gott/Services/AnthropicService.Agents.cs

using System;
using System.Collections.Generic;
using System.Linq;
using System.Net.Http;
using System.Text;
using System.Text.Json;
using System.Threading;
using System.Threading.Tasks;
using Gott.Models;
using Microsoft.AspNetCore.Http;

namespace Gott.Services;

public partial class AnthropicService
{
    // ── Preparer-Agent Logik ──

    public class PreparerResult
    {
        public bool ShouldForward { get; set; }
        public string ForwardedPrompt { get; set; }
        public string ResponseText { get; set; }
        public int IterationCount { get; set; }
    }

    /// <summary>
    /// Führt den Preparer-Agent aus. Der Preparer bekommt die User-Nachricht und die bisherige Konversation.
    /// Antwortet er mit [FORWARD] → der Text nach [FORWARD] wird an den Session-Agent weitergegeben.
    /// Sonst → seine Antwort wird als Rückfrage an den User gestreamt.
    /// </summary>
    public async Task<PreparerResult> RunPreparer(
        HttpResponse response,
        SubAgentContext preparer,
        List<object> historyMessages,
        string chainId,
        string workspace,
        string sessionId,
        string apiToken,
        CancellationToken cancellationToken)
    {
        // Preparer läuft VOR dem Worker und nutzt den Request-CancellationToken direkt.
        // Bei Client-Disconnect wird der Preparer abgebrochen (kein Sinn weiterzumachen,
        // da der Worker noch nicht gestartet hat).
        cancellationToken.ThrowIfCancellationRequested();

        await SseHelper.TryWriteSse(response, "preparer_start", new
        {
            agentId = preparer.AgentId,
            agentName = preparer.Agent?.Name ?? "Vorbereiter-Agent"
        }, cancellationToken);

        // Preparer bekommt die History als Kontext
        var preparerMessages = new List<object>(historyMessages);

        var usage = new TokenUsage();
        string text;
        int iterations;

        if (!string.IsNullOrEmpty(preparer.Provider) && _openAiService != null)
        {
            // OpenAI-kompatibler Preparer
            var subCtx = BuildSubAgentChatContext(preparer, chainId, sessionId, apiToken);
            text = await _openAiService.ChatNonStreaming(subCtx, preparerMessages);
            iterations = 1;
        }
        else if (string.IsNullOrEmpty(apiToken) && _cliService != null && _cliService.IsAvailable)
        {
            // Kein API-Key → CLI-Fallback
            (text, iterations) = await RunSubAgentViaCli(
                response, preparer, preparerMessages, chainId, sessionId,
                "preparer", cancellationToken, cancellationToken);
        }
        else
        {
            (text, iterations) = await RunAgentChat(
                response, preparer.ModelId, preparer.SystemPrompt, preparer.AgentId,
                preparer.Workspace, preparer.UseMemory, preparer.WebSearchVersion,
                preparer.ToolChoice, preparerMessages, chainId, sessionId, apiToken, usage,
                0, "preparer", preparer.Agent?.Name ?? "Vorbereiter-Agent",
                cancellationToken, cancellationToken);
        }

        var result = new PreparerResult { IterationCount = iterations, ResponseText = text };

        // Logik: Standardmäßig wird IMMER weitergeleitet.
        // Nur wenn der Preparer explizit [QUESTION] nutzt, wird gestoppt (= Rückfrage an User).
        // [FORWARD] wird weiterhin erkannt und entfernt, ist aber nicht mehr zwingend nötig.
        if (!string.IsNullOrEmpty(text) && text.Contains("[QUESTION]"))
        {
            // Explizite Rückfrage an User
            var questionText = text.Replace("[QUESTION]", "").Trim();
            result.ShouldForward = false;
            result.ResponseText = questionText;

            await SseHelper.TryWriteSse(response, "preparer_done", new
            {
                agentName = preparer.Agent?.Name ?? "Vorbereiter-Agent",
                forwarded = false,
                responseText = questionText
            }, cancellationToken);
        }
        else
        {
            // Weiterleitung (Standard): [FORWARD]-Tag entfernen falls vorhanden
            var forwardedText = text ?? "";
            if (forwardedText.Contains("[FORWARD]"))
            {
                var idx = forwardedText.IndexOf("[FORWARD]");
                forwardedText = forwardedText.Substring(idx + "[FORWARD]".Length).Trim();
            }

            result.ShouldForward = true;
            result.ForwardedPrompt = forwardedText;

            await SseHelper.TryWriteSse(response, "preparer_done", new
            {
                agentName = preparer.Agent?.Name ?? "Vorbereiter-Agent",
                forwarded = true
            }, cancellationToken);
        }

        return result;
    }

    // ── Review-Agent Logik ──

    private class ReviewResult
    {
        public bool HasFeedback { get; set; }
        public string Feedback { get; set; }
        public int ReviewIterationCount { get; set; }
    }

