merge: AI settings (coach timeout, local streaming, OCR provider) into v1.22

docs/api/openapi.yaml

@@ -14374,6 +14374,43 @@ components:
             - type: string
             - type: "null"
           description: "`...` + last 4 chars of the stored OpenAI key, or null."
+        responseTimeoutSeconds:
+          anyOf:
+            - type: integer
+              minimum: -9007199254740991
+              maximum: 9007199254740991
+            - type: "null"
+          description: v1.22 (#89) — per-user response timeout for AI generation, in seconds (10–600). Null = the generous
+            built-in default (~180 s). Mainly for slow local/self-hosted backends.
+        ocrEnabled:
+          type: boolean
+          description: v1.22 (#90) — whether a dedicated document-scan (Lab-OCR) provider is active. When false the OCR path uses
+            the main provider chain.
+        ocrProvider:
+          anyOf:
+            - type: string
+              enum:
+                - OPENAI
+                - ANTHROPIC
+                - LOCAL
+            - type: "null"
+          description: The dedicated document-scan provider, or null when unset.
+        ocrModel:
+          anyOf:
+            - type: string
+            - type: "null"
+        ocrBaseUrl:
+          anyOf:
+            - type: string
+            - type: "null"
+          description: Custom base URL (LOCAL OCR provider only); null otherwise.
+        hasOcrKey:
+          type: boolean
+        ocrKeyPreview:
+          anyOf:
+            - type: string
+            - type: "null"
+          description: "`...` + last 4 chars of the stored document-scan provider key, or null."
       required:
         - provider
         - model
@@ -14385,6 +14422,13 @@ components:
         - hasLocalKey
         - hasOpenaiKey
         - openaiKeyPreview
+        - responseTimeoutSeconds
+        - ocrEnabled
+        - ocrProvider
+        - ocrModel
+        - ocrBaseUrl
+        - hasOcrKey
+        - ocrKeyPreview
       additionalProperties: false
       description: The calling user's AI-provider configuration plus the effective availability (`aiAvailable` + `managedBy`).
         Reports key presence + a masked preview only — never a plaintext key.

messages/de.json

@@ -4494,7 +4494,24 @@
         "completeness": "{answered} von {total} beantwortet",
         "questionsTitle": "Der Coach hat Rückfragen",
         "questionsOpenCoach": "Im Coach beantworten"
-      }
+      },
+      "responseTimeoutHeading": "Antwort-Timeout",
+      "responseTimeoutBody": "Wie lange auf die Antwort der KI gewartet wird, bevor abgebrochen wird. Erhöhe den Wert für ein lokales oder selbst gehostetes Modell, das beim ersten Aufruf Zeit zum Laden braucht.",
+      "responseTimeoutLabel": "Antwort-Timeout (Sekunden)",
+      "responseTimeoutPlaceholder": "Standard (~180)",
+      "responseTimeoutInvalid": "Gib eine Zahl zwischen 10 und 600 Sekunden ein.",
+      "ocrHeading": "Dokumentenscan (Laborberichte)",
+      "ocrBody": "Verwende einen eigenen Anbieter zum Auslesen hochgeladener Laborberichte. Ist die Option aus, nutzt der Scan deinen Hauptanbieter.",
+      "ocrEnableLabel": "Eigenen Anbieter für den Dokumentenscan aktivieren",
+      "ocrProviderLabel": "Scan-Anbieter",
+      "ocrProviderOption": {
+        "OPENAI": "OpenAI (dein API-Key)",
+        "ANTHROPIC": "Anthropic (Claude)",
+        "LOCAL": "Lokales Modell (OpenAI-kompatibel)"
+      },
+      "ocrModelPlaceholder": "z. B. gpt-4o",
+      "ocrVisionHint": "Dieses Modell kann möglicherweise keine Bilder lesen. Wähle für den Dokumentenscan ein Modell mit Bildverständnis.",
+      "ocrKeyLabel": "API-Key"
     },
     "dashboard": {
       "units": {

messages/en.json

@@ -4494,7 +4494,24 @@
         "completeness": "{answered} of {total} answered",
         "questionsTitle": "The Coach has follow-up questions",
         "questionsOpenCoach": "Answer them in the Coach"
-      }
+      },
+      "responseTimeoutHeading": "Response timeout",
+      "responseTimeoutBody": "How long to wait for the AI to respond before giving up. Raise it for a local or self-hosted model that needs time to load on the first request.",
+      "responseTimeoutLabel": "Response timeout (seconds)",
+      "responseTimeoutPlaceholder": "Default (~180)",
+      "responseTimeoutInvalid": "Enter a number between 10 and 600 seconds.",
+      "ocrHeading": "Document scanning (lab reports)",
+      "ocrBody": "Use a dedicated provider for reading uploaded lab reports. When off, scanning uses your main provider.",
+      "ocrEnableLabel": "Enable a dedicated document-scanning provider",
+      "ocrProviderLabel": "Scanning provider",
+      "ocrProviderOption": {
+        "OPENAI": "OpenAI (your API key)",
+        "ANTHROPIC": "Anthropic (Claude)",
+        "LOCAL": "Local model (OpenAI-compatible)"
+      },
+      "ocrModelPlaceholder": "e.g. gpt-4o",
+      "ocrVisionHint": "This model may not be able to read images. Choose a vision-capable model for document scanning.",
+      "ocrKeyLabel": "API key"
     },
     "dashboard": {
       "units": {

messages/es.json

@@ -4436,7 +4436,24 @@
         "completeness": "{answered} de {total} respondidas",
         "questionsTitle": "El Coach tiene preguntas",
         "questionsOpenCoach": "Responderlas en el Coach"
-      }
+      },
+      "responseTimeoutHeading": "Tiempo de espera de la respuesta",
+      "responseTimeoutBody": "Cuánto esperar a que responda la IA antes de cancelar. Aumenta el valor para un modelo local o autoalojado que necesita tiempo para cargarse en la primera solicitud.",
+      "responseTimeoutLabel": "Tiempo de espera de la respuesta (segundos)",
+      "responseTimeoutPlaceholder": "Predeterminado (~180)",
+      "responseTimeoutInvalid": "Introduce un número entre 10 y 600 segundos.",
+      "ocrHeading": "Escaneo de documentos (informes de laboratorio)",
+      "ocrBody": "Usa un proveedor dedicado para leer los informes de laboratorio subidos. Si está desactivado, el escaneo usa tu proveedor principal.",
+      "ocrEnableLabel": "Activar un proveedor dedicado para el escaneo de documentos",
+      "ocrProviderLabel": "Proveedor de escaneo",
+      "ocrProviderOption": {
+        "OPENAI": "OpenAI (tu clave de API)",
+        "ANTHROPIC": "Anthropic (Claude)",
+        "LOCAL": "Modelo local (compatible con OpenAI)"
+      },
+      "ocrModelPlaceholder": "p. ej. gpt-4o",
+      "ocrVisionHint": "Puede que este modelo no pueda leer imágenes. Elige un modelo con capacidad de visión para el escaneo de documentos.",
+      "ocrKeyLabel": "Clave de API"
     },
     "dashboard": {
       "units": {

messages/fr.json

@@ -4436,7 +4436,24 @@
         "completeness": "{answered} sur {total} renseignés",
         "questionsTitle": "Le Coach a des questions",
         "questionsOpenCoach": "Y répondre dans le Coach"
-      }
+      },
+      "responseTimeoutHeading": "Délai de réponse",
+      "responseTimeoutBody": "Durée d'attente de la réponse de l'IA avant d'abandonner. Augmentez-la pour un modèle local ou auto-hébergé qui a besoin de temps pour se charger à la première requête.",
+      "responseTimeoutLabel": "Délai de réponse (secondes)",
+      "responseTimeoutPlaceholder": "Par défaut (~180)",
+      "responseTimeoutInvalid": "Saisissez un nombre entre 10 et 600 secondes.",
+      "ocrHeading": "Numérisation de documents (comptes rendus de laboratoire)",
+      "ocrBody": "Utilisez un fournisseur dédié pour lire les comptes rendus de laboratoire importés. Désactivé, la numérisation utilise votre fournisseur principal.",
+      "ocrEnableLabel": "Activer un fournisseur dédié à la numérisation de documents",
+      "ocrProviderLabel": "Fournisseur de numérisation",
+      "ocrProviderOption": {
+        "OPENAI": "OpenAI (votre clé API)",
+        "ANTHROPIC": "Anthropic (Claude)",
+        "LOCAL": "Modèle local (compatible OpenAI)"
+      },
+      "ocrModelPlaceholder": "p. ex. gpt-4o",
+      "ocrVisionHint": "Ce modèle ne peut peut-être pas lire les images. Choisissez un modèle capable de vision pour la numérisation de documents.",
+      "ocrKeyLabel": "Clé API"
     },
     "dashboard": {
       "units": {

messages/it.json

@@ -4436,7 +4436,24 @@
         "completeness": "{answered} su {total} compilati",
         "questionsTitle": "Il Coach ha alcune domande",
         "questionsOpenCoach": "Rispondi nel Coach"
-      }
+      },
+      "responseTimeoutHeading": "Timeout della risposta",
+      "responseTimeoutBody": "Quanto attendere la risposta dell'IA prima di annullare. Aumentalo per un modello locale o self-hosted che ha bisogno di tempo per caricarsi alla prima richiesta.",
+      "responseTimeoutLabel": "Timeout della risposta (secondi)",
+      "responseTimeoutPlaceholder": "Predefinito (~180)",
+      "responseTimeoutInvalid": "Inserisci un numero tra 10 e 600 secondi.",
+      "ocrHeading": "Scansione documenti (referti di laboratorio)",
+      "ocrBody": "Usa un provider dedicato per leggere i referti di laboratorio caricati. Se disattivato, la scansione usa il tuo provider principale.",
+      "ocrEnableLabel": "Attiva un provider dedicato per la scansione dei documenti",
+      "ocrProviderLabel": "Provider di scansione",
+      "ocrProviderOption": {
+        "OPENAI": "OpenAI (la tua chiave API)",
+        "ANTHROPIC": "Anthropic (Claude)",
+        "LOCAL": "Modello locale (compatibile con OpenAI)"
+      },
+      "ocrModelPlaceholder": "es. gpt-4o",
+      "ocrVisionHint": "Questo modello potrebbe non essere in grado di leggere le immagini. Scegli un modello con capacità di visione per la scansione dei documenti.",
+      "ocrKeyLabel": "Chiave API"
     },
     "dashboard": {
       "units": {

messages/pl.json

@@ -4436,7 +4436,24 @@
         "completeness": "Wypełnione: {answered} z {total}",
         "questionsTitle": "Coach ma pytania",
         "questionsOpenCoach": "Odpowiedz w Coachu"
-      }
+      },
+      "responseTimeoutHeading": "Limit czasu odpowiedzi",
+      "responseTimeoutBody": "Jak długo czekać na odpowiedź AI, zanim nastąpi przerwanie. Zwiększ wartość dla lokalnego lub samodzielnie hostowanego modelu, który potrzebuje czasu na załadowanie przy pierwszym żądaniu.",
+      "responseTimeoutLabel": "Limit czasu odpowiedzi (sekundy)",
+      "responseTimeoutPlaceholder": "Domyślnie (~180)",
+      "responseTimeoutInvalid": "Podaj liczbę od 10 do 600 sekund.",
+      "ocrHeading": "Skanowanie dokumentów (wyniki badań)",
+      "ocrBody": "Użyj dedykowanego dostawcy do odczytu przesłanych wyników badań. Gdy wyłączone, skanowanie używa Twojego głównego dostawcy.",
+      "ocrEnableLabel": "Włącz dedykowanego dostawcę do skanowania dokumentów",
+      "ocrProviderLabel": "Dostawca skanowania",
+      "ocrProviderOption": {
+        "OPENAI": "OpenAI (Twój klucz API)",
+        "ANTHROPIC": "Anthropic (Claude)",
+        "LOCAL": "Model lokalny (zgodny z OpenAI)"
+      },
+      "ocrModelPlaceholder": "np. gpt-4o",
+      "ocrVisionHint": "Ten model może nie odczytywać obrazów. Wybierz model obsługujący obraz do skanowania dokumentów.",
+      "ocrKeyLabel": "Klucz API"
     },
     "dashboard": {
       "units": {

