Fixed reasoning tokens inflating the exact token count

app/MindWork AI Studio/Chat/ChatThread.cs

@@ -416,22 +416,25 @@ public sealed record ChatThread
     /// is a lot of machinery for a number which corrects itself one answer later.
     /// </remarks>
     /// <param name="model">The model the next request would go to.</param>
-    /// <returns>What the provider reported, or unknown when no report describes this conversation.</returns>
-    public TokenUsage ReportedUsageFor(Model model)
+    /// <returns>
+    /// What the provider reported, together with the answer which followed it, or
+    /// ReportedHistory.UNKNOWN when no report describes this conversation.
+    /// </returns>
+    public ReportedHistory ReportedHistoryFor(Model model)
     {
         if (this.Blocks.Count is 0)
-            return TokenUsage.UNKNOWN;
+            return ReportedHistory.UNKNOWN;
 
-        if (this.Blocks[^1].Content is not ContentText { IsStreaming: false, ReportedUsage: { } reported })
-            return TokenUsage.UNKNOWN;
+        if (this.Blocks[^1].Content is not ContentText { IsStreaming: false, ReportedUsage: { } reported } answer)
+            return ReportedHistory.UNKNOWN;
 
         if (reported.BlockCount != this.Blocks.Count)
-            return TokenUsage.UNKNOWN;
+            return ReportedHistory.UNKNOWN;
 
         if (!string.Equals(reported.ModelId, model.Id, StringComparison.Ordinal))
-            return TokenUsage.UNKNOWN;
+            return ReportedHistory.UNKNOWN;
 
-        return reported.ToTokenUsage();
+        return ReportedHistory.Of(reported.ToTokenUsage(), answer.Text);
     }
 
     private static void DeleteManagedAttachments(ContentBlock block)

app/MindWork AI Studio/Chat/ContentText.cs

@@ -62,7 +62,7 @@ public sealed class ContentText : IContent
     /// answer written before this was recorded, and at every provider which reports nothing.
     ///
     /// Having a report is not the same as the report still being true. Whether it still describes
-    /// the conversation is decided by ChatThread.ReportedUsageFor, which looks at the thread around
+    /// the conversation is decided by ChatThread.ReportedHistoryFor, which looks at the thread around
     /// the answer, not just at the answer.
     /// </remarks>
     public ReportedTokenUsage? ReportedUsage { get; set; }
@@ -132,7 +132,6 @@ public sealed class ContentText : IContent
         this.ReportedUsage = new()
         {
             PromptTokens = usage.PromptTokens,
-            CompletionTokens = usage.CompletionTokens,
             ModelId = modelId,
             BlockCount = blockCount,
         };

app/MindWork AI Studio/Chat/ConversationTokens.cs

@@ -61,8 +61,11 @@ public readonly record struct ConversationTokens
     /// </summary>
     /// <remarks>
     /// True once a provider has stated what a request of this conversation cost, which makes
-    /// everything but the draft an exact number. It says nothing about the draft, which stays an
-    /// estimate either way -- nobody has charged for that one yet.
+    /// everything up to the last answer an exact number. That answer is counted by this app, since
+    /// the provider's number for it includes reasoning which no request carries; next to the rest
+    /// of the conversation, its share of the error is small enough to still call the history exact.
+    /// It says nothing about the draft, which stays an estimate either way -- nobody has charged
+    /// for that one yet.
     /// </remarks>
     public bool HistoryIsReported { get; init; }
 

