feat(chat): per-model vision flag for custom/BYOK image attachments (#3282) (#3420)

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

app/src/components/chat/ChatComposer.tsx

@@ -81,7 +81,12 @@ export default function ChatComposer({
         <input
           ref={fileInputRef}
           type="file"
-          accept={allowedMimeTypes.join(',')}
+          // No `accept` filter: Chromium 146 / CEF on macOS greys out valid files
+          // at the native open panel regardless of the filter shape (MIME, mixed,
+          // or extension-only). Selection is gated in `onAttachFiles` →
+          // `validateAndReadFile` instead (rejects unsupported types + images on
+          // non-vision models). `allowedMimeTypes` is kept for that JS validation.
+          accept={allowedMimeTypes.length ? allowedMimeTypes.join(',') : undefined}
           multiple
           className="hidden"
           onChange={e => {

app/src/components/settings/panels/AIPanel.tsx

@@ -26,11 +26,14 @@ import {
   loadLocalProviderSnapshot,
   type LocalProviderSnapshot,
   type ModelInfo,
+  type ModelRegistryEntry,
+  modelRegistryVision,
   OPENAI_CODEX_OAUTH_MISSING_AUTH_URL,
   OPENAI_CODEX_OAUTH_MISSING_CALLBACK_URL,
   saveAISettings,
   setCloudProviderKey,
   testProviderModel,
+  upsertModelRegistryVision,
 } from '../../../services/api/aiSettingsApi';
 import {
   creditsApi,
@@ -144,12 +147,17 @@ const BUILTIN_PROVIDER_META: Record<string, { tone: string; label: string }> = {
 };
 
 const WORKLOADS: Workload[] = [
-  { id: 'chat', group: 'chat', label: 'Chat', description: 'Direct conversational back-and-forth' },
+  {
+    id: 'chat',
+    group: 'chat',
+    label: 'Chat',
+    description: 'Direct conversational back-and-forth — “Quick” mode in Conversations',
+  },
   {
     id: 'reasoning',
     group: 'chat',
     label: 'Reasoning',
-    description: 'Main chat agent, meeting summarizer',
+    description: 'Main chat agent, meeting summarizer — “Reasoning” mode in Conversations',
   },
   {
     id: 'agentic',
@@ -217,7 +225,11 @@ const WORKLOAD_MODEL_HINTS: Record<WorkloadId, string> = {
 // just derives the `maskedKey` display string from `has_api_key`.
 // ─────────────────────────────────────────────────────────────────────────────
 
-type AISettings = { cloudProviders: CloudProvider[]; routing: RoutingMap };
+type AISettings = {
+  cloudProviders: CloudProvider[];
+  routing: RoutingMap;
+  modelRegistry: ModelRegistryEntry[];
+};
 
 const EMPTY_ROUTING: RoutingMap = {
   chat: { kind: 'default' },
@@ -230,7 +242,11 @@ const EMPTY_ROUTING: RoutingMap = {
   subconscious: { kind: 'default' },
 };
 
-const EMPTY_SETTINGS: AISettings = { cloudProviders: [], routing: EMPTY_ROUTING };
+const EMPTY_SETTINGS: AISettings = {
+  cloudProviders: [],
+  routing: EMPTY_ROUTING,
+  modelRegistry: [],
+};
 
 function maskKeyLabel(hasKey: boolean): string {
   return hasKey ? '•••• configured' : 'Not configured';
@@ -280,7 +296,7 @@ function toPanelRoutingFromApi(api: ApiAISettings): { panel: AISettings } {
     learning: liftRef(api.routing.learning),
     subconscious: liftRef(api.routing.subconscious),
   };
-  return { panel: { cloudProviders, routing } };
+  return { panel: { cloudProviders, routing, modelRegistry: api.modelRegistry } };
 }
 
 function toApiSettings(panel: AISettings): ApiAISettings {
@@ -303,6 +319,7 @@ function toApiSettings(panel: AISettings): ApiAISettings {
       learning: panel.routing.learning,
       subconscious: panel.routing.subconscious,
     },
+    modelRegistry: panel.modelRegistry,
   };
 }
 
@@ -1725,8 +1742,11 @@ interface CustomRoutingDialogProps {
   cloudProviders: CloudProvider[];
   localModels: OllamaModel[];
   ollamaRunning: boolean;
+  /** Current per-model vision registry, used to prefill the vision checkbox. */
+  modelRegistry: ModelRegistryEntry[];
   onClose: () => void;
-  onSubmit: (next: ProviderRef) => void;
+  /** Emits the chosen provider ref plus the user's vision flag for that model. */
+  onSubmit: (next: ProviderRef, vision: boolean) => void;
 }
 
 type CustomDialogSource =
@@ -1811,6 +1831,7 @@ const CustomRoutingDialog = ({
   cloudProviders,
   localModels,
   ollamaRunning,
+  modelRegistry,
   onClose,
   onSubmit,
 }: CustomRoutingDialogProps) => {
@@ -1862,6 +1883,35 @@ const CustomRoutingDialog = ({
       : null
   );
 
+  // Registry slug for the selected source — keys the per-model vision flag.
+  // Cloud uses the provider slug; local → `ollama`; claude-code → `claude-code`.
+  const registrySlug =
+    source?.kind === 'cloud'
+      ? source.providerSlug
+      : source?.kind === 'local'
+        ? 'ollama'
+        : source?.kind === 'claude-code'
+          ? 'claude-code'
+          : null;
+
+  // User-set vision flag for this (provider, model). Prefilled from the registry,
+  // re-prefilled whenever the selected provider/model changes.
+  const [vision, setVision] = useState<boolean>(() =>
+    registrySlug && model.trim()
+      ? modelRegistryVision(modelRegistry, registrySlug, model.trim())
+      : false
+  );
+  useEffect(() => {
+    // eslint-disable-next-line react-hooks/set-state-in-effect
+    setVision(
+      registrySlug && model.trim()
+        ? modelRegistryVision(modelRegistry, registrySlug, model.trim())
+        : false
+    );
+    // modelRegistry is stable for the dialog's lifetime (prop doesn't change mid-open).
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+  }, [registrySlug, model]);
+
   const selectedCloud =
     source?.kind === 'cloud' ? customCloud.find(c => c.slug === source.providerSlug) : undefined;
 
@@ -1935,16 +1985,19 @@ const CustomRoutingDialog = ({
     if (!source || !canSave) return;
     const temp = temperature == null || !Number.isFinite(temperature) ? null : temperature;
     if (source.kind === 'cloud') {
-      onSubmit({
-        kind: 'cloud',
-        providerSlug: source.providerSlug,
-        model: model.trim(),
-        temperature: temp,
-      });
+      onSubmit(
+        {
+          kind: 'cloud',
+          providerSlug: source.providerSlug,
+          model: model.trim(),
+          temperature: temp,
+        },
+        vision
+      );
     } else if (source.kind === 'claude-code') {
-      onSubmit({ kind: 'claude-code', model: model.trim(), temperature: temp });
+      onSubmit({ kind: 'claude-code', model: model.trim(), temperature: temp }, vision);
     } else {
-      onSubmit({ kind: 'local', model: model.trim(), temperature: temp });
+      onSubmit({ kind: 'local', model: model.trim(), temperature: temp }, vision);
     }
   };
 
@@ -2220,6 +2273,26 @@ const CustomRoutingDialog = ({
               </p>
             </div>
 
+            {/* Vision capability (optional). Marks a custom/BYOK model as
+                accepting image input so the chat composer offers image
+                attachments for it. Only shown once a concrete model is chosen. */}
+            {registrySlug && model.trim().length > 0 && (
+              <div className="flex flex-col gap-1.5">
+                <label className="inline-flex items-center gap-2 text-xs font-medium text-stone-700 dark:text-neutral-200">
+                  <input
+                    type="checkbox"
+                    checked={vision}
+                    onChange={e => setVision(e.target.checked)}
+                    className="h-3.5 w-3.5 rounded border-stone-300 dark:border-neutral-700 text-primary-500 focus:ring-primary-500"
+                  />
+                  {t('settings.ai.modelVision')}
+                </label>
+                <p className="text-[11px] text-stone-400 dark:text-neutral-500">
+                  {t('settings.ai.modelVisionDesc')}
+                </p>
+              </div>
+            )}
+
             {(testBusy || testReply || testError || testStartedAt) && (
               <div
                 role={testError ? 'alert' : 'status'}
@@ -2357,6 +2430,7 @@ const GlobalOwnModelSelector = ({
   cloudProviders,
   localModels,
   ollamaRunning,
+  modelRegistry,
   onApply,
 }: {
   current: ProviderRef | null;
@@ -2364,7 +2438,8 @@ const GlobalOwnModelSelector = ({
   cloudProviders: CloudProvider[];
   localModels: OllamaModel[];
   ollamaRunning: boolean;
-  onApply: (next: ProviderRef) => Promise<void>;
+  modelRegistry: ModelRegistryEntry[];
+  onApply: (next: ProviderRef, vision: boolean) => Promise<void>;
 }) => {
   const { t } = useT();
   const customCloud = cloudProviders.filter(p => p.slug !== 'openhuman');
@@ -2385,6 +2460,29 @@ const GlobalOwnModelSelector = ({
   const [model, setModel] = useState<string>(
     current?.kind === 'cloud' || current?.kind === 'local' ? current.model : ''
   );
+  // Registry slug for the selected source — keys the per-model vision flag.
+  const registrySlug =
+    source?.kind === 'cloud'
+      ? source.providerSlug
+      : source?.kind === 'local'
+        ? 'ollama'
+        : source?.kind === 'claude-code'
+          ? 'claude-code'
+          : null;
+  const [vision, setVision] = useState<boolean>(() =>
+    registrySlug && model.trim()
+      ? modelRegistryVision(modelRegistry, registrySlug, model.trim())
+      : false
+  );
+  useEffect(() => {
+    // eslint-disable-next-line react-hooks/set-state-in-effect
+    setVision(
+      registrySlug && model.trim()
+        ? modelRegistryVision(modelRegistry, registrySlug, model.trim())
+        : false
+    );
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+  }, [registrySlug, model]);
   const [cloudModels, setCloudModels] = useState<ModelInfo[]>([]);
   const [cloudModelsLoading, setCloudModelsLoading] = useState(false);
   const [cloudModelsError, setCloudModelsError] = useState<string | null>(null);
@@ -2453,15 +2551,14 @@ const GlobalOwnModelSelector = ({
     setSaving(true);
     try {
       if (nextSource.kind === 'local') {
-        await onApply({ kind: 'local', model: nextModel.trim() });
+        await onApply({ kind: 'local', model: nextModel.trim() }, vision);
       } else if (nextSource.kind === 'claude-code') {
-        await onApply({ kind: 'claude-code', model: nextModel.trim() });
+        await onApply({ kind: 'claude-code', model: nextModel.trim() }, vision);
       } else {
-        await onApply({
-          kind: 'cloud',
-          providerSlug: nextSource.providerSlug,
-          model: nextModel.trim(),
-        });
+        await onApply(
+          { kind: 'cloud', providerSlug: nextSource.providerSlug, model: nextModel.trim() },
+          vision
+        );
       }
     } finally {
       setSaving(false);
@@ -2566,6 +2663,23 @@ const GlobalOwnModelSelector = ({
               ) : null}
             </div>
           </div>
+          {registrySlug && model.trim().length > 0 && (
+            <label className="flex items-start gap-2 text-xs font-medium text-stone-700 dark:text-neutral-200">
+              <input
+                type="checkbox"
+                checked={vision}
+                onChange={e => setVision(e.target.checked)}
+                className="mt-0.5 h-3.5 w-3.5 rounded border-stone-300 dark:border-neutral-700 text-primary-500 focus:ring-primary-500"
+              />
+              <span>
+                {t('settings.ai.modelVision')}
+                <span className="block font-normal text-[11px] text-stone-400 dark:text-neutral-500">
+                  {t('settings.ai.modelVisionDesc')}
+                </span>
+              </span>
+            </label>
+          )}
+
           <div className="rounded-lg bg-stone-50 dark:bg-neutral-800/60 px-3 py-2 text-xs text-stone-500 dark:text-neutral-400">
             {t('settings.ai.globalModel.appliesToAll')}
           </div>
@@ -3103,8 +3217,23 @@ const AIPanel = ({ embedded = false }: AIPanelProps = {}) => {
                 cloudProviders={draft.cloudProviders}
                 localModels={installed}
                 ollamaRunning={ollama.state === 'running'}
-                onApply={async next => {
-                  await persist({ ...draft, routing: routingWithAllWorkloads(next) });
+                modelRegistry={draft.modelRegistry}
+                onApply={async (next, vision) => {
+                  const reg =
+                    next.kind === 'cloud'
+                      ? { slug: next.providerSlug, model: next.model }
+                      : next.kind === 'local'
+                        ? { slug: 'ollama', model: next.model }
+                        : next.kind === 'claude-code'
+                          ? { slug: 'claude-code', model: next.model }
+                          : null;
+                  await persist({
+                    ...draft,
+                    routing: routingWithAllWorkloads(next),
+                    modelRegistry: reg
+                      ? upsertModelRegistryVision(draft.modelRegistry, reg.slug, reg.model, vision)
+                      : draft.modelRegistry,
+                  });
                 }}
               />
             ) : null}
@@ -3290,11 +3419,24 @@ const AIPanel = ({ embedded = false }: AIPanelProps = {}) => {
               cloudProviders={draft.cloudProviders}
               localModels={installed}
               ollamaRunning={ollama.state === 'running'}
+              modelRegistry={draft.modelRegistry}
               onClose={() => setCustomDialogFor(null)}
-              onSubmit={async next => {
+              onSubmit={async (next, vision) => {
+                // (provider slug, model id) the vision flag keys on.
+                const reg =
+                  next.kind === 'cloud'
+                    ? { slug: next.providerSlug, model: next.model }
+                    : next.kind === 'local'
+                      ? { slug: 'ollama', model: next.model }
+                      : next.kind === 'claude-code'
+                        ? { slug: 'claude-code', model: next.model }
+                        : null;
                 const nextDraft = {
                   ...draft,
                   routing: { ...draft.routing, [customDialogFor]: next },
+                  modelRegistry: reg
+                    ? upsertModelRegistryVision(draft.modelRegistry, reg.slug, reg.model, vision)
+                    : draft.modelRegistry,
                 };
                 await persist(nextDraft);
                 setCustomDialogFor(null);

app/src/components/settings/panels/__tests__/AIPanel.test.tsx

@@ -15,6 +15,7 @@ import {
   setCloudProviderKey,
   startOpenAiCodexOAuth,
   testProviderModel,
+  upsertModelRegistryVision,
 } from '../../../../services/api/aiSettingsApi';
 import { creditsApi } from '../../../../services/api/creditsApi';
 import { renderWithProviders } from '../../../../test/test-utils';
@@ -45,6 +46,8 @@ vi.mock('../../../../services/api/aiSettingsApi', () => ({
   saveAISettings: vi.fn(),
   loadLocalProviderSnapshot: vi.fn(),
   testProviderModel: vi.fn(),
+  modelRegistryVision: vi.fn(() => false),
+  upsertModelRegistryVision: vi.fn((registry: unknown[]) => registry),
   setCloudProviderKey: vi.fn().mockResolvedValue(undefined),
   clearCloudProviderKey: vi.fn().mockResolvedValue(undefined),
   serializeProviderRef: vi.fn((r: { kind: string; providerSlug?: string; model?: string }) =>
@@ -123,6 +126,7 @@ const baseSettings = {
     learning: { kind: 'openhuman' as const },
     subconscious: { kind: 'openhuman' as const },
   },
+  modelRegistry: [],
 };
 
 const baseLocalSnapshot = { status: null, diagnostics: null, presets: null, installedModels: [] };
@@ -304,6 +308,52 @@ describe('AIPanel', () => {
     }
   });
 
