gott/Endpoints/ChatEndpoint.cs

using System;
using System.Collections.Generic;
using System.Linq;
using System.Text.Json;
using System.Threading;
using System.Threading.Tasks;
using Gott.Extensions;
using Gott.Models;
using Gott.Services;
using Microsoft.AspNetCore.Builder;
using Microsoft.AspNetCore.Http;
using Microsoft.AspNetCore.Routing;

namespace Gott.Endpoints;

public static class ChatEndpoint
{
    public static void MapChatEndpoint(this IEndpointRouteBuilder app)
    {
        app.MapPost("/api/chat", HandleChat);
        app.MapPost("/api/chat/stop", HandleStop);
    }

    private static async Task HandleChat(HttpRequest req, HttpResponse res, ServerSettingsService settings, ChatService chatService, DatabaseService db, StreamingTracker streamingTracker, AnthropicService anthropic, OpenAiService openAi, RemoteConnectionManager remoteMgr, EmbeddingBackgroundService embeddingBg, CancellationToken cancellationToken)
    {
        var body = await JsonSerializer.DeserializeAsync<ChatRequest>(req.Body);
        if (body == null)
        {
            res.StatusCode = 400;
            await res.WriteAsJsonAsync(new { error = "body required" });
            return;
        }

        var sessionId = body.SessionId;
        if (string.IsNullOrEmpty(sessionId))
        {
            res.StatusCode = 400;
            await res.WriteAsJsonAsync(new { error = "sessionId required" });
            return;
        }

        // Prüfe ob dies eine Choice-Antwort ist (ask_user_choice response)
        var isChoiceResponse = !string.IsNullOrEmpty(body.ChoiceToolCallId) && !string.IsNullOrEmpty(body.ChoiceResult);

        if (!isChoiceResponse && body.Content.ValueKind == JsonValueKind.Undefined)
        {
            res.StatusCode = 400;
            await res.WriteAsJsonAsync(new { error = "content required" });
            return;
        }

        var userId = req.HttpContext.GetUserId();

        var maxIterations = settings.GetInt("MaxIterations", 100);

        var remoteConn = remoteMgr.GetBySessionId(sessionId);
        var hasRemote = remoteConn != null;
        var ctx = await chatService.PrepareContext(sessionId, userId, maxIterations, planMode: body.PlanMode, hasRemoteConnection: hasRemote, remoteOs: remoteConn?.OsVersion);
        if (ctx == null)
        {
            res.StatusCode = 400;
            await res.WriteAsJsonAsync(new { error = "Session/Agent nicht gefunden oder kein Agent ausgewählt" });
            return;
        }

        // API-Key des Users für diesen Agent auflösen (kann null sein)
        var userApiKey = await db.GetAgentApiKeyRecord(userId, ctx.AgentId);

        if (isChoiceResponse)
        {
            // Choice-Antwort: tool_result in die VORHERIGE Chain speichern (wo der tool_use liegt)
            // Damit BuildHistoryFromDb die tool_use + tool_result korrekt zusammenbauen kann.
            var previousChainId = await chatService.FindChainWithToolCall(sessionId, body.ChoiceToolCallId);
            if (previousChainId != null)
            {
                await db.SaveMessage(new MessageRecord
                {
                    ChainId = previousChainId,
                    Iteration = 999, // Hohe Iteration damit es nach dem tool_use kommt
                    Role = "user",
                    Type = "tool_result",
                    ToolCallId = body.ChoiceToolCallId,
                    Content = body.ChoiceResult
                });
            }
            else
            {
                // Fallback: In aktuelle Chain speichern
                await db.SaveMessage(new MessageRecord
                {
                    ChainId = ctx.ChainId,
                    Iteration = 0,
                    Role = "user",
                    Type = "tool_result",
                    ToolCallId = body.ChoiceToolCallId,
                    Content = body.ChoiceResult
                });
            }
        }
        else
        {
            // Normale User-Message aufbereiten und speichern
            var userMessage = new ChatMessage { Role = "user", Content = body.Content };
            var textContent = userMessage.GetTextContent();
            var attachmentBlocks = userMessage.GetAttachmentBlocks();

            var attachmentRecords = attachmentBlocks.Select(b => new AttachmentRecord
            {
                Type = b.Type,
                MediaType = b.MediaType,
                FileName = b.FileName,
                Data = b.Data
            }).ToList();

            var userMsgId = await db.SaveMessageWithAttachments(new MessageRecord
            {
                ChainId = ctx.ChainId,
                Iteration = 0,
                Role = "user",
                Type = "text",
                Content = textContent
            }, attachmentRecords);
            embeddingBg?.Enqueue(userMsgId, ctx.ChainId, ctx.SessionId, ctx.AgentId, "user", textContent ?? "");
        }

        res.Headers["Content-Type"] = "text/event-stream";
        res.Headers["Cache-Control"] = "no-cache";
        res.Headers["Connection"] = "keep-alive";

