Agent: Max Output Tokens + Thinking deaktivieren einstellbar

Zwei neue Agent-Einstellungen (Frontend, DB-Migration additiv):
- MaxOutputTokens: hartes Ausgabelimit. CLI: CLAUDE_CODE_MAX_OUTPUT_TOKENS,
  API: max_tokens (statt hart 16384). 0 = aus.
- DisableThinking: CLI: CLAUDE_CODE_DISABLE_THINKING=1 + MAX_THINKING_TOKENS=0;
  entfernt CLAUDE_CODE_EFFORT_LEVEL (Explizit-Aus schlaegt Effort).
  API-Pfad ohne Aequivalent (Hint im UI).

Die einzigen zwei Parameter, die z.ai/GLM nachweislich honoriert
(Sampling-Parameter wie temperature/top_p werden dort still verworfen).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

ARCHITECTURE.md

@@ -863,6 +863,8 @@ Ein **Agent** definiert die Persönlichkeit und Fähigkeiten der AI:
 | `AllowUserChoice` | `ask_user_choice`-Tool verfügbar (Default: true)                 |
 | `TaskInstruction` | Aufgabe/Anweisung – optionaler Text, wird als `## Anweisungen`-Block in System-Prompt eingefügt |
 | `Effort`          | Reasoning-Effort-Level für Claude. `""` (=high, Default), `low`, `medium`, `high`, `xhigh` (nur Opus 4.7+), `max` (nur Opus 4.6+/Sonnet 4.6). API: als `output_config.effort` gesendet. CLI: via Env-Var `CLAUDE_CODE_EFFORT_LEVEL`. Bei OpenAI-Providern ignoriert; bei z.ai/GLM nur ab **GLM 5.2** durchgereicht (sonst entfernt). Ultracode hebt es auf `xhigh` an. |
+| `MaxOutputTokens` | Hartes Limit der Ausgabelänge in Tokens. `0` = aus (Modell-Default bzw. API-Default 16384). CLI: via Env-Var `CLAUDE_CODE_MAX_OUTPUT_TOKENS`. API: als `max_tokens` im Payload. Wirkt auch über z.ai/GLM. |
+| `DisableThinking` | Reasoning/Thinking komplett aus. CLI: `CLAUDE_CODE_DISABLE_THINKING=1` + `MAX_THINKING_TOKENS=0`, und `CLAUDE_CODE_EFFORT_LEVEL` wird entfernt (Explizit-Aus schlägt Effort). API-Pfad: kein Äquivalent → still ignoriert. |
 | `IsPublic`        | Andere User können den Agent in ihren Sessions verwenden (aber nicht bearbeiten) |
 
 ### Model-Auflösung

Endpoints/AgentEndpoint.cs

@@ -72,6 +72,8 @@ public static class AgentEndpoint
                 a.AllowUserChoice,
                 a.TaskInstruction,
                 a.Effort,
+                a.MaxOutputTokens,
+                a.DisableThinking,
                 a.AvatarScale,
                 a.AvatarOffsetX,
                 a.AvatarOffsetY,
@@ -87,6 +89,18 @@ public static class AgentEndpoint
         return results;
     }
 
+    /// <summary>
+    /// Max-Output-Tokens robust parsen: Zahl, Zahl-als-String oder leer/null (= 0 = aus).
+    /// Negative Werte werden auf 0 geklemmt; das Frontend schickt bei leerem Feld null.
+    /// </summary>
+    private static int ParseMaxOutputTokens(JsonElement el)
+    {
+        int val = 0;
+        if (el.ValueKind == JsonValueKind.Number) el.TryGetInt32(out val);
+        else if (el.ValueKind == JsonValueKind.String) int.TryParse(el.GetString(), out val);
+        return val > 0 ? val : 0;
+    }
+
     private static async Task HandleGet(HttpContext ctx, DatabaseService db)
     {
         var agentId = ctx.Request.RouteValues["agentId"]?.ToString();
@@ -139,6 +153,8 @@ public static class AgentEndpoint
             AllowUserChoice = !body.TryGetProperty("allowUserChoice", out var auc) || auc.GetBoolean(),
             TaskInstruction = body.TryGetProperty("taskInstruction", out var ti) ? ti.GetString() : null,
             Effort = body.TryGetProperty("effort", out var ef) ? ef.GetString() ?? "" : "",
+            MaxOutputTokens = body.TryGetProperty("maxOutputTokens", out var mot) ? ParseMaxOutputTokens(mot) : 0,
+            DisableThinking = body.TryGetProperty("disableThinking", out var dt) && dt.ValueKind == JsonValueKind.True,
             AvatarScale = body.TryGetProperty("avatarScale", out var asc) ? asc.GetDouble() : 1.0,
             AvatarOffsetX = body.TryGetProperty("avatarOffsetX", out var aox) ? aox.GetDouble() : 0,
             AvatarOffsetY = body.TryGetProperty("avatarOffsetY", out var aoy) ? aoy.GetDouble() : 0,
@@ -187,6 +203,8 @@ public static class AgentEndpoint
         if (body.TryGetProperty("allowUserChoice", out var auc)) agent.AllowUserChoice = auc.GetBoolean();
         if (body.TryGetProperty("taskInstruction", out var ti)) agent.TaskInstruction = ti.GetString();
         if (body.TryGetProperty("effort", out var ef)) agent.Effort = ef.GetString() ?? "";
+        if (body.TryGetProperty("maxOutputTokens", out var mot)) agent.MaxOutputTokens = ParseMaxOutputTokens(mot);
+        if (body.TryGetProperty("disableThinking", out var dt)) agent.DisableThinking = dt.ValueKind == JsonValueKind.True;
         if (body.TryGetProperty("avatarScale", out var asc)) agent.AvatarScale = asc.GetDouble();
         if (body.TryGetProperty("avatarOffsetX", out var aox)) agent.AvatarOffsetX = aox.GetDouble();
         if (body.TryGetProperty("avatarOffsetY", out var aoy)) agent.AvatarOffsetY = aoy.GetDouble();

