fix(gui): let the framed custom domain dialog finish closing before it answers

The answer removes the frame from the page, so the dialog's exit transition
never showed. The framed action waits for the overlay to leave first, with the
wait the feedback embed already had, now one shared helper.

.github/workflows/ai-provider-integration-tests.yaml

@@ -53,7 +53,7 @@ jobs:
             # step above fails; the tests skip themselves without the key.
             - name: Run live realtime conversation tests
               if: ${{ !cancelled() }}
-              run: npm run setupExtensions && (cd src/puter-js && npm run build) && npx vitest run --config src/puter-js/tests/api/vitest.config.ts src/puter-js/tests/api/realtime/convoLive.test.ts
+              run: npm run setupExtensions && (cd src/puter-js && npm run build) && npx vitest run --config src/puter-js/tests/api/vitest.config.ts src/puter-js/tests/api/realtime/realtimeLive.test.ts
               env:
                   CI: 'true'
                   PUTER_TEST_AI_OPENAI_API_KEY: ${{ secrets.PUTER_TEST_AI_OPENAI_API_KEY }}

CLAUDE.md

@@ -1 +1 @@
-FOLLOW ./AGENTS.md
+FOLLOW @AGENTS.md

doc/embedding-app-browser.md

@@ -124,7 +124,7 @@ localization system.
 </script>
 ```
 
-## Implementation and checks
+## Implementation
 
 `createAppBrowser` in `src/gui/src/UI/Dashboard/AppBrowser.js` owns the shared UI
 and data behavior; its `guestApps` option is the signed-out list. `TabApps.js`
@@ -134,14 +134,232 @@ injects desktop actions. `appsEmbed.js` supplies new-tab launching through
 `launch_app`, and loads only the embed entry, its SDK, and shared styles. `embedHostUser` in
 `src/gui/src/helpers/embedHostUser.js` reads the host's user from the fragment, and
 `embedHostTheme.js` beside it applies the host's theme.
-The normal GUI webpack build emits `bundle.min.js` and `apps-embed.min.js`.
+The normal GUI webpack build emits `bundle.min.js`, `apps-embed.min.js`, and `feedback-embed.min.js`.
 The `/embed/apps` route puts the recommended list in the page's config as
 `guestApps`. No puter.js API or app-list endpoint changes are needed.
 
+## Feedback dialog
+
+`GET /embed/feedback` is the app feedback dialog (`puter.ui.showFeedbackDialog()`)
+for a first-party page, with the same framing rules, query redirect, and
+host-user binding as `/embed/apps`. It sends the feedback for the framing
+page's origin, which the server reads from the referrer and pins with
+`frame-ancestors`; opened top-level, it does nothing. Its page is transparent
+so the dialog's own scrim covers the host.
+
+puter.js uses it on a page one label under the GUI origin, other than the
+subdomains the server refuses, when the page is signed in, framing it in a
+modal `<dialog>` that stays hidden until the embed is ready. A session check that takes longer than 200ms shows a loading scrim
+with a spinner in its place; Escape or a click on it cancels. The embed posts `{ msg: 'feedbackEmbed', state }` to its parent
+(`loading`, then `ready` or `unavailable`) and finally
+`{ msg: 'feedbackDialogClosed', sent }`. It reports `unavailable` without the
+named user's session, or while that user's email still needs confirming, and
+puter.js then uses the `/action/send-feedback` popup, which can sign in. So does
+a frame that never answers, such as one the host's CSP blocks. `feedbackEmbed.js`
+is the entry, sharing `embedSession` in `src/gui/src/helpers/embedSession.js`
+with the app browser; `showFeedbackEmbed` in `src/puter-js/src/lib/guiEmbed.js`
+is the host side.
+
+## Sign-in and file pickers
+
+`GET /embed/sign-in`, `/embed/show-open-file-picker`, `/embed/show-save-file-picker`
+and `/embed/show-directory-picker` serve the same GUI popups as `/action/<action>`
+for a first-party page, with the same framing rules, query redirect, and
+`no-store` as `/embed/apps`. Any other spelling of the path (another case, a
+trailing slash) redirects to it, since the shell boots the desktop for it. They render the normal GUI shell, which names the
+framing page (read from the referrer, pinned with `frame-ancestors`) as
+`gui_params.embed_host_origin`. The GUI then runs the popup boot unchanged, with
+the framing page as `window.opener`, and `window.close()` asks the page to
+remove the frame (`src/gui/src/helpers/popupFrame.js`). Without a framing page
+to answer (opened top-level, or framed by a refused page) it does nothing. The
+popup's parameters come from the fragment instead of the query: `msg_id`,
+`options` and `request_auth`, plus `user` for a picker.
+
+The frame posts `{ msg: 'puterPopupFrame', state }` to its parent: `loading` as
+soon as the shell boots, `ready` once it shows a window or dialog, `unavailable`
+when the page has to use the popup, `provider` when the user picks a sign-in
+provider (below), and `closed` where a popup would close. Its
+answers (`puter.token`, `fileOpenPicked`, `directoryPicked`, `fileSaved`, the
+cancel messages, `sendMeFileData`) are the popup's.
+
+puter.js uses them on a page one label under the GUI origin, other than the
+subdomains the server refuses, unless the page is cross-origin isolated or in a
+frame of another site (which `frame-ancestors` refuses too).
+`showPopupFrame` in `src/puter-js/src/lib/guiEmbed.js` shows the frame in a box
+the size of the popup window, over a scrim, hidden until `ready` and behind the
+loading spinner after 200ms. Escape closes it, on the page or in the frame. One shows per popup, as each
+popup has its own window name: a second picker request of the same kind
+replaces the one showing, and a `signIn()` call made while a framed sign-in is
+open settles with it. On
+screens up to 600px wide the box fills the screen. A frame that is refused,
+never answers, or reports `unavailable` falls back to the popup, opened through
+a consent dialog when the click that asked for it has expired. A reload or navigation the frame
+starts itself (logging out, a sign-out in another tab) names the frame as its
+referrer, so the server refuses it; the page then shows its spinner and opens
+the frame again, once per `ready`.
+
+- **Sign-in** (`puter.auth.signIn()`) has no host-user binding: it is how the
+  page gets a user. The box has a close button, which rejects with
+  `auth_window_closed` as closing the popup does. Providers refuse to be
+  framed, so choosing one posts `{ state: 'provider', url }` with the
+  provider's start URL for this sign-in's `msg_id`, then `unavailable`. The
+  page opens the popup straight at that URL, once it checks the URL is the
+  GUI's `/auth/oidc/<provider>/start` for this `msg_id`, and the provider
+  returns it as this sign-in's popup. `attempt_temp_user_creation` keeps the popup, since
+  only the popup creates a temporary account.
+- **Pickers** frame only for a signed-in page. The GUI reports `unavailable`
+  unless its session belongs to the user named in `user`, and the page then uses
+  the popup, which can sign in and switch accounts. The token the GUI sends
+  before a picker's answer isn't adopted, since it is for the page's own user.
+  Dismissing the loading scrim settles the picker as cancelled.
+
+## Upgrade
+
+`GET /embed/upgrade` is the deployment's upgrade dialog for a first-party page,
+served like the popups above: the GUI shell with the same framing rules, query
+redirect, exact-path redirect and `no-store`, naming the framing page as
+`gui_params.embed_host_origin`. It runs only when framed by an allowed page.
+
+puter.js frames it on a signed-in first-party page (the same pages as the
+pickers) for `puter.ui.requestUpgrade(details)`, and for the Upgrade button of
+the usage limit dialog a refused call shows there. Elsewhere both keep their
+behavior: inside an app `requestUpgrade()` asks the desktop, and on any other
+page the usage limit dialog opens the dashboard's upgrade page in a new tab.
+
+The frame covers the page with no box or close button, and the GUI page is
+transparent (`html.popup-frame-transparent`) so the dialog's own scrim covers
+the host. The frame allows `payment` as well as `otp-credentials`. Escape on the
+page is left to the dialog, since one there shouldn't drop a payment in
+progress. Its fragment names `user` (the page's signed-in user) and, when the
+refusal gave them, `reason` (`funds`, `subscription` or `storage`) and `method`.
+
+The GUI reports `unavailable` unless its session belongs to `user`, the account
+has no verification or password change waiting (which the frame couldn't be
+closed from), and the deployment defines `window.UIUpgradeAccount` with
+`UIUpgradeAccount.embeddable === true`. Otherwise it runs:
+
+```js
+new UIUpgradeAccount({ attribution: { appUid, reason, method }, onClose }).open_as_window();
+```
+
+`appUid` is the app the framing page's origin resolves to, never one the
+fragment names. `open_as_window()` resolves once the dialog shows, and the GUI
+then posts `ready`. While the page still hides the frame, which Chromium and
+WebKit lay out at 0x0, the GUI posts `ready` first and opens the dialog once
+the frame has a viewport, so a window it opens doesn't center on 0x0.
+`onClose({ upgraded })` must run exactly once, when the
+dialog closes or when it declines to open for the account; the GUI then posts
+`{ msg: 'puterPopupFrame', state: 'closed', upgraded }` and the page removes
+the frame. If `open_as_window()` rejects, the GUI posts `unavailable` instead. `requestUpgrade()` resolves `{ upgraded }`. A call made while the
+frame shows settles with it rather than replacing it, and a call refused while
+it shows still rejects but opens no usage limit dialog over it.
+
+When the frame can't show the upgrade (refused, never ready, `unavailable`, or
+no user to name), puter.js opens the dashboard's upgrade page in a new tab with
+what was refused, through a consent dialog when the click has expired or the
+tab is blocked, and `requestUpgrade()` resolves `{ upgraded: false }`.
+
+## Top-up
+
+`GET /embed/top-up` is the deployment's top-up dialog, where a user with a plan
+buys add-on credit, served and framed like the upgrade: a transparent GUI page
+covering the host with no box or close button, allowing `payment` and
+`otp-credentials`, with Escape left to the dialog. Its fragment names `user`.
+
+puter.js frames it for the **Add credits** item of `<puter-user-menu>`, which
+shows only when whoami reports `subscribed: true` and no `team` paying for the
+account. On a page that can't frame the GUI, the item opens the dashboard with
+`?topup=` in a new tab instead.
