feat: Nvidia API direkt-Anbindung mit Thinking-Mode Support

Neues Server-Setting NvidiaApiKey – wenn gesetzt, wird direkt
integrate.api.nvidia.com verwendet statt g4f.space Proxy.
Thinking-Modelle (kimi-k2, qwq, deepseek-r1) erhalten automatisch
chat_template_kwargs und reasoning-Output wird als klappbare
Denkprozess-Bubble im Chat angezeigt.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Services/OpenAiService.cs

@@ -127,7 +127,11 @@ public class OpenAiService
     /// </summary>
     public async Task RefreshModelsFromApis()
     {
-        _nvidiaModels       = await FetchModels("https://g4f.space/api/nvidia/models", "Nvidia");
+        var nvidiaApiKey = _settings.Get("NvidiaApiKey");
+        if (!string.IsNullOrEmpty(nvidiaApiKey))
+            _nvidiaModels = await FetchModels("https://integrate.api.nvidia.com/v1/models", "Nvidia", nvidiaApiKey);
+        else
+            _nvidiaModels = await FetchModels("https://g4f.space/api/nvidia/models", "Nvidia");
         _pollinationsModels = await FetchModels("https://text.pollinations.ai/models", "Pollinations");
 
         var geminiApiKey = _settings.Get("GeminiApiKey");
@@ -303,6 +307,9 @@ public class OpenAiService
                     ["tool_choice"] = "auto"
                 };
 
+                if (provider == "nvidia" && IsNvidiaThinkingModel(modelId))
+                    payload["chat_template_kwargs"] = new { thinking = true };
+
                 var jsonPayload = JsonSerializer.Serialize(payload);
                 var httpContent = new StringContent(jsonPayload, Encoding.UTF8, "application/json");
 
@@ -333,10 +340,12 @@ public class OpenAiService
                 using var reader = new System.IO.StreamReader(stream);
 
                 var fullText = new StringBuilder();
+                var thinkingText = new StringBuilder();
                 // Tool-Call Akkumulatoren: id → (name, arguments)
                 var toolCallAccumulators = new Dictionary<int, (string Id, string Name, StringBuilder Args, string ThoughtSignature)>();
                 string thoughtSignature = null;
                 var textStarted = false;
+                var thinkingStarted = false;
                 var stopReason = "end_turn";
 
                 string line;
@@ -367,12 +376,34 @@ public class OpenAiService
 
                         if (!choice.TryGetProperty("delta", out var delta)) continue;
 
+                        // Reasoning/Thinking-Delta (Nvidia Thinking-Modelle)
+                        if (delta.TryGetProperty("reasoning", out var reasoningProp) && reasoningProp.ValueKind == JsonValueKind.String)
+                        {
+                            var chunk = reasoningProp.GetString();
+                            if (!string.IsNullOrEmpty(chunk))
+                            {
+                                if (!thinkingStarted)
+                                {
+                                    await SseHelper.TryWriteSse(response, "thinking_start", new { }, sseCt);
+                                    thinkingStarted = true;
+                                }
+                                thinkingText.Append(chunk);
+                                await SseHelper.TryWriteSse(response, "thinking_delta", new { text = chunk }, sseCt);
+                            }
+                        }
+
                         // Text-Delta
                         if (delta.TryGetProperty("content", out var contentProp) && contentProp.ValueKind == JsonValueKind.String)
                         {
                             var chunk = contentProp.GetString();
                             if (!string.IsNullOrEmpty(chunk))
                             {
+                                // Thinking-Block beenden falls noch offen
+                                if (thinkingStarted)
+                                {
+                                    await SseHelper.TryWriteSse(response, "thinking_end", new { }, sseCt);
+                                    thinkingStarted = false;
+                                }
                                 if (!textStarted)
                                 {
                                     await SseHelper.TryWriteSse(response, "text_start", new { index = 0 }, sseCt);
@@ -383,8 +414,6 @@ public class OpenAiService
                             }
                         }
 