+  // ─── per-model vision flag (BYOK) ───────────────────────────────────────────
+
+  it('flags a custom BYOK model as vision-capable via the Own-model selector', async () => {
+    vi.mocked(loadAISettings).mockResolvedValue({
+      ...baseSettings,
+      cloudProviders: [
+        ...baseSettings.cloudProviders,
+        {
+          id: 'p_custom_openai',
+          slug: 'openai',
+          label: 'OpenAI',
+          endpoint: 'https://api.openai.com/v1',
+          auth_style: 'bearer' as const,
+          has_api_key: true,
+        },
+      ],
+    });
+    renderWithProviders(<AIPanel />);
+    await waitFor(() =>
+      expect(screen.getByRole('button', { name: /Use Your Own Models/i })).toBeInTheDocument()
+    );
+    fireEvent.click(screen.getByRole('button', { name: /Use Your Own Models/i }));
+
+    // Enter a model id → the per-model "Supports vision" checkbox appears.
+    const modelInput = await screen.findByPlaceholderText('Enter model id');
+    fireEvent.change(modelInput, { target: { value: 'gpt-4o' } });
+
+    const visionCheckbox = await screen.findByRole('checkbox', { name: /Supports vision/i });
+    expect(visionCheckbox).not.toBeChecked();
+    fireEvent.click(visionCheckbox);
+    expect(visionCheckbox).toBeChecked();
+
+    fireEvent.click(screen.getByRole('button', { name: /^Save$/ }));
+
+    // The vision flag is threaded through to the registry upsert + persisted.
+    await waitFor(() =>
+      expect(vi.mocked(upsertModelRegistryVision)).toHaveBeenCalledWith(
+        expect.anything(),
+        'openai',
+        'gpt-4o',
+        true
+      )
+    );
+    expect(saveAISettings).toHaveBeenCalled();
+  });
+
   // ─── auth_style preservation ────────────────────────────────────────────────
 
   it('preserves auth_style: "anthropic" through save when Anthropic provider is configured', async () => {
@@ -333,6 +383,7 @@ describe('AIPanel', () => {
         learning: { kind: 'openhuman' as const },
         subconscious: { kind: 'openhuman' as const },
       },
+      modelRegistry: [],
     };
 
     vi.mocked(loadAISettings).mockResolvedValue(settingsWithAnthropic);
@@ -557,6 +608,7 @@ describe('AIPanel', () => {
         learning: { kind: 'openhuman' as const },
         subconscious: { kind: 'openhuman' as const },
       },
+      modelRegistry: [],
     };
     vi.mocked(loadAISettings).mockResolvedValue(settingsWithOpenAI);
     vi.mocked(saveAISettings).mockResolvedValue(undefined);
@@ -612,6 +664,7 @@ describe('AIPanel', () => {
         learning: { kind: 'openhuman' as const },
         subconscious: { kind: 'openhuman' as const },
       },
+      modelRegistry: [],
     };
     vi.mocked(loadAISettings).mockResolvedValue(settingsWithOllama);
     vi.mocked(saveAISettings).mockResolvedValue(undefined);
@@ -986,6 +1039,7 @@ describe('AIPanel', () => {
           has_api_key: false,
         },
       ],
+      modelRegistry: [],
     };
     vi.mocked(loadAISettings).mockResolvedValue(settingsWithOllama);
     renderWithProviders(<AIPanel />);
@@ -1008,6 +1062,7 @@ describe('AIPanel', () => {
           has_api_key: false,
         },
       ],
+      modelRegistry: [],
     };
     vi.mocked(loadAISettings).mockResolvedValue(settingsWithOllama);
     renderWithProviders(<AIPanel />);
@@ -1047,6 +1102,7 @@ describe('AIPanel', () => {
         ...baseSettings.routing,
         reasoning: { kind: 'cloud' as const, providerSlug: 'openai', model: 'gpt-4o' },
       },
+      modelRegistry: [],
     };
     vi.mocked(loadAISettings).mockResolvedValue(settingsWithOpenAI);
     vi.mocked(saveAISettings).mockResolvedValue(undefined);
@@ -1103,6 +1159,7 @@ describe('AIPanel', () => {
         ...baseSettings.routing,
         reasoning: { kind: 'cloud' as const, providerSlug: 'openai', model: 'gpt-4o' },
       },
+      modelRegistry: [],
     };
     vi.mocked(loadAISettings).mockResolvedValue(settingsWithOpenAI);
     vi.mocked(listProviderModels).mockResolvedValue([{ id: 'gpt-4o' }, { id: 'gpt-4o-mini' }]);
@@ -1146,6 +1203,7 @@ describe('AIPanel', () => {
         ...baseSettings.routing,
         reasoning: { kind: 'cloud' as const, providerSlug: 'openai', model: 'gpt-4o' },
       },
+      modelRegistry: [],
     };
     vi.mocked(loadAISettings).mockResolvedValue(settingsWithOpenAI);
     vi.mocked(listProviderModels).mockResolvedValue([{ id: 'gpt-4o' }]);
@@ -1190,6 +1248,7 @@ describe('AIPanel', () => {
         ...baseSettings.routing,
         reasoning: { kind: 'cloud' as const, providerSlug: 'openai', model: 'gpt-4o' },
       },
+      modelRegistry: [],
     };
     vi.mocked(loadAISettings).mockResolvedValue(settingsWithOpenAI);
     vi.mocked(listProviderModels).mockResolvedValue([{ id: 'gpt-4o' }]);

app/src/lib/attachments.test.ts

@@ -85,6 +85,40 @@ describe('validateAndReadFile', () => {
     }
   });
 
+  it('rejects non-extractable documents (docx/pptx/xlsx) the agent cannot read', async () => {
+    for (const mime of [
+      'application/vnd.openxmlformats-officedocument.wordprocessingml.document',
+      'application/vnd.openxmlformats-officedocument.presentationml.presentation',
+      'application/vnd.openxmlformats-officedocument.spreadsheetml.sheet',
+      'application/zip',
+    ]) {
+      const result = await validateAndReadFile(makeFile('f', mime, 8), 0);
+      expect('error' in result && result.error.code).toBe('unsupported_type');
+    }
+  });
+
+  it('accepts the text-extractable document formats (pdf/txt/markdown)', async () => {
+    for (const mime of ['application/pdf', 'text/plain', 'text/markdown']) {
+      const result = await validateAndReadFile(makeFile('f', mime, 8), 0);
+      expect('attachment' in result && result.attachment.kind === 'file').toBe(true);
+    }
+  });
+
+  it('rejects an image when allowImages is false (non-vision model)', async () => {
+    const result = await validateAndReadFile(makeFile('p.png', 'image/png', 8), 0, 0, false);
+    expect('error' in result && result.error.code).toBe('image_not_supported');
+  });
+
+  it('still accepts a document when allowImages is false', async () => {
+    const result = await validateAndReadFile(makeFile('d.pdf', 'application/pdf', 8), 0, 0, false);
+    expect('attachment' in result && result.attachment.kind === 'file').toBe(true);
+  });
+
+  it('accepts an image when allowImages is true (vision model)', async () => {
+    const result = await validateAndReadFile(makeFile('p.png', 'image/png', 8), 0, 0, true);
+    expect('attachment' in result && result.attachment.kind === 'image').toBe(true);
+  });
+
   it('rejects files that exceed the file size limit', async () => {
     const oversizedFile = makeFile(
       'big.pdf',

app/src/lib/attachments.ts

@@ -41,6 +41,17 @@ export const ALLOWED_ATTACHMENT_MIME_TYPES = [
   ...ALLOWED_FILE_MIME_TYPES,
 ] as const;
 
+// Document formats the backend actually text-extracts (PDF via pdf_extract;
+// TXT/Markdown via UTF-8). DOCX/PPTX/XLSX/ZIP are intentionally excluded — the
+// agent would only see a reference stub, not their content. `text/csv` is also
+// deliberately left out: the backend *can* extract it, but the chat composer is
+// scoped to PDF/TXT/Markdown by product decision (revisit here if CSV is wanted).
+// Used by the ingest validator below, not by a native `accept` filter:
+// Chromium/CEF on macOS greys valid files at the open panel regardless of the
+// filter shape, so selection is gated in `validateAndReadFile` after the user
+// picks, not at the dialog.
+const EXTRACTABLE_FILE_MIME_TYPES = ['application/pdf', 'text/plain', 'text/markdown'] as const;
+
 export const ATTACHMENT_MAX_IMAGES = 4;
 export const ATTACHMENT_MAX_FILES = 4;
 export const ATTACHMENT_MAX_IMAGE_SIZE_BYTES = 8 * 1024 * 1024; // 8 MB
@@ -63,6 +74,7 @@ export type AttachmentError =
   | { code: 'unsupported_type'; mimeType: string }
   | { code: 'too_large'; sizeBytes: number; maxBytes: number }
   | { code: 'too_many'; kind: AttachmentKind; max: number }
+  | { code: 'image_not_supported' }
   | { code: 'read_failed'; reason: string };
 
 export function isAllowedMimeType(mime: string): mime is AllowedImageMimeType {
@@ -73,6 +85,29 @@ export function isAllowedAttachmentMimeType(mime: string): mime is AllowedAttach
   return (ALLOWED_ATTACHMENT_MIME_TYPES as readonly string[]).includes(mime);
 }
 
+/**
+ * The exact MIME set the ingest validator accepts — images plus the
+ * text-extractable documents. A strict subset of {@link AllowedAttachmentMimeType}
+ * (which also lists reference-only types like CSV/DOCX/ZIP that we reject).
+ */
+export type SupportedAttachmentMimeType =
+  | AllowedImageMimeType
+  | (typeof EXTRACTABLE_FILE_MIME_TYPES)[number];
+
+/**
+ * Stricter gate than {@link isAllowedAttachmentMimeType}: only the formats the
+ * backend actually reads — images, plus the text-extractable documents (PDF via
+ * pdf_extract; TXT/Markdown via UTF-8). DOCX/PPTX/XLSX/ZIP are excluded so they
+ * can't be attached as content-less reference stubs. Applied on every ingest
+ * path (picker, drag-drop, paste).
+ */
+export function isSupportedAttachmentMimeType(mime: string): mime is SupportedAttachmentMimeType {
+  return (
+    (ALLOWED_IMAGE_MIME_TYPES as readonly string[]).includes(mime) ||
+    (EXTRACTABLE_FILE_MIME_TYPES as readonly string[]).includes(mime)
+  );
+}
+
 export function attachmentKindForMime(mime: AllowedAttachmentMimeType): AttachmentKind {
   return isAllowedMimeType(mime) ? 'image' : 'file';
 }
@@ -148,13 +183,20 @@ async function buildAttachmentDataUri(
 export async function validateAndReadFile(
   file: File,
   existingCount: number,
-  existingFileCount = 0
+  existingFileCount = 0,
+  // When `false` (the active chat model isn't vision-capable), image files are
+  // rejected; documents (PDF/Word/etc.) still flow. Defaults `true` so non-chat
+  // callers are unaffected.
+  allowImages = true
 ): Promise<{ attachment: Attachment } | { error: AttachmentError }> {
-  if (!isAllowedAttachmentMimeType(file.type)) {
+  if (!isSupportedAttachmentMimeType(file.type)) {
     return { error: { code: 'unsupported_type', mimeType: file.type || 'unknown' } };
   }
 
   const kind = attachmentKindForMime(file.type);
+  if (!allowImages && kind === 'image') {
+    return { error: { code: 'image_not_supported' } };
+  }
   const maxCount = kind === 'image' ? ATTACHMENT_MAX_IMAGES : ATTACHMENT_MAX_FILES;
   const count = kind === 'image' ? existingCount : existingFileCount;
   if (count >= maxCount) {

