fix(openai): Auto-Context, Tool-Result-Matching, Payload-Preview, leere Bubbles

- Auto-Context (Embedding-Suche) für OpenAI-Provider aktiviert: BuildAutoContextAsync auf internal static refactored und in StreamChat + ChatNonStreaming integriert - Tool-Results mit Role="tool" (OpenAI-Konvention) werden in BuildHistoryFromDb jetzt korrekt erkannt (vorher nur Role="user"/Anthropic) - Payload-Preview zeigt für OpenAI-Provider echtes API-Format (function-Wrapper, tool_choice:"auto", max_tokens, stream_options, extra_body) - Tool-Input (Argumente) wird im SSE tool_end-Event mitgesendet (fehlte im OpenAI-Pfad) - Leere Bubbles durch Whitespace-only Text nach </think>-Extraktion behoben (Trim in DB-Save, BuildHistoryFromDb und Frontend-Guard) - History-Toggle-Bug: var→const in for-Loop (Block-Scoping für Closures), tool_result-Matching per data-tool-call-id statt blindes letztes Element - Non-Streaming max_tokens für MiniMax auf 32768 + extra_body.reasoning_split Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-05-14 11:16:40 +02:00
1 parent ecddbe5f36
commit 1fe04025d3
@@ -240,7 +240,7 @@ public static class ChatEndpoint
                    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);
+                await openAi.StreamChat(res, ctx.ModelId, ctx.SystemPromptText, messages, ctx.ChainId, workspace: ctx.Workspace, sessionId: sessionId, agentId: ctx.AgentId, apiKeyOverride: openAiKeyOverride, useAutoContext: ctx.Agent?.UseAutoContext ?? true, workCt: workCt, sseCt: sseCt);
            }
            else
            {
@@ -317,8 +317,54 @@ public static class SessionEndpoint
        var toolChoice = hasTools || !string.IsNullOrEmpty(agent.WebSearchVersion) ? "any" : null;
        var effectiveChoice = toolChoice == "any" ? "any" : "auto";

-        // Payload zusammenbauen (Anthropic-Format)
-        var payload = new Dictionary<string, object>
+        // OpenAI-kompatible Provider: Payload im tatsächlichen API-Format anzeigen
+        if (previewProvider)
+        {
+            var (provider, _) = OpenAiService.ParseModelString(modelId);
+            var maxTokens = provider == "minimax" ? 32768 : 16384;
+
+            // Messages konvertieren (Anthropic → OpenAI-Format)
+            var openAiMessages = OpenAiService.BuildApiMessages(systemPromptText, messages);
+
+            // Tools konvertieren (input_schema → function/parameters)
+            var openAiTools = tools.Count > 0
+                ? OpenAiService.ConvertTo_openAiToolDefs(tools.ToArray())
+                : new List<object>();
+
+            var payload = new Dictionary<string, object>
+            {
+                ["model"] = modelId,
+                ["max_tokens"] = maxTokens,
+                ["messages"] = openAiMessages,
+                ["stream"] = true,
+                ["stream_options"] = new { include_usage = true }
+            };
+
+            if (openAiTools.Count > 0)
+            {
+                payload["tools"] = openAiTools;
+                payload["tool_choice"] = "auto";
+            }
+
+            if (provider == "minimax")
+                payload["extra_body"] = new { reasoning_split = true };
+
+            if (provider == "nvidia")
+            {
+                var nvidiaModelId = modelId.Contains(":") ? modelId.Split(':')[1] : modelId;
+                if (nvidiaModelId.Contains("kimi-k2", StringComparison.OrdinalIgnoreCase)
+                    || nvidiaModelId.Contains("qwq", StringComparison.OrdinalIgnoreCase)
+                    || nvidiaModelId.Contains("deepseek-r1", StringComparison.OrdinalIgnoreCase))
+                    payload["chat_template_kwargs"] = new { thinking = true };
+            }
+
+            ctx.Response.ContentType = "application/json; charset=utf-8";
+            await ctx.Response.WriteAsync(JsonSerializer.Serialize(payload, new JsonSerializerOptions { WriteIndented = true }));
+            return;
+        }
+
+        // Anthropic-Format Payload
+        var anthropicPayload = new Dictionary<string, object>
        {
            ["model"] = modelId,
            ["max_tokens"] = 16384,
@@ -327,16 +373,16 @@ public static class SessionEndpoint
        };

