diff --git a/Endpoints/ChatEndpoint.cs b/Endpoints/ChatEndpoint.cs
index c7356f5..9aea217 100644
--- a/Endpoints/ChatEndpoint.cs
+++ b/Endpoints/ChatEndpoint.cs
@@ -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, useAutoContext: ctx.Agent?.UseAutoContext ?? true, workCt: workCt, sseCt: sseCt);
+                await openAi.StreamChat(res, ctx.ModelId, ctx.SystemPromptText, messages, ctx.ChainId, workspace: ctx.Workspace, sessionId: sessionId, agentId: ctx.AgentId, maxIterations: ctx.MaxIterations, apiKeyOverride: openAiKeyOverride, useAutoContext: ctx.Agent?.UseAutoContext ?? true, workCt: workCt, sseCt: sseCt);
             }
             else
             {
diff --git a/Endpoints/SessionEndpoint.cs b/Endpoints/SessionEndpoint.cs
index 1d37577..7e1ad16 100644
--- a/Endpoints/SessionEndpoint.cs
+++ b/Endpoints/SessionEndpoint.cs
@@ -347,7 +347,11 @@ public static class SessionEndpoint
             }
 
             if (provider == "minimax")
-                payload["extra_body"] = new { reasoning_split = true };
+            {
+                payload["temperature"] = 1.0;
+                payload["top_p"] = 0.95;
+                payload["parallel_tool_calls"] = false;
+            }
 
             if (provider == "nvidia")
             {
diff --git a/Services/OpenAiService.cs b/Services/OpenAiService.cs
index 1aadba1..1e46379 100644
--- a/Services/OpenAiService.cs
+++ b/Services/OpenAiService.cs
@@ -370,6 +370,9 @@ public class OpenAiService
                 var client = CreateClient(providerConfig.ApiKey);
                 // MiniMax-spezifisch: höheres Output-Budget (Thinking-Modelle brauchen viel)
                 var maxTokens = provider == "minimax" ? 32768 : 16384;
+                // Bei Auto-Continue: tool_choice = "required" erzwingt einen Tool-Call,
+                // statt dass das Modell erneut nur eine Ankündigung als Text schreibt.
+                var toolChoice = autoContinueCount > 0 ? "required" : "auto";
                 var payload = new Dictionary<string, object>
                 {
                     ["model"] = modelId,
@@ -378,20 +381,23 @@ public class OpenAiService
                     ["max_tokens"] = maxTokens,
                     ["stream_options"] = new { include_usage = true },
                     ["tools"] = _openAiToolDefs,
-                    ["tool_choice"] = "auto"
+                    ["tool_choice"] = toolChoice
                 };
 
                 if (provider == "nvidia" && IsNvidiaThinkingModel(modelId))
                     payload["chat_template_kwargs"] = new { thinking = true };
 
-                // MiniMax-spezifisch: reasoning_split → Thinking landet in delta.reasoning_details
-                // statt als <think>...</think> inline in delta.content (siehe MiniMax-OpenAI-Compat-Docs).
-                // Sampling-Parameter: temperature=1.0, top_p=0.95 (MiniMax-Empfehlung für Agentic Performance).
+                // MiniMax-spezifisch:
+                // - temperature=1.0, top_p=0.95: MiniMax-Empfehlung für agentic Performance
+                // - parallel_tool_calls=false: Modell muss nur einen Tool-Call planen statt mehrere parallel,
+                //   was bei komplexen Multi-Step-Tasks die Zuverlässigkeit erhöht
+                // Hinweis: reasoning_split=true wurde entfernt – wird von der gehosteten MiniMax-API ignoriert
+                // (Thinking kommt trotzdem als inline <think>...</think> in delta.content)
                 if (provider == "minimax")
                 {
-                    payload["extra_body"] = new { reasoning_split = true };
                     payload["temperature"] = 1.0;
                     payload["top_p"] = 0.95;
+                    payload["parallel_tool_calls"] = false;
                 }
 
                 var jsonPayload = JsonSerializer.Serialize(payload);
@@ -857,8 +863,9 @@ public class OpenAiService
                         if (!string.IsNullOrEmpty(continueContent))
                             apiMessages.Add(new { role = "assistant", content = continueContent });
 
-                        // Synthetische User-Message injizieren
-                        apiMessages.Add(new { role = "user", content = "Continue." });
+                        // Synthetische User-Message injizieren – direktiv, nicht nur "Continue."
+                        // Damit das Modell den Tool-Call tatsächlich ausführt statt nur erneut anzukündigen.
+                        apiMessages.Add(new { role = "user", content = "You announced an action but didn't execute it. Call the appropriate tool now to carry out what you described. Do not repeat the announcement." });
 
