feat(chat): per-conversation model selection (#150)

mateclaw-server/src/main/java/vip/mate/agent/AgentGraphBuilder.java

@@ -126,9 +126,41 @@ public class AgentGraphBuilder {
     }
 
     /**
-     * 根据 AgentEntity 构建完整的 Agent 实例
+     * 根据 AgentEntity 构建完整的 Agent 实例（沿用 Agent / 全局默认模型）。
      */
     public BaseAgent build(AgentEntity entity) {
+        return build(entity, null, null);
+    }
+
+    /**
+     * Resolve the model the runtime should use, honouring the precedence
+     * <em>conversation pin &gt; Agent model override &gt; global default</em>.
+     * A conversation pin that no longer resolves to an enabled model (the model
+     * was disabled or deleted after it was picked) silently degrades to the
+     * Agent / global default rather than failing the chat.
+     */
+    private ModelConfigEntity resolveRuntimeBaseModel(String modelProvider, String modelName,
+                                                      String agentModelName) {
+        if (modelProvider != null && !modelProvider.isBlank()
+                && modelName != null && !modelName.isBlank()) {
+            ModelConfigEntity pinned = modelConfigService.findEnabledModel(modelProvider, modelName);
+            if (pinned != null) {
+                return pinned;
+            }
+            log.info("Conversation model pin {}/{} is no longer an enabled model — "
+                    + "falling back to the Agent / global default", modelProvider, modelName);
+        }
+        return modelConfigService.resolveModel(agentModelName);
+    }
+
+    /**
+     * 根据 AgentEntity 构建完整的 Agent 实例。
+     *
+     * <p>{@code modelProvider} / {@code modelName} carry an optional
+     * per-conversation model pin; when both are blank the build falls back to
+     * the Agent's model override, then the global default.</p>
+     */
+    public BaseAgent build(AgentEntity entity, String modelProvider, String modelName) {
         AgentToolSet toolSet = toolRegistry.getEnabledToolSet();
 
         // 过滤掉 denied 工具，使模型完全看不到它们（防止 prompt injection 利用 schema）
@@ -142,17 +174,16 @@ public class AgentGraphBuilder {
         Set<String> boundTools = agentBindingService.getEffectiveToolNames(entity.getId());
         toolSet = toolSet.withAllowedToolsOnly(boundTools); // null = 全局默认
 
-        // RFC-090 §9.2 调整 C — pick a primary model that satisfies
+        // Resolve the base model with the precedence: per-conversation pin >
-        // the agent's bound-skill requires-model. Falls back to the
+        // per-Agent model override > global default. resolveRuntimeBaseModel
-        // global default when no preferred provider satisfies, so the
+        // looks up enabled-only models and silently degrades an unmatched pin /
-        // existing "no default model" error path stays intact.
+        // override to the global default, preserving the legacy behaviour for
-        // Honor per-Agent model override when set.
+        // Agents and conversations without an explicit choice.
-        // resolveModel() looks up entity.modelName in enabled-only models;
+        // providerRouter.selectPrimary below may still swap this for a model
-        // null / blank / unmatched silently fall back to getDefaultModel(),
+        // that satisfies a bound skill's requires-model constraint.
-        // preserving the legacy behavior for Agents without an override.
         ModelConfigEntity globalDefault;
         try {
-            globalDefault = modelConfigService.resolveModel(entity.getModelName());
+            globalDefault = resolveRuntimeBaseModel(modelProvider, modelName, entity.getModelName());
         } catch (Exception e) {
             throw new MateClawException("err.agent.no_default_model", "无法构建 Agent：请先在「设置 → 模型」中配置并启用默认模型");
         }

mateclaw-server/src/main/java/vip/mate/agent/AgentService.java

@@ -21,6 +21,8 @@ import vip.mate.memory.MemoryProperties;
 import vip.mate.memory.lifecycle.MemoryLifecycleMediator;
 import vip.mate.memory.lifecycle.TurnContext;
 import vip.mate.memory.service.MemoryRecallTracker;
+import vip.mate.workspace.conversation.model.ConversationEntity;
+import vip.mate.workspace.conversation.repository.ConversationMapper;
 