        if (systemPrompt.Count > 0)
-            payload["system"] = systemPrompt;
+            anthropicPayload["system"] = systemPrompt;

        if (tools.Count > 0)
        {
-            payload["tools"] = tools;
-            payload["tool_choice"] = new { type = effectiveChoice };
+            anthropicPayload["tools"] = tools;
+            anthropicPayload["tool_choice"] = new { type = effectiveChoice };
        }

        ctx.Response.ContentType = "application/json; charset=utf-8";
-        await ctx.Response.WriteAsync(JsonSerializer.Serialize(payload, new JsonSerializerOptions { WriteIndented = true }));
+        await ctx.Response.WriteAsync(JsonSerializer.Serialize(anthropicPayload, new JsonSerializerOptions { WriteIndented = true }));
    }

    private static async Task HandleStreaming(HttpContext ctx, DatabaseService db, StreamingTracker streamingTracker)
@@ -95,7 +95,7 @@ builder.Services.AddSingleton<AnthropicService>(sp =>

 // OpenAI-kompatibler Service (Nvidia etc.)
 builder.Services.AddSingleton<OpenAiService>(sp =>
-    new OpenAiService(sp.GetRequiredService<IHttpClientFactory>(), dbService, sp.GetRequiredService<ToolService>(), serverSettings, dataDir));
+    new OpenAiService(sp.GetRequiredService<IHttpClientFactory>(), dbService, sp.GetRequiredService<ToolService>(), serverSettings, dataDir, vectorSearchService));

 // Telegram Bot (pro Session – Token aus Session-Config)
 builder.Services.AddSingleton<TelegramService>(sp =>
@@ -92,7 +92,7 @@ public partial class AnthropicService
            {
                try
                {
-                    var autoCtx = await BuildAutoContextAsync(messages, chainId, agentId);
+                    var autoCtx = await BuildAutoContextAsync(_vectorSearch, messages, chainId, agentId);
                    if (autoCtx != null)
                    {
                        systemText.AppendLine(autoCtx);
@@ -62,7 +62,7 @@ public partial class AnthropicService
            {
                try
                {
-                    var autoCtx = await BuildAutoContextAsync(messages, chainId, agentId);
+                    var autoCtx = await BuildAutoContextAsync(_vectorSearch, messages, chainId, agentId);
                    if (autoCtx != null)
                    {
                        systemPrompt.Add(new { type = "text", text = autoCtx });
@@ -138,7 +138,7 @@ public partial class AnthropicService
    }

    /// <summary>Extrahiert die letzten N User-Texte für Query-Erweiterung.</summary>
-    private static string ExtractRecentUserTexts(List<object> messages, int count = 3, int maxLength = 500)
+    internal static string ExtractRecentUserTexts(List<object> messages, int count = 3, int maxLength = 500)
    {
        var texts = new List<string>();
        for (int i = messages.Count - 1; i >= 0 && texts.Count < count; i--)
@@ -171,21 +171,21 @@ public partial class AnthropicService

    /// <summary>
    /// Baut den AutoContext-String aus Embedding-Suche.
-    /// Shared zwischen StreamChat und StreamChatViaCli.
+    /// Shared zwischen StreamChat, StreamChatViaCli und OpenAiService.
    /// </summary>
-    private async Task<string> BuildAutoContextAsync(List<object> messages, string chainId, string agentId)
+    internal static async Task<string> BuildAutoContextAsync(VectorSearchService vectorSearch, List<object> messages, string chainId, string agentId)
    {
-        if (_vectorSearch == null || messages.Count == 0) return null;
+        if (vectorSearch == null || messages.Count == 0) return null;

        // Query-Erweiterung: letzte 2-3 User-Messages statt nur die letzte
        var queryText = ExtractRecentUserTexts(messages, count: 3, maxLength: 500);
        if (string.IsNullOrWhiteSpace(queryText)) return null;