-
-
                         // Tool-Call Deltas akkumulieren
                         if (delta.TryGetProperty("tool_calls", out var toolCallsProp))
                         {
@@ -465,10 +494,26 @@ public class OpenAiService
                     }
                 }
 
+                // Thinking-End senden falls noch offen
+                if (thinkingStarted)
+                    await SseHelper.TryWriteSse(response, "thinking_end", new { }, sseCt);
+
                 // Text-End senden falls Text gestreamt wurde
                 if (textStarted)
                     await SseHelper.TryWriteSse(response, "text_end", new { }, sseCt);
 
+                // Thinking-Text in DB speichern
+                var finalThinking = thinkingText.ToString();
+                if (!string.IsNullOrEmpty(finalThinking))
+                {
+                    await _db.SaveMessage(new MessageRecord
+                    {
+                        ChainId = chainId, Iteration = iteration,
+                        Role = "assistant", Type = "thinking",
+                        Content = finalThinking
+                    });
+                }
+
                 // Text in DB speichern
                 var finalText = fullText.ToString();
                 if (!string.IsNullOrEmpty(finalText))
@@ -669,6 +714,9 @@ public class OpenAiService
                     ["tool_choice"] = "auto"
                 };
 
+                if (provider == "nvidia" && IsNvidiaThinkingModel(modelId))
+                    payload["chat_template_kwargs"] = new { thinking = true };
+
                 var jsonPayload = JsonSerializer.Serialize(payload);
                 var httpContent = new StringContent(jsonPayload, Encoding.UTF8, "application/json");
 
@@ -709,6 +757,21 @@ public class OpenAiService
                 var message = choice.GetProperty("message");
                 var finishReason = choice.TryGetProperty("finish_reason", out var frProp) ? frProp.GetString() ?? "" : "";
 
+                // Reasoning/Thinking extrahieren (Nvidia Thinking-Modelle)
+                if (message.TryGetProperty("reasoning", out var reasoningProp) && reasoningProp.ValueKind == JsonValueKind.String)
+                {
+                    var reasoning = reasoningProp.GetString() ?? "";
+                    if (!string.IsNullOrEmpty(reasoning))
+                    {
+                        await _db.SaveMessage(new MessageRecord
+                        {
+                            ChainId = ctx.ChainId, Iteration = iterations,
+                            Role = "assistant", Type = "thinking",
+                            Content = reasoning
+                        });
+                    }
+                }
+
                 // Text extrahieren
                 if (message.TryGetProperty("content", out var contentProp) && contentProp.ValueKind == JsonValueKind.String)
                 {
@@ -909,12 +972,30 @@ public class OpenAiService
         if (!string.IsNullOrEmpty(apiKeyOverride))
             providerConfig = (providerConfig.BaseUrl, apiKeyOverride);
 
+        if (provider == "nvidia" && string.IsNullOrEmpty(providerConfig.ApiKey))
+        {
+            var nvidiaKey = _settings.Get("NvidiaApiKey");
+            if (!string.IsNullOrEmpty(nvidiaKey))
+                providerConfig = ("https://integrate.api.nvidia.com/v1", nvidiaKey);
+        }
+
         if (provider == "gemini" && string.IsNullOrEmpty(providerConfig.ApiKey))
             providerConfig = (providerConfig.BaseUrl, _settings.Get("GeminiApiKey"));
 
         return (provider, modelId, providerConfig);
     }
 
+    /// <summary>
+    /// Prüft ob ein Nvidia-Modell Thinking-Mode unterstützt (chat_template_kwargs).
+    /// </summary>
+    private static bool IsNvidiaThinkingModel(string modelId)
+    {
+        // Modelle mit Extended-Thinking-Support bei Nvidia
+        return modelId.Contains("kimi-k2", StringComparison.OrdinalIgnoreCase)
+            || modelId.Contains("qwq", StringComparison.OrdinalIgnoreCase)
+            || modelId.Contains("deepseek-r1", StringComparison.OrdinalIgnoreCase);
+    }
+
     /// <summary>
     /// Konvertiert Anthropic-Format Messages in OpenAI-kompatibles Format.
     /// </summary>