Surface model provider errors as run errors

Extend the existing model-error boundary beyond rate limits so known Bailian provider failures are emitted as AG-UI RUN_ERROR with original message text and optional diagnostic metadata.

Constraint: Aone 83711637 asks model-side stream errors to surface as run errors instead of text messages.

Rejected: status-code-only text promotion | generic assistant text can mention HTTP/status codes and must not terminate the stream.

Rejected: broad provider error framework | the current PR only needs a compact SDK-side classifier for known Bailian model errors.

Confidence: high

Scope-risk: moderate

Directive: Keep raw-text model-error patterns narrow; add negative tests whenever expanding provider text matching.

Tested: uv run ruff check changed files; uv run pytest tests/unittests/server tests/unittests/integration => 781 passed, 2 skipped; uv run mypy --config-file mypy.ini . => 377 source files clean; independent code-reviewer APPROVE; architect WATCH with no blockers after heuristic tightening.

Change-Id: Ibf1cbd9505b7064cbebd804383710fd7333455cd
Not-tested: GitHub CI and MR review-comment readback are pending because push/comment touch shared remote state.
Signed-off-by: congxiao.wxx <congxiao.wxx@alibaba-inc.com>

Author: congxiao-wxx

agentrun/integration/langgraph/agent_converter.py

@@ -30,10 +30,7 @@
 from typing import Any, Dict, Iterator, List, Optional, Union
 
 from agentrun.server.model import AgentResult, EventType
-from agentrun.utils.error_utils import (
-    build_error_event_data,
-    is_rate_limited_error,
-)
+from agentrun.utils.error_utils import build_error_event_data, is_model_error
 from agentrun.utils.log import logger
 
 # 需要从工具输入中过滤掉的内部字段（LangGraph/MCP 注入的运行时对象）
@@ -961,7 +958,7 @@ def _convert_astream_events_event(
                     fallback_code="LLM_ERROR",
                     fallback_message=(
                         error_message
-                        if is_rate_limited_error(error)
+                        if is_model_error(error)
                         else f"LLM error: {error_message}"
                     ),
                 ),

agentrun/server/agui_protocol.py

@@ -53,7 +53,14 @@
 # ============================================================================
 
 DEFAULT_PREFIX = "/ag-ui/agent"
-RUN_ERROR_EXTRA_FIELDS = ("retryable", "retryAfterMs", "traceId")
+RUN_ERROR_EXTRA_FIELDS = (
+    "retryable",
+    "retryAfterMs",
+    "traceId",
+    "requestId",
+    "statusCode",
+    "providerCode",
+)
 
 
 @dataclass

agentrun/server/invoker.py

@@ -24,17 +24,15 @@
 )
 import uuid
 
-from agentrun.utils.error_utils import (
-    build_error_event_data,
-    is_rate_limited_error,
-)
+from agentrun.utils.error_utils import build_error_event_data, is_model_error
+from agentrun.utils.reasoning import get_reasoning_content
+
 from .model import AgentEvent, AgentRequest, EventType
 from .protocol import (
     AsyncInvokeAgentHandler,
     InvokeAgentHandler,
     SyncInvokeAgentHandler,
 )
-from agentrun.utils.reasoning import get_reasoning_content
 
 
 class AgentInvoker:
@@ -343,7 +341,7 @@ def _wrap_model_chunk(self, item: Any) -> List[AgentEvent]:
         return events
 
     def _wrap_text(self, text: str) -> AgentEvent:
-        if is_rate_limited_error(text):
+        if is_model_error(text):
             return AgentEvent(
                 event=EventType.ERROR,
                 data=build_error_event_data(