+
+The GUI reports `unavailable` unless its session belongs to `user`, the account
+has no verification or password change waiting, and the deployment defines
+`window.UITopUp` with `UITopUp.embeddable === true`. Otherwise it runs:
+
+```js
+new UITopUp({ onClose }).open();
+```
+
+`open()` resolves once the dialog has loaded the account and shows what it can
+buy, and `onClose({ toppedUp })` must run exactly once, when the dialog closes
+or when it declines to open for the account. The GUI then posts `{ msg: 'puterPopupFrame', state: 'closed', toppedUp }`
+and the page removes the frame. When the frame can't show the top-up, puter.js
+opens the dashboard's top-up in a new tab, as it does the upgrade.
+
+While the top-up shows, a call refused for funds still rejects but opens no
+usage limit dialog over it, and `requestUpgrade()` frames nothing over it: it
+resolves `{ upgraded: false }` once the top-up closes.
+
+## Email confirmation
+
+`GET /embed/request-email-confirmation` walks a first-party page's user through
+confirming their email, served like the upgrade: the GUI shell with the same
+framing rules, redirects and `no-store`, run only when framed by an allowed
+page. puter.js frames it on a signed-in first-party page for
+`puter.ui.requestEmailConfirmation()`. Inside an app the desktop shows the same
+windows over the app instead.
+
+Like the upgrade, the frame covers the page with no box or close button, over a
+transparent GUI page, and its fragment names `user`. Escape closes it, on the
+page or on the save-account or code window. On a window, Escape closes that
+window as its close button does, so a code already sent still answers. The GUI
+reports `unavailable` unless its session belongs to `user` and the account has
+no phone, card or password check waiting.
+An email check is fine: the boot skips its own email gate for this action, so
+the user gets one code window they can close.
+
+The GUI then runs `ensureConfirmedEmail` (`src/gui/src/helpers/ensureConfirmedEmail.js`),
+which the desktop's handler uses too: a temporary account gets the save-account
+window first, then an unconfirmed email gets the code window, each at most
+once. When the page's whoami already reads the email as confirmed,
+`requestEmailConfirmation()` resolves `{ response: true }` without framing
+anything, and a confirmation in the frame marks that whoami confirmed for the
+calls after it. An email the GUI reads as confirmed shows nothing. Before
+opening a window the GUI waits for the frame to have a viewport, as the upgrade
+does. It then posts
+`{ msg: 'puterPopupFrame', state: 'closed', confirmed }`, and
+`requestEmailConfirmation()` resolves `{ response: confirmed }`, the field an
+app reads from the desktop's answer. A call made while the frame shows settles
+with it. A frame that is refused, never ready, or `unavailable` resolves
+`{ response: false }`, since there's no popup to fall back to.
+
+## Custom domains
+
+`GET /embed/attach-custom-domain` is the custom domain dialog of
+`puter.hosting.attach()`, the popup at `/action/attach-custom-domain`, served
+like the upgrade: the GUI shell with the same framing rules, redirects and
+`no-store`, run only when framed by an allowed page. puter.js frames it on a
+signed-in first-party page (the same pages as the pickers).
+
+The frame covers the page with no box or close button, over a transparent GUI
+page, so the dialog floats over the host with its own scrim as it does on the
+desktop. The frame allows `clipboard-write`, for the dialog's copy buttons, as
+well as `otp-credentials`. Escape on the page is left to the dialog, which
+doesn't close while a request is in flight. Its fragment names `user` (the
+page's signed-in user), `subdomain` (the site to attach a domain to) and
+`msg_id`.
+
+The GUI reports `unavailable` unless its session belongs to `user` and the
+account has no verification or password change waiting. It then runs the
+popup's action unchanged, with the framing page as the requester the dialog
+names: it posts `ready` once the dialog is on the page, answers
+`{ msg: 'customDomainAttached', domain, original_msg_id }` once the dialog has
+closed and its exit has played, and then posts `closed`. `domain` is the domain the user chose, or
+null. `attach()` resolves it. A frame that closes without an answer, or a
+loading scrim the user dismisses, resolves null.
+
+A call made while the frame shows resolves null without replacing it, as a
+second call does on the desktop. A frame that is refused, never ready, or
+`unavailable` falls back to the popup, which can sign in, through a consent
+dialog when the click has expired.
+
+## Checks
+
 Run the GUI component/browser tests (with the SDK bundle built and Playwright
 Chromium installed), and the backend route tests:
 
 ```sh
-npx vitest run --config src/gui/vitest.config.js src/gui/src/UI/Dashboard/AppBrowser
+npx vitest run --config src/gui/vitest.config.js src/gui/src/UI/Dashboard/AppBrowser src/feedbackEmbed
+npx vitest run src/puter-js/src/lib/guiEmbed.test.js src/puter-js/src/modules/UI.showFeedbackDialog.test.js src/puter-js/src/modules/Auth.signInFrame.test.js src/puter-js/src/modules/UI.pickerFrame.test.js src/puter-js/src/modules/UI.requestUpgrade.test.js src/puter-js/src/modules/UI.showTopUp.test.js src/puter-js/src/modules/UI.attachCustomDomainFrame.test.js src/puter-js/src/lib/upgradePrompt.test.js src/puter-js/src/modules/UsageLimitDialog.upgrade.test.js
+npx vitest run --config src/gui/vitest.config.js src/gui/src/helpers/popupFrame.test.js
 npx vitest run --config src/backend/vitest.config.ts src/backend/server.test.ts
 ```

src/backend/clients/event/types.ts

@@ -420,7 +420,11 @@ export type EventMap = {
         old_username: string;
         new_username: string;
     };
-    'user.email-changed': { user_id: number; new_email: string };
+    'user.email-changed': {
+        user_id: number;
+        old_email: string | null;
+        new_email: string;
+    };
     /** An account was suspended, or its suspension lifted. Local events. */
     'user.suspended': { userId: number };
     'user.unsuspended': { userId: number };

src/backend/clients/s3/S3Client.test.ts

@@ -21,6 +21,8 @@
 import { GetObjectCommand } from '@aws-sdk/client-s3';
 import { mkdirSync, mkdtempSync, rmSync, writeFileSync } from 'node:fs';
 import { existsSync } from 'node:fs';
+import { createServer, type Server } from 'node:http';
+import type { AddressInfo } from 'node:net';
 import { tmpdir } from 'node:os';
 import { join } from 'node:path';
 import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest';
@@ -200,6 +202,77 @@ describe('S3Client — remote endpoint configuration', () => {
     });
 });
 
+describe('S3Client — bucket in another region', () => {
+    let server: Server;
+    let client: S3Client;
+    let signedRegions: string[];
+
+    // Answers like a bucket that lives in us-west-2: a request signed for any
+    // other region gets S3's 301 naming the bucket's region.
+    beforeEach(async () => {
+        signedRegions = [];
+        server = createServer((req, res) => {
+            const region =
+                /Credential=[^/]+\/\d{8}\/([^/]+)\//.exec(
+                    req.headers.authorization ?? '',
+                )?.[1] ?? '';
+            signedRegions.push(region);
+            if (region !== 'us-west-2') {
+                res.writeHead(301, {
+                    'content-type': 'application/xml',
+                    'x-amz-bucket-region': 'us-west-2',
+                });
+                res.end(
+                    '<?xml version="1.0" encoding="UTF-8"?><Error><Code>PermanentRedirect</Code><Message>The bucket you are attempting to access must be addressed using the specified endpoint.</Message></Error>',
+                );
+                return;
+            }
+            res.writeHead(200, {
+                'content-type': 'text/plain',
+                'content-length': '5',
+            });
+            res.end('hello');
+        });
+        await new Promise<void>((resolve) =>
+            server.listen(0, '127.0.0.1', resolve),
+        );
+        const { port } = server.address() as AddressInfo;
+        client = new S3Client({
+            port: 0,
+            extensions: [],
+            s3: {
+                s3Config: {
+                    endpoint: `http://127.0.0.1:${port}`,
+                    accessKeyId: 'fakeAccessKeyId',
+                    secretAccessKey: 'fakeSecretAccessKey',
+                    forcePathStyle: true,
+                },
+            },
+        } as unknown as IConfig);
+        await client.onServerStart();
+    });
+
+    afterEach(async () => {
+        await client.onServerShutdown();
+        server.closeAllConnections();
+        await new Promise<void>((resolve) => server.close(() => resolve()));
+    });
+
+    it("retries a read in the bucket's region when S3 redirects it", async () => {
+        const result = await client
+            .get('ap-northeast-2')
+            .send(
+                new GetObjectCommand({
+                    Bucket: 'other-bucket',
+                    Key: 'index.html',
+                }),
+            );
+
+        expect(await result.Body?.transformToString()).toBe('hello');
+        expect(signedRegions).toEqual(['ap-northeast-2', 'us-west-2']);
+    });
+});
+
 describe('S3Client — legacy storage migration', () => {
     let client: S3Client;
     let dir: string;

src/backend/clients/s3/S3Client.ts

@@ -235,6 +235,9 @@ export class S3Client extends PuterClient {
 
         const client = new AwsS3Client({
             region,
+            // Rows without a stored region are read with the server's own, so
+            // retry where S3's 301 says the bucket actually lives.