 import java.util.List;
 import java.util.Map;
@@ -46,14 +48,22 @@ public class AgentService {
     private final MemoryRecallTracker memoryRecallTracker;
     private final MemoryLifecycleMediator lifecycleMediator;
     private final MemoryProperties memoryProperties;
+    /** Read-only lookup of a conversation's pinned model. Mapper (not service)
+     *  to keep this a leaf dependency with no risk of a bean cycle. */
+    private final ConversationMapper conversationMapper;
 
     /** Field-injected publisher for agent_lifecycle trigger events; the
      *  trigger module's bridge listens and forwards into ingest. */
     @Autowired(required = false)
     private ApplicationEventPublisher events;
 
-    /** 运行时 Agent 实例缓存（agentId -> BaseAgent） */
+    /**
-    private final Map<Long, BaseAgent> agentInstances = new ConcurrentHashMap<>();
+     * Runtime Agent instance cache. Keyed first by agentId, then by a model
+     * key, so a conversation that pins a non-default model gets its own graph
+     * variant instead of mutating the one every other conversation shares.
+     * The model key is {@code ""} for the Agent / global-default model.
+     */
+    private final Map<Long, Map<String, BaseAgent>> agentInstances = new ConcurrentHashMap<>();
 
     // ==================== CRUD ====================
 
@@ -217,7 +227,7 @@ public class AgentService {
      */
     public String chat(Long agentId, String message, String conversationId, ChatOrigin origin) {
         memoryRecallTracker.trackRecalls(agentId, message);
-        BaseAgent agent = getOrBuildAgent(agentId);
+        BaseAgent agent = getOrBuildAgentForConversation(agentId, conversationId);
         ChatOriginHolder.set(origin != null ? origin : ChatOrigin.EMPTY);
         try {
             return withLifecycleSync(agentId, message, conversationId,
@@ -233,7 +243,7 @@ public class AgentService {
 
     public Flux<String> chatStream(Long agentId, String message, String conversationId, ChatOrigin origin) {
         memoryRecallTracker.trackRecalls(agentId, message);
-        BaseAgent agent = getOrBuildAgent(agentId);
+        BaseAgent agent = getOrBuildAgentForConversation(agentId, conversationId);
         // Capture the origin into a request-scoped holder; cleared on Flux
         // termination so the next reactive subscriber doesn't inherit stale state.
         ChatOrigin captured = origin != null ? origin : ChatOrigin.EMPTY;
@@ -269,7 +279,7 @@ public class AgentService {
                                                    String requesterId, String thinkingLevel,
                                                    ChatOrigin origin) {
         memoryRecallTracker.trackRecalls(agentId, message);
-        BaseAgent agent = getOrBuildAgent(agentId);
+        BaseAgent agent = getOrBuildAgentForConversation(agentId, conversationId);
 
         // 设置请求级思考深度（通过 ThreadLocal 传递到 StateGraph 执行）
         if (thinkingLevel != null && !thinkingLevel.isBlank()) {
@@ -315,7 +325,7 @@ public class AgentService {
 
