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

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

app/Tests/Provider/ChatCompletionUsageTests.cs

@@ -0,0 +1,129 @@
+using System.Text.Json;
+
+using AIStudio.Provider;
+using AIStudio.Provider.OpenAI;
+
+namespace AIStudio.Tests.Provider;
+
+/// <summary>
+/// Checks that what a provider says a request cost is read off the stream, and asked for.
+/// </summary>
+/// <remarks>
+/// Both halves matter and neither is visible from the other: an OpenAI-compatible provider says
+/// nothing about the cost of a streamed request unless the request asks for it, and the line it
+/// then sends carries no content, so the reading side has to look for it apart from the text.
+/// Get either half wrong and the app silently keeps estimating, which looks exactly like a
+/// provider which reports nothing.
+/// </remarks>
+[TestFixture]
+public sealed class ChatCompletionUsageTests
+{
+    /// <summary>
+    /// The last line of a streamed answer at a provider which was asked for the usage.
+    /// </summary>
+    private const string USAGE_LINE =
+        """
+        {"id":"chatcmpl-1","object":"chat.completion.chunk","created":1,"model":"gpt-5","choices":[],"usage":{"prompt_tokens":1200,"completion_tokens":345,"total_tokens":1545}}
+        """;
+
+    /// <summary>
+    /// An ordinary line carrying a piece of the answer.
+    /// </summary>
+    private const string CONTENT_LINE =
+        """
+        {"id":"chatcmpl-1","object":"chat.completion.chunk","created":1,"model":"gpt-5","choices":[{"index":0,"delta":{"content":"Hi"}}]}
+        """;
+
+    [Test]
+    public void TheFinalLineStatesWhatTheRequestCost()
+    {
+        var line = JsonSerializer.Deserialize<ChatCompletionDeltaStreamLine>(USAGE_LINE, ProviderJsonOptions.OPTIONS);
+
+        Assert.Multiple(() =>
+        {
+            Assert.That(line!.GetUsage().IsKnown, Is.True);
+            Assert.That(line.GetUsage().PromptTokens, Is.EqualTo(1200));
+
+            //
+            // The line which carries the usage carries no answer, which is why it has to be read
+            // before the content check drops it:
+            //
+            Assert.That(line.ContainsContent(), Is.False);
+        });
+    }
+
+    [Test]
+    public void ALineOfTheAnswerStatesNoCost()
+    {
+        var line = JsonSerializer.Deserialize<ChatCompletionDeltaStreamLine>(CONTENT_LINE, ProviderJsonOptions.OPTIONS);
+
+        Assert.Multiple(() =>
+        {
+            Assert.That(line!.GetUsage().IsKnown, Is.False);
+            Assert.That(line.ContainsContent(), Is.True);
+        });
+    }
+
+    [Test]
+    public void AStreamedRequestAsksForTheUsage()
+    {
+        var request = new ChatCompletionAPIRequest("gpt-5", [], true);
+        var json = JsonSerializer.Serialize(request, ProviderJsonOptions.OPTIONS);
+
+        Assert.That(json, Does.Contain("""
+                                       "stream_options":{"include_usage":true}
+                                       """));
+    }
+
+    [Test]
+    public void ARequestWhichIsNotStreamedDoesNot()
+    {
+        var request = new ChatCompletionAPIRequest("gpt-5", [], false);
+        var json = JsonSerializer.Serialize(request, ProviderJsonOptions.OPTIONS);
+
+        Assert.That(json, Does.Not.Contain("stream_options"));
+    }
+
+    /// <summary>
+    /// A provider which sends the block but fills in nothing usable states nothing.
+    /// </summary>
+    /// <remarks>
+    /// Several OpenAI-compatible servers send an empty or zeroed usage block on every line while
+    /// streaming and the real numbers only at the end. Reading a zero as a fact would replace an
+    /// estimate with a statement that the conversation costs nothing.
+    /// </remarks>
+    [Test]
+    public void AnEmptyUsageBlockStatesNothing()
+    {
+        var line = JsonSerializer.Deserialize<ChatCompletionDeltaStreamLine>(
+            """
+            {"id":"chatcmpl-1","object":"chat.completion.chunk","created":1,"model":"gpt-5","choices":[],"usage":{"prompt_tokens":0,"completion_tokens":0}}
+            """, ProviderJsonOptions.OPTIONS);
+
+        Assert.That(line!.GetUsage().IsKnown, Is.False);
+    }
+
+    /// <summary>
+    /// The line a real LM Studio server sends, taken off the wire.
+    /// </summary>
+    /// <remarks>
+    /// It carries far more than we read, and it shows why the prompt is all we take: 102 of the 116
+    /// completion tokens are the model's reasoning, which the next request never carries. Counting
+    /// the completion would have put the history at 133 tokens, when what travels on is the prompt
+    /// and an answer of a handful of tokens.
+    /// </remarks>
+    [Test]
+    public void ARealServerLineIsRead()
+    {
+        var line = JsonSerializer.Deserialize<ChatCompletionDeltaStreamLine>(
+            """
+            {"id":"chatcmpl-xb8mn282eff3tiu46xz8t3","object":"chat.completion.chunk","created":1789917744,"model":"google/gemma-4-12b-qat","system_fingerprint":"google/gemma-4-12b-qat","choices":[],"usage":{"prompt_tokens":17,"completion_tokens":116,"total_tokens":133,"completion_tokens_details":{"reasoning_tokens":102}}}
+            """, ProviderJsonOptions.OPTIONS);
+
+        Assert.Multiple(() =>
+        {
+            Assert.That(line!.GetUsage().IsKnown, Is.True);
+            Assert.That(line.GetUsage().PromptTokens, Is.EqualTo(17));
+        });
+    }
+}