+            followRegionRedirects: true,
             requestStreamBufferSize: 32 * 1024,
             requestHandler: new NodeHttpHandler({
                 socketTimeout: 5000,

src/backend/controllers/auth/AuthController.test.ts

@@ -7468,6 +7468,11 @@ describe('AuthController user-protected mutations (validation paths)', () => {
             { expiresIn: '1h' },
         );
 
+        const heard: Array<Record<string, unknown>> = [];
+        eventClient.on('user.email-changed', (_k, data) => {
+            heard.push(data as Record<string, unknown>);
+        });
+
         const req = makeReq({});
         (req as unknown as { query: Record<string, string> }).query = {
             token: linkJwt,
@@ -7475,6 +7480,11 @@ describe('AuthController user-protected mutations (validation paths)', () => {
         const res = makeRes();
         await controller.handleChangeEmailConfirm(req, res);
         expect(res.sent).toMatch(/Email changed successfully/);
+        expect(heard).toContainEqual({
+            user_id: user.id,
+            old_email: user.email ?? null,
+            new_email: newEmail,
+        });
 
         const after = await server.stores.user.getById(user.id, {
             force: true,

src/backend/controllers/auth/AuthController.ts

@@ -2647,6 +2647,7 @@ export class AuthController extends PuterController {
                 'user.email-changed' as never,
                 {
                     user_id: user.id,
+                    old_email: user.email ?? null,
                     new_email: newEmail,
                 } as never,
                 {},

src/backend/controllers/auth/oidcLoginReturn.ts

@@ -15,7 +15,6 @@ export interface OidcLoginContinuation {
     embeddedInPopup: boolean;
     openerOrigin: string | null;
     messageId: string | null;
-    switchedToLogin: boolean;
 }
 
 interface OidcReturnDeps {
@@ -56,7 +55,6 @@ export function completeOidcLoginReturn(
         data.userUuid !== user.uuid ||
         typeof data.redirectUri !== 'string' ||
         typeof data.embeddedInPopup !== 'boolean' ||
-        typeof data.switchedToLogin !== 'boolean' ||
         (data.openerOrigin !== null && typeof data.openerOrigin !== 'string') ||
         (data.messageId !== null && typeof data.messageId !== 'string')
     ) {
@@ -91,6 +89,5 @@ export function completeOidcLoginReturn(
             ),
         );
     }
-    if (data.switchedToLogin) append('oidc_switched', 'login');
     return target;
 }

src/backend/controllers/drivers/DriverController.test.ts

@@ -501,6 +501,29 @@ describe('DriverController.#handleCall (via captured router)', () => {
         expect(body.service.name).toBe('puter-kvstore');
     });
 
+    it('treats null `args` as no arguments', async () => {
+        const actor = await makeUserActor();
+        await server.stores.permission.setFlatUserPerm(
+            actor.user!.id!,
+            'service:puter-subdomains:ii:puter-subdomains',
+            {
+                permission: 'service:puter-subdomains:ii:puter-subdomains',
+                deleted: false,
+                issuer_user_id: actor.user!.id!,
+            } as never,
+        );
+
+        const res = makeRes();
+        const req = makeReq(
+            { interface: 'puter-subdomains', method: 'select', args: null },
+            actor,
+        );
+        await runWithContext({ actor }, () =>
+            routes['POST /call'](req, res as unknown as Response, () => {}),
+        );
+        expect((res.body as { success: boolean }).success).toBe(true);
+    });
+
     it('pipes a stream-shaped result instead of JSON', async () => {
         // Use an ephemeral driver registered onto the controller's bag.
         const streamPayload = Buffer.from('hello-world');

src/backend/controllers/drivers/DriverController.ts

@@ -276,8 +276,10 @@ export class DriverController extends PuterController {
             interface: ifaceName,
             method,
             driver: driverName,
-            args = {},
+            args: rawArgs,
         } = (req.body ?? {}) as Record<string, unknown>;
+        // Drivers read fields off `args`, so a null one counts as absent.
+        const args = rawArgs ?? {};
 
         if (!ifaceName || typeof ifaceName !== 'string') {
             throw new HttpError(400, 'Missing or invalid `interface`', {

src/backend/controllers/homepage/HomepageController.test.ts

@@ -27,7 +27,7 @@ import type { Actor } from '../../core/actor.js';
 import { PuterRouter } from '../../core/http/PuterRouter.js';
 import { PuterServer } from '../../server.js';
 import { setupTestServer } from '../../testUtil.js';
-import { embedFrameAncestors } from './HomepageController.js';
+import { embedFrameAncestors, embedHostOrigin } from './HomepageController.js';
 
 // ── Test harness ────────────────────────────────────────────────────
 //
@@ -550,3 +550,26 @@ describe('embedFrameAncestors', () => {
         ).toBe("'self'");
     });
 });
+
+describe('embedHostOrigin', () => {
+    const config = {
+        origin: 'https://example.test',
+        domain: 'example.test',
+        api_base_url: 'https://backend.example.test',
+        static_hosting_domain: 'site.example.test',
+    };
+
+    it.each([
+        ['https://sheets.example.test/doc/1', 'https://sheets.example.test'],
+        ['https://backend.example.test/x', null],
+        ['https://api.example.test/', null],
+        ['https://someone.site.example.test/', null],
+        ['https://a.sheets.example.test/', null],
+        ['http://sheets.example.test/', null],
+        ['https://example.test/', null],
+        ['not a url', null],
+        [undefined, null],
+    ])('%s frames as %s', (referer, expected) => {
+        expect(embedHostOrigin(config, referer)).toBe(expected);
+    });
+});

src/backend/controllers/homepage/HomepageController.ts

@@ -31,8 +31,26 @@ import type {
     PuterHomepageService,
     PageMeta,
     LaunchOptions,
+    PopupFrame,
 } from '../../services/homepage/PuterHomepageService';
 
+/**
+ * GUI popups a first-party page can show in a frame at `/embed/<action>`,
+ * instead of opening `/action/<action>` in a window, plus the upgrade and
+ * top-up dialogs, which it would otherwise open in a dashboard tab, and the
+ * email confirmation, which it otherwise can't show.
+ */
+export const FRAMED_POPUP_ACTIONS = [
+    'sign-in',
+    'show-open-file-picker',
+    'show-save-file-picker',
+    'show-directory-picker',
+    'upgrade',
+    'top-up',
+    'request-email-confirmation',
+    'attach-custom-domain',
+] as const;
+
 /** Longest description/og:description/twitter:description on `/app/<name>`. */
 export const APP_META_DESCRIPTION_MAX = 150;
 
@@ -61,32 +79,34 @@ export function appMetaDescription(text: string): string {
     return `${cut.replace(/[\s,;:.!?…—-]+$/, '')}…`;
 }
 
+type EmbedFramingConfig = Pick<
+    IConfig,
+    | 'origin'
+    | 'domain'
+    | 'api_base_url'
+    | 'static_hosting_domain'
+    | 'static_hosting_domain_alt'
+    | 'private_app_hosting_domain'
+    | 'private_app_hosting_domain_alt'
+>;
+
 /**
- * `frame-ancestors` for the app browser embed. A wildcard over `*.<domain>`
+ * The origin of a first-party page that may frame an `/embed/*` page, named
+ * exactly from the request's referrer, or null. A wildcard over `*.<domain>`
  * would also admit user-hosted sites whenever the hosting domains sit under it,
- * so the embedding page is named exactly, from the request's referrer. Only a
- * single label under `<domain>` qualifies, and not one this server answers
- * itself (the API, WebDAV, SDK and native-app hosts).
+ * so only a single label under `<domain>` qualifies, and not one this server
+ * answers itself (the API, WebDAV, SDK and native-app hosts).
  */
-export function embedFrameAncestors(
-    config: Pick<
-        IConfig,
-        | 'origin'
-        | 'domain'
-        | 'api_base_url'
-        | 'static_hosting_domain'
-        | 'static_hosting_domain_alt'
-        | 'private_app_hosting_domain'
-        | 'private_app_hosting_domain_alt'
-    >,
+export function embedHostOrigin(
+    config: EmbedFramingConfig,
     referer: string | undefined,
-): string {
+): string | null {
     const gui = new URL(config.origin ?? `https://${config.domain}`);
     let parent: URL;
     try {
         parent = new URL(referer ?? '');
     } catch {
-        return "'self'";
+        return null;
     }
     const host = parent.hostname.toLowerCase();
     const label = subdomainLabelOf(host, config.domain);
@@ -105,7 +125,16 @@ export function embedFrameAncestors(
         !isReservedSubdomain(label) &&
         host !== new URL(config.api_base_url).hostname &&
         !hostingDomains.some((d) => host === d || host.endsWith(`.${d}`));
-    return allowed ? `'self' ${parent.origin}` : "'self'";
+    return allowed ? parent.origin : null;
+}
+
+/** `frame-ancestors` for an `/embed/*` page: the GUI and {@link embedHostOrigin}. */
+export function embedFrameAncestors(
+    config: EmbedFramingConfig,
+    referer: string | undefined,
+): string {
+    const host = embedHostOrigin(config, referer);
+    return host ? `'self' ${host}` : "'self'";
 }
 
 /**
@@ -142,6 +171,7 @@ export class HomepageController extends PuterController {
             res: express.Response,
             metaOverrides: Partial<PageMeta> = {},
             launch: LaunchOptions = {},
+            popupFrame?: PopupFrame,
         ) => {
             const meta = { ...defaultMeta(req), ...metaOverrides };
             const actor =
@@ -150,19 +180,27 @@ export class HomepageController extends PuterController {
                         actor?: Parameters<typeof homepage.send>[0]['actor'];
                     }
                 ).actor ?? null;
-            await homepage.send({ req, res, actor }, meta, launch);
+            await homepage.send({ req, res, actor }, meta, launch, popupFrame);
         };
 
         // -- Root + path-aliased shell routes ------------------------
 
         router.get('/', {}, (req, res) => sendShell(req, res));
 
-        router.get('/embed/apps', {}, async (req, res) => {
-            // SDK bootstrap parameters must not change this page's account or API origin.
-            // Checked on the raw URL; `req.query` drops pairs past the parser's limit.
+        /**
+         * Headers shared by the `/embed/*` pages; false when the request was
+         * redirected instead. SDK bootstrap parameters must not change an
+         * embed's account or API origin, so a query string is dropped. Checked
+         * on the raw URL; `req.query` drops pairs past the parser's limit.
+         */
+        const prepareEmbed = (
+            req: express.Request,
+            res: express.Response,
+            path: string,
+        ): boolean => {
             if (req.originalUrl.includes('?')) {
-                res.redirect('/embed/apps');
-                return;
+                res.redirect(path);
+                return false;
             }
             res.removeHeader('X-Frame-Options');
             res.setHeader(
@@ -170,6 +208,11 @@ export class HomepageController extends PuterController {
                 `frame-ancestors ${embedFrameAncestors(this.config, req.get('referer'))}`,
             );
             res.setHeader('Cache-Control', 'no-store');
+            return true;
+        };
+
+        router.get('/embed/apps', {}, async (req, res) => {
+            if (!prepareEmbed(req, res, '/embed/apps')) return;
             // Shown to visitors without a session; a signed-in embed fetches
             // its own list, so a failure here only costs signed-out visitors.
             const guestApps = await this.services.recommendedApps
@@ -178,6 +221,48 @@ export class HomepageController extends PuterController {
             res.send(homepage.renderAppBrowserEmbed({ guestApps }));
         });
 
+        router.get('/embed/feedback', {}, (req, res) => {
+            if (!prepareEmbed(req, res, '/embed/feedback')) return;
+            // The feedback is recorded against this origin, which
+            // frame-ancestors has just pinned to the actual framing page.
