ClawdDotNet/tests/ClawdDotNet.Core.Tests/Infrastructure/FakeChatClient.cs

using System.Text.Json;
using ClawdDotNet.Core.Api;
using ClawdDotNet.Core.Api.Models;

namespace ClawdDotNet.Core.Tests.Infrastructure;

/// <summary>
/// Replaces the OpenRouterClient in tests. Delivers a preprogrammed sequence of answers
/// and records every request received.
///
/// Important: requests are deep-copied. The engine passes on the same List&lt;ChatMessage&gt;
/// and changes it afterwards — without a copy tests would check the final state
/// instead of what was actually sent.
/// </summary>
internal sealed class FakeChatClient : IChatCompletionClient
{
    private readonly Queue<Func<ChatRequest, ChatResponse>> _responses = new();
    private Func<ChatRequest, ChatResponse>? _fallback;
    private readonly Lock _gate = new();

    /// <summary>All requests received, as deep copies.</summary>
    public List<ChatRequest> ReceivedRequests { get; } = new();

    public int CallCount => ReceivedRequests.Count;

    // ─── Programming the answers ───

    public FakeChatClient RespondsWithText(string text, Usage? usage = null)
    {
        _responses.Enqueue(_ => TextResponse(text, usage));
        return this;
    }

    public FakeChatClient RespondsWithToolCall(string toolName, string argumentsJson = "{}", string? id = null)
        => RespondsWithToolCalls((toolName, argumentsJson, id));

    public FakeChatClient RespondsWithToolCalls(params (string Tool, string Args, string? Id)[] calls)
    {
        _responses.Enqueue(_ => ToolCallResponse(calls));
        return this;
    }

    private static ChatResponse ToolCallResponse((string Tool, string Args, string? Id)[] calls)
    {
        // IDs must be unique per answer, otherwise parallel runs collide.
        var toolCalls = calls.Select(c => new ToolCall
        {
            Id = c.Id ?? "call_" + Guid.NewGuid().ToString("N")[..12],
            Type = "function",
            Function = new ToolCallFunction { Name = c.Tool, Arguments = c.Args }
        }).ToList();

        return new ChatResponse
        {
            Id = "resp",
            Choices = [new Choice { Index = 0, Message = new ChatMessage { Role = "assistant", ToolCalls = toolCalls } }],
            Usage = new Usage { PromptTokens = 100, CompletionTokens = 20, TotalTokens = 120 }
        };
    }

    /// <summary>Answers with a particular prompt token count — for compaction thresholds.</summary>
    public FakeChatClient RespondsWithTokens(int promptTokens, string text = "fertig")
    {
        _responses.Enqueue(_ => TextResponse(text, new Usage
        {
            PromptTokens = promptTokens,
            CompletionTokens = 10,
            TotalTokens = promptTokens + 10
        }));
        return this;
    }

    public FakeChatClient Throws(Exception ex)
    {
        _responses.Enqueue(_ => throw ex);
        return this;
    }

    /// <summary>
    /// Answers with a text cut off at the output limit (#548).
    /// </summary>
    public FakeChatClient RespondsWithTruncatedText(string text, string finishReason = "length")
    {
        _responses.Enqueue(_ => new ChatResponse
        {
            Id = "resp",
            Choices =
            [
                new Choice
                {
                    Index = 0,
                    Message = new ChatMessage { Role = "assistant", Content = text },
                    FinishReason = finishReason
                }
            ],
            Usage = new Usage { PromptTokens = 100, CompletionTokens = 20, TotalTokens = 120 }
        });
        return this;
    }