     public String execute(Long agentId, String goal, String conversationId, ChatOrigin origin) {
         memoryRecallTracker.trackRecalls(agentId, goal);
-        BaseAgent agent = getOrBuildAgent(agentId);
+        BaseAgent agent = getOrBuildAgentForConversation(agentId, conversationId);
         ChatOriginHolder.set(origin != null ? origin : ChatOrigin.EMPTY);
         try {
             return withLifecycleSync(agentId, goal, conversationId,
@@ -342,7 +352,7 @@ public class AgentService {
     public String chatWithReplay(Long agentId, String userMessage, String conversationId,
                                   String toolCallPayload, ChatOrigin origin) {
         memoryRecallTracker.trackRecalls(agentId, userMessage);
-        BaseAgent agent = getOrBuildAgent(agentId);
+        BaseAgent agent = getOrBuildAgentForConversation(agentId, conversationId);
         ChatOriginHolder.set(origin != null ? origin : ChatOrigin.EMPTY);
         try {
             return withLifecycleSync(agentId, userMessage, conversationId,
@@ -370,7 +380,7 @@ public class AgentService {
                                                    String toolCallPayload, String requesterId,
                                                    ChatOrigin origin) {
         memoryRecallTracker.trackRecalls(agentId, userMessage);
-        BaseAgent agent = getOrBuildAgent(agentId);
+        BaseAgent agent = getOrBuildAgentForConversation(agentId, conversationId);
         ChatOrigin captured = origin != null ? origin : ChatOrigin.EMPTY;
         return Flux.defer(() -> {
                     ChatOriginHolder.set(captured);
@@ -383,8 +393,20 @@ public class AgentService {
     }
 
     public AgentState getAgentState(Long agentId) {
-        BaseAgent agent = agentInstances.get(agentId);
+        Map<String, BaseAgent> variants = agentInstances.get(agentId);
-        return agent != null ? agent.getState() : AgentState.IDLE;
+        if (variants == null || variants.isEmpty()) {
+            return AgentState.IDLE;
+        }
+        // An Agent may have several cached graph variants (one per pinned
+        // model). Report the first non-IDLE state so a turn running on any
+        // variant stays visible.
+        for (BaseAgent agent : variants.values()) {
+            AgentState state = agent.getState();
+            if (state != AgentState.IDLE) {
+                return state;
+            }
+        }
+        return AgentState.IDLE;
     }
 
     // ==================== 缓存管理 ====================
@@ -473,14 +495,44 @@ public class AgentService {
 
     // ==================== 内部方法 ====================
 
-    private BaseAgent getOrBuildAgent(Long agentId) {
+    /**
-        return agentInstances.computeIfAbsent(agentId, id -> {
+     * Resolve (and cache) the Agent graph for a conversation, honouring the
-            AgentEntity entity = getAgent(id);
+     * conversation's pinned model. Conversations with no pin — IM channels,
-            if (!Boolean.TRUE.equals(entity.getEnabled())) {
+     * cron, sub-tasks, or rows not yet created — resolve to the shared Agent /
-                throw new MateClawException("err.agent.disabled", "Agent 已禁用: " + entity.getName());
+     * global-default graph.
+     */
+    private BaseAgent getOrBuildAgentForConversation(Long agentId, String conversationId) {
+        String provider = null;
+        String modelName = null;
+        if (conversationId != null && !conversationId.isBlank()) {
+            ConversationEntity conv = conversationMapper.selectOne(
+                    new LambdaQueryWrapper<ConversationEntity>()
+                            .eq(ConversationEntity::getConversationId, conversationId));
+            if (conv != null) {
+                provider = conv.getModelProvider();
+                modelName = conv.getModelName();
             }
-            return agentGraphBuilder.build(entity);
+        }
-        });
+        return getOrBuildAgent(agentId, provider, modelName);
+    }
+
+    private BaseAgent getOrBuildAgent(Long agentId) {
+        return getOrBuildAgent(agentId, null, null);
+    }
+
+    private BaseAgent getOrBuildAgent(Long agentId, String modelProvider, String modelName) {
+        boolean pinned = modelProvider != null && !modelProvider.isBlank()
+                && modelName != null && !modelName.isBlank();
+        String modelKey = pinned ? modelProvider + "::" + modelName : "";
+        return agentInstances
+                .computeIfAbsent(agentId, id -> new ConcurrentHashMap<>())
+                .computeIfAbsent(modelKey, key -> {
+                    AgentEntity entity = getAgent(agentId);
+                    if (!Boolean.TRUE.equals(entity.getEnabled())) {
+                        throw new MateClawException("err.agent.disabled", "Agent 已禁用: " + entity.getName());
+                    }
+                    return agentGraphBuilder.build(entity, modelProvider, modelName);
+                });
     }
 
