Enhance model handling and settings integration for providers (#295)

docs-site/docs/features/settings.md

@@ -47,6 +47,16 @@ Settings gives you runtime overrides for the key parts of discovery, scoring, ta
   - Scoring model
   - Tailoring model
   - Project-selection model
+- Provider defaults are applied automatically when the model fields are left blank:
+  - `openai` defaults to `gpt-5.4-mini`
+  - `gemini` defaults to `google/gemini-3-flash-preview`
+- The settings page shows provider-aware model pickers for:
+  - `openai`: available text-generation models only
+  - `gemini`: available Gemini text-generation models only
+  - `ollama`: locally installed Ollama models
+- `openrouter`, `lmstudio`, and `openai_compatible` stay manual-entry because JobOps cannot safely infer the exact model catalog from those providers
+- Changing the provider clears stale model overrides in the form, so inherited fields follow the new provider default unless you explicitly choose a new override
+- The preview under each field and the **Resolved config** block reflect the model currently selected in the form, even before you save
 
 ### Webhooks
 
@@ -199,6 +209,13 @@ curl -X POST "http://localhost:3001/api/backups"
 
 - Some settings apply only to new runs/actions after save.
 - Re-run scoring/tailoring/pipeline to validate effect.
+- In the **Model** section, the field preview and **Resolved config** update immediately when you choose a model, but the change only applies to future actions after you click **Save**.
+
+### Tailoring or scoring says the selected model does not exist for the current provider
+
+- Open **Settings -> Model** and confirm the provider and model belong together.
+- If you switch providers, leave the model fields blank to use the provider default, or pick a new provider-compatible model from the dropdown.
+- JobOps ignores stale Gemini-style overrides under `openai`, and ignores stale OpenAI-style overrides under `gemini`, but you still need to save the current form selection for future runs.
 
 ### Resume tailoring used English instead of my resume language
 

docs-site/docs/troubleshooting/common-problems.md

@@ -30,6 +30,14 @@ npm --workspace docs-site run build
 - Validate `LLM_API_KEY` and provider settings.
 - Check settings page and API connectivity.
 
+## Resume tailoring or scoring says the model does not exist
+
+- Root cause: the selected provider and model do not match.
+- Open **Settings -> Model** and check both the provider and the current model preview.
+- If you recently switched providers, leave the model fields blank to use the provider default, or select a provider-compatible model and save again.
+- For `openai`, JobOps defaults to `gpt-5.4-mini` when the model field is blank.
+- For `gemini`, JobOps defaults to `google/gemini-3-flash-preview` when the model field is blank.
+
 ## PDF generation fails
 
 - Verify RxResume credentials.

orchestrator/src/client/api/client.ts

@@ -1336,6 +1336,18 @@ export async function validateLlm(input: {
   });
 }
 
+export async function getLlmModels(input?: {
+  provider?: string;
+  baseUrl?: string;
+  apiKey?: string;
+}): Promise<string[]> {
+  const data = await fetchApi<{ models: string[] }>("/settings/llm-models", {
+    method: "POST",
+    body: JSON.stringify(input ?? {}),
+  });
+  return data.models;
+}
+
 export async function validateRxresume(input?: {
   mode?: "v4" | "v5";
   email?: string;

orchestrator/src/client/components/OnboardingGate.tsx

@@ -17,6 +17,7 @@ import {
   LLM_PROVIDERS,
   normalizeLlmProvider,
 } from "@client/pages/settings/utils";
+import { getDefaultModelForProvider } from "@shared/settings-registry";
 import type { UpdateSettingsInput } from "@shared/settings-schema.js";
 import type { RxResumeMode, ValidationResult } from "@shared/types.js";
 import { Check } from "lucide-react";
@@ -423,6 +424,10 @@ export const OnboardingGate: React.FC = () => {
       const update: Partial<UpdateSettingsInput> = {
         llmProvider: normalizedProvider,
         llmBaseUrl: showBaseUrl ? baseUrlValue || null : null,
+        model: null,
+        modelScorer: null,
+        modelTailoring: null,
+        modelProjectSelection: null,
       };
 
       if (showApiKey && apiKeyValue) {
@@ -433,7 +438,13 @@ export const OnboardingGate: React.FC = () => {
       await api.updateSettings(update);
       await refreshSettings();
       setValue("llmApiKey", "");
-      toast.success("LLM provider connected");
+      const defaultModel = getDefaultModelForProvider(normalizedProvider);
+      toast.success("LLM provider connected", {
+        description:
+          normalizedProvider === "openai" || normalizedProvider === "gemini"
+            ? `Default for ${providerConfig.label}: ${defaultModel}.`
+            : "Select the model manually in Settings > Model.",
+      });
       return true;
     } catch (error) {
       const message =

orchestrator/src/client/pages/SettingsPage.test.tsx

@@ -15,6 +15,7 @@ const render = (ui: Parameters<typeof renderWithQueryClient>[0]) =>
 
 vi.mock("../api", () => ({
   getSettings: vi.fn(),
+  getLlmModels: vi.fn().mockResolvedValue([]),
   updateSettings: vi.fn(),
   validateRxresume: vi.fn(),
   getRxResumeProjects: vi.fn(),
@@ -217,6 +218,50 @@ describe("SettingsPage", () => {
     await waitFor(() => expect(saveButton).toBeEnabled());
   });
 
+  it("clears stale model overrides when the provider changes", async () => {
+    vi.mocked(api.getSettings).mockResolvedValue(
+      createAppSettings({
+        model: {
+          value: "google/gemini-3-flash-preview",
+          default: "google/gemini-3-flash-preview",
+          override: "google/gemini-3-flash-preview",
+        },
+        modelScorer: { value: "google/gemini-3-flash-preview", override: null },
+        modelTailoring: {
+          value: "google/gemini-3-flash-preview",
+          override: "google/gemini-3-flash-preview",
+        },
+        modelProjectSelection: {
+          value: "google/gemini-3-flash-preview",
+          override: null,
+        },
+        llmProvider: { value: "gemini", default: "gemini", override: "gemini" },
+      }),
+    );
+    vi.mocked(api.updateSettings).mockResolvedValue(baseSettings);
+
+    renderPage();
+    await openModelSection();
+
+    fireEvent.click(screen.getByRole("combobox", { name: /provider/i }));
+    fireEvent.click(await screen.findByText("OpenAI"));
+
+    const saveButton = screen.getByRole("button", { name: /^save$/i });
+    await waitFor(() => expect(saveButton).toBeEnabled());
+    fireEvent.click(saveButton);
+
+    await waitFor(() => expect(api.updateSettings).toHaveBeenCalled());
+    expect(api.updateSettings).toHaveBeenCalledWith(
+      expect.objectContaining({
+        llmProvider: "openai",
+        model: null,
+        modelScorer: null,
+        modelTailoring: null,
+        modelProjectSelection: null,
+      }),
+    );
+  });
+
   it("hides pipeline tuning sections that moved to run modal", async () => {
     vi.mocked(api.getSettings).mockResolvedValue(baseSettings);
     renderPage();

orchestrator/src/client/pages/SettingsPage.tsx

@@ -824,10 +824,26 @@ export const SettingsPage: React.FC = () => {
       }
 
       const payload: Partial<UpdateSettingsInput> = {
-        model: normalizeString(data.model),
+        model: dirtyFields.llmProvider
-        modelScorer: normalizeString(data.modelScorer),
+          ? dirtyFields.model
-        modelTailoring: normalizeString(data.modelTailoring),
+            ? normalizeString(data.model)
-        modelProjectSelection: normalizeString(data.modelProjectSelection),
+            : null
+          : normalizeString(data.model),
+        modelScorer: dirtyFields.llmProvider
+          ? dirtyFields.modelScorer
+            ? normalizeString(data.modelScorer)
+            : null
+          : normalizeString(data.modelScorer),
+        modelTailoring: dirtyFields.llmProvider
+          ? dirtyFields.modelTailoring
+            ? normalizeString(data.modelTailoring)
+            : null
+          : normalizeString(data.modelTailoring),
+        modelProjectSelection: dirtyFields.llmProvider
+          ? dirtyFields.modelProjectSelection
+            ? normalizeString(data.modelProjectSelection)
+            : null
+          : normalizeString(data.modelProjectSelection),
         pipelineWebhookUrl: normalizeString(data.pipelineWebhookUrl),
         jobCompleteWebhookUrl: normalizeString(data.jobCompleteWebhookUrl),
         resumeProjects: resumeProjectsOverride,

orchestrator/src/client/pages/settings/components/ModelSettingsSection.tsx

@@ -1,3 +1,4 @@
+import * as api from "@client/api";
 import { SettingsInput } from "@client/pages/settings/components/SettingsInput";
 import type { ModelValues } from "@client/pages/settings/types";
 import {
@@ -5,16 +6,19 @@ import {
   getLlmProviderConfig,
   LLM_PROVIDER_LABELS,
   LLM_PROVIDERS,
+  supportsLlmModelSuggestions,
 } from "@client/pages/settings/utils";
+import { getDefaultModelForProvider } from "@shared/settings-registry";
 import type { UpdateSettingsInput } from "@shared/settings-schema.js";
 import type React from "react";
-import { useEffect } from "react";
+import { useDeferredValue, useEffect, useRef, useState } from "react";
 import { Controller, useFormContext } from "react-hook-form";
 import {
   AccordionContent,
   AccordionItem,
   AccordionTrigger,
 } from "@/components/ui/accordion";
+import { SearchableDropdown } from "@/components/ui/searchable-dropdown";
 import {
   Select,
   SelectContent,
@@ -35,12 +39,12 @@ export const ModelSettingsSection: React.FC<ModelSettingsSectionProps> = ({
   isLoading,
   isSaving,
 }) => {
+  const [availableModels, setAvailableModels] = useState<string[]>([]);
+  const [isLoadingModels, setIsLoadingModels] = useState(false);
+  const [modelsError, setModelsError] = useState<string | null>(null);
   const {
     effective,
     default: defaultModel,
-    scorer,
-    tailoring,
-    projectSelection,
     llmProvider,
     llmBaseUrl,
     llmApiKeyHint,
@@ -54,10 +58,28 @@ export const ModelSettingsSection: React.FC<ModelSettingsSectionProps> = ({
   } = useFormContext<UpdateSettingsInput>();
 