-        var contextResults = await _vectorSearch.SearchAsync(queryText, agentId, topN: 5, minScore: 0.35f);
+        var contextResults = await vectorSearch.SearchAsync(queryText, agentId, topN: 5, minScore: 0.35f);
        // Note: threshold lowered from 0.45 to 0.35 because hybrid search is now more precise
        contextResults.RemoveAll(r => r.ChainId == chainId || (r.TextPreview?.Length ?? 0) < 20);

-        var memoryResults = await _vectorSearch.SearchMemoryFilesAsync(queryText, agentId, topN: 3, minScore: 0.35f);
+        var memoryResults = await vectorSearch.SearchMemoryFilesAsync(queryText, agentId, topN: 3, minScore: 0.35f);

        if (contextResults.Count == 0 && memoryResults.Count == 0) return null;

@@ -299,12 +299,12 @@ Für welche Plattform soll die App entwickelt werden? Web, Mobile oder Desktop?
                            ["content"] = new List<object> { new { type = "text", text = msg.Content } }
                        });
                    }
-                    else if (msg.Role == "assistant" && msg.Type == "text" && !string.IsNullOrEmpty(msg.Content))
+                    else if (msg.Role == "assistant" && msg.Type == "text" && !string.IsNullOrWhiteSpace(msg.Content))
                    {
                        history.Add(new Dictionary<string, object>
                        {
                            ["role"] = "assistant",
-                            ["content"] = new List<object> { new { type = "text", text = msg.Content } }
+                            ["content"] = new List<object> { new { type = "text", text = msg.Content.Trim() } }
                        });
                    }
                }
@@ -383,11 +383,14 @@ Für welche Plattform soll die App entwickelt werden? Web, Mobile oder Desktop?
                    }
                    else if (msg.Role == "assistant" && msg.Type == "text")
                    {
-                        if (!string.IsNullOrEmpty(msg.Content))
+                        // Trim: Nach </think>-Extraktion bleibt oft nur "\n" als Text übrig.
+                        // Whitespace-only Text nicht in die History aufnehmen (verursacht leere Bubbles).
+                        var trimmedContent = msg.Content?.Trim();
+                        if (!string.IsNullOrEmpty(trimmedContent))
                        {
                            // Wenn die letzte Message bereits assistant ist (z.B. mit Thinking-Block
                            // aus derselben Iteration), Text dort anhängen statt neue Message zu starten.
-                            AppendToLastMessage(history, "assistant", new { type = "text", text = msg.Content });
+                            AppendToLastMessage(history, "assistant", new { type = "text", text = trimmedContent });
                        }
                    }
                    else if (msg.Role == "assistant" && msg.Type == "tool_use")
@@ -436,8 +439,11 @@ Für welche Plattform soll die App entwickelt werden? Web, Mobile oder Desktop?

                        AppendToLastMessage(history, "assistant", resultEl);
                    }
-                    else if (msg.Role == "user" && msg.Type == "tool_result")
+                    else if ((msg.Role == "user" || msg.Role == "tool") && msg.Type == "tool_result")
                    {
+                        // Anthropic speichert tool_results mit Role="user", OpenAI mit Role="tool".
+                        // Beide werden in Anthropic-Format (user-Message mit tool_result-Block) eingefügt,
+                        // da BuildApiMessages später das Format für OpenAI-Provider konvertiert.
                        var toolResultBlock = new { type = "tool_result", tool_use_id = msg.ToolCallId, content = SseHelper.TruncateToolResult(msg.Content ?? ""), is_error = msg.IsError };
                        AppendToLastMessage(history, "user", toolResultBlock);
                    }
@@ -23,6 +23,7 @@ public class OpenAiService
    private readonly IHttpClientFactory _httpFactory;
    private readonly DatabaseService _db;
    private readonly ToolService _toolService;
+    private readonly VectorSearchService _vectorSearch;

