fix: chat works with any configured provider + surfaces errors

The chat defaulted to a Claude model, so a user who saved only a Gemini key got
a silent failure (backend derived Anthropic, found no key, errored — and the UI
showed nothing).

- Backend: resolveModel now falls back to whichever provider actually has a key
  (preferring the configured one), so a Gemini key just works regardless of the
  picked model. Clear 400 if no provider is configured at all.
- Frontend: the chat seeds its model/effort from the saved AI config, and now
  surfaces request failures as both a persistent alert banner and a toast —
  never silent.
- Settings: switching provider auto-selects that provider's default model.
- Refreshed the model catalog to current ids (Gemini 2.5 Pro/Flash + 2.0 Flash;
  dropped the retired gemini-1.5-pro); per-provider default model updated.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

backend/src/services/ai/provider.ts

@@ -39,6 +39,28 @@ const PROVIDER_LABELS: Record<ApiProvider, string> = {
   gemini: "Google Gemini",
 };
 
+const PROVIDER_ORDER: ApiProvider[] = ["anthropic", "openai", "gemini"];
+
+// A safe default model id for each provider, used when the picked model doesn't
+// belong to the provider we end up calling.
+const DEFAULT_MODEL: Record<ApiProvider, string> = {
+  anthropic: "claude-sonnet-4-6",
+  openai: "gpt-4o",
+  gemini: "gemini-2.5-flash",
+};
+
+// Choose which provider to actually call: prefer the one the picked model maps
+// to (if it has a key), else the user's configured provider (if keyed), else any
+// provider that has a key. Returns null when no key is configured at all.
+function chooseProvider(
+  settings: AiSettingsRow,
+  requested: ApiProvider | null,
+): ApiProvider | null {
+  if (requested && getApiKey(settings, requested)) return requested;
+  if (getApiKey(settings, settings.provider)) return settings.provider;
+  return PROVIDER_ORDER.find((p) => getApiKey(settings, p)) ?? null;
+}
+
 // Resolve a concrete LanguageModel for a request. `requestedModelId` is the id
 // the user picked in the chat input; when it maps to a cloud provider we use
 // that provider's stored key, otherwise we fall back to local Ollama (also used
@@ -47,11 +69,8 @@ export function resolveModel(
   settings: AiSettingsRow,
   requestedModelId: string,
 ): ResolvedModel {
-  const provider =
-    settings.mode === "local" ? null : providerForModel(requestedModelId);
-
-  if (!provider) {
-    // Local mode via Ollama's OpenAI-compatible endpoint. No key required.
+  // Local mode (or the local sentinel) → Ollama's OpenAI-compatible endpoint.
+  if (settings.mode === "local" || requestedModelId === OLLAMA_SENTINEL) {
     const ollama = createOpenAICompatible({
       name: "ollama",
       baseURL: `${settings.ollamaBaseUrl.replace(/\/$/, "")}/v1`,
@@ -63,20 +82,34 @@ export function resolveModel(
     };
   }
 
-  const apiKey = getApiKey(settings, provider);
-  if (!apiKey) {
+  // API mode. Pick a provider that actually has a key, falling back gracefully
+  // so a configured key (e.g. Gemini) is used even if the picked model belongs
+  // to a different, unconfigured provider.
+  const requested = providerForModel(requestedModelId);
+  const provider = chooseProvider(settings, requested);
+  if (!provider) {
     throw new HttpError(
       400,
-      `No API key configured for ${PROVIDER_LABELS[provider]}. Add one in Settings → AI.`,
+      "No AI provider API key is configured. Add one in Settings → AI, or switch to a local model.",
     );
   }
 