+            res.send(
+                homepage.renderFeedbackEmbed({
+                    hostOrigin: embedHostOrigin(
+                        this.config,
+                        req.get('referer'),
+                    ),
+                }),
+            );
+        });
+
+        // The GUI runs the popup's own boot, answering the framing page
+        // instead of an opener.
+        for (const action of FRAMED_POPUP_ACTIONS) {
+            const embedPath = `/embed/${action}`;
+            router.get(embedPath, {}, async (req, res) => {
+                // Routes match case-insensitively and with a trailing slash,
+                // and the shell boots the desktop for any other spelling, so
+                // only the exact one may be framed.
+                if (req.originalUrl !== embedPath) {
+                    res.redirect(embedPath);
+                    return;
+                }
+                if (!prepareEmbed(req, res, embedPath)) return;
+                await sendShell(
+                    req,
+                    res,
+                    {},
+                    {},
+                    {
+                        hostOrigin: embedHostOrigin(
+                            this.config,
+                            req.get('referer'),
+                        ),
+                    },
+                );
+            });
+        }
+
         router.get('/settings', {}, (req, res) => sendShell(req, res));
         router.get('/settings/*splat', {}, (req, res) => sendShell(req, res));
 

src/backend/controllers/oidc/OIDCController.test.ts

@@ -1620,7 +1620,7 @@ describe('OIDCController signup callback', () => {
         expect(captured.redirectUrl).toContain('action=signup');
     });
 
-    it('creates a fresh user without oidc_switched on first signup', async () => {
+    it('creates a fresh user on first signup', async () => {
         const state = oidc().signState({
             provider: 'custom',
             redirect_uri: TEST_ORIGIN + '/',
@@ -1646,11 +1646,9 @@ describe('OIDCController signup callback', () => {
         );
         expect(captured.cookies).toHaveLength(1);
         expect(captured.redirectUrl).toBe(TEST_ORIGIN + '/');
-        // Fresh creation — must NOT advertise "you've actually been signed in".
-        expect(captured.redirectUrl).not.toContain('oidc_switched=login');
     });
 
-    it('adds oidc_switched=login when signup hits an already-linked account', async () => {
+    it('signs in to an already-linked account on signup, returning as a login would', async () => {
         // Seed an account linked on (provider, sub) so the signup callback
         // resolves to `linked-sub` rather than creating a new user.
         const sub = `sub-${Math.random().toString(36).slice(2, 8)}`;
@@ -1686,7 +1684,7 @@ describe('OIDCController signup callback', () => {
             res,
         );
         expect(captured.cookies).toHaveLength(1);
-        expect(captured.redirectUrl).toContain('oidc_switched=login');
+        expect(captured.redirectUrl).toBe(TEST_ORIGIN + '/');
     });
 
     it('redirects to action=signup with auth_error when user resolution fails (unconfirmed email)', async () => {

src/backend/controllers/oidc/OIDCController.ts

@@ -688,14 +688,7 @@ export class OIDCController extends PuterController {
                 );
             }
 
-            // If we landed on an existing account (either via the
-            // provider_sub or via email match), signal the GUI so it can
-            // render a "signed in" flow rather than "account created".
-            const extra =
-                resolved.origin === 'created'
-                    ? undefined
-                    : { oidc_switched: 'login' };
-            await this.#finishLogin(req, res, user, stateDecoded, extra);
+            await this.#finishLogin(req, res, user, stateDecoded);
         };
         router.get('/auth/oidc/callback/signup', cbOpts, signupCb);
         router.post('/auth/oidc/callback/signup', cbOpts, signupCb);
@@ -834,17 +827,14 @@ if (window.opener) {
         attempt = 0,
     ): Promise<
         | { error: string; code?: string; requestCode?: string }
-        | {
-              user: import('../../stores/user/UserStore.js').UserRow;
-              origin: 'linked-sub' | 'linked-email' | 'created';
-          }
+        | { user: import('../../stores/user/UserStore.js').UserRow }
     > {
         // 1. Existing provider/sub link.
         const linked = await this.services.oidc.findUserByProviderSub(
             provider,
             userinfo.sub,
         );
-        if (linked) return { user: linked, origin: 'linked-sub' };
+        if (linked) return { user: linked };
 
         // 2/3. Email match branch.
         const claimedEmail =
@@ -881,7 +871,7 @@ if (window.opener) {
                         error: outcome.error ?? 'Failed to link provider.',
                     };
                 }
-                return { user: byEmail, origin: 'linked-email' };
+                return { user: byEmail };
             }
         }
 
@@ -911,7 +901,7 @@ if (window.opener) {
                 requestCode: outcome.requestCode,
             };
         }
-        return { user: outcome.user, origin: 'created' };
+        return { user: outcome.user };
     }
 
     async #processCallback(
@@ -1000,7 +990,6 @@ if (window.opener) {
         res: Response,
         user: UserRow,
         stateDecoded: Record<string, unknown>,
-        extraQueryParams?: Record<string, string>,
     ): Promise<void> {
         const origin = (this.config.origin ?? '').replace(/\/$/, '');
         const continuation: OidcLoginContinuation = {
@@ -1018,7 +1007,6 @@ if (window.opener) {
                 stateDecoded.msg_id == null
                     ? null
                     : String(stateDecoded.msg_id),
-            switchedToLogin: extraQueryParams?.oidc_switched === 'login',
         };
         if (user.otp_enabled) {
             const challenge = this.services.token.sign(

src/backend/controllers/oidc/OIDCController.twoFactor.http.test.ts

@@ -139,7 +139,7 @@ describe('OIDC account second-factor login', () => {
             expect(await response.json()).toMatchObject({
                 proceed: true,
                 next_step: 'complete',
-                redirectUrl: `${env.origin}/dashboard${flow === 'signup' ? '?oidc_switched=login' : ''}`,
+                redirectUrl: `${env.origin}/dashboard`,
             });
             expect(
                 response.headers

src/backend/controllers/webdav/WebDAVController.upload.http.test.ts

@@ -0,0 +1,89 @@
+/*
+ * Copyright (C) 2024-present Puter Technologies Inc.
+ *
+ * This file is part of Puter.
+ *
+ * Puter is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU Affero General Public License as published
+ * by the Free Software Foundation, either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU Affero General Public License for more details.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program.  If not, see <https://www.gnu.org/licenses/>.
+ */
+
+import { randomUUID } from 'node:crypto';
+import { afterAll, beforeAll, describe, expect, it } from 'vitest';
+import {
+    setupPuterTestEnv,
+    TEST_USER_CREDENTIALS,
+    type PuterTestEnv,
+} from '../../testUtil.js';
+
+/** A PUT body is the file, whatever content type the client labels it with. */
+describe('PUT stores the body as sent', () => {
+    let env: PuterTestEnv;
+
+    beforeAll(async () => {
+        env = await setupPuterTestEnv();
+    }, 60_000);
+
+    afterAll(async () => {
+        await env?.shutdown();
+    });
+
+    const dav = (path: string, init: RequestInit) => {
+        const port = new URL(env.apiOrigin).port;
+        const auth = Buffer.from(
+            `${TEST_USER_CREDENTIALS.username}:${TEST_USER_CREDENTIALS.password}`,
+        ).toString('base64');
+        return fetch(`http://dav.puter.localhost:${port}${encodeURI(path)}`, {
+            ...init,
+            headers: { Authorization: `Basic ${auth}`, ...init.headers },
+        });
+    };
+
+    it.each([
+        ['application/json', '{"name":"report","rows":[1,2,3]}'],
+        // Not JSON at all: the label alone must not get the body rejected.
+        ['application/json', 'not json'],
+        ['text/plain;actually=json', '{"a":1}'],
+        ['application/x-www-form-urlencoded', 'a=1&b=two'],
+        ['application/octet-stream', 'plain bytes'],
+    ])(
+        'with Content-Type %s',
+        async (contentType, content) => {
+            const path = `/${env.users.user.username}/Documents/put-${randomUUID().slice(0, 8)}.bin`;
+            const put = await dav(path, {
+                method: 'PUT',
+                headers: { 'Content-Type': contentType },
+                body: content,
+            });
+            expect(put.status).toBe(201);
+
+            const get = await dav(path, { method: 'GET' });
+            expect(get.status).toBe(200);
+            expect(await get.text()).toBe(content);
+        },
+        30_000,
+    );
+
+    it('with a form body over the form parser limit', async () => {
+        const path = `/${env.users.user.username}/Documents/put-${randomUUID().slice(0, 8)}.bin`;
+        const content = `a=${'x'.repeat(200 * 1024)}`;
+        const put = await dav(path, {
+            method: 'PUT',
+            headers: { 'Content-Type': 'application/x-www-form-urlencoded' },
+            body: content,
+        });
+        expect(put.status).toBe(201);
+
+        const get = await dav(path, { method: 'GET' });
+        expect(await get.text()).toBe(content);
+    }, 30_000);
+});

src/backend/core/http/middleware/hostRedirects.test.ts

@@ -187,9 +187,9 @@ describe('createUserSubdomainNotFound', () => {
     });
 
     it('passes through reserved subdomains (api, js, native apps, etc.)', () => {
-        // `api`, `js`, `dav`, `docs`, `developer`, `editor`, `pdf`,
+        // `api`, `js`, `dav`, `docs`, `developer`, `about`, `apps`,
         // `puter-app-icons`, etc. all pass through.