    // Provider-Keys (statisch für ChatService.ResolveModelId)
    public static readonly HashSet<string> ProviderKeys = new() { "nvidia", "pollinations", "gemini", "openai", "minimax" };
@@ -36,7 +37,7 @@ public class OpenAiService
    /// Konvertiert Anthropic-Format Tool-Definitionen (input_schema) in OpenAI-Format (function/parameters).
    /// Single source of truth: ToolService.GetToolDefinitions()
    /// </summary>
-    private static List<object> ConvertTo_openAiToolDefs(object[] anthropicTools)
+    public static List<object> ConvertTo_openAiToolDefs(object[] anthropicTools)
    {
        var result = new List<object>();
        foreach (var tool in anthropicTools)
@@ -86,11 +87,12 @@ public class OpenAiService

    private readonly ServerSettingsService _settings;

-    public OpenAiService(IHttpClientFactory httpFactory, DatabaseService db, ToolService toolService, ServerSettingsService settings, string dataDir)
+    public OpenAiService(IHttpClientFactory httpFactory, DatabaseService db, ToolService toolService, ServerSettingsService settings, string dataDir, VectorSearchService vectorSearch = null)
    {
        _httpFactory = httpFactory;
        _db = db;
        _toolService = toolService;
+        _vectorSearch = vectorSearch;
        _settings = settings;
        _cacheFile = Path.Combine(dataDir, "models_cache.json");
        Providers = new Dictionary<string, (string BaseUrl, string ApiKey)>
@@ -295,6 +297,7 @@ public class OpenAiService
        string agentId = null,
        int maxIterations = 10,
        string apiKeyOverride = null,
+        bool useAutoContext = true,
        CancellationToken workCt = default,
        CancellationToken sseCt = default)
    {
@@ -328,6 +331,32 @@ public class OpenAiService

        try
        {
+            // ── Automatische Kontext-Injektion via Embedding-Suche ──
+            if (useAutoContext)
+            {
+                try
+                {
+                    var autoCtx = await AnthropicService.BuildAutoContextAsync(_vectorSearch, messages, chainId, agentId);
+                    if (autoCtx != null)
+                    {
+                        // Auto-Context an System-Prompt anhängen
+                        systemPromptText = (systemPromptText ?? "") + "\n" + autoCtx;
+                        await SseHelper.TryWriteSse(response, "auto_context", new { context = autoCtx }, sseCt);
+                        await _db.SaveMessage(new MessageRecord
+                        {
+                            ChainId = chainId,
+                            Role = "system",
+                            Type = "auto_context",
+                            Content = autoCtx
+                        });
+                    }
+                }
+                catch (Exception ex)
+                {
+                    Console.WriteLine($"[AutoContext/OpenAI] Error: {ex.Message}");
+                }
+            }
+
            var apiMessages = BuildApiMessages(systemPromptText, messages);

            // Agentic Loop
@@ -786,8 +815,8 @@ public class OpenAiService
                    });
                }

-                // Text in DB speichern
-                var finalText = fullText.ToString();
+                // Text in DB speichern (Whitespace-only nach </think> ignorieren)
+                var finalText = fullText.ToString().Trim();
                if (!string.IsNullOrEmpty(finalText))
                {
                    await _db.SaveMessage(new MessageRecord
@@ -925,6 +954,7 @@ public class OpenAiService
                    {
                        name = tcName,
                        id = tcId,
+                        input = tcArgs.ToString(),
                        result = toolResult,
                        isError
                    }, sseCt);
@@ -1029,7 +1059,32 @@ public class OpenAiService
            Model = ctx.ModelId
        };

-        var apiMessages = BuildApiMessages(ctx.SystemPromptText, messages);
+        // ── Automatische Kontext-Injektion via Embedding-Suche ──
+        var systemPromptText = ctx.SystemPromptText;
+        if (ctx.Agent?.UseAutoContext ?? true)
+        {
+            try
+            {
+                var autoCtx = await AnthropicService.BuildAutoContextAsync(_vectorSearch, messages, ctx.ChainId, ctx.AgentId);
+                if (autoCtx != null)
+                {
+                    systemPromptText = (systemPromptText ?? "") + "\n" + autoCtx;
+                    await _db.SaveMessage(new MessageRecord
+                    {
+                        ChainId = ctx.ChainId,
+                        Role = "system",
+                        Type = "auto_context",
+                        Content = autoCtx
+                    });
+                }
+            }
+            catch (Exception ex)
+            {
+                Console.WriteLine($"[AutoContext/OpenAI-NonStreaming] Error: {ex.Message}");
+            }
+        }
+
+        var apiMessages = BuildApiMessages(systemPromptText, messages);