    /// <summary>
    /// Führt den Review-Agent aus. Der Review-Agent bekommt die Text-Konversation
    /// und kann Tools nutzen (z.B. read_file um die Arbeit zu prüfen).
    /// Default = Approved. Nur bei explizitem [FEEDBACK]-Tag wird Feedback an den Worker zurückgegeben.
    /// reviewRound = aktuelle Runde (0-basiert), wird vom Caller hochgezählt.
    /// </summary>
    private async Task<ReviewResult> RunReview(
        HttpResponse response,
        SubAgentContext review,
        List<object> workerMessages,
        string chainId,
        string workspace,
        string sessionId,
        string apiToken,
        TokenUsage usage,
        int baseIterations,
        int reviewRound,
        CancellationToken workCt,
        CancellationToken sseCt)
    {
        await SseHelper.TryWriteSse(response, "review_start", new
        {
            agentId = review.AgentId,
            agentName = review.Agent?.Name ?? "Review-Agent",
            round = reviewRound
        }, sseCt);

        // Review-Agent bekommt die Worker-Konversation als Kontext
        var reviewMessages = BuildReviewMessages(workerMessages);

        string text;
        int iterations;

        if (!string.IsNullOrEmpty(review.Provider) && _openAiService != null)
        {
            // OpenAI-kompatibler Reviewer
            var subCtx = BuildSubAgentChatContext(review, chainId, sessionId, apiToken);
            text = await _openAiService.ChatNonStreaming(subCtx, reviewMessages);
            iterations = 1;
        }
        else if (string.IsNullOrEmpty(apiToken) && _cliService != null && _cliService.IsAvailable)
        {
            // Kein API-Key → CLI-Fallback
            (text, iterations) = await RunSubAgentViaCli(
                response, review, reviewMessages, chainId, sessionId,
                "review", workCt, sseCt);
        }
        else
        {
            (text, iterations) = await RunAgentChat(
                response, review.ModelId, review.SystemPrompt, review.AgentId,
                review.Workspace, review.UseMemory, review.WebSearchVersion,
                review.ToolChoice, reviewMessages, chainId, sessionId, apiToken, usage,
                baseIterations, "review", review.Agent?.Name ?? "Review-Agent",
                workCt, sseCt);
        }

        var result = new ReviewResult { ReviewIterationCount = iterations };

        if (!string.IsNullOrEmpty(text) && text.Contains("[FEEDBACK]"))
        {
            var feedbackText = text.Substring(text.IndexOf("[FEEDBACK]") + "[FEEDBACK]".Length).Trim();
            result.HasFeedback = true;
            result.Feedback = $"[Review-Feedback, Runde {reviewRound + 1}]: {feedbackText}";
        }

        await SseHelper.TryWriteSse(response, "review_done", new
        {
            agentName = review.Agent?.Name ?? "Review-Agent",
            round = reviewRound,
            hasFeedback = result.HasFeedback
        }, sseCt);

        return result;
    }

    /// <summary>
    /// Führt den Finalizer-Agent aus. Wird einmal nach der Review-Phase (oder direkt nach dem Worker) aufgerufen.
    /// Der Finalizer kann Tools nutzen, gibt aber kein Feedback an den Worker zurück.
    /// </summary>
    private async Task<int> RunFinalizer(
        HttpResponse response,
        SubAgentContext finalizer,
        List<object> workerMessages,
        string chainId,
        string workspace,
        string sessionId,
        string apiToken,
        TokenUsage usage,
        int baseIterations,
        CancellationToken workCt,
        CancellationToken sseCt)
    {
        await SseHelper.TryWriteSse(response, "finalizer_start", new
        {
            agentId = finalizer.AgentId,
            agentName = finalizer.Agent?.Name ?? "Finalizer-Agent"
        }, sseCt);