prisma/migrations/0198_v122_ai_settings/migration.sql

@@ -0,0 +1,26 @@
+-- v1.22 (#89 + #90) — AI-settings wave.
+--
+-- Two additive, UI/DB-backed features on the per-user AI config (no new env
+-- vars — every value is configured in the app):
+--
+--   #89 ai_response_timeout_seconds — per-user response timeout for AI
+--       generation, in SECONDS. Threaded onto CompletionParams.timeoutMs
+--       (x1000). NULL = the generous built-in default (~180 s). Surfaced in
+--       Settings -> AI mainly for local/self-hosted servers where the first
+--       request can take >60 s while the model loads.
+--
+--   #90 dedicated document-scan (Lab-OCR) provider. OFF by default
+--       (ai_ocr_enabled = false) -> the OCR ingestion uses the main provider
+--       chain unchanged. When enabled, the OCR path resolves this dedicated
+--       provider/model/key instead. ai_ocr_key_encrypted is AES-256-GCM
+--       ciphertext like every other *_encrypted column (registered in the
+--       encrypted-columns registry + the rotation script).
+--
+-- All columns are nullable / defaulted, so existing rows need no backfill.
+ALTER TABLE "User"
+  ADD COLUMN "ai_response_timeout_seconds" INTEGER,
+  ADD COLUMN "ai_ocr_enabled" BOOLEAN NOT NULL DEFAULT false,
+  ADD COLUMN "ai_ocr_provider" TEXT,
+  ADD COLUMN "ai_ocr_model" TEXT,
+  ADD COLUMN "ai_ocr_base_url" TEXT,
+  ADD COLUMN "ai_ocr_key_encrypted" TEXT;

prisma/schema.prisma

@@ -254,6 +254,27 @@ model User {
   // behaviour (single-provider resolution via `aiProvider`).
   aiProviderChain Json? @map("ai_provider_chain")
 
+  // v1.22 (#89) — per-user response timeout for AI generation, in SECONDS.
+  // Threaded onto `CompletionParams.timeoutMs` (×1000). Null = the generous
+  // built-in default (~180 s). Surfaced in Settings → AI mainly for
+  // local/self-hosted servers (an MLX/exo backend can take >60 s on the
+  // first request while it loads the model), where the legacy 60 s client
+  // default timed the Coach out before the first token landed.
+  aiResponseTimeoutSeconds Int? @map("ai_response_timeout_seconds")
+
+  // v1.22 (#90) — dedicated document-scan (Lab-OCR) provider. OFF by default:
+  // the lab-report OCR path uses the main provider chain unchanged. When
+  // `aiOcrEnabled` is true the OCR ingestion resolves THIS provider/model/key
+  // instead, so a user can keep a cheap text Coach on one provider and route
+  // expensive vision OCR to a vision-capable one. The key is AES-256-GCM at
+  // rest like every other `*Encrypted` column (registered in
+  // `src/lib/crypto/encrypted-columns.ts` + the rotation script).
+  aiOcrEnabled      Boolean @default(false) @map("ai_ocr_enabled")
+  aiOcrProvider     String? @map("ai_ocr_provider") // "OPENAI" | "ANTHROPIC" | "LOCAL"
+  aiOcrModel        String? @map("ai_ocr_model")
+  aiOcrBaseUrl      String? @map("ai_ocr_base_url") // LOCAL OCR provider only
+  aiOcrKeyEncrypted String? @map("ai_ocr_key_encrypted")
+
   // Onboarding
   onboardingCompletedAt   DateTime? @map("onboarding_completed_at")
   // v1.4.25 W14b — last persisted onboarding wizard step.

scripts/rotate-encryption-key.ts

@@ -191,6 +191,7 @@ async function main() {
     "aiAnthropicKeyEncrypted",
     "aiLocalKeyEncrypted",
     "aiOpenaiKeyEncrypted",
+    "aiOcrKeyEncrypted",
     "insuranceNumberEncrypted",
   ];
   for (const field of userFields) {

src/app/api/insights/chat/__tests__/route-snapshot-once.test.ts

@@ -50,12 +50,12 @@ vi.mock("@/lib/i18n/server-locale", () => ({
   resolveServerLocale: vi.fn(async () => "en"),
 }));
 
-const { runRawCompletionWithFallback } = vi.hoisted(() => ({
-  runRawCompletionWithFallback: vi.fn(),
+const { runStreamingRawCompletionWithFallback } = vi.hoisted(() => ({
+  runStreamingRawCompletionWithFallback: vi.fn(),
 }));
 vi.mock("@/lib/ai/provider-runner", () => ({
   AllProvidersFailedError: class extends Error {},
-  runRawCompletionWithFallback,
+  runStreamingRawCompletionWithFallback,
 }));
 vi.mock("@/lib/ai/provider", () => ({
   // v1.20.0 (F1) — pin a no-tools provider so this suite exercises the legacy
@@ -138,17 +138,18 @@ vi.mock("@/lib/validations/coach-prefs", () => ({
   parseCoachPrefs: vi.fn(() => ({ defaultWindow: undefined })),
   DEFAULT_REMINDER_SUGGESTION_PREFS: {},
 }));
+// v1.22 (#89) — the provider call now runs INSIDE the stream producer. Capture
+// the producer promise so the test can await it before asserting on call args.
+const sse = vi.hoisted(() => ({ done: Promise.resolve() as Promise<unknown> }));
 vi.mock("@/lib/sse/create-stream", () => ({
-  // Drain the producer synchronously so the handler completes; we only care
-  // about the provider call args, not the wire frames.
   createSseStream: (
     producer: (c: {
       signal: { aborted: boolean };
       enqueue: () => void;
     }) => void | Promise<void>,
   ) => {
-    void Promise.resolve(
-      producer({ signal: { aborted: true }, enqueue: () => {} }),
+    sse.done = Promise.resolve(
+      producer({ signal: { aborted: false }, enqueue: () => {} }),
     );
     return new ReadableStream();
   },
@@ -158,6 +159,12 @@ import { POST } from "../route";
 
 const post = POST as unknown as (req: Request) => Promise<Response>;
 
+async function postAndDrain(body: Record<string, unknown>): Promise<Response> {
+  const res = await post(chatReq(body));
+  await sse.done;
+  return res;
+}
+
 function chatReq(body: Record<string, unknown>): Request {
   return new Request("http://localhost/api/insights/chat", {
     method: "POST",
@@ -167,7 +174,8 @@ function chatReq(body: Record<string, unknown>): Request {
 }
 
 function lastUserPrompt(): string {
-  const calls = runRawCompletionWithFallback.mock.calls as unknown as Array<
+  const calls = runStreamingRawCompletionWithFallback.mock
+    .calls as unknown as Array<
     [{ params: { messages: Array<{ role: string; content: string }> } }]
   >;
   const last = calls[calls.length - 1];
@@ -178,14 +186,14 @@ function lastUserPrompt(): string {
 describe("coach chat — snapshot sent once per conversation (C4)", () => {
   beforeEach(() => {
     vi.clearAllMocks();
-    runRawCompletionWithFallback.mockResolvedValue({
+    runStreamingRawCompletionWithFallback.mockResolvedValue({
       result: { content: "Your BP looks stable.", tokensUsed: 42, model: "m" },
       workingProvider: { providerType: "admin-openai" },
     });
   });
 
   it("ships the full SNAPSHOT + grounding on the first turn", async () => {
-    await post(chatReq({ message: "How is my BP?" }));
+    await postAndDrain({ message: "How is my BP?" });
     const prompt = lastUserPrompt();
     expect(prompt).toContain(SNAPSHOT_JSON);
     expect(prompt).toContain(GROUNDING);
@@ -201,7 +209,7 @@ describe("coach chat — snapshot sent once per conversation (C4)", () => {
       ],
     });
 
-    await post(chatReq({ conversationId: "c1", message: "High?" }));
+    await postAndDrain({ conversationId: "c1", message: "High?" });
     const prompt = lastUserPrompt();
     // The expensive figures + grounding must be gone …
     expect(prompt).not.toContain(SNAPSHOT_JSON);

src/app/api/insights/chat/__tests__/route-tool-mode.test.ts

@@ -52,12 +52,12 @@ vi.mock("@/lib/i18n/server-locale", () => ({
   resolveServerLocale: vi.fn(async () => "en"),
 }));
 
-const { runRawCompletionWithFallback } = vi.hoisted(() => ({
-  runRawCompletionWithFallback: vi.fn(),
+const { runStreamingRawCompletionWithFallback } = vi.hoisted(() => ({
+  runStreamingRawCompletionWithFallback: vi.fn(),
 }));
 vi.mock("@/lib/ai/provider-runner", () => ({
   AllProvidersFailedError: class extends Error {},
-  runRawCompletionWithFallback,
+  runStreamingRawCompletionWithFallback,
 }));
 
 const { resolveProviderChain } = vi.hoisted(() => ({
@@ -194,6 +194,10 @@ const { parseKeyValuesSentinel } = await import("@/lib/ai/coach/keyvalues");
 const { parseSuggestReminder } =
   await import("@/lib/ai/coach/suggest-reminder");
 
+// v1.22 (#89) — the provider call + guards + persistence now run INSIDE the
+// stream producer (heartbeat-fronted). The mock captures the producer's promise
+// so tests can await it after `post()` before asserting on persisted effects.
+const sse = vi.hoisted(() => ({ done: Promise.resolve() as Promise<unknown> }));
 vi.mock("@/lib/sse/create-stream", () => ({
   createSseStream: (
     producer: (c: {
@@ -201,8 +205,9 @@ vi.mock("@/lib/sse/create-stream", () => ({
       enqueue: () => void;
     }) => void | Promise<void>,
   ) => {
-    void Promise.resolve(
-      producer({ signal: { aborted: true }, enqueue: () => {} }),
+    // aborted:false so the producer streams to completion (persist + frames).
+    sse.done = Promise.resolve(
+      producer({ signal: { aborted: false }, enqueue: () => {} }),
     );
     return new ReadableStream();
   },
@@ -212,6 +217,13 @@ import { POST } from "../route";
 
 const post = POST as unknown as (req: Request) => Promise<Response>;
 
