.NET: Add context window size compaction strategy for harness (#5304)

* Add context window size compaction strategy for harness * Apply suggestions from code review Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com> * Address PR comments --------- Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
2026-06-16 21:04:09 +08:00 · 2026-04-16 16:44:45 +01:00
parent 9d89353818
commit d0ac1d83bc
6 changed files with 468 additions and 20 deletions
@@ -19,7 +19,9 @@ public static class HarnessConsole
    /// <param name="agent">The agent to interact with.</param>
    /// <param name="title">The title displayed in the console header.</param>
    /// <param name="userPrompt">A short prompt to the user, displayed below the title.</param>
-    public static async Task RunAgentAsync(AIAgent agent, string title, string userPrompt)
+    /// <param name="maxContextWindowTokens">Optional max context window size in tokens. When set, usage is displayed as a percentage.</param>
+    /// <param name="maxOutputTokens">Optional max output tokens. Used with <paramref name="maxContextWindowTokens"/> to show input/output budget breakdown.</param>
+    public static async Task RunAgentAsync(AIAgent agent, string title, string userPrompt, int? maxContextWindowTokens = null, int? maxOutputTokens = null)
    {
        var todoProvider = agent.GetService<TodoProvider>();
        var modeProvider = agent.GetService<AgentModeProvider>();
@@ -46,7 +48,7 @@ public static class HarnessConsole
            }
            else
            {
-                await StreamAgentResponseAsync(agent, session, modeProvider, userInput);
+                await StreamAgentResponseAsync(agent, session, modeProvider, userInput, maxContextWindowTokens, maxOutputTokens);
            }

            WritePrompt(modeProvider, session);
@@ -57,7 +59,7 @@ public static class HarnessConsole
        System.Console.WriteLine("Goodbye!");
    }

-    private static async Task StreamAgentResponseAsync(AIAgent agent, AgentSession session, AgentModeProvider? modeProvider, string userInput)
+    private static async Task StreamAgentResponseAsync(AIAgent agent, AgentSession session, AgentModeProvider? modeProvider, string userInput, int? maxContextWindowTokens, int? maxOutputTokens)
    {
        string mode = modeProvider?.GetMode(session) ?? "unknown";
        System.Console.ForegroundColor = GetModeColor(mode);
@@ -106,6 +108,37 @@ public static class HarnessConsole

                        System.Console.ForegroundColor = GetModeColor(mode);
                    }
+                    else if (content is TextReasoningContent reasoning && !string.IsNullOrEmpty(reasoning.Text))
+                    {
+                        await spinner.StopAsync();
+
+                        if (!hasTextOutput)
+                        {
+                            System.Console.Write("\n");
+                            hasTextOutput = true;
+                            hasReceivedAnyText = true;
+                        }
+
+                        System.Console.ForegroundColor = ConsoleColor.DarkMagenta;
+                        System.Console.Write(reasoning.Text);
+                        System.Console.ForegroundColor = GetModeColor(mode);
+                    }
+                    else if (content is UsageContent usage)
+                    {
+                        await spinner.StopAsync();
+                        System.Console.ForegroundColor = ConsoleColor.DarkGray;
+                        System.Console.Write("\n\n  📊 Tokens");
+                        if (usage.Details is not null)
+                        {
+                            WriteUsageBreakdown(usage.Details, maxContextWindowTokens, maxOutputTokens);
+                        }
+                        else
+                        {
+                            System.Console.Write(" —");
+                        }
+                        System.Console.ForegroundColor = GetModeColor(mode);
+                        hasTextOutput = false;
+                    }
                }

                if (string.IsNullOrEmpty(update.Text))
@@ -136,7 +169,7 @@ public static class HarnessConsole
        {
            await spinner.StopAsync();
            System.Console.ForegroundColor = ConsoleColor.Red;
-            System.Console.Write($"\n  ❌ Stream error: {ex.GetType().Name}: {ex.Message}");
+            System.Console.Write($"\n  ❌ Stream error: {ex.GetType().Name}:\n{ex}");
        }

        await spinner.StopAsync();
@@ -190,7 +223,7 @@ public static class HarnessConsole
        catch (ArgumentException ex)
        {
            System.Console.ForegroundColor = ConsoleColor.Red;
-            System.Console.WriteLine($"\n  {ex.Message}\n");
+            System.Console.WriteLine($"\n  {ex}\n");
            System.Console.ResetColor();
        }
    }
@@ -237,6 +270,38 @@ public static class HarnessConsole
        System.Console.WriteLine();
    }