app/Tests/Provider/ToolCalling/ChatCompletionToolCallAccumulatorTests.cs

@@ -183,7 +183,38 @@ public sealed class ChatCompletionToolCallAccumulatorTests
         
         Assert.That(part.Sources.Select(x => x.URL), Is.EqualTo(new[] { "https://example.org/" }), "Whatever the provider announced on that line reaches the user with it.");
     }
-    
+
+    [Test]
+    public void TheLineWithoutChoicesStatesWhatTheRequestCost()
+    {
+        var accumulator = new ChatCompletionToolCallAccumulator();
+        var part = accumulator.Process(Event("""{"choices":[],"usage":{"prompt_tokens":1200,"completion_tokens":345,"total_tokens":1545}}"""));
+
+        Assert.Multiple(() =>
+        {
+            Assert.That(part.Usage.IsKnown, Is.True, "The last line of the stream has no choices, and it must not be dropped for that.");
+            Assert.That(part.Usage.PromptTokens, Is.EqualTo(1200));
+            Assert.That(part.HasContent, Is.False, "It has nothing to show, though.");
+        });
+    }
+
+    [Test]
+    public void AUsageNextToTheLastTextIsReadAsWell()
+    {
+        //
+        // Some providers put the usage on the line which carries the last piece of the answer
+        // rather than on a line of its own.
+        //
+        var accumulator = new ChatCompletionToolCallAccumulator();
+        var part = accumulator.Process(Event("""{"choices":[{"index":0,"delta":{"content":"Bye"}}],"usage":{"prompt_tokens":1200,"completion_tokens":2}}"""));
+
+        Assert.Multiple(() =>
+        {
+            Assert.That(part.TextDelta, Is.EqualTo("Bye"));
+            Assert.That(part.Usage.PromptTokens, Is.EqualTo(1200));
+        });
+    }
+
     private static ChatCompletionResponseMessage? Read(params string[] data)
     {
         var accumulator = new ChatCompletionToolCallAccumulator();