+/** Run a POST and wait for the in-stream producer to finish. */
+async function postAndDrain(body: Record<string, unknown>): Promise<Response> {
+  const res = await post(chatReq(body));
+  await sse.done;
+  return res;
+}
+
 function chatReq(body: Record<string, unknown>): Request {
   return new Request("http://localhost/api/insights/chat", {
     method: "POST",
@@ -233,10 +245,10 @@ describe("coach chat — tool-mode routing (F1)", () => {
     resolveProviderChain.mockResolvedValue([
       { providerType: "anthropic", instance: {} }, // no supportsTools=false ⇒ tools on
     ]);
-    await post(chatReq({ message: "How is my BP?" }));
+    await postAndDrain({ message: "How is my BP?" });
 
     expect(runCoachToolLoop).toHaveBeenCalledTimes(1);
-    expect(runRawCompletionWithFallback).not.toHaveBeenCalled();
+    expect(runStreamingRawCompletionWithFallback).not.toHaveBeenCalled();
     // The base context carries the inventory, NOT the snapshot figures.
     const loopArgs = (runCoachToolLoop.mock.calls[0] as unknown[])[0] as {
       messages: Array<{ content: string }>;
@@ -277,6 +289,7 @@ describe("coach chat — tool-mode routing (F1)", () => {
     );
 
     const res = await post(chatReq({ message: "How is my BP?" }));
+    await sse.done;
 
     // No throw → no 500. The provider-error frame streams over a 200.
     expect(res.status).toBe(200);
@@ -293,7 +306,7 @@ describe("coach chat — tool-mode routing (F1)", () => {
     resolveProviderChain.mockResolvedValue([
       { providerType: "anthropic", instance: {} },
     ]);
-    await post(chatReq({ message: "How is my BP?" }));
+    await postAndDrain({ message: "How is my BP?" });
     const calls = (appendMessage as ReturnType<typeof vi.fn>).mock.calls;
     const assistantCall = calls.find(
       (c) => (c[0] as { role: string }).role === "assistant",
@@ -338,7 +351,7 @@ describe("coach chat — tool-mode routing (F1)", () => {
         ],
       }),
     );
-    await post(chatReq({ message: "How is my BP?" }));
+    await postAndDrain({ message: "How is my BP?" });
     const calls = (appendMessage as ReturnType<typeof vi.fn>).mock.calls;
     const assistantCall = calls.find(
       (c) => (c[0] as { role: string }).role === "assistant",
@@ -353,14 +366,14 @@ describe("coach chat — tool-mode routing (F1)", () => {
       { providerType: "anthropic", instance: {} },
       { providerType: "local", instance: { supportsTools: false } },
     ]);
-    runRawCompletionWithFallback.mockResolvedValue({
+    runStreamingRawCompletionWithFallback.mockResolvedValue({
       result: { content: "Your BP is steady.", tokensUsed: 42, model: "m" },
       workingProvider: { providerType: "anthropic" },
     });
-    await post(chatReq({ message: "How is my BP?" }));
+    await postAndDrain({ message: "How is my BP?" });
 
     expect(runCoachToolLoop).not.toHaveBeenCalled();
-    expect(runRawCompletionWithFallback).toHaveBeenCalledTimes(1);
+    expect(runStreamingRawCompletionWithFallback).toHaveBeenCalledTimes(1);
     // Single-round budget reservation in the fallback path; the 4th arg is the
     // provider-aware daily cap (F1).
     expect(reserveBudget).toHaveBeenCalledWith(
@@ -371,7 +384,7 @@ describe("coach chat — tool-mode routing (F1)", () => {
     );
     // The legacy path ships the snapshot figures in the user turn.
     const params = (
-      (runRawCompletionWithFallback.mock.calls[0] as unknown[])[0] as {
+      (runStreamingRawCompletionWithFallback.mock.calls[0] as unknown[])[0] as {
         params: { messages: Array<{ content: string }> };
       }
     ).params;

src/app/api/user/ai-provider/__tests__/route.test.ts

@@ -18,7 +18,7 @@ vi.mock("@/lib/api-handler", () => ({
 
 vi.mock("@/lib/db", () => ({
   prisma: {
-    user: { findUnique: vi.fn() },
+    user: { findUnique: vi.fn(), update: vi.fn() },
   },
 }));
 
@@ -35,14 +35,26 @@ vi.mock("@/lib/validations/notifications", () => ({
   isPublicUrl: vi.fn(() => true),
 }));
 
+vi.mock("@/lib/ai/local-host-allowlist", () => ({
+  isLocalAiHostAllowed: vi.fn(() => false),
+}));
+
 vi.mock("@/lib/logging/context", () => ({
   annotate: vi.fn(),
 }));
 
-import { GET } from "../route";
+import { GET, PATCH } from "../route";
 import { prisma } from "@/lib/db";
 import { resolveProviderAvailability } from "@/lib/ai/provider";
 
+function patchRequest(body: unknown): Request {
+  return new Request("http://localhost/api/user/ai-provider", {
+    method: "PATCH",
+    headers: { "content-type": "application/json" },
+    body: JSON.stringify(body),
+  });
+}
+
 interface AiProviderResponse {
   data: {
     provider: string | null;
@@ -126,4 +138,101 @@ describe("GET /api/user/ai-provider availability", () => {
     expect(body.data.aiAvailable).toBe(false);
     expect(body.data.managedBy).toBeNull();
   });
+
+  it("returns the v1.22 timeout + OCR fields", async () => {
+    vi.mocked(prisma.user.findUnique).mockResolvedValue({
+      aiProvider: "OPENAI",
+      aiModel: "gpt-4o",
+      aiBaseUrl: null,
+      aiAnthropicKeyEncrypted: null,
+      aiLocalKeyEncrypted: null,
+      aiOpenaiKeyEncrypted: null,
+      aiResponseTimeoutSeconds: 240,
+      aiOcrEnabled: true,
+      aiOcrProvider: "ANTHROPIC",
+      aiOcrModel: "claude-sonnet-4-6",
+      aiOcrBaseUrl: null,
+      aiOcrKeyEncrypted: "enc-ocr-1234",
+    } as never);
+    vi.mocked(resolveProviderAvailability).mockResolvedValue({
+      aiAvailable: true,
+      managedBy: "user",
+    });
+
+    const res = await (GET as () => Promise<Response>)();
+    const body = (await res.json()) as {
+      data: {
+        responseTimeoutSeconds: number | null;
+        ocrEnabled: boolean;
+        ocrProvider: string | null;
+        hasOcrKey: boolean;
+        ocrKeyPreview: string | null;
+      };
+    };
+
+    expect(body.data.responseTimeoutSeconds).toBe(240);
+    expect(body.data.ocrEnabled).toBe(true);
+    expect(body.data.ocrProvider).toBe("ANTHROPIC");
+    expect(body.data.hasOcrKey).toBe(true);
+    // Preview is `...` + last 4 of the decrypted value (decrypt → `dec:...`).
+    expect(body.data.ocrKeyPreview).toMatch(/^\.\.\./);
+  });
+});
+
+describe("PATCH /api/user/ai-provider — v1.22 fields", () => {
+  const patch = PATCH as (req: Request) => Promise<Response>;
+
+  it("persists a valid response timeout", async () => {
+    const res = await patch(patchRequest({ responseTimeoutSeconds: 120 }));
+    expect(res.status).toBe(200);
+    expect(vi.mocked(prisma.user.update)).toHaveBeenCalledWith(
+      expect.objectContaining({
+        data: expect.objectContaining({ aiResponseTimeoutSeconds: 120 }),
+      }),
+    );
+  });
+
+  it("rejects an out-of-range response timeout", async () => {
+    await expect(
+      patch(patchRequest({ responseTimeoutSeconds: 5 })),
+    ).rejects.toThrow(/between 10 and 600/);
+    expect(vi.mocked(prisma.user.update)).not.toHaveBeenCalled();
+  });
+
+  it("clears the timeout when null is sent", async () => {
+    await patch(patchRequest({ responseTimeoutSeconds: null }));
+    expect(vi.mocked(prisma.user.update)).toHaveBeenCalledWith(
+      expect.objectContaining({
+        data: expect.objectContaining({ aiResponseTimeoutSeconds: null }),
+      }),
+    );
+  });
+
+  it("persists OCR provider config and encrypts the key", async () => {
+    const res = await patch(
+      patchRequest({
+        ocrEnabled: true,
+        ocrProvider: "OPENAI",
+        ocrModel: "gpt-4o",
+        ocrKey: "sk-ocr-secret",
+      }),
+    );
+    expect(res.status).toBe(200);
+    expect(vi.mocked(prisma.user.update)).toHaveBeenCalledWith(
+      expect.objectContaining({
+        data: expect.objectContaining({
+          aiOcrEnabled: true,
+          aiOcrProvider: "OPENAI",
+          aiOcrModel: "gpt-4o",
+          aiOcrKeyEncrypted: "enc:sk-ocr-secret",
+        }),
+      }),
+    );
+  });
+
+  it("rejects an invalid OCR provider", async () => {
+    await expect(
+      patch(patchRequest({ ocrProvider: "CHATGPT_OAUTH" })),
+    ).rejects.toThrow(/Invalid OCR provider/);
+  });
 });

src/app/api/user/ai-provider/route.ts

@@ -11,6 +11,9 @@ import { annotate } from "@/lib/logging/context";
 export const dynamic = "force-dynamic";
 
 const ALLOWED = new Set(["OPENAI", "ANTHROPIC", "LOCAL", "CHATGPT_OAUTH"]);
+// v1.22 (#90) — the dedicated document-scan provider has no OAuth path, so the
+// allowlist is the three keyed/self-hosted providers only (no CHATGPT_OAUTH).
+const OCR_PROVIDERS = new Set(["OPENAI", "ANTHROPIC", "LOCAL"]);
 
 export const GET = apiHandler(async () => {
   const { user } = await requireAuth();
@@ -25,6 +28,13 @@ export const GET = apiHandler(async () => {
       aiAnthropicKeyEncrypted: true,
       aiLocalKeyEncrypted: true,
       aiOpenaiKeyEncrypted: true,
+      // v1.22 (#89 + #90)
+      aiResponseTimeoutSeconds: true,
+      aiOcrEnabled: true,
+      aiOcrProvider: true,
+      aiOcrModel: true,
+      aiOcrBaseUrl: true,
+      aiOcrKeyEncrypted: true,
     },
   });
 
@@ -50,6 +60,18 @@ export const GET = apiHandler(async () => {
     openaiKeyPreview: u?.aiOpenaiKeyEncrypted
       ? `...${decrypt(u.aiOpenaiKeyEncrypted).slice(-4)}`
       : null,
+    // v1.22 (#89) — per-user response timeout, in seconds (null = default).
+    responseTimeoutSeconds: u?.aiResponseTimeoutSeconds ?? null,
+    // v1.22 (#90) — dedicated document-scan (Lab-OCR) provider config. Key
+    // material is presence + 4-char preview only, like every other provider.
+    ocrEnabled: Boolean(u?.aiOcrEnabled),
+    ocrProvider: u?.aiOcrProvider ?? null,
+    ocrModel: u?.aiOcrModel ?? null,
+    ocrBaseUrl: u?.aiOcrBaseUrl ?? null,
+    hasOcrKey: Boolean(u?.aiOcrKeyEncrypted),
+    ocrKeyPreview: u?.aiOcrKeyEncrypted
+      ? `...${decrypt(u.aiOcrKeyEncrypted).slice(-4)}`
+      : null,
   });
 });
 
@@ -132,6 +154,87 @@ export const PATCH = apiHandler(async (request: NextRequest) => {
     }
   }
 