     // ==================== StreamDelta ====================

mateclaw-server/src/main/java/vip/mate/channel/web/ChatController.java

@@ -520,6 +520,11 @@ public class ChatController {
             AtomicBoolean finalized = new AtomicBoolean(false);
             try {
                 conversationService.getOrCreateConversation(conversationId, agentId, username, workspaceId);
+                // Pin the model the user picked for this conversation so later
+                // turns (and the runtime model resolver) honour it independently
+                // of every other conversation.
+                conversationService.updateConversationModel(conversationId,
+                        request.getModelProvider(), request.getModelName());
                 List<MessageContentPart> requestParts = normalizeRequestParts(request);
                 String promptText = buildPromptText(message, requestParts);
                 conversationService.saveMessage(conversationId, "user", message, requestParts);
@@ -1151,6 +1156,14 @@ public class ChatController {
         private Long lastEventId;
         /** 思考深度：off / low / medium / high / max，null 表示跟随 Agent 默认 */
         private String thinkingLevel;
+        /**
+         * Provider id of the model the user picked for this conversation.
+         * Paired with {@link #modelName}; null means "no per-conversation
+         * override — use the agent / global default".
+         */
+        private String modelProvider;
+        /** Model id the user picked for this conversation. See {@link #modelProvider}. */
+        private String modelName;
     }
 
     /**

mateclaw-server/src/main/java/vip/mate/llm/service/ModelConfigService.java

@@ -323,6 +323,24 @@ public class ModelConfigService {
         return getDefaultModel();
     }
 
+    /**
+     * Resolve an enabled model by its exact (provider, modelName) pair. Unlike
+     * {@link #resolveModel(String)} this does NOT fall back to the default —
+     * it returns {@code null} when nothing matches, leaving the fallback
+     * decision to the caller. Used to honour a per-conversation model pin while
+     * still degrading gracefully when that model was later disabled or deleted.
+     */
+    public ModelConfigEntity findEnabledModel(String provider, String modelName) {
+        if (!StringUtils.hasText(provider) || !StringUtils.hasText(modelName)) {
+            return null;
+        }
+        return modelConfigMapper.selectOne(new LambdaQueryWrapper<ModelConfigEntity>()
+                .eq(ModelConfigEntity::getProvider, provider)
+                .eq(ModelConfigEntity::getModelName, modelName)
+                .eq(ModelConfigEntity::getEnabled, true)
+                .last("LIMIT 1"));
+    }
+
     private void validateModel(ModelConfigEntity entity, Long currentId) {
         if (!StringUtils.hasText(entity.getName())) {
             throw new MateClawException("err.llm.name_required", "模型名称不能为空");

mateclaw-server/src/main/java/vip/mate/workspace/conversation/ConversationService.java

@@ -349,6 +349,32 @@ public class ConversationService {
         }
     }
 
+    /**
+     * Pin the model a conversation uses. A blank provider or model id is a
+     * no-op (no override supplied — the conversation keeps inheriting the
+     * agent / global default). The write is skipped when the stored value
+     * already matches, so persisting the same model on every turn costs only
+     * a SELECT.
+     */
+    @Transactional
+    public void updateConversationModel(String conversationId, String modelProvider, String modelName) {
+        if (modelProvider == null || modelProvider.isBlank()
+                || modelName == null || modelName.isBlank()) {
+            return;
+        }
+        ConversationEntity conv = conversationMapper.selectOne(new LambdaQueryWrapper<ConversationEntity>()
+                .eq(ConversationEntity::getConversationId, conversationId));
+        if (conv == null) {
+            return;
+        }
+        if (modelProvider.equals(conv.getModelProvider()) && modelName.equals(conv.getModelName())) {
+            return;
+        }
+        conv.setModelProvider(modelProvider);
+        conv.setModelName(modelName);
+        conversationMapper.updateById(conv);
+    }
+
     /**
      * Persist an assistant placeholder marker only when the last message is a
      * user turn (i.e., the assistant never got to reply). Used by the admin

mateclaw-server/src/main/java/vip/mate/workspace/conversation/model/ConversationEntity.java