    /// <summary>
    /// Answers with tool calls AND accompanying text or reasoning — the case
    /// #547 and #549 concern.
    /// </summary>
    public FakeChatClient RespondsWithToolCallAndContent(
        string toolName,
        string content,
        string? reasoning = null,
        string? reasoningDetailsJson = null,
        string argumentsJson = "{}")
    {
        _responses.Enqueue(_ =>
        {
            var response = ToolCallResponse([(toolName, argumentsJson, null)]);
            var message = response.Choices[0].Message;
            message.Content = content;
            message.Reasoning = reasoning;
            if (reasoningDetailsJson is not null)
                message.ReasoningDetails = JsonDocument.Parse(reasoningDetailsJson).RootElement.Clone();
            return response;
        });
        return this;
    }

    /// <summary>Answer for all calls that go beyond the programmed sequence.</summary>
    public FakeChatClient AlwaysRespondsWithText(string text)
    {
        _fallback = _ => TextResponse(text);
        return this;
    }

    public FakeChatClient AlwaysRespondsWithToolCall(string toolName = "TestTool")
    {
        _fallback = _ => ToolCallResponse([(toolName, "{}", null)]);
        return this;
    }

    /// <summary>
    /// Answers matching the respective state of the conversation: after a tool answer
    /// comes text, otherwise a tool call. That way the fake behaves sensibly
    /// even when several conversations run interleaved.
    /// </summary>
    public FakeChatClient RespondsContextually(string toolName = "TestTool")
    {
        _fallback = req => req.Messages.LastOrDefault()?.Role == "tool"
            ? TextResponse("Fertig")
            : ToolCallResponse([(toolName, "{}", null)]);
        return this;
    }

    // ─── IChatCompletionClient ───

    public Task<ChatResponse> CompleteAsync(ChatRequest request, CancellationToken ct)
    {
        ct.ThrowIfCancellationRequested();

        // Copy first, then evaluate: the engine keeps mutating the original list,
        // and in concurrency tests the fake should not fail because of that.
        var snapshot = DeepClone(request);
        lock (_gate)
            ReceivedRequests.Add(snapshot);

        Func<ChatRequest, ChatResponse>? responder;
        lock (_gate)
            responder = _responses.Count > 0 ? _responses.Dequeue() : _fallback;

        if (responder is null)
            throw new InvalidOperationException(
                $"FakeChatClient: unerwarteter Aufruf Nr. {ReceivedRequests.Count} — " +
                "keine Antwort mehr programmiert.");

        return Task.FromResult(responder(snapshot));
    }

    // ─── Helpers ───

    private static ChatResponse TextResponse(string text, Usage? usage = null) => new()
    {
        Id = "resp",
        Choices = [new Choice { Index = 0, Message = new ChatMessage { Role = "assistant", Content = text }, FinishReason = "stop" }],
        Usage = usage ?? new Usage { PromptTokens = 100, CompletionTokens = 20, TotalTokens = 120 }
    };

    private static ChatRequest DeepClone(ChatRequest request) => new()
    {
        Model = request.Model,
        Stream = request.Stream,
        Temperature = request.Temperature,
        MaxTokens = request.MaxTokens,
        ToolChoice = request.ToolChoice,
        Tools = request.Tools?.ToList(),
        // The gateway fields must be copied too, otherwise a test checks the hardening
        // past the wire format — the same trap as with CacheBreakpoint.
        Provider = request.Provider,
        Usage = request.Usage,
        Reasoning = request.Reasoning,
        Messages = request.Messages.Select(CloneMessage).ToList()
    };

    private static ChatMessage CloneMessage(ChatMessage m) => new()
    {
        Role = m.Role,
        Content = m.Content,
        ToolCallId = m.ToolCallId,
        // Must be copied too, otherwise caching tests check past the wire format.
        CacheBreakpoint = m.CacheBreakpoint,
        Reasoning = m.Reasoning,
        ReasoningDetails = m.ReasoningDetails,
        ToolCalls = m.ToolCalls?.Select(tc => new ToolCall
        {
            Id = tc.Id,
            Type = tc.Type,
            Function = new ToolCallFunction { Name =tc.Function.Name, Arguments = tc.Function.Arguments }
        }).ToList()
    };
}