app/src/lib/i18n/ar.ts

@@ -1822,7 +1822,10 @@ const messages: TranslationMap = {
   'chat.attachment.tooMany': 'الحد الأقصى {max} صور لكل رسالة',
   'chat.attachment.tooManyFiles': 'الحد الأقصى {max} ملفات لكل رسالة',
   'chat.attachment.tooLarge': 'حجم الصورة يتجاوز الحد المسموح {max}',
-  'chat.attachment.unsupportedType': 'نوع ملف غير مدعوم. استخدم PNG أو JPEG أو WebP أو GIF أو BMP.',
+  'chat.attachment.unsupportedType':
+    'نوع ملف غير مدعوم. استخدم صورة (PNG أو JPEG أو WebP أو GIF أو BMP) أو ملف PDF أو TXT أو Markdown.',
+  'chat.attachment.imageNotSupported':
+    'لا يمكن لهذا النموذج قراءة الصور. أرفق ملف PDF أو TXT أو Markdown بدلاً من ذلك.',
   'chat.attachment.readFailed': 'تعذر قراءة الملف',
   'memory.searchAria': 'البحث في الذاكرة',
   'memory.searchPlaceholder': 'البحث في إدخالات الذاكرة...',
@@ -3046,6 +3049,9 @@ const messages: TranslationMap = {
   'settings.ai.temperatureOverrideValue': 'تجاوز درجة الحرارة (القيمة)',
   'settings.ai.temperatureOverrideDesc':
     'أقل = أكثر تحديدا. إتركْ غير مُمْكَنَّلَ لإسْتِعْمال قَبْل المُقدّمِ.',
+  'settings.ai.modelVision': 'يدعم الرؤية (إدخال الصور)',
+  'settings.ai.modelVisionDesc':
+    'فعّل هذا إذا كان النموذج يقبل الصور. يتيح لمربع الدردشة إرفاق الصور عند اختيار هذا النموذج.',
   'settings.ai.testFailed': 'فشل الاختبار',
   'settings.ai.testingModel': 'نموذج الاختبار...',
   'settings.ai.modelResponse': 'استجابة النموذج',

app/src/lib/i18n/bn.ts

@@ -1865,7 +1865,9 @@ const messages: TranslationMap = {
   'chat.attachment.tooManyFiles': 'প্রতি বার্তায় সর্বোচ্চ {max}টি ফাইল',
   'chat.attachment.tooLarge': 'ছবি {max} আকারের সীমা অতিক্রম করেছে',
   'chat.attachment.unsupportedType':
-    'অসমর্থিত ফাইল প্রকার। PNG, JPEG, WebP, GIF, বা BMP ব্যবহার করুন।',
+    'অসমর্থিত ফাইল প্রকার। একটি ছবি (PNG, JPEG, WebP, GIF, BMP) অথবা একটি PDF, TXT, বা Markdown ফাইল ব্যবহার করুন।',
+  'chat.attachment.imageNotSupported':
+    'এই মডেলটি ছবি পড়তে পারে না। পরিবর্তে একটি PDF, TXT, বা Markdown ফাইল সংযুক্ত করুন।',
   'chat.attachment.readFailed': 'ফাইল পড়া যায়নি',
   'memory.searchAria': 'মেমোরি খুঁজুন',
   'memory.searchPlaceholder': 'মেমোরি এন্ট্রি খুঁজুন...',
@@ -3112,6 +3114,9 @@ const messages: TranslationMap = {
   'settings.ai.temperatureOverrideValue': 'তাপমাত্রা ওভাররাইড (মান)',
   'settings.ai.temperatureOverrideDesc':
     'নিচে = আরও বুদ্ধিমান. পরিসেবা উপলব্ধকারীর ডিফল্ট ব্যবহারের জন্য সীমা ধার্য না করা হলে, বন্ধ করুন ।',
+  'settings.ai.modelVision': 'ভিশন সমর্থন করে (ছবি ইনপুট)',
+  'settings.ai.modelVisionDesc':
+    'এই মডেল ছবি গ্রহণ করলে সক্ষম করুন। এই মডেল নির্বাচিত থাকলে চ্যাটে ছবি সংযুক্ত করতে দেয়।',
   'settings.ai.testFailed': 'পরীক্ষা ব্যর্থ হয়েছে',
   'settings.ai.testingModel': 'পরীক্ষার মডেল...',
   'settings.ai.modelResponse': 'মডেল প্রতিক্রিয়া',

app/src/lib/i18n/de.ts

@@ -1912,7 +1912,9 @@ const messages: TranslationMap = {
   'chat.attachment.tooManyFiles': 'Maximal {max} Dateien pro Nachricht',
   'chat.attachment.tooLarge': 'Bild überschreitet die Größenbeschränkung von {max}',
   'chat.attachment.unsupportedType':
-    'Nicht unterstützter Dateityp. Verwenden Sie PNG, JPEG, WebP, GIF oder BMP.',
+    'Nicht unterstützter Dateityp. Verwenden Sie ein Bild (PNG, JPEG, WebP, GIF, BMP) oder eine PDF-, TXT- oder Markdown-Datei.',
+  'chat.attachment.imageNotSupported':
+    'Dieses Modell kann keine Bilder lesen. Hängen Sie stattdessen eine PDF-, TXT- oder Markdown-Datei an.',
   'chat.attachment.readFailed': 'Datei konnte nicht gelesen werden',
   'memory.searchAria': 'Speicher durchsuchen',
   'memory.searchPlaceholder': 'Speichereinträge durchsuchen...',
@@ -3191,6 +3193,9 @@ const messages: TranslationMap = {
   'settings.ai.temperatureOverrideValue': 'Temperatur-Override (Wert)',
   'settings.ai.temperatureOverrideDesc':
     'Niedriger = deterministischer. Deaktivieren Sie diese Option, um den Anbieterstandard zu verwenden.',
+  'settings.ai.modelVision': 'Unterstützt Bildeingabe (Vision)',
+  'settings.ai.modelVisionDesc':
+    'Aktivieren, wenn dieses Modell Bilder akzeptiert. Ermöglicht das Anhängen von Bildern im Chat, wenn dieses Modell ausgewählt ist.',
   'settings.ai.testFailed': 'Test fehlgeschlagen.',
   'settings.ai.testingModel': 'Modell wird getestet...',
   'settings.ai.modelResponse': 'Modellantwort',

app/src/lib/i18n/en.ts

@@ -2238,7 +2238,10 @@ const en: TranslationMap = {
   'chat.attachment.tooMany': 'Maximum {max} images per message',
   'chat.attachment.tooManyFiles': 'Maximum {max} files per message',
   'chat.attachment.tooLarge': 'Image exceeds {max} size limit',
-  'chat.attachment.unsupportedType': 'Unsupported file type. Use PNG, JPEG, WebP, GIF, or BMP.',
+  'chat.attachment.unsupportedType':
+    'Unsupported file type. Use an image (PNG, JPEG, WebP, GIF, BMP) or a PDF, TXT, or Markdown file.',
+  'chat.attachment.imageNotSupported':
+    'This model can’t read images. You can attach a pdf, txt, or md file.',
   'chat.attachment.readFailed': 'Could not read file',
 
   // Memory (additional)
@@ -3633,6 +3636,9 @@ const en: TranslationMap = {
   'settings.ai.temperatureOverrideValue': 'Temperature override (value)',
   'settings.ai.temperatureOverrideDesc':
     'Lower = more deterministic. Leave unchecked to use the provider default.',
+  'settings.ai.modelVision': 'Supports vision (image input)',
+  'settings.ai.modelVisionDesc':
+    'Enable if this model accepts images. Lets the chat composer attach images when this model is selected.',
   'settings.ai.testFailed': 'Test failed',
   'settings.ai.testingModel': 'Testing model...',
   'settings.ai.modelResponse': 'Model response',

app/src/lib/i18n/es.ts

@@ -1904,7 +1904,9 @@ const messages: TranslationMap = {
   'chat.attachment.tooManyFiles': 'Máximo {max} archivos por mensaje',
   'chat.attachment.tooLarge': 'La imagen supera el límite de tamaño de {max}',
   'chat.attachment.unsupportedType':
-    'Tipo de archivo no compatible. Use PNG, JPEG, WebP, GIF o BMP.',
+    'Tipo de archivo no compatible. Use una imagen (PNG, JPEG, WebP, GIF, BMP) o un archivo PDF, TXT o Markdown.',
+  'chat.attachment.imageNotSupported':
+    'Este modelo no puede leer imágenes. Adjunta un archivo PDF, TXT o Markdown en su lugar.',
   'chat.attachment.readFailed': 'No se pudo leer el archivo',
   'memory.searchAria': 'Buscar en memoria',
   'memory.searchPlaceholder': 'Buscar entradas de memoria...',
@@ -3173,6 +3175,9 @@ const messages: TranslationMap = {
   'settings.ai.temperatureOverrideValue': 'Anulación de temperatura (valor)',
   'settings.ai.temperatureOverrideDesc':
     'Más bajo = más determinista. Déjelo sin marcar para usar el valor predeterminado del proveedor.',
+  'settings.ai.modelVision': 'Admite visión (entrada de imágenes)',
+  'settings.ai.modelVisionDesc':
+    'Actívalo si este modelo acepta imágenes. Permite adjuntar imágenes en el chat cuando este modelo está seleccionado.',
   'settings.ai.testFailed': 'Prueba fallida',
   'settings.ai.testingModel': 'Modelo de prueba...',
   'settings.ai.modelResponse': 'Respuesta modelo',

app/src/lib/i18n/fr.ts

@@ -1912,7 +1912,9 @@ const messages: TranslationMap = {
   'chat.attachment.tooManyFiles': 'Maximum {max} fichiers par message',
   'chat.attachment.tooLarge': "L'image dépasse la taille limite de {max}",
   'chat.attachment.unsupportedType':
-    'Type de fichier non pris en charge. Utilisez PNG, JPEG, WebP, GIF ou BMP.',
+    'Type de fichier non pris en charge. Utilise une image (PNG, JPEG, WebP, GIF, BMP) ou un fichier PDF, TXT ou Markdown.',
+  'chat.attachment.imageNotSupported':
+    'Ce modèle ne peut pas lire les images. Joins plutôt un fichier PDF, TXT ou Markdown.',
   'chat.attachment.readFailed': 'Impossible de lire le fichier',
   'memory.searchAria': 'Rechercher dans la mémoire',
   'memory.searchPlaceholder': 'Rechercher des entrées de mémoire…',
@@ -3184,6 +3186,9 @@ const messages: TranslationMap = {
   'settings.ai.temperatureOverrideValue': 'Override de température (valeur)',
   'settings.ai.temperatureOverrideDesc':
     'Plus bas = plus déterministe. Laissez décoché pour utiliser la valeur par défaut du fournisseur.',
+  'settings.ai.modelVision': 'Prend en charge la vision (entrée d’images)',
+  'settings.ai.modelVisionDesc':
+    'Activez si ce modèle accepte les images. Permet de joindre des images dans le chat lorsque ce modèle est sélectionné.',
   'settings.ai.testFailed': 'Test échoué',
   'settings.ai.testingModel': 'Modèle de test...',
   'settings.ai.modelResponse': 'Réponse du modèle',

app/src/lib/i18n/hi.ts

@@ -1864,7 +1864,9 @@ const messages: TranslationMap = {
   'chat.attachment.tooManyFiles': 'प्रति संदेश अधिकतम {max} फ़ाइलें',
   'chat.attachment.tooLarge': 'छवि {max} आकार सीमा से अधिक है',
   'chat.attachment.unsupportedType':
-    'असमर्थित फ़ाइल प्रकार। PNG, JPEG, WebP, GIF, या BMP का उपयोग करें।',
+    'असमर्थित फ़ाइल प्रकार। कोई छवि (PNG, JPEG, WebP, GIF, BMP) या PDF, TXT, या Markdown फ़ाइल का उपयोग करें।',
+  'chat.attachment.imageNotSupported':
+    'यह मॉडल छवियाँ नहीं पढ़ सकता। इसके बजाय कोई PDF, TXT या Markdown फ़ाइल संलग्न करें।',
   'chat.attachment.readFailed': 'फ़ाइल पढ़ नहीं सकी',
   'memory.searchAria': 'मेमोरी सर्च करें',
   'memory.searchPlaceholder': 'मेमोरी एंट्रीज़ सर्च करें...',
@@ -3120,6 +3122,9 @@ const messages: TranslationMap = {
   'settings.ai.temperatureOverrideValue': 'तापमान ओवरराइड (मान)',
   'settings.ai.temperatureOverrideDesc':
     'लोअर = अधिक नियतिवादी। प्रदाता डिफ़ॉल्ट का उपयोग करने के लिए अनचेक छोड़ दें।',
+  'settings.ai.modelVision': 'विज़न समर्थित है (छवि इनपुट)',
+  'settings.ai.modelVisionDesc':
+    'यदि यह मॉडल छवियाँ स्वीकार करता है तो सक्षम करें। यह मॉडल चुने जाने पर चैट में छवियाँ संलग्न करने देता है।',
   'settings.ai.testFailed': 'परीक्षण विफल रहा',
   'settings.ai.testingModel': 'परीक्षण मॉडल...',
   'settings.ai.modelResponse': 'मॉडल प्रतिक्रिया',

