feat: Finalizer-Agent – dritter Agent der nach Worker + Review einschreitet

Backend:
- SessionRecord + DB-Migration: finalizer_agent_id auf sessions
- ChatContext.FinalizerContext (Model, SystemPrompt, Tools, Workspace)
- ChatService: FinalizerContext aufbauen (analog zu ReviewContext)
- AnthropicService: RunFinalizer() nach Review-Phase (oder direkt nach Worker)
- RunReviewAgentChat() refactored zu generischem RunAgentChat()
  mit agentType-Parameter ('review' | 'finalizer') für SSE-Events und DB-Tags
- SSE-Events: finalizer_start, finalizer_text, finalizer_tool_use, finalizer_done

Frontend:
- Session-Page: Finalizer-Agent-Picker (im 'Weitere Einstellungen' Bereich)
- Chat: SSE-Events für Finalizer-Phase verarbeiten
- UI: addReviewBubble/finalizeReviewBubble unterstützen agentType 'finalizer'
  mit eigener Farbe (success-green) und Flag-Checkered-Icon
- renderMarkdown: [Finalizer: AgentName] Tags parsen
- History: Finalizer-Messages als .finalizer-bubble rendern
- Shared-Page: Finalizer-Messages korrekt erkennen und stylen
- Dashboard: Finalizer-Badge (flag-checkered) auf Session-Cards

Services/AnthropicService.cs

@@ -48,6 +48,7 @@ public class AnthropicService
         int maxIterations = 100,
         string toolChoice = "any",
         ReviewContext review = null,
+        FinalizerContext finalizer = null,
         CancellationToken cancellationToken = default)
     {
         var sw = System.Diagnostics.Stopwatch.StartNew();
@@ -202,6 +203,13 @@ public class AnthropicService
                         }
                     }
 
+                    // ── Finalizer-Agent Phase ──
+                    if (finalizer != null && stopReason == "end_turn")
+                    {
+                        var finalizerIterations = await RunFinalizer(response, finalizer, messages, chainId, workspace, sessionId, usage, iterations, cancellationToken);
+                        iterations += finalizerIterations;
+                    }
+
                     var finalStopReason = string.IsNullOrEmpty(stopReason) ? "unexpected_end" : stopReason;
                     await FinalizeChain(chain, usage, iterations, finalStopReason, sw, streamErrorInfo);
                     await TryWriteSse(response, "done", new
@@ -311,9 +319,12 @@ public class AnthropicService
         var reviewMessages = BuildReviewMessages(workerMessages);
 
         // Review-Agent iteriert mit eigenen Tools
-        var (reviewText, reviewIterations) = await RunReviewAgentChat(
-            response, review, reviewMessages, chainId, workspace, sessionId,
-            usage, baseIterations, cancellationToken);
+        var (reviewText, reviewIterations) = await RunAgentChat(
+            response, review.ModelId, review.SystemPrompt, review.AgentId,
+            workspace, review.UseMemory, review.WebSearchVersion,
+            review.ToolChoice, reviewMessages, chainId, sessionId, usage,
+            baseIterations, "review", review.Agent?.Name ?? "Review-Agent",
+            cancellationToken);
 
         var result = new ReviewResult { ReviewIterationCount = reviewIterations };
 
@@ -344,6 +355,47 @@ public class AnthropicService
         return result;
     }
 
+    // ── Finalizer-Agent Logik ──
+
+    /// <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,
+        FinalizerContext finalizer,
+        List<object> workerMessages,
+        string chainId,
+        string workspace,
+        string sessionId,
+        TokenUsage usage,
+        int baseIterations,
+        CancellationToken cancellationToken)
+    {
+        await TryWriteSse(response, "finalizer_start", new
+        {
+            agentId = finalizer.AgentId,
+            agentName = finalizer.Agent?.Name ?? "Finalizer-Agent"
+        }, cancellationToken);
+
+        // Finalizer bekommt die Konversation als Kontext (gleiche Methode wie Review)
+        var finalizerMessages = BuildReviewMessages(workerMessages);
+
+        var (text, iterations) = await RunAgentChat(
+            response, finalizer.ModelId, finalizer.SystemPrompt, finalizer.AgentId,
+            finalizer.Workspace, finalizer.UseMemory, finalizer.WebSearchVersion,
+            finalizer.ToolChoice, finalizerMessages, chainId, sessionId, usage,
+            baseIterations, "finalizer", finalizer.Agent?.Name ?? "Finalizer-Agent",
+            cancellationToken);
+
+        await TryWriteSse(response, "finalizer_done", new
+        {
+            agentName = finalizer.Agent?.Name ?? "Finalizer-Agent"
+        }, cancellationToken);
+
+        return iterations;
+    }
+
     /// <summary>
     /// Baut die Messages für den Review-Agent auf.
     /// Der Review-Agent bekommt die komplette Worker-Konversation als eine User-Message.
@@ -411,23 +463,32 @@ public class AnthropicService
     }
 