+  // Use the picked model only if it belongs to the chosen provider; otherwise
+  // use the user's default (or a per-provider default).
+  const modelName =
+    requested === provider
+      ? requestedModelId
+      : providerForModel(settings.defaultModel) === provider
+        ? settings.defaultModel
+        : DEFAULT_MODEL[provider];
+
+  const apiKey = getApiKey(settings, provider)!;
   const model: LanguageModel =
     provider === "anthropic"
-      ? createAnthropic({ apiKey })(requestedModelId)
+      ? createAnthropic({ apiKey })(modelName)
       : provider === "gemini"
-        ? createGoogleGenerativeAI({ apiKey })(requestedModelId)
-        : createOpenAI({ apiKey })(requestedModelId);
+        ? createGoogleGenerativeAI({ apiKey })(modelName)
+        : createOpenAI({ apiKey })(modelName);
 
   return { model, isExternal: true, providerLabel: PROVIDER_LABELS[provider] };
 }

frontend/components/chat/chat-panel.tsx

@@ -2,6 +2,7 @@
 
 import { useChat } from "@ai-sdk/react";
 import { DefaultChatTransport } from "ai";
+import { AlertTriangle } from "lucide-react";
 import { nanoid } from "nanoid";
 import { useSearchParams } from "next/navigation";
 import { useCallback, useEffect, useMemo, useRef, useState } from "react";
@@ -38,8 +39,10 @@ import {
   getModel,
 } from "@/lib/ai-models";
 import type { TemetroUIMessage } from "@/lib/ai-chat";
+import { getAiConfig } from "@/lib/ai-settings";
 import { API_BASE_URL } from "@/lib/api-client";
 import { getPatient } from "@/lib/patients";
+import { notify } from "@/lib/toast";
 
 // Trigger: `/patient 10293` or just `/10293` — a client-side fast-path that
 // pulls records instantly without the LLM (also works offline).
@@ -65,9 +68,36 @@ export function ChatPanel() {
     [],
   );
 
-  const { messages, setMessages, sendMessage, status, stop } =
+  const { messages, setMessages, sendMessage, status, stop, error } =
     useChat<TemetroUIMessage>({ transport });
 
+  // Seed the model + effort from the user's saved AI config so the chat uses the
+  // provider they actually configured (e.g. their Gemini default), not a stale
+  // hardcoded default.
+  useEffect(() => {
+    let cancelled = false;
+    getAiConfig()
+      .then((cfg) => {
+        if (cancelled) return;
+        setModel(cfg.mode === "local" ? "ollama" : cfg.defaultModel);
+        setEffort(cfg.defaultEffort);
+      })
+      .catch(() => {
+        // Keep defaults; the chat still works and the backend falls back to any
+        // configured provider.
+      });
+    return () => {
+      cancelled = true;
+    };
+  }, []);
+
+  // Pop a toast whenever a request errors, so failures are never silent.
+  useEffect(() => {
+    if (error) {
+      notify.error(t("chat.error.title"), error.message || t("chat.error.body"));
+    }
+  }, [error, t]);
+
   const isCloudModel = (getModel(model)?.provider ?? "ollama") !== "ollama";
 
   // Run the LLM agent for a message (after any consent gate).
@@ -158,6 +188,21 @@ export function ChatPanel() {
     />
   );
 
+  const errorAlert = error ? (
+    <div
+      className="flex w-full items-start gap-2 rounded-2xl border border-destructive/40 bg-destructive/8 px-4 py-3 text-sm text-destructive-foreground"
+      role="alert"
+    >
+      <AlertTriangle className="mt-0.5 size-4 shrink-0" />
+      <div className="space-y-0.5">
+        <p className="font-medium">{t("chat.error.title")}</p>
+        <p className="text-destructive-foreground/90">
+          {error.message || t("chat.error.body")}
+        </p>
+      </div>
+    </div>
+  ) : null;
+
   const consentDialog = (
     <Dialog onOpenChange={setConsentOpen} open={consentOpen}>
       <DialogPopup>
@@ -191,7 +236,10 @@ export function ChatPanel() {
           <h1 className="text-center text-3xl font-semibold tracking-tight text-balance sm:text-4xl">
             {t("chat.heading")}
           </h1>
-          {promptInput}
+          <div className="flex w-full flex-col gap-3">
+            {errorAlert}
+            {promptInput}
+          </div>
         </div>
         {consentDialog}
       </div>
@@ -240,7 +288,10 @@ export function ChatPanel() {
         </ConversationContent>
         <ConversationScrollButton />
       </Conversation>
-      <div className="mx-auto w-full max-w-3xl px-4 pb-4">{promptInput}</div>
+      <div className="mx-auto flex w-full max-w-3xl flex-col gap-3 px-4 pb-4">
+        {errorAlert}
+        {promptInput}
+      </div>
       {consentDialog}
     </div>
   );

frontend/components/settings/settings-ai.tsx

@@ -71,6 +71,19 @@ export function AIPanel() {
   const set = <K extends keyof AiConfig>(key: K, value: AiConfig[K]) =>
     setConfig((prev) => ({ ...prev, [key]: value }));
 