        // Response-Buffering deaktivieren – wichtig für SSE unter Windows
        var bufferingFeature = res.HttpContext.Features.Get<Microsoft.AspNetCore.Http.Features.IHttpResponseBodyFeature>();
        bufferingFeature?.DisableBuffering();

        // Kestrel MinResponseDataRate deaktivieren für diesen SSE-Stream.
        // Ohne dies killt Kestrel die Verbindung wenn bei langen Tool-Ausführungen
        // keine Daten fließen (Default: 240 bytes/s, 5s Grace → "client_disconnected").
        var minRateFeature = res.HttpContext.Features.Get<Microsoft.AspNetCore.Server.Kestrel.Core.Features.IHttpMinResponseDataRateFeature>();
        if (minRateFeature != null)
            minRateFeature.MinDataRate = null;

        // ── Preparer-Agent Phase (vor dem Worker) ──
        if (ctx.Preparer != null && !isChoiceResponse)
        {
            var preparerApiToken = userApiKey?.ApiKey;

            // Kein Key und kein CLI → Preparer überspringen (main agent wird auch CLI nutzen oder fehlermelden)
            bool skipPreparer = string.IsNullOrEmpty(ctx.Preparer.Provider)
                && string.IsNullOrEmpty(preparerApiToken)
                && !anthropic.IsCliAvailable;

            if (!skipPreparer)
            {
            // Preparer bekommt History MIT seinen eigenen vorherigen Messages (Tag "preparer"),
            // damit er den Kontext seiner Rückfragen kennt
            var preparerHistory = await chatService.BuildHistoryFromDb(sessionId, includeTag: "preparer", promptBlocks: ctx.PromptBlocks);

            var preparerResult = await anthropic.RunPreparer(
                res, ctx.Preparer, preparerHistory, ctx.ChainId, ctx.Workspace, sessionId, preparerApiToken, cancellationToken);

            if (!preparerResult.ShouldForward)
            {
                // Preparer hat eine Rückfrage → Antwort als Text streamen, Flow beenden
                // Preparer-Antwort wird schon in RunAgentChat in DB gespeichert
                await SseHelper.WriteSseEvent(res, "done", new
                {
                    chainId = ctx.ChainId,
                    stopReason = "preparer_query",
                    planMode = ctx.PlanMode
                }, cancellationToken);
                return;
            }

            // Preparer ist zufrieden → aufbereiteten Prompt in DB speichern
            if (!string.IsNullOrWhiteSpace(preparerResult.ForwardedPrompt))
            {
                await db.SaveMessage(new MessageRecord
                {
                    ChainId = ctx.ChainId,
                    Iteration = 0,
                    Role = "user",
                    Type = "text",
                    Content = $"[Vorbereitet von {ctx.Preparer.Agent?.Name ?? "Vorbereiter-Agent"}]:\n{preparerResult.ForwardedPrompt}"
                });
            }
            } // !skipPreparer
        }

        // History für den Session-Agent: OHNE Sub-Agent-Messages (saubere History).
        // includeThinking=true für OpenAI-kompatible Provider → Thinking wird als
        // <think>...</think>-Prefix dem Assistant-Text vorangestellt (laut MiniMax-Docs essentiell
        // für Tool-Calling-Reliability / Interleaved Thinking). Anthropic-Pfad braucht das nicht,
        // weil dort thinking-Blöcke nur mit Signatur akzeptiert werden (die wir nicht persistieren).
        var includeThinkingForProvider = !string.IsNullOrEmpty(ctx.Provider);
        var messages = await chatService.BuildHistoryFromDb(sessionId, promptBlocks: ctx.PromptBlocks, includeThinking: includeThinkingForProvider);

        // Bei Choice-Response: toolChoice auf "auto" setzen, damit Claude frei antworten kann
        var effectiveToolChoice = isChoiceResponse ? "auto" : ctx.ToolChoice;

        // Drei separate CancellationTokens:
        // - workCt: Nur Stop-Button (graceful) → für Anthropic API-Calls, Tool-Ausführungen, DB-Writes
        //   Agent-Arbeit läuft bei Client-Disconnect weiter (Client pollt ja)
        // - hardCt: Hard-Stop (Doppel-Klick) → sofortiger Abbruch aus while-Loop
        // - sseCt: Stop-Button + Client-Disconnect → für SSE-Writes an den Browser
        //   Bei Client-Disconnect werden SSE-Writes zu stillen No-Ops
        var tokens = streamingTracker.Start(sessionId, ctx.ChainId);
        var workCt = tokens.StopToken;
        var hardCt = tokens.HardToken;
        using var sseLinkCts = CancellationTokenSource.CreateLinkedTokenSource(cancellationToken, tokens.StopToken);
        var sseCt = sseLinkCts.Token;