        try
        {
@@ -1039,12 +1094,13 @@ public class OpenAiService
                onIteration?.Invoke(iterations);

                var client = CreateClient(providerConfig.ApiKey);
+                var maxTokens = provider == "minimax" ? 32768 : 16384;
                var payload = new Dictionary<string, object>
                {
                    ["model"] = modelId,
                    ["messages"] = apiMessages,
                    ["stream"] = false,
-                    ["max_tokens"] = 16384,
+                    ["max_tokens"] = maxTokens,
                    ["tools"] = _openAiToolDefs,
                    ["tool_choice"] = "auto"
                };
@@ -1052,6 +1108,9 @@ public class OpenAiService
                if (provider == "nvidia" && IsNvidiaThinkingModel(modelId))
                    payload["chat_template_kwargs"] = new { thinking = true };

+                if (provider == "minimax")
+                    payload["extra_body"] = new { reasoning_split = true };
+
                var jsonPayload = JsonSerializer.Serialize(payload);
                var httpContent = new StringContent(jsonPayload, Encoding.UTF8, "application/json");

@@ -1364,7 +1423,7 @@ public class OpenAiService
    /// <summary>
    /// Konvertiert Anthropic-Format Messages in OpenAI-kompatibles Format.
    /// </summary>
-    private static List<object> BuildApiMessages(string systemPromptText, List<object> messages)
+    public static List<object> BuildApiMessages(string systemPromptText, List<object> messages)
    {
        var apiMessages = new List<object>();
        if (!string.IsNullOrEmpty(systemPromptText))
@@ -220,7 +220,7 @@ function _renderMessagesIntoFragment(fragment, msgs, imageMessageIds, opts) {
    // ── Grouped Review/Preparer/Finalizer Bubble (tag-based) ──
    if (msg.tag && (msg.tag === 'review' || msg.tag === 'preparer' || msg.tag === 'finalizer')) {
      // Collect all consecutive messages with the same tag
-      var tagGroup = [msg];
+      const tagGroup = [msg];
      while (_msgIdx + 1 < msgs.length && msgs[_msgIdx + 1].tag === msg.tag) {
        _msgIdx++;
        tagGroup.push(msgs[_msgIdx]);
@@ -250,6 +250,8 @@ function _renderMessagesIntoFragment(fragment, msgs, imageMessageIds, opts) {
          _addBubbleActions(div, { timestamp: msg.createdAt, messageId: msg.id, deletedOn: msg.deletedOn, isMarked: msg.isMarked });
          fragment.appendChild(div);
        } else if (msg.role === 'assistant') {
+          // Whitespace-only Content nicht rendern (z.B. "\n" nach </think>-Extraktion)
+          if (!msg.content || !msg.content.trim()) break;
          const div = _createMessageBubble({ role: 'assistant', timestamp: msg.createdAt, messageId: msg.id, isMarked: msg.isMarked, deletedOn: msg.deletedOn, archived });
          div.innerHTML = renderMarkdown(msg.content);
          parseFluentEmojis(div);
@@ -261,12 +263,12 @@ function _renderMessagesIntoFragment(fragment, msgs, imageMessageIds, opts) {
      }
      case 'auto_context': {
        if (msg.content) {
-          var acDiv = document.createElement('div');
+          const acDiv = document.createElement('div');
          acDiv.className = 'tool-box auto-context-box' + (archived ? ' archived' : '');
-          var acHeader = document.createElement('div');
+          const acHeader = document.createElement('div');
          acHeader.className = 'tool-box-header clickable';
          acHeader.innerHTML = '<span class="tool-name">Auto-Context</span>';
-          var acContent = document.createElement('div');
+          const acContent = document.createElement('div');
          acContent.className = 'auto-context-content markdown';
          acContent.innerHTML = renderMarkdown(msg.content.replace(/<\/?recalled_context>/g, '').trim());
          acContent.hidden = true;
@@ -279,12 +281,12 @@ function _renderMessagesIntoFragment(fragment, msgs, imageMessageIds, opts) {
