feat(mind): add chat screen model config dialog

packages/feature_mind/lib/feature_mind.dart

@@ -129,6 +129,8 @@ export 'src/assistant/presentation/screens/prompt_lab_screen.dart';
 
 // Agent chat + model management
 export 'src/agent_chat/application/assistant_model_preferences.dart';
+export 'src/agent_chat/application/chat_model_config_preferences.dart';
+export 'src/agent_chat/domain/models/chat_model_config.dart';
 export 'src/agent_chat/application/assistant_runtime_readiness.dart';
 export 'src/agent_chat/data/services/agent_notification_scheduler.dart';
 export 'src/agent_chat/data/services/assistant_runtime_service.dart';

packages/feature_mind/lib/src/agent_chat/application/chat_model_config_preferences.dart

@@ -0,0 +1,64 @@
+import 'package:flutter_riverpod/legacy.dart';
+import 'package:shared_preferences/shared_preferences.dart';
+
+import '../domain/models/chat_model_config.dart';
+
+const String chatModelConfigMaxTokensKey = 'chat_model_config.max_tokens';
+const String chatModelConfigTopKKey = 'chat_model_config.top_k';
+const String chatModelConfigTopPKey = 'chat_model_config.top_p';
+const String chatModelConfigTemperatureKey = 'chat_model_config.temperature';
+const String chatModelConfigAcceleratorKey = 'chat_model_config.accelerator';
+const String chatModelConfigSystemPromptKey = 'chat_model_config.system_prompt';
+
+final chatModelConfigProvider =
+    StateNotifierProvider<ChatModelConfigNotifier, ChatModelConfig>((ref) {
+      return ChatModelConfigNotifier();
+    });
+
+class ChatModelConfigNotifier extends StateNotifier<ChatModelConfig> {
+  ChatModelConfigNotifier() : super(ChatModelConfig.defaults) {
+    _load();
+  }
+
+  Future<void> _load() async {
+    final prefs = await SharedPreferences.getInstance();
+    state = ChatModelConfig(
+      maxTokens:
+          prefs.getInt(chatModelConfigMaxTokensKey) ??
+          ChatModelConfig.defaults.maxTokens,
+      topK:
+          prefs.getInt(chatModelConfigTopKKey) ?? ChatModelConfig.defaults.topK,
+      topP:
+          prefs.getDouble(chatModelConfigTopPKey) ??
+          ChatModelConfig.defaults.topP,
+      temperature:
+          prefs.getDouble(chatModelConfigTemperatureKey) ??
+          ChatModelConfig.defaults.temperature,
+      accelerator: _acceleratorFromName(
+        prefs.getString(chatModelConfigAcceleratorKey),
+      ),
+      systemPrompt:
+          prefs.getString(chatModelConfigSystemPromptKey) ??
+          ChatModelConfig.defaults.systemPrompt,
+    ).normalized();
+  }
+
+  Future<void> save(ChatModelConfig config) async {
+    final next = config.normalized();
+    state = next;
+    final prefs = await SharedPreferences.getInstance();
+    await prefs.setInt(chatModelConfigMaxTokensKey, next.maxTokens);
+    await prefs.setInt(chatModelConfigTopKKey, next.topK);
+    await prefs.setDouble(chatModelConfigTopPKey, next.topP);
+    await prefs.setDouble(chatModelConfigTemperatureKey, next.temperature);
+    await prefs.setString(chatModelConfigAcceleratorKey, next.accelerator.name);
+    await prefs.setString(chatModelConfigSystemPromptKey, next.systemPrompt);
+  }
+
+  static ChatAccelerator _acceleratorFromName(String? raw) {
+    return ChatAccelerator.values.firstWhere(
+      (value) => value.name == raw,
+      orElse: () => ChatModelConfig.defaults.accelerator,
+    );
+  }
+}

packages/feature_mind/lib/src/agent_chat/data/services/assistant_runtime_service.dart