app/src/lib/i18n/id.ts

@@ -1867,7 +1867,9 @@ const messages: TranslationMap = {
   'chat.attachment.tooManyFiles': 'Maksimal {max} file per pesan',
   'chat.attachment.tooLarge': 'Gambar melebihi batas ukuran {max}',
   'chat.attachment.unsupportedType':
-    'Jenis file tidak didukung. Gunakan PNG, JPEG, WebP, GIF, atau BMP.',
+    'Jenis file tidak didukung. Gunakan gambar (PNG, JPEG, WebP, GIF, BMP) atau file PDF, TXT, atau Markdown.',
+  'chat.attachment.imageNotSupported':
+    'Model ini tidak dapat membaca gambar. Lampirkan file PDF, TXT, atau Markdown sebagai gantinya.',
   'chat.attachment.readFailed': 'Tidak dapat membaca file',
   'memory.searchAria': 'Cari memori',
   'memory.searchPlaceholder': 'Cari entri memori...',
@@ -3124,6 +3126,9 @@ const messages: TranslationMap = {
   'settings.ai.temperatureOverrideValue': 'Penggantian suhu (nilai)',
   'settings.ai.temperatureOverrideDesc':
     'Turunkan = lebih deterministik. Biarkan tak diperiksa untuk memakai penyedia bawaan.',
+  'settings.ai.modelVision': 'Mendukung visi (input gambar)',
+  'settings.ai.modelVisionDesc':
+    'Aktifkan jika model ini menerima gambar. Memungkinkan penyusun obrolan melampirkan gambar saat model ini dipilih.',
   'settings.ai.testFailed': 'Pengujian gagal',
   'settings.ai.testingModel': 'Pengujian model...',
   'settings.ai.modelResponse': 'Respons model',

app/src/lib/i18n/it.ts

@@ -1895,7 +1895,10 @@ const messages: TranslationMap = {
   'chat.attachment.tooMany': 'Massimo {max} immagini per messaggio',
   'chat.attachment.tooManyFiles': 'Massimo {max} file per messaggio',
   'chat.attachment.tooLarge': "L'immagine supera il limite di dimensione di {max}",
-  'chat.attachment.unsupportedType': 'Tipo di file non supportato. Usa PNG, JPEG, WebP, GIF o BMP.',
+  'chat.attachment.unsupportedType':
+    'Tipo di file non supportato. Usa un file immagine (PNG, JPEG, WebP, GIF, BMP) oppure PDF, TXT o Markdown.',
+  'chat.attachment.imageNotSupported':
+    'Questo modello non può leggere le immagini. Allega invece un file PDF, TXT o Markdown.',
   'chat.attachment.readFailed': 'Impossibile leggere il file',
   'memory.searchAria': 'Cerca memoria',
   'memory.searchPlaceholder': 'Cerca voci di memoria...',
@@ -3166,6 +3169,9 @@ const messages: TranslationMap = {
   'settings.ai.temperatureOverrideValue': 'Ignora temperatura (valore)',
   'settings.ai.temperatureOverrideDesc':
     'Più basso = più deterministico. Lascia deselezionato per usare il valore predefinito del provider.',
+  'settings.ai.modelVision': 'Supporta la visione (input immagini)',
+  'settings.ai.modelVisionDesc':
+    'Attiva se questo modello accetta immagini. Consente di allegare immagini nella chat quando è selezionato questo modello.',
   'settings.ai.testFailed': 'Test non riuscito',
   'settings.ai.testingModel': 'Test del modello...',
   'settings.ai.modelResponse': 'Risposta del modello',

app/src/lib/i18n/ko.ts

@@ -1844,7 +1844,9 @@ const messages: TranslationMap = {
   'chat.attachment.tooManyFiles': '메시지당 최대 {max}개 파일',
   'chat.attachment.tooLarge': '이미지가 {max} 크기 제한을 초과합니다',
   'chat.attachment.unsupportedType':
-    '지원되지 않는 파일 형식입니다. PNG, JPEG, WebP, GIF 또는 BMP를 사용하세요.',
+    '지원되지 않는 파일 형식입니다. 이미지(PNG, JPEG, WebP, GIF, BMP) 또는 PDF, TXT, Markdown 파일을 사용하세요.',
+  'chat.attachment.imageNotSupported':
+    '이 모델은 이미지를 읽을 수 없습니다. 대신 PDF, TXT 또는 Markdown 파일을 첨부하세요.',
   'chat.attachment.readFailed': '파일을 읽을 수 없습니다',
   'memory.searchAria': '메모리 검색',
   'memory.searchPlaceholder': '메모리 항목 검색...',
@@ -3086,6 +3088,9 @@ const messages: TranslationMap = {
   'settings.ai.temperatureOverrideValue': '온도 재정의(값)',
   'settings.ai.temperatureOverrideDesc':
     '낮을수록 더 결정적입니다. 제공업체 기본값을 사용하려면 선택하지 않은 상태로 두세요.',
+  'settings.ai.modelVision': '비전 지원 (이미지 입력)',
+  'settings.ai.modelVisionDesc':
+    '이 모델이 이미지를 지원하면 활성화하세요. 이 모델을 선택하면 채팅 작성기에서 이미지를 첨부할 수 있습니다.',
   'settings.ai.testFailed': '테스트 실패',
   'settings.ai.testingModel': '모델 테스트 중...',
   'settings.ai.modelResponse': '모델 응답',

app/src/lib/i18n/pl.ts

@@ -1884,7 +1884,10 @@ const messages: TranslationMap = {
   'chat.attachment.tooMany': 'Maksymalnie {max} obrazów na wiadomość',
   'chat.attachment.tooManyFiles': 'Maksymalnie {max} plików na wiadomość',
   'chat.attachment.tooLarge': 'Obraz przekracza limit rozmiaru {max}',
-  'chat.attachment.unsupportedType': 'Nieobsługiwany typ pliku. Użyj PNG, JPEG, WebP, GIF lub BMP.',
+  'chat.attachment.unsupportedType':
+    'Nieobsługiwany typ pliku. Użyj obrazu (PNG, JPEG, WebP, GIF, BMP) lub pliku PDF, TXT albo Markdown.',
+  'chat.attachment.imageNotSupported':
+    'Ten model nie może odczytywać obrazów. Zamiast tego dołącz plik PDF, TXT lub Markdown.',
   'chat.attachment.readFailed': 'Nie można odczytać pliku',
   'memory.searchAria': 'Szukaj w pamięci',
   'memory.searchPlaceholder': 'Szukaj wpisów pamięci...',
@@ -3166,6 +3169,9 @@ const messages: TranslationMap = {
   'settings.ai.temperatureOverrideValue': 'Nadpisanie temperatury (wartość)',
   'settings.ai.temperatureOverrideDesc':
     'Niższa = bardziej deterministyczna. Pozostaw niezaznaczone, aby użyć domyślnej wartości dostawcy.',
+  'settings.ai.modelVision': 'Obsługuje wizję (wejście obrazu)',
+  'settings.ai.modelVisionDesc':
+    'Włącz, jeśli ten model akceptuje obrazy. Pozwala dołączać obrazy w czacie, gdy ten model jest wybrany.',
   'settings.ai.testFailed': 'Test nie powiódł się',
   'settings.ai.testingModel': 'Testowanie modelu...',
   'settings.ai.modelResponse': 'Odpowiedź modelu',

app/src/lib/i18n/pt.ts

@@ -1903,7 +1903,9 @@ const messages: TranslationMap = {
   'chat.attachment.tooManyFiles': 'Máximo de {max} arquivos por mensagem',
   'chat.attachment.tooLarge': 'A imagem excede o limite de tamanho de {max}',
   'chat.attachment.unsupportedType':
-    'Tipo de arquivo não suportado. Use PNG, JPEG, WebP, GIF ou BMP.',
+    'Tipo de arquivo não suportado. Use uma imagem (PNG, JPEG, WebP, GIF, BMP) ou um arquivo PDF, TXT ou Markdown.',
+  'chat.attachment.imageNotSupported':
+    'Este modelo não consegue ler imagens. Anexe um arquivo PDF, TXT ou Markdown.',
   'chat.attachment.readFailed': 'Não foi possível ler o arquivo',
   'memory.searchAria': 'Pesquisar memória',
   'memory.searchPlaceholder': 'Pesquisar entradas de memória...',
@@ -3169,6 +3171,9 @@ const messages: TranslationMap = {
   'settings.ai.temperatureOverrideValue': 'Substituição de temperatura (valor)',
   'settings.ai.temperatureOverrideDesc':
     'Mais baixo = mais determinístico. Deixe desmarcado para usar o padrão do provedor.',
+  'settings.ai.modelVision': 'Suporta visão (entrada de imagem)',
+  'settings.ai.modelVisionDesc':
+    'Ative se este modelo aceitar imagens. Permite anexar imagens no chat quando este modelo está selecionado.',
   'settings.ai.testFailed': 'Teste falhou',
   'settings.ai.testingModel': 'Modelo de teste...',
   'settings.ai.modelResponse': 'Resposta do modelo',

app/src/lib/i18n/ru.ts

@@ -1879,7 +1879,9 @@ const messages: TranslationMap = {
   'chat.attachment.tooManyFiles': 'Максимум {max} файлов на сообщение',
   'chat.attachment.tooLarge': 'Изображение превышает ограничение размера {max}',
   'chat.attachment.unsupportedType':
-    'Неподдерживаемый тип файла. Используйте PNG, JPEG, WebP, GIF или BMP.',
+    'Неподдерживаемый тип файла. Используйте изображение (PNG, JPEG, WebP, GIF, BMP) или файл PDF, TXT либо Markdown.',
+  'chat.attachment.imageNotSupported':
+    'Эта модель не может читать изображения. Вместо этого прикрепите файл PDF, TXT или Markdown.',
   'chat.attachment.readFailed': 'Не удалось прочитать файл',
   'memory.searchAria': 'Поиск в памяти',
   'memory.searchPlaceholder': 'Поиск записей памяти...',
@@ -3140,6 +3142,9 @@ const messages: TranslationMap = {
   'settings.ai.temperatureOverrideValue': 'Переопределение температуры (значение)',
   'settings.ai.temperatureOverrideDesc':
     'Ниже = более детерминированно. Оставьте флажок неактивным, чтобы использовать поставщика по умолчанию.',
+  'settings.ai.modelVision': 'Поддерживает зрение (ввод изображений)',
+  'settings.ai.modelVisionDesc':
+    'Включите, если модель принимает изображения. Позволяет прикреплять изображения в чате, когда выбрана эта модель.',
   'settings.ai.testFailed': 'Тест не пройден.',
   'settings.ai.testingModel': 'Модель тестирования...',
   'settings.ai.modelResponse': 'Ответ модели',

app/src/lib/i18n/zh-CN.ts

@@ -1760,7 +1760,9 @@ const messages: TranslationMap = {
   'chat.attachment.tooMany': '每条消息最多 {max} 张图片',
   'chat.attachment.tooManyFiles': '每条消息最多 {max} 个文件',
   'chat.attachment.tooLarge': '图片超过 {max} 大小限制',
-  'chat.attachment.unsupportedType': '不支持的文件类型。请使用 PNG、JPEG、WebP、GIF 或 BMP。',
+  'chat.attachment.unsupportedType':
+    '不支持的文件类型。请使用图片（PNG、JPEG、WebP、GIF、BMP）或 PDF、TXT、Markdown 文件。',
+  'chat.attachment.imageNotSupported': '此模型无法读取图像。请改为附加 PDF、TXT 或 Markdown 文件。',
   'chat.attachment.readFailed': '无法读取文件',
   'memory.searchAria': '搜索记忆',
   'memory.searchPlaceholder': '搜索记忆条目...',
@@ -2955,6 +2957,8 @@ const messages: TranslationMap = {
   'settings.ai.temperatureOverrideSlider': '温度超控（滑块）',
   'settings.ai.temperatureOverrideValue': '温度超控（值）',
   'settings.ai.temperatureOverrideDesc': '值越低，结果越确定。保持未勾选则使用提供商默认值。',
+  'settings.ai.modelVision': '支持视觉（图像输入）',
+  'settings.ai.modelVisionDesc': '如果此模型接受图像，请启用。选择此模型后可在聊天框中附加图像。',
   'settings.ai.testFailed': '测试失败',
   'settings.ai.testingModel': '测试模型...',
   'settings.ai.modelResponse': '模型响应',

app/src/pages/Conversations.tsx

@@ -17,7 +17,6 @@ import MicComposer from '../features/human/MicComposer';
 import { useStickToBottom } from '../hooks/useStickToBottom';
 import { useUsageState } from '../hooks/useUsageState';
 import {
-  ALLOWED_ATTACHMENT_MIME_TYPES,
   type Attachment,
   ATTACHMENT_MAX_FILES,
   ATTACHMENT_MAX_IMAGES,
@@ -270,6 +269,11 @@ const Conversations = ({
     onCancel: () => {},
   });
   const [resolvedModel, setResolvedModel] = useState<string | null>(null);
+  // Whether the resolved model for the active profile accepts image input.
+  // Managed tiers do; custom/BYOK models only when the user flagged them. Gates
+  // the composer's image-attachment affordance (docs flow regardless). Resolved
+  // against the non-attachment hint so the affordance is stable as you attach.
+  const [modelSupportsVision, setModelSupportsVision] = useState(false);
 