@@ -51,6 +51,16 @@ public class ConversationEntity {
     /** Pin flag: 0 = normal, 1 = pinned to the top of the sidebar list */
     private Integer pinned;
 
+    /**
+     * Provider id of the model this conversation is pinned to. NULL means
+     * "inherit" — fall back to the agent's model override, then the global
+     * default. Paired with {@link #modelName}.
+     */
+    private String modelProvider;
+
+    /** Model id this conversation is pinned to. See {@link #modelProvider}. */
+    private String modelName;
+
     @TableField(fill = FieldFill.INSERT)
     private LocalDateTime createTime;
 

mateclaw-server/src/main/java/vip/mate/workspace/conversation/vo/ConversationVO.java

@@ -66,6 +66,8 @@ public class ConversationVO extends ConversationEntity {
         vo.setLastActiveTime(entity.getLastActiveTime());
         vo.setWorkspaceId(entity.getWorkspaceId());
         vo.setPinned(entity.getPinned() != null ? entity.getPinned() : 0);
+        vo.setModelProvider(entity.getModelProvider());
+        vo.setModelName(entity.getModelName());
         vo.setCreateTime(entity.getCreateTime());
         vo.setUpdateTime(entity.getUpdateTime());
         // 补充关联字段

mateclaw-server/src/main/resources/db/migration/h2/V116__conversation_model.sql

@@ -0,0 +1,7 @@
+-- Per-conversation model selection. Each conversation pins the LLM it uses, so
+-- switching the model in one chat no longer changes every other conversation.
+-- NULL means "inherit": fall back to the agent's model override, then to the
+-- global default model.
+
+ALTER TABLE mate_conversation ADD COLUMN IF NOT EXISTS model_provider VARCHAR(64);
+ALTER TABLE mate_conversation ADD COLUMN IF NOT EXISTS model_name VARCHAR(128);

mateclaw-server/src/main/resources/db/migration/mysql/V116__conversation_model.sql

@@ -0,0 +1,29 @@
+-- See the H2 file for context. MySQL 8.0 doesn't support
+-- `ADD COLUMN IF NOT EXISTS`, so the existence check goes through
+-- INFORMATION_SCHEMA + a prepared statement.
+
+SET @col_exists := (
+    SELECT COUNT(*) FROM information_schema.COLUMNS
+    WHERE TABLE_SCHEMA = DATABASE()
+      AND TABLE_NAME = 'mate_conversation'
+      AND COLUMN_NAME = 'model_provider'
+);
+SET @ddl := IF(@col_exists = 0,
+    'ALTER TABLE mate_conversation ADD COLUMN model_provider VARCHAR(64)',
+    'SELECT 1');
+PREPARE stmt FROM @ddl;
+EXECUTE stmt;
+DEALLOCATE PREPARE stmt;
+
+SET @col_exists := (
+    SELECT COUNT(*) FROM information_schema.COLUMNS
+    WHERE TABLE_SCHEMA = DATABASE()
+      AND TABLE_NAME = 'mate_conversation'
+      AND COLUMN_NAME = 'model_name'
+);
+SET @ddl := IF(@col_exists = 0,
+    'ALTER TABLE mate_conversation ADD COLUMN model_name VARCHAR(128)',
+    'SELECT 1');
+PREPARE stmt FROM @ddl;
+EXECUTE stmt;
+DEALLOCATE PREPARE stmt;

mateclaw-server/src/test/java/vip/mate/memory/lifecycle/LifecycleRecallCountIT.java