                         // SSE-Event an Frontend senden
                         await SseHelper.TryWriteSse(response, "auto_continue", new
@@ -885,7 +892,10 @@ public class OpenAiService
                     return;
                 }
 
-                // Tool-Calls ausführen
+                // Tool-Calls ausführen → autoContinueCount zurücksetzen, damit tool_choice
+                // wieder "auto" wird (Modell darf wieder frei zwischen Text und Tool-Call wählen)
+                autoContinueCount = 0;
+
                 // Assistant-Message mit tool_calls in die History einfügen
                 var toolCallsForHistory = new List<object>();
                 foreach (var (_, (tcId, tcName, tcArgs, tcTs)) in toolCallAccumulators)
@@ -1150,7 +1160,11 @@ public class OpenAiService
                     payload["chat_template_kwargs"] = new { thinking = true };
 
                 if (provider == "minimax")
-                    payload["extra_body"] = new { reasoning_split = true };
+                {
+                    payload["temperature"] = 1.0;
+                    payload["top_p"] = 0.95;
+                    payload["parallel_tool_calls"] = false;
+                }
 
                 var jsonPayload = JsonSerializer.Serialize(payload);
                 var httpContent = new StringContent(jsonPayload, Encoding.UTF8, "application/json");
@@ -1513,9 +1527,74 @@ public class OpenAiService
             }
             apiMessages.Add(converted);
         }
+
+        // Thinking aus älteren Turns strippen (Kontext-Spar-Maßnahme):
+        // MiniMax/DeepSeek-R1/QwQ brauchen ihr eigenes Thinking in der History,
+        // aber nur die letzten paar Turns sind relevant – ältere Thinking-Blöcke
+        // blähen den Kontext auf und verursachen Tool-Call-Degradation ab ~20k Tokens.
+        StripOldThinking(apiMessages, keepRecentCount: 4);
+
         return apiMessages;
     }
 
+    /// <summary>
+    /// Entfernt <think>...</think>-Blöcke aus älteren Assistant-Messages,
+    /// behält sie aber in den letzten <paramref name="keepRecentCount"/> Turns.
+    /// Reduziert Kontext-Inflation bei Thinking-Modellen (MiniMax-M2, DeepSeek-R1, etc.).
+    /// </summary>
+    private static void StripOldThinking(List<object> apiMessages, int keepRecentCount)
+    {
+        // Finde alle Indizes von Assistant-Messages mit <think>-Blöcken
+        var assistantIndices = new List<int>();
+        for (int i = 0; i < apiMessages.Count; i++)
+        {
+            var json = JsonSerializer.Serialize(apiMessages[i]);
+            if (json.Contains("\"role\":\"assistant\"") && json.Contains("<think>"))
+                assistantIndices.Add(i);
+        }
+
+        // Nur die ältesten strippen, die neuesten keepRecentCount behalten
+        var stripCount = assistantIndices.Count - keepRecentCount;
+        if (stripCount <= 0) return;
+
+        for (int j = 0; j < stripCount; j++)
+        {
+            var idx = assistantIndices[j];
+            var json = JsonSerializer.Serialize(apiMessages[idx]);
+            var el = JsonSerializer.Deserialize<JsonElement>(json);
+
+            if (!el.TryGetProperty("content", out var contentProp) || contentProp.ValueKind != JsonValueKind.String)
+                continue;
+
+            var content = contentProp.GetString() ?? "";
+            // <think>...\n</think>\n\n entfernen (greedy innerhalb eines Blocks)
+            var stripped = System.Text.RegularExpressions.Regex.Replace(
+                content, @"<think>[\s\S]*?</think>\s*", "", System.Text.RegularExpressions.RegexOptions.None);
+            stripped = stripped.TrimStart();
+
+            if (string.IsNullOrWhiteSpace(stripped))
+            {
+                // Kein sichtbarer Text übrig nach Strip → Message trotzdem behalten falls tool_calls dran sind
+                if (json.Contains("\"tool_calls\""))
+                    stripped = "";
+                else
+                    continue; // Komplett leer → überspringen
+            }
+
+            // Neue Message mit gestripptem Content erstellen
+            var dict = JsonSerializer.Deserialize<Dictionary<string, JsonElement>>(json);
+            var newDict = new Dictionary<string, object>();
+            foreach (var kv in dict)
+            {
+                if (kv.Key == "content")
+                    newDict["content"] = stripped;
+                else
+                    newDict[kv.Key] = kv.Value;
+            }
+            apiMessages[idx] = newDict;
+        }
+    }
+
     private Task FinalizeChain(ChainRecord chain, TokenUsage usage, int iterations, string stopReason, System.Diagnostics.Stopwatch sw)
         => SseHelper.FinalizeChain(_db, chain, usage, iterations, stopReason, sw);