Endpoints/ChatEndpoint.cs

@@ -294,7 +294,7 @@ public static class ChatEndpoint
 
                     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, ultracode: body.Ultracode, webSearch: !string.IsNullOrEmpty(ctx.WebSearchVersion), sessionId: sessionId, agentId: ctx.AgentId, remoteMgr: remoteMgr, mcpServers: ctx.Session?.McpServers, useAutoContext: ctx.Agent?.UseAutoContext ?? true, effort: ctx.Effort, seedUsage: panelUsage, workCt: workCt, sseCt: sseCt);
+                        await anthropic.StreamChatViaCli(res, ctx.ModelId, ctx.SystemPrompt, messages, ctx.ChainId, workspace: ctx.Workspace, allowUserChoice: ctx.Agent?.AllowUserChoice ?? true, planMode: body.PlanMode, ultracode: body.Ultracode, webSearch: !string.IsNullOrEmpty(ctx.WebSearchVersion), sessionId: sessionId, agentId: ctx.AgentId, remoteMgr: remoteMgr, mcpServers: ctx.Session?.McpServers, useAutoContext: ctx.Agent?.UseAutoContext ?? true, effort: ctx.Effort, maxOutputTokens: ctx.MaxOutputTokens, disableThinking: ctx.DisableThinking, seedUsage: panelUsage, workCt: workCt, sseCt: sseCt);
                     }
                     else
                     {
@@ -304,7 +304,7 @@ public static class ChatEndpoint
                     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, effort: ctx.Effort, seedUsage: panelUsage, workCt: workCt, hardCt: hardCt, sseCt: sseCt, remoteMgr: remoteMgr);
+                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, effort: ctx.Effort, maxOutputTokens: ctx.MaxOutputTokens, seedUsage: panelUsage, workCt: workCt, hardCt: hardCt, sseCt: sseCt, remoteMgr: remoteMgr);
             }
         }
         finally

Models/AgentRecord.cs

@@ -30,6 +30,18 @@ public class AgentRecord
     /// Bei OpenAI-Providern (MiniMax/Nvidia/etc.) wird das Feld stillschweigend ignoriert.
     /// </summary>
     public string Effort { get; set; } = "";
+    /// <summary>
+    /// Hartes Limit für die Ausgabelänge (Tokens). 0 = aus (Modell-Default).
+    /// CLI-Pfad: wird als CLAUDE_CODE_MAX_OUTPUT_TOKENS gesetzt. API-Pfad: als max_tokens gesendet.
+    /// Wirkt auch bei z.ai/GLM (dort einer der wenigen Parameter, die tatsächlich honoriert werden).
+    /// </summary>
+    public int MaxOutputTokens { get; set; }
+    /// <summary>
+    /// Reasoning/Thinking komplett abschalten (Modell spielt/antwortet statt zu analysieren).
+    /// CLI-Pfad: CLAUDE_CODE_DISABLE_THINKING=1 + MAX_THINKING_TOKENS=0.
+    /// API-Pfad: kein Äquivalent (dort steuert Effort) → wird still ignoriert.
+    /// </summary>
+    public bool DisableThinking { get; set; }
     public double AvatarScale { get; set; } = 1.0;
     public double AvatarOffsetX { get; set; }
     public double AvatarOffsetY { get; set; }