   useEffect(() => {
     let cancelled = false;
@@ -280,13 +284,19 @@ const Conversations = ({
           attachments.length > 0
             ? MULTIMODAL_MODEL_HINT
             : (profile?.modelOverride ?? CHAT_MODEL_HINT);
-        const res = await callCoreRpc<{ model: string }>({
+        const res = await callCoreRpc<{ model: string; vision?: boolean }>({
           method: 'openhuman.inference_resolve_model',
           params: { hint },
         });
-        if (!cancelled) setResolvedModel(res.model);
+        if (!cancelled) {
+          setResolvedModel(res.model);
+          setModelSupportsVision(res.vision === true);
+        }
       } catch {
-        if (!cancelled) setResolvedModel(null);
+        if (!cancelled) {
+          setResolvedModel(null);
+          setModelSupportsVision(false);
+        }
       }
     })();
     return () => {
@@ -680,10 +690,19 @@ const Conversations = ({
     let acceptedImageCount = attachments.filter(attachment => attachment.kind === 'image').length;
     let acceptedFileCount = attachments.filter(attachment => attachment.kind === 'file').length;
     for (const file of Array.from(files)) {
-      const result = await validateAndReadFile(file, acceptedImageCount, acceptedFileCount);
+      const result = await validateAndReadFile(
+        file,
+        acceptedImageCount,
+        acceptedFileCount,
+        modelSupportsVision
+      );
       if ('error' in result) {
         const { error } = result;
-        if (error.code === 'too_many') {
+        if (error.code === 'image_not_supported') {
+          setAttachError(
+            chatSendError('attachment_invalid', t('chat.attachment.imageNotSupported'))
+          );
+        } else if (error.code === 'too_many') {
           const key =
             error.kind === 'image' ? 'chat.attachment.tooMany' : 'chat.attachment.tooManyFiles';
           setAttachError(
@@ -1724,6 +1743,18 @@ const Conversations = ({
                                 ? (msg.extraMetadata.attachmentDataUris as string[])
                                 : parseMessageImages(msg.content ?? '').dataUris;
                               const hasImages = dataUris.length > 0;
+                              // Document attachments carry no image data-URI (only
+                              // images do); surface them as filename chips from the
+                              // persisted attachmentKinds/attachmentNames metadata.
+                              const kinds = Array.isArray(msg.extraMetadata?.attachmentKinds)
+                                ? (msg.extraMetadata.attachmentKinds as string[])
+                                : [];
+                              const names = Array.isArray(msg.extraMetadata?.attachmentNames)
+                                ? (msg.extraMetadata.attachmentNames as string[])
+                                : [];
+                              const fileNames = kinds
+                                .map((k, i) => (k === 'file' ? names[i] : null))
+                                .filter((n): n is string => Boolean(n));
                               const showTime = latestVisibleMessage?.id === msg.id;
                               return (
                                 <>
@@ -1739,6 +1770,35 @@ const Conversations = ({
                                       ))}
                                     </div>
                                   )}
+                                  {fileNames.length > 0 && (
+                                    <div className="flex flex-wrap gap-1.5 justify-end">
+                                      {fileNames.map((name, i) => (
+                                        <div
+                                          key={i}
+                                          className="flex items-center gap-2 rounded-lg border border-stone-200 dark:border-neutral-700 bg-stone-50 dark:bg-neutral-800 px-2.5 py-1.5 text-xs text-stone-700 dark:text-neutral-300 max-w-[220px]">
+                                          <svg
+                                            className="w-4 h-4 flex-shrink-0 text-stone-500 dark:text-neutral-400"
+                                            fill="none"
+                                            stroke="currentColor"
+                                            viewBox="0 0 24 24">
+                                            <path
+                                              strokeLinecap="round"
+                                              strokeLinejoin="round"
+                                              strokeWidth={1.8}
+                                              d="M14 2H6a2 2 0 00-2 2v16a2 2 0 002 2h12a2 2 0 002-2V8z"
+                                            />
+                                            <path
+                                              strokeLinecap="round"
+                                              strokeLinejoin="round"
+                                              strokeWidth={1.8}
+                                              d="M14 2v6h6"
+                                            />
+                                          </svg>
+                                          <span className="truncate font-medium">{name}</span>
+                                        </div>
+                                      ))}
+                                    </div>
+                                  )}
                                   {(msg.content || showTime) && (
                                     <div className="rounded-2xl px-4 py-2.5 bg-primary-500 text-white rounded-br-md break-words overflow-hidden">
                                       {msg.content && (
@@ -2213,7 +2273,10 @@ const Conversations = ({
               inlineCompletionSuffix={inlineCompletionSuffix}
               isComposingTextRef={isComposingTextRef}
               maxAttachments={ATTACHMENT_MAX_IMAGES + ATTACHMENT_MAX_FILES}
-              allowedMimeTypes={ALLOWED_ATTACHMENT_MIME_TYPES}
+              // Empty → no native `accept` filter (it greys valid files on
+              // macOS/CEF). Type enforcement happens in handleAttachFiles via
+              // validateAndReadFile, which honors modelSupportsVision.
+              allowedMimeTypes={[]}
               attachmentsEnabled={CHAT_ATTACHMENTS_ENABLED}
             />
           ) : (

app/src/pages/__tests__/Conversations.attachments.test.tsx

@@ -140,6 +140,18 @@ vi.mock('../../features/autocomplete/useAutocompleteSkillStatus', () => ({
 
 vi.mock('../../utils/openUrl', () => ({ openUrl: vi.fn() }));
 
+// The composer gates image attachments on the resolved model's vision capability
+// (inference_resolve_model). These tests exercise image attachments, so resolve
+// the active model as vision-capable.
+vi.mock('../../services/coreRpcClient', () => ({
+  callCoreRpc: vi.fn(async ({ method }: { method: string }) =>
+    method === 'openhuman.inference_resolve_model'
+      ? { model: 'test-vision-model', vision: true }
+      : {}
+  ),
+  CoreRpcError: class CoreRpcError extends Error {},
+}));
+
 vi.mock('../../lib/coreState/store', () => ({
   getCoreStateSnapshot: vi.fn(() => ({
     isBootstrapping: false,
@@ -506,4 +518,53 @@ describe('Conversations — attachment feature', () => {
       expect(img).not.toBeNull();
     });
   });
+
+  it('renders a document filename chip in the user bubble from attachmentKinds/Names', async () => {
+    const thread = makeThread({ id: 'file-thread', title: 'File Thread' });
+    const message = {
+      id: 'msg-file-1',
+      content: 'whats in this file',
+      type: 'text' as const,
+      sender: 'user' as const,
+      createdAt: new Date().toISOString(),
+      extraMetadata: {
+        attachmentCount: 1,
+        attachmentKinds: ['file'],
+        attachmentNames: ['report.pdf'],
+      },
+    };
+
+    mockGetThreads.mockResolvedValue({ threads: [thread], count: 1 });
+    mockGetThreadMessages.mockResolvedValue({ messages: [message], count: 1 });
+
+    const store = buildStore({
+      thread: {
+        threads: [thread],
+        selectedThreadId: thread.id,
+        activeThreadId: null,
+        welcomeThreadId: null,
+        messagesByThreadId: { [thread.id]: [message] },
+        messages: [message],
+        isLoadingThreads: false,
+        isLoadingMessages: false,
+        messagesError: null,
+      },
+      socket: socketState('connected'),
+    });
+
+    const { default: Conversations } = await import('../Conversations');
+
+    render(
+      <Provider store={store}>
+        <MemoryRouter>
+          <Conversations />
+        </MemoryRouter>
+      </Provider>
+    );
+
+    // The document attachment surfaces as a filename chip (not an <img>).
+    await waitFor(() => {
+      expect(document.body.textContent).toContain('report.pdf');
+    });
+  });
 });

app/src/services/api/__tests__/aiSettingsApi.test.ts

@@ -18,6 +18,7 @@ import {
   loadAISettings,
   loadLocalProviderSnapshot,
   localProvider,
+  modelRegistryVision,
   OPENAI_CODEX_OAUTH_MISSING_AUTH_URL,
   OPENAI_CODEX_OAUTH_MISSING_CALLBACK_URL,
   parseProviderString,
@@ -28,6 +29,7 @@ import {
   setLocalRuntimeEnabled,
   startOpenAiCodexOAuth,
   testProviderModel,
+  upsertModelRegistryVision,
 } from '../aiSettingsApi';
 
 // ─── Mock declarations (must be hoisted before imports) ───────────────────────
@@ -83,6 +85,7 @@ function makeClientConfigResult(overrides: Record<string, unknown> = {}) {
       api_key_set: false,
       model_routes: [],
       cloud_providers: [],
+      model_registry: [],
       primary_cloud: null,
       reasoning_provider: null,
       agentic_provider: null,
@@ -587,6 +590,7 @@ describe('saveAISettings', () => {
         learning: { kind: 'openhuman' },
         subconscious: { kind: 'openhuman' },
       },
+      modelRegistry: [],
       ...overrides,
     };
   }
@@ -673,8 +677,13 @@ describe('saveAISettings', () => {
         learning: { kind: 'openhuman' },
         subconscious: { kind: 'openhuman' },
       },
+      modelRegistry: [],
+    };
+    const next: AISettings = {
+      cloudProviders: [anthropicProvider],
+      routing: { ...prev.routing },
+      modelRegistry: [],
     };
-    const next: AISettings = { cloudProviders: [anthropicProvider], routing: { ...prev.routing } };
 
     await saveAISettings(prev, next);
 
@@ -697,6 +706,34 @@ describe('saveAISettings', () => {
     expect(patch.cloud_providers).toBeDefined();
     expect(patch.coding_provider).toBe('openai:gpt-4o-mini');
   });
+
+  it('sends model_registry when the vision flag changes', async () => {
+    const prev = makeSettings({ modelRegistry: [] });
+    const next = makeSettings({
+      modelRegistry: [{ id: 'my-llava', provider: 'openai', cost_per_1m_output: 0, vision: true }],
+    });
+    await saveAISettings(prev, next);
+    const patch = mockOpenhumanUpdateModelSettings.mock.calls[0][0];
+    expect(patch.model_registry).toEqual([
+      { id: 'my-llava', provider: 'openai', cost_per_1m_output: 0, vision: true },
+    ]);
+  });
+
+  it('omits model_registry when unchanged', async () => {
+    const registry = [{ id: 'my-llava', provider: 'openai', cost_per_1m_output: 0, vision: true }];
+    const prev = makeSettings({ modelRegistry: registry });
+    const next = makeSettings({
+      modelRegistry: [...registry],
+      routing: {
+        ...makeSettings().routing,
+        coding: { kind: 'cloud', providerSlug: 'openai', model: 'gpt-4o-mini' },
+      },
+    });
+    await saveAISettings(prev, next);
+    const patch = mockOpenhumanUpdateModelSettings.mock.calls[0][0];
+    expect(patch.model_registry).toBeUndefined();
+    expect(patch.coding_provider).toBe('openai:gpt-4o-mini');
+  });
 });
 
 // ─── setCloudProviderKey ──────────────────────────────────────────────────────
@@ -918,3 +955,32 @@ describe('flushCloudProviders', () => {
     expect(mockOpenhumanUpdateModelSettings).not.toHaveBeenCalled();
   });
 });
+
+describe('model registry vision helpers', () => {
+  const reg = [
+    { id: 'gpt-4o', provider: 'openai', cost_per_1m_output: 0, vision: true },
+    { id: 'text-only', provider: 'openai', cost_per_1m_output: 0, vision: false },
+  ];
+
+  it('modelRegistryVision matches by (provider, id)', () => {
+    expect(modelRegistryVision(reg, 'openai', 'gpt-4o')).toBe(true);
+    expect(modelRegistryVision(reg, 'openai', 'text-only')).toBe(false);
+    expect(modelRegistryVision(reg, 'openai', 'unlisted')).toBe(false);
+    expect(modelRegistryVision(reg, 'azure', 'gpt-4o')).toBe(false);
+  });
+
+  it('upsertModelRegistryVision adds, flips, and removes entries', () => {
+    const added = upsertModelRegistryVision([], 'openai', 'my-llava', true);
+    expect(added).toEqual([
+      { id: 'my-llava', provider: 'openai', cost_per_1m_output: 0, vision: true },
+    ]);
+    // vision:false removes the entry (absence ⇒ no vision).
+    const removed = upsertModelRegistryVision(reg, 'openai', 'gpt-4o', false);
+    expect(removed.find(e => e.id === 'gpt-4o')).toBeUndefined();
+    expect(removed.find(e => e.id === 'text-only')).toBeDefined();
+    // Flipping an existing entry on stays idempotent on the key.
+    const flipped = upsertModelRegistryVision(reg, 'openai', 'text-only', true);
+    expect(flipped.filter(e => e.id === 'text-only')).toHaveLength(1);
+    expect(modelRegistryVision(flipped, 'openai', 'text-only')).toBe(true);
+  });
+});

app/src/services/api/aiSettingsApi.ts

@@ -26,6 +26,7 @@ import { isTauri } from '../../utils/tauriCommands/common';
 import {
   type ClientConfig,
   type CloudProviderCreds,
+  type ModelRegistryEntry,
   type ModelSettingsUpdate,
   openhumanGetClientConfig,
   openhumanUpdateLocalAiSettings,
@@ -132,6 +133,47 @@ const PROVIDER_MODEL_TEST_TIMEOUT_MS = 120_000;
 export interface AISettings {
   cloudProviders: CloudProviderView[];
   routing: Record<WorkloadId, ProviderRef>;
+  /**
+   * Per-model registry carrying each model's user-set `vision` flag, keyed by
+   * `(provider slug, model id)`. Surfaced in the custom-model dialog; gates chat
+   * image attachments for custom/BYOK models.
+   */
+  modelRegistry: ModelRegistryEntry[];
+}
+
+/** Re-export so callers (e.g. the AI panel) can reference the entry type. */
+export type { ModelRegistryEntry };
+
+/** Find a model's vision flag in the registry, matching by (provider, id).
+ *  Tolerates an undefined registry (older snapshots / transient load state). */
+export function modelRegistryVision(
+  registry: ModelRegistryEntry[] | undefined,
+  provider: string,
+  id: string
+): boolean {
+  return (registry ?? []).some(e => e.provider === provider && e.id === id && e.vision);
+}
+
+/**
+ * Upsert a model's vision flag, returning a new array. Matches by
+ * `(provider, id)`; a `vision: false` entry is removed (absence ⇒ no vision).
+ */
+export function upsertModelRegistryVision(
+  registry: ModelRegistryEntry[] | undefined,
+  provider: string,
+  id: string,
+  vision: boolean
+): ModelRegistryEntry[] {
+  const base = registry ?? [];
+  const without = base.filter(e => !(e.provider === provider && e.id === id));
+  if (!vision) {
+    return without;
+  }
+  const existing = base.find(e => e.provider === provider && e.id === id);
+  return [
+    ...without,
+    { id, provider, cost_per_1m_output: existing?.cost_per_1m_output ?? 0, vision: true },
+  ];
 }
 
 // ─── Read path: load + parse ───────────────────────────────────────────────
@@ -264,7 +306,10 @@ export async function loadAISettings(): Promise<AISettings> {
     );
   }
 