@@ -692,6 +692,11 @@ class AssistantRuntimeService {
     required String prompt,
     String? systemPrompt,
     String? grammar,
+    int? maxOutputTokens,
+    double? temperature,
+    double? topP,
+    int? topK,
+    bool preferGpu = true,
   }) async {
     await _ensureCheckpointsHydrated();
     lastGenerationStats = null;
@@ -794,6 +799,11 @@ class AssistantRuntimeService {
             package: package,
             prompt: prompt,
             systemPrompt: systemPrompt,
+            maxOutputTokens: maxOutputTokens,
+            temperature: temperature,
+            topP: topP,
+            topK: topK,
+            preferGpu: preferGpu,
             emitRequestTrace: false,
             grammar: constrained ? grammar : null,
           )) {
@@ -835,6 +845,10 @@ class AssistantRuntimeService {
     required String prompt,
     String? systemPrompt,
     int? maxOutputTokens,
+    double? temperature,
+    double? topP,
+    int? topK,
+    bool preferGpu = true,
     GenerationConstraint? constraint,
   }) async* {
     await _ensureCheckpointsHydrated();
@@ -854,6 +868,10 @@ class AssistantRuntimeService {
             prompt: prompt,
             systemPrompt: systemPrompt,
             maxOutputTokens: maxOutputTokens,
+            temperature: temperature,
+            topP: topP,
+            topK: topK,
+            preferGpu: preferGpu,
             assistantPrefill: constraint?.forcedPrefix,
           );
           return;
@@ -863,6 +881,11 @@ class AssistantRuntimeService {
         selectedModelId: runtimeId,
         prompt: constrainedPrompt,
         systemPrompt: systemPrompt,
+        maxOutputTokens: maxOutputTokens,
+        temperature: temperature,
+        topP: topP,
+        topK: topK,
+        preferGpu: preferGpu,
       );
       return;
     }
@@ -914,6 +937,10 @@ class AssistantRuntimeService {
     required String prompt,
     String? systemPrompt,
     int? maxOutputTokens,
+    double? temperature,
+    double? topP,
+    int? topK,
+    bool preferGpu = true,
     bool emitRequestTrace = true,
     String? grammar,
     String? assistantPrefill,
@@ -930,7 +957,7 @@ class AssistantRuntimeService {
         ),
       );
     }
-    await _ensureGgufReady(runtimeId, package);
+    await _ensureGgufReady(runtimeId, package, preferGpu: preferGpu);
     final instructPrompt = formatGgufInstructPrompt(
       prompt: prompt,
       systemPrompt: systemPrompt,
@@ -950,6 +977,9 @@ class AssistantRuntimeService {
     await for (final token in _llamaGguf.generate(
       prompt: instructPrompt,
       maxTokens: maxOutputTokens ?? ggufMaxOutputTokens(package),
+      temperature: temperature ?? 0.7,
+      topP: topP ?? 0.9,
+      topK: topK ?? 40,
       grammar: grammar,
     )) {
       accumulated += token;
@@ -997,8 +1027,9 @@ class AssistantRuntimeService {
 
   Future<void> _ensureGgufReady(
     String runtimeId,
-    OfflineModelInfo package,
-  ) async {
+    OfflineModelInfo package, {
+    bool preferGpu = true,
+  }) async {
     if (!await _llamaGguf.isAvailable()) {
       throw AssistantRuntimeUnavailableException(
         runtimeId,
@@ -1008,6 +1039,7 @@ class AssistantRuntimeService {
     final loaded = await _llamaGguf.loadModelOutcome(
       package,
       contextSize: _effectiveContextLength(package),
+      preferGpu: preferGpu,
     );
     if (loaded.succeeded) return;
     final copy = GgufLoadDiagnostics.describe(

packages/feature_mind/lib/src/agent_chat/domain/models/chat_model_config.dart

@@ -0,0 +1,92 @@
+import 'package:equatable/equatable.dart';
+
+/// CPU vs GPU preference for on-device chat generation.
+enum ChatAccelerator { cpu, gpu }
+
+/// Per-conversation sampling and prompt overrides for the chat screen.
+class ChatModelConfig extends Equatable {
+  const ChatModelConfig({
+    required this.maxTokens,
+    required this.topK,
+    required this.topP,
+    required this.temperature,
+    required this.accelerator,
+    required this.systemPrompt,
+  });
+
+  static const int minMaxTokens = 64;
+  static const int maxMaxTokens = 8192;
+  static const int minTopK = 1;
+  static const int maxTopK = 128;
+  static const double minTopP = 0;
+  static const double maxTopP = 1;
+  static const double minTemperature = 0;
+  static const double maxTemperature = 2;
+
+  /// Defaults match the on-device Gemma gallery config the chat UI copies.
+  static const ChatModelConfig defaults = ChatModelConfig(
+    maxTokens: 4000,
+    topK: 1,
+    topP: 0.95,
+    temperature: 1,
+    accelerator: ChatAccelerator.gpu,
+    systemPrompt: '',
+  );
+
+  final int maxTokens;
+  final int topK;
+  final double topP;
+  final double temperature;
+  final ChatAccelerator accelerator;
+  final String systemPrompt;
+
+  bool get preferGpu => accelerator == ChatAccelerator.gpu;
+
+  ChatModelConfig copyWith({
+    int? maxTokens,
+    int? topK,
+    double? topP,
+    double? temperature,
+    ChatAccelerator? accelerator,
+    String? systemPrompt,
+  }) {
+    return ChatModelConfig(
+      maxTokens: maxTokens ?? this.maxTokens,
+      topK: topK ?? this.topK,
+      topP: topP ?? this.topP,
+      temperature: temperature ?? this.temperature,
+      accelerator: accelerator ?? this.accelerator,
+      systemPrompt: systemPrompt ?? this.systemPrompt,
+    ).normalized();
+  }
+
+  ChatModelConfig normalized() {
+    return ChatModelConfig(
+      maxTokens: maxTokens.clamp(minMaxTokens, maxMaxTokens).toInt(),
+      topK: topK.clamp(minTopK, maxTopK).toInt(),
+      topP: topP.clamp(minTopP, maxTopP).toDouble(),
+      temperature: temperature.clamp(minTemperature, maxTemperature).toDouble(),
+      accelerator: accelerator,
+      systemPrompt: systemPrompt,
+    );
+  }
+
+  /// Prepends a user-authored system prompt without replacing assembled context.
+  String mergeSystemPrompt(String assembled) {
+    final custom = systemPrompt.trim();
+    final existing = assembled.trim();
+    if (custom.isEmpty) return assembled;
+    if (existing.isEmpty) return custom;
+    return '$custom\n\n$assembled';
+  }
+
+  @override
+  List<Object?> get props => [
+    maxTokens,
+    topK,
+    topP,
+    temperature,
+    accelerator,
+    systemPrompt,
+  ];
+}