app/MindWork AI Studio/Chat/ReportedHistory.cs

@@ -0,0 +1,59 @@
+using AIStudio.Provider;
+
+namespace AIStudio.Chat;
+
+/// <summary>
+/// What a conversation carries into its next request, as far as a provider has stated it.
+/// </summary>
+/// <remarks>
+/// Two parts, and only one of them is the provider's statement. PromptTokens is what the provider
+/// counted for the request behind the last answer: the system prompt, the tools, and every message
+/// up to the question. The answer travels in the next request as well, though not in the shape the
+/// provider charged for: its completion included the model's reasoning and whatever the model wrote
+/// between think tags. So the answer comes along as the text it will be sent as, and is counted the
+/// same way every other text is.
+///
+/// That text is the answer alone. Reasoning never belongs to it, whether it was thrown away or kept
+/// to be read next to the answer: it is there for a person, and no request carries it. An answer
+/// which consists of reasoning only has no text at all, is not sent, and adds nothing to the prompt.
+///
+/// What is estimated that way is one answer, next to a prompt which holds the whole conversation
+/// before it. The error is the tokenizer's error on that one answer, not on the chat.
+/// </remarks>
+public sealed record ReportedHistory
+{
+    /// <summary>
+    /// The history of a conversation no report describes.
+    /// </summary>
+    public static readonly ReportedHistory UNKNOWN = new();
+
+    /// <summary>
+    /// Whether a report describes the conversation. When false, nothing else here means anything.
+    /// </summary>
+    public bool IsKnown { get; private init; }
+
+    /// <summary>
+    /// What the provider counted for the request behind the last answer.
+    /// </summary>
+    public int PromptTokens { get; private init; }
+
+    /// <summary>
+    /// The text of the last answer, as the next request will carry it, without any reasoning.
+    /// </summary>
+    public string LastAnswer { get; private init; } = string.Empty;
+
+    /// <summary>
+    /// States what a report says about the conversation.
+    /// </summary>
+    /// <param name="usage">What the provider reported for the request behind the last answer.</param>
+    /// <param name="lastAnswer">The text of that answer.</param>
+    /// <returns>The history, or UNKNOWN when the usage states nothing.</returns>
+    public static ReportedHistory Of(TokenUsage usage, string lastAnswer) => usage.IsKnown
+        ? new()
+        {
+            IsKnown = true,
+            PromptTokens = usage.PromptTokens,
+            LastAnswer = lastAnswer,
+        }
+        : UNKNOWN;
+}

app/MindWork AI Studio/Chat/ReportedTokenUsage.cs

@@ -6,10 +6,11 @@ namespace AIStudio.Chat;
 /// What a provider said the request behind one answer cost, as it is stored on that answer.
 /// </summary>
 /// <remarks>
-/// The numbers and the model they were charged for travel as one value. Each of them is meaningless
-/// without the others, and loose fields on the answer could be set, cleared, or copied apart.
+/// The number, the model it was charged for, and the conversation it was counted on travel as one
+/// value. Each of them is meaningless without the others, and loose fields on the answer could be
+/// set, cleared, or copied apart.
 ///
-/// Plain numbers rather than a TokenUsage: that type only ever comes out of its own factory, which
+/// A plain number rather than a TokenUsage: that type only ever comes out of its own factory, which
 /// is what keeps an impossible usage from existing, while a stored value has to be readable back by
 /// the serializer. ToTokenUsage is the way back, and it treats a chat file somebody edited by hand
 /// the same way the reading side treats a provider's JSON.
@@ -22,12 +23,7 @@ public sealed record ReportedTokenUsage
     public int PromptTokens { get; init; }
 
     /// <summary>
-    /// What the model wrote in answer.
-    /// </summary>
-    public int CompletionTokens { get; init; }
-
-    /// <summary>
-    /// Which model the numbers were charged for.
+    /// Which model the number was charged for.
     /// </summary>
     /// <remarks>
     /// A token count belongs to the tokenizer which produced it. Switch the model of a chat, and
@@ -52,8 +48,8 @@ public sealed record ReportedTokenUsage
     public int BlockCount { get; init; }
 
     /// <summary>
-    /// States the stored numbers as a usage again.
+    /// States the stored number as a usage again.
     /// </summary>
-    /// <returns>The usage, or TokenUsage.UNKNOWN when the stored numbers state nothing usable.</returns>
-    public TokenUsage ToTokenUsage() => TokenUsage.OfReported(this.PromptTokens, this.CompletionTokens);
+    /// <returns>The usage, or TokenUsage.UNKNOWN when the stored number states nothing usable.</returns>
+    public TokenUsage ToTokenUsage() => TokenUsage.OfReported(this.PromptTokens);
 }