     /// <summary>
-    /// Führt den Review-Agent Chat aus (non-streaming für den Agent, aber SSE-Events für den Client).
-    /// Gibt den finalen Text und die Anzahl der Iterationen zurück.
+    /// Generischer Agent-Chat (non-streaming für den Agent, aber SSE-Events für den Client).
+    /// Wird für Review-Agent und Finalizer-Agent verwendet.
+    /// agentType steuert SSE-Event-Prefixe und DB-Message-Tags ("review" oder "finalizer").
     /// </summary>
-    private async Task<(string text, int iterations)> RunReviewAgentChat(
+    private async Task<(string text, int iterations)> RunAgentChat(
         HttpResponse response,
-        ReviewContext review,
+        string modelId,
+        List<object> systemPrompt,
+        string agentId,
+        string agentWorkspace,
+        bool useMemory,
+        string webSearchVersion,
+        string agentToolChoice,
         List<object> messages,
         string chainId,
-        string workspace,
         string sessionId,
         TokenUsage usage,
         int baseIterations,
+        string agentType,
+        string agentName,
         CancellationToken cancellationToken)
     {
         int iterations = 0;
         var finalText = new StringBuilder();
-        int maxToolIterations = 20; // Review-Agent bekommt max 20 Tool-Iterationen
+        int maxToolIterations = 20;
+        var tag = agentType == "finalizer" ? "Finalizer" : "Review";
 
         while (iterations < maxToolIterations)
         {
@@ -436,19 +497,19 @@ public class AnthropicService
 
             var payload = new Dictionary<string, object>
             {
-                ["model"] = review.ModelId,
+                ["model"] = modelId,
                 ["max_tokens"] = 8192,
                 ["messages"] = messages,
                 ["stream"] = false
             };
 
-            if (review.SystemPrompt != null && review.SystemPrompt.Count > 0)
-                payload["system"] = review.SystemPrompt;
+            if (systemPrompt != null && systemPrompt.Count > 0)
+                payload["system"] = systemPrompt;
 
-            var hasWorkspace = !string.IsNullOrEmpty(workspace);
-            var includeMemory = review.UseMemory && _memoryService != null && !string.IsNullOrEmpty(review.AgentId);
+            var hasWorkspace = !string.IsNullOrEmpty(agentWorkspace);
+            var includeMemory = useMemory && _memoryService != null && !string.IsNullOrEmpty(agentId);
             var hasTools = hasWorkspace || includeMemory;
-            var hasWebSearch = !string.IsNullOrEmpty(review.WebSearchVersion);
+            var hasWebSearch = !string.IsNullOrEmpty(webSearchVersion);
 
             if (hasTools || hasWebSearch)
             {
@@ -456,9 +517,9 @@ public class AnthropicService
                 if (hasTools)
                     tools.AddRange(_toolService.GetToolDefinitions(includeFileTools: hasWorkspace, includeMemory: includeMemory));
                 if (hasWebSearch)
-                    tools.Add(new { type = review.WebSearchVersion, name = "web_search", max_uses = 5 });
+                    tools.Add(new { type = webSearchVersion, name = "web_search", max_uses = 5 });
                 payload["tools"] = tools;
-                var effectiveChoice = (review.ToolChoice == "any" && iterations == 1) ? "any" : "auto";
+                var effectiveChoice = (agentToolChoice == "any" && iterations == 1) ? "any" : "auto";
                 payload["tool_choice"] = new { type = effectiveChoice };
             }
 
@@ -475,12 +536,11 @@ public class AnthropicService
 
             if (!httpResp.IsSuccessStatusCode)
             {
-                return ($"Review-Agent API Fehler: {(int)httpResp.StatusCode}: {respBody}", iterations);
+                return ($"{tag}-Agent API Fehler: {(int)httpResp.StatusCode}: {respBody}", iterations);
             }
 
             var resp = JsonSerializer.Deserialize<JsonElement>(respBody);
 
-            // Usage tracken (auf den gemeinsamen Zähler)
             if (resp.TryGetProperty("usage", out var u))
             {
                 if (u.TryGetProperty("input_tokens", out var it)) usage.InputTokens += it.GetInt64();
@@ -505,22 +565,20 @@ public class AnthropicService
                     assistantContent.Add(new { type = "text", text });
                     finalText.Append(text);
 
-                    // Review-Text per SSE an den Client streamen
-                    await TryWriteSse(response, "review_text", new
+                    await TryWriteSse(response, $"{agentType}_text", new
                     {
                         text,
-                        agentName = review.Agent?.Name ?? "Review-Agent",
-                        agentId = review.AgentId
+                        agentName,
+                        agentId
                     }, cancellationToken);
 
-                    // In DB speichern (als Review-Message markiert)
                     await _db.SaveMessage(new MessageRecord
                     {
                         ChainId = chainId,
                         Iteration = baseIterations + iterations,
                         Role = "assistant",
                         Type = "text",
-                        Content = $"[Review: {review.Agent?.Name ?? "Review-Agent"}] {text}"
+                        Content = $"[{tag}: {agentName}] {text}"
                     });
                 }
                 else if (blockType == "tool_use")
@@ -531,13 +589,12 @@ public class AnthropicService
 
                     assistantContent.Add(new { type = "tool_use", id = toolId, name = toolName, input = toolInput });
 
-                    // Tool-Use per SSE
-                    await TryWriteSse(response, "review_tool_use", new
+                    await TryWriteSse(response, $"{agentType}_tool_use", new
                     {
                         toolName,
                         toolId,
                         input = toolInput.ToString(),
-                        agentName = review.Agent?.Name ?? "Review-Agent"
+                        agentName
                     }, cancellationToken);
 
                     await _db.SaveMessage(new MessageRecord
@@ -553,7 +610,7 @@ public class AnthropicService
 
                     string toolResult;
                     bool isError = false;
-                    try { toolResult = await _toolService.ExecuteTool(toolName, toolInput, workspace, _memoryService, review.AgentId, cancellationToken); }
+                    try { toolResult = await _toolService.ExecuteTool(toolName, toolInput, agentWorkspace, _memoryService, agentId, cancellationToken); }
                     catch (Exception ex) { toolResult = $"Fehler: {ex.Message}"; isError = true; }
 
                     const int MaxToolResultChars = 20_000;