+  // ── v1.22 (#89) — response timeout (seconds) ──────────────────
+  if (body.responseTimeoutSeconds !== undefined) {
+    if (body.responseTimeoutSeconds === null) {
+      updates.aiResponseTimeoutSeconds = null;
+    } else if (
+      typeof body.responseTimeoutSeconds === "number" &&
+      Number.isInteger(body.responseTimeoutSeconds)
+    ) {
+      // Sane bounds: 10 s floor (a stricter value just trips the timeout) and
+      // 600 s ceiling (a generous cap for a slow self-hosted backend).
+      if (
+        body.responseTimeoutSeconds < 10 ||
+        body.responseTimeoutSeconds > 600
+      ) {
+        throw new HttpError(
+          422,
+          "Response timeout must be between 10 and 600 seconds",
+        );
+      }
+      updates.aiResponseTimeoutSeconds = body.responseTimeoutSeconds;
+    } else {
+      throw new HttpError(422, "Invalid response timeout");
+    }
+  }
+
+  // ── v1.22 (#90) — dedicated document-scan (Lab-OCR) provider ──
+  if (body.ocrEnabled !== undefined) {
+    if (typeof body.ocrEnabled === "boolean") {
+      updates.aiOcrEnabled = body.ocrEnabled;
+    } else {
+      throw new HttpError(422, "Invalid ocrEnabled");
+    }
+  }
+
+  if (body.ocrProvider !== undefined) {
+    if (body.ocrProvider === null || body.ocrProvider === "") {
+      updates.aiOcrProvider = null;
+    } else if (
+      typeof body.ocrProvider === "string" &&
+      OCR_PROVIDERS.has(body.ocrProvider)
+    ) {
+      updates.aiOcrProvider = body.ocrProvider;
+    } else {
+      throw new HttpError(422, "Invalid OCR provider");
+    }
+  }
+
+  if (body.ocrModel !== undefined) {
+    if (body.ocrModel === null || body.ocrModel === "") {
+      updates.aiOcrModel = null;
+    } else if (typeof body.ocrModel === "string") {
+      updates.aiOcrModel = body.ocrModel.trim() || null;
+    }
+  }
+
+  if (body.ocrBaseUrl !== undefined) {
+    if (body.ocrBaseUrl === null || body.ocrBaseUrl === "") {
+      updates.aiOcrBaseUrl = null;
+    } else if (typeof body.ocrBaseUrl === "string") {
+      const trimmed = body.ocrBaseUrl.trim();
+      // Same SSRF guard as the main LOCAL base URL: reject private/internal
+      // hosts unless the operator opted in via ALLOW_LOCAL_AI_PRIVATE_HOSTS.
+      const allowPrivate = isLocalAiHostAllowed(trimmed);
+      if (!allowPrivate && !isPublicUrl(trimmed)) {
+        return apiError(
+          "OCR base URL points to an internal/private host. Ops must allow it on this instance via ALLOW_LOCAL_AI_PRIVATE_HOSTS.",
+          422,
+        );
+      }
+      updates.aiOcrBaseUrl = trimmed;
+    }
+  }
+
+  if (body.ocrKey !== undefined) {
+    if (body.ocrKey === null || body.ocrKey === "") {
+      updates.aiOcrKeyEncrypted = null;
+    } else if (typeof body.ocrKey === "string") {
+      updates.aiOcrKeyEncrypted = encrypt(body.ocrKey.trim());
+    }
+  }
+
   if (Object.keys(updates).length === 0) {
     throw new HttpError(422, "No valid fields");
   }

src/components/settings/ai/ai-insights-card.tsx

@@ -31,9 +31,11 @@ import { queryKeys } from "@/lib/query-keys";
 import { AdminOpenAIProviderForm } from "./admin-openai-provider-form";
 import { AnthropicProviderForm } from "./anthropic-provider-form";
 import { CodexProviderForm } from "./codex-provider-form";
+import { DocumentScanCard } from "./document-scan-card";
 import { FallbackChainCard } from "./fallback-chain-card";
 import { LocalProviderForm } from "./local-provider-form";
 import { OpenAIProviderForm } from "./openai-provider-form";
+import { ResponseTimeoutCard } from "./response-timeout-card";
 import { RuntimeActionsRow } from "./runtime-actions-row";
 import {
   PROVIDER_TYPES,
@@ -171,6 +173,13 @@ export function AiInsightsCard({
           onSelect={pickProvider}
         />
 
+        {/* v1.22 (#89) — per-user response timeout (mainly for slow
+            local/self-hosted backends). */}
+        <ResponseTimeoutCard userProvider={userProvider} />
+
+        {/* v1.22 (#90) — dedicated document-scan (Lab-OCR) provider. */}
+        <DocumentScanCard userProvider={userProvider} />
+
         <RuntimeActionsRow
           provider={selectedProvider}
           userProvider={userProvider}

src/components/settings/ai/document-scan-card.tsx

@@ -0,0 +1,212 @@
+"use client";
+
+/* ────────────────────────────────────────────────────────────────
+ * v1.22 (#90) — Document scanning (Lab reports).
+ *
+ * Opt-in dedicated provider for the Lab-OCR ingestion path. OFF by default:
+ * scanning uses the main provider chain. When enabled, document/lab-report
+ * processing routes to THIS provider/model/key instead — so a user can keep a
+ * cheap text Coach on one provider and send expensive vision OCR to a
+ * vision-capable one. Surfaces a hint when the chosen model can't read images.
+ * ──────────────────────────────────────────────────────────────── */
+
+import { useState } from "react";
+import { useMutation, useQueryClient } from "@tanstack/react-query";
+import { Loader2, Save } from "lucide-react";
+
+import { Button } from "@/components/ui/button";
+import { Input } from "@/components/ui/input";
+import { Label } from "@/components/ui/label";
+import { NativeSelect } from "@/components/ui/native-select";
+import { PasswordInput } from "@/components/ui/password-input";
+import { Switch } from "@/components/ui/switch";
+import { apiPatch } from "@/lib/api/api-fetch";
+import {
+  supportsVisionForConfig,
+  type VisionProviderType,
+} from "@/lib/ai/vision-capability";
+import { useTranslations } from "@/lib/i18n/context";
+import { queryKeys } from "@/lib/query-keys";
+
+import { OCR_PROVIDER_TYPES, type UserAIProvider } from "./shared";
+
+/** Map the OCR provider enum to the vision-capability tag. */
+function toVisionType(provider: string): VisionProviderType {
+  switch (provider) {
+    case "ANTHROPIC":
+      return "anthropic";
+    case "OPENAI":
+      return "openai";
+    case "LOCAL":
+      return "local";
+    default:
+      return "none";
+  }
+}
+
+export function DocumentScanCard({
+  userProvider,
+}: {
+  userProvider: UserAIProvider | null | undefined;
+}) {
+  const { t } = useTranslations();
+  const queryClient = useQueryClient();
+
+  const [enabled, setEnabled] = useState(false);
+  const [provider, setProvider] = useState<string>("OPENAI");
+  const [model, setModel] = useState("");
+  const [baseUrl, setBaseUrl] = useState("");
+  const [apiKey, setApiKey] = useState("");
+  const [msg, setMsg] = useState<string | null>(null);
+  const [ok, setOk] = useState(false);
+
+  const seededKey =
+    userProvider != null
+      ? `${userProvider.ocrEnabled}|${userProvider.ocrProvider ?? ""}|${userProvider.ocrModel ?? ""}|${userProvider.ocrBaseUrl ?? ""}`
+      : null;
+  const [previousSeed, setPreviousSeed] = useState<string | null>(null);
+  if (seededKey != null && seededKey !== previousSeed) {
+    setPreviousSeed(seededKey);
+    setEnabled(userProvider?.ocrEnabled ?? false);
+    setProvider(userProvider?.ocrProvider ?? "OPENAI");
+    setModel(userProvider?.ocrModel ?? "");
+    setBaseUrl(userProvider?.ocrBaseUrl ?? "");
+  }
+
+  // Vision-capability hint: warn when a model is chosen that can't read images
+  // (local is trust-by-default; the warning surfaces only for cloud providers).
+  const visionOk =
+    !enabled ||
+    model.trim() === "" ||
+    supportsVisionForConfig(toVisionType(provider), model.trim());
+
+  const saveMutation = useMutation({
+    mutationFn: async () => {
+      const body: Record<string, unknown> = {
+        ocrEnabled: enabled,
+        ocrProvider: provider,
+        ocrModel: model.trim() || null,
+        ocrBaseUrl: provider === "LOCAL" ? baseUrl.trim() || null : null,
+      };
+      if (apiKey.trim()) body.ocrKey = apiKey.trim();
+      await apiPatch("/api/user/ai-provider", body);
+    },
+    onSuccess: () => {
+      setOk(true);
+      setMsg(t("settings.ai.saved"));
+      setApiKey("");
+      queryClient.invalidateQueries({ queryKey: queryKeys.userAiProvider() });
+    },
+    onError: (e) => {
+      setOk(false);
+      setMsg(e instanceof Error ? e.message : t("settings.ai.errorGeneric"));
+    },
+  });
+
+  return (
+    <div className="bg-muted/50 space-y-4 rounded-lg p-4">
+      <div className="flex items-start justify-between gap-3">
+        <div>
+          <p className="text-sm font-medium">{t("settings.ai.ocrHeading")}</p>
+          <p className="text-muted-foreground text-xs">
+            {t("settings.ai.ocrBody")}
+          </p>
+        </div>
+        <Switch
+          checked={enabled}
+          onCheckedChange={setEnabled}
+          aria-label={t("settings.ai.ocrEnableLabel")}
+          data-testid="ai-ocr-enable"
+        />
+      </div>
+
+      {enabled && (
+        <div className="space-y-4">
+          <div>
+            <Label htmlFor="ai-ocr-provider">
+              {t("settings.ai.ocrProviderLabel")}
+            </Label>
+            <NativeSelect
+              id="ai-ocr-provider"
+              value={provider}
+              onChange={(e) => setProvider(e.target.value)}
+              className="mt-1 sm:max-w-md"
+            >
+              {OCR_PROVIDER_TYPES.map((p) => (
+                <option key={p} value={p}>
+                  {t(`settings.ai.ocrProviderOption.${p}` as const)}
+                </option>
+              ))}
+            </NativeSelect>
+          </div>
+
+          <div>
+            <Label htmlFor="ai-ocr-model">{t("settings.ai.modelLabel")}</Label>
+            <Input
+              id="ai-ocr-model"
+              value={model}
+              onChange={(e) => setModel(e.target.value)}
+              placeholder={t("settings.ai.ocrModelPlaceholder")}
+              className="mt-1"
+            />
+            {!visionOk && (
+              <p className="text-warning mt-1 text-xs">
+                {t("settings.ai.ocrVisionHint")}
+              </p>
+            )}
+          </div>
+
+          {provider === "LOCAL" && (
+            <div>
+              <Label htmlFor="ai-ocr-base-url">
+                {t("settings.ai.baseUrlLabel")}
+              </Label>
+              <Input
+                id="ai-ocr-base-url"
+                value={baseUrl}
+                onChange={(e) => setBaseUrl(e.target.value)}
+                placeholder="http://localhost:11434/v1"
+                className="mt-1"
+              />
+            </div>
+          )}
+
+          <div>
+            <Label htmlFor="ai-ocr-key">{t("settings.ai.ocrKeyLabel")}</Label>
+            <PasswordInput
+              id="ai-ocr-key"
+              value={apiKey}
+              onChange={(e) => setApiKey(e.target.value)}
+              placeholder={
+                userProvider?.hasOcrKey ? t("settings.ai.savedShort") : ""
+              }
+              className="mt-1"
+            />
+          </div>
+        </div>
+      )}
+
+      <div>
+        <Button
+          size="sm"
+          className="min-h-11 sm:min-h-9"
+          onClick={() => saveMutation.mutate()}
+          disabled={saveMutation.isPending}
+        >
+          {saveMutation.isPending ? (
+            <Loader2 className="h-4 w-4 animate-spin motion-reduce:animate-none" />
+          ) : (
+            <Save className="h-4 w-4" />
+          )}
+          {t("settings.ai.saveCta")}
+        </Button>
+      </div>
+
+      {msg && (
+        <p className={`text-xs ${ok ? "text-success" : "text-destructive"}`}>
+          {msg}
+        </p>
+      )}
+    </div>
+  );
+}