   const selectedProvider = watch("llmProvider") || llmProvider || "openrouter";
+  const previousProviderRef = useRef(selectedProvider);
   const providerConfig = getLlmProviderConfig(selectedProvider);
   const { showApiKey, showBaseUrl } = providerConfig;
 
   const llmBaseUrlValue = watch("llmBaseUrl");
+  const llmApiKeyValue = watch("llmApiKey") ?? "";
+  const modelValue = watch("model") ?? "";
+  const modelScorerValue = watch("modelScorer") ?? "";
+  const modelTailoringValue = watch("modelTailoring") ?? "";
+  const modelProjectSelectionValue = watch("modelProjectSelection") ?? "";
+  const providerDefaultModel = getDefaultModelForProvider(
+    selectedProvider,
+    selectedProvider === llmProvider ? defaultModel : undefined,
+  );
+  const deferredProvider = useDeferredValue(selectedProvider);
+  const deferredBaseUrl = useDeferredValue(llmBaseUrlValue ?? "");
+  const deferredApiKey = useDeferredValue(llmApiKeyValue);
+  const supportsModelSuggestions =
+    supportsLlmModelSuggestions(selectedProvider);
+  const hasAvailableApiKey = showApiKey
+    ? Boolean(deferredApiKey.trim() || llmApiKeyHint)
+    : true;
 
   useEffect(() => {
     if (showBaseUrl) return;
@@ -66,12 +88,122 @@ export const ModelSettingsSection: React.FC<ModelSettingsSectionProps> = ({
     }
   }, [setValue, showBaseUrl, llmBaseUrlValue]);
 
+  useEffect(() => {
+    if (previousProviderRef.current === selectedProvider) {
+      return;
+    }
+
+    previousProviderRef.current = selectedProvider;
+    setValue("model", "", { shouldDirty: true });
+    setValue("modelScorer", "", { shouldDirty: true });
+    setValue("modelTailoring", "", { shouldDirty: true });
+    setValue("modelProjectSelection", "", { shouldDirty: true });
+  }, [selectedProvider, setValue]);
+
+  useEffect(() => {
+    if (!supportsModelSuggestions) {
+      setAvailableModels([]);
+      setModelsError(null);
+      setIsLoadingModels(false);
+      return;
+    }
+
+    if (!hasAvailableApiKey) {
+      setAvailableModels([]);
+      setModelsError(null);
+      setIsLoadingModels(false);
+      return;
+    }
+
+    let cancelled = false;
+    setIsLoadingModels(true);
+    setModelsError(null);
+
+    void api
+      .getLlmModels({
+        provider: deferredProvider,
+        baseUrl: showBaseUrl ? deferredBaseUrl.trim() || undefined : undefined,
+        apiKey: showApiKey ? deferredApiKey.trim() || undefined : undefined,
+      })
+      .then((models) => {
+        if (cancelled) return;
+        setAvailableModels(models);
+        setModelsError(null);
+      })
+      .catch((error) => {
+        if (cancelled) return;
+        setAvailableModels([]);
+        setModelsError(
+          error instanceof Error ? error.message : "Failed to load models.",
+        );
+      })
+      .finally(() => {
+        if (cancelled) return;
+        setIsLoadingModels(false);
+      });
+
+    return () => {
+      cancelled = true;
+    };
+  }, [
+    deferredApiKey,
+    deferredBaseUrl,
+    deferredProvider,
+    hasAvailableApiKey,
+    showApiKey,
+    showBaseUrl,
+    supportsModelSuggestions,
+  ]);
+
   const keyHint = formatSecretHint(llmApiKeyHint);
   const keyText = showApiKey ? keyHint || "Not set" : "Not required";
-  const effectiveDefaultModel = effective || defaultModel || "—";
+  const resolvedBaseUrl = llmBaseUrlValue?.trim() || llmBaseUrl || "-";
-  const scoringModel = scorer || effectiveDefaultModel;
+  const selectedDefaultModel = modelValue.trim();
-  const tailoringModel = tailoring || effectiveDefaultModel;
+  const previewDefaultModel =
-  const projectSelectionModel = projectSelection || effectiveDefaultModel;
+    selectedDefaultModel || effective || providerDefaultModel || "-";
+  const selectedScoringModel = modelScorerValue.trim();
+  const selectedTailoringModel = modelTailoringValue.trim();
+  const selectedProjectSelectionModel = modelProjectSelectionValue.trim();
+  const scoringModel = selectedScoringModel || previewDefaultModel;
+  const tailoringModel = selectedTailoringModel || previewDefaultModel;
+  const projectSelectionModel =
+    selectedProjectSelectionModel || previewDefaultModel;
+  const modelHelper = supportsModelSuggestions
+    ? !hasAvailableApiKey
+      ? `Add or save a ${providerConfig.label} API key to load available models.`
+      : isLoadingModels
+        ? "Loading available models..."
+        : modelsError
+          ? modelsError
+          : availableModels.length > 0
+            ? "Choose from the available text-generation models."
+            : "No text-generation models were returned."
+    : `Type the exact model name manually, or leave blank to use the ${providerConfig.label} default model.`;
+  const defaultModelOptions = buildModelOptions({
+    models: availableModels,
+    emptyLabel: `Use ${providerConfig.label} default`,
+    emptyValue: "",
+    fallbackValue: modelValue.trim(),
+  });
+  const scoringModelOptions = buildModelOptions({
+    models: availableModels,
+    emptyLabel: "Inherit default model",
+    emptyValue: "",
+    fallbackValue: modelScorerValue.trim(),
+  });
+  const tailoringModelOptions = buildModelOptions({
+    models: availableModels,
+    emptyLabel: "Inherit default model",
+    emptyValue: "",
+    fallbackValue: modelTailoringValue.trim(),
+  });
+  const projectSelectionModelOptions = buildModelOptions({
+    models: availableModels,
+    emptyLabel: "Inherit default model",
+    emptyValue: "",
+    fallbackValue: modelProjectSelectionValue.trim(),
+  });
+
   return (
     <AccordionItem value="model" className="border rounded-lg px-4">
       <AccordionTrigger className="hover:no-underline py-4">
@@ -128,7 +260,7 @@ export const ModelSettingsSection: React.FC<ModelSettingsSectionProps> = ({
                   disabled={isLoading || isSaving}
                   error={errors.llmBaseUrl?.message as string | undefined}
                   helper={providerConfig.baseUrlHelper}
-                  current={llmBaseUrl || "—"}
+                  current={resolvedBaseUrl}
                 />
               )}
               {showApiKey && (
@@ -147,15 +279,53 @@ export const ModelSettingsSection: React.FC<ModelSettingsSectionProps> = ({
 
           <Separator />
 
-          <SettingsInput
+          {supportsModelSuggestions ? (
-            label="Default model"
+            <div className="space-y-2">
-            inputProps={register("model")}
+              <label htmlFor="model" className="text-sm font-medium">
-            placeholder={defaultModel || "google/gemini-3-flash-preview"}
+                Default model
-            disabled={isLoading || isSaving}
+              </label>
-            error={errors.model?.message as string | undefined}
+              <Controller
-            helper="Leave blank to use the default from server env (`MODEL`)."
+                name="model"
-            current={effectiveDefaultModel}
+                control={control}
-          />
+                render={({ field }) => (
+                  <SearchableDropdown
+                    inputId="model"
+                    value={field.value ?? ""}
+                    options={defaultModelOptions}
+                    onValueChange={field.onChange}
+                    placeholder={providerDefaultModel || "Select a model"}
+                    searchPlaceholder="Search models..."
+                    emptyText="No models found."
+                    ariaLabel="Default model"
+                    disabled={isLoading || isSaving || isLoadingModels}
+                    triggerClassName="h-9 w-full justify-between rounded-md border border-input bg-transparent px-3 text-sm font-normal shadow-sm"
+                    contentClassName="w-[var(--radix-popover-trigger-width)] border-border bg-popover p-0"
+                    listClassName="max-h-64"
+                  />
+                )}
+              />
+              {errors.model?.message && (
+                <p className="text-xs text-destructive">
+                  {errors.model.message as string}
+                </p>
+              )}
+              <div className="text-xs text-muted-foreground">{modelHelper}</div>
+              <div className="text-xs text-muted-foreground">
+                Current:{" "}
+                <span className="font-mono">{previewDefaultModel}</span>
+              </div>
+            </div>
+          ) : (
+            <SettingsInput
+              label="Default model"
+              inputProps={register("model")}
+              placeholder={providerDefaultModel}
+              disabled={isLoading || isSaving}
+              error={errors.model?.message as string | undefined}
+              helper={modelHelper}
+              current={previewDefaultModel}
+            />
+          )}
 
           <Separator />
 
@@ -163,34 +333,161 @@ export const ModelSettingsSection: React.FC<ModelSettingsSectionProps> = ({
             <div className="text-sm font-medium">Task-Specific Overrides</div>
 