-        for (const sub of ['api', 'js', 'docs', 'editor', 'puter-app-icons']) {
+        for (const sub of ['api', 'js', 'docs', 'about', 'puter-app-icons']) {
             const { out, next } = run(
                 createUserSubdomainNotFound(config),
                 makeReq({
@@ -865,13 +865,13 @@ describe('createNativeAppStatic', () => {
 
     beforeAll(() => {
         root = mkdtempSync(nodePath.join(tmpdir(), 'native-apps-'));
-        mkdirSync(nodePath.join(root, 'editor', 'assets'), { recursive: true });
+        mkdirSync(nodePath.join(root, 'about', 'assets'), { recursive: true });
         writeFileSync(
-            nodePath.join(root, 'editor', 'index.html'),
-            '<h1>editor</h1>',
+            nodePath.join(root, 'about', 'index.html'),
+            '<h1>about</h1>',
         );
         writeFileSync(
-            nodePath.join(root, 'editor', 'assets', 'app.js'),
+            nodePath.join(root, 'about', 'assets', 'app.js'),
             'console.log(1);',
         );
         mkdirSync(nodePath.join(root, 'docs', 'dist'), { recursive: true });
@@ -893,7 +893,7 @@ describe('createNativeAppStatic', () => {
 
     it('leaves a native app name alone on a host outside our domains', async () => {
         const req = staticReq({
-            subdomains: ['com', 'example', 'editor'],
+            subdomains: ['com', 'example', 'about'],
             path: '/index.html',
         });
         (req as Request & { is_custom_domain?: boolean }).is_custom_domain =
@@ -907,7 +907,7 @@ describe('createNativeAppStatic', () => {
         const { out, next } = await runStatic(
             createNativeAppStatic(config),
             staticReq({
-                subdomains: ['localhost', 'puter', 'editor'],
+                subdomains: ['localhost', 'puter', 'about'],
                 path: '/index.html',
             }),
         );
@@ -915,17 +915,24 @@ describe('createNativeAppStatic', () => {
         expect(next).toHaveBeenCalledTimes(1);
     });
 
-    it('passes through subdomains that are not native apps', async () => {
-        const { out, next } = await runStatic(
-            withRoot(),
-            staticReq({
-                subdomains: ['localhost', 'puter', 'api'],
-                path: '/index.html',
-            }),
-        );
-        expect(out.sent).toBeUndefined();
-        expect(next).toHaveBeenCalledTimes(1);
-    });
+    // `editor`, `markus` and `pdf` are web apps on standard host routing,
+    // even with a folder of their own under the native apps root.
+    it.each(['api', 'editor', 'markus', 'pdf'])(
+        'passes %s through, as it is not a native app',
+        async (sub) => {
+            mkdirSync(nodePath.join(root, sub), { recursive: true });
+            writeFileSync(nodePath.join(root, sub, 'index.html'), sub);
+            const { out, next } = await runStatic(
+                withRoot(),
+                staticReq({
+                    subdomains: ['localhost', 'puter', sub],
+                    path: '/index.html',
+                }),
+            );
+            expect(out.sent).toBeUndefined();
+            expect(next).toHaveBeenCalledTimes(1);
+        },
+    );
 
     it('passes through when there is no subdomain at all', async () => {
         const { next } = await runStatic(
@@ -939,13 +946,13 @@ describe('createNativeAppStatic', () => {
         const { out, next } = await runStatic(
             withRoot(),
             staticReq({
-                subdomains: ['localhost', 'puter', 'editor'],
+                subdomains: ['localhost', 'puter', 'about'],
                 path: '/index.html',
             }),
         );
         expect(out.sent).toEqual({
             path: '/index.html',
-            root: nodePath.join(root, 'editor'),
+            root: nodePath.join(root, 'about'),
         });
         expect(next).not.toHaveBeenCalled();
     });
@@ -954,11 +961,11 @@ describe('createNativeAppStatic', () => {
         const { out } = await runStatic(
             withRoot(),
             staticReq({
-                subdomains: ['localhost', 'puter', 'EDITOR'],
+                subdomains: ['localhost', 'puter', 'ABOUT'],
                 path: '/index.html',
             }),
         );
-        expect(out.sent?.root).toBe(nodePath.join(root, 'editor'));
+        expect(out.sent?.root).toBe(nodePath.join(root, 'about'));
     });
 
     it('serves docs out of its dist/ subdirectory, not the app root', async () => {
@@ -979,7 +986,7 @@ describe('createNativeAppStatic', () => {
         const { out, next } = await runStatic(
             withRoot(),
             staticReq({
-                subdomains: ['localhost', 'puter', 'editor'],
+                subdomains: ['localhost', 'puter', 'about'],
                 path: '/assets',
                 originalUrl: '/assets?v=2',
             }),
@@ -995,7 +1002,7 @@ describe('createNativeAppStatic', () => {
         const { out } = await runStatic(
             withRoot(),
             staticReq({
-                subdomains: ['localhost', 'puter', 'editor'],
+                subdomains: ['localhost', 'puter', 'about'],
                 path: '/assets/',
             }),
         );
@@ -1007,7 +1014,7 @@ describe('createNativeAppStatic', () => {
         const { out, next } = await runStatic(
             withRoot(),
             staticReq({
-                subdomains: ['localhost', 'puter', 'editor'],
+                subdomains: ['localhost', 'puter', 'about'],
                 path: '/nope.html',
             }),
         );
@@ -1019,7 +1026,7 @@ describe('createNativeAppStatic', () => {
         const { next } = await runStatic(
             withRoot(),
             staticReq({
-                subdomains: ['localhost', 'puter', 'editor'],
+                subdomains: ['localhost', 'puter', 'about'],
                 path: '/index.html',
             }),
             new Error('send failed'),
@@ -1032,7 +1039,7 @@ describe('createNativeAppStatic', () => {
             runStatic(
                 withRoot(),
                 staticReq({
-                    subdomains: ['localhost', 'puter', 'editor'],
+                    subdomains: ['localhost', 'puter', 'about'],
                     path: '/../../etc/passwd',
                 }),
             ),

src/backend/core/http/middleware/jsonBody.http.test.ts

@@ -0,0 +1,263 @@
+/*
+ * Copyright (C) 2024-present Puter Technologies Inc.
+ *
+ * This file is part of Puter.
+ *
+ * Puter is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU Affero General Public License as published
+ * by the Free Software Foundation, either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU Affero General Public License for more details.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program.  If not, see <https://www.gnu.org/licenses/>.
+ */
+
+import type { Server } from 'node:http';
+import type { AddressInfo } from 'node:net';
+import { brotliCompressSync, deflateSync, gzipSync } from 'node:zlib';
+import express, {
+    type ErrorRequestHandler,
+    type RequestHandler,
+} from 'express';
+import { afterAll, beforeAll, describe, expect, it } from 'vitest';
+import { SLICE_BYTES } from '../../../util/slicedJson';
+import { createJsonBodyParser } from './jsonBody';
+
+const LIMIT = 4 * 1024 * 1024;
+const TEXT_AS_JSON = 'text/plain;actually=json';
+
+type Verify = NonNullable<Parameters<typeof express.json>[0]>['verify'];
+const captureRawBody: Verify = (req, _res, buf) => {
+    (req as { rawBody?: Buffer }).rawBody = buf;
+};
+
+/** An app that answers with what its parsers left on the request. */
+const listen = (parsers: RequestHandler[]): Promise<Server> => {
+    const app = express();
+    app.use(...parsers);
+    app.post('/', (req, res) => {
+        const rawBody = (req as { rawBody?: Buffer }).rawBody;
+        res.json({
+            body: req.body,
+            hasBody: req.body !== undefined,
+            rawBody: rawBody?.toString('base64'),
+        });
+    });
+    const onError: ErrorRequestHandler = (err, _req, res, _next) => {
+        res.status(err.statusCode ?? 500).json({ type: err.type });
+    };
+    app.use(onError);
+    return new Promise((resolve) => {
+        const server = app.listen(0, '127.0.0.1', () => resolve(server));
+    });
+};
+
+describe('createJsonBodyParser', () => {
+    let sliced: Server;
+    let framework: Server;
+
+    beforeAll(async () => {
+        sliced = await listen([
+            createJsonBodyParser({ limit: LIMIT, verify: captureRawBody }),
+        ]);
+        // What the server ran before: one parser per content type.
