fix: model picker in chat no longer overwrites the global default model

src/main/hermes.ts

@@ -1109,6 +1109,7 @@ function sendMessageViaApi(
   history?: Array<{ role: string; content: string }>,
   attachments?: Attachment[],
   contextFolder?: string,
+  modelOverride?: string,
 ): ChatHandle {
   const mc = getModelConfig(profile);
   const controller = new AbortController();
@@ -1138,7 +1139,7 @@ function sendMessageViaApi(
 
   const reasoningEffort = reasoningEffortForProfile(profile);
   const bodyObj: Record<string, unknown> = {
-    model: mc.model || "hermes-agent",
+    model: modelOverride || mc.model || "hermes-agent",
     messages,
     stream: true,
     ...(_resumeSessionId ? { session_id: _resumeSessionId } : {}),
@@ -1224,7 +1225,7 @@ function sendMessageViaApi(
   function probeRealError(): void {
     // When streaming returns empty, make a non-streaming request to surface the real error
     const probeBodyObj: Record<string, unknown> = {
-      model: mc.model || "hermes-agent",
+      model: modelOverride || mc.model || "hermes-agent",
       messages: [{ role: "user", content: userContent }],
       stream: false,
     };
@@ -1513,6 +1514,7 @@ function sendMessageViaRuns(
   history?: Array<{ role: string; content: string }>,
   attachments?: Attachment[],
   contextFolder?: string,
+  modelOverride?: string,
 ): ChatHandle {
   const mc = getModelConfig(profile);
   const controller = new AbortController();
@@ -1523,7 +1525,7 @@ function sendMessageViaRuns(
     (headersForAuth.Authorization ? `desk-${Date.now()}-${randomUUID()}` : "");
   const ctxSystem = contextFolderSystemMessage(contextFolder);
   const bodyObj: Record<string, unknown> = {
-    model: mc.model || "hermes-agent",
+    model: modelOverride || mc.model || "hermes-agent",
     input: message,
     conversation_history: apiHistory(history),
   };
@@ -1567,6 +1569,7 @@ function sendMessageViaRuns(
       history,
       attachments,
       contextFolder,
+      modelOverride,
     );
   }
 
@@ -2077,6 +2080,7 @@ function sendMessageViaCli(
   profile?: string,
   resumeSessionId?: string,
   attachments?: Attachment[],
+  modelOverride?: string,
 ): ChatHandle {
   // CLI fallback can't pipe multimodal content; inline text-file attachments
   // and ignore images.  The gateway is the supported attachment path; this
@@ -2108,8 +2112,8 @@ function sendMessageViaCli(
     args.push("--resume", resumeSessionId);
   }
 
-  if (mc.model) {
-    args.push("-m", mc.model);
+  if (modelOverride || mc.model) {
+    args.push("-m", modelOverride || mc.model);
   }
 
   const cliProvider = CLI_COMPAT_PROVIDER_OVERRIDE[mc.provider];
@@ -2404,6 +2408,7 @@ async function sendMessageViaNonGatewayApi(
   history?: Array<{ role: string; content: string }>,
   attachments?: Attachment[],
   contextFolder?: string,
+  modelOverride?: string,
 ): Promise<ChatHandle> {
   const approvalCommand = /^\/(?:approve|deny)\b/i.test(message.trim());
   if (!attachments?.length && !approvalCommand) {
@@ -2417,6 +2422,7 @@ async function sendMessageViaNonGatewayApi(
           history,
           attachments,
           contextFolder,
+          modelOverride,
         );
       }
   }
@@ -2429,6 +2435,7 @@ async function sendMessageViaNonGatewayApi(
     history,
     attachments,
     contextFolder,
+    modelOverride,
   );
 }
 
@@ -2440,13 +2447,18 @@ async function sendMessageViaBestApi(
   history?: Array<{ role: string; content: string }>,
   attachments?: Attachment[],
   contextFolder?: string,
+  modelOverride?: string,
 ): Promise<ChatHandle> {
   const approvalCommand = /^\/(?:approve|deny)\b/i.test(message.trim());
+  // Skip the TUI gateway when a session-scoped model override is active — the
+  // TUI gateway reads its model from config.yaml and has no per-request
+  // override mechanism. The API path below already honours modelOverride.
   if (
     shouldUseTuiGatewayClient() &&
     !isRemoteMode() &&
     !attachments?.length &&
-    !approvalCommand
+    !approvalCommand &&
+    !modelOverride
   ) {
     try {
       return await sendMessageViaTuiGateway(
@@ -2473,6 +2485,7 @@ async function sendMessageViaBestApi(
     history,
     attachments,
     contextFolder,
+    modelOverride,
   );
 }
 