Models/ChatContext.cs

@@ -18,6 +18,8 @@ public class ChatContext
     public bool UseMemory { get; set; }
     public string WebSearchVersion { get; set; }
     public string Effort { get; set; }
+    public int MaxOutputTokens { get; set; }   // 0 = aus (Modell-Default)
+    public bool DisableThinking { get; set; }  // nur CLI-Pfad wirksam
     public int MaxIterations { get; set; }
     public AgentRecord Agent { get; set; }
     public SessionRecord Session { get; set; }

Services/AnthropicService.Cli.cs

@@ -38,6 +38,8 @@ public partial class AnthropicService
         string mcpServers = null,
         bool useAutoContext = true,
         string effort = null,
+        int maxOutputTokens = 0,
+        bool disableThinking = false,
         TokenUsage seedUsage = null,
         CancellationToken workCt = default,
         CancellationToken sseCt = default)
@@ -377,6 +379,8 @@ public partial class AnthropicService
                         mcpConfigPath: mcpConfigPath,
                         mcpServerNames: mcpServerList.Select(s => s.Name).ToList(),
                         effort: effort,
+                        maxOutputTokens: maxOutputTokens,
+                        disableThinking: disableThinking,
                         sessionId: sessionId,
                         chainId: chainId,
                         cancellationToken: segCts.Token

Services/AnthropicService.Streaming.cs

@@ -76,6 +76,7 @@ public partial class AnthropicService
         bool useGitHubSearch = false,
         bool useSerperSearch = false,
         string effort = null,
+        int maxOutputTokens = 0,
         CancellationToken workCt = default,
         CancellationToken hardCt = default,
         CancellationToken sseCt = default,
@@ -136,7 +137,8 @@ public partial class AnthropicService
                 var payload = new Dictionary<string, object>
                 {
                     ["model"] = modelId,
-                    ["max_tokens"] = 16384,
+                    // Agent-Einstellung "Max Output Tokens" (0 = aus → bisheriger Default 16384).
+                    ["max_tokens"] = maxOutputTokens > 0 ? maxOutputTokens : 16384,
                     ["messages"] = messages,
                     ["stream"] = true
                 };

Services/ChatService.cs

@@ -252,6 +252,8 @@ Für welche Plattform soll die App entwickelt werden? Web, Mobile oder Desktop?
             UseMemory = agent.UseMemory,
             WebSearchVersion = agent.WebSearchVersion,
             Effort = agent.Effort,
+            MaxOutputTokens = agent.MaxOutputTokens,
+            DisableThinking = agent.DisableThinking,
             MaxIterations = maxIterations,
             Agent = agent,
             Session = session,

Services/ClaudeCliService.cs

@@ -136,7 +136,7 @@ public class ClaudeCliService
         catch { /* Telemetrie darf den Stream nie stören */ }
     }
 
-    private ProcessStartInfo CreateCliStartInfo(List<string> args, bool redirectInput = false, string effort = null, string workingDirectory = null, string modelId = null)
+    private ProcessStartInfo CreateCliStartInfo(List<string> args, bool redirectInput = false, string effort = null, string workingDirectory = null, string modelId = null, int maxOutputTokens = 0, bool disableThinking = false)
     {
         var psi = new ProcessStartInfo
         {
@@ -168,6 +168,23 @@ public class ClaudeCliService
         if (!string.IsNullOrEmpty(effort) && effort != "high")
             psi.Environment["CLAUDE_CODE_EFFORT_LEVEL"] = effort;
 
+        // Hartes Ausgabelimit (Agent-Einstellung, 0 = aus). Bremst "Wall of Text".
+        // Wirkt auch über die z.ai/Anthropic-Route – dort einer der wenigen Parameter,
+        // die tatsächlich honoriert werden (Sampling-Parameter wirft z.ai still weg).
+        if (maxOutputTokens > 0)
+            psi.Environment["CLAUDE_CODE_MAX_OUTPUT_TOKENS"] = maxOutputTokens.ToString();
+
+        // Thinking/Reasoning komplett abschalten (Agent-Einstellung). MAX_THINKING_TOKENS=0
+        // zusätzlich, damit auch Backends ohne den Disable-Schalter kein Budget bekommen.
+        if (disableThinking)
+        {
+            psi.Environment["CLAUDE_CODE_DISABLE_THINKING"] = "1";
+            psi.Environment["MAX_THINKING_TOKENS"] = "0";
+            // Effort steuert genau das Reasoning-Budget, das hier abgeschaltet wird –
+            // beides gleichzeitig zu setzen wäre widersprüchlich. Explizit-Aus gewinnt.
+            psi.Environment.Remove("CLAUDE_CODE_EFFORT_LEVEL");
+        }
+
         // z.ai / GLM: das Anthropic-kompatible z.ai-Backend NUR für diesen Prozess aktivieren.
         // Da gott pro Request einen eigenen CLI-Prozess startet, gilt das ausschließlich für
         // diesen Lauf – das Claude-OAuth-Login aller anderen Requests bleibt unberührt.
@@ -324,6 +341,8 @@ public class ClaudeCliService
         string mcpConfigPath = null,
         List<string> mcpServerNames = null,
         string effort = null,
+        int maxOutputTokens = 0,
+        bool disableThinking = false,
         string sessionId = null,
         string chainId = null,
         CancellationToken cancellationToken = default)
@@ -489,7 +508,7 @@ public class ClaudeCliService
 
