.NET: [BREAKING] Add auto-approval rules (heuristics) to ToolApprovalAgent (#6335)

* Add support for approving tools via heuristic rules * Address PR comments * Address PR comments * Apply suggestion from @SergeyMenshykh Co-authored-by: SergeyMenshykh <68852919+SergeyMenshykh@users.noreply.github.com> --------- Co-authored-by: SergeyMenshykh <68852919+SergeyMenshykh@users.noreply.github.com>
2026-06-16 21:04:09 +08:00 · 2026-06-05 18:43:07 +01:00
parent ab8ba8fc61
commit 6bd2cfec03
8 changed files with 497 additions and 48 deletions
@@ -644,6 +644,51 @@ public class HarnessAgentTests
        Assert.Null(agent.GetService<ToolApprovalAgent>());
    }

+    /// <summary>
+    /// Verify that ToolApprovalAgentOptions auto-approval rules are passed through and actually used.
+    /// </summary>
+    [Fact]
+    public async Task ToolApproval_AutoApprovalRulesAreAppliedAsync()
+    {
+        // Arrange — inner client returns an approval request on first call, then final response on second.
+        var callCount = 0;
+        var approvalRequest = new ToolApprovalRequestContent("req1", new FunctionCallContent("call1", "ReadTool"));
+
+        var mockClient = new Mock<IChatClient>();
+        mockClient
+            .Setup(c => c.GetResponseAsync(
+                It.IsAny<IEnumerable<ChatMessage>>(),
+                It.IsAny<ChatOptions>(),
+                It.IsAny<CancellationToken>()))
+            .ReturnsAsync(() =>
+            {
+                callCount++;
+                if (callCount == 1)
+                {
+                    return new ChatResponse(new ChatMessage(ChatRole.Assistant, [approvalRequest]));
+                }
+
+                return new ChatResponse(new ChatMessage(ChatRole.Assistant, "Done"));
+            });
+
+        var options = CreateAllDisabledOptions();
+        options.DisableToolApproval = false;
+        options.ToolApprovalAgentOptions = new ToolApprovalAgentOptions
+        {
+            AutoApprovalRules = [fcc => new ValueTask<bool>(fcc.Name == "ReadTool")]
+        };
+
+        var agent = new HarnessAgent(mockClient.Object, TestMaxContextWindowTokens, TestMaxOutputTokens, options);
+        var session = await agent.CreateSessionAsync();
+
+        // Act
+        var response = await agent.RunAsync([new ChatMessage(ChatRole.User, "Hi")], session);
+
+        // Assert — the auto-approval rule approved the request, so we get "Done" (not an approval request)
+        Assert.Equal(2, callCount);
+        Assert.Equal("Done", response.Text);
+    }
+
    #endregion

    #region Feature: OpenTelemetry
@@ -59,15 +59,15 @@ public class ToolApprovalAgentBuilderExtensionsTests
    /// Verify that UseToolApproval with custom JsonSerializerOptions works correctly.
    /// </summary>
    [Fact]
-    public void UseToolApproval_WithCustomJsonSerializerOptions_ReturnsToolApprovalAgent()
+    public void UseToolApproval_WithCustomOptions_ReturnsToolApprovalAgent()
    {
        // Arrange
        var mockAgent = new Mock<AIAgent>();
        var builder = new AIAgentBuilder(mockAgent.Object);
-        var options = new JsonSerializerOptions();
+        var options = new ToolApprovalAgentOptions { JsonSerializerOptions = new JsonSerializerOptions() };

        // Act
-        var result = builder.UseToolApproval(jsonSerializerOptions: options).Build();
+        var result = builder.UseToolApproval(options: options).Build();

        // Assert
        Assert.IsType<ToolApprovalAgent>(result);
@@ -47,14 +47,14 @@ public class ToolApprovalAgentTests
    }

    /// <summary>
-    /// Verify that constructor accepts custom JsonSerializerOptions.
+    /// Verify that constructor accepts custom options.
    /// </summary>
    [Fact]
-    public void Constructor_CustomJsonSerializerOptions_CreatesInstanceAsync()
+    public void Constructor_CustomOptions_CreatesInstance()
    {
        // Arrange
        var innerAgent = new Mock<AIAgent>().Object;
-        var options = new JsonSerializerOptions();
+        var options = new ToolApprovalAgentOptions { JsonSerializerOptions = new JsonSerializerOptions() };

        // Act
        var agent = new ToolApprovalAgent(innerAgent, options);
@@ -1535,4 +1535,311 @@ public class ToolApprovalAgentTests
    }

    #endregion
+
+    #region Auto-Approval Rules (Heuristics)
+
+    /// <summary>
+    /// Verify that an auto-approval rule can approve a function call that would otherwise need user approval.
+    /// </summary>
+    [Fact]
+    public async Task RunAsync_AutoApprovalRule_ApprovesMatchingToolAsync()
+    {
+        // Arrange
+        var session = new ChatClientAgentSession();
+        var approvalRequest = new ToolApprovalRequestContent("req1", new FunctionCallContent("call1", "ReadTool"));
+
+        // Inner agent: first call returns approval request, second returns final response.
+        var callCount = 0;
+        var innerAgent = new Mock<AIAgent>();
+        innerAgent
+            .Protected()
+            .Setup<Task<AgentResponse>>("RunCoreAsync",
+                ItExpr.IsAny<IEnumerable<ChatMessage>>(),
+                ItExpr.IsAny<AgentSession?>(),
+                ItExpr.IsAny<AgentRunOptions?>(),
+                ItExpr.IsAny<CancellationToken>())
+            .ReturnsAsync(() =>
+            {
+                callCount++;
+                if (callCount == 1)
+                {
+                    return new AgentResponse([new ChatMessage(ChatRole.Assistant, [approvalRequest])]);
+                }
+
+                return new AgentResponse([new ChatMessage(ChatRole.Assistant, "Done")]);
+            });
+
+        var options = new ToolApprovalAgentOptions
+        {
+            AutoApprovalRules = [fcc => new ValueTask<bool>(fcc.Name == "ReadTool")]
+        };
+        var agent = new ToolApprovalAgent(innerAgent.Object, options);
+
+        // Act
+        var response = await agent.RunAsync(
+            [new ChatMessage(ChatRole.User, "Hi")],
+            session);
+
+        // Assert — the approval request was auto-approved, inner agent called twice
+        Assert.Equal(2, callCount);
+        Assert.Equal("Done", response.Text);
+    }
+
+    /// <summary>
+    /// Verify that when auto-approval rule does not match, request is surfaced to the caller.
+    /// </summary>
+    [Fact]
+    public async Task RunAsync_AutoApprovalRule_DoesNotMatchSurfacesToCallerAsync()
+    {
+        // Arrange
+        var session = new ChatClientAgentSession();
+        var approvalRequest = new ToolApprovalRequestContent("req1", new FunctionCallContent("call1", "DangerousTool"));
+
+        var innerAgent = CreateMockAgent(new AgentResponse([new ChatMessage(ChatRole.Assistant, [approvalRequest])]));
+
+        var options = new ToolApprovalAgentOptions
+        {
+            AutoApprovalRules = [fcc => new ValueTask<bool>(fcc.Name == "ReadTool")]  // Only approves ReadTool
+        };
+        var agent = new ToolApprovalAgent(innerAgent.Object, options);
+
+        // Act
+        var response = await agent.RunAsync(
+            [new ChatMessage(ChatRole.User, "Hi")],
+            session);
+
+        // Assert — request surfaced to caller since heuristic doesn't match
+        var requests = response.Messages.SelectMany(m => m.Contents).OfType<ToolApprovalRequestContent>().ToList();
+        Assert.Single(requests);
+        Assert.Equal("DangerousTool", ((FunctionCallContent)requests[0].ToolCall).Name);
+    }
+
+    /// <summary>
+    /// Verify that multiple auto-approval rules are evaluated in order; first match wins.
+    /// </summary>
+    [Fact]
+    public async Task RunAsync_MultipleAutoApprovalRules_FirstMatchWinsAsync()
+    {
+        // Arrange
+        var session = new ChatClientAgentSession();
+        var approvalRequest = new ToolApprovalRequestContent("req1", new FunctionCallContent("call1", "SpecialTool"));
+
+        var callCount = 0;
+        var innerAgent = new Mock<AIAgent>();
+        innerAgent
+            .Protected()
+            .Setup<Task<AgentResponse>>("RunCoreAsync",
+                ItExpr.IsAny<IEnumerable<ChatMessage>>(),
+                ItExpr.IsAny<AgentSession?>(),
+                ItExpr.IsAny<AgentRunOptions?>(),
+                ItExpr.IsAny<CancellationToken>())
+            .ReturnsAsync(() =>
+            {
+                callCount++;
+                if (callCount == 1)
+                {
+                    return new AgentResponse([new ChatMessage(ChatRole.Assistant, [approvalRequest])]);
+                }
+
+                return new AgentResponse([new ChatMessage(ChatRole.Assistant, "Done")]);
+            });
+
+        var rule1Called = false;
+        var rule2Called = false;
+        var options = new ToolApprovalAgentOptions
+        {
+            AutoApprovalRules =
+            [
+                fcc => { rule1Called = true; return new ValueTask<bool>(fcc.Name == "SpecialTool"); },
+                fcc => { rule2Called = true; return new ValueTask<bool>(true); }  // Should not be reached
+            ]
+        };
+        var agent = new ToolApprovalAgent(innerAgent.Object, options);
+
+        // Act
+        await agent.RunAsync([new ChatMessage(ChatRole.User, "Hi")], session);
+
+        // Assert — first rule matched, second was never called
+        Assert.True(rule1Called);
+        Assert.False(rule2Called);
+    }
+
+    /// <summary>
+    /// Verify that standing rules are evaluated before auto-approval rules.
+    /// </summary>
+    [Fact]
+    public async Task RunAsync_StandingRuleTakesPrecedenceOverAutoApprovalRuleAsync()
+    {
+        // Arrange
+        var session = new ChatClientAgentSession();
+        var approvalRequest = new ToolApprovalRequestContent("req1", new FunctionCallContent("call1", "MyTool"));
+
+        var callCount = 0;
+        var innerAgent = new Mock<AIAgent>();
+        innerAgent
+            .Protected()
+            .Setup<Task<AgentResponse>>("RunCoreAsync",
+                ItExpr.IsAny<IEnumerable<ChatMessage>>(),
+                ItExpr.IsAny<AgentSession?>(),
+                ItExpr.IsAny<AgentRunOptions?>(),
+                ItExpr.IsAny<CancellationToken>())
+            .ReturnsAsync(() =>
+            {
+                callCount++;
+                if (callCount <= 2)
+                {
+                    return new AgentResponse([new ChatMessage(ChatRole.Assistant, [approvalRequest])]);
+                }
+
+                return new AgentResponse([new ChatMessage(ChatRole.Assistant, "Done")]);
+            });
+
+        var heuristicCalled = false;
+        var options = new ToolApprovalAgentOptions
+        {
+            AutoApprovalRules = [fcc => { heuristicCalled = true; return new ValueTask<bool>(true); }]
+        };
+        var agent = new ToolApprovalAgent(innerAgent.Object, options);
+
+        // Call 1: heuristic should be called (no standing rule yet)
+        var response1 = await agent.RunAsync([new ChatMessage(ChatRole.User, "Hi")], session);
+        Assert.True(heuristicCalled);
+        Assert.Equal("Done", response1.Text);
+
+        // Now establish a standing rule by sending AlwaysApprove
+        heuristicCalled = false;
+        callCount = 0;
+        var alwaysApprove = new AlwaysApproveToolApprovalResponseContent(
+            approvalRequest.CreateResponse(approved: true),
+            alwaysApproveTool: true,
+            alwaysApproveToolWithArguments: false);
+
+        // Call 2: standing rule should match first, heuristic should NOT be called
+        var response2 = await agent.RunAsync(
+            [new ChatMessage(ChatRole.User, [alwaysApprove])],
+            session);
+        Assert.False(heuristicCalled);
+        Assert.Equal("Done", response2.Text);
+    }
+
+    /// <summary>
+    /// Verify that when a batch contains a mix of heuristic-approved and standing-rule-approved
+    /// requests, the collected approval responses preserve the original request order rather than
+    /// being grouped by approval kind.
+    /// </summary>
+    [Fact]
+    public async Task RunAsync_MixedAutoApprovals_PreserveOriginalOrderAsync()
+    {
+        // Arrange
+        var session = new ChatClientAgentSession();
+
+        // Batch ordering: first request is approved by a heuristic, second by a standing rule.
+        var heuristicRequest = new ToolApprovalRequestContent("reqA", new FunctionCallContent("callA", "HeuristicTool"));
+        var standingRequest = new ToolApprovalRequestContent("reqB", new FunctionCallContent("callB", "StandingTool"));
+
+        var batchResponse = new AgentResponse([new ChatMessage(ChatRole.Assistant, [heuristicRequest, standingRequest])]);
+        var finalResponse = new AgentResponse([new ChatMessage(ChatRole.Assistant, "Done")]);