@@ -2484,6 +2497,7 @@ async function sendMessageViaBestApiWithLocalRecovery(
   history?: Array<{ role: string; content: string }>,
   attachments?: Attachment[],
   contextFolder?: string,
+  modelOverride?: string,
 ): Promise<ChatHandle> {
   let aborted = false;
   let retrying = false;
@@ -2528,6 +2542,7 @@ async function sendMessageViaBestApiWithLocalRecovery(
         history,
         attachments,
         contextFolder,
+        modelOverride,
       );
       return;
     }
@@ -2538,6 +2553,7 @@ async function sendMessageViaBestApiWithLocalRecovery(
       profile,
       resumeSessionId,
       attachments,
+      modelOverride,
     );
   };
 
@@ -2615,6 +2631,7 @@ async function sendMessageViaBestApiWithLocalRecovery(
     history,
     attachments,
     contextFolder,
+    modelOverride,
   );
 
   return handle;
@@ -2628,6 +2645,7 @@ export async function sendMessage(
   history?: Array<{ role: string; content: string }>,
   attachments?: Attachment[],
   contextFolder?: string,
+  modelOverride?: string,
 ): Promise<ChatHandle> {
   ensureInitialized();
 
@@ -2641,6 +2659,7 @@ export async function sendMessage(
       history,
       attachments,
       contextFolder,
+      modelOverride,
     );
   }
 
@@ -2652,6 +2671,7 @@ export async function sendMessage(
       profile,
       resumeSessionId,
       attachments,
+      modelOverride,
     );
   }
 
@@ -2675,11 +2695,12 @@ export async function sendMessage(
       history,
       attachments,
       contextFolder,
+      modelOverride,
     );
   }
 
   // Fallback to CLI
-  return sendMessageViaCli(message, cb, profile, resumeSessionId, attachments);
+  return sendMessageViaCli(message, cb, profile, resumeSessionId, attachments, modelOverride);
 }
 
 // Lazy init — called on first sendMessage or gateway start

src/main/index.ts

@@ -1251,6 +1251,7 @@ function setupIPC(): void {
       attachments?: Attachment[],
       contextFolder?: string,
       runId?: string,
+      modelOverride?: string,
     ) => {
       // Each conversation has a stable runId minted by the renderer. Fall back
       // to a generated id for legacy callers so the run is still tracked.
@@ -1395,6 +1396,7 @@ function setupIPC(): void {
         history,
         attachments,
         contextFolder,
+        modelOverride,
       );
 
       activeRuns.set(chatRunId, handle.abort);

src/preload/index.d.ts

@@ -368,6 +368,7 @@ interface HermesAPI {
     attachments?: Attachment[],
     contextFolder?: string,
     runId?: string,
+    modelOverride?: string,
   ) => Promise<{ response: string; sessionId?: string }>;
   abortChat: (runId?: string) => Promise<void>;
   transcribeAudio: (

src/preload/index.ts

@@ -368,6 +368,7 @@ const hermesAPI = {
     attachments?: Attachment[],
     contextFolder?: string,
     runId?: string,
+    modelOverride?: string,
   ): Promise<{ response: string; sessionId?: string }> =>
     ipcRenderer.invoke(
       "send-message",
@@ -378,6 +379,7 @@ const hermesAPI = {
       attachments,
       contextFolder,
       runId,
+      modelOverride,
     ),
 