         try
         {
-            return await RunCliStreamingAsync(args, prompt, onEvent, cancellationToken, effortToUse, workspace, modelId, sessionId, chainId);
+            return await RunCliStreamingAsync(args, prompt, onEvent, cancellationToken, effortToUse, workspace, modelId, sessionId, chainId, maxOutputTokens, disableThinking);
         }
         finally
         {
@@ -510,7 +529,9 @@ public class ClaudeCliService
         string workspace = null,
         string modelId = null,
         string sessionId = null,
-        string chainId = null)
+        string chainId = null,
+        int maxOutputTokens = 0,
+        bool disableThinking = false)
     {
         var result = new CliChatResult();
         var fullOutput = new StringBuilder();
@@ -530,7 +551,7 @@ public class ClaudeCliService
 
         try
         {
-            var startInfo = CreateCliStartInfo(args, redirectInput: true, effort: effort, workingDirectory: workspace, modelId: modelId);
+            var startInfo = CreateCliStartInfo(args, redirectInput: true, effort: effort, workingDirectory: workspace, modelId: modelId, maxOutputTokens: maxOutputTokens, disableThinking: disableThinking);
 
             using var process = new Process { StartInfo = startInfo };
 

Services/DatabaseService.Agents.cs

@@ -19,8 +19,8 @@ public partial class DatabaseService
         agent.CreatedAt = now;
         agent.UpdatedAt = now;
         await conn.ExecuteAsync(@"
-            INSERT INTO agents (agent_id, user_id, name, avatar, avatar_small, description, model, tool_choice, custom_commands, auto_commit, use_soul, use_identity, use_user_info, use_memory, use_auto_context, use_github_search, use_serper_search, use_web_search, web_search_version, allow_user_choice, task_instruction, effort, avatar_scale, avatar_offset_x, avatar_offset_y, is_public, is_deleted, created_at, updated_at)
-            VALUES (@AgentId, @UserId, @Name, @Avatar, @AvatarSmall, @Description, @Model, @ToolChoice, @CustomCommands, @AutoCommit, @UseSoul, @UseIdentity, @UseUserInfo, @UseMemory, @UseAutoContext, @UseGitHubSearch, @UseSerperSearch, @UseWebSearch, @WebSearchVersion, @AllowUserChoice, @TaskInstruction, @Effort, @AvatarScale, @AvatarOffsetX, @AvatarOffsetY, @IsPublic, 0, @CreatedAt, @UpdatedAt)
+            INSERT INTO agents (agent_id, user_id, name, avatar, avatar_small, description, model, tool_choice, custom_commands, auto_commit, use_soul, use_identity, use_user_info, use_memory, use_auto_context, use_github_search, use_serper_search, use_web_search, web_search_version, allow_user_choice, task_instruction, effort, max_output_tokens, disable_thinking, avatar_scale, avatar_offset_x, avatar_offset_y, is_public, is_deleted, created_at, updated_at)
+            VALUES (@AgentId, @UserId, @Name, @Avatar, @AvatarSmall, @Description, @Model, @ToolChoice, @CustomCommands, @AutoCommit, @UseSoul, @UseIdentity, @UseUserInfo, @UseMemory, @UseAutoContext, @UseGitHubSearch, @UseSerperSearch, @UseWebSearch, @WebSearchVersion, @AllowUserChoice, @TaskInstruction, @Effort, @MaxOutputTokens, @DisableThinking, @AvatarScale, @AvatarOffsetX, @AvatarOffsetY, @IsPublic, 0, @CreatedAt, @UpdatedAt)
         ", agent);
         return agent;
     }
@@ -56,6 +56,7 @@ public partial class DatabaseService
                 custom_commands = @CustomCommands, auto_commit = @AutoCommit,
                 use_soul = @UseSoul, use_identity = @UseIdentity, use_user_info = @UseUserInfo, use_memory = @UseMemory, use_auto_context = @UseAutoContext, use_github_search = @UseGitHubSearch, use_serper_search = @UseSerperSearch, use_web_search = @UseWebSearch, web_search_version = @WebSearchVersion,
                 allow_user_choice = @AllowUserChoice, task_instruction = @TaskInstruction, effort = @Effort,
+                max_output_tokens = @MaxOutputTokens, disable_thinking = @DisableThinking,
                 avatar_scale = @AvatarScale, avatar_offset_x = @AvatarOffsetX, avatar_offset_y = @AvatarOffsetY,
                 is_public = @IsPublic, updated_at = @UpdatedAt
             WHERE agent_id = @AgentId AND user_id = @UserId