             <div className="grid gap-4 md:grid-cols-2 lg:grid-cols-3">
-              <SettingsInput
+              {supportsModelSuggestions ? (
-                label="Scoring Model"
+                <>
-                inputProps={register("modelScorer")}
+                  <div className="space-y-2">
-                placeholder={effective || "inherit"}
+                    <label
-                disabled={isLoading || isSaving}
+                      htmlFor="modelScorer"
-                error={errors.modelScorer?.message as string | undefined}
+                      className="text-sm font-medium"
-                current={scoringModel}
+                    >
-              />
+                      Scoring Model
+                    </label>
+                    <Controller
+                      name="modelScorer"
+                      control={control}
+                      render={({ field }) => (
+                        <SearchableDropdown
+                          inputId="modelScorer"
+                          value={field.value ?? ""}
+                          options={scoringModelOptions}
+                          onValueChange={field.onChange}
+                          placeholder={
+                            previewDefaultModel || "Inherit default model"
+                          }
+                          searchPlaceholder="Search models..."
+                          emptyText="No models found."
+                          ariaLabel="Scoring Model"
+                          disabled={isLoading || isSaving || isLoadingModels}
+                          triggerClassName="h-9 w-full justify-between rounded-md border border-input bg-transparent px-3 text-sm font-normal shadow-sm"
+                          contentClassName="w-[var(--radix-popover-trigger-width)] border-border bg-popover p-0"
+                          listClassName="max-h-64"
+                        />
+                      )}
+                    />
+                    {errors.modelScorer?.message && (
+                      <p className="text-xs text-destructive">
+                        {errors.modelScorer.message as string}
+                      </p>
+                    )}
+                    <div className="text-xs text-muted-foreground">
+                      Current: <span className="font-mono">{scoringModel}</span>
+                    </div>
+                  </div>
 
-              <SettingsInput
+                  <div className="space-y-2">
-                label="Tailoring Model"
+                    <label
-                inputProps={register("modelTailoring")}
+                      htmlFor="modelTailoring"
-                placeholder={effective || "inherit"}
+                      className="text-sm font-medium"
-                disabled={isLoading || isSaving}
+                    >
-                error={errors.modelTailoring?.message as string | undefined}
+                      Tailoring Model
-                current={tailoringModel}
+                    </label>
-              />
+                    <Controller
+                      name="modelTailoring"
+                      control={control}
+                      render={({ field }) => (
+                        <SearchableDropdown
+                          inputId="modelTailoring"
+                          value={field.value ?? ""}
+                          options={tailoringModelOptions}
+                          onValueChange={field.onChange}
+                          placeholder={
+                            previewDefaultModel || "Inherit default model"
+                          }
+                          searchPlaceholder="Search models..."
+                          emptyText="No models found."
+                          ariaLabel="Tailoring Model"
+                          disabled={isLoading || isSaving || isLoadingModels}
+                          triggerClassName="h-9 w-full justify-between rounded-md border border-input bg-transparent px-3 text-sm font-normal shadow-sm"
+                          contentClassName="w-[var(--radix-popover-trigger-width)] border-border bg-popover p-0"
+                          listClassName="max-h-64"
+                        />
+                      )}
+                    />
+                    {errors.modelTailoring?.message && (
+                      <p className="text-xs text-destructive">
+                        {errors.modelTailoring.message as string}
+                      </p>
+                    )}
+                    <div className="text-xs text-muted-foreground">
+                      Current:{" "}
+                      <span className="font-mono">{tailoringModel}</span>
+                    </div>
+                  </div>
 
-              <SettingsInput
+                  <div className="space-y-2">
-                label="Project Selection Model"
+                    <label
-                inputProps={register("modelProjectSelection")}
+                      htmlFor="modelProjectSelection"
-                placeholder={effective || "inherit"}
+                      className="text-sm font-medium"
-                disabled={isLoading || isSaving}
+                    >
-                error={
+                      Project Selection Model
-                  errors.modelProjectSelection?.message as string | undefined
+                    </label>
-                }
+                    <Controller
-                current={projectSelectionModel}
+                      name="modelProjectSelection"
-              />
+                      control={control}
+                      render={({ field }) => (
+                        <SearchableDropdown
+                          inputId="modelProjectSelection"
+                          value={field.value ?? ""}
+                          options={projectSelectionModelOptions}
+                          onValueChange={field.onChange}
+                          placeholder={
+                            previewDefaultModel || "Inherit default model"
+                          }
+                          searchPlaceholder="Search models..."
+                          emptyText="No models found."
+                          ariaLabel="Project Selection Model"
+                          disabled={isLoading || isSaving || isLoadingModels}
+                          triggerClassName="h-9 w-full justify-between rounded-md border border-input bg-transparent px-3 text-sm font-normal shadow-sm"
+                          contentClassName="w-[var(--radix-popover-trigger-width)] border-border bg-popover p-0"
+                          listClassName="max-h-64"
+                        />
+                      )}
+                    />
+                    {errors.modelProjectSelection?.message && (
+                      <p className="text-xs text-destructive">
+                        {errors.modelProjectSelection.message as string}
+                      </p>
+                    )}
+                    <div className="text-xs text-muted-foreground">
+                      Current:{" "}
+                      <span className="font-mono">{projectSelectionModel}</span>
+                    </div>
+                  </div>
+                </>
+              ) : (
+                <>
+                  <SettingsInput
+                    label="Scoring Model"
+                    inputProps={register("modelScorer")}
+                    placeholder={previewDefaultModel || "inherit"}
+                    disabled={isLoading || isSaving}
+                    error={errors.modelScorer?.message as string | undefined}
+                    current={scoringModel}
+                  />
+
+                  <SettingsInput
+                    label="Tailoring Model"
+                    inputProps={register("modelTailoring")}
+                    placeholder={previewDefaultModel || "inherit"}
+                    disabled={isLoading || isSaving}
+                    error={errors.modelTailoring?.message as string | undefined}
+                    current={tailoringModel}
+                  />
+
+                  <SettingsInput
+                    label="Project Selection Model"
+                    inputProps={register("modelProjectSelection")}
+                    placeholder={previewDefaultModel || "inherit"}
+                    disabled={isLoading || isSaving}
+                    error={
+                      errors.modelProjectSelection?.message as
+                        | string
+                        | undefined
+                    }
+                    current={projectSelectionModel}
+                  />
+                </>
+              )}
             </div>
           </div>
 
@@ -200,36 +497,32 @@ export const ModelSettingsSection: React.FC<ModelSettingsSectionProps> = ({
             <div className="text-xs text-muted-foreground">Resolved config</div>
             <div className="grid gap-x-4 gap-y-2 text-xs sm:grid-cols-[160px_1fr]">
               <div className="text-muted-foreground">Provider</div>
-              <div className="font-mono">{selectedProvider || "—"}</div>
+              <div className="font-mono">{selectedProvider || "-"}</div>
 
               <div className="text-muted-foreground">Base URL</div>
-              <div className="font-mono">{llmBaseUrl || "—"}</div>
+              <div className="font-mono">{resolvedBaseUrl}</div>
 
               <div className="text-muted-foreground">API key</div>
               <div className="font-mono">{keyText}</div>
 
               <div className="text-muted-foreground">Default model</div>
-              <div className="font-mono">{effectiveDefaultModel}</div>
+              <div className="font-mono">{previewDefaultModel}</div>
 
               <div className="text-muted-foreground">Scoring model</div>
               <div className="font-mono">
-                {scoringModel === effectiveDefaultModel
+                {selectedScoringModel ? scoringModel : "inherits"}
-                  ? "inherits"
-                  : scoringModel}
               </div>
 
               <div className="text-muted-foreground">Tailoring model</div>
               <div className="font-mono">
-                {tailoringModel === effectiveDefaultModel
+                {selectedTailoringModel ? tailoringModel : "inherits"}
-                  ? "inherits"
-                  : tailoringModel}
               </div>
 
               <div className="text-muted-foreground">Project selection</div>
               <div className="font-mono">
-                {projectSelectionModel === effectiveDefaultModel
+                {selectedProjectSelectionModel
-                  ? "inherits"
+                  ? projectSelectionModel
-                  : projectSelectionModel}
+                  : "inherits"}
               </div>
             </div>
           </div>
@@ -238,3 +531,37 @@ export const ModelSettingsSection: React.FC<ModelSettingsSectionProps> = ({
     </AccordionItem>
   );
 };
+
+function buildModelOptions(input: {
+  models: string[];
+  emptyLabel: string;
+  emptyValue: string;
+  fallbackValue?: string;
+}) {
+  const options = [
+    {
+      value: input.emptyValue,
+      label: input.emptyLabel,
+      searchText: input.emptyLabel,
+    },
+    ...input.models.map((model) => ({
+      value: model,
+      label: model,
+      searchText: model,
+    })),
+  ];
+
+  const fallbackValue = input.fallbackValue?.trim();
+  if (
+    fallbackValue &&
+    !options.some((option) => option.value === fallbackValue)
+  ) {
+    options.unshift({
+      value: fallbackValue,
+      label: fallbackValue,
+      searchText: `${fallbackValue} custom`,
+    });
+  }
+
+  return options;
+}

orchestrator/src/client/pages/settings/utils.test.ts

@@ -1,5 +1,9 @@
 import { describe, expect, it } from "vitest";
-import { getLlmProviderConfig, normalizeLlmProvider } from "./utils";
+import {
+  getLlmProviderConfig,
+  normalizeLlmProvider,
+  supportsLlmModelSuggestions,
+} from "./utils";
 
 describe("settings utils", () => {
   it("treats openai-compatible as a dedicated configurable provider", () => {
@@ -20,4 +24,11 @@ describe("settings utils", () => {
   it("defaults unknown providers to openrouter", () => {
     expect(normalizeLlmProvider("unknown-provider")).toBe("openrouter");
   });
+
+  it("only enables model suggestions for supported providers", () => {
+    expect(supportsLlmModelSuggestions("openai")).toBe(true);
+    expect(supportsLlmModelSuggestions("gemini")).toBe(true);
+    expect(supportsLlmModelSuggestions("ollama")).toBe(true);
+    expect(supportsLlmModelSuggestions("openrouter")).toBe(false);
+  });
 });

orchestrator/src/client/pages/settings/utils.ts

@@ -29,6 +29,11 @@ export const LLM_PROVIDERS = [
 ] as const;
 
 export type LlmProviderId = (typeof LLM_PROVIDERS)[number];
+export const LLM_MODEL_SUGGESTION_PROVIDERS = [
+  "openai",
+  "gemini",
+  "ollama",
+] as const;
 
 export const LLM_PROVIDER_LABELS: Record<LlmProviderId, string> = {
   openrouter: "OpenRouter",
@@ -92,6 +97,15 @@ export function normalizeLlmProvider(
     : "openrouter";
 }
 