+        framework = await listen([
+            express.json({ limit: LIMIT, verify: captureRawBody }),
+            express.json({
+                limit: LIMIT,
+                type: (req) => req.headers['content-type'] === TEXT_AS_JSON,
+                verify: captureRawBody,
+            }),
+        ]);
+    });
+
+    afterAll(() => {
+        sliced?.close();
+        framework?.close();
+    });
+
+    const post = async (
+        server: Server,
+        body: string | Buffer | undefined,
+        headers: Record<string, string>,
+    ) => {
+        const { port } = server.address() as AddressInfo;
+        const response = await fetch(`http://127.0.0.1:${port}/`, {
+            method: 'POST',
+            headers,
+            body,
+        });
+        return { status: response.status, answer: await response.json() };
+    };
+
+    const large = `${JSON.stringify({
+        text: 'é😀 "quoted" \\ line\n'.repeat(SLICE_BYTES / 8),
+        list: Array.from({ length: 20_000 }, (_, i) => ({
+            i,
+            on: i % 2 === 0,
+        })),
+    }).slice(0, -1)},"__proto__":{"nested":true}}`;
+
+    it.each<[string, string | Buffer | undefined, Record<string, string>]>([
+        ['an object', '{"a":1}', { 'Content-Type': 'application/json' }],
+        ['an array', '[1,2]', { 'Content-Type': 'application/json' }],
+        [
+            'a body sent as text',
+            '{"interface":"x","auth_token":"t"}',
+            { 'Content-Type': TEXT_AS_JSON },
+        ],
+        ['a large body', large, { 'Content-Type': 'application/json' }],
+        ['a large body sent as text', large, { 'Content-Type': TEXT_AS_JSON }],
+        [
+            'a charset parameter',
+            large,
+            { 'Content-Type': 'application/json; charset=UTF-8' },
+        ],
+        [
+            'a quoted charset parameter',
+            '{"a":"é"}',
+            { 'Content-Type': 'application/json;charset="utf-8"' },
+        ],
+        [
+            'another parameter',
+            '{"a":1}',
+            { 'Content-Type': 'application/json; profile=x' },
+        ],
+        ['an empty body', '', { 'Content-Type': 'application/json' }],
+        ['no body', undefined, { 'Content-Type': 'application/json' }],
+        [
+            'a byte order mark',
+            `\ufeff${large}`,
+            { 'Content-Type': 'application/json' },
+        ],
+        [
+            'only a byte order mark',
+            '\ufeff',
+            { 'Content-Type': 'application/json' },
+        ],
+        ['only whitespace', '  \n', { 'Content-Type': 'application/json' }],
+        ['a bare string', '"text"', { 'Content-Type': 'application/json' }],
+        ['a bare number', ' 12', { 'Content-Type': 'application/json' }],
+        ['null', 'null', { 'Content-Type': 'application/json' }],
+        ['malformed JSON', '{"a":', { 'Content-Type': 'application/json' }],
+        [
+            'malformed large JSON',
+            `${large.slice(0, -1)},]`,
+            { 'Content-Type': 'application/json' },
+        ],
+        [
+            'a large bare string',
+            JSON.stringify('x'.repeat(4 * SLICE_BYTES)),
+            { 'Content-Type': 'application/json' },
+        ],
+        [
+            'invalid UTF-8',
+            Buffer.concat([
+                Buffer.from('{"a":"'),
+                Buffer.from([0xff, 0xe2, 0x82]),
+                Buffer.from('"}'),
+            ]),
+            { 'Content-Type': 'application/json' },
+        ],
+        [
+            'a body over the limit',
+            `["${'x'.repeat(LIMIT)}"]`,
+            { 'Content-Type': 'application/json' },
+        ],
+        [
+            'a gzipped body',
+            gzipSync(large),
+            { 'Content-Type': 'application/json', 'Content-Encoding': 'gzip' },
+        ],
+        [
+            'a deflated body',
+            deflateSync(large),
+            {
+                'Content-Type': 'application/json',
+                'Content-Encoding': 'deflate',
+            },
+        ],
+        [
+            'a brotli body',
+            brotliCompressSync(large),
+            { 'Content-Type': 'application/json', 'Content-Encoding': 'br' },
+        ],
+        [
+            'a body that inflates past the limit',
+            gzipSync(`["${'x'.repeat(LIMIT)}"]`),
+            { 'Content-Type': 'application/json', 'Content-Encoding': 'gzip' },
+        ],
+        [
+            'an unknown content encoding',
+            '{"a":1}',
+            { 'Content-Type': 'application/json', 'Content-Encoding': 'lzw' },
+        ],
+        ['plain text', '{"a":1}', { 'Content-Type': 'text/plain' }],
+        [
+            'text with the marker and a charset',
+            '{"a":1}',
+            { 'Content-Type': `${TEXT_AS_JSON}; charset=utf-8` },
+        ],
+        [
+            'a JSON-suffixed type',
+            '{"a":1}',
+            { 'Content-Type': 'application/vnd.api+json' },
+        ],
+        [
+            'form data',
+            'a=1',
+            { 'Content-Type': 'application/x-www-form-urlencoded' },
+        ],
+        ['no content type', '{"a":1}', { 'Content-Type': '' }],
+    ])('answers %s the way express.json does', async (_name, body, headers) => {
+        const got = await post(sliced, body, headers);
+        const want = await post(framework, body, headers);
+        expect(got).toEqual(want);
+    });
+
+    it('leaves a body in another charset unread', async () => {
+        for (const charset of ['utf-16', 'utf8', 'latin1']) {
+            const { status, answer } = await post(
+                sliced,
+                Buffer.from('{"a":1}', 'utf16le'),
+                { 'Content-Type': `application/json; charset=${charset}` },
+            );
+            expect(status).toBe(200);
+            expect(answer.hasBody).toBe(false);
+        }
+        // The last charset named is the one that counts.
+        const { answer } = await post(sliced, '{"a":1}', {
+            'Content-Type': 'application/json; charset=utf-8; charset=utf-16',
+        });
+        expect(answer.hasBody).toBe(false);
+    });
+
+    it('parses bodies that arrive together, each to its own result', async () => {
+        const bodies = Array.from({ length: 6 }, (_, n) =>
+            JSON.stringify({ n, pad: 'x'.repeat((n + 1) * SLICE_BYTES) }),
+        );
+        const answers = await Promise.all(
+            bodies.map((body) =>
+                post(sliced, body, { 'Content-Type': 'application/json' }),
+            ),
+        );
+        answers.forEach(({ status, answer }, n) => {
+            expect(status).toBe(200);
+            expect(answer.body).toEqual(JSON.parse(bodies[n]!));
+        });
+        // A failed parse hands its turn on like a finished one.
+        const mixed = await Promise.all([
+            post(sliced, `${bodies[5]!},`, {
+                'Content-Type': 'application/json',
+            }),
+            post(sliced, bodies[4], { 'Content-Type': 'application/json' }),
+            post(sliced, bodies[3], { 'Content-Type': 'application/json' }),
+        ]);
+        expect(mixed.map(({ status }) => status)).toEqual([400, 200, 200]);
+    });
+});

src/backend/core/http/middleware/jsonBody.ts

@@ -0,0 +1,127 @@
+/*
+ * Copyright (C) 2024-present Puter Technologies Inc.
+ *
+ * This file is part of Puter.
+ *
+ * Puter is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU Affero General Public License as published
+ * by the Free Software Foundation, either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU Affero General Public License for more details.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program.  If not, see <https://www.gnu.org/licenses/>.
+ */
+
+import express, { type Request, type RequestHandler } from 'express';
+import { parseJsonInSlices, SLICE_BYTES } from '../../../util/slicedJson';
+
+/** What puter.js labels a JSON driver call, to keep it a simple CORS request. */
+const TEXT_AS_JSON = 'text/plain;actually=json';
+
+const UTF8_CHARSET = /;\s*charset=("?)utf-8\1\s*(?:;|$)/i;
+
+/** No charset, or exactly UTF-8. Anything else is left to `express.json`. */
+const declaresUtf8 = (contentType: string): boolean => {
+    const mentions = contentType.match(/charset/gi)?.length ?? 0;
+    return mentions === 0 || (mentions === 1 && UTF8_CHARSET.test(contentType));
+};
+
+const isUtf8Json = (req: Request): boolean => {
+    const contentType = req.headers['content-type'];
+    if (contentType === TEXT_AS_JSON) return true;
+    return (
+        typeof contentType === 'string' &&
+        Boolean(req.is('application/json')) &&
+        declaresUtf8(contentType)
+    );
+};
+
+/** The error `express.json` raises, which the error handler answers with a 400. */
+const parseFailed = (message: string): Error =>
+    Object.assign(new SyntaxError(message), {
+        status: 400,
+        statusCode: 400,
+        expose: true,
+        type: 'entity.parse.failed',
+    });
+
+const waiting: Array<() => void> = [];
+let parsing = false;
+
+/**
+ * Bodies too large for one `JSON.parse` call parse one at a time, so a single
+ * tree is under construction however many arrive together.
+ */
+const parseInTurn = async (body: Buffer): Promise<unknown> => {
+    if (body.length <= SLICE_BYTES) return parseJsonInSlices(body);
+    if (parsing) await new Promise<void>((resolve) => waiting.push(resolve));
+    parsing = true;
+    try {
+        return await parseJsonInSlices(body);
+    } finally {
+        // Handing over keeps `parsing` set, so a new arrival can't cut in.
+        const next = waiting.shift();
+        if (next) next();
+        else parsing = false;
+    }
+};
+
+interface JsonBodyOptions {
+    limit: number;
+    verify?: NonNullable<Parameters<typeof express.raw>[0]>['verify'];
+}
+
+/**
+ * `express.json` for UTF-8 bodies, with the parse done in slices so a large
+ * body can't hold the event loop. Reading, limits, decompression and `verify`
+ * are the framework's; the result and the errors match `express.json`.
+ */
+export const createJsonBodyParser = ({
+    limit,
+    verify,
+}: JsonBodyOptions): RequestHandler => {
+    const read = express.raw({
+        limit,
+        verify,
+        type: (req) => isUtf8Json(req as Request),
+    });
+
+    return (req, res, next) => {
+        read(req, res, async (err?: unknown) => {
+            // Not a Buffer: the reader skipped this request.
+            if (err || !Buffer.isBuffer(req.body)) {
+                next(err);
+                return;
+            }
+            let body: Buffer = req.body;
+            req.body = undefined;
+            // The framework's decoder drops a byte order mark.
+            if (body[0] === 0xef && body[1] === 0xbb && body[2] === 0xbf) {
+                body = body.subarray(3);
+            }
+            if (body.length === 0) {
+                req.body = {};
+                next();
+                return;
+            }
+            let parsed: unknown;
+            try {
+                parsed = await parseInTurn(body);
+            } catch (e) {
+                next(parseFailed((e as Error).message));
+                return;
+            }
+            if (typeof parsed !== 'object' || parsed === null) {
+                next(parseFailed('JSON body must be an object or an array'));
+                return;
+            }
+            req.body = parsed;
+            next();
+        });
+    };
+};

src/backend/core/http/middleware/largeJsonBody.http.test.ts

@@ -0,0 +1,184 @@
+/*
+ * Copyright (C) 2024-present Puter Technologies Inc.
+ *
+ * This file is part of Puter.
+ *
+ * Puter is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU Affero General Public License as published
+ * by the Free Software Foundation, either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU Affero General Public License for more details.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program.  If not, see <https://www.gnu.org/licenses/>.
+ */
+
+import { request, type Server } from 'node:http';
+import type { AddressInfo } from 'node:net';
+import { setImmediate as nextTurn } from 'node:timers/promises';
+import express, {
+    type ErrorRequestHandler,
+    type RequestHandler,
+} from 'express';
+import { afterAll, beforeAll, describe, expect, it } from 'vitest';
+import { createJsonBodyParser } from './jsonBody';
+import {
+    createLargeJsonBodyParser,
+    DEFAULT_JSON_LIMIT_BYTES,
+} from './largeJsonBody';
+
+const MiB = 1024 * 1024;
+const TOKEN = 'valid-token';
+const PAD = 'x'.repeat(DEFAULT_JSON_LIMIT_BYTES);
+
+/** Stands in for the auth probe: one token is valid, wherever it is sent. */
+const authProbe: RequestHandler = (req, _res, next) => {
+    const token =
+        (req.body as { auth_token?: string } | undefined)?.auth_token ??