        // Finalizer bekommt die Konversation als Kontext (gleiche Methode wie Review)
        var finalizerMessages = BuildReviewMessages(workerMessages);

        string text;
        int iterations;

        if (!string.IsNullOrEmpty(finalizer.Provider) && _openAiService != null)
        {
            var subCtx = BuildSubAgentChatContext(finalizer, chainId, sessionId, apiToken);
            text = await _openAiService.ChatNonStreaming(subCtx, finalizerMessages);
            iterations = 1;
        }
        else if (string.IsNullOrEmpty(apiToken) && _cliService != null && _cliService.IsAvailable)
        {
            // Kein API-Key → CLI-Fallback
            (text, iterations) = await RunSubAgentViaCli(
                response, finalizer, finalizerMessages, chainId, sessionId,
                "finalizer", workCt, sseCt);
        }
        else
        {
            (text, iterations) = await RunAgentChat(
                response, finalizer.ModelId, finalizer.SystemPrompt, finalizer.AgentId,
                finalizer.Workspace, finalizer.UseMemory, finalizer.WebSearchVersion,
                finalizer.ToolChoice, finalizerMessages, chainId, sessionId, apiToken, usage,
                baseIterations, "finalizer", finalizer.Agent?.Name ?? "Finalizer-Agent",
                workCt, sseCt);
        }

        await SseHelper.TryWriteSse(response, "finalizer_done", new
        {
            agentName = finalizer.Agent?.Name ?? "Finalizer-Agent"
        }, sseCt);

        return iterations;
    }

    /// <summary>
    /// Baut die Messages für den Review-Agent auf.
    /// Der Review-Agent bekommt die komplette Worker-Konversation als eine User-Message.
    /// </summary>
    private List<object> BuildReviewMessages(List<object> workerMessages)
    {
        // Nur die Text-Konversation extrahieren – Tool-Calls und Tool-Results weglassen.
        // Der Review/Finalizer-Agent hat eigene Tools und kann bei Bedarf selbst Dateien lesen.
        // Aufeinanderfolgende Messages gleicher Rolle werden zusammengeführt,
        // da die Anthropic API abwechselnde user/assistant-Rollen erfordert.
        var messages = new List<object>();
        string lastRole = null;
        var currentParts = new List<string>();

        foreach (var msg in workerMessages)
        {
            var json = JsonSerializer.Serialize(msg);
            var el = JsonSerializer.Deserialize<JsonElement>(json);

            if (!el.TryGetProperty("role", out var role)) continue;
            var roleStr = role.GetString() ?? "";

            if (!el.TryGetProperty("content", out var content)) continue;

            string textContent = null;

            if (content.ValueKind == JsonValueKind.String)
            {
                textContent = content.GetString();
            }
            else if (content.ValueKind == JsonValueKind.Array)
            {
                var parts = new List<string>();
                foreach (var block in content.EnumerateArray())
                {
                    var blockType = block.TryGetProperty("type", out var bt) ? bt.GetString() ?? "" : "";
                    if (blockType == "text" && block.TryGetProperty("text", out var txt))
                        parts.Add(txt.GetString() ?? "");
                }
                if (parts.Count > 0)
                    textContent = string.Join("\n", parts);
            }

            if (string.IsNullOrEmpty(textContent)) continue;

            if (roleStr == lastRole)
            {
                // Gleiche Rolle → Text anhängen
                currentParts.Add(textContent);
            }
            else
            {
                // Rolle wechselt → vorherige Message abschließen
                if (lastRole != null && currentParts.Count > 0)
                {
                    messages.Add(new
                    {
                        role = lastRole,
                        content = string.Join("\n\n", currentParts)
                    });
                }
                lastRole = roleStr;
                currentParts = new List<string> { textContent };
            }
        }

        // Letzte Message abschließen
        if (lastRole != null && currentParts.Count > 0)
        {
            messages.Add(new
            {
                role = lastRole,
                content = string.Join("\n\n", currentParts)
            });
        }