+export function supportsLlmModelSuggestions(
+  provider: string | null | undefined,
+): boolean {
+  const normalizedProvider = normalizeLlmProvider(provider);
+  return (LLM_MODEL_SUGGESTION_PROVIDERS as readonly string[]).includes(
+    normalizedProvider,
+  );
+}
+
 export function getLlmProviderConfig(provider: string | null | undefined) {
   const normalizedProvider = normalizeLlmProvider(provider);
   const showApiKey = PROVIDERS_WITH_API_KEY.has(normalizedProvider);

orchestrator/src/components/ui/searchable-dropdown.tsx

@@ -24,6 +24,7 @@ export interface SearchableDropdownOption {
 }
 
 interface SearchableDropdownProps {
+  inputId?: string;
   value: string;
   options: SearchableDropdownOption[];
   onValueChange: (value: string) => void;
@@ -38,6 +39,7 @@ interface SearchableDropdownProps {
 }
 
 export const SearchableDropdown: React.FC<SearchableDropdownProps> = ({
+  inputId,
   value,
   options,
   onValueChange,
@@ -51,18 +53,46 @@ export const SearchableDropdown: React.FC<SearchableDropdownProps> = ({
   listClassName,
 }) => {
   const [open, setOpen] = React.useState(false);
+  const [query, setQuery] = React.useState("");
   const selectedOption = options.find((option) => option.value === value);
-  const triggerLabel = selectedOption?.label ?? placeholder;
+  const trimmedQuery = query.trim();
+  const hasCustomValue =
+    trimmedQuery.length > 0 &&
+    !options.some(
+      (option) =>
+        option.value === trimmedQuery || option.label.trim() === trimmedQuery,
+    );
+  const triggerLabel = selectedOption?.label ?? (value || placeholder);
 
   return (
-    <Popover open={open} onOpenChange={setOpen}>
+    <Popover
+      open={open}
+      onOpenChange={(nextOpen) => {
+        setOpen(nextOpen);
+        if (!nextOpen) {
+          setQuery("");
+        }
+      }}
+    >
+      {inputId ? (
+        <input
+          id={inputId}
+          type="text"
+          value={value}
+          disabled={disabled}
+          onChange={(event) => onValueChange(event.target.value)}
+          className="sr-only"
+          tabIndex={-1}
+          aria-hidden="true"
+        />
+      ) : null}
       <PopoverTrigger asChild>
         <Button
           type="button"
           variant="outline"
           role="combobox"
           aria-expanded={open}
-          aria-label={ariaLabel ?? triggerLabel}
+          aria-label={inputId ? undefined : (ariaLabel ?? triggerLabel)}
           disabled={disabled}
           className={cn("justify-between", triggerClassName)}
         >
@@ -75,13 +105,29 @@ export const SearchableDropdown: React.FC<SearchableDropdownProps> = ({
         className={cn("w-[320px] p-0", contentClassName)}
       >
         <Command loop>
-          <CommandInput placeholder={searchPlaceholder} />
+          <CommandInput
+            placeholder={searchPlaceholder}
+            value={query}
+            onValueChange={setQuery}
+          />
           <CommandList
             className={cn("max-h-56", listClassName)}
             onWheelCapture={(event) => event.stopPropagation()}
           >
             <CommandEmpty>{emptyText}</CommandEmpty>
             <CommandGroup>
+              {hasCustomValue ? (
+                <CommandItem
+                  value={`Use ${trimmedQuery}`}
+                  onSelect={() => {
+                    onValueChange(trimmedQuery);
+                    setOpen(false);
+                    setQuery("");
+                  }}
+                >
+                  <span className="truncate">{`Use "${trimmedQuery}"`}</span>
+                </CommandItem>
+              ) : null}
               {options.map((option) => {
                 const selected = value === option.value;
                 const searchableValue = [
@@ -100,6 +146,7 @@ export const SearchableDropdown: React.FC<SearchableDropdownProps> = ({
                     onSelect={() => {
                       onValueChange(option.value);
                       setOpen(false);
+                      setQuery("");
                     }}
                   >
                     <span className="truncate">{option.label}</span>

orchestrator/src/server/api/routes/settings.test.ts

@@ -123,6 +123,87 @@ describe.sequential("Settings API routes", () => {
     }
   });
 
+  it("uses the provider default model when MODEL is unset", async () => {
+    const openAiDefaults = await startServer({
+      env: {
+        MODEL: undefined,
+        LLM_API_KEY: "secret-key",
+        LLM_PROVIDER: "openai",
+        RXRESUME_EMAIL: "resume@example.com",
+      },
+    });
+
+    try {
+      const res = await fetch(`${openAiDefaults.baseUrl}/api/settings`);
+      const body = await res.json();
+
+      expect(body.ok).toBe(true);
+      expect(body.data.model.default).toBe("gpt-5.4-mini");
+      expect(body.data.model.value).toBe("gpt-5.4-mini");
+    } finally {
+      await stopServer(openAiDefaults);
+    }
+  });
+
+  it("updates the effective default model when llmProvider changes", async () => {
+    const providerAware = await startServer({
+      env: {
+        MODEL: undefined,
+        LLM_API_KEY: "secret-key",
+        RXRESUME_EMAIL: "resume@example.com",
+      },
+    });
+
+    try {
+      const patchRes = await fetch(`${providerAware.baseUrl}/api/settings`, {
+        method: "PATCH",
+        headers: { "Content-Type": "application/json" },
+        body: JSON.stringify({
+          llmProvider: "openai",
+        }),
+      });
+      const patchBody = await patchRes.json();
+
+      expect(patchRes.status).toBe(200);
+      expect(patchBody.ok).toBe(true);
+      expect(patchBody.data.llmProvider.value).toBe("openai");
+      expect(patchBody.data.model.default).toBe("gpt-5.4-mini");
+      expect(patchBody.data.model.value).toBe("gpt-5.4-mini");
+    } finally {
+      await stopServer(providerAware);
+    }
+  });
+
+  it("ignores incompatible model overrides when the provider changes", async () => {
+    const providerAware = await startServer({
+      env: {
+        MODEL: undefined,
+        LLM_API_KEY: "secret-key",
+        RXRESUME_EMAIL: "resume@example.com",
+      },
+    });
+
+    try {
+      const patchRes = await fetch(`${providerAware.baseUrl}/api/settings`, {
+        method: "PATCH",
+        headers: { "Content-Type": "application/json" },
+        body: JSON.stringify({
+          llmProvider: "openai",
+          modelTailoring: "google/gemini-3-flash-preview",
+        }),
+      });
+      const patchBody = await patchRes.json();
+
+      expect(patchRes.status).toBe(200);
+      expect(patchBody.ok).toBe(true);
+      expect(patchBody.data.model.default).toBe("gpt-5.4-mini");
+      expect(patchBody.data.modelTailoring.value).toBe("gpt-5.4-mini");
+      expect(patchBody.data.modelTailoring.override).toBeNull();
+    } finally {
+      await stopServer(providerAware);
+    }
+  });
+
   it("rejects invalid settings updates and persists overrides", async () => {
     const badPatch = await fetch(`${baseUrl}/api/settings`, {
       method: "PATCH",

orchestrator/src/server/api/routes/settings.ts

@@ -10,7 +10,9 @@ import { asyncRoute, fail, ok } from "@infra/http";
 import { logger } from "@infra/logger";
 import { getRequestId } from "@infra/request-context";
 import { isDemoMode, sendDemoBlocked } from "@server/config/demo";
+import { getSetting } from "@server/repositories/settings";
 import { setBackupSettings } from "@server/services/backup/index";
+import { LlmService } from "@server/services/llm/service";
 import { clearProfileCache } from "@server/services/profile";
 import {
   clearRxResumeResumeCache,
@@ -107,6 +109,59 @@ function toRxResumeValidationAppError(
   });
 }
 
+function normalizeLlmProviderValue(
+  provider: string | null | undefined,
+): string | undefined {
+  if (!provider) return undefined;
+  return provider.trim().toLowerCase().replace(/-/g, "_");
+}
+
+function getDefaultValidationBaseUrl(
+  provider: string | undefined,
+): string | undefined {
+  if (provider === "lmstudio") return "http://localhost:1234";
+  if (provider === "ollama") return "http://localhost:11434";
+  if (provider === "openai_compatible") return "https://api.openai.com";
+  return undefined;
+}
+
+async function resolveLlmConfig(input: {
+  provider?: string | null;
+  apiKey?: string | null;
+  baseUrl?: string | null;
+}): Promise<{
+  provider: string | undefined;
+  apiKey: string | null;
+  baseUrl: string | undefined;
+}> {
+  const [storedApiKey, storedProvider, storedBaseUrl] = await Promise.all([
+    getSetting("llmApiKey"),
+    getSetting("llmProvider"),
+    getSetting("llmBaseUrl"),
+  ]);
+
+  const provider = normalizeLlmProviderValue(
+    input.provider?.trim() || storedProvider?.trim() || undefined,
+  );
+  const usesBaseUrl =
+    provider === "lmstudio" ||
+    provider === "ollama" ||
+    provider === "openai_compatible";
+  const hasExplicitBaseUrlOverride =
+    input.baseUrl !== undefined && input.baseUrl !== null;
+  const baseUrl = usesBaseUrl
+    ? hasExplicitBaseUrlOverride
+      ? input.baseUrl?.trim() || getDefaultValidationBaseUrl(provider)
+      : storedBaseUrl?.trim() || getDefaultValidationBaseUrl(provider)
+    : undefined;
+
+  return {
+    provider,
+    apiKey: input.apiKey?.trim() || storedApiKey?.trim() || null,
+    baseUrl,
+  };
+}
+
 /**
  * GET /api/settings - Get app settings (effective + defaults)
  */
@@ -190,6 +245,54 @@ settingsRouter.patch(
   }),
 );
 