+        req.headers.authorization;
+    if (!req.actor && token === TOKEN) {
+        req.actor = { user: { uuid: 'u1' } } as typeof req.actor;
+    }
+    next();
+};
+
+describe('a JSON body over the default limit', () => {
+    let server: Server;
+
+    beforeAll(async () => {
+        const app = express();
+        app.post(
+            '/call',
+            createLargeJsonBodyParser({
+                authProbe,
+                limitBytes: 128 * MiB,
+                budgetBytes: 256 * MiB,
+                perUserBytes: 128 * MiB,
+            }),
+        );
+        app.use(createJsonBodyParser({ limit: DEFAULT_JSON_LIMIT_BYTES }));
+        app.use(authProbe);
+        app.post('/call', (req, res) => {
+            const body = req.body as { auth_token?: string; pad: string };
+            res.json({
+                user: req.actor?.user.uuid ?? null,
+                token: body.auth_token ?? null,
+                padLength: body.pad.length,
+            });
+        });
+        const onError: ErrorRequestHandler = (err, _req, res, _next) => {
+            res.status(err.statusCode ?? 500).json({});
+        };
+        app.use(onError);
+        await new Promise<void>((resolve) => {
+            server = app.listen(0, '127.0.0.1', resolve);
+        });
+    });
+
+    afterAll(() => {
+        server?.close();
+    });
+
+    /** Sends `chunks` one write at a time, so the server sees them apart. */
+    const post = (chunks: string[], headers: Record<string, string> = {}) =>
+        new Promise<{ status: number; answer: Record<string, unknown> }>(
+            (resolve, reject) => {
+                const { port } = server.address() as AddressInfo;
+                const req = request(
+                    {
+                        host: '127.0.0.1',
+                        port,
+                        path: '/call',
+                        method: 'POST',
+                        headers: {
+                            'Content-Type': 'text/plain;actually=json',
+                            'Content-Length': chunks.reduce(
+                                (total, chunk) =>
+                                    total + Buffer.byteLength(chunk),
+                                0,
+                            ),
+                            ...headers,
+                        },
+                    },
+                    (res) => {
+                        const parts: Buffer[] = [];
+                        res.on('data', (part: Buffer) => parts.push(part));
+                        res.on('end', () =>
+                            resolve({
+                                status: res.statusCode!,
+                                answer: JSON.parse(
+                                    Buffer.concat(parts).toString(),
+                                ),
+                            }),
+                        );
+                    },
+                );
+                req.on('error', reject);
+                void (async () => {
+                    for (const chunk of chunks) {
+                        req.write(chunk);
+                        await nextTurn();
+                    }
+                    req.end();
+                })();
+            },
+        );
+
+    it('is accepted when it opens with a valid auth_token', async () => {
+        const { status, answer } = await post([
+            `{"auth_token":"${TOKEN}","pad":"${PAD}"}`,
+        ]);
+        expect(status).toBe(200);
+        expect(answer).toEqual({
+            user: 'u1',
+            token: TOKEN,
+            padLength: PAD.length,
+        });
+    }, 60_000);
+
+    it('is accepted when the token arrives in pieces, with spacing', async () => {
+        const opening = ` {\n  "auth_token" : "${TOKEN}" ,`;
+        const { status, answer } = await post([...opening, `"pad":"${PAD}"}`]);
+        expect(status).toBe(200);
+        expect(answer.user).toBe('u1');
+    }, 60_000);
+
+    it('is accepted with a header credential, as before', async () => {
+        const { status, answer } = await post([`{"pad":"${PAD}"}`], {
+            Authorization: TOKEN,
+        });
+        expect(status).toBe(200);
+        expect(answer).toEqual({
+            user: 'u1',
+            token: null,
+            padLength: PAD.length,
+        });
+    }, 60_000);
+
+    it.each([
+        [
+            'its token comes after other fields',
+            `{"pad":"${PAD}","auth_token":"${TOKEN}"}`,
+        ],
+        ['its token is not valid', `{"auth_token":"expired","pad":"${PAD}"}`],
+        ['it has no token', `{"pad":"${PAD}"}`],
+        [
+            'a later field replaces the token it opens with',
+            `{"auth_token":"${TOKEN}","pad":"${PAD}","auth_token":"other"}`,
+        ],
+        [
+            'the token it opens with is nested',
+            `{"args":{"auth_token":"${TOKEN}"},"pad":"${PAD}"}`,
+        ],
+    ])(
+        'is refused with 413 when %s',
+        async (_name, body) => {
+            expect((await post([body])).status).toBe(413);
+        },
+        60_000,
+    );
+});

src/backend/core/http/middleware/largeJsonBody.test.ts

@@ -18,6 +18,7 @@
  */
 
 import { EventEmitter } from 'node:events';
+import { Readable } from 'node:stream';
 import type { NextFunction, Request, RequestHandler, Response } from 'express';
 import { describe, expect, it, vi } from 'vitest';
 import { isHttpError } from '../HttpError';
@@ -35,7 +36,7 @@ const makeReq = (
     user: string | null = 'u1',
     body?: unknown,
 ) =>
-    ({
+    Object.assign(Readable.from([]), {
         complete: true,
         body,
         headers: {

src/backend/core/http/middleware/largeJsonBody.ts

@@ -17,13 +17,54 @@
  * along with this program.  If not, see <https://www.gnu.org/licenses/>.
  */
 
-import express, { type Request, type RequestHandler } from 'express';
+import type { Request, RequestHandler, Response } from 'express';
 import { actorUid } from '../../actor';
 import { HttpError } from '../HttpError';
+import { createJsonBodyParser } from './jsonBody';
 
 /** The global JSON parser's limit, which applies before auth. */
 export const DEFAULT_JSON_LIMIT_BYTES = 50 * 1024 * 1024;
 
+/** How much of a body is read to find a token it opens with. */
+const LEADING_TOKEN_BYTES = 8 * 1024;
+const LEADING_TOKEN =
+    /^[ \t\n\r]*\{[ \t\n\r]*"auth_token"[ \t\n\r]*:[ \t\n\r]*"([^"\\]+)"/;
+
+/**
+ * The first `bytes` of the body, put back for the parser to read. Empty when
+ * the body ends or fails before that many arrive.
+ */
+const peekBody = (req: Request, bytes: number): Promise<Buffer> =>
+    new Promise((resolve) => {
+        const settle = (head: Buffer | null) => {
+            req.off('readable', onReadable);
+            req.off('end', onGone);
+            req.off('close', onGone);
+            req.off('error', onGone);
+            if (head) req.unshift(head);
+            resolve(head ?? Buffer.alloc(0));
+        };
+        const onReadable = () => {
+            // Null until `bytes` are buffered.
+            const head = req.read(bytes) as Buffer | null;
+            if (head) settle(head);
+        };
+        const onGone = () => settle(null);
+        req.on('readable', onReadable);
+        req.on('end', onGone);
+        req.on('close', onGone);
+        req.on('error', onGone);
+    });
+
+/**
+ * The `auth_token` a JSON body opens with, which puter.js sends first so it can
+ * be read ahead of the body.
+ */
+const leadingBodyToken = async (req: Request): Promise<string | undefined> => {
+    const head = await peekBody(req, LEADING_TOKEN_BYTES);
+    return LEADING_TOKEN.exec(head.toString('latin1'))?.[1];
+};
+
 interface LargeJsonBodyOptions {
     /** The global auth probe; it skips requests that already carry an actor. */
     authProbe: RequestHandler;
@@ -38,9 +79,9 @@ interface LargeJsonBodyOptions {
 
 /**
  * Parses JSON bodies over the default limit, up to `limitBytes`, once the
- * caller's header or cookie credential resolves. Unauthenticated requests and
- * requests without a Content-Length fall through to the global parser and its
- * limit.
+ * caller's credential resolves: a header, a cookie, or an `auth_token` the body
+ * opens with. Unauthenticated requests and requests without a Content-Length
+ * fall through to the global parser and its limit.
  *
  * The parsed body lives as long as the request, so a body's bytes count against
  * both budgets until its response closes. Over the user's budget gets a 429,
@@ -53,82 +94,91 @@ export const createLargeJsonBodyParser = ({
     perUserBytes,
     accepts,
 }: LargeJsonBodyOptions): RequestHandler => {
-    const json = express.json({ limit: limitBytes });
-    const textAsJson = express.json({
-        limit: limitBytes,
-        type: (req) =>
-            req.headers['content-type'] === 'text/plain;actually=json',
-    });
+    const json = createJsonBodyParser({ limit: limitBytes });
     let inFlightBytes = 0;
     const inFlightByUser = new Map<string, number>();
 
-    return (req, res, next) => {
+    const probe = (req: Request, res: Response) =>
+        new Promise<void>((resolve) => {
+            void authProbe(req, res, () => resolve());
+        });
+
+    return async (req, res, next) => {
         const declared = Number(req.headers['content-length']);
         if (!(declared > DEFAULT_JSON_LIMIT_BYTES && declared <= limitBytes)) {
             next();
             return;
         }
-        void authProbe(req, res, () => {
-            // `res.closed`: the client left during the probe, so `close` has
-            // already fired and would never release a charge.
-            if (!req.actor || res.closed) {
-                next();
+        await probe(req, res);
+        let bodyToken: string | undefined;
+        if (!req.actor && !res.closed) {
+            bodyToken = await leadingBodyToken(req);
+            if (bodyToken) {
+                // Where the probe looks for a body token.
+                req.body = { auth_token: bodyToken };
+                await probe(req, res);
+                req.body = undefined;
+            }
+        }
+        // `res.closed`: the client left during the probe, so `close` has
+        // already fired and would never release a charge.
+        if (!req.actor || res.closed) {
+            next();
+            return;
+        }
+        const user = req.actor.user.uuid ?? actorUid(req.actor);
+        const userBytes = inFlightByUser.get(user) ?? 0;
+        if (userBytes + declared > perUserBytes) {
+            res.setHeader('Retry-After', '5');
+            next(
+                new HttpError(
+                    429,
+                    'Too many large requests in progress for this account',
+                    { legacyCode: 'too_many_requests' },
+                ),
+            );
+            return;
+        }
+        if (inFlightBytes + declared > budgetBytes) {
+            res.setHeader('Retry-After', '5');
+            next(
+                new HttpError(503, 'Too many large requests in progress', {
+                    code: 'large_body_capacity',
+                    // Load shedding, not a fault: record it, don't page.
