Added the exact token count where the provider reports it (#989)

Co-authored-by: Thorsten Sommer <SommerEngineering@users.noreply.github.com>

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

@@ -16,7 +16,7 @@ namespace AIStudio.Provider.OpenAI;
 public sealed class ChatCompletionToolCallingAdapter<TRequest>(
     Func<TextMessage, IDictionary<string, object>, IList<object>?, Task<TRequest>> requestFactory,
     TextMessage systemPrompt, IDictionary<string, object> apiParameters,
-    IList<object> providerTools,
+    IList<object> providerTools, bool mayAskForSequentialToolCalls,
     IReadOnlyList<(ToolDefinition Definition, IToolImplementation Implementation)> runnableTools,
     Func<ChatCompletionAPIRequest, CancellationToken, IAsyncEnumerable<ServerSentEvent>> streamRequestAsync,
     Func<ServerSentEvent, IList<ISource>> readSources,
@@ -49,11 +49,23 @@ public sealed class ChatCompletionToolCallingAdapter<TRequest>(
             //
             // AI Studio runs tool calls one after another, so asking for parallel calls would
             // only produce work it then has to serialize anyway. Requests without tools omit the
-            // parameter because some providers reject it then.
+            // parameter because some providers reject it then. So does every request to a provider
+            // which rejects the parameter altogether: its models may then ask for several calls at
+            // once, and the loop works through them one by one, checking the limits per call.
             //
-            ParallelToolCalls = requestDtoBase.Tools is null ? null : false,
+            ParallelToolCalls = requestDtoBase.Tools is null || !mayAskForSequentialToolCalls ? null : false,
         };
 
+        //
+        // Only the first round passes on what its request cost. Its prompt is the conversation up
+        // to the question, which is exactly what the next question will be sent after. Every later
+        // round carries the tool calls and their results on top, and none of that is sent again
+        // once the answer stands -- a report of such a round would count a chat far larger than
+        // the one the next request carries. What the answer adds, all rounds of text together, is
+        // counted from its text afterwards, cf. ReportedHistory.
+        //
+        var passesOnUsage = this.internalMessages.Count is 0;
+
         //
         // The text goes out while it is being written; the tool calls are put back together
         // behind it, fragment by fragment.
@@ -62,8 +74,9 @@ public sealed class ChatCompletionToolCallingAdapter<TRequest>(
         await foreach (var serverSentEvent in streamRequestAsync(requestDto, token))
         {
             var part = accumulator.Process(serverSentEvent);
-            if (part.HasContent)
-                yield return ToolCallingStreamEvent.TextDelta(new ContentStreamChunk(part.TextDelta, part.Sources));
+            var usage = passesOnUsage ? part.Usage : TokenUsage.UNKNOWN;
+            if (part.HasContent || usage.IsKnown)
+                yield return ToolCallingStreamEvent.TextDelta(new ContentStreamChunk(part.TextDelta, part.Sources, Usage: usage));
         }
 
         var message = accumulator.Build();