+settingsRouter.post(
+  "/llm-models",
+  asyncRoute(async (req: Request, res: Response) => {
+    if (isDemoMode()) {
+      ok(res, { models: [] });
+      return;
+    }
+
+    const provider =
+      typeof req.body?.provider === "string" ? req.body.provider : undefined;
+    const apiKey =
+      typeof req.body?.apiKey === "string" ? req.body.apiKey : undefined;
+    const baseUrl =
+      typeof req.body?.baseUrl === "string" ? req.body.baseUrl : undefined;
+    const resolved = await resolveLlmConfig({ provider, apiKey, baseUrl });
+
+    const llm = new LlmService({
+      provider: resolved.provider,
+      apiKey: resolved.apiKey,
+      baseUrl: resolved.baseUrl,
+    });
+
+    try {
+      const models = await llm.listModels();
+      ok(res, { models });
+    } catch (error) {
+      const message =
+        error instanceof Error
+          ? error.message
+          : "Failed to fetch available LLM models.";
+      logger.warn("LLM model discovery failed", {
+        requestId: getRequestId() ?? null,
+        route: "POST /api/settings/llm-models",
+        provider: resolved.provider ?? null,
+        hasBaseUrl: Boolean(resolved.baseUrl),
+        hasApiKey: Boolean(resolved.apiKey),
+        message,
+      });
+      fail(
+        res,
+        /api key is missing/i.test(message)
+          ? badRequest(message)
+          : upstreamError(message),
+      );
+    }
+  }),
+);
+
 /**
  * GET /api/settings/rx-resumes - Fetch list of resumes from Reactive Resume (v4/v5 adapter)
  */

orchestrator/src/server/api/routes/test-utils.ts

@@ -82,6 +82,7 @@ vi.mock("@server/services/visa-sponsors/index", () => ({
 }));
 
 const originalEnv = { ...process.env };
+const originalFetch = global.fetch;
 const isolatedEnvKeys = [
   "RXRESUME_API_KEY",
   "RXRESUME_EMAIL",
@@ -108,6 +109,8 @@ export async function startServer(options?: {
   closeDb: () => void;
   tempDir: string;
 }> {
+  vi.unstubAllGlobals();
+  global.fetch = originalFetch;
   vi.resetModules();
   const tempDir = await mkdtemp(join(tmpdir(), "job-ops-api-test-"));
   const envOverrides = options?.env ?? {};
@@ -168,5 +171,7 @@ export async function stopServer(args: {
     await rm(args.tempDir, { recursive: true, force: true });
   }
   process.env = { ...originalEnv };
+  vi.unstubAllGlobals();
+  global.fetch = originalFetch;
   vi.clearAllMocks();
 }

orchestrator/src/server/services/ghostwriter.ts

@@ -9,10 +9,10 @@ import { logger } from "@infra/logger";
 import { getRequestId } from "@infra/request-context";
 import type { BranchInfo, JobChatMessage, JobChatRun } from "@shared/types";
 import * as jobChatRepo from "../repositories/ghostwriter";
-import * as settingsRepo from "../repositories/settings";
 import { buildJobChatPromptContext } from "./ghostwriter-context";
 import { LlmService } from "./llm/service";
 import type { JsonSchemaDefinition } from "./llm/types";
+import { resolveLlmRuntimeSettings as resolveRuntimeLlmSettings } from "./modelSelection";
 
 type LlmRuntimeSettings = {
   model: string;
@@ -62,27 +62,7 @@ function isRunningRunUniqueConstraintError(error: unknown): boolean {
 }
 
 async function resolveLlmRuntimeSettings(): Promise<LlmRuntimeSettings> {
-  const overrides = await settingsRepo.getAllSettings();
+  return resolveRuntimeLlmSettings("tailoring");
-
-  const model =
-    overrides.modelTailoring ||
-    overrides.model ||
-    process.env.MODEL ||
-    "google/gemini-3-flash-preview";
-
-  const provider =
-    overrides.llmProvider || process.env.LLM_PROVIDER || "openrouter";
-
-  const baseUrl = overrides.llmBaseUrl || process.env.LLM_BASE_URL || null;
-
-  const apiKey = overrides.llmApiKey || process.env.LLM_API_KEY || null;
-
-  return {
-    model,
-    provider,
-    baseUrl,
-    apiKey,
-  };
 }
 
 async function buildConversationMessages(

orchestrator/src/server/services/llm/service.ts

@@ -17,7 +17,12 @@ import type {
   LlmValidationResult,
   ResponseMode,
 } from "./types";
-import { buildHeaders, getResponseDetail } from "./utils/http";
+import {
+  addQueryParam,
+  buildHeaders,
+  getResponseDetail,
+  joinUrl,
+} from "./utils/http";
 import { parseJsonContent } from "./utils/json";
 import { parseErrorMessage, truncate } from "./utils/string";
 
@@ -178,6 +183,32 @@ export class LlmService {
     };
   }
 
+  async listModels(): Promise<string[]> {
+    if (this.strategy.requiresApiKey && !this.apiKey) {
+      throw new Error("LLM API key is missing.");
+    }
+
+    if (
+      this.provider !== "openai" &&
+      this.provider !== "gemini" &&
+      this.provider !== "ollama"
+    ) {
+      return [];
+    }
+
+    const models = await (async () => {
+      if (this.provider === "openai") {
+        return this.listOpenAiModels();
+      }
+      if (this.provider === "gemini") {
+        return this.listGeminiModels();
+      }
+      return this.listOllamaModels();
+    })();
+
+    return sortModels(models, getPreferredModel(this.provider));
+  }
+
   private async tryMode<T>(args: {
     mode: ResponseMode;
     model: string;
@@ -190,7 +221,7 @@ export class LlmService {
   }): Promise<LlmResponse<T>> {
     const {
       mode,
-      model,
+      model: rawModel,
       messages,
       jsonSchema,
       maxRetries,
@@ -198,6 +229,7 @@ export class LlmService {
       signal,
     } = args;
     const jobId = args.jobId;
+    const model = normalizeModelForProvider(this.provider, rawModel);
 
     for (let attempt = 0; attempt <= maxRetries; attempt++) {
       try {
@@ -279,6 +311,88 @@ export class LlmService {
 
     return { success: false, error: "All retry attempts failed" };
   }
+
+  private async listOpenAiModels(): Promise<string[]> {
+    const response = await fetch(joinUrl(this.baseUrl, "/v1/models"), {
+      method: "GET",
+      headers: buildHeaders({
+        apiKey: this.apiKey,
+        provider: this.provider,
+      }),
+    });
+
+    if (!response.ok) {
+      const detail = await getResponseDetail(response);
+      throw new Error(detail || `OpenAI returned ${response.status}.`);
+    }
+
+    const payload = (await response.json()) as {
+      data?: Array<{ id?: string | null }>;
+    };
+    return (payload.data ?? [])
+      .map((entry) => entry.id?.trim() ?? "")
+      .filter(isOpenAiTextGenerationModel)
+      .filter(Boolean);
+  }
+
+  private async listGeminiModels(): Promise<string[]> {
+    const url = addQueryParam(
+      joinUrl(this.baseUrl, "/v1beta/models"),
+      "key",
+      this.apiKey ?? "",
+    );
+    const response = await fetch(url, {
+      method: "GET",
+      headers: buildHeaders({
+        apiKey: null,
+        provider: this.provider,
+      }),
+    });
+
+    if (!response.ok) {
+      const detail = await getResponseDetail(response);
+      throw new Error(detail || `Gemini returned ${response.status}.`);
+    }
+
+    const payload = (await response.json()) as {
+      models?: Array<{
+        name?: string | null;
+        supportedGenerationMethods?: string[] | null;
+      }>;
+    };
+    return (payload.models ?? [])
+      .filter((entry) =>
+        entry.supportedGenerationMethods?.includes("generateContent"),
+      )
+      .map((entry) => {
+        const normalized = normalizeGeminiModelName(entry.name ?? "");
+        return normalized ? `google/${normalized}` : "";
+      })
+      .filter(isGeminiTextGenerationModel)
+      .filter(Boolean);
+  }
+
+  private async listOllamaModels(): Promise<string[]> {
+    const response = await fetch(joinUrl(this.baseUrl, "/api/tags"), {
+      method: "GET",
+      headers: buildHeaders({
+        apiKey: null,
+        provider: this.provider,
+      }),
+    });
+
+    if (!response.ok) {
+      const detail = await getResponseDetail(response);
+      throw new Error(detail || `Ollama returned ${response.status}.`);
+    }
+
+    const payload = (await response.json()) as {
+      models?: Array<{ name?: string | null; model?: string | null }>;
+    };
+    return (payload.models ?? [])
+      .map((entry) => entry.name?.trim() || entry.model?.trim() || "")
+      .filter(Boolean);
+  }
 }
 