Services/DatabaseService.cs

@@ -201,6 +201,8 @@ public partial class DatabaseService
         TryAddColumn(conn, "messages", "tag", "TEXT");
         TryAddColumn(conn, "agents", "effort", "TEXT NOT NULL DEFAULT ''");
         TryAddColumn(conn, "chains", "effort", "TEXT NOT NULL DEFAULT ''");
+        TryAddColumn(conn, "agents", "max_output_tokens", "INTEGER NOT NULL DEFAULT 0");
+        TryAddColumn(conn, "agents", "disable_thinking", "INTEGER NOT NULL DEFAULT 0");
         TryAddColumn(conn, "sessions", "multi_model_mode", "TEXT");
         TryAddColumn(conn, "sessions", "panel_agent_ids", "TEXT");
         TryAddColumn(conn, "sessions", "moa_layers", "INTEGER NOT NULL DEFAULT 2");

wwwroot/index.html

@@ -262,6 +262,12 @@
           <option value="xhigh">Extra – noch tiefer (Opus 4.7+ · GLM 5.2 → maximales Reasoning)</option>
           <option value="max">Max – maximal viele Tokens (Opus 4.6+/Sonnet 4.6+ · GLM 5.2 → maximales Reasoning)</option>
         </select>
+        <label>Max Output Tokens <span class="checkbox-hint">(leer = Modell-Default · hartes Limit gegen Wall-of-Text)</span></label>
+        <input type="number" id="agentMaxOutputTokens" min="1" step="1" placeholder="z. B. 350 – leer lassen für Standard">
+        <label class="checkbox-label">
+          <input type="checkbox" id="agentDisableThinking">
+          Thinking deaktivieren <span class="checkbox-hint">(kein Reasoning – Modell antwortet direkt statt zu analysieren · übersteuert den Effort · nur CLI-Pfad: Claude/GLM)</span>
+        </label>
         <div class="more">
           <a>Weitere Einstellungen</a>
           <div hidden>

wwwroot/js/pages/agent-page.js

@@ -88,6 +88,8 @@ registerPage('agent', {
     const allowUserChoiceEl = document.getElementById('agentAllowUserChoice');
     const autoCommitEl = document.getElementById('agentAutoCommit');
     const effortEl = document.getElementById('agentEffort');
+    const maxOutputTokensEl = document.getElementById('agentMaxOutputTokens');
+    const disableThinkingEl = document.getElementById('agentDisableThinking');
     const useSoulEl = document.getElementById('agentUseSoul');
     const useIdentityEl = document.getElementById('agentUseIdentity');
     const useUserInfoEl = document.getElementById('agentUseUserInfo');
@@ -141,6 +143,8 @@ registerPage('agent', {
       _agentTaskMde.value(agent.taskInstruction || '');
       autoCommitEl.value = agent.autoCommit || '';
       effortEl.value = agent.effort || '';
+      maxOutputTokensEl.value = agent.maxOutputTokens > 0 ? agent.maxOutputTokens : '';
+      disableThinkingEl.checked = agent.disableThinking || false;
       useSoulEl.checked = agent.useSoul !== false;
       useIdentityEl.checked = agent.useIdentity !== false;
       useUserInfoEl.checked = agent.useUserInfo !== false;
@@ -186,6 +190,9 @@ registerPage('agent', {
         allowUserChoice: allowUserChoiceEl.checked,
         autoCommit: autoCommitEl.value,
         effort: effortEl.value,
+        // Leer / 0 / ungültig → 0 = aus (Modell-Default)
+        maxOutputTokens: Math.max(0, parseInt(maxOutputTokensEl.value, 10) || 0),
+        disableThinking: disableThinkingEl.checked,
         useSoul: useSoulEl.checked,
         useIdentity: useIdentityEl.checked,
         useUserInfo: useUserInfoEl.checked,