src/components/settings/ai/response-timeout-card.tsx

@@ -0,0 +1,126 @@
+"use client";
+
+/* ────────────────────────────────────────────────────────────────
+ * v1.22 (#89) — Response timeout (seconds).
+ *
+ * A per-user upstream timeout for AI generation, threaded onto the provider
+ * call (`CompletionParams.timeoutMs`). Surfaced here mainly for local /
+ * self-hosted backends: an MLX/exo server can take >60 s on the first request
+ * while it loads the model, which the legacy 60 s default timed out before the
+ * first token landed. Empty = the generous built-in default (~180 s).
+ * ──────────────────────────────────────────────────────────────── */
+
+import { useState } from "react";
+import { useMutation, useQueryClient } from "@tanstack/react-query";
+import { Loader2, Save } from "lucide-react";
+
+import { Button } from "@/components/ui/button";
+import { Input } from "@/components/ui/input";
+import { Label } from "@/components/ui/label";
+import { apiPatch } from "@/lib/api/api-fetch";
+import { useTranslations } from "@/lib/i18n/context";
+import { queryKeys } from "@/lib/query-keys";
+
+import type { UserAIProvider } from "./shared";
+
+export function ResponseTimeoutCard({
+  userProvider,
+}: {
+  userProvider: UserAIProvider | null | undefined;
+}) {
+  const { t } = useTranslations();
+  const queryClient = useQueryClient();
+
+  const [value, setValue] = useState("");
+  const [msg, setMsg] = useState<string | null>(null);
+  const [ok, setOk] = useState(false);
+
+  // Seed from the persisted value once it arrives (seed-on-data pattern).
+  const seededKey =
+    userProvider != null
+      ? `${userProvider.responseTimeoutSeconds ?? ""}`
+      : null;
+  const [previousSeed, setPreviousSeed] = useState<string | null>(null);
+  if (seededKey != null && seededKey !== previousSeed) {
+    setPreviousSeed(seededKey);
+    setValue(
+      userProvider?.responseTimeoutSeconds != null
+        ? String(userProvider.responseTimeoutSeconds)
+        : "",
+    );
+  }
+
+  const saveMutation = useMutation({
+    mutationFn: async () => {
+      const trimmed = value.trim();
+      const responseTimeoutSeconds = trimmed === "" ? null : Number(trimmed);
+      if (
+        responseTimeoutSeconds !== null &&
+        (!Number.isInteger(responseTimeoutSeconds) ||
+          responseTimeoutSeconds < 10 ||
+          responseTimeoutSeconds > 600)
+      ) {
+        throw new Error(t("settings.ai.responseTimeoutInvalid"));
+      }
+      await apiPatch("/api/user/ai-provider", { responseTimeoutSeconds });
+    },
+    onSuccess: () => {
+      setOk(true);
+      setMsg(t("settings.ai.saved"));
+      queryClient.invalidateQueries({ queryKey: queryKeys.userAiProvider() });
+    },
+    onError: (e) => {
+      setOk(false);
+      setMsg(e instanceof Error ? e.message : t("settings.ai.errorGeneric"));
+    },
+  });
+
+  return (
+    <div className="bg-muted/50 space-y-3 rounded-lg p-4">
+      <div>
+        <p className="text-sm font-medium">
+          {t("settings.ai.responseTimeoutHeading")}
+        </p>
+        <p className="text-muted-foreground text-xs">
+          {t("settings.ai.responseTimeoutBody")}
+        </p>
+      </div>
+      <div>
+        <Label htmlFor="ai-response-timeout">
+          {t("settings.ai.responseTimeoutLabel")}
+        </Label>
+        <Input
+          id="ai-response-timeout"
+          type="number"
+          inputMode="numeric"
+          min={10}
+          max={600}
+          value={value}
+          onChange={(e) => setValue(e.target.value)}
+          placeholder={t("settings.ai.responseTimeoutPlaceholder")}
+          className="mt-1 sm:max-w-xs"
+        />
+      </div>
+      <div>
+        <Button
+          size="sm"
+          className="min-h-11 sm:min-h-9"
+          onClick={() => saveMutation.mutate()}
+          disabled={saveMutation.isPending}
+        >
+          {saveMutation.isPending ? (
+            <Loader2 className="h-4 w-4 animate-spin motion-reduce:animate-none" />
+          ) : (
+            <Save className="h-4 w-4" />
+          )}
+          {t("settings.ai.saveCta")}
+        </Button>
+      </div>
+      {msg && (
+        <p className={`text-xs ${ok ? "text-success" : "text-destructive"}`}>
+          {msg}
+        </p>
+      )}
+    </div>
+  );
+}

src/components/settings/ai/shared.ts

@@ -27,8 +27,21 @@ export interface UserAIProvider {
   hasLocalKey: boolean;
   hasOpenaiKey: boolean;
   openaiKeyPreview: string | null;
+  // v1.22 (#89) — per-user response timeout, in seconds (null = default).
+  responseTimeoutSeconds: number | null;
+  // v1.22 (#90) — dedicated document-scan (Lab-OCR) provider config.
+  ocrEnabled: boolean;
+  ocrProvider: string | null;
+  ocrModel: string | null;
+  ocrBaseUrl: string | null;
+  hasOcrKey: boolean;
+  ocrKeyPreview: string | null;
 }
 
+/** v1.22 (#90) — providers the dedicated document-scan path accepts. */
+export const OCR_PROVIDER_TYPES = ["OPENAI", "ANTHROPIC", "LOCAL"] as const;
+export type OcrProviderType = (typeof OCR_PROVIDER_TYPES)[number];
+
 /**
  * v1.4.16 phase B2 — provider tags exposed to the UI. Mirrors
  * `PROVIDER_CHAIN_TYPES` server-side; kept as a const-array here to

src/lib/ai/__tests__/local-client.test.ts

@@ -200,6 +200,135 @@ describe("LocalOpenAICompatibleClient", () => {
     ).rejects.toThrow("Local AI returned empty content");
   });
 
+  // ── v1.22 (#89) — true token streaming ──────────────────────────
+  function sseStreamResponse(chunks: string[]) {
+    const encoder = new TextEncoder();
+    const body = new ReadableStream<Uint8Array>({
+      start(controller) {
+        for (const c of chunks) controller.enqueue(encoder.encode(c));
+        controller.close();
+      },
+    });
+    return {
+      ok: true,
+      status: 200,
+      headers: {
+        get: (k: string) => (k === "content-type" ? "text/event-stream" : null),
+      },
+      body,
+    };
+  }
+
+  it("streams token deltas over stream:true and assembles the full reply", async () => {
+    const mockFetch = vi
+      .fn()
+      .mockResolvedValue(
+        sseStreamResponse([
+          'data: {"choices":[{"delta":{"content":"Hel"}}]}\n\n',
+          'data: {"choices":[{"delta":{"content":"lo"}}]}\n\n',
+          'data: {"choices":[{"delta":{"content":" world"}}]}\n\n',
+          'data: {"choices":[{"delta":{}}],"usage":{"total_tokens":42}}\n\n',
+          "data: [DONE]\n\n",
+        ]),
+      );
+    vi.stubGlobal("fetch", mockFetch);
+
+    const client = new LocalOpenAICompatibleClient({
+      apiKey: null,
+      model: "llama3:8b",
+      baseUrl: "http://localhost:11434/v1",
+    });
+
+    const deltas: string[] = [];
+    const result = await client.generateCompletionStream(
+      singleUserTurn({ system: "s", user: "u" }),
+      (d) => deltas.push(d),
+    );
+
+    expect(deltas).toEqual(["Hel", "lo", " world"]);
+    expect(result.content).toBe("Hello world");
+    expect(result.tokensUsed).toBe(42);
+    expect(result.providerType).toBe("local");
+
+    // The request opted into streaming.
+    const body = JSON.parse(mockFetch.mock.calls[0][1].body);
+    expect(body.stream).toBe(true);
+    expect(body.stream_options).toEqual({ include_usage: true });
+  });
+
+  it("falls back to the buffered body when the server ignores stream:true", async () => {
+    // 200 OK but a normal JSON completion (not event-stream).
+    const mockFetch = vi.fn().mockResolvedValue({
+      ok: true,
+      status: 200,
+      headers: { get: () => "application/json" },
+      body: null,
+      json: () =>
+        Promise.resolve({
+          choices: [{ message: { content: "buffered reply" } }],
+          usage: { total_tokens: 7 },
+        }),
+    });
+    vi.stubGlobal("fetch", mockFetch);
+
+    const client = new LocalOpenAICompatibleClient({
+      apiKey: null,
+      model: "llama3",
+      baseUrl: "http://localhost:11434/v1",
+    });
+
+    const deltas: string[] = [];
+    const result = await client.generateCompletionStream(
+      singleUserTurn({ system: "s", user: "u" }),
+      (d) => deltas.push(d),
+    );
+
+    expect(result.content).toBe("buffered reply");
+    expect(result.tokensUsed).toBe(7);
+    // The whole buffered body is emitted as one delta.
+    expect(deltas).toEqual(["buffered reply"]);
+  });
+
+  it("falls back to non-streaming generateCompletion when streaming is rejected", async () => {
+    const mockFetch = vi
+      .fn()
+      // 1st call: the streaming attempt is rejected (server 400s on stream).
+      .mockResolvedValueOnce({
+        ok: false,
+        status: 400,
+        headers: { get: () => "application/json" },
+        text: () => Promise.resolve("stream not supported"),
+      })
+      // 2nd call: the buffered retry succeeds.
+      .mockResolvedValueOnce({
+        ok: true,
+        status: 200,
+        json: () =>
+          Promise.resolve({
+            choices: [{ message: { content: "non-stream reply" } }],
+            usage: { total_tokens: 5 },
+          }),
+      });
+    vi.stubGlobal("fetch", mockFetch);
+
+    const client = new LocalOpenAICompatibleClient({
+      apiKey: null,
+      model: "llama3",
+      baseUrl: "http://localhost:11434/v1",
+    });
+
+    const result = await client.generateCompletionStream(
+      singleUserTurn({ system: "s", user: "u" }),
+      () => {},
+    );
+
+    expect(result.content).toBe("non-stream reply");
+    expect(mockFetch).toHaveBeenCalledTimes(2);
+    // The retry was the buffered (non-streaming) request shape.
+    const retryBody = JSON.parse(mockFetch.mock.calls[1][1].body);
+    expect(retryBody).not.toHaveProperty("stream");
+  });
+
   it("degrades tools silently — never forwards a `tools` field", async () => {
     const mockFetch = vi.fn().mockResolvedValue({
       ok: true,