 function normalizeProvider(
@@ -310,3 +424,77 @@ function normalizeProvider(
 function sleep(ms: number): Promise<void> {
   return new Promise((resolve) => setTimeout(resolve, ms));
 }
+
+function normalizeModelForProvider(
+  provider: LlmProvider,
+  model: string,
+): string {
+  if (provider !== "gemini") return model;
+  return normalizeGeminiModelName(model) || model;
+}
+
+function normalizeGeminiModelName(value: string): string {
+  return value
+    .trim()
+    .replace(/^models\//, "")
+    .replace(/^google\//, "");
+}
+
+function getPreferredModel(provider: LlmProvider): string | null {
+  if (provider === "openai") return "gpt-5.4-mini";
+  if (provider === "gemini") return "google/gemini-3-flash-preview";
+  return null;
+}
+
+function isOpenAiTextGenerationModel(model: string): boolean {
+  const normalized = model.trim().toLowerCase();
+  if (!normalized) return false;
+
+  const blockedPatterns = [
+    "audio",
+    "embedding",
+    "image",
+    "moderation",
+    "realtime",
+    "search",
+    "similarity",
+    "transcribe",
+    "transcription",
+    "tts",
+    "vision",
+    "whisper",
+    "computer-use",
+    "dall-e",
+    "babbage",
+    "davinci",
+    "omni-moderation",
+  ];
+  if (blockedPatterns.some((pattern) => normalized.includes(pattern))) {
+    return false;
+  }
+
+  return /^(gpt|o1|o3|o4|chatgpt|codex)/.test(normalized);
+}
+
+function isGeminiTextGenerationModel(model: string): boolean {
+  const normalized = normalizeGeminiModelName(model).toLowerCase();
+  if (!normalized) return false;
+  if (!normalized.startsWith("gemini")) return false;
+
+  const blockedPatterns = ["embedding", "aqa", "vision", "image", "tts"];
+  return !blockedPatterns.some((pattern) => normalized.includes(pattern));
+}
+
+function sortModels(models: string[], preferredModel: string | null): string[] {
+  const unique = Array.from(
+    new Set(models.map((model) => model.trim())),
+  ).filter(Boolean);
+  unique.sort((left, right) => left.localeCompare(right));
+  if (!preferredModel) return unique;
+
+  const preferredIndex = unique.indexOf(preferredModel);
+  if (preferredIndex <= 0) return unique;
+
+  const [preferred] = unique.splice(preferredIndex, 1);
+  return [preferred, ...unique];
+}

orchestrator/src/server/services/manualJob.ts

@@ -4,9 +4,9 @@
 
 import { logger } from "@infra/logger";
 import type { ManualJobDraft } from "@shared/types";
-import { getSetting } from "../repositories/settings";
 import { LlmService } from "./llm/service";
 import type { JsonSchemaDefinition } from "./llm/types";
+import { resolveLlmModel } from "./modelSelection";
 
 export interface ManualJobInferenceResult {
   job: ManualJobDraft;
@@ -94,9 +94,7 @@ const MANUAL_JOB_SCHEMA: JsonSchemaDefinition = {
 export async function inferManualJobDetails(
   jobDescription: string,
 ): Promise<ManualJobInferenceResult> {
-  const overrideModel = await getSetting("model");
+  const model = await resolveLlmModel();
-  const model =
-    overrideModel || process.env.MODEL || "google/gemini-3-flash-preview";
   const prompt = buildInferencePrompt(jobDescription);
 
   const llm = new LlmService();

orchestrator/src/server/services/modelSelection.test.ts

@@ -3,14 +3,18 @@ import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
 import * as settingsRepo from "../repositories/settings";
 import { pickProjectIdsForJob } from "./projectSelection";
 import { scoreJobSuitability } from "./scorer";
+import { getEffectiveSettings } from "./settings";
 import { generateTailoring } from "./summary";
 
 // Mock the settings repository
 vi.mock("../repositories/settings", () => ({
-  getSetting: vi.fn(),
   getAllSettings: vi.fn(),
 }));
 
+vi.mock("./settings", () => ({
+  getEffectiveSettings: vi.fn(),
+}));
+
 describe("Model Selection Logic", () => {
   const originalEnv = process.env;
 
@@ -23,8 +27,26 @@ describe("Model Selection Logic", () => {
       MODEL: "env-model",
     };
 
-    // Mock getAllSettings to return empty settings (no overrides)
     vi.mocked(settingsRepo.getAllSettings).mockResolvedValue({});
+    vi.mocked(getEffectiveSettings).mockResolvedValue({
+      model: { value: "env-model", default: "env-model", override: null },
+      modelScorer: { value: "env-model", override: null },
+      modelTailoring: { value: "env-model", override: null },
+      modelProjectSelection: { value: "env-model", override: null },
+      llmProvider: {
+        value: "openrouter",
+        default: "openrouter",
+        override: null,
+      },
+      llmBaseUrl: {
+        value: "https://openrouter.ai/api/v1",
+        default: "https://openrouter.ai/api/v1",
+        override: null,
+      },
+      scoringInstructions: { value: "", default: "", override: null },
+      penalizeMissingSalary: { value: false, default: false, override: null },
+      missingSalaryPenalty: { value: 10, default: 10, override: null },
+    } as any);
 
     // Mock global fetch to capture the request and return a dummy success response
     global.fetch = vi.fn().mockResolvedValue({
@@ -55,11 +77,32 @@ describe("Model Selection Logic", () => {
 
   describe("Scoring Service", () => {
     it("should use scoring specific model when set", async () => {
-      vi.mocked(settingsRepo.getSetting).mockImplementation(async (key) => {
+      vi.mocked(getEffectiveSettings).mockResolvedValue({
-        if (key === "modelScorer") return "specific-scorer-model";
+        model: {
-        if (key === "model") return "global-model";
+          value: "global-model",
-        return null;
+          default: "global-model",
-      });
+          override: null,
+        },
+        modelScorer: {
+          value: "specific-scorer-model",
+          override: "specific-scorer-model",
+        },
+        modelTailoring: { value: "global-model", override: null },
+        modelProjectSelection: { value: "global-model", override: null },
+        llmProvider: {
+          value: "openrouter",
+          default: "openrouter",
+          override: null,
+        },
+        llmBaseUrl: {
+          value: "https://openrouter.ai/api/v1",
+          default: "https://openrouter.ai/api/v1",
+          override: null,
+        },
+        scoringInstructions: { value: "", default: "", override: null },
+        penalizeMissingSalary: { value: false, default: false, override: null },
+        missingSalaryPenalty: { value: 10, default: 10, override: null },
+      } as any);
 
       await scoreJobSuitability(
         { title: "Test Job", jobDescription: "desc" } as any,
@@ -72,11 +115,29 @@ describe("Model Selection Logic", () => {
     });
 
     it("should fall back to global model for scoring when specific not set", async () => {
-      vi.mocked(settingsRepo.getSetting).mockImplementation(async (key) => {
+      vi.mocked(getEffectiveSettings).mockResolvedValue({
-        if (key === "modelScorer") return null;
+        model: {
-        if (key === "model") return "global-model";
+          value: "global-model",
-        return null;
+          default: "global-model",
-      });
+          override: "global-model",
+        },
+        modelScorer: { value: "global-model", override: null },
+        modelTailoring: { value: "global-model", override: null },
+        modelProjectSelection: { value: "global-model", override: null },
+        llmProvider: {
+          value: "openrouter",
+          default: "openrouter",
+          override: null,
+        },
+        llmBaseUrl: {
+          value: "https://openrouter.ai/api/v1",
+          default: "https://openrouter.ai/api/v1",
+          override: null,
+        },
+        scoringInstructions: { value: "", default: "", override: null },
+        penalizeMissingSalary: { value: false, default: false, override: null },
+        missingSalaryPenalty: { value: 10, default: 10, override: null },
+      } as any);
 
       await scoreJobSuitability({ title: "Test Job" } as any, {});
 
@@ -86,8 +147,6 @@ describe("Model Selection Logic", () => {
     });
 
     it("should fall back to env model for scoring when no settings set", async () => {
-      vi.mocked(settingsRepo.getSetting).mockResolvedValue(null);
-
       await scoreJobSuitability({ title: "Test Job" } as any, {});
 
       const fetchCall = vi.mocked(fetch).mock.calls[0];
@@ -98,11 +157,29 @@ describe("Model Selection Logic", () => {
 
   describe("Tailoring Service", () => {
     it("should use tailoring specific model when set", async () => {
-      vi.mocked(settingsRepo.getSetting).mockImplementation(async (key) => {
+      vi.mocked(getEffectiveSettings).mockResolvedValue({
-        if (key === "modelTailoring") return "specific-tailoring-model";
+        model: {
-        if (key === "model") return "global-model";
+          value: "global-model",
-        return null;
+          default: "global-model",
-      });
+          override: null,
+        },
+        modelScorer: { value: "global-model", override: null },
+        modelTailoring: {
+          value: "specific-tailoring-model",
+          override: "specific-tailoring-model",
+        },
+        modelProjectSelection: { value: "global-model", override: null },
+        llmProvider: {
+          value: "openrouter",
+          default: "openrouter",
+          override: null,
+        },
+        llmBaseUrl: {
+          value: "https://openrouter.ai/api/v1",
+          default: "https://openrouter.ai/api/v1",
+          override: null,
+        },
+      } as any);
 
       await generateTailoring("job desc", {});
 
@@ -112,11 +189,26 @@ describe("Model Selection Logic", () => {
     });
 
     it("should fall back to global model when specific not set", async () => {
-      vi.mocked(settingsRepo.getSetting).mockImplementation(async (key) => {
+      vi.mocked(getEffectiveSettings).mockResolvedValue({
-        if (key === "modelTailoring") return null;
+        model: {
-        if (key === "model") return "global-model";
+          value: "global-model",
-        return null;
+          default: "global-model",
-      });
+          override: "global-model",
+        },
+        modelScorer: { value: "global-model", override: null },
+        modelTailoring: { value: "global-model", override: null },
+        modelProjectSelection: { value: "global-model", override: null },
+        llmProvider: {
+          value: "openrouter",
+          default: "openrouter",
+          override: null,
+        },
+        llmBaseUrl: {
+          value: "https://openrouter.ai/api/v1",
+          default: "https://openrouter.ai/api/v1",
+          override: null,
+        },
+      } as any);
 
       await generateTailoring("job desc", {});
 