      }
      case 'thinking': {
        if (msg.role === 'assistant' && msg.content) {
-          var thinkDiv = document.createElement('div');
+          const thinkDiv = document.createElement('div');
          thinkDiv.className = 'tool-box thinking-box' + (archived ? ' archived' : '');
-          var thinkHeader = document.createElement('div');
+          const thinkHeader = document.createElement('div');
          thinkHeader.className = 'tool-box-header clickable';
          thinkHeader.innerHTML = '<span class="tool-name">Denkprozess</span>';
-          var thinkContent = document.createElement('div');
+          const thinkContent = document.createElement('div');
          thinkContent.className = 'thinking-content markdown';
          thinkContent.innerHTML = renderMarkdown(msg.content);
          thinkContent.hidden = true;
@@ -301,17 +303,18 @@ function _renderMessagesIntoFragment(fragment, msgs, imageMessageIds, opts) {
          continue;
        }
        if (!isToolVisible(msg.toolName)) continue;
-        var toolDiv = document.createElement('div');
+        const toolDiv = document.createElement('div');
        toolDiv.className = 'tool-box' + (archived ? ' archived' : '');
        toolDiv.dataset.toolName = msg.toolName;
+        if (msg.toolCallId) toolDiv.dataset.toolCallId = msg.toolCallId;
        renderToolBoxContent(toolDiv, msg.toolName, msg.content, null, false);
        fragment.appendChild(toolDiv);
        break;
      case 'web_search_result': {
        if (!isToolVisible('web_search')) continue;
-        var wsSources = _extractWebSearchSources(msg.content);
+        const wsSources = _extractWebSearchSources(msg.content);
        if (wsSources.length === 0) continue;
-        var wsDiv = document.createElement('div');
+        const wsDiv = document.createElement('div');
        wsDiv.className = 'tool-box' + (archived ? ' archived' : '');
        wsDiv.dataset.toolName = 'web_search';
        renderToolBoxContent(wsDiv, 'web_search', null, null, false, wsSources);
@@ -320,10 +323,18 @@ function _renderMessagesIntoFragment(fragment, msgs, imageMessageIds, opts) {
      }
      case 'tool_result': {
        if (msg._choiceHandled) continue;
-        const prevBoxes = fragment.querySelectorAll('.tool-box');
-        const lastToolBox = prevBoxes[prevBoxes.length - 1];
+        // Passendes Tool-Box per toolCallId finden (falls verfügbar),
+        // sonst Fallback auf letztes Tool-Box (ältere DB-Einträge ohne toolCallId)
+        let lastToolBox = null;
+        if (msg.toolCallId) {
+          lastToolBox = fragment.querySelector('.tool-box[data-tool-call-id="' + msg.toolCallId + '"]');
+        }
+        if (!lastToolBox) {
+          const prevBoxes = fragment.querySelectorAll('.tool-box');
+          lastToolBox = prevBoxes[prevBoxes.length - 1];
+        }
        if (lastToolBox) {
-          var existingResult = lastToolBox.querySelector('.tool-result');
+          const existingResult = lastToolBox.querySelector('.tool-result');
          // todo_write: nachträglich die Todo-Liste rendern (sichtbar)
          if (lastToolBox.dataset.toolName === 'todo_write' && !lastToolBox.querySelector('.todo-list')) {
            lastToolBox.appendChild(buildTodoList(msg.content || '', msg.isError));
@@ -339,7 +350,7 @@ function _renderMessagesIntoFragment(fragment, msgs, imageMessageIds, opts) {
            rd.hidden = true;
            lastToolBox.appendChild(rd);
            // Header clickable machen
-            var hdr = lastToolBox.querySelector('.tool-box-header');
+            const hdr = lastToolBox.querySelector('.tool-box-header');
            if (hdr && !hdr.classList.contains('clickable')) {
              hdr.classList.add('clickable');
              hdr.addEventListener('click', function() { rd.hidden = !rd.hidden; });