-  return { cloudProviders, routing };
+  // Per-model registry (vision flags). Defensive default for older snapshots.
+  const modelRegistry: ModelRegistryEntry[] = config.model_registry ?? [];
+
+  return { cloudProviders, routing, modelRegistry };
 }
 // ─── Write path: diff + save ───────────────────────────────────────────────
 
@@ -310,12 +355,37 @@ export async function saveAISettings(prev: AISettings, next: AISettings): Promis
     }
   }
 
+  // Per-model registry (vision flags): any change → send the full list.
+  if (!modelRegistriesEqual(prev.modelRegistry, next.modelRegistry)) {
+    patch.model_registry = next.modelRegistry.map(
+      ({ id, provider, cost_per_1m_output, vision }) => ({
+        id,
+        provider,
+        cost_per_1m_output,
+        vision,
+      })
+    );
+  }
+
   if (Object.keys(patch).length === 0) {
     return;
   }
   await openhumanUpdateModelSettings(patch);
 }
 
+/** Order-insensitive structural equality for two model registries. */
+function modelRegistriesEqual(a: ModelRegistryEntry[], b: ModelRegistryEntry[]): boolean {
+  if (a.length !== b.length) {
+    return false;
+  }
+  const key = (e: ModelRegistryEntry) => `${e.provider} ${e.id}`;
+  const bByKey = new Map(b.map(e => [key(e), e]));
+  return a.every(e => {
+    const m = bByKey.get(key(e));
+    return !!m && m.vision === e.vision && m.cost_per_1m_output === e.cost_per_1m_output;
+  });
+}
+
 // ─── API key management (per cloud provider slug) ──────────────────────────
 
 /**

app/src/utils/tauriCommands/config.ts

@@ -49,6 +49,18 @@ export interface CloudProviderCreds {
   auth_style: AuthStyle;
 }
 
+/**
+ * Per-model registry entry. Mirrors the Rust `ModelRegistryEntry`
+ * (`config/schema/types.rs`). Carries the user-set `vision` flag that lets a
+ * custom/BYOK model accept chat image attachments.
+ */
+export interface ModelRegistryEntry {
+  id: string;
+  provider: string;
+  cost_per_1m_output: number;
+  vision: boolean;
+}
+
 export interface ModelSettingsUpdate {
   /**
    * OpenHuman product backend URL. Almost always left untouched; the
@@ -77,6 +89,12 @@ export interface ModelSettingsUpdate {
    * Each entry: { id?, slug, label?, endpoint, auth_style? }
    */
   cloud_providers?: CloudProviderCreds[] | null;
+  /**
+   * When present, REPLACES `config.model_registry` wholesale. Carries each
+   * model's `vision` flag (Settings → Advanced LLM → custom model → "Supports
+   * vision"). Send `[]` to clear; omit to leave untouched.
+   */
+  model_registry?: ModelRegistryEntry[] | null;
   /** @deprecated No longer used — slug-based routing replaces primary_cloud. */
   primary_cloud?: string | null;
   /** Per-workload provider strings — see Rust `providers::factory` grammar. */
@@ -215,6 +233,8 @@ export interface ClientConfig {
   model_routes: ModelRoute[];
   /** Configured cloud providers (no API keys — those live in auth-profiles.json). */
   cloud_providers: CloudProviderCreds[];
+  /** Per-model registry carrying each model's `vision` flag. */
+  model_registry: ModelRegistryEntry[];
   /** Id of the `cloud_providers` entry resolved by the `"cloud"` sentinel. */
   primary_cloud: string | null;
   /** Per-workload provider strings (e.g. `"cloud"`, `"ollama:llama3.1:8b"`, `"openai:gpt-4o"`). */

app/test/playwright/specs/chat-management-functional.spec.ts

@@ -2,7 +2,6 @@ import { expect, type Page, test } from '@playwright/test';
 
 import {
   bootAuthenticatedPage,
-  callCoreRpc,
   dismissWalkthroughIfPresent,
   waitForAppReady,
 } from '../helpers/core-rpc';
@@ -73,7 +72,7 @@ async function newThread(page: Page): Promise<string> {
 }
 
 test.describe('Chat management functional coverage', () => {
-  test('attachment preview, remove, and multimodal send path remain interactive', async ({
+  test('attachment preview, remove, and document send path remain interactive', async ({
     page,
   }) => {
     await resetMock();
@@ -89,34 +88,24 @@ test.describe('Chat management functional coverage', () => {
 
     const fileInput = page.locator('input[type="file"]');
     await expect(fileInput).toHaveCount(1);
-    await fileInput.setInputFiles({
-      name: 'pixel.png',
-      mimeType: 'image/png',
-      buffer: Buffer.from(
-        'iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mP8/x8AAwMCAO+/p9sAAAAASUVORK5CYII=',
-        'base64'
-      ),
-    });
 
-    await expect(page.getByText('pixel.png')).toBeVisible();
-    await page.getByRole('button', { name: /Remove pixel\.png/ }).click();
-    await expect(page.getByText('pixel.png')).toHaveCount(0);
+    // Documents work on every model; images require a vision-capable model.
+    // Managed OpenHuman tiers are text-only (see `oh_tier_supports_vision`), so
+    // the attachment mechanics here are exercised with a text document. Image
+    // upload is gated to vision-flagged models and validated in unit tests.
+    const txtBuffer = Buffer.from('renderer uploaded text document', 'utf8');
+    const attachTxt = () =>
+      fileInput.setInputFiles({ name: 'notes.txt', mimeType: 'text/plain', buffer: txtBuffer });
 
-    const pngBuffer = Buffer.from(
-      'iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mP8/x8AAwMCAO+/p9sAAAAASUVORK5CYII=',
-      'base64'
-    );
-    await fileInput.setInputFiles({ name: 'pixel.png', mimeType: 'image/png', buffer: pngBuffer });
-    await expect(page.getByText('pixel.png')).toBeVisible();
+    await attachTxt();
+    await expect(page.getByText('notes.txt')).toBeVisible();
+    await page.getByRole('button', { name: /Remove notes\.txt/ }).click();
+    await expect(page.getByText('notes.txt')).toHaveCount(0);
 
-    await fileInput.setInputFiles({
-      name: 'notes.txt',
-      mimeType: 'text/plain',
-      buffer: Buffer.from('renderer uploaded text document', 'utf8'),
-    });
+    await attachTxt();
     await expect(page.getByText('notes.txt')).toBeVisible();
 
-    await page.getByPlaceholder('How can I help you today?').fill('Describe this image');
+    await page.getByPlaceholder('How can I help you today?').fill('Summarize this file');
     await page.getByTestId('send-message-button').click();
     await expect(page.getByText('Attachment received by the assistant.')).toBeVisible({
       timeout: 30_000,
@@ -132,13 +121,13 @@ test.describe('Chat management functional coverage', () => {
               request.method === 'POST' &&
               request.url?.includes('/chat/completions') &&
               typeof request.body === 'string' &&
-              request.body.includes('Describe this image')
+              request.body.includes('Summarize this file')
           );
           return completion?.body ?? '';
         },
         { timeout: 10_000 }
       )
-      .toContain('image_url');
+      .toContain('[FILE-EXTRACTED:');
 
     const log = await requests();
     const completionBody =
@@ -147,13 +136,21 @@ test.describe('Chat management functional coverage', () => {
           request.method === 'POST' &&
           request.url?.includes('/chat/completions') &&
           typeof request.body === 'string' &&
-          request.body.includes('Describe this image')
+          request.body.includes('Summarize this file')
       )?.body ?? '';
     expect(completionBody).toContain('reasoning');
     expect(completionBody).toContain('[FILE-EXTRACTED:');
     expect(completionBody).toContain('renderer uploaded text document');
   });
 
+  // NOTE: image attachments require a vision-capable model. Managed tiers are
+  // text-only by default (`oh_tier_supports_vision`); a model is flagged
+  // vision-capable via `model_registry` (Settings → "Supports vision"). The
+  // `image_url` promotion path is covered by Rust unit tests
+  // (`inference::provider::compatible` MessageContent). An E2E that toggles the
+  // vision flag at runtime is a follow-up (needs a page re-mount so the
+  // composer's resolve picks up the flag — out of scope here).
+
   test('thread rename and delete remain usable from the conversation UI', async ({ page }) => {
     await resetMock();
     await openChat(page, 'pw-chat-rename-delete');

src/openhuman/agent/harness/engine/core.rs

@@ -248,7 +248,14 @@ pub(crate) async fn run_turn_engine(
 
         tracing::debug!(iteration, "[agent_loop] sending LLM request");
         let image_marker_count = multimodal::count_image_markers(history);
-        if image_marker_count > 0 && !provider.supports_vision() {
+        // A model accepts images when its provider advertises vision (managed
+        // backend) OR the user marked it vision-capable in `model_registry`
+        // (custom/BYOK) — surfaced via the `current_model_vision` task-local set
+        // at session build. Unset (CLI/tests) falls back to the provider flag.
+        let has_vision = provider.supports_vision()
+            || crate::openhuman::agent::harness::model_vision_context::current_model_vision()
+                .unwrap_or(false);
+        if image_marker_count > 0 && !has_vision {
             let cap_err = ProviderCapabilityError {
                 provider: provider_name.to_string(),
                 capability: "vision".to_string(),

src/openhuman/agent/harness/mod.rs

@@ -30,6 +30,7 @@ mod instructions;
 pub mod interrupt;
 pub(crate) mod memory_context;
 pub(crate) mod memory_context_safety;
+pub mod model_vision_context;
 mod parse;
 pub(crate) mod payload_summarizer;
 pub mod run_queue;
@@ -50,6 +51,7 @@ pub use definition::{
 };
 pub use fork_context::{current_parent, with_parent_context, ParentExecutionContext};
 pub use interrupt::{check_interrupt, InterruptFence, InterruptedError};
+pub use model_vision_context::{current_model_vision, with_current_model_vision};
 pub use sandbox_context::{current_sandbox_mode, with_current_sandbox_mode};
 pub(crate) use spawn_depth_context::{current_spawn_depth, with_spawn_depth, MAX_SPAWN_DEPTH};
 pub use subagent_runner::{run_subagent, SubagentRunError, SubagentRunOptions};

src/openhuman/agent/harness/model_vision_context.rs

@@ -0,0 +1,75 @@
+//! Task-local carrier for the **current session/sub-agent's user-configured
+//! vision capability** so the deep turn engine's image gate can honor a custom
+//! (BYOK) model's `model_registry.vision` flag without widening
+//! [`crate::openhuman::agent::harness::engine::run_turn_engine`]'s signature.
+//!
+//! Managed-backend models advertise vision via `Provider::supports_vision()`, so
+//! the gate accepts their image turns already. Custom OpenAI-compatible providers
+//! report `supports_vision() == false` (a provider endpoint can't know a per-model
+//! property), so without this the gate would reject image turns for a model the
+//! user explicitly marked vision-capable. This task-local surfaces that per-model
+//! flag — computed once at session build (where the full `Config` / `model_registry`
+//! and the resolved model id coexist) — to the gate.
+//!
+//! Mirrors [`super::sandbox_context`]. When unset (CLI / direct invocation / tests
+//! that never wrapped the call) [`current_model_vision`] returns `None` and the
+//! gate falls back to the provider capability only — strictly additive.
+
+tokio::task_local! {
+    /// User-configured vision capability for the currently-executing
+    /// session/sub-agent's model. Scoped per turn by the turn loop + subagent
+    /// runner. `None` when unset.
+    pub static CURRENT_MODEL_VISION: bool;
+}
+
+/// Returns the current model's user-configured vision flag, if scope is active.
+pub fn current_model_vision() -> Option<bool> {
+    CURRENT_MODEL_VISION.try_with(|v| *v).ok()
+}
+
+/// Run `future` with `vision` installed as the current model's vision flag.
+/// Intended call site is around each `run_turn_engine` invocation.
+pub async fn with_current_model_vision<F, R>(vision: bool, future: F) -> R
+where
+    F: std::future::Future<Output = R>,
+{
+    CURRENT_MODEL_VISION.scope(vision, future).await
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[tokio::test]
+    async fn current_model_vision_returns_none_outside_scope() {
+        assert_eq!(current_model_vision(), None);
+    }
+
+    #[tokio::test]
+    async fn with_current_model_vision_installs_value() {
+        let observed = with_current_model_vision(true, async { current_model_vision() }).await;
+        assert_eq!(observed, Some(true));
+    }
+
+    #[tokio::test]
+    async fn with_current_model_vision_does_not_leak_across_scopes() {
+        with_current_model_vision(true, async {
+            assert_eq!(current_model_vision(), Some(true));
+        })
+        .await;
+        assert_eq!(current_model_vision(), None);
+    }
+
+    #[tokio::test]
+    async fn nested_scope_overrides_outer() {
+        with_current_model_vision(false, async {
+            assert_eq!(current_model_vision(), Some(false));
+            with_current_model_vision(true, async {
+                assert_eq!(current_model_vision(), Some(true));
+            })
+            .await;
+            assert_eq!(current_model_vision(), Some(false));
+        })
+        .await;
+    }
+}

src/openhuman/agent/harness/session/builder/factory.rs

@@ -439,6 +439,15 @@ impl Agent {
             model_name = pinned_model.to_string();
         }
 
