Keep tool results readable for the model

app/MindWork AI Studio/Tools/ToolCallingSystem/ToolExecutionResult.cs

@@ -1,3 +1,5 @@
+using System.Text.Encodings.Web;
+using System.Text.Json;
 using System.Text.Json.Nodes;
 
 using AIStudio.Provider;
@@ -7,6 +9,18 @@ namespace AIStudio.Tools.ToolCallingSystem;
 
 public sealed class ToolExecutionResult
 {
+    /// <summary>
+    /// How a JSON result is written for the model.
+    /// </summary>
+    /// <remarks>
+    /// The result goes into the request as a string, and the request is serialized once more on its
+    /// way to the provider, so the model reads whatever this escapes as the escape itself. The
+    /// default encoder escapes every character outside ASCII and those HTML treats specially, for
+    /// JSON embedded in a web page, which this never is: a German document would reach the model
+    /// with every umlaut as six characters. The relaxed encoder escapes only what JSON requires.
+    /// </remarks>
+    private static readonly JsonSerializerOptions MODEL_CONTENT_OPTIONS = new() { Encoder = JavaScriptEncoder.UnsafeRelaxedJsonEscaping };
+
     public string? TextContent { get; init; }
 
     public JsonNode? JsonContent { get; init; }
@@ -29,7 +43,7 @@ public sealed class ToolExecutionResult
     public string ToModelContent()
     {
         if (this.JsonContent is not null)
-            return this.JsonContent.ToJsonString();
+            return this.JsonContent.ToJsonString(MODEL_CONTENT_OPTIONS);
 
         return this.TextContent ?? string.Empty;
     }

app/Tests/Tools/ToolCalling/ToolExecutionResultTests.cs

@@ -0,0 +1,39 @@
+using System.Text.Json;
+using System.Text.Json.Nodes;
+
+using AIStudio.Tools.ToolCallingSystem;
+
+namespace AIStudio.Tests.Tools.ToolCalling;
+
+/// <summary>
+/// Checks what a model reads of a tool result.
+/// </summary>
+/// <remarks>
+/// The result goes into the request as a string, and the request is serialized once more on its way
+/// to the provider. Whatever the first serialization escapes therefore reaches the model as the
+/// escape itself: a German document would arrive with every umlaut spelled out as six characters,
+/// and a piece of code with every angle bracket. That costs tokens, the budget of all tool results
+/// counts it, and a model quoting a name from it may quote the escape.
+/// </remarks>
+[TestFixture]
+public sealed class ToolExecutionResultTests
+{
+    [TestCase("Größe der Übersicht")]
+    [TestCase("if (a < b && c > d) return 'x';")]
+    [TestCase("日本語のテキスト")]
+    public void TextReachesTheModelAsWritten(string text)
+    {
+        var result = new ToolExecutionResult { JsonContent = new JsonObject { ["text_content"] = text } };
+        Assert.That(result.ToModelContent(), Does.Contain(text));
+    }
+
+    [Test]
+    public void TheResultStaysValidJson()
+    {
+        const string TEXT = "A \"quoted\" line\nand a backslash \\ at its end.";
+        var result = new ToolExecutionResult { JsonContent = new JsonObject { ["text_content"] = TEXT } };
+
+        var readBack = JsonSerializer.Deserialize<JsonObject>(result.ToModelContent());
+        Assert.That(readBack?["text_content"]?.GetValue<string>(), Is.EqualTo(TEXT), "What JSON itself has to escape, it still escapes.");
+    }
+}