Fixed OpenAI o models (#239)

app/MindWork AI Studio/Provider/BaseProvider.cs

@@ -1,3 +1,5 @@
+using System.Net;
+
 using AIStudio.Chat;
 
 using RustService = AIStudio.Tools.RustService;
@@ -103,6 +105,13 @@ public abstract class BaseProvider : IProvider, ISecretId
                 break;
             }
             
+            if(nextResponse.StatusCode is HttpStatusCode.BadRequest)
+            {
+                this.logger.LogError($"Failed request with status code {nextResponse.StatusCode} (message = '{nextResponse.ReasonPhrase}').");
+                errorMessage = nextResponse.ReasonPhrase;
+                break;
+            }
+
             errorMessage = nextResponse.ReasonPhrase;
             var timeSeconds = Math.Pow(RETRY_DELAY_SECONDS, retry + 1);
             if(timeSeconds > 90)
@@ -112,7 +121,7 @@ public abstract class BaseProvider : IProvider, ISecretId
             await Task.Delay(TimeSpan.FromSeconds(timeSeconds), token);
         }
         
-        if(retry >= MAX_RETRIES)
+        if(retry >= MAX_RETRIES || !string.IsNullOrWhiteSpace(errorMessage))
             return new HttpRateLimitedStreamResult(false, true, errorMessage ?? $"Failed after {MAX_RETRIES} retries; no provider message available", response);
         
         return new HttpRateLimitedStreamResult(true, false, string.Empty, response);

app/MindWork AI Studio/Provider/OpenAI/ProviderOpenAI.cs

@@ -33,10 +33,35 @@ public sealed class ProviderOpenAI(ILogger logger) : BaseProvider("https://api.o
         if(!requestedSecret.Success)
             yield break;
         
+        // Unfortunately, OpenAI changed the name of the system prompt based on the model.
+        // All models that start with "o" (the omni aka reasoning models) and all GPT4o models
+        // have the system prompt named "developer". All other models have the system prompt
+        // named "system". We need to check this to get the correct system prompt.
+        //
+        // To complicate it even more: The early versions of reasoning models, which are released
+        // before the 17th of December 2024, have no system prompt at all. We need to check this
+        // as well.
+        
+        // Apply the basic rule first:
+        var systemPromptRole = chatModel.Id.StartsWith('o') || chatModel.Id.Contains("4o") ? "developer" : "system";
+        
+        // Check if the model is an early version of the reasoning models:
+        systemPromptRole = chatModel.Id switch
+        {
+            "o1-mini" => "user",
+            "o1-mini-2024-09-12" => "user",
+            "o1-preview" => "user",
+            "o1-preview-2024-09-12" => "user",
+            
+            _ => systemPromptRole,
+        };
+        
+        this.logger.LogInformation($"Using the system prompt role '{systemPromptRole}' for model '{chatModel.Id}'.");
+
         // Prepare the system prompt:
         var systemPrompt = new Message
         {
-            Role = "system",
+            Role = systemPromptRole,
             Content = chatThread.SystemPrompt,
         };
         
@@ -55,7 +80,7 @@ public sealed class ProviderOpenAI(ILogger logger) : BaseProvider("https://api.o
                     ChatRole.USER => "user",
                     ChatRole.AI => "assistant",
                     ChatRole.AGENT => "assistant",
-                    ChatRole.SYSTEM => "system",
+                    ChatRole.SYSTEM => systemPromptRole,
 
                     _ => "user",
                 },
@@ -71,7 +96,6 @@ public sealed class ProviderOpenAI(ILogger logger) : BaseProvider("https://api.o
             
             // Right now, we only support streaming completions:
             Stream = true,
-            FrequencyPenalty = 0f,
         }, JSON_SERIALIZER_OPTIONS);
 
         async Task<HttpRequestMessage> RequestBuilder()

app/MindWork AI Studio/wwwroot/changelog/v0.9.23.md

@@ -3,6 +3,7 @@
 - Improved provider requests by handling rate limits by retrying requests. 
 - Improved the creation of the "the bias of the day" workspace; create that workspace only when the bias of the day feature is used.
 - Improved the save operation of settings by using a temporary file to avoid data loss in rare cases.
+- Fixed OpenAI `o` (aka omni, aka reasoning) models. The early preview versions (released before 17th December 2024) could not use any system prompts —- we translated the system prompts to be user prompts. Final versions of the OpenAI `o` models can now use system prompts, by they are named `developer` instead of `system`.
 - Fixed layout issues when selecting `other` items (e.g., programming languages).
 - Fixed a bug about the bias of the day workspace when the workspace component was hidden.
 - Upgraded dependencies to the latest versions.