TIMEOUT_SECONDS}}"
+export CLAW_MODEL_IDLE_TIMEOUT_SECONDS="${CLAW_MODEL_IDLE_TIMEOUT_SECONDS:-60}"
 
 # 后端也会调用 Node 生态工具，例如 feishu-mcp-pro。systemd 不读取交互式
 # shell 的 nvm 配置，所以需要把部署时记录的 Node 路径显式传给 Python 进程。
diff --git a/scripts/start-webui.sh b/scripts/start-webui.sh
index cc9f9c2..4386894 100755
--- a/scripts/start-webui.sh
+++ b/scripts/start-webui.sh
@@ -142,6 +142,7 @@ export OPENAI_API_KEY="${OPENAI_API_KEY:-}"
 export OPENAI_BASE_URL="${OPENAI_BASE_URL:-http://model.mify.ai.srv/v1}"
 export OPENAI_MODEL="${OPENAI_MODEL:-tongyi/deepseek-v4-pro}"
 export OPENAI_TIMEOUT_SECONDS="${OPENAI_TIMEOUT_SECONDS:-3600}"
+export CLAW_MODEL_IDLE_TIMEOUT_SECONDS="${CLAW_MODEL_IDLE_TIMEOUT_SECONDS:-60}"
 
 BACKEND_LOG="${RUN_DIR}/webui-backend.log"
 FRONTEND_LOG="${RUN_DIR}/webui-frontend.log"
diff --git a/src/evaluation_runtime.py b/src/evaluation_runtime.py
index 330edec..8f6897f 100644
--- a/src/evaluation_runtime.py
+++ b/src/evaluation_runtime.py
@@ -2936,6 +2936,9 @@ def transient_model_error_message(value: Any) -> str:
         'connection reset by peer',
         'connectionreseterror',
         'broken pipe',
+        'timed out',
+        'timeouterror',
+        'socket timeout',
     )
     if any(marker in normalized for marker in transport_markers):
         return (
diff --git a/tests/test_evaluation_runtime.py b/tests/test_evaluation_runtime.py
index 8f3e86d..4139c4a 100644
--- a/tests/test_evaluation_runtime.py
+++ b/tests/test_evaluation_runtime.py
@@ -392,6 +392,12 @@ class EvaluationRuntimeTests(unittest.TestCase):
             '响应流意外中断',
             transient_model_error_message('IncompleteRead(199 bytes read)'),
         )
+        self.assertIn(
+            '响应流意外中断',
+            transient_model_error_message(
+                'Unable to reach local model backend: timed out'
+            ),
+        )
 
     def test_transient_model_failure_retries_before_validation(self) -> None:
         with tempfile.TemporaryDirectory() as directory: