meeting-assistant/MeetingAssistant.Tests/LiteLlmResponsesChatClientTests.cs

using MeetingAssistant.Summary;
using Microsoft.Extensions.AI;
using System.Net;
using System.Text;
using System.Text.Json;

namespace MeetingAssistant.Tests;

public sealed class LiteLlmResponsesChatClientTests
{
    [Fact]
    public async Task ClientAssemblesStreamedTextResponseWithMetadataAndUsage()
    {
        var handler = new SequencedHttpMessageHandler(
            new HttpResponseMessage(HttpStatusCode.OK)
            {
                Content = new StringContent(
                    """
                    data: {"type":"response.created","response":{"id":"resp_stream","created_at":1779147100,"model":"gpt-5.5","object":"response","output":[],"parallel_tool_calls":true,"status":"in_progress","store":false},"sequence_number":0}

                    data: {"type":"response.output_item.added","output_index":0,"item":{"id":"msg_stream","type":"message","status":"in_progress","content":[],"role":"assistant"},"sequence_number":1}

                    data: {"type":"response.content_part.added","item_id":"msg_stream","output_index":0,"content_index":0,"part":{"type":"output_text","annotations":[],"text":""},"sequence_number":2}

                    data: {"type":"response.output_text.delta","item_id":"msg_stream","output_index":0,"content_index":0,"delta":"Streamed OK","sequence_number":3}

                    data: {"type":"response.output_text.done","item_id":"msg_stream","output_index":0,"content_index":0,"text":"Streamed OK","sequence_number":4}

                    data: {"type":"response.content_part.done","item_id":"msg_stream","output_index":0,"content_index":0,"part":{"type":"output_text","annotations":[],"text":"Streamed OK"},"sequence_number":5}

                    data: {"type":"response.output_item.done","output_index":0,"item":{"id":"msg_stream","type":"message","status":"completed","content":[{"type":"output_text","annotations":[],"text":"Streamed OK"}],"role":"assistant"},"sequence_number":6}

                    data: {"type":"response.completed","response":{"id":"resp_stream","created_at":1779147100,"model":"gpt-5.5","object":"response","output":[],"parallel_tool_calls":true,"status":"completed","store":false,"usage":{"input_tokens":12,"output_tokens":3,"total_tokens":15}},"sequence_number":7}

                    data: [DONE]

                    """,
                    Encoding.UTF8,
                    "text/event-stream")
            });
        using var client = CreateClient(handler, reconnectionAttempts: 0);

        var response = await client.GetResponseAsync([new ChatMessage(ChatRole.User, "reply")]);

        Assert.Equal("Streamed OK", response.Text);
        Assert.Equal("resp_stream", response.ResponseId);
        Assert.Equal("gpt-5.5", response.ModelId);
        Assert.Equal(DateTimeOffset.FromUnixTimeSeconds(1779147100), response.CreatedAt);
        Assert.NotNull(response.Usage);
        Assert.Equal(12, response.Usage.InputTokenCount);
        Assert.Equal(3, response.Usage.OutputTokenCount);
        Assert.Equal(15, response.Usage.TotalTokenCount);
    }

    [Fact]
    public async Task ClientAssemblesStreamedFunctionCall()
    {
        var handler = new SequencedHttpMessageHandler(
            new HttpResponseMessage(HttpStatusCode.OK)
            {
                Content = new StringContent(
                    """
                    data: {"type":"response.output_item.done","output_index":0,"item":{"type":"function_call","call_id":"call_stream","name":"write_summary","arguments":"{\"markdown\":\"# Summary\\nDone\"}","status":"completed"}}

                    data: {"type":"response.completed","response":{"id":"resp_stream","model":"gpt-5.5","output":[]}}

                    data: [DONE]

                    """,
                    Encoding.UTF8,
                    "text/event-stream")
            });
        using var client = CreateClient(handler, reconnectionAttempts: 0);

        var response = await client.GetResponseAsync([new ChatMessage(ChatRole.User, "write summary")]);
        var call = Assert.IsType<FunctionCallContent>(Assert.Single(response.Messages[0].Contents));

        Assert.Equal("call_stream", call.CallId);
        Assert.Equal("write_summary", call.Name);
        Assert.NotNull(call.Arguments);
        Assert.Equal("# Summary\nDone", call.Arguments["markdown"]?.ToString());
    }

    [Fact]
    public async Task ClientUsesNonStreamingResponsesWhenConfigured()
    {
        var handler = new RecordingHttpMessageHandler(_ => new HttpResponseMessage(HttpStatusCode.OK)
        {
            Content = new StringContent(
                """
                {
                  "id": "resp_nonstream",
                  "created_at": 1779147100,
                  "model": "gpt-5.5",
                  "object": "response",
                  "output": [
                    {
                      "id": "msg_nonstream",
                      "type": "message",
                      "status": "completed",
                      "content": [
                        {
                          "type": "output_text",
                          "annotations": [],
                          "text": "Non-streamed OK"
                        }
                      ],
                      "role": "assistant"
                    }
                  ],
                  "parallel_tool_calls": true,
                  "status": "completed",
                  "store": false,
                  "usage": {
                    "input_tokens": 12,
                    "output_tokens": 4,
                    "total_tokens": 16
                  }
                }
                """,
                Encoding.UTF8,
                "application/json")
        });
        using var client = CreateClient(handler, reconnectionAttempts: 0, useStreaming: false);

        var response = await client.GetResponseAsync([new ChatMessage(ChatRole.User, "reply")]);

        Assert.Equal("Non-streamed OK", response.Text);
        Assert.Contains("\"stream\":false", Assert.Single(handler.RequestBodies));
    }

    [Fact]
    public async Task ClientSendsChatMessagesAsResponsesMessageItems()
    {
        var handler = new RecordingHttpMessageHandler(_ => CreateStreamedTextResponse("Done."));
        using var client = CreateClient(handler, reconnectionAttempts: 0);

        await client.GetResponseAsync(
        [
            new ChatMessage(ChatRole.User, "write summary"),
            new ChatMessage(ChatRole.Assistant, "I will inspect the meeting.")
        ]);

        using var request = JsonDocument.Parse(Assert.Single(handler.RequestBodies));
        var inputItems = request.RootElement.GetProperty("input").EnumerateArray().ToArray();
        Assert.Equal(2, inputItems.Length);
        Assert.All(inputItems, item => Assert.Equal("message", item.GetProperty("type").GetString()));
        Assert.Equal("user", inputItems[0].GetProperty("role").GetString());
        Assert.Equal(
            "input_text",
            Assert.Single(inputItems[0].GetProperty("content").EnumerateArray())
                .GetProperty("type")
                .GetString());
        Assert.Equal("assistant", inputItems[1].GetProperty("role").GetString());
        Assert.Equal(
            "output_text",
            Assert.Single(inputItems[1].GetProperty("content").EnumerateArray())
                .GetProperty("type")
                .GetString());
        Assert.DoesNotContain("\"type\":\"unknown\"", handler.RequestBodies[0]);
    }

    [Fact]
    public async Task AgentLoopReturnsMalformedFunctionArgumentsWithoutInvokingTool()
    {
        var responses = new Queue<HttpResponseMessage>(
        [
            new HttpResponseMessage(HttpStatusCode.OK)
            {
                Content = new StringContent(
                    """
                    data: {"type":"response.output_item.done","output_index":0,"item":{"type":"function_call","call_id":"call_invalid","name":"write_summary","arguments":"{not-json","status":"completed"}}

                    data: {"type":"response.completed","response":{"id":"resp_invalid","model":"gpt-5.5","output":[]}}

                    data: [DONE]

                    """,
                    Encoding.UTF8,
                    "text/event-stream")
            },
            CreateStreamedTextResponse("Recovered.")
        ]);
        var handler = new RecordingHttpMessageHandler(_ => responses.Dequeue());
        var toolInvoked = false;
        var tool = AIFunctionFactory.Create(
            (string markdown) =>
            {
                toolInvoked = true;
                return markdown;
            },
            "write_summary",
            "Writes a summary.");
        using var innerClient = CreateClient(handler, reconnectionAttempts: 0);
        using var functionClient = innerClient
            .AsBuilder()
            .UseFunctionInvocation(
                loggerFactory: null,
                client => client.FunctionInvoker = FunctionInvocationGuard.InvokeAsync)
            .Build();

        var response = await functionClient.GetResponseAsync(
            [new ChatMessage(ChatRole.User, "write summary")],
            new ChatOptions { Tools = [tool] });

        Assert.False(toolInvoked);
        Assert.Equal("Recovered.", response.Text);
        Assert.Equal(2, handler.RequestBodies.Count);
        Assert.Contains("invalid_tool_arguments", handler.RequestBodies[1]);
        Assert.Contains("call_invalid", handler.RequestBodies[1]);
    }

    [Fact]
    public async Task ClientReportsVisibleReasoningSummariesSeparatelyFromResponseText()
    {
        var handler = new SequencedHttpMessageHandler(
            CreateStreamedTextResponse(
                "Done.",
                "Checked the configured rules file.",
                "Prepared a targeted update."));
        var reasoningSummaries = new List<string>();
        using var client = CreateClient(
            handler,
            reconnectionAttempts: 0,
            reasoningSummaryChanged: reasoningSummaries.Add);

        var response = await client.GetResponseAsync([new ChatMessage(ChatRole.User, "update rules")]);

        Assert.Equal("Done.", response.Text);
        Assert.Equal(
            [
                "Checked the configured rules file.",
                "Prepared a targeted update."
            ],
            reasoningSummaries);
    }

    [Fact]
    public async Task ClientDoesNotFailResponseWhenReasoningSummaryCallbackFails()
    {
        var handler = new SequencedHttpMessageHandler(
            CreateStreamedTextResponse("Done.", "Checked the configured rules file."));
        using var client = CreateClient(
            handler,
            reconnectionAttempts: 0,
            reasoningSummaryChanged: _ => throw new InvalidOperationException("UI failed"));

        var response = await client.GetResponseAsync([new ChatMessage(ChatRole.User, "update rules")]);

        Assert.Equal("Done.", response.Text);
    }

    [Fact]
    public async Task ClientRetriesTransientServerFailure()
    {
        var handler = new SequencedHttpMessageHandler(
            new HttpResponseMessage(HttpStatusCode.InternalServerError)
            {
                Content = new StringContent("Internal Server Error")
            },
            CreateStreamedTextResponse("Done."));
        var retryCount = 0;
        using var client = CreateClient(handler, reconnectionAttempts: 1, retrying: () => retryCount++);

        var response = await client.GetResponseAsync([new ChatMessage(ChatRole.User, "write summary")]);

        Assert.Equal("Done.", response.Text);
        Assert.Equal(2, handler.RequestCount);
        Assert.Equal(1, retryCount);
    }

    [Fact]
    public async Task ClientDoesNotRetryNonTransientClientFailure()
    {
        var handler = new SequencedHttpMessageHandler(
            new HttpResponseMessage(HttpStatusCode.BadRequest)
            {
                Content = new StringContent("bad request")
            });
        using var client = CreateClient(handler, reconnectionAttempts: 3);

        var exception = await Assert.ThrowsAsync<InvalidOperationException>(
            () => client.GetResponseAsync([new ChatMessage(ChatRole.User, "write summary")]));

        Assert.Contains("400", exception.Message);
        Assert.Equal(1, handler.RequestCount);
    }

    [Fact]
    public async Task ClientSendsUserInitiatorOnceThenAgentInitiator()
    {
        var handler = new RecordingHttpMessageHandler(_ => CreateStreamedTextResponse("Done."));
        using var client = CreateClient(handler, reconnectionAttempts: 0);

        await client.GetResponseAsync([new ChatMessage(ChatRole.User, "write summary")]);
        await client.GetResponseAsync([new ChatMessage(ChatRole.Tool, "tool result")]);
        await client.GetResponseAsync([new ChatMessage(ChatRole.Assistant, "continue")]);

        Assert.Equal(["user", "agent", "agent"], handler.RequestHeaders["X-Initiator"]);
    }

    [Fact]
    public async Task ClientUsesResponsesCompactionEndpointWhenContextThresholdIsReached()
    {
        var handler = new RecordingHttpMessageHandler(request =>
        {
            if (request.RequestUri?.PathAndQuery == "/v1/responses/compact")
            {
                return new HttpResponseMessage(HttpStatusCode.OK)
                {
                    Content = new StringContent("""
                        {
                          "input": [
                            {
                              "role": "user",
                              "content": "compacted"
                            }
                          ]
                        }
                        """)
                };
            }

            return CreateStreamedTextResponse("Done.");
        });
        using var client = CreateClient(
            handler,
            reconnectionAttempts: 0,
            new LiteLlmResponsesCompactionOptions
            {
                Enabled = true,
                ContextWindowTokens = 100,
                MaxOutputTokens = 10,
                RemainingRatio = 0.10,
                CompactPath = "responses/compact"
            });

        var response = await client.GetResponseAsync([new ChatMessage(ChatRole.User, new string('x', 800))]);

        Assert.Equal("Done.", response.Text);
        Assert.Equal(["/v1/responses/compact", "/v1/responses"], handler.RequestPaths);
        Assert.Contains("\"content\":\"compacted\"", handler.RequestBodies[1]);
        Assert.Equal(["agent", "user"], handler.RequestHeaders["X-Initiator"]);
    }

    [Fact]
    public async Task ClientFallsBackToAgentFrameworkCompactionWhenResponsesCompactionFails()
    {
        var handler = new RecordingHttpMessageHandler(request =>
        {
            if (request.RequestUri?.PathAndQuery == "/v1/responses/compact")
            {
                return new HttpResponseMessage(HttpStatusCode.NotFound)
                {
                    Content = new StringContent("missing")
                };
            }

            return CreateStreamedTextResponse("Done.");
        });
        using var client = CreateClient(
            handler,
            reconnectionAttempts: 0,
            new LiteLlmResponsesCompactionOptions
            {
                Enabled = true,
                ContextWindowTokens = 100,
                MaxOutputTokens = 10,
                RemainingRatio = 0.10,
                CompactPath = "responses/compact",
                FallbackStrategy = OpenAiMeetingSummaryAgentPipeline.CreateFallbackCompactionStrategyForTests(
                    contextWindowTokens: 100,
                    maxOutputTokens: 10,
                    remainingRatio: 0.10,
                    summaryClient: null)
            });

        var response = await client.GetResponseAsync(
            Enumerable.Range(0, 8)
                .Select(index => new ChatMessage(ChatRole.User, $"turn {index} {new string('x', 200)}")));