+        // Resolve the user-configured vision flag for the (now-final) model while
+        // the full `Config` / `model_registry` is in scope — the turn engine only
+        // sees `MultimodalConfig`. Stored on the session and surfaced to the image
+        // gate via the `current_model_vision` task-local (covers custom/BYOK models
+        // the provider can't introspect). Computed with `&model_name` since it's
+        // moved into the builder below.
+        let model_vision =
+            crate::openhuman::inference::model_context::model_supports_vision(&model_name, config);
+
         // Dispatcher selection is deferred until after the tool list is
         // finalised (orchestrator tools are appended below). We capture
         // the choice string now so the provider borrow doesn't conflict
@@ -1047,6 +1056,7 @@ impl Agent {
             .config(config.agent.clone())
             .context_config(config.context.clone())
             .model_name(model_name)
+            .model_vision(model_vision)
             .temperature(effective_temperature)
             .workspace_dir(config.workspace_dir.clone())
             .action_dir(config.action_dir.clone())

src/openhuman/agent/harness/session/builder/setters.rs

@@ -29,6 +29,7 @@ impl AgentBuilder {
             config: None,
             context_config: None,
             model_name: None,
+            model_vision: None,
             temperature: None,
             workspace_dir: None,
             action_dir: None,
@@ -142,6 +143,14 @@ impl AgentBuilder {
         self
     }
 
+    /// Sets the user-configured vision capability for the resolved model.
+    /// Surfaced to the turn engine's image gate via the `current_model_vision`
+    /// task-local. Defaults to `false` when unset.
+    pub fn model_vision(mut self, model_vision: bool) -> Self {
+        self.model_vision = Some(model_vision);
+        self
+    }
+
     /// Sets the temperature for chat requests.
     pub fn temperature(mut self, temperature: f64) -> Self {
         self.temperature = Some(temperature);
@@ -514,6 +523,7 @@ impl AgentBuilder {
                 .unwrap_or_else(|| Box::new(DefaultMemoryLoader::default())),
             config,
             model_name,
+            model_vision: self.model_vision.unwrap_or(false),
             temperature: self.temperature.unwrap_or(0.7),
             workspace_dir,
             action_dir,

src/openhuman/agent/harness/session/turn/core.rs

@@ -323,6 +323,10 @@ impl Agent {
         // model prompt, not in the Rust-side classifier. Sub-agents pick
         // their own tier via `ModelSpec::Hint(...)` in their definition.
         let effective_model = self.model_name.clone();
+        // Capture before `self` is borrowed by the turn observer below, so it can
+        // be installed as the `current_model_vision` task-local around the engine
+        // call (read by the image gate for custom/BYOK vision models).
+        let model_vision = self.model_vision;
         log::info!(
             "[agent_loop] model pinned model={} (per-turn classification disabled for KV cache stability)",
             effective_model
@@ -436,25 +440,28 @@ impl Agent {
             // overflow happens during the parent's poll on the way in
             // — verified against the `chat-harness-subagent` Playwright
             // lane crash on PR #3151.
-            let outcome = Box::pin(super::super::super::engine::run_turn_engine(
-                provider.as_ref(),
-                &mut buf,
-                &mut tool_source,
-                &progress,
-                &mut observer,
-                &checkpoint,
-                &parser,
-                &provider_name,
-                &effective_model,
-                temperature,
-                true, // silent — the channel/UI renders via progress + the return value
-                &multimodal,
-                &multimodal_files,
-                max_iterations,
-                None, // the web bridge streams via on_progress deltas, not on_delta
-                &[],
-                turn_run_queue,
-            ))
+            let outcome = super::super::super::model_vision_context::with_current_model_vision(
+                model_vision,
+                Box::pin(super::super::super::engine::run_turn_engine(
+                    provider.as_ref(),
+                    &mut buf,
+                    &mut tool_source,
+                    &progress,
+                    &mut observer,
+                    &checkpoint,
+                    &parser,
+                    &provider_name,
+                    &effective_model,
+                    temperature,
+                    true, // silent — the channel/UI renders via progress + the return value
+                    &multimodal,
+                    &multimodal_files,
+                    max_iterations,
+                    None, // the web bridge streams via on_progress deltas, not on_delta
+                    &[],
+                    turn_run_queue,
+                )),
+            )
             .await?;
 
             // Pull the observer's accounting out, then drop it to release the

src/openhuman/agent/harness/session/types.rs

@@ -53,6 +53,11 @@ pub struct Agent {
     pub(super) memory_loader: Box<dyn MemoryLoader>,
     pub(super) config: crate::openhuman::config::AgentConfig,
     pub(super) model_name: String,
+    /// User-configured vision capability for [`Self::model_name`], evaluated at
+    /// session build from `model_vision_enabled(&model, config)`. Surfaced to the
+    /// turn engine's image gate via the `current_model_vision` task-local so a
+    /// custom/BYOK model the user flagged can forward images. Defaults to `false`.
+    pub(super) model_vision: bool,
     pub(super) temperature: f64,
     pub(super) workspace_dir: std::path::PathBuf,
     pub(super) action_dir: std::path::PathBuf,
@@ -272,6 +277,8 @@ pub struct AgentBuilder {
     /// [`crate::openhuman::config::ContextConfig::default`].
     pub(super) context_config: Option<crate::openhuman::config::ContextConfig>,
     pub(super) model_name: Option<String>,
+    /// User vision flag for the resolved model; `None` → `false` in `build()`.
+    pub(super) model_vision: Option<bool>,
     pub(super) temperature: Option<f64>,
     pub(super) workspace_dir: Option<std::path::PathBuf>,
     pub(super) action_dir: Option<std::path::PathBuf>,

src/openhuman/agent/harness/subagent_runner/ops/loop_.rs

@@ -48,6 +48,10 @@ pub(super) async fn run_inner_loop(
     allowed_names: HashSet<String>,
     lazy_resolver: Option<LazyToolkitResolver>,
     model: &str,
+    // User-configured vision flag for `model` (computed at the call site), set as
+    // the `current_model_vision` task-local around the engine call so a flagged
+    // custom/BYOK sub-agent model can forward images.
+    model_vision: bool,
     temperature: f64,
     max_iterations: usize,
     task_id: &str,
@@ -176,25 +180,28 @@ pub(super) async fn run_inner_loop(
     // `nested_subagent_dispatch_runs_on_a_constrained_worker_stack`;
     // the deep end-to-end catcher is the `chat-harness-subagent`
     // Playwright spec.
-    let outcome = Box::pin(super::super::super::engine::run_turn_engine(
-        provider,
-        history,
-        &mut tool_source,
-        &progress,
-        &mut observer,
-        &checkpoint,
-        &parser,
-        "subagent",
-        model,
-        temperature,
-        true, // silent — sub-agents never echo to stdout
-        &crate::openhuman::config::MultimodalConfig::default(),
-        &crate::openhuman::config::MultimodalFileConfig::default(),
-        max_iterations,
-        None, // sub-agents don't stream a draft
-        &["ask_user_clarification"],
-        None, // sub-agents don't support run-queue steering
-    ))
+    let outcome = super::super::super::model_vision_context::with_current_model_vision(
+        model_vision,
+        Box::pin(super::super::super::engine::run_turn_engine(
+            provider,
+            history,
+            &mut tool_source,
+            &progress,
+            &mut observer,
+            &checkpoint,
+            &parser,
+            "subagent",
+            model,
+            temperature,
+            true, // silent — sub-agents never echo to stdout
+            &crate::openhuman::config::MultimodalConfig::default(),
+            &crate::openhuman::config::MultimodalFileConfig::default(),
+            max_iterations,
+            None, // sub-agents don't stream a draft
+            &["ask_user_clarification"],
+            None, // sub-agents don't support run-queue steering
+        )),
+    )
     .await?;
 
     Ok((

src/openhuman/agent/harness/subagent_runner/ops/runner.rs

@@ -645,6 +645,20 @@ async fn run_typed_mode(
         };
 
     // ── Run the inner tool-call loop ───────────────────────────────────
+    // Resolve the sub-agent model's user-configured vision flag; defaults to
+    // `false` when config can't be loaded. Combined with the provider capability
+    // at the gate, this lets a flagged custom/BYOK sub-agent model forward images.
+    let model_vision = crate::openhuman::config::Config::load_or_init()
+        .await
+        .ok()
+        .map(|cfg| crate::openhuman::inference::model_context::model_supports_vision(&model, &cfg))
+        .unwrap_or(false);
+    tracing::debug!(
+        target: "subagent_runner",
+        model = %model,
+        model_vision,
+        "[subagent_runner] resolved sub-agent model vision capability"
+    );
     let (output, iterations, _agg_usage, early_exit_tool) = Box::pin(run_inner_loop(
         subagent_provider.as_ref(),
         &mut history,
@@ -654,6 +668,7 @@ async fn run_typed_mode(
         allowed_names,
         lazy_resolver,
         &model,
+        model_vision,
         temperature,
         definition.effective_max_iterations(),
         task_id,

src/openhuman/channels/providers/web/session.rs

@@ -11,6 +11,14 @@ pub(super) fn autonomy_signature(config: &Config) -> String {
     serde_json::to_string(&config.autonomy).unwrap_or_default()
 }
 
+/// Signature of `config.model_registry` for the session-cache fingerprint.
+/// Captures every per-model `vision` flag so toggling one in Settings forces a
+/// rebuild (picking up the new build-time `model_vision`). Mirrors
+/// [`autonomy_signature`].
+pub(super) fn model_registry_signature(config: &Config) -> String {
+    serde_json::to_string(&config.model_registry).unwrap_or_default()
+}
+
 pub(super) fn pick_target_agent_id(_config: &Config, profile: &AgentProfile) -> String {
     if profile.id == DEFAULT_PROFILE_ID {
         "orchestrator".to_string()
@@ -186,5 +194,6 @@ pub(super) fn build_session_fingerprint(
         ),
         target_agent_id,
         autonomy_signature: autonomy_signature(config),
+        model_registry_signature: model_registry_signature(config),
     }
 }

src/openhuman/channels/providers/web/types.rs

@@ -19,6 +19,12 @@ pub(crate) struct SessionCacheFingerprint {
     pub(super) target_agent_id: String,
     pub(super) provider_binding: String,
     pub(super) autonomy_signature: String,
+    /// Signature of `config.model_registry`. The cached `Agent` stores a
+    /// build-time `model_vision` bool; toggling a model's "Supports vision" flag
+    /// keeps the same model id (so neither `model_override` nor `provider_binding`
+    /// change) — without this the stale session would be reused. Mirrors
+    /// [`Self::autonomy_signature`].
+    pub(super) model_registry_signature: String,
 }
 
 pub(super) struct SessionEntry {

src/openhuman/channels/providers/web_tests.rs

@@ -1208,6 +1208,7 @@ fn fp(
         target_agent_id: target.to_string(),
         provider_binding: provider_binding.to_string(),
         autonomy_signature: "sig-default".to_string(),
+        model_registry_signature: "registry-default".to_string(),
     }
 }
 
@@ -1225,6 +1226,20 @@ fn fingerprint_autonomy_change_is_cache_miss() {
     );
 }
 
+#[test]
+fn fingerprint_model_registry_change_is_cache_miss() {
+    // Toggling a model's "Supports vision" flag keeps the same model id, so it
+    // changes neither model_override nor provider_binding. Without the registry
+    // signature the stale Agent (old build-time model_vision) would be reused.
+    let base = fp(None, None, "orchestrator", "openai:my-llava");
+    let mut changed = fp(None, None, "orchestrator", "openai:my-llava");
+    changed.model_registry_signature = "registry-after-vision-toggle".to_string();
+    assert_ne!(
+        base, changed,
+        "a model_registry change (vision toggle) must produce a cache miss → rebuild"
+    );
+}
+
 #[test]
 fn fingerprint_identical_inputs_are_cache_hit() {
     let a = fp(None, None, "orchestrator", "anthropic:claude-sonnet-4-6");

src/openhuman/config/ops/loader.rs

@@ -254,6 +254,18 @@ pub fn client_config_json(config: &Config) -> serde_json::Value {
             })
         })
         .collect();
+    let model_registry: Vec<serde_json::Value> = config
+        .model_registry
+        .iter()
+        .map(|m| {
+            serde_json::json!({
+                "id": m.id,
+                "provider": m.provider,
+                "cost_per_1m_output": m.cost_per_1m_output,
+                "vision": m.vision,
+            })
+        })
+        .collect();
 
     serde_json::json!({
         "api_url": config.api_url,
@@ -263,6 +275,7 @@ pub fn client_config_json(config: &Config) -> serde_json::Value {
         "api_key_set": api_key_set,
         "model_routes": model_routes,
         "cloud_providers": cloud_providers,
+        "model_registry": model_registry,
         "primary_cloud": config.primary_cloud,
         "chat_provider": config.chat_provider,
         "reasoning_provider": config.reasoning_provider,

src/openhuman/config/ops/model.rs

@@ -26,6 +26,10 @@ pub struct ModelSettingsPatch {
     /// Pass `Some(vec![])` to clear all third-party cloud providers.
     pub cloud_providers:
         Option<Vec<crate::openhuman::config::schema::cloud_providers::CloudProviderCreds>>,
+    /// When `Some`, REPLACES the entire `config.model_registry` array. Carries
+    /// each model's user-set `vision` flag (Settings → Advanced LLM → custom
+    /// model → "Supports vision"). Pass `Some(vec![])` to clear; `None` keeps it.
+    pub model_registry: Option<Vec<crate::openhuman::config::schema::ModelRegistryEntry>>,
     /// Id of the `cloud_providers` entry used when a workload routes to
     /// `"cloud"`. Empty string clears (factory falls back to OpenHuman).
     pub primary_cloud: Option<String>,
@@ -139,6 +143,24 @@ pub async fn apply_model_settings(
     if let Some(routes) = update.model_routes {
         config.model_routes = routes;
     }
+    if let Some(registry) = update.model_registry {
+        // Full replacement — the UI sends the canonical per-model registry
+        // (carrying each model's `vision` flag). Empty vec clears it.
+        log::debug!(
+            "[config] apply_model_settings: replacing model_registry ({} entries)",
+            registry.len()
+        );
+        // Normalize ids: `model_vision_enabled` matches the resolved model id
+        // exactly, so stray surrounding whitespace would silently disable vision
+        // for an otherwise valid model.
+        config.model_registry = registry
+            .into_iter()
+            .map(|mut entry| {
+                entry.id = entry.id.trim().to_string();
+                entry
+            })
+            .collect();
+    }
     if let Some(providers) = update.cloud_providers {
         use crate::openhuman::config::schema::cloud_providers::is_slug_reserved;
         let preserved: Vec<_> = config

src/openhuman/config/ops_tests.rs

@@ -588,6 +588,77 @@ async fn apply_model_settings_replaces_model_routes_when_some_and_keeps_when_non
     assert!(cfg.model_routes.is_empty());
 }
 
