feat(provider): make outer OpenAI retry count configurable

astrbot/core/config/default.py

@@ -104,6 +104,7 @@
         "default_provider_id": "",
         "fallback_chat_models": [],
         "request_max_retries": 5,
+        "provider_error_retries": 1,
         "default_image_caption_provider_id": "",
         "image_caption_prompt": "Please describe the image using Chinese.",
         "provider_pool": ["*"],  # "*" 表示使用所有可用的提供者
@@ -2803,6 +2804,9 @@
                     "request_max_retries": {
                         "type": "int",
                     },
+                    "provider_error_retries": {
+                        "type": "int",
+                    },
                     "wake_prefix": {
                         "type": "string",
                     },

astrbot/core/provider/sources/openai_source.py

@@ -398,6 +398,24 @@ def __init__(self, provider_config, provider_settings) -> None:
 
         self.reasoning_key = "reasoning_content"
 
+    def _provider_error_retries(self) -> int:
+        """Return retry attempts for provider-level recovery paths.
+
+        This outer retry loop handles payload mutation and key rotation (for
+        example context trimming, tool removal, image fallback, or switching to
+        another configured API key). Transport/status-code retries are handled
+        separately by ``request_max_retries`` in ``retry_provider_request``.
+        Keeping this value configurable prevents nested retry loops from
+        multiplying latency for proxy/aggregator providers that already perform
+        their own upstream retry and fallback.
+        """
+        raw = self.provider_settings.get("provider_error_retries", 1)
+        try:
+            retries = int(raw)
+        except (TypeError, ValueError):
+            retries = 1
+        return max(1, retries)
+
     def _ollama_disable_thinking_enabled(self) -> bool:
         value = self.provider_config.get("ollama_disable_thinking", False)
         if isinstance(value, str):
@@ -1188,7 +1206,7 @@ async def text_chat(
             payloads["tool_choice"] = tool_choice
 
         llm_response = None
-        max_retries = 10
+        max_retries = self._provider_error_retries()
         available_api_keys = self.api_keys.copy()
         chosen_key = random.choice(available_api_keys)
         image_fallback_used = False
@@ -1264,7 +1282,7 @@ async def text_chat_stream(
         if func_tool and not func_tool.empty():
             payloads["tool_choice"] = tool_choice
 
-        max_retries = 10
+        max_retries = self._provider_error_retries()
         available_api_keys = self.api_keys.copy()
         chosen_key = random.choice(available_api_keys)
         image_fallback_used = False

tests/test_openai_source.py

@@ -120,6 +120,22 @@ def fake_import(name, globals=None, locals=None, fromlist=(), level=0):
     assert captured["httpx_module"] is openai_source_module.httpx
 
 
+def test_provider_error_retries_defaults_and_coerces_values():
+    provider = ProviderOpenAIOfficial.__new__(ProviderOpenAIOfficial)
+
+    provider.provider_settings = {}
+    assert provider._provider_error_retries() == 1
+
+    provider.provider_settings = {"provider_error_retries": "3"}
+    assert provider._provider_error_retries() == 3
+
+    provider.provider_settings = {"provider_error_retries": 0}
+    assert provider._provider_error_retries() == 1
+
+    provider.provider_settings = {"provider_error_retries": "invalid"}
+    assert provider._provider_error_retries() == 1
+
+
 @pytest.mark.asyncio
 async def test_get_models_retries_transient_request_error(monkeypatch):
     monkeypatch.setattr(request_retry, "REQUEST_RETRY_WAIT_MIN_S", 0)