src/lib/ai/__tests__/ocr-provider-resolution.test.ts

@@ -0,0 +1,120 @@
+import { describe, it, expect, vi, beforeEach } from "vitest";
+
+// v1.22 (#90) — dedicated document-scan provider resolution.
+
+vi.mock("@/lib/db", () => ({
+  prisma: {
+    user: { findUnique: vi.fn() },
+    appSettings: { findUnique: vi.fn() },
+  },
+}));
+vi.mock("@/lib/crypto", () => ({
+  decrypt: vi.fn((v: string) => `decrypted:${v}`),
+  encrypt: vi.fn((v: string) => `encrypted:${v}`),
+}));
+vi.mock("@/lib/ai/codex-oauth", () => ({
+  refreshDeviceTokens: vi.fn(),
+  encryptCodexCreds: vi.fn(),
+  decryptCodexCreds: vi.fn(),
+}));
+
+import { resolveOcrProviderChain } from "../provider";
+import { prisma } from "@/lib/db";
+
+describe("resolveOcrProviderChain", () => {
+  beforeEach(() => {
+    vi.resetAllMocks();
+  });
+
+  it("uses the dedicated OCR provider when enabled with a key", async () => {
+    // Single read: the OCR columns. The dedicated provider builds, so the main
+    // chain is never resolved.
+    vi.mocked(prisma.user.findUnique).mockResolvedValueOnce({
+      aiOcrEnabled: true,
+      aiOcrProvider: "OPENAI",
+      aiOcrModel: "gpt-4o",
+      aiOcrBaseUrl: null,
+      aiOcrKeyEncrypted: "enc-key",
+    } as never);
+
+    const res = await resolveOcrProviderChain("u1");
+
+    expect(res.dedicated).toBe(true);
+    expect(res.ocrModelOverride).toBe("gpt-4o");
+    expect(res.chain).toHaveLength(1);
+    expect(res.chain[0].providerType).toBe("openai");
+    // Only the OCR-column read happened — no main-chain resolution.
+    expect(prisma.user.findUnique).toHaveBeenCalledTimes(1);
+  });
+
+  it("maps LOCAL OCR provider to the local chain tag via base URL", async () => {
+    vi.mocked(prisma.user.findUnique).mockResolvedValueOnce({
+      aiOcrEnabled: true,
+      aiOcrProvider: "LOCAL",
+      aiOcrModel: "llava",
+      aiOcrBaseUrl: "http://localhost:11434/v1",
+      aiOcrKeyEncrypted: null,
+    } as never);
+
+    const res = await resolveOcrProviderChain("u1");
+
+    expect(res.dedicated).toBe(true);
+    expect(res.chain[0].providerType).toBe("local");
+  });
+
+  it("falls back to the main chain when OCR is disabled", async () => {
+    // First read: OCR columns (disabled). Second read: the main provider chain
+    // resolution (no enabled chain / creds) → empty main chain.
+    vi.mocked(prisma.user.findUnique)
+      .mockResolvedValueOnce({
+        aiOcrEnabled: false,
+        aiOcrProvider: null,
+        aiOcrModel: null,
+        aiOcrBaseUrl: null,
+        aiOcrKeyEncrypted: null,
+      } as never)
+      .mockResolvedValueOnce({
+        aiProvider: null,
+        aiModel: null,
+        aiBaseUrl: null,
+        aiAnthropicKeyEncrypted: null,
+        aiLocalKeyEncrypted: null,
+        aiOpenaiKeyEncrypted: null,
+        aiProviderChain: null,
+      } as never);
+
+    const res = await resolveOcrProviderChain("u1");
+
+    expect(res.dedicated).toBe(false);
+    expect(res.ocrModelOverride).toBeNull();
+    // The OCR-column read plus the main-chain resolution both ran (the
+    // fallback path), so more than the single dedicated-path read happened.
+    expect(
+      vi.mocked(prisma.user.findUnique).mock.calls.length,
+    ).toBeGreaterThanOrEqual(2);
+  });
+
+  it("falls back to the main chain when enabled but the key is missing", async () => {
+    vi.mocked(prisma.user.findUnique)
+      .mockResolvedValueOnce({
+        aiOcrEnabled: true,
+        aiOcrProvider: "OPENAI",
+        aiOcrModel: "gpt-4o",
+        aiOcrBaseUrl: null,
+        aiOcrKeyEncrypted: null, // missing credential → cannot build
+      } as never)
+      .mockResolvedValueOnce({
+        aiProvider: null,
+        aiModel: null,
+        aiBaseUrl: null,
+        aiAnthropicKeyEncrypted: null,
+        aiLocalKeyEncrypted: null,
+        aiOpenaiKeyEncrypted: null,
+        aiProviderChain: null,
+      } as never);
+
+    const res = await resolveOcrProviderChain("u1");
+
+    expect(res.dedicated).toBe(false);
+  });
+});

src/lib/ai/__tests__/provider-runner.test.ts

@@ -32,6 +32,7 @@ import {
   getLastWorkingProvider,
   isHardProviderFailure,
   runRawCompletionWithFallback,
+  runStreamingRawCompletionWithFallback,
   runWithFallback,
 } from "../provider-runner";
 import {
@@ -720,3 +721,96 @@ describe("AllProvidersFailedError — primaryCredentialExpired", () => {
     expect(e.primaryCredentialExpired).toBe(false);
   });
 });
+
+// ── v1.22 (#89) — streaming runner ───────────────────────────────────
+/** A provider that streams: emits each char of `text` through onDelta. */
+class StreamingProvider implements AIProvider {
+  readonly type: AIProvider["type"] = "local";
+  streamCalls = 0;
+  bufferedCalls = 0;
+  lastStreamParams: CompletionParams | null = null;
+  constructor(private readonly text: string) {}
+  async generateCompletion(): Promise<CompletionResult> {
+    this.bufferedCalls += 1;
+    return {
+      content: this.text,
+      tokensUsed: 1,
+      model: "stream-model",
+      providerType: this.type,
+    };
+  }
+  async generateCompletionStream(
+    params: CompletionParams,
+    onDelta: (d: string) => void,
+  ): Promise<CompletionResult> {
+    this.streamCalls += 1;
+    this.lastStreamParams = params;
+    for (const ch of this.text) onDelta(ch);
+    return {
+      content: this.text,
+      tokensUsed: 3,
+      model: "stream-model",
+      providerType: this.type,
+    };
+  }
+}
+
+describe("runStreamingRawCompletionWithFallback", () => {
+  it("drives a streaming provider via generateCompletionStream + onDelta and forwards timeoutMs", async () => {
+    const ledger = createInMemoryProviderHealthLedger();
+    const provider = new StreamingProvider("Hi");
+    const deltas: string[] = [];
+    const result = await runStreamingRawCompletionWithFallback({
+      userId: "u-stream",
+      providers: [{ providerType: "local", instance: provider }],
+      params: singleUserTurn({ system: "s", user: "u", timeoutMs: 12345 }),
+      onDelta: (d) => deltas.push(d),
+      ledger,
+    });
+    expect(deltas).toEqual(["H", "i"]);
+    expect(result.result.content).toBe("Hi");
+    expect(provider.streamCalls).toBe(1);
+    expect(provider.bufferedCalls).toBe(0);
+    // Per-request timeout is threaded through to the provider call.
+    expect(provider.lastStreamParams?.timeoutMs).toBe(12345);
+  });
+
+  it("falls back to generateCompletion for a provider that does not stream", async () => {
+    const ledger = createInMemoryProviderHealthLedger();
+    // ScriptedProvider has no generateCompletionStream → buffered path.
+    const provider = new ScriptedProvider({
+      script: [{ ok: true, content: "x" }],
+    });
+    const deltas: string[] = [];
+    const result = await runStreamingRawCompletionWithFallback({
+      userId: "u-nostream",
+      providers: [{ providerType: "local", instance: provider }],
+      params: singleUserTurn({ system: "s", user: "u" }),
+      onDelta: (d) => deltas.push(d),
+      ledger,
+    });
+    expect(result.result.content).toBe("x");
+    expect(provider.callCount).toBe(1);
+    expect(deltas).toEqual([]); // no streaming → no deltas
+  });
+
+  it("cascades past a hard failure to the next (streaming) provider", async () => {
+    const ledger = createInMemoryProviderHealthLedger();
+    const broken = new ScriptedProvider({
+      script: [{ ok: false, error: err(503) }],
+    });
+    const good = new StreamingProvider("ok");
+    const result = await runStreamingRawCompletionWithFallback({
+      userId: "u-cascade",
+      providers: [
+        { providerType: "openai", instance: broken },
+        { providerType: "local", instance: good },
+      ],
+      params: singleUserTurn({ system: "s", user: "u" }),
+      onDelta: () => {},
+      ledger,
+    });
+    expect(result.workingProvider.providerType).toBe("local");
+    expect(result.fallbackHops.map((h) => h.providerType)).toEqual(["openai"]);
+  });
+});

src/lib/ai/coach/tools/loop.ts

@@ -96,6 +96,12 @@ export async function runCoachToolLoop(args: {
   ledger?: ProviderHealthLedger;
   /** Aborts the per-round provider calls on client disconnect. */
   signal?: AbortSignal;
+  /**
+   * v1.22 (#89) — per-user upstream timeout (ms) for each tool-round provider
+   * call. Forwarded to `CompletionParams.timeoutMs`; unset → the client's
+   * shared default holds.
+   */
+  timeoutMs?: number;
 }): Promise<CoachToolLoopResult> {
   const {
     userId,
@@ -108,6 +114,7 @@ export async function runCoachToolLoop(args: {
     sharedScope,
     ledger,
     signal,
+    timeoutMs,
   } = args;
 
   const messages: AiMessage[] = [...args.messages];
@@ -135,6 +142,7 @@ export async function runCoachToolLoop(args: {
         temperature,
         maxTokens,
         signal,
+        timeoutMs,
         ...(offerTools
           ? { tools, toolChoice: "auto" as const }
           : { toolChoice: "none" as const }),

src/lib/ai/local-client.ts

@@ -17,6 +17,14 @@ interface LocalClientConfig {
 const STRICT_JSON_PREFIX =
   "Return strict JSON only, no markdown, no commentary.\n\n";
 
+/**
+ * v1.22 (#89) — absolute backstop for the streaming path. The real ceiling is
+ * the per-idle-gap timer (`CompletionParams.timeoutMs`); this only stops a
+ * server that streams forever from pinning a worker. Ten minutes is far beyond
+ * any legitimate single Coach turn.
+ */
+const STREAM_ABSOLUTE_CEILING_MS = 10 * 60_000;
+
 /**
  * Prepend the strict-JSON instruction to the FIRST user turn's text. Local
  * model templates respect an in-message instruction better than the
@@ -64,9 +72,16 @@ export class LocalOpenAICompatibleClient implements AIProvider {
     this.config = config;
   }
 
-  async generateCompletion(
+  /**
+   * Build the shared request shape (url + headers + body) for both the
+   * buffered and the streaming paths. `stream` adds the OpenAI-compatible
+   * `stream: true` + `stream_options.include_usage` so the final SSE chunk
+   * carries the token total.
+   */
+  private buildRequest(
     params: CompletionParams,
-  ): Promise<CompletionResult> {
+    stream: boolean,
+  ): { url: string; headers: Record<string, string>; body: string } {
     const url = `${this.config.baseUrl.replace(/\/$/, "")}/chat/completions`;
 
     const headers: Record<string, string> = {
@@ -89,6 +104,36 @@ export class LocalOpenAICompatibleClient implements AIProvider {
       withStrictJsonPrefix(params.messages),
     );
 