        Assert.Equal("Done.", response.Text);
        Assert.Equal(["/v1/responses/compact", "/v1/responses"], handler.RequestPaths);
        Assert.DoesNotContain("turn 0", handler.RequestBodies[1]);
    }

    private static LiteLlmResponsesChatClient CreateClient(
        HttpMessageHandler handler,
        int reconnectionAttempts,
        LiteLlmResponsesCompactionOptions? compactionOptions = null,
        Action? retrying = null,
        Action<string>? reasoningSummaryChanged = null,
        bool useStreaming = true)
    {
        return new LiteLlmResponsesChatClient(
            new HttpClient(handler)
            {
                BaseAddress = new Uri("https://litellm.example/v1/")
            },
            "test-key",
            "test-model",
            enableThinking: false,
            reasoningEffort: "none",
            reconnectionAttempts,
            TimeSpan.Zero,
            compactionOptions,
            retrying: retrying,
            reasoningSummaryChanged: reasoningSummaryChanged,
            useStreaming: useStreaming);
    }

    private static HttpResponseMessage CreateStreamedTextResponse(
        string text,
        params string[] reasoningSummaries)
    {
        var events = new List<string>();
        for (var index = 0; index < reasoningSummaries.Length; index++)
        {
            events.Add("data: " + JsonSerializer.Serialize(new
            {
                type = "response.reasoning_summary_text.delta",
                item_id = "reasoning_stream",
                output_index = 0,
                summary_index = index,
                delta = reasoningSummaries[index]
            }));
        }

        var outputIndex = reasoningSummaries.Length > 0 ? 1 : 0;
        events.Add("data: " + JsonSerializer.Serialize(new
        {
            type = "response.output_text.delta",
            item_id = "msg_stream",
            output_index = outputIndex,
            content_index = 0,
            delta = text
        }));
        events.Add(
            """data: {"type":"response.completed","response":{"id":"resp_stream","created_at":1779147100,"model":"gpt-5.5","object":"response","output":[],"parallel_tool_calls":true,"status":"completed","store":false,"usage":{"input_tokens":12,"output_tokens":3,"total_tokens":15}}}""");
        events.Add("data: [DONE]");

        return new HttpResponseMessage(HttpStatusCode.OK)
        {
            Content = new StringContent(
                string.Join($"{Environment.NewLine}{Environment.NewLine}", events)
                    + Environment.NewLine
                    + Environment.NewLine,
                Encoding.UTF8,
                "text/event-stream")
        };
    }

    private sealed class SequencedHttpMessageHandler : HttpMessageHandler
    {
        private readonly Queue<HttpResponseMessage> responses;

        public SequencedHttpMessageHandler(params HttpResponseMessage[] responses)
        {
            this.responses = new Queue<HttpResponseMessage>(responses);
        }

        public int RequestCount { get; private set; }

        protected override Task<HttpResponseMessage> SendAsync(
            HttpRequestMessage request,
            CancellationToken cancellationToken)
        {
            RequestCount++;
            return Task.FromResult(responses.Dequeue());
        }
    }

    private sealed class RecordingHttpMessageHandler : HttpMessageHandler
    {
        private readonly Func<HttpRequestMessage, HttpResponseMessage> handler;

        public RecordingHttpMessageHandler(Func<HttpRequestMessage, HttpResponseMessage> handler)
        {
            this.handler = handler;
        }

        public List<string> RequestPaths { get; } = [];

        public List<string> RequestBodies { get; } = [];

        public Dictionary<string, List<string>> RequestHeaders { get; } = new(StringComparer.OrdinalIgnoreCase);

        protected override async Task<HttpResponseMessage> SendAsync(
            HttpRequestMessage request,
            CancellationToken cancellationToken)
        {
            RequestPaths.Add(request.RequestUri?.PathAndQuery ?? string.Empty);
            RequestBodies.Add(request.Content is null
                ? string.Empty
                : await request.Content.ReadAsStringAsync(cancellationToken));
            foreach (var header in request.Headers)
            {
                if (!RequestHeaders.TryGetValue(header.Key, out var values))
                {
                    values = [];
                    RequestHeaders[header.Key] = values;
                }

                values.AddRange(header.Value);
            }

            return handler(request);
        }
    }
}