+
+        var callCount = 0;
+        List<ChatMessage>? secondCallMessages = null;
+        var innerAgent = new Mock<AIAgent>();
+        innerAgent
+            .Protected()
+            .Setup<Task<AgentResponse>>("RunCoreAsync",
+                ItExpr.IsAny<IEnumerable<ChatMessage>>(),
+                ItExpr.IsAny<AgentSession?>(),
+                ItExpr.IsAny<AgentRunOptions?>(),
+                ItExpr.IsAny<CancellationToken>())
+            .Callback<IEnumerable<ChatMessage>, AgentSession?, AgentRunOptions?, CancellationToken>((msgs, _, _, _) =>
+            {
+                callCount++;
+                if (callCount == 2)
+                {
+                    secondCallMessages = msgs.ToList();
+                }
+            })
+            .ReturnsAsync(() => callCount == 1 ? batchResponse : finalResponse);
+
+        var options = new ToolApprovalAgentOptions
+        {
+            AutoApprovalRules = [fcc => new ValueTask<bool>(fcc.Name == "HeuristicTool")]
+        };
+        var agent = new ToolApprovalAgent(innerAgent.Object, options);
+
+        // Establish a standing rule for "StandingTool" via an AlwaysApprove response in the same call.
+        var alwaysApprove = standingRequest.CreateAlwaysApproveToolResponse("User said always");
+
+        // Act — both requests auto-approve (heuristic + standing rule), so the inner agent is re-invoked.
+        var response = await agent.RunAsync(
+            [new ChatMessage(ChatRole.User, [alwaysApprove])],
+            session);
+
+        // Assert — inner agent re-called and final response returned.
+        Assert.Equal(2, callCount);
+        Assert.Equal("Done", response.Text);
+
+        // The injected approval responses must preserve the original request order: reqA before reqB,
+        // even though reqA was approved by a heuristic and reqB by a standing rule.
+        Assert.NotNull(secondCallMessages);
+        var injected = secondCallMessages!
+            .SelectMany(m => m.Contents)
+            .OfType<ToolApprovalResponseContent>()
+            .Where(r => r.RequestId is "reqA" or "reqB")
+            .ToList();
+        Assert.Equal(2, injected.Count);
+        Assert.Equal("reqA", injected[0].RequestId);
+        Assert.Equal("reqB", injected[1].RequestId);
+        Assert.All(injected, r => Assert.True(r.Approved));
+    }
+
+    /// <summary>
+    /// Verify that auto-approval rules work in the streaming path.
+    /// </summary>
+    [Fact]
+    public async Task RunStreamingAsync_AutoApprovalRule_ApprovesMatchingToolAsync()
+    {
+        // Arrange
+        var session = new ChatClientAgentSession();
+        var approvalRequest = new ToolApprovalRequestContent("req1", new FunctionCallContent("call1", "ReadTool"));
+
+        var callCount = 0;
+        var innerAgent = new Mock<AIAgent>();
+        innerAgent
+            .Protected()
+            .Setup<IAsyncEnumerable<AgentResponseUpdate>>("RunCoreStreamingAsync",
+                ItExpr.IsAny<IEnumerable<ChatMessage>>(),
+                ItExpr.IsAny<AgentSession?>(),
+                ItExpr.IsAny<AgentRunOptions?>(),
+                ItExpr.IsAny<CancellationToken>())
+            .Returns(() =>
+            {
+                callCount++;
+                if (callCount == 1)
+                {
+                    return ToAsyncEnumerableAsync([new AgentResponseUpdate(ChatRole.Assistant, [approvalRequest])]);
+                }
+
+                return ToAsyncEnumerableAsync([new AgentResponseUpdate(ChatRole.Assistant, "Done")]);
+            });
+
+        var options = new ToolApprovalAgentOptions
+        {
+            AutoApprovalRules = [fcc => new ValueTask<bool>(fcc.Name == "ReadTool")]
+        };
+        var agent = new ToolApprovalAgent(innerAgent.Object, options);
+
+        // Act
+        var updates = new List<AgentResponseUpdate>();
+        await foreach (var update in agent.RunStreamingAsync([new ChatMessage(ChatRole.User, "Hi")], session))
+        {
+            updates.Add(update);
+        }
+
+        // Assert — the approval request was auto-approved, inner agent streamed twice
+        Assert.Equal(2, callCount);
+        Assert.Single(updates);
+        Assert.Equal("Done", updates[0].Text);
+    }
+
+    #endregion
 }