        // Sicherstellen, dass die Messages mit "user" beginnen
        if (messages.Count > 0)
        {
            var firstJson = JsonSerializer.Serialize(messages[0]);
            var firstEl = JsonSerializer.Deserialize<JsonElement>(firstJson);
            if (firstEl.GetProperty("role").GetString() != "user")
            {
                messages.Insert(0, new { role = "user", content = "Bitte überprüfe die folgende Konversation und die Arbeit des Agents." });
            }
        }
        else
        {
            messages.Add(new { role = "user", content = "Es gibt keine Konversation zum Überprüfen." });
        }

        return messages;
    }

    /// <summary>
    /// Baut einen ChatContext für Sub-Agents (Preparer/Review/Finalizer) auf,
    /// wenn diese einen OpenAI-kompatiblen Provider nutzen.
    /// </summary>
    private ChatContext BuildSubAgentChatContext(SubAgentContext sub, string chainId, string sessionId, string apiToken)
    {
        // SystemPrompt als Text extrahieren (SubAgent hat List<object> mit { type, text } Blöcken)
        var systemText = "";
        if (sub.SystemPrompt != null)
        {
            var parts = new List<string>();
            foreach (var block in sub.SystemPrompt)
            {
                var json = JsonSerializer.Serialize(block);
                var el = JsonSerializer.Deserialize<JsonElement>(json);
                if (el.TryGetProperty("text", out var t))
                    parts.Add(t.GetString() ?? "");
            }
            systemText = string.Join("\n\n", parts);
        }

        return new ChatContext
        {
            ChainId = chainId,
            SessionId = sessionId,
            ModelId = sub.ModelId,
            Provider = sub.Provider,
            Workspace = sub.Workspace,
            AgentId = sub.AgentId,
            SystemPromptText = systemText,
            UseMemory = sub.UseMemory,
            WebSearchVersion = sub.WebSearchVersion,
            MaxIterations = 20,
            ApiToken = apiToken
        };
    }

    // ── Shared Non-Streaming Tool Loop ──

    /// <summary>
    /// Optionale Callbacks für RunNonStreamingToolLoop.
    /// </summary>
    private class NonStreamingLoopCallbacks
    {
        /// <summary>Wird zu Beginn jeder Iteration aufgerufen (z.B. Typing-Indicator).</summary>
        public Func<int, Task> OnIteration { get; set; }

        /// <summary>SSE-Response für Client-Events. Null = keine SSE-Events.</summary>
        public HttpResponse SseResponse { get; set; }
        public CancellationToken SseCt { get; set; }

        /// <summary>SSE-Event-Prefix (z.B. "review" → "review_text", "review_tool").</summary>
        public string SsePrefix { get; set; }
        public string AgentName { get; set; }

        /// <summary>Tag für DB-Records (z.B. "review", "finalizer").</summary>
        public string Tag { get; set; }

        /// <summary>Ob Text-Embeddings in die Queue geschrieben werden.</summary>
        public bool EnqueueEmbeddings { get; set; }

        /// <summary>
        /// Optionaler Tool-Executor. Parameter: (toolName, input, ct) → result.
        /// Bei Fehler Exception werfen. Wenn null, wird Standard-Executor verwendet.
        /// </summary>
        public Func<string, JsonElement, CancellationToken, Task<string>> ExecuteTool { get; set; }
    }

    /// <summary>
    /// Gemeinsamer Non-Streaming Tool-Loop für RunAgentChat und ChatNonStreaming.
    /// Baut Payload, sendet API-Request, verarbeitet Content-Blöcke, führt Tools aus.
    /// systemPrompt kann List&lt;object&gt; (Agent) oder string (ChatContext) sein.
    /// </summary>
    private async Task<(string text, int iterations)> RunNonStreamingToolLoop(
        string modelId,
        object systemPrompt,
        ToolPayloadOptions toolOpts,
        List<object> messages,
        string chainId,
        string sessionId,
        string apiToken,
        TokenUsage usage,
        int baseIteration,
        int maxIterations,
        CancellationToken ct,
        NonStreamingLoopCallbacks cb = null)
    {
        cb ??= new NonStreamingLoopCallbacks();
        int iterations = 0;
        var finalText = new StringBuilder();

        while (iterations < maxIterations)
        {
            ct.ThrowIfCancellationRequested();
            iterations++;
            if (cb.OnIteration != null) await cb.OnIteration(iterations);

            var payload = new Dictionary<string, object>
            {
                ["model"] = modelId,
                ["max_tokens"] = 16384,
                ["messages"] = messages
            };