app/Tests/Provider/ToolCalling/ChatCompletionToolCallingAdapterTests.cs

@@ -0,0 +1,169 @@
+using System.Runtime.CompilerServices;
+using System.Text.Json;
+
+using AIStudio.Provider;
+using AIStudio.Provider.OpenAI;
+
+using Microsoft.Extensions.Logging.Abstractions;
+
+namespace AIStudio.Tests.Provider.ToolCalling;
+
+/// <summary>
+/// Checks what a round of a tool calling conversation asks for, and what it passes on.
+/// </summary>
+/// <remarks>
+/// Every round of a tool conversation is a request of its own, and every one of them reports what
+/// it cost. Only the first one describes what the next question will be sent after: every later
+/// round carries the tool calls and their results on top, none of which is sent again once the
+/// answer stands. Passing on the last report instead would put the chat at the size of everything
+/// the tools returned, which is the one number a person watching their context window must not
+/// see as exact.
+///
+/// What a round asks for is one tool call at a time, wherever the provider lets it ask: a provider
+/// which rejects the question fails the whole request, so it is not asked at all.
+/// </remarks>
+[TestFixture]
+public sealed class ChatCompletionToolCallingAdapterTests
+{
+    private const string FIRST_ROUND_USAGE = """{"choices":[],"usage":{"prompt_tokens":1200,"completion_tokens":20}}""";
+
+    private const string SECOND_ROUND_USAGE = """{"choices":[],"usage":{"prompt_tokens":9800,"completion_tokens":150}}""";
+
+    [Test]
+    public async Task OnlyTheFirstRoundPassesOnWhatItsRequestCost()
+    {
+        var adapter = Adapter(
+        [
+            """{"choices":[{"index":0,"delta":{"tool_calls":[{"index":0,"id":"call_1","type":"function","function":{"name":"web_search","arguments":"{\"query\":\"weather\"}"}}]}}]}""",
+            FIRST_ROUND_USAGE,
+            "[DONE]",
+        ],
+        [
+            """{"choices":[{"index":0,"delta":{"content":"It is sunny."}}]}""",
+            SECOND_ROUND_USAGE,
+            "[DONE]",
+        ]);
+
+        var firstRound = await Usages(adapter);
+
+        //
+        // What the loop does between two rounds: the model's turn and the tool's result become part
+        // of the next request.
+        //
+        adapter.RecordAssistantTurn();
+        adapter.RecordToolResult("call_1", "Sunny, 24 degrees.");
+
+        var secondRound = await Usages(adapter);
+
+        Assert.Multiple(() =>
+        {
+            Assert.That(firstRound, Is.EqualTo(new[] { 1200 }), "The first round's prompt is the conversation up to the question.");
+            Assert.That(secondRound, Is.Empty, "The second round's prompt holds the tool result as well, which the next question is not sent with.");
+        });
+    }
+
+    [Test]
+    public async Task ARoundWithoutToolCallsPassesItOnAsWell()
+    {
+        //
+        // Offering tools does not mean the model uses them. Then the first round is the only one,
+        // and its report is as good as the one of a request which offered none.
+        //
+        var adapter = Adapter(
+        [
+            """{"choices":[{"index":0,"delta":{"content":"Hello."}}]}""",
+            FIRST_ROUND_USAGE,
+            "[DONE]",
+        ]);
+
+        Assert.That(await Usages(adapter), Is.EqualTo(new[] { 1200 }));
+    }
+
+    [Test]
+    public async Task ARoundWhichOffersToolsAsksForOneCallAtATime()
+    {
+        Assert.That(await SentRequest(mayAskForSequentialToolCalls: true, includeTools: true), Does.Contain("\"parallel_tool_calls\":false"));
+    }
+
+    [Test]
+    public async Task AProviderWhichRejectsTheQuestionIsNotAskedIt()
+    {
+        //
+        // Hugging Face answers the question with a bad request. Its models may then ask for several
+        // calls at once, which the loop works through one by one anyway.
+        //
+        Assert.That(await SentRequest(mayAskForSequentialToolCalls: false, includeTools: true), Does.Not.Contain("parallel_tool_calls"));
+    }
+
+    [Test]
+    public async Task ARoundWithoutToolsDoesNotAskAboutToolCalls()
+    {
+        Assert.That(await SentRequest(mayAskForSequentialToolCalls: true, includeTools: false), Does.Not.Contain("parallel_tool_calls"));
+    }
+
+    /// <summary>
+    /// Runs one round and returns the request it sent, as it goes over the wire.
+    /// </summary>
+    private static async Task<string> SentRequest(bool mayAskForSequentialToolCalls, bool includeTools)
+    {
+        ChatCompletionAPIRequest? sent = null;
+        var adapter = Adapter(mayAskForSequentialToolCalls, request => sent = request, ["[DONE]"]);
+        await foreach (var _ in adapter.ExecuteRoundAsync(null, includeTools))
+        {
+        }
+
+        return JsonSerializer.Serialize(sent, ProviderJsonOptions.OPTIONS);
+    }
+
+    /// <summary>
+    /// Runs the next round and returns the prompt of every usage it passed on.
+    /// </summary>
+    private static async Task<List<int>> Usages(ChatCompletionToolCallingAdapter<ChatCompletionAPIRequest> adapter)
+    {
+        var usages = new List<int>();
+        await foreach (var streamEvent in adapter.ExecuteRoundAsync(null, true))
+            if (streamEvent.Delta is { Usage.IsKnown: true } delta)
+                usages.Add(delta.Usage.PromptTokens);
+
+        return usages;
+    }
+
+    /// <summary>
+    /// Builds an adapter whose requests are answered by the given rounds, one after another.
+    /// </summary>
+    private static ChatCompletionToolCallingAdapter<ChatCompletionAPIRequest> Adapter(params string[][] rounds) => Adapter(true, _ => { }, rounds);
+
+    /// <summary>
+    /// Builds an adapter whose requests are answered by the given rounds, and which hands every
+    /// request it sends to the given observer.
+    /// </summary>
+    private static ChatCompletionToolCallingAdapter<ChatCompletionAPIRequest> Adapter(bool mayAskForSequentialToolCalls, Action<ChatCompletionAPIRequest> sent, params string[][] rounds)
+    {
+        var nextRound = 0;
+        return new(
+            (_, _, tools) => Task.FromResult(new ChatCompletionAPIRequest("model-a", [], true) { Tools = tools }),
+            new TextMessage("You are a helpful assistant.", "system"),
+            new Dictionary<string, object>(),
+            [],
+            mayAskForSequentialToolCalls,
+            [],
+            (request, token) =>
+            {
+                sent(request);
+                return Lines(rounds[nextRound++], token);
+            },
+            _ => [],
+            NullLogger.Instance);
+    }
+
+    private static async IAsyncEnumerable<ServerSentEvent> Lines(string[] data, [EnumeratorCancellation] CancellationToken token = default)
+    {
+        foreach (var line in data)
+        {
+            token.ThrowIfCancellationRequested();
+            yield return new ServerSentEvent($"data: {line}", line);
+        }
+
+        await Task.CompletedTask;
+    }
+}