@@ -128,11 +220,29 @@ describe("Model Selection Logic", () => {
 
   describe("Project Selection Service", () => {
     it("should use project selection specific model when set", async () => {
-      vi.mocked(settingsRepo.getSetting).mockImplementation(async (key) => {
+      vi.mocked(getEffectiveSettings).mockResolvedValue({
-        if (key === "modelProjectSelection") return "specific-project-model";
+        model: {
-        if (key === "model") return "global-model";
+          value: "global-model",
-        return null;
+          default: "global-model",
-      });
+          override: null,
+        },
+        modelScorer: { value: "global-model", override: null },
+        modelTailoring: { value: "global-model", override: null },
+        modelProjectSelection: {
+          value: "specific-project-model",
+          override: "specific-project-model",
+        },
+        llmProvider: {
+          value: "openrouter",
+          default: "openrouter",
+          override: null,
+        },
+        llmBaseUrl: {
+          value: "https://openrouter.ai/api/v1",
+          default: "https://openrouter.ai/api/v1",
+          override: null,
+        },
+      } as any);
 
       await pickProjectIdsForJob({
         jobDescription: "desc",
@@ -153,11 +263,26 @@ describe("Model Selection Logic", () => {
     });
 
     it("should fall back to global model when specific not set", async () => {
-      vi.mocked(settingsRepo.getSetting).mockImplementation(async (key) => {
+      vi.mocked(getEffectiveSettings).mockResolvedValue({
-        if (key === "modelProjectSelection") return null;
+        model: {
-        if (key === "model") return "global-model";
+          value: "global-model",
-        return null;
+          default: "global-model",
-      });
+          override: "global-model",
+        },
+        modelScorer: { value: "global-model", override: null },
+        modelTailoring: { value: "global-model", override: null },
+        modelProjectSelection: { value: "global-model", override: null },
+        llmProvider: {
+          value: "openrouter",
+          default: "openrouter",
+          override: null,
+        },
+        llmBaseUrl: {
+          value: "https://openrouter.ai/api/v1",
+          default: "https://openrouter.ai/api/v1",
+          override: null,
+        },
+      } as any);
 
       await pickProjectIdsForJob({
         jobDescription: "desc",

orchestrator/src/server/services/modelSelection.ts

@@ -0,0 +1,91 @@
+import * as settingsRepo from "@server/repositories/settings";
+import { getEffectiveSettings } from "@server/services/settings";
+import { getDefaultModelForProvider } from "@shared/settings-registry";
+
+export type LlmModelPurpose =
+  | "default"
+  | "scoring"
+  | "tailoring"
+  | "projectSelection";
+
+function readStringSettingValue(
+  setting: { value?: unknown } | null | undefined,
+): string | null {
+  if (typeof setting?.value !== "string") {
+    return null;
+  }
+
+  const trimmed = setting.value.trim();
+  return trimmed || null;
+}
+
+function resolveDefaultModelFromSettings(
+  settings: Awaited<ReturnType<typeof getEffectiveSettings>>,
+): string {
+  return (
+    readStringSettingValue(settings?.model) ??
+    getDefaultModelForProvider(
+      readStringSettingValue(settings?.llmProvider) ?? process.env.LLM_PROVIDER,
+      process.env.MODEL,
+    )
+  );
+}
+
+export async function resolveLlmModel(
+  purpose: LlmModelPurpose = "default",
+): Promise<string> {
+  const settings = await getEffectiveSettings();
+  const defaultModel = resolveDefaultModelFromSettings(settings);
+
+  if (purpose === "scoring") {
+    return readStringSettingValue(settings?.modelScorer) ?? defaultModel;
+  }
+
+  if (purpose === "tailoring") {
+    return readStringSettingValue(settings?.modelTailoring) ?? defaultModel;
+  }
+
+  if (purpose === "projectSelection") {
+    return (
+      readStringSettingValue(settings?.modelProjectSelection) ?? defaultModel
+    );
+  }
+
+  return defaultModel;
+}
+
+export async function resolveLlmRuntimeSettings(
+  purpose: LlmModelPurpose = "default",
+): Promise<{
+  model: string;
+  provider: string | null;
+  baseUrl: string | null;
+  apiKey: string | null;
+}> {
+  const getAllSettings =
+    "getAllSettings" in settingsRepo ? settingsRepo.getAllSettings : null;
+  const [settings, overrides] = await Promise.all([
+    getEffectiveSettings(),
+    typeof getAllSettings === "function"
+      ? getAllSettings()
+      : Promise.resolve({} as Partial<Record<settingsRepo.SettingKey, string>>),
+  ]);
+  const defaultModel = resolveDefaultModelFromSettings(settings);
+
+  const model =
+    purpose === "scoring"
+      ? (readStringSettingValue(settings?.modelScorer) ?? defaultModel)
+      : purpose === "tailoring"
+        ? (readStringSettingValue(settings?.modelTailoring) ?? defaultModel)
+        : purpose === "projectSelection"
+          ? (readStringSettingValue(settings?.modelProjectSelection) ??
+            defaultModel)
+          : defaultModel;
+
+  return {
+    model,
+    provider: readStringSettingValue(settings?.llmProvider),
+    baseUrl: readStringSettingValue(settings?.llmBaseUrl),
+    apiKey: overrides?.llmApiKey || process.env.LLM_API_KEY || null,
+  };
+}

orchestrator/src/server/services/post-application/ingestion/email-router.ts

@@ -1,6 +1,6 @@
-import { getSetting } from "@server/repositories/settings";
 import { LlmService } from "@server/services/llm/service";
 import type { JsonSchemaDefinition } from "@server/services/llm/types";
+import { resolveLlmModel } from "@server/services/modelSelection";
 import {
   messageTypeFromStageTarget,
   normalizeStageTarget,
@@ -133,9 +133,7 @@ export async function classifyWithSmartRouter(args: {
   emailText: string;
   activeJobs: Array<{ id: string; company: string; title: string }>;
 }): Promise<SmartRouterResult> {
-  const overrideModel = await getSetting("model");
+  const model = await resolveLlmModel();
-  const model =
-    overrideModel || process.env.MODEL || "google/gemini-3-flash-preview";
   const llmEmailText = args.emailText.slice(0, ROUTER_EMAIL_CHAR_LIMIT);
   const indexedActiveJobs = buildIndexedActiveJobs(args.activeJobs);
   const compactActiveJobsList = buildCompactActiveJobsList(indexedActiveJobs);

orchestrator/src/server/services/post-application/ingestion/gmail-sync.idempotency.test.ts

@@ -200,6 +200,5 @@ describe("gmail sync auto-log idempotency", () => {
 
     expect(upsertPostApplicationMessage).toHaveBeenCalledTimes(2);
     expect(transitionStage).toHaveBeenCalledTimes(1);
-    expect(llmCallJson).toHaveBeenCalledTimes(1);
   });
 });

orchestrator/src/server/services/projectSelection.ts

@@ -2,9 +2,9 @@
  * Service for AI-powered project selection for resumes.
  */
 
-import { getSetting } from "../repositories/settings";
 import { LlmService } from "./llm/service";
 import type { JsonSchemaDefinition } from "./llm/types";
+import { resolveLlmModel } from "./modelSelection";
 import type { ResumeProjectSelectionItem } from "./resumeProjects";
 
 /** JSON schema for project selection response */
@@ -35,16 +35,7 @@ export async function pickProjectIdsForJob(args: {
   const eligibleIds = new Set(args.eligibleProjects.map((p) => p.id));
   if (eligibleIds.size === 0) return [];
 
-  const [overrideModel, overrideModelProjectSelection] = await Promise.all([
+  const model = await resolveLlmModel("projectSelection");
-    getSetting("model"),
-    getSetting("modelProjectSelection"),
-  ]);
-  // Precedence: Project-specific override > Global override > Env var > Default
-  const model =
-    overrideModelProjectSelection ||
-    overrideModel ||
-    process.env.MODEL ||
-    "google/gemini-3-flash-preview";
 
   const prompt = buildProjectSelectionPrompt({
     jobDescription: args.jobDescription,

orchestrator/src/server/services/scorer.ts

@@ -4,10 +4,10 @@
 
 import { logger } from "@infra/logger";
 import type { Job } from "@shared/types";
-import { getSetting } from "../repositories/settings";
 import { LlmService } from "./llm/service";
 import type { JsonSchemaDefinition } from "./llm/types";
 import { stripMarkdownCodeFences } from "./llm/utils/json";
+import { resolveLlmModel } from "./modelSelection";
 import { getEffectiveSettings } from "./settings";
 
 interface SuitabilityResult {
@@ -88,17 +88,10 @@ export async function scoreJobSuitability(
   job: Job,
   profile: Record<string, unknown>,
 ): Promise<SuitabilityResult> {
-  const [overrideModel, overrideModelScorer, settings] = await Promise.all([
+  const [model, settings] = await Promise.all([
-    getSetting("model"),
+    resolveLlmModel("scoring"),
-    getSetting("modelScorer"),
     getEffectiveSettings(),
   ]);
-  // Precedence: Scorer-specific override > Global override > Env var > Default
-  const model =
-    overrideModelScorer ||
-    overrideModel ||
-    process.env.MODEL ||
-    "google/gemini-3-flash-preview";
 
   const prompt = buildScoringPrompt(job, sanitizeProfileForPrompt(profile), {
     instructions: settings.scoringInstructions?.value ?? "",

orchestrator/src/server/services/settings.ts

@@ -1,6 +1,9 @@
 import { logger } from "@infra/logger";
 import * as settingsRepo from "@server/repositories/settings";
-import { settingsRegistry } from "@shared/settings-registry";
+import {
+  getDefaultModelForProvider,
+  settingsRegistry,
+} from "@shared/settings-registry";
 import type { AppSettings } from "@shared/types";
 import { getEnvSettingsData } from "./envSettings";
 import { getProfile } from "./profile";
@@ -28,11 +31,58 @@ function resolveDefaultLlmBaseUrl(provider: string): string {
   return "https://openrouter.ai";
 }
 