            if (systemPrompt is List<object> spBlocks && spBlocks.Count > 0)
            {
                // Cache-Breakpoint auf den letzten System-Prompt-Block (1h TTL).
                ApplyCacheControlToLast(spBlocks);
                payload["system"] = spBlocks;
            }
            else if (systemPrompt is string spText && !string.IsNullOrEmpty(spText))
            {
                // String-System-Prompt → in Block-Form mit cache_control konvertieren.
                payload["system"] = new List<object>
                {
                    new Dictionary<string, object>
                    {
                        ["type"] = "text",
                        ["text"] = spText,
                        ["cache_control"] = new { type = "ephemeral", ttl = "1h" }
                    }
                };
            }

            AddToolsToPayload(payload, toolOpts with { Iteration = iterations });

            var client = CreateClient(apiToken);
            var json = JsonSerializer.Serialize(payload);
            LogPayload(json, sessionId, chainId, baseIteration + iterations);
            var httpContent = new StringContent(json, Encoding.UTF8, "application/json");

            using var httpResp = await client.SendAsync(
                new HttpRequestMessage(HttpMethod.Post, "https://api.anthropic.com/v1/messages") { Content = httpContent },
                HttpCompletionOption.ResponseHeadersRead,
                ct
            );

            TryUpdateKeyMeta(httpResp, apiToken);

            if (!httpResp.IsSuccessStatusCode)
            {
                var errBody = await httpResp.Content.ReadAsStringAsync(ct);
                var errMsg = $"API {(int)httpResp.StatusCode}: {errBody}";
                await _db.SaveMessage(new MessageRecord
                {
                    ChainId = chainId, Iteration = baseIteration + iterations,
                    Role = "system", Type = "error",
                    Content = errMsg, IsError = true, Tag = cb.Tag
                });
                if (cb.SseResponse != null)
                    await SseHelper.TryWriteSse(cb.SseResponse, $"{cb.SsePrefix}_error", new { error = errMsg }, cb.SseCt);
                break;
            }

            var respJson = await httpResp.Content.ReadAsStringAsync(ct);
            var respDoc = JsonSerializer.Deserialize<JsonElement>(respJson);

            if (respDoc.TryGetProperty("usage", out var u))
                TrackUsage(usage, u);

            var stopReason = respDoc.TryGetProperty("stop_reason", out var sr) ? sr.GetString() ?? "" : "";
            var contentArr = respDoc.TryGetProperty("content", out var c) ? c : default;
            if (contentArr.ValueKind != JsonValueKind.Array) break;

            var assistantContent = new List<object>();
            var toolResults = new List<object>();

            foreach (var block in contentArr.EnumerateArray())
            {
                var blockType = block.TryGetProperty("type", out var bt) ? bt.GetString() ?? "" : "";
                if (blockType == "text")
                {
                    var text = block.GetProperty("text").GetString() ?? "";
                    finalText.Append(text);
                    assistantContent.Add(new { type = "text", text });

                    var msgId = await _db.SaveMessage(new MessageRecord
                    {
                        ChainId = chainId, Iteration = baseIteration + iterations,
                        Role = "assistant", Type = "text",
                        Content = text, Tag = cb.Tag
                    });
                    if (cb.EnqueueEmbeddings)
                        _embeddingBg?.Enqueue(msgId, chainId, sessionId, toolOpts.AgentId, "assistant", text);
                    if (cb.SseResponse != null)
                        await SseHelper.TryWriteSse(cb.SseResponse, $"{cb.SsePrefix}_text", new { text, agentName = cb.AgentName }, cb.SseCt);
                }
                else if (blockType == "tool_use")
                {
                    var toolId = block.GetProperty("id").GetString() ?? "";
                    var toolName = block.GetProperty("name").GetString() ?? "";
                    JsonElement inputEl;
                    try { inputEl = block.GetProperty("input"); }
                    catch { inputEl = JsonSerializer.Deserialize<JsonElement>("{}"); }

                    assistantContent.Add(new { type = "tool_use", id = toolId, name = toolName, input = inputEl });