@@ -15,6 +15,7 @@ import vip.mate.agent.repository.AgentMapper;
 import vip.mate.memory.MemoryProperties;
 import vip.mate.memory.service.MemoryRecallTracker;
 import vip.mate.memory.spi.MemoryManager;
+import vip.mate.workspace.conversation.repository.ConversationMapper;
 
 import static org.mockito.ArgumentMatchers.any;
 import static org.mockito.ArgumentMatchers.eq;
@@ -41,6 +42,7 @@ class LifecycleRecallCountIT {
     @Mock private MemoryManager memoryManager;
     @Mock private ApplicationEventPublisher eventPublisher;
     @Mock private BaseAgent mockAgent;
+    @Mock private ConversationMapper conversationMapper;
 
     private MemoryProperties props;
     private AgentService agentService;
@@ -50,14 +52,14 @@ class LifecycleRecallCountIT {
         props = new MemoryProperties();
         MemoryLifecycleMediator mediator = new MemoryLifecycleMediator(memoryManager, eventPublisher);
         agentService = new AgentService(agentMapper, agentGraphBuilder,
-                memoryRecallTracker, mediator, props);
+                memoryRecallTracker, mediator, props, conversationMapper);
 
         // Stub agent resolution (lenient for structural-only tests)
         AgentEntity entity = new AgentEntity();
         entity.setId(1L);
         entity.setEnabled(true);
         lenient().when(agentMapper.selectById(1L)).thenReturn(entity);
-        lenient().when(agentGraphBuilder.build(any(AgentEntity.class))).thenReturn(mockAgent);
+        lenient().when(agentGraphBuilder.build(any(AgentEntity.class), any(), any())).thenReturn(mockAgent);
         lenient().when(mockAgent.chat(any(), any())).thenReturn("reply");
     }
 

mateclaw-ui/src/composables/chat/useChat.ts

@@ -147,6 +147,10 @@ export interface SendMessageOptions {
   contentParts?: MessageContentPart[]
   /** Thinking depth: off / low / medium / high / max */
   thinkingLevel?: string
+  /** Provider id of the model picked for this conversation. */
+  modelProvider?: string
+  /** Model id picked for this conversation. Paired with modelProvider. */
+  modelName?: string
 }
 
 export function useChat(options: UseChatOptions): UseChatReturn {
@@ -1644,6 +1648,12 @@ export function useChat(options: UseChatOptions): UseChatReturn {
       if (options.thinkingLevel) {
         body.thinkingLevel = options.thinkingLevel
       }
+      // Per-conversation model: the backend pins it onto the conversation row
+      // so switching the model here never leaks into other conversations.
+      if (options.modelProvider && options.modelName) {
+        body.modelProvider = options.modelProvider
+        body.modelName = options.modelName
+      }
       await stream.connect(body)
     } catch (e) {
       error.value = e instanceof Error ? e : new Error(String(e))

mateclaw-ui/src/types/index.ts

@@ -66,6 +66,10 @@ export interface Conversation {
   streamStatus?: 'idle' | 'running'
   source?: string
   pinned?: number
+  /** Provider id of the model this conversation is pinned to (per-conversation model). */
+  modelProvider?: string
+  /** Model id this conversation is pinned to. Paired with modelProvider. */
+  modelName?: string
   lastActiveTime?: string
   updateTime?: string
   createTime?: string

mateclaw-ui/src/views/ChatConsole.vue

@@ -342,7 +342,11 @@ const providersUnavailable = ref(false)
 // otherwise the model selector trigger would show its 配置模型 fallback even
 // though there IS an active model.
 const enabledModels = ref<ModelConfig[]>([])
+// The model the CURRENT conversation uses. Per-conversation — switching it
+// never leaks into other conversations (see selectModel / applyConversationModel).
 const activeModels = ref<ActiveModelsInfo | null>(null)
+// Global default model — seeds the selector for conversations with no pin yet.
+const globalDefaultModel = ref<{ providerId: string; model: string } | null>(null)
 const pendingAttachments = ref<ChatAttachment[]>([])
 const uploadingAttachment = ref(false)
 
