fix: 收敛 Agent 会话队列与 worker 关闭 (#6386)

app/api/endpoints/anthropic.py

@@ -13,6 +13,7 @@ from app.schemas.openai import AnthropicTextBlock as _SchemaAnthropicTextBlock
 from app.api.endpoints.openai import (
     MODEL_ID,
     _is_manager_unavailable,
+    _is_manager_queue_full,
     _run_managed_agent,
 )
 from app.api.openai_utils import (
@@ -246,6 +247,12 @@ async def messages(
                 503,
                 error_type="api_error",
             )
+        if _is_manager_queue_full(exc):
+            return _anthropic_error_response(
+                str(exc),
+                429,
+                error_type="rate_limit_error",
+            )
         return _anthropic_error_response(str(exc), 500, error_type="api_error")
     finally:
         await manager.clear_session(session_id=session_id, user_id=session_id)

app/api/endpoints/openai.py

@@ -346,6 +346,11 @@ def _is_manager_unavailable(error: BaseException) -> bool:
     return getattr(error, "code", None) == "agent_manager_unavailable"
 
 
+def _is_manager_queue_full(error: BaseException) -> bool:
+    """识别 Agent 会话排队已满，供兼容 API 返回可重试状态。"""
+    return getattr(error, "code", None) == "agent_manager_queue_full"
+
+
 async def _run_managed_agent(
     *,
     manager,
@@ -552,6 +557,13 @@ async def chat_completions(
                 error_type="server_error",
                 code="ai_agent_unavailable",
             )
+        if _is_manager_queue_full(exc):
+            return _error_response(
+                str(exc),
+                429,
+                error_type="rate_limit_error",
+                code="ai_agent_queue_full",
+            )
         return _error_response(
             str(exc),
             500,
@@ -652,6 +664,13 @@ async def responses(
                 error_type="server_error",
                 code="ai_agent_unavailable",
             )
+        if _is_manager_queue_full(exc):
+            return _error_response(
+                str(exc),
+                429,
+                error_type="rate_limit_error",
+                code="ai_agent_queue_full",
+            )
         return _error_response(
             str(exc),
             500,