                    await _db.SaveMessage(new MessageRecord
                    {
                        ChainId = chainId, Iteration = baseIteration + iterations,
                        Role = "assistant", Type = "tool_use",
                        ToolName = toolName, ToolCallId = toolId,
                        Content = inputEl.ToString(), Tag = cb.Tag
                    });

                    ct.ThrowIfCancellationRequested();

                    string toolResult;
                    bool isError = false;
                    try
                    {
                        toolResult = cb.ExecuteTool != null
                            ? await cb.ExecuteTool(toolName, inputEl, ct)
                            : await _toolService.ExecuteTool(toolName, inputEl, new ToolExecutionContext
                            {
                                Workspace = toolOpts.Workspace, MemoryService = _memoryService,
                                AgentId = toolOpts.AgentId, CancellationToken = ct,
                                VectorSearch = _vectorSearch, EmbeddingBg = _embeddingBg
                            });
                    }
                    catch (OperationCanceledException) { throw; }
                    catch (Exception ex) { toolResult = $"Fehler: {ex.Message}"; isError = true; }

                    ct.ThrowIfCancellationRequested();

                    toolResults.Add(new
                    {
                        type = "tool_result",
                        tool_use_id = toolId,
                        content = TruncateToolResult(toolResult),
                        is_error = isError
                    });

                    await _db.SaveMessage(new MessageRecord
                    {
                        ChainId = chainId, Iteration = baseIteration + iterations,
                        Role = "user", Type = "tool_result",
                        ToolCallId = toolId,
                        Content = toolResult, IsError = isError, Tag = cb.Tag
                    });

                    if (cb.SseResponse != null)
                        await SseHelper.TryWriteSse(cb.SseResponse, $"{cb.SsePrefix}_tool", new { name = toolName, result = TruncateToolResult(toolResult), isError, agentName = cb.AgentName }, cb.SseCt);
                }
            }

            messages.Add(new { role = "assistant", content = assistantContent });

            if (stopReason != "tool_use" || toolResults.Count == 0)
                break;

            messages.Add(new { role = "user", content = toolResults });
        }

        return (finalText.ToString(), iterations);
    }

    /// <summary>
    /// Agent-Chat (non-streaming API, SSE-Events für Client).
    /// Wrapper um RunNonStreamingToolLoop für Review-/Finalizer-Agents.
    /// </summary>
    private Task<(string text, int iterations)> RunAgentChat(
        HttpResponse response,
        string modelId,
        List<object> systemPrompt,
        string agentId,
        string agentWorkspace,
        bool useMemory,
        string webSearchVersion,
        string agentToolChoice,
        List<object> messages,
        string chainId,
        string sessionId,
        string apiToken,
        TokenUsage usage,
        int baseIterations,
        string agentType,
        string agentName,
        CancellationToken workCt,
        CancellationToken sseCt)
    {
        return RunNonStreamingToolLoop(
            modelId, systemPrompt,
            new ToolPayloadOptions
            {
                Workspace = agentWorkspace, AgentId = agentId,
                UseMemory = useMemory, WebSearchVersion = webSearchVersion,
                ToolChoice = agentToolChoice
            },
            messages, chainId, sessionId, apiToken, usage,
            baseIterations, 20, workCt,
            new NonStreamingLoopCallbacks
            {
                SseResponse = response, SseCt = sseCt,
                SsePrefix = agentType, AgentName = agentName,
                Tag = agentType
            });
    }

    /// <summary>
    /// Führt einen Sub-Agent (Preparer/Review/Finalizer) über die Claude CLI aus.
    /// Sammelt den gesamten Text und gibt ihn zurück. Wird als Fallback genutzt,
    /// wenn kein API-Key vorhanden ist.
    /// </summary>
    private async Task<(string text, int iterations)> RunSubAgentViaCli(
        HttpResponse response,
        SubAgentContext sub,
        List<object> historyMessages,
        string chainId,
        string sessionId,
        string agentType,
        CancellationToken workCt,
        CancellationToken sseCt)
    {
        if (_cliService == null || !_cliService.IsAvailable)
            return ("", 0);

        // System-Prompt als Text extrahieren
        var systemText = new StringBuilder();
        if (sub.SystemPrompt != null)
        {
            foreach (var block in sub.SystemPrompt)
            {
                var json = JsonSerializer.Serialize(block);
                var el = JsonSerializer.Deserialize<JsonElement>(json);
                if (el.TryGetProperty("text", out var t))
                    systemText.AppendLine(t.GetString());
            }
        }