@@ -379,18 +383,30 @@ function onAgentPicked(value: string | number | null) {
   }
 }
 
-async function selectModel(value: string) {
+function selectModel(value: string) {
   const [providerId, model] = value.split('::')
   if (!providerId || !model) return
-  modelSaving.value = true
+  // Per-conversation model: switching here only affects THIS conversation.
-  try {
+  // The backend pins it onto the conversation row when the next message is
-    const res: any = await modelApi.setActive({ providerId, model })
+  // sent (see sendChatMessage payload); we also patch the local list entry so
-    activeModels.value = res.data || { activeLlm: { providerId, model } }
+  // re-opening the conversation restores the choice without a round-trip.
-    await loadModelState()
+  activeModels.value = { activeLlm: { providerId, model } }
-  } catch (e) {
+  const conv = conversations.value.find(c => c.conversationId === currentConversationId.value)
-    mcToast.error(t('chat.switchModelFailed'))
+  if (conv) {
-  } finally {
+    conv.modelProvider = providerId
-    modelSaving.value = false
+    conv.modelName = model
+  }
+}
+
+/**
+ * Point the model selector at a conversation's pinned model, or the global
+ * default when the conversation has no pin yet (fresh chat, IM, cron).
+ */
+function applyConversationModel(conv?: Conversation | null) {
+  if (conv?.modelProvider && conv?.modelName) {
+    activeModels.value = { activeLlm: { providerId: conv.modelProvider, model: conv.modelName } }
+  } else if (globalDefaultModel.value) {
+    activeModels.value = { activeLlm: { ...globalDefaultModel.value } }
   }
 }
 
@@ -1013,7 +1029,16 @@ async function loadModelState() {
       modelApi.listEnabled(),
     ])
     defaultModel.value = defaultRes.data || null
-    activeModels.value = activeRes.data || null
+    const ga = activeRes.data?.activeLlm
+    globalDefaultModel.value = ga?.providerId && ga?.model
+      ? { providerId: ga.providerId, model: ga.model }
+      : null
+    // Seed the selector when no conversation has set it yet (fresh chat, or
+    // before a conversation is selected). A conversation that already has a
+    // model keeps it — selectConversation/applyConversationModel own that.
+    if (!activeModels.value && globalDefaultModel.value) {
+      activeModels.value = { activeLlm: { ...globalDefaultModel.value } }
+    }
     enabledModels.value = enabledRes.data || []
   } catch (e) {
     mcToast.error(t('chat.loadModelFailed'))
@@ -1177,6 +1202,8 @@ async function selectConversation(conv: Conversation) {
   }
   currentConversationId.value = conv.conversationId
   selectedAgentId.value = conv.agentId || selectedAgentId.value
+  // Restore this conversation's pinned model into the selector.
+  applyConversationModel(conv)
   // Reset cron placeholder state up front; the immediate fetch below repopulates
   // it for cron conversations so the user doesn't wait up to 4s for the next tick.
   activeCronRuns.value = []
@@ -1308,6 +1335,8 @@ function newConversation() {
   resetForNewConversation()
   currentConversationId.value = `conv_${Date.now()}_${Math.random().toString(36).slice(2, 8)}`
   messages.value = []
+  // A fresh conversation starts on the global default model.
+  applyConversationModel()
 }
 
 // The sidebar performs the delete API call(s) and emits the removed ids.
@@ -1475,6 +1504,8 @@ async function handleSendMessage(content: string) {
       agentId: selectedAgentId.value,
       contentParts,
       thinkingLevel: thinkingLevel.value,
+      modelProvider: activeModels.value?.activeLlm?.providerId,
+      modelName: activeModels.value?.activeLlm?.model,
       attachments: outgoingAttachments.map(a => ({
         type: 'file' as const,
         fileUrl: a.url,