   abortChat: (runId?: string): Promise<void> =>

src/renderer/src/screens/Chat/Chat.tsx

@@ -117,6 +117,13 @@ function Chat({
   // Whether the worktree panel is visible (only applies when contextFolder is set)
   // Default false so the panel doesn't open automatically and interfere with scrolling
   const [worktreeVisible, setWorktreeVisible] = useState<boolean>(false);
+  // Explicit session-scoped model override — set only when the user picks
+  // from the chat-screen picker (persist:false). Undefined until then so the
+  // TUI gateway bypass in sendMessageViaBestApi is not triggered for normal
+  // chats where the user never changed the model (issue #688).
+  const [sessionModelOverride, setSessionModelOverride] = useState<
+    string | undefined
+  >(undefined);
   const dragCounter = useRef(0);
   const chatInputRef = useRef<ChatInputHandle>(null);
   const queueRef = useRef<QueuedMessage[]>([]);
@@ -424,6 +431,7 @@ function Chat({
     localCommands,
     activeTurnRef,
     contextFolder,
+    sessionModel: sessionModelOverride,
     sendViaDashboard: dashboardTransport.enabled
       ? dashboardTransport.sendMessage
       : undefined,
@@ -634,7 +642,12 @@ function Chat({
                 modelGroups={modelConfig.modelGroups}
                 displayModel={modelConfig.displayModel}
                 onOpen={modelConfig.reload}
-                onSelectModel={modelConfig.selectModel}
+                onSelectModel={(provider, model, baseUrl) => {
+                  void modelConfig.selectModel(provider, model, baseUrl, {
+                    persist: false,
+                  });
+                  setSessionModelOverride(model || undefined);
+                }}
               />
               <ReasoningEffortPicker
                 value={reasoningEffort}

src/renderer/src/screens/Chat/hooks/useChatActions.ts

@@ -44,6 +44,9 @@ interface UseChatActionsArgs {
   activeTurnRef: React.MutableRefObject<ActiveTurn | null>;
   /** Working folder bound to this conversation (issue #27), or null. */
   contextFolder: string | null;
+  /** Session-local model override — selected via the chat picker without
+   *  persisting to config.yaml (issue #688). */
+  sessionModel?: string;
   sendViaDashboard?: (
     text: string,
     attachments?: Attachment[],
@@ -106,6 +109,7 @@ export function useChatActions({
   localCommands,
   activeTurnRef,
   contextFolder,
+  sessionModel,
   sendViaDashboard,
   execSlashViaDashboard,
   runBackgroundViaDashboard,
@@ -115,9 +119,11 @@ export function useChatActions({
 }: UseChatActionsArgs): UseChatActionsResult {
   const messagesRef = useRef(messages);
   const isLoadingRef = useRef(isLoading);
+  const sessionModelRef = useRef(sessionModel);
   useEffect(() => {
     messagesRef.current = messages;
     isLoadingRef.current = isLoading;
+    sessionModelRef.current = sessionModel;
   });
 
   const pushUser = useCallback(
@@ -156,6 +162,7 @@ export function useChatActions({
           attachments,
           contextFolder ?? undefined,
           runId,
+          sessionModelRef.current || undefined,
         );
       } catch {
         // onChatError IPC already surfaces this to the user

src/renderer/src/screens/Chat/hooks/useModelConfig.ts

@@ -49,6 +49,7 @@ interface UseModelConfigResult {
     provider: string,
     model: string,
     baseUrl: string,
+    options?: { persist?: boolean },
   ) => Promise<void>;
 }
 
@@ -134,8 +135,12 @@ export function useModelConfig(profile?: string): UseModelConfigResult {
   }, [reload]);
 
   const selectModel = useCallback(
-    async (provider: string, model: string, baseUrl: string): Promise<void> => {
-      const seq = ++loadSeqRef.current;
+    async (
+      provider: string,
+      model: string,
+      baseUrl: string,
+      { persist = true }: { persist?: boolean } = {},
+    ): Promise<void> => {
       // Named providers (deepseek, groq, anthropic, …) have a hardcoded
       // canonical base_url in `hermes-agent`'s PROVIDER_REGISTRY.  A stored
       // model entry that carries a stale `baseUrl` from an earlier confused
@@ -148,6 +153,16 @@ export function useModelConfig(profile?: string): UseModelConfigResult {
       setCurrentModel(model);
       setCurrentProvider(provider);
       setCurrentBaseUrl(effectiveBaseUrl);
+      // Session-only selection: update local state only, do not write to
+      // config.yaml so the global default model is preserved (issue #688).
+      // Advance the sequence counter so any in-flight reload() triggered by
+      // onConnectionConfigChanged / onModelLibraryChanged cannot clobber the
+      // session-scoped selection with the persisted value.
+      if (!persist) {
+        ++loadSeqRef.current;
+        return;
+      }
+      const seq = ++loadSeqRef.current;
       try {
         await window.hermesAPI.setModelConfig(
           provider,