+  // Switching provider: if the current default model doesn't belong to the new
+  // provider, pick that provider's first model so they never end up mismatched.
+  const setProvider = (provider: ApiProvider) =>
+    setConfig((prev) => {
+      const models = AI_MODELS.filter((m) => m.provider === provider);
+      const stillValid = models.some((m) => m.id === prev.defaultModel);
+      return {
+        ...prev,
+        provider,
+        defaultModel: stillValid ? prev.defaultModel : (models[0]?.id ?? prev.defaultModel),
+      };
+    });
+
   // Models available for the currently selected cloud provider.
   const providerModels = useMemo(
     () => AI_MODELS.filter((m) => m.provider === config.provider),
@@ -169,9 +182,7 @@ export function AIPanel() {
               <div className="space-y-1.5">
                 <FieldLabel>{t("settings.ai.provider")}</FieldLabel>
                 <Select
-                  onValueChange={(value) =>
-                    set("provider", value as ApiProvider)
-                  }
+                  onValueChange={(value) => setProvider(value as ApiProvider)}
                   value={config.provider}
                 >
                   <SelectTrigger>

frontend/lib/ai-models.ts

@@ -51,7 +51,27 @@ export const AI_MODELS: AiModel[] = [
     provider: "anthropic",
     label: "Haiku 4.5",
     descriptionKey: "chat.input.models.haiku45",
+  },
+  {
+    id: "gemini-2.5-pro",
+    provider: "gemini",
+    label: "Gemini 2.5 Pro",
+    descriptionKey: "chat.input.models.gemini25Pro",
     primary: true,
+    supportsEffort: true,
+  },
+  {
+    id: "gemini-2.5-flash",
+    provider: "gemini",
+    label: "Gemini 2.5 Flash",
+    descriptionKey: "chat.input.models.gemini25Flash",
+    primary: true,
+  },
+  {
+    id: "gemini-2.0-flash",
+    provider: "gemini",
+    label: "Gemini 2.0 Flash",
+    descriptionKey: "chat.input.models.gemini20Flash",
   },
   {
     id: "gpt-4o",
@@ -66,19 +86,6 @@ export const AI_MODELS: AiModel[] = [
     label: "GPT-4o mini",
     descriptionKey: "chat.input.models.gpt4oMini",
   },
-  {
-    id: "gemini-2.0-flash",
-    provider: "gemini",
-    label: "Gemini 2.0 Flash",
-    descriptionKey: "chat.input.models.gemini20Flash",
-  },
-  {
-    id: "gemini-1.5-pro",
-    provider: "gemini",
-    label: "Gemini 1.5 Pro",
-    descriptionKey: "chat.input.models.gemini15Pro",
-    supportsEffort: true,
-  },
   {
     id: "ollama",
     provider: "ollama",

frontend/lib/i18n/locales/en/translation.json

@@ -634,10 +634,11 @@
         "opus48": "For complex clinical reasoning",
         "sonnet46": "Balanced for everyday tasks",
         "haiku45": "Fastest for quick answers",
+        "gemini25Pro": "Google’s most capable model",
+        "gemini25Flash": "Fast Google model",
+        "gemini20Flash": "Lightweight Google model",
         "gpt4o": "OpenAI multimodal flagship",
         "gpt4oMini": "Fast, low-cost OpenAI model",
-        "gemini20Flash": "Fast Google model",
-        "gemini15Pro": "Google long-context model",
         "ollama": "Runs locally on your infrastructure"
       }
     },
@@ -673,6 +674,10 @@
       "failedCount": "{{count}} failed",
       "failedTitle": "Import failed",
       "failedBody": "The records could not be imported. Please try again."
+    },
+    "error": {
+      "title": "AI request failed",
+      "body": "Something went wrong reaching the AI. Check your provider and API key in Settings → AI."
     }
   },
   "patientCard": {