        // SSE-Heartbeat: Alle 15s ein Kommentar senden, damit die Verbindung
        // bei langen Tool-Ausführungen nicht von Browser/Netzwerk getrennt wird.
        var heartbeatTask = Task.Run(async () =>
        {
            try
            {
                while (!sseCt.IsCancellationRequested)
                {
                    await Task.Delay(15000, sseCt);
                    await res.WriteAsync(": heartbeat\n\n", sseCt);
                    await res.Body.FlushAsync(sseCt);
                }
            }
            catch (OperationCanceledException) { }
            catch { }
        }, sseCt);

        try
        {
            if (!string.IsNullOrEmpty(ctx.Provider))
            {
                // OpenAI-kompatibler Provider – Key des Users für diesen Agent verwenden
                string openAiKeyOverride = userApiKey?.ApiKey;

                // OpenAI-Provider braucht zwingend einen Key
                if (ctx.Provider == "openai" && string.IsNullOrEmpty(openAiKeyOverride))
                {
                    await SseHelper.TryWriteSse(res, "error", new { error = "Kein API-Key für diesen Agent zugewiesen. Bitte einen OpenAI-Key zuweisen." }, sseCt);
                    await SseHelper.TryWriteSse(res, "done", new { chainId = ctx.ChainId }, sseCt);
                    return;
                }

                await openAi.StreamChat(res, ctx.ModelId, ctx.SystemPromptText, messages, ctx.ChainId, workspace: ctx.Workspace, sessionId: sessionId, agentId: ctx.AgentId, apiKeyOverride: openAiKeyOverride, workCt: workCt, sseCt: sseCt);
            }
            else
            {
                // Standard: Anthropic
                var anthropicToken = userApiKey?.ApiKey;

                // Kein API-Key → Claude CLI als Fallback verwenden
                if (string.IsNullOrEmpty(anthropicToken))
                {
                    // CLI-Berechtigung prüfen
                    var chatUser = await db.GetUser(userId);
                    if (chatUser != null && !chatUser.CanUseCli && !chatUser.IsAdmin)
                    {
                        await SseHelper.TryWriteSse(res, "error", new { error = "Kein API-Key zugewiesen und keine CLI-Berechtigung. Bitte einen Key zuweisen." }, sseCt);
                        await SseHelper.TryWriteSse(res, "done", new { chainId = ctx.ChainId }, sseCt);
                        return;
                    }

                    if (anthropic.IsCliAvailable)
                    {
                        await anthropic.StreamChatViaCli(res, ctx.ModelId, ctx.SystemPrompt, messages, ctx.ChainId, workspace: ctx.Workspace, allowUserChoice: ctx.Agent?.AllowUserChoice ?? true, planMode: body.PlanMode, webSearch: !string.IsNullOrEmpty(ctx.WebSearchVersion), sessionId: sessionId, agentId: ctx.AgentId, remoteMgr: remoteMgr, unityMcpUrl: ctx.Session?.UnityMcpUrl, unityMcpToken: ctx.Session?.UnityMcpToken, useAutoContext: ctx.Agent?.UseAutoContext ?? true, workCt: workCt, sseCt: sseCt);
                    }
                    else
                    {
                        await SseHelper.TryWriteSse(res, "error", new { error = "Kein API-Key zugewiesen und Claude CLI nicht verfügbar. Bitte einen Key zuweisen oder 'claude login' ausführen." }, sseCt);
                        await SseHelper.TryWriteSse(res, "done", new { chainId = ctx.ChainId }, sseCt);
                    }
                    return;
                }

                await anthropic.StreamChat(res, ctx.ModelId, ctx.SystemPrompt, messages, ctx.ChainId, ctx.Workspace, apiToken: anthropicToken, sessionId: sessionId, agentId: ctx.AgentId, userId: userId, useMemory: ctx.UseMemory, webSearchVersion: ctx.WebSearchVersion, maxIterations: ctx.MaxIterations, toolChoice: effectiveToolChoice, review: ctx.Review, finalizer: ctx.Finalizer, allowUserChoice: ctx.Agent?.AllowUserChoice ?? true, planMode: ctx.PlanMode, useAutoContext: ctx.Agent?.UseAutoContext ?? true, useGitHubSearch: ctx.Agent?.UseGitHubSearch ?? false, useSerperSearch: ctx.Agent?.UseSerperSearch ?? false, workCt: workCt, hardCt: hardCt, sseCt: sseCt, remoteMgr: remoteMgr);
            }
        }
        finally
        {
            streamingTracker.Stop(sessionId);
            // Heartbeat-Task beenden (sseLinkCts wird disposed → Token cancelled)
        }
    }

    private static async Task HandleStop(HttpRequest req, HttpResponse res, StreamingTracker streamingTracker)
    {
        var body = await JsonSerializer.DeserializeAsync<JsonElement>(req.Body);
        var sessionId = body.TryGetProperty("sessionId", out var sid) ? sid.GetString() : null;

        if (string.IsNullOrEmpty(sessionId))
        {
            res.StatusCode = 400;
            await res.WriteAsJsonAsync(new { error = "sessionId required" });
            return;
        }

        // Erster Klick = "graceful" (laufende Iteration fertig), zweiter = "hard" (sofort)
        var stopLevel = streamingTracker.Cancel(sessionId);
        await res.WriteAsJsonAsync(new { stopped = stopLevel != null, level = stopLevel ?? "none" });
    }
}