+#[tokio::test]
+async fn apply_model_settings_replaces_model_registry_when_some_and_keeps_when_none() {
+    // Per-model vision registry follows Some=replace / None=keep / empty=clear —
+    // this persists the "Supports vision" flag set in Settings → Advanced LLM.
+    use crate::openhuman::config::schema::ModelRegistryEntry;
+    let tmp = tempdir().unwrap();
+    let mut cfg = tmp_config(&tmp);
+
+    let set = ModelSettingsPatch {
+        model_registry: Some(vec![ModelRegistryEntry {
+            id: "my-llava".into(),
+            provider: "openai".into(),
+            cost_per_1m_output: 0.0,
+            vision: true,
+        }]),
+        ..Default::default()
+    };
+    let _ = apply_model_settings(&mut cfg, set).await.expect("set");
+    assert_eq!(cfg.model_registry.len(), 1);
+    assert!(cfg
+        .model_registry
+        .iter()
+        .any(|e| e.id == "my-llava" && e.vision));
+
+    // None — leave registry alone.
+    let _ = apply_model_settings(
+        &mut cfg,
+        ModelSettingsPatch {
+            api_url: Some("https://x.test/v1".into()),
+            ..Default::default()
+        },
+    )
+    .await
+    .expect("touch");
+    assert_eq!(cfg.model_registry.len(), 1);
+
+    // Empty vec — clear.
+    let _ = apply_model_settings(
+        &mut cfg,
+        ModelSettingsPatch {
+            model_registry: Some(vec![]),
+            ..Default::default()
+        },
+    )
+    .await
+    .expect("clear");
+    assert!(cfg.model_registry.is_empty());
+}
+
+#[tokio::test]
+async fn apply_model_settings_trims_model_registry_ids() {
+    // `model_vision_enabled` matches the resolved id exactly, so persisted ids
+    // must be trimmed or stray whitespace would silently disable vision.
+    use crate::openhuman::config::schema::ModelRegistryEntry;
+    let tmp = tempdir().unwrap();
+    let mut cfg = tmp_config(&tmp);
+
+    let set = ModelSettingsPatch {
+        model_registry: Some(vec![ModelRegistryEntry {
+            id: "  spaced-model  ".into(),
+            provider: "openai".into(),
+            cost_per_1m_output: 0.0,
+            vision: true,
+        }]),
+        ..Default::default()
+    };
+    let _ = apply_model_settings(&mut cfg, set).await.expect("set");
+    assert_eq!(cfg.model_registry.len(), 1);
+    assert_eq!(cfg.model_registry[0].id, "spaced-model");
+}
+
 #[tokio::test]
 async fn apply_model_settings_empty_strings_clear_optional_fields() {
     let tmp = tempdir().unwrap();

src/openhuman/config/schemas/controllers.rs

@@ -360,6 +360,10 @@ fn handle_update_model_settings(params: Map<String, Value>) -> ControllerFuture
                         .collect::<Result<Vec<_>, String>>()
                 })
                 .transpose()?,
+            // The config-domain RPC doesn't carry a model-registry payload — the
+            // per-model vision registry is updated via the inference-domain
+            // `inference_update_model_settings` path.
+            model_registry: None,
             primary_cloud: update.primary_cloud,
             chat_provider: update.chat_provider,
             reasoning_provider: update.reasoning_provider,

src/openhuman/inference/model_context.rs

@@ -137,6 +137,46 @@ pub fn context_window_for_model_with_local_fallback(
     None
 }
 
+/// Whether the model resolved for a chat hint/agent/profile accepts image input
+/// according to the **user-configured** vision flag in `config.model_registry`.
+///
+/// This is the per-model override that lets a user mark a **custom / BYOK** model
+/// as vision-capable (Settings → Advanced LLM → custom model → "Supports
+/// vision"). Managed-backend models already advertise vision via
+/// [`crate::openhuman::inference::provider::Provider::supports_vision`]; this flag
+/// covers OpenAI-compatible providers the backend can't introspect per-model.
+/// Returns `false` for models the user has not flagged.
+pub fn model_vision_enabled(model: &str, config: &crate::openhuman::config::Config) -> bool {
+    let normalized = model.trim();
+    if normalized.is_empty() {
+        return false;
+    }
+    let enabled = config
+        .model_registry
+        .iter()
+        .any(|entry| entry.id == normalized && entry.vision);
+    tracing::debug!(
+        model = normalized,
+        vision_enabled = enabled,
+        "[model_context] resolved user-configured vision flag"
+    );
+    enabled
+}
+
+/// Whether a resolved model accepts image input. The single predicate shared by
+/// the chat UI resolve and the server-side session/sub-agent gates.
+///
+/// - **Managed OpenHuman tiers** consult the hardcoded per-tier map
+///   ([`crate::openhuman::inference::provider::factory::oh_tier_supports_vision`]) —
+///   the remote backend does not advertise per-tier capability, so the core owns
+///   it. Currently every managed tier is `false`.
+/// - **Custom/BYOK models** consult the user-set `model_registry.vision` flag
+///   ([`model_vision_enabled`]).
+pub fn model_supports_vision(model: &str, config: &crate::openhuman::config::Config) -> bool {
+    crate::openhuman::inference::provider::factory::oh_tier_supports_vision(model)
+        || model_vision_enabled(model, config)
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -210,6 +250,50 @@ mod tests {
         assert_eq!(context_window_for_model("   "), None);
     }
 
+    #[test]
+    fn model_vision_enabled_reads_registry_only() {
+        use crate::openhuman::config::schema::ModelRegistryEntry;
+        let mut config = crate::openhuman::config::Config::default();
+        config.model_registry = vec![
+            ModelRegistryEntry {
+                id: "my-llava".into(),
+                provider: "openai".into(),
+                cost_per_1m_output: 0.0,
+                vision: true,
+            },
+            ModelRegistryEntry {
+                id: "text-only".into(),
+                provider: "openai".into(),
+                cost_per_1m_output: 0.0,
+                vision: false,
+            },
+        ];
+        assert!(model_vision_enabled("my-llava", &config));
+        assert!(!model_vision_enabled("text-only", &config));
+        assert!(!model_vision_enabled("unlisted", &config));
+        assert!(!model_vision_enabled("   ", &config));
+    }
+
+    #[test]
+    fn model_supports_vision_combines_tier_map_and_registry() {
+        use crate::openhuman::config::schema::ModelRegistryEntry;
+        let mut config = crate::openhuman::config::Config::default();
+        config.model_registry = vec![ModelRegistryEntry {
+            id: "my-llava".into(),
+            provider: "openai".into(),
+            cost_per_1m_output: 0.0,
+            vision: true,
+        }];
+        // Managed tiers are non-vision (the per-tier map is all `false` today).
+        assert!(!model_supports_vision("chat-v1", &config));
+        assert!(!model_supports_vision("reasoning-v1", &config));
+        assert!(!model_supports_vision("hint:chat", &config));
+        // BYOK model flagged in the registry is vision-capable.
+        assert!(model_supports_vision("my-llava", &config));
+        // Unlisted custom model is not.
+        assert!(!model_supports_vision("gpt-5", &config));
+    }
+
     #[test]
     fn o1_o3_segment_match_does_not_overmatch() {
         // Real OpenAI o1/o3 model ids must still resolve.

src/openhuman/inference/provider/factory.rs

@@ -172,6 +172,34 @@ pub(crate) fn is_known_openhuman_tier(model: &str) -> bool {
     )
 }
 
+/// Per-tier vision (image-input) capability for the managed OpenHuman backend.
+///
+/// The remote managed backend (`api.tinyhumans.ai`) does not advertise per-tier
+/// capabilities, so the core maintains this map itself. Accepts both the tier
+/// constants and their `hint:*` forms (callers may pass either pre- or
+/// post-resolution).
+///
+/// Every managed tier currently returns `false` — flip an individual arm to
+/// `true` once that tier is confirmed multimodal on the backend. This is the
+/// **only** place to change managed-model vision; BYOK/custom models are handled
+/// separately by the user-set `model_registry.vision` flag
+/// ([`crate::openhuman::inference::model_context::model_vision_enabled`]).
+pub(crate) fn oh_tier_supports_vision(model: &str) -> bool {
+    use crate::openhuman::config::{
+        MODEL_AGENTIC_V1, MODEL_CHAT_V1, MODEL_CODING_V1, MODEL_REASONING_QUICK_V1,
+        MODEL_REASONING_V1, MODEL_SUMMARIZATION_V1,
+    };
+    match model {
+        MODEL_CHAT_V1 | "hint:chat" => false,
+        MODEL_REASONING_V1 | "hint:reasoning" => false,
+        MODEL_REASONING_QUICK_V1 => false,
+        MODEL_AGENTIC_V1 | "hint:agentic" => false,
+        MODEL_CODING_V1 | "hint:coding" => false,
+        MODEL_SUMMARIZATION_V1 | "hint:summarization" => false,
+        _ => false,
+    }
+}
+
 /// Return the configured provider string for a named workload role.
 ///
 /// Empty / `"cloud"` resolves through BYOK fallback first for the three

src/openhuman/inference/provider/factory_tests.rs

@@ -867,6 +867,40 @@ fn invalid_models_fail() {
     assert!(!is_known_openhuman_tier("hint:"));
 }
 
+// ── oh_tier_supports_vision ──────────────────────────────────────────────────────
+
+#[test]
+fn no_managed_tier_is_vision_capable_yet() {
+    // Every managed tier (and its hint form) is non-vision until confirmed
+    // multimodal on the backend. Flip the corresponding arm in
+    // `oh_tier_supports_vision` to enable one.
+    for model in [
+        "reasoning-v1",
+        "chat-v1",
+        "agentic-v1",
+        "coding-v1",
+        "reasoning-quick-v1",
+        "summarization-v1",
+        "hint:reasoning",
+        "hint:chat",
+        "hint:agentic",
+        "hint:coding",
+        "hint:summarization",
+    ] {
+        assert!(
+            !oh_tier_supports_vision(model),
+            "expected managed tier '{model}' to be non-vision"
+        );
+    }
+}
+
+#[test]
+fn unknown_models_are_not_vision_capable() {
+    assert!(!oh_tier_supports_vision("gpt-5"));
+    assert!(!oh_tier_supports_vision("claude-opus-4-7"));
+    assert!(!oh_tier_supports_vision(""));
+}
+
 #[test]
 fn make_openhuman_backend_forwards_unknown_hint_verbatim() {
     // Unrecognised hint:* strings (e.g. hint:reaction for lightweight models)

src/openhuman/inference/schemas.rs

@@ -80,6 +80,8 @@ struct InferenceUpdateModelSettingsParams {
     default_temperature: Option<f64>,
     model_routes: Option<Vec<InferenceModelRouteUpdate>>,
     cloud_providers: Option<Vec<InferenceCloudProviderUpdate>>,
+    #[serde(default)]
+    model_registry: Option<Vec<crate::openhuman::config::schema::ModelRegistryEntry>>,
     primary_cloud: Option<String>,
     chat_provider: Option<String>,
     reasoning_provider: Option<String>,
@@ -255,7 +257,13 @@ pub fn schemas(function: &str) -> ControllerSchema {
             function: "resolve_model",
             description: "Resolve a model hint or tier name to the concrete model the provider router would use.",
             inputs: vec![required_string("hint", "Model hint (e.g. hint:reasoning) or tier name (e.g. reasoning-v1).")],
-            outputs: vec![json_output("model", "Resolved concrete model id.")],
+            outputs: vec![
+                json_output("model", "Resolved concrete model id."),
+                json_output(
+                    "vision",
+                    "Whether the resolved model accepts image input (vision-capable).",
+                ),
+            ],
         },
         "status" => ControllerSchema {
             namespace: "inference",
@@ -283,6 +291,7 @@ pub fn schemas(function: &str) -> ControllerSchema {
                 optional_f64("default_temperature", "Optional default temperature override."),
                 optional_json("model_routes", "Optional full replacement for legacy model routes."),
                 optional_json("cloud_providers", "Optional full replacement for configured cloud providers."),
+                optional_json("model_registry", "Optional full replacement for the per-model registry (carries each model's `vision` flag)."),
                 optional_string("primary_cloud", "Optional primary cloud provider id."),
                 optional_string("chat_provider", "Optional chat workload provider string."),
                 optional_string("reasoning_provider", "Optional reasoning workload provider string."),
@@ -557,8 +566,14 @@ fn handle_inference_resolve_model(params: Map<String, Value>) -> ControllerFutur
         let resolved = crate::openhuman::inference::provider::factory::resolve_model_for_hint(
             &p.hint, &config,
         );
+        // Whether the resolved model accepts image input — drives the chat UI's
+        // image-attachment affordance. Managed OpenHuman tiers consult the
+        // core-owned per-tier map (currently all `false`); custom/BYOK models are
+        // covered by the user's per-model `model_registry.vision` flag.
+        let vision =
+            crate::openhuman::inference::model_context::model_supports_vision(&resolved, &config);
         to_json(RpcOutcome::new(
-            serde_json::json!({ "model": resolved }),
+            serde_json::json!({ "model": resolved, "vision": vision }),
             vec![],
         ))
     })
@@ -676,6 +691,7 @@ fn handle_inference_update_model_settings(params: Map<String, Value>) -> Control
                         .collect::<Result<Vec<_>, String>>()
                 })
                 .transpose()?,
+            model_registry: update.model_registry,
             primary_cloud: update.primary_cloud,
             chat_provider: update.chat_provider,
             reasoning_provider: update.reasoning_provider,