+    const body = JSON.stringify({
+      model: this.config.model,
+      messages,
+      temperature: params.temperature ?? 0.3,
+      max_tokens: params.maxTokens ?? 1000,
+      // Deterministic seed for reproducible reference output. Ollama +
+      // most OpenAI-compatible local servers honour it; servers that
+      // ignore it simply disregard the field. Omitted when unset.
+      ...(params.seed !== undefined ? { seed: params.seed } : {}),
+      // v1.18.7 — Ollama's native structured-output switch. The JSON
+      // surfaces opt in via `responseFormat: "json"`; it cuts the
+      // first-pass JSON-failure rate that the fence-stripping safety net
+      // otherwise has to catch. The Coach (prose) leaves it unset.
+      ...(params.responseFormat === "json" ? { format: "json" } : {}),
+      // v1.22 (#89) — streaming opt-in. `stream_options.include_usage` asks
+      // OpenAI-compatible servers (vLLM / LM Studio / exo) to append a final
+      // chunk carrying the usage block so the token footer survives streaming.
+      ...(stream
+        ? { stream: true, stream_options: { include_usage: true } }
+        : {}),
+    });
+
+    return { url, headers, body };
+  }
+
+  async generateCompletion(
+    params: CompletionParams,
+  ): Promise<CompletionResult> {
+    const { url, headers, body } = this.buildRequest(params, false);
+
     // safeFetch defaults: no redirect-follow (the local endpoint is the
     // most exploitable on this surface — a user-controlled baseUrl that
     // 302s to 169.254.169.254 would otherwise leak the bearer on the
@@ -109,21 +154,7 @@ export class LocalOpenAICompatibleClient implements AIProvider {
       {
         method: "POST",
         headers,
-        body: JSON.stringify({
-          model: this.config.model,
-          messages,
-          temperature: params.temperature ?? 0.3,
-          max_tokens: params.maxTokens ?? 1000,
-          // Deterministic seed for reproducible reference output. Ollama +
-          // most OpenAI-compatible local servers honour it; servers that
-          // ignore it simply disregard the field. Omitted when unset.
-          ...(params.seed !== undefined ? { seed: params.seed } : {}),
-          // v1.18.7 — Ollama's native structured-output switch. The JSON
-          // surfaces opt in via `responseFormat: "json"`; it cuts the
-          // first-pass JSON-failure rate that the fence-stripping safety net
-          // otherwise has to catch. The Coach (prose) leaves it unset.
-          ...(params.responseFormat === "json" ? { format: "json" } : {}),
-        }),
+        body,
       },
       // v1.20.1 — compose the caller's cancel signal (Coach SSE disconnect) so
       // a mid-generation abort tears the upstream call down early.
@@ -181,4 +212,182 @@ export class LocalOpenAICompatibleClient implements AIProvider {
       providerType: "local",
     };
   }
+
+  /**
+   * v1.22 (#89) — true token streaming over the OpenAI-compatible
+   * `stream: true` wire. POSTs with `stream: true`, reads the SSE response body
+   * incrementally, and emits each `choices[0].delta.content` chunk through
+   * `onDelta` as it arrives. The returned `CompletionResult` carries the FULL
+   * assembled reply so every downstream guard still runs on the complete text.
+   *
+   * Timeout model: per-IDLE-gap, not whole-call. An idle timer (`timeoutMs`,
+   * default 60 s) is armed before the request and reset on every chunk, so a
+   * slow first token (an MLX/exo server loading the model) and a long total
+   * generation are both fine as long as bytes keep flowing. A genuine stall
+   * (no byte for the idle window) aborts the upstream. A generous absolute
+   * ceiling backstops a server that streams forever.
+   *
+   * Graceful fallback: a server that rejects `stream: true` (non-2xx), or
+   * ignores it and returns a buffered JSON body, or has no readable stream
+   * body, degrades to the non-streaming {@link generateCompletion} path. The
+   * Coach gets real tokens where the backend supports streaming and identical
+   * behaviour where it does not.
+   */
+  async generateCompletionStream(
+    params: CompletionParams,
+    onDelta: (delta: string) => void,
+  ): Promise<CompletionResult> {
+    const { url, headers, body } = this.buildRequest(params, true);
+    const allowPrivate = isLocalAiHostAllowed(url);
+
+    // Idle controller: aborts the upstream when no chunk arrives within the
+    // per-idle window. Composed with the caller's cancel signal so a client
+    // disconnect still tears the call down immediately.
+    const idle = new AbortController();
+    const idleMs = params.timeoutMs ?? 60_000;
+    const signal = params.signal
+      ? AbortSignal.any([params.signal, idle.signal])
+      : idle.signal;
+
+    let res: Response;
+    try {
+      res = await safeFetch(
+        url,
+        { method: "POST", headers, body },
+        {
+          // Absolute backstop only — the per-idle timer below is the real
+          // ceiling. A server that streams forever still cannot pin a worker.
+          timeoutMs: STREAM_ABSOLUTE_CEILING_MS,
+          requirePublicHost: !allowPrivate,
+          signal,
+        },
+      );
+    } catch {
+      // Network / private-host / abort before the stream opened → fall back to
+      // the buffered path, which surfaces the structured error the chain runner
+      // understands (or succeeds on a server that simply doesn't stream).
+      return this.generateCompletion(params);
+    }
+
+    if (!res.ok) {
+      // The server rejected the streaming request (some servers 400 on
+      // `stream`/`stream_options`). Retry once on the buffered path so a
+      // non-streaming server still works and surfaces its real status.
+      return this.generateCompletion(params);
+    }
+
+    const contentType = res.headers.get("content-type") ?? "";
+    if (!contentType.includes("text/event-stream") || !res.body) {
+      // The server ignored the stream flag and returned a buffered completion
+      // (or there is no readable body). Parse it like the non-streaming path
+      // rather than paying a second round-trip.
+      const json = (await res.json().catch(() => null)) as {
+        choices?: Array<{ message?: { content?: string } }>;
+        usage?: { total_tokens?: number };
+      } | null;
+      const buffered = json?.choices?.[0]?.message?.content;
+      if (!buffered) {
+        const err = new Error("Local AI returned empty content");
+        Object.assign(err, {
+          httpStatus: 0,
+          kind: "empty_response",
+          upstream: "local",
+        });
+        throw err;
+      }
+      onDelta(buffered);
+      return {
+        content: buffered,
+        tokensUsed: json?.usage?.total_tokens ?? null,
+        model: this.config.model,
+        providerType: "local",
+      };
+    }
+
+    const reader = res.body.getReader();
+    const decoder = new TextDecoder();
+    let sseBuffer = "";
+    let content = "";
+    let tokensUsed: number | null = null;
+
+    let idleTimer: ReturnType<typeof setTimeout> | null = null;
+    const armIdle = () => {
+      if (idleTimer) clearTimeout(idleTimer);
+      idleTimer = setTimeout(() => idle.abort(), idleMs);
+    };
+    armIdle();
+
+    try {
+      for (;;) {
+        const { done, value } = await reader.read();
+        if (done) break;
+        armIdle();
+        sseBuffer += decoder.decode(value, { stream: true });
+        // SSE frames are separated by a blank line; each frame is one or more
+        // `data:` lines. Split on newlines and process complete lines, keeping
+        // any partial trailing line in the buffer.
+        let nl: number;
+        while ((nl = sseBuffer.indexOf("\n")) !== -1) {
+          const line = sseBuffer.slice(0, nl).trim();
+          sseBuffer = sseBuffer.slice(nl + 1);
+          if (!line.startsWith("data:")) continue;
+          const payload = line.slice("data:".length).trim();
+          if (!payload || payload === "[DONE]") continue;
+          let chunk: {
+            choices?: Array<{ delta?: { content?: string } }>;
+            usage?: { total_tokens?: number };
+          };
+          try {
+            chunk = JSON.parse(payload);
+          } catch {
+            // A malformed chunk line is skipped, not fatal — keep reading.
+            continue;
+          }
+          const delta = chunk.choices?.[0]?.delta?.content;
+          if (delta) {
+            content += delta;
+            onDelta(delta);
+          }
+          if (typeof chunk.usage?.total_tokens === "number") {
+            tokensUsed = chunk.usage.total_tokens;
+          }
+        }
+      }
+    } catch (err) {
+      // An abort (idle stall or client disconnect) mid-stream: if we already
+      // assembled usable prose, return it — the user got a real (if clipped)
+      // answer and every guard still runs on it. With nothing assembled, this
+      // is a hard transport failure for the chain runner to cascade on.
+      if (!content) {
+        const e = new Error("Local AI stream failed before any content");
+        Object.assign(e, {
+          httpStatus: 0,
+          kind: "stream_aborted",
+          upstream: "local",
+          cause: err,
+        });
+        throw e;
+      }
+    } finally {
+      if (idleTimer) clearTimeout(idleTimer);
+      reader.cancel().catch(() => {});
+    }
+
+    if (!content) {
+      const err = new Error("Local AI returned empty content");
+      Object.assign(err, {
+        httpStatus: 0,
+        kind: "empty_response",
+        upstream: "local",
+      });
+      throw err;
+    }
+
+    return {
+      content,
+      tokensUsed,
+      model: this.config.model,
+      providerType: "local",
+    };
+  }
 }

src/lib/ai/provider-runner.ts

@@ -424,6 +424,92 @@ export async function runWithFallback(
   throw new AllProvidersFailedError(hops);
 }
 
+/**
+ * v1.22 (#89) — streaming sibling of {@link runRawCompletionWithFallback}.
+ * Identical chain / fallback / health-ledger semantics, but each candidate is
+ * driven through its `generateCompletionStream` (emitting `onDelta` per token)
+ * when the provider implements one, and through the buffered
+ * `generateCompletion` otherwise. The caller (the Coach SSE route) uses the
+ * delta callback to keep the connection warm and — once safety guards have run
+ * on the full reply — to confirm real tokens flowed. The returned shape matches
+ * the buffered runner so the route is path-agnostic.
+ */
+export async function runStreamingRawCompletionWithFallback(args: {
+  userId: string;
+  providers: ProviderChainResolved[];
+  params: CompletionParams;
+  /** Called once per streamed token/chunk for providers that stream. */
+  onDelta: (delta: string) => void;
+  /** See `RunWithFallbackParams.ledger`. */
+  ledger?: ProviderHealthLedger;
+}): Promise<RunRawWithFallbackResult> {
+  const { userId, providers, params, onDelta } = args;
+  const ledger = args.ledger ?? postgresProviderHealthLedger;
+
+  if (providers.length === 0) {
+    throw new AllProvidersFailedError([]);
+  }
+
+  const ordered = await resolveChainOrder(userId, providers, ledger);
+  const hops: FallbackHop[] = [];
+
+  for (let i = 0; i < ordered.length; i += 1) {
+    const candidate = ordered[i];
+    try {
+      // Prefer the provider's streaming wire; fall back to the buffered call
+      // for providers (cloud clients today) that do not implement streaming.
+      const result = candidate.instance.generateCompletionStream
+        ? await candidate.instance.generateCompletionStream(params, onDelta)
+        : await candidate.instance.generateCompletion(params);
+      rememberWorkingProvider(userId, candidate.providerType);
+      void ledger.recordSuccess(userId, candidate.providerType);
+      annotate({
+        meta: {
+          ai_chain_working_provider: candidate.providerType,
+          ai_chain_fallback_count: hops.length,
+          ai_chain_streamed: Boolean(
+            candidate.instance.generateCompletionStream,
+          ),
+        },
+      });
+      return {
+        result,
+        workingProvider: candidate,
+        fallbackHops: hops,
+      };
+    } catch (error) {
+      if (!isHardProviderFailure(error)) {
+        throw error;
+      }
+      const summary = summariseError(error);
+      void ledger.recordFailure(userId, candidate.providerType, summary.status);
+      const hop: FallbackHop = {
+        providerType: candidate.providerType,
+        attempt: i + 1,
+        failureReason: summary.reason,
+        httpStatus: summary.status,
+      };
+      hops.push(hop);
+      annotate({
+        meta: {
+          [`ai_chain_hop_${i + 1}_provider`]: candidate.providerType,
+          [`ai_chain_hop_${i + 1}_status`]: summary.status,
+          [`ai_chain_hop_${i + 1}_reason`]: summary.reason.slice(0, 240),
+          [`ai_chain_hop_${i + 1}_body`]: summary.bodyExcerpt,
+        },
+      });
+    }
+  }
+
+  annotate({
+    meta: {
+      ai_chain_outcome: "all-failed",
+      ai_chain_fallback_count: hops.length,
+    },
+  });
+  throw new AllProvidersFailedError(hops);
+}
+
 export interface RunRawWithFallbackResult {
   /** Raw provider response — the legacy `/api/insights/generate` route
    *  consumes the rich shape from `result.content`, so the runner does