+                    alarmSeverity: 'info',
+                }),
+            );
+            return;
+        }
+        inFlightBytes += declared;
+        inFlightByUser.set(user, userBytes + declared);
+        res.once('close', () => {
+            inFlightBytes -= declared;
+            const left = (inFlightByUser.get(user) ?? 0) - declared;
+            if (left > 0) inFlightByUser.set(user, left);
+            else inFlightByUser.delete(user);
+        });
+        json(req, res, (err?: unknown) => {
+            if (err) {
+                next(err);
                 return;
             }
-            const user = req.actor.user.uuid ?? actorUid(req.actor);
-            const userBytes = inFlightByUser.get(user) ?? 0;
-            if (userBytes + declared > perUserBytes) {
-                res.setHeader('Retry-After', '5');
+            // The caller was resolved from the token the body opens with,
+            // so the parsed body has to name the same one.
+            const tokenMoved =
+                bodyToken !== undefined &&
+                (req.body as { auth_token?: unknown } | undefined)
+                    ?.auth_token !== bodyToken;
+            if (tokenMoved || (accepts && !accepts(req))) {
+                // The same answer the default limit gives.
                 next(
-                    new HttpError(
-                        429,
-                        'Too many large requests in progress for this account',
-                        { legacyCode: 'too_many_requests' },
-                    ),
-                );
-                return;
-            }
-            if (inFlightBytes + declared > budgetBytes) {
-                res.setHeader('Retry-After', '5');
-                next(
-                    new HttpError(503, 'Too many large requests in progress', {
-                        code: 'large_body_capacity',
-                        // Load shedding, not a fault: record it, don't page.
-                        alarmSeverity: 'info',
+                    new HttpError(413, 'request entity too large', {
+                        legacyCode: 'client_bad_request',
                     }),
                 );
                 return;
             }
-            inFlightBytes += declared;
-            inFlightByUser.set(user, userBytes + declared);
-            res.once('close', () => {
-                inFlightBytes -= declared;
-                const left = (inFlightByUser.get(user) ?? 0) - declared;
-                if (left > 0) inFlightByUser.set(user, left);
-                else inFlightByUser.delete(user);
-            });
-            json(req, res, (err?: unknown) => {
-                if (err) {
-                    next(err);
-                    return;
-                }
-                textAsJson(req, res, (err?: unknown) => {
-                    if (err) {
-                        next(err);
-                        return;
-                    }
-                    if (accepts && !accepts(req)) {
-                        // The same answer the default limit gives.
-                        next(
-                            new HttpError(413, 'request entity too large', {
-                                legacyCode: 'client_bad_request',
-                            }),
-                        );
-                        return;
-                    }
-                    next();
-                });
-            });
+            next();
         });
     };
 };

src/backend/drivers/ai-chat/ChatCompletionDriver.test.ts

@@ -355,6 +355,17 @@ describe('ChatCompletionDriver.complete auth and model resolution', () => {
         ).rejects.toMatchObject({ statusCode: 400 });
     });
 
+    it('rejects a model that is not a string with 400', async () => {
+        await expect(
+            withTestActor(() =>
+                driver.complete({
+                    messages: [{ role: 'user', content: 'hi' }],
+                    model: { id: 'gpt' },
+                } as unknown as ICompleteArguments),
+            ),
+        ).rejects.toMatchObject({ statusCode: 400 });
+    });
+
     it('routes by alias when a provider exposes one (puterId is auto-aliased on the bucket)', async () => {
         // Inject a model that carries a `puterId` so we can verify the
         // alias bucket resolves back to the canonical id. Spy on
@@ -2222,3 +2233,166 @@ describe('ChatCompletionDriver.countTokens', () => {
         ).rejects.toMatchObject({ statusCode: 400 });
     });
 });
+
+describe('ChatCompletionDriver service speed', () => {
+    it.each([
+        [{ serviceSpeed: 'ultrafast' }, 'fast', 0.028],
+        [{ speed: 'fast' }, 'fast', 0.028],
+        [{ serviceSpeed: 'standard', speed: 'fast' }, 'standard', 0.014],
+    ] as const)('uses the Claude catalog rate card for %j', async (options, speed, cents) => {
+        const sent: Record<string, unknown>[] = [];
+        vi.stubGlobal('fetch', async (input: Request | string, init?: RequestInit) => {
+            const body = input instanceof Request
+                ? await input.clone().json()
+                : JSON.parse(String(init?.body));
+            sent.push(body);
+            return new Response(JSON.stringify({
+                id: 'claude-speed-test', type: 'message', role: 'assistant',
+                model: body.model, stop_reason: 'end_turn',
+                content: [{ type: 'text', text: 'OK' }],
+                usage: { input_tokens: 10, output_tokens: 5, speed },
+            }), { headers: { 'content-type': 'application/json' } });
+        });
+        const d = new ChatCompletionDriver(
+            { providers: { claude: { apiKey: 'test-key' }, ollama: { enabled: false } } } as never,
+            server.clients, server.stores, server.services,
+        );
+        try {
+            await d.onServerStart();
+            const result = await withTestActor(() => d.complete({
+                model: 'claude-opus-5-5', provider: 'claude',
+                messages: [{ role: 'user', content: 'Hello' }],
+                fallback: false, max_tokens: 32, normalize: true, ...options,
+            }));
+            expect(sent).toHaveLength(1);
+            expect(sent[0]!.speed).toBe(speed === 'fast' ? 'fast' : undefined);
+            expect(result).toMatchObject({
+                message: { content: 'OK' }, usage: { usd_cents: cents },
+            });
+        } finally {
+            d.onServerShutdown();
+            vi.unstubAllGlobals();
+        }
+    });
+
+    it.each([
+        null, false, true, 0, 1, -1, NaN, Infinity, '', 'invalid', 'FAST',
+        'fast ', ' fast', 'ultraFast', [], ['fast'], {}, { speed: 'fast' },
+        { toString: 'fast' },
+    ])('rejects invalid speed %j before sending a request', async (serviceSpeed) => {
+        await expect(
+            withTestActor(() =>
+                driver.complete({
+                    model: 'fake',
+                    messages: [{ role: 'user', content: 'Hello' }],
+                    serviceSpeed: serviceSpeed as never,
+                }),
+            ),
+        ).rejects.toMatchObject({
+            statusCode: 400,
+            legacyCode: 'invalid_service_speed',
+        });
+    });
+
+    it.each([
+        ['openai-responses', true],
+        ['openai-responses', false],
+        ['openai-completion', false],
+    ] as const)(
+        'settles unavailable tiers at standard rates on %s with provider fallback disabled (admission rejection: %s)',
+        async (provider, rejectFast) => {
+            const sent: Record<string, unknown>[] = [];
+            vi.stubGlobal(
+                'fetch',
+                async (_input: unknown, init?: RequestInit) => {
+                    const body =
+                        _input instanceof Request
+                            ? await _input.clone().json()
+                            : JSON.parse(String(init?.body));
+                    sent.push(body);
+                    if (rejectFast && body.service_tier !== 'default') {
+                        return new Response(
+                            JSON.stringify({
+                                error: {
+                                    message: `${body.service_tier} service_tier capacity unavailable`,
+                                    type: 'invalid_request_error',
+                                },
+                            }),
+                            {
+                                status: 400,
+                                headers: { 'content-type': 'application/json' },
+                            },
+                        );
+                    }
+                    return new Response(
+                        JSON.stringify({
+                            id: 'response-test',
+                            object: 'response',
+                            model: body.model,
+                            status: 'completed',
+                            service_tier: 'default',
+                            output: [
+                                {
+                                    type: 'message',
+                                    role: 'assistant',
+                                    content: [
+                                        { type: 'output_text', text: 'OK' },
+                                    ],
+                                },
+                            ],
+                            usage: {
+                                input_tokens: 10,
+                                output_tokens: 2,
+                                input_tokens_details: { cached_tokens: 0 },
+                            },
+                        }),
+                        { headers: { 'content-type': 'application/json' } },
+                    );
+                },
+            );
+            const d = new ChatCompletionDriver(
+                {
+                    providers: {
+                        'openai-completion': { apiKey: 'test-key' },
+                        ollama: { enabled: false },
+                    },
+                } as never,
+                server.clients,
+                server.stores,
+                server.services,
+            );
+            try {
+                await d.onServerStart();
+                const result = (await withTestActor(() =>
+                    d.complete({
+                        model: 'gpt-6.1-sol',
+                        provider,
+                        messages: [{ role: 'user', content: 'Hello' }],
+                        serviceSpeed: 'ultrafast',
+                        fallback: false,
+                        max_tokens: 32,
+                        normalize: true,
+                    }),
+                )) as {
+                    message: { content: string };
+                    usageDetails: { serviceSpeed: string };
+                    usage: { usd_cents: number };
+                };
+                expect(sent.map((body) => body.service_tier)).toEqual(
+                    rejectFast
+                        ? ['ultrafast', 'priority', 'default']
+                        : ['ultrafast'],
+                );
+                expect(sent.every((body) => body.model === 'gpt-6.1-sol')).toBe(
+                    true,
+                );
+                expect(result.message.content).toBe('OK');
+                expect(result.usageDetails.serviceSpeed).toBe('standard');
+                expect(result.usage.usd_cents).toBeCloseTo(0.004, 8);
+            } finally {
+                d.onServerShutdown();
+                vi.unstubAllGlobals();
+            }
+        },
+    );
+});

src/backend/drivers/ai-chat/ChatCompletionDriver.ts

@@ -78,6 +78,11 @@ import type {
     UsageDetails,
 } from './types.js';
 import { normalize_tools_object } from './utils/FunctionCalling.js';
+import {
+    selectServiceSpeed,
+    lowerServiceSpeed,
+    isServiceSpeedUnavailable,
+} from './utils/serviceSpeed.js';
 import {
     messagesHavePuterPaths,
     modelSupportsModality,
@@ -674,6 +679,16 @@ export class ChatCompletionDriver extends PuterDriver {
     }
 
     async complete(args: ICompleteArguments): Promise<IChatCompleteResult> {
+        if (
+            args.serviceSpeed !== undefined &&
+            !['standard', 'fast', 'ultrafast'].includes(args.serviceSpeed)
+        