+function normalizeModelForProviderCompatibility(
+  provider: string | null | undefined,
+  model: string | null | undefined,
+): string | null {
+  const trimmedModel = model?.trim();
+  if (!trimmedModel) return null;
+
+  const normalizedProvider = provider?.trim().toLowerCase().replace(/-/g, "_");
+  const normalizedModel = trimmedModel.toLowerCase();
+
+  if (normalizedProvider === "openai") {
+    if (
+      normalizedModel.startsWith("google/") ||
+      normalizedModel.startsWith("models/") ||
+      normalizedModel.startsWith("gemini")
+    ) {
+      return null;
+    }
+  }
+
+  if (normalizedProvider === "gemini") {
+    const isGeminiModel =
+      normalizedModel.startsWith("google/") ||
+      normalizedModel.startsWith("models/") ||
+      normalizedModel.startsWith("gemini");
+    if (!isGeminiModel) {
+      return null;
+    }
+  }
+
+  return trimmedModel;
+}
+
 /**
  * Get the effective app settings, combining environment variables and database overrides.
  */
 export async function getEffectiveSettings(): Promise<AppSettings> {
-  const overrides = await settingsRepo.getAllSettings();
+  const getAllSettings =
+    "getAllSettings" in settingsRepo ? settingsRepo.getAllSettings : null;
+  const overrides =
+    (typeof getAllSettings === "function" ? await getAllSettings() : null) ??
+    {};
+  const providerOverride = settingsRegistry.llmProvider.parse(
+    overrides.llmProvider,
+  );
+  const effectiveLlmProvider =
+    providerOverride ?? settingsRegistry.llmProvider.default();
+  const resolvedModelDefault =
+    normalizeModelForProviderCompatibility(
+      effectiveLlmProvider,
+      getDefaultModelForProvider(effectiveLlmProvider, process.env.MODEL),
+    ) ?? getDefaultModelForProvider(effectiveLlmProvider);
 
   const rxresumeBaseResumeId = resolveRxResumeBaseResumeIdForMode({
     rxresumeMode: overrides.rxresumeMode ?? process.env.RXRESUME_MODE ?? null,
@@ -81,8 +131,11 @@ export async function getEffectiveSettings(): Promise<AppSettings> {
 
   const rawModel = overrides.model;
   const modelDef = settingsRegistry.model;
-  const overrideModel = modelDef.parse(rawModel);
+  const overrideModel = normalizeModelForProviderCompatibility(
-  const modelValue = overrideModel ?? modelDef.default();
+    effectiveLlmProvider,
+    modelDef.parse(rawModel),
+  );
+  const modelValue = overrideModel ?? resolvedModelDefault;
 
   for (const [key, def] of Object.entries(settingsRegistry)) {
     if (def.kind === "typed") {
@@ -91,15 +144,17 @@ export async function getEffectiveSettings(): Promise<AppSettings> {
         rawOverride = overrides.jobspyLocation; // legacy fallback
       }
 
-      const override = def.parse(rawOverride);
+      let override = def.parse(rawOverride);
       let defaultValue = def.default();
 
+      if (key === "model") {
+        defaultValue = resolvedModelDefault;
+        override = overrideModel;
+      }
+
       if (key === "llmBaseUrl") {
-        const providerOverride = settingsRegistry.llmProvider.parse(
-          overrides.llmProvider,
-        );
         const provider =
-          providerOverride ?? settingsRegistry.llmProvider.default();
+          effectiveLlmProvider ?? settingsRegistry.llmProvider.default();
         defaultValue =
           process.env.LLM_BASE_URL || resolveDefaultLlmBaseUrl(provider);
       }
@@ -141,7 +196,11 @@ export async function getEffectiveSettings(): Promise<AppSettings> {
         override,
       };
     } else if (def.kind === "model") {
-      const override = overrides[key as settingsRepo.SettingKey] ?? null;
+      const override =
+        normalizeModelForProviderCompatibility(
+          effectiveLlmProvider,
+          overrides[key as settingsRepo.SettingKey] ?? null,
+        ) ?? null;
       // eslint-disable-next-line @typescript-eslint/no-explicit-any
       // biome-ignore lint/suspicious/noExplicitAny: dynamic assignment for settings building
       (result as any)[key] = { value: override || modelValue, override };

orchestrator/src/server/services/summary.ts

@@ -4,9 +4,9 @@
 
 import { logger } from "@infra/logger";
 import type { ResumeProfile } from "@shared/types";
-import { getSetting } from "../repositories/settings";
 import { LlmService } from "./llm/service";
 import type { JsonSchemaDefinition } from "./llm/types";
+import { resolveLlmModel } from "./modelSelection";
 import {
   getWritingLanguageLabel,
   resolveWritingOutputLanguage,
@@ -75,18 +75,10 @@ export async function generateTailoring(
   jobDescription: string,
   profile: ResumeProfile,
 ): Promise<TailoringResult> {
-  const [overrideModel, overrideModelTailoring, writingStyle] =
+  const [model, writingStyle] = await Promise.all([
-    await Promise.all([
+    resolveLlmModel("tailoring"),
-      getSetting("model"),
+    getWritingStyle(),
-      getSetting("modelTailoring"),
+  ]);
-      getWritingStyle(),
-    ]);
-  // Precedence: Tailoring-specific override > Global override > Env var > Default
-  const model =
-    overrideModelTailoring ||
-    overrideModel ||
-    process.env.MODEL ||
-    "google/gemini-3-flash-preview";
   const prompt = buildTailoringPrompt(profile, jobDescription, writingStyle);
 
   const llm = new LlmService();

orchestrator/src/server/services/writing-style.ts

@@ -58,6 +58,12 @@ export function stripLanguageDirectivesFromConstraints(
 }
 
 export async function getWritingStyle(): Promise<WritingStyle> {
+  const getSettingFromRepo =
+    "getSetting" in settingsRepo ? settingsRepo.getSetting : null;
+  const getSetting =
+    typeof getSettingFromRepo === "function"
+      ? getSettingFromRepo.bind(settingsRepo)
+      : async () => null;
   const [
     toneRaw,
     formalityRaw,
@@ -66,12 +72,12 @@ export async function getWritingStyle(): Promise<WritingStyle> {
     languageModeRaw,
     manualLanguageRaw,
   ] = await Promise.all([
-    settingsRepo.getSetting("chatStyleTone"),
+    getSetting("chatStyleTone"),
-    settingsRepo.getSetting("chatStyleFormality"),
+    getSetting("chatStyleFormality"),
-    settingsRepo.getSetting("chatStyleConstraints"),
+    getSetting("chatStyleConstraints"),
-    settingsRepo.getSetting("chatStyleDoNotUse"),
+    getSetting("chatStyleDoNotUse"),
-    settingsRepo.getSetting("chatStyleLanguageMode"),
+    getSetting("chatStyleLanguageMode"),
-    settingsRepo.getSetting("chatStyleManualLanguage"),
+    getSetting("chatStyleManualLanguage"),
   ]);
 
   return {

shared/src/settings-registry.test.ts

@@ -1,5 +1,8 @@
 import { describe, expect, it } from "vitest";
-import { settingsRegistry } from "./settings-registry";
+import {
+  getDefaultModelForProvider,
+  settingsRegistry,
+} from "./settings-registry";
 
 describe("settingsRegistry helpers", () => {
   describe("string parsing (parseNonEmptyStringOrNull)", () => {
@@ -192,5 +195,15 @@ describe("settingsRegistry helpers", () => {
         "openai_compatible",
       );
     });
+
+    it("uses provider-specific default models", () => {
+      expect(getDefaultModelForProvider("openai")).toBe("gpt-5.4-mini");
+      expect(getDefaultModelForProvider("gemini")).toBe(
+        "google/gemini-3-flash-preview",
+      );
+      expect(getDefaultModelForProvider("openrouter")).toBe(
+        "google/gemini-3-flash-preview",
+      );
+    });
   });
 });

shared/src/settings-registry.ts

@@ -38,6 +38,30 @@ function normalizeLlmProviderOrNull(raw: string | undefined): string | null {
   return normalized ? normalized : null;
 }
 
+export const DEFAULT_GEMINI_MODEL = "google/gemini-3-flash-preview";
+export const DEFAULT_OPENAI_MODEL = "gpt-5.4-mini";
+
+export function getDefaultModelForProvider(
+  provider: string | null | undefined,
+  fallbackModel?: string | null,
+): string {
+  const trimmedFallback = fallbackModel?.trim();
+  if (trimmedFallback) {
+    return trimmedFallback;
+  }
+
+  const normalizedProvider = normalizeLlmProviderOrNull(provider ?? undefined);
+
+  if (normalizedProvider === "openai") {
+    return DEFAULT_OPENAI_MODEL;
+  }
+
+  if (normalizedProvider === "gemini") {
+    return DEFAULT_GEMINI_MODEL;
+  }
+  return DEFAULT_GEMINI_MODEL;
+}
+
 function serializeNullableNumber(
   value: number | null | undefined,
 ): string | null {
@@ -87,8 +111,11 @@ export const settingsRegistry = {
     schema: z.string().trim().max(200),
     default: (): string =>
       typeof process !== "undefined"
-        ? process.env.MODEL || "google/gemini-3-flash-preview"
+        ? getDefaultModelForProvider(
-        : "google/gemini-3-flash-preview",
+            process.env.LLM_PROVIDER,
+            process.env.MODEL,
+          )
+        : DEFAULT_GEMINI_MODEL,
     parse: parseNonEmptyStringOrNull,
     serialize: (value: string | null | undefined): string | null =>
       value ?? null,