src/lib/ai/provider.ts

@@ -314,6 +314,114 @@ export async function resolveProviderChain(
   return resolved;
 }
 
+/**
+ * v1.22 (#90) — dedicated document-scan (Lab-OCR) provider columns.
+ */
+type UserOcrRow = {
+  aiOcrEnabled: boolean | null;
+  aiOcrProvider: string | null;
+  aiOcrModel: string | null;
+  aiOcrBaseUrl: string | null;
+  aiOcrKeyEncrypted: string | null;
+};
+
+/**
+ * Build the dedicated OCR provider instance from the per-user `aiOcr*` columns.
+ * Returns the constructed instance paired with the chain tag the consent gate +
+ * vision-capability resolver understand, or null when the OCR provider is
+ * unconfigured / missing its credential (the caller then falls back to the main
+ * provider chain so document scanning is never silently lost).
+ */
+function buildOcrProvider(
+  row: UserOcrRow,
+): { providerType: ProviderChainType; instance: AIProvider } | null {
+  if (!row.aiOcrEnabled) return null;
+  const choice = row.aiOcrProvider?.toUpperCase();
+  switch (choice) {
+    case "ANTHROPIC": {
+      if (!row.aiOcrKeyEncrypted) return null;
+      return {
+        providerType: "anthropic",
+        instance: new AnthropicClient({
+          apiKey: decrypt(row.aiOcrKeyEncrypted),
+          model: row.aiOcrModel ?? "claude-sonnet-4-6",
+        }),
+      };
+    }
+    case "OPENAI": {
+      if (!row.aiOcrKeyEncrypted) return null;
+      return {
+        providerType: "openai",
+        instance: new OpenAIClient({
+          apiKey: decrypt(row.aiOcrKeyEncrypted),
+          model: row.aiOcrModel ?? "gpt-4o",
+          baseUrl: "https://api.openai.com/v1",
+        }),
+      };
+    }
+    case "LOCAL": {
+      if (!row.aiOcrBaseUrl) return null;
+      return {
+        providerType: "local",
+        instance: new LocalOpenAICompatibleClient({
+          apiKey: row.aiOcrKeyEncrypted ? decrypt(row.aiOcrKeyEncrypted) : null,
+          model: row.aiOcrModel ?? "local-model",
+          baseUrl: row.aiOcrBaseUrl,
+        }),
+      };
+    }
+    default:
+      return null;
+  }
+}
+
+/**
+ * v1.22 (#90) — resolve the provider chain for the Lab-OCR ingestion path.
+ *
+ * When the user enabled a dedicated document-scan provider (Settings → AI →
+ * Document scanning), this returns a single-entry chain built from the `aiOcr*`
+ * columns plus the OCR model that drives the vision-capability allowlist.
+ * Otherwise (the default) it returns the MAIN provider chain unchanged, so the
+ * OCR path behaves exactly as before the feature existed. A dedicated provider
+ * that fails to build (missing key / base URL) also falls through to the main
+ * chain so scanning is never silently dropped.
+ */
+export async function resolveOcrProviderChain(userId: string): Promise<{
+  chain: ProviderChainResolved[];
+  /** The model driving the dedicated provider's vision allowlist, else null. */
+  ocrModelOverride: string | null;
+  /** True when the dedicated document-scan provider served this chain. */
+  dedicated: boolean;
+}> {
+  const row = await prisma.user.findUnique({
+    where: { id: userId },
+    select: {
+      aiOcrEnabled: true,
+      aiOcrProvider: true,
+      aiOcrModel: true,
+      aiOcrBaseUrl: true,
+      aiOcrKeyEncrypted: true,
+    },
+  });
+
+  if (row?.aiOcrEnabled) {
+    const built = buildOcrProvider(row);
+    if (built) {
+      return {
+        chain: [built],
+        ocrModelOverride: row.aiOcrModel ?? null,
+        dedicated: true,
+      };
+    }
+  }
+
+  return {
+    chain: await resolveProviderChain(userId),
+    ocrModelOverride: null,
+    dedicated: false,
+  };
+}
+
 /**
  * Credential-presence subset of the `User` row that
  * `userRowHasProviderCredential` evaluates. Mirrors the columns

src/lib/ai/types.ts

@@ -121,6 +121,24 @@ export interface AIProvider {
    */
   supportsTools?: boolean;
   generateCompletion(params: CompletionParams): Promise<CompletionResult>;
+  /**
+   * v1.22 (#89) — optional true-streaming variant. A provider that can emit
+   * tokens incrementally (the local OpenAI-compatible client over `stream:true`)
+   * implements this so a slow self-hosted backend (e.g. an MLX/exo server that
+   * takes >60 s to load the model before the first token) produces output as it
+   * arrives rather than buffering the whole reply behind a single fetch — which
+   * is what tripped the legacy total-request timeout and the reverse-proxy idle
+   * cut. `onDelta` is called once per token/chunk as it arrives; the returned
+   * `CompletionResult` carries the FULL assembled reply so every downstream
+   * guard (grounding, refusal, sentinel-stripping) still runs on the complete
+   * text. The timeout becomes a per-idle-gap ceiling (time-to-first-token +
+   * inter-token gap), not a whole-call ceiling. Providers without a streaming
+   * wire omit this; the runner falls back to {@link generateCompletion}.
+   */
+  generateCompletionStream?(
+    params: CompletionParams,
+    onDelta: (delta: string) => void,
+  ): Promise<CompletionResult>;
 }
 
 /**

src/lib/crypto/encrypted-columns.ts

@@ -68,6 +68,8 @@ export const ENCRYPTED_COLUMNS: readonly EncryptedColumn[] = [
   { model: "User", field: "aiAnthropicKeyEncrypted", kind: "string" },
   { model: "User", field: "aiLocalKeyEncrypted", kind: "string" },
   { model: "User", field: "aiOpenaiKeyEncrypted", kind: "string" },
+  // v1.22 (#90) — dedicated document-scan (Lab-OCR) provider key.
+  { model: "User", field: "aiOcrKeyEncrypted", kind: "string" },
   { model: "User", field: "insuranceNumberEncrypted", kind: "string" },
 
   // ───── OAuth token tables (`*_token`-convention columns) ─────

src/lib/labs/ocr-capability.ts

@@ -10,7 +10,7 @@
  * vision allowlist.
  */
 import type { ProviderChainResolved } from "@/lib/ai/provider-runner";
-import { resolveProvider, resolveProviderChain } from "@/lib/ai/provider";
+import { resolveProvider, resolveOcrProviderChain } from "@/lib/ai/provider";
 import { resolveCodexVisionSlug } from "@/lib/ai/codex-client";
 import { prisma } from "@/lib/db";
 import {
@@ -70,14 +70,21 @@ export async function resolveVisionProvider(
     where: { id: "singleton" },
     select: { adminAiModel: true },
   });
-  const ctx: ModelContext = {
-    userModel: userRow?.aiModel ?? null,
-    adminModel: settings?.adminAiModel ?? null,
-  };
 
   const localOcrEnabled = userRow?.labsLocalOcrEnabled ?? false;
 
-  const chain = await resolveProviderChain(userId);
+  // v1.22 (#90) — resolve the OCR chain: the dedicated document-scan provider
+  // when the user enabled one, else the main provider chain unchanged. When a
+  // dedicated provider is active, ITS model drives the vision allowlist.
+  const ocr = await resolveOcrProviderChain(userId);
+  const ctx: ModelContext = {
+    userModel: ocr.dedicated
+      ? (ocr.ocrModelOverride ?? userRow?.aiModel ?? null)
+      : (userRow?.aiModel ?? null),
+    adminModel: settings?.adminAiModel ?? null,
+  };
+
+  const chain = ocr.chain;
   if (chain.length === 0) {
     // Mirror the Coach: fall back to the legacy single-provider resolution and
     // tag it as the admin-managed entry the consent gate recognises.
@@ -116,7 +123,9 @@ export async function resolveTextProvider(userId: string): Promise<{
   chain: ProviderChainResolved[];
   pick: { entry: ProviderChainResolved; providerType: string } | null;
 }> {
-  const chain = await resolveProviderChain(userId);
+  // v1.22 (#90) — use the dedicated document-scan provider when enabled, else
+  // the main chain (the text-mode structuring pass needs no vision).
+  const chain = (await resolveOcrProviderChain(userId)).chain;
   if (chain.length === 0) {
     const legacy = await resolveProvider(userId);
     if (legacy.type !== "none") {

src/lib/openapi/routes/profile.ts

@@ -334,6 +334,34 @@ const aiProviderResponse = z
       .string()
       .nullable()
       .describe("`...` + last 4 chars of the stored OpenAI key, or null."),
+    responseTimeoutSeconds: z
+      .number()
+      .int()
+      .nullable()
+      .describe(
+        "v1.22 (#89) — per-user response timeout for AI generation, in seconds (10–600). Null = the generous built-in default (~180 s). Mainly for slow local/self-hosted backends.",
+      ),
+    ocrEnabled: z
+      .boolean()
+      .describe(
+        "v1.22 (#90) — whether a dedicated document-scan (Lab-OCR) provider is active. When false the OCR path uses the main provider chain.",
+      ),
+    ocrProvider: z
+      .enum(["OPENAI", "ANTHROPIC", "LOCAL"])
+      .nullable()
+      .describe("The dedicated document-scan provider, or null when unset."),
+    ocrModel: z.string().nullable(),
+    ocrBaseUrl: z
+      .string()
+      .nullable()
+      .describe("Custom base URL (LOCAL OCR provider only); null otherwise."),
+    hasOcrKey: z.boolean(),
+    ocrKeyPreview: z
+      .string()
+      .nullable()
+      .describe(
+        "`...` + last 4 chars of the stored document-scan provider key, or null.",
+      ),
   })
   .meta({
     id: "AiProviderResponse",