+    private static void WriteUsageBreakdown(UsageDetails details, int? maxContextWindowTokens, int? maxOutputTokens)
+    {
+        int? inputBudget = (maxContextWindowTokens is not null && maxOutputTokens is not null)
+            ? maxContextWindowTokens.Value - maxOutputTokens.Value
+            : null;
+
+        System.Console.Write(" — input: ");
+        WriteTokenCount(details.InputTokenCount, inputBudget);
+
+        System.Console.Write(" | output: ");
+        WriteTokenCount(details.OutputTokenCount, maxOutputTokens);
+
+        System.Console.Write(" | total: ");
+        WriteTokenCount(details.TotalTokenCount, maxContextWindowTokens);
+    }
+
+    private static void WriteTokenCount(long? count, int? budget)
+    {
+        if (count is null)
+        {
+            System.Console.Write("—");
+            return;
+        }
+
+        System.Console.Write($"{count.Value:N0}");
+        if (budget is not null && budget.Value > 0)
+        {
+            double pct = (double)count.Value / budget.Value * 100;
+            System.Console.Write($"/{budget.Value:N0} ({pct:F1}%)");
+        }
+    }
+
    private static ConsoleColor GetModeColor(string mode) => mode switch
    {
        AgentModeProvider.PlanMode => ConsoleColor.Cyan,
@@ -13,7 +13,7 @@
  </ItemGroup>

  <ItemGroup>
-    <ProjectReference Include="..\..\..\..\src\Microsoft.Agents.AI.Foundry\Microsoft.Agents.AI.Foundry.csproj" />
+    <ProjectReference Include="..\..\..\..\src\Microsoft.Agents.AI.OpenAI\Microsoft.Agents.AI.OpenAI.csproj" />
    <ProjectReference Include="..\Harness_Shared_Console\Harness_Shared_Console.csproj" />
  </ItemGroup>

@@ -11,28 +11,40 @@
 //   exit    — End the session.

 #pragma warning disable OPENAI001 // Suppress experimental API warnings for Responses API usage.
+#pragma warning disable MAAI001  // Suppress experimental API warnings for Agents AI experiments.

-using Azure.AI.Projects;
+using System.ClientModel.Primitives;
 using Azure.Identity;
 using Harness.Shared.Console;
 using Microsoft.Agents.AI;
-using Microsoft.Agents.AI.Foundry;
+using Microsoft.Agents.AI.Compaction;
 using Microsoft.Extensions.AI;
+using OpenAI;
+using OpenAI.Responses;
 using SampleApp;

-var endpoint = Environment.GetEnvironmentVariable("AZURE_AI_PROJECT_ENDPOINT") ?? throw new InvalidOperationException("AZURE_AI_PROJECT_ENDPOINT is not set.");
+var endpoint = Environment.GetEnvironmentVariable("AZURE_FOUNDRY_OPENAI_ENDPOINT") ?? throw new InvalidOperationException("AZURE_FOUNDRY_OPENAI_ENDPOINT is not set.");
 var deploymentName = Environment.GetEnvironmentVariable("AZURE_AI_MODEL_DEPLOYMENT_NAME") ?? "gpt-5.4";