        // Letzte User-Message als Prompt extrahieren
        var prompt = ExtractLastUserText(historyMessages) ?? "";

        // Konversations-History als Kontext
        var historyContext = BuildHistoryContext(historyMessages);
        if (!string.IsNullOrEmpty(historyContext))
            prompt = $"<conversation_history>\n{historyContext}\n</conversation_history>\n\n<current_request>\n{prompt}\n</current_request>";

        var fullText = new StringBuilder();

        var result = await _cliService.ChatStreamingAsync(
            prompt,
            async evt =>
            {
                switch (evt.Type)
                {
                    case CliEventType.TextDelta:
                        fullText.Append(evt.Text);
                        await SseHelper.TryWriteSse(response, $"{agentType}_text",
                            new { text = evt.Text, agentName = sub.Agent?.Name ?? agentType }, sseCt);
                        break;
                    case CliEventType.ToolResult:
                        await SseHelper.TryWriteSse(response, $"{agentType}_tool",
                            new { name = evt.ToolName, result = evt.ToolResultContent, isError = evt.ToolResultIsError, agentName = sub.Agent?.Name ?? agentType }, sseCt);
                        break;
                }
            },
            modelId: sub.ModelId,
            systemPrompt: systemText.ToString(),
            workspace: sub.Workspace,
            hasWorkspace: !string.IsNullOrEmpty(sub.Workspace),
            cancellationToken: workCt
        );

        // Text in DB speichern
        if (fullText.Length > 0)
        {
            await _db.SaveMessage(new MessageRecord
            {
                ChainId = chainId,
                Iteration = 0,
                Role = "assistant",
                Type = "text",
                Content = fullText.ToString(),
                Tag = agentType
            });
        }

        return (fullText.ToString(), 1);
    }

    /// <summary>
    /// Non-Streaming Chat für Telegram. Gibt den finalen Antworttext zurück.
    /// Wrapper um RunNonStreamingToolLoop mit Chain-Management und Embedding-Enqueue.
    /// onIteration wird nach jeder Iteration aufgerufen (z.B. für Typing-Indicator).
    /// </summary>
    public async Task<string> ChatNonStreaming(
        ChatContext ctx,
        List<object> messages,
        Func<int, Task> onIteration = null)
    {
        var sw = System.Diagnostics.Stopwatch.StartNew();
        var usage = new TokenUsage();

        var chain = new ChainRecord
        {
            ChainId = ctx.ChainId,
            SessionId = ctx.SessionId,
            Model = ctx.ModelId
        };

        try
        {
            var (text, iterations) = await RunNonStreamingToolLoop(
                ctx.ModelId, ctx.SystemPromptText,
                new ToolPayloadOptions
                {
                    Workspace = ctx.Workspace, AgentId = ctx.AgentId,
                    UseMemory = ctx.UseMemory, WebSearchVersion = ctx.WebSearchVersion,
                    ToolChoice = "auto"
                },
                messages, ctx.ChainId, ctx.SessionId, ctx.ApiToken, usage,
                0, ctx.MaxIterations, CancellationToken.None,
                new NonStreamingLoopCallbacks
                {
                    OnIteration = onIteration,
                    EnqueueEmbeddings = true,
                    ExecuteTool = (name, input, ct) =>
                        _toolService.ExecuteTool(name, input, new ToolExecutionContext
                        {
                            Workspace = ctx.Workspace, MemoryService = _memoryService, AgentId = ctx.AgentId,
                            CancellationToken = ct, VectorSearch = _vectorSearch, EmbeddingBg = _embeddingBg,
                            Db = _db, ChatService = _chatService, AnthropicService = this, OpenAiService = _openAiService
                        })
                });

            var finalStopReason = text.Length > 0 ? "end_turn" : "error";
            await FinalizeChain(chain, usage, iterations, finalStopReason, sw);
            return text;
        }
        catch (Exception ex)
        {
            await _db.SaveMessage(new MessageRecord
            {
                ChainId = ctx.ChainId, Iteration = 0,
                Role = "system", Type = "error",
                Content = $"Fehler: {ex.Message}", IsError = true
            });
            await FinalizeChain(chain, usage, 0, "error", sw);
            return "";
        }
    }
}