-// Create the Azure AI Project client and get an IChatClient with stored output disabled
+// Create a compaction strategy based on the model's context window.
+// gpt-5.4: 1,050,000 token context window, 128,000 max output tokens.
+// Defaults: tool result eviction at 50% of input budget, truncation at 80%.
+var compactionStrategy = new ContextWindowCompactionStrategy(
+    maxContextWindowTokens: 1_050_000,
+    maxOutputTokens: 128_000);
+
+// Create an OpenAIClient that communicates with the Foundry responses service and get an IChatClient with stored output disabled
 // so that chat history is managed locally by the agent framework.
 // WARNING: DefaultAzureCredential is convenient for development but requires careful consideration in production.
 // In production, consider using a specific credential (e.g., ManagedIdentityCredential) to avoid
 // latency issues, unintended credential probing, and potential security risks from fallback mechanisms.
-var aiProjectClient = new AIProjectClient(new Uri(endpoint), new DefaultAzureCredential());
-IChatClient chatClient = aiProjectClient
-    .GetProjectOpenAIClient()
-    .GetProjectResponsesClient()
-    .AsIChatClient(deploymentName);
+OpenAIClientOptions clientOptions = new() { Endpoint = new Uri(endpoint) };
+IChatClient chatClient = new OpenAIClient(new BearerTokenPolicy(new DefaultAzureCredential(), "https://ai.azure.com/.default"), clientOptions)
+    .GetResponsesClient()
+    .AsIChatClientWithStoredOutputDisabled(deploymentName)
+    .AsBuilder()
+    .UseAIContextProviders(new CompactionProvider(compactionStrategy))
+    .Build();

 // Create web browsing tools for downloading and converting HTML pages to markdown.
 var webBrowsingTools = new WebBrowsingTools();
@@ -79,6 +91,10 @@ AIAgent agent = new ChatClientAgent(
        Name = "ResearchAgent",
        Description = "A research assistant that plans and executes research tasks.",
        AIContextProviders = [new TodoProvider(), new AgentModeProvider()],
+        ChatHistoryProvider = new InMemoryChatHistoryProvider(new InMemoryChatHistoryProviderOptions
+        {
+            ChatReducer = compactionStrategy.AsChatReducer(),
+        }),
        ChatOptions = new ChatOptions
        {
            // Set a high token limit for long research tasks with many tool calls and long outputs.
@@ -86,9 +102,9 @@ AIAgent agent = new ChatClientAgent(
            MaxOutputTokens = 128_000,
            Instructions = instructions,
            Reasoning = new() { Effort = ReasoningEffort.High },
-            Tools = [FoundryAITool.CreateWebSearchTool(), .. webBrowsingTools.Tools],
+            Tools = [ResponseTool.CreateWebSearchTool().AsAITool(), .. webBrowsingTools.Tools],
        },
    });

 // Run the interactive console session using the shared HarnessConsole helper.
-await HarnessConsole.RunAgentAsync(agent, title: "Research Assistant", userPrompt: "Enter a research topic to get started.");
+await HarnessConsole.RunAgentAsync(agent, title: "Research Assistant", userPrompt: "Enter a research topic to get started.", maxContextWindowTokens: 1_050_000, maxOutputTokens: 128_000);
@@ -5,7 +5,7 @@ This sample demonstrates how to use a `ChatClientAgent` with the Harness `AICont
 Key features showcased:

 - **ChatClientAgent** — configured directly with Harness providers for planning and task management
- **Web Search** — the agent can search the web for current information via `FoundryAITool.CreateWebSearchTool()`
+- **Web Search** — the agent can search the web for current information via `ResponseTool.CreateWebSearchTool()`
 - **TodoProvider** — the agent creates and manages a todo list to track research questions
 - **AgentModeProvider** — the agent switches between "plan" mode (breaking down the topic) and "execute" mode (answering each research question)
 - **Interactive conversation** — you can review the agent's plan, provide feedback, and approve before execution begins
@@ -25,8 +25,8 @@ Before running this sample, ensure you have:
 Set the following environment variables:

 ```bash
-# Required: Your Azure AI Foundry project endpoint
-export AZURE_AI_PROJECT_ENDPOINT="https://your-project.services.ai.azure.com/api/projects/your-project-name"
+# Required: Your Azure AI Foundry OpenAI endpoint
+export AZURE_FOUNDRY_OPENAI_ENDPOINT="https://your-project.services.ai.azure.com/openai/v1/"

 # Optional: Model deployment name (defaults to gpt-5.4)
 export AZURE_AI_MODEL_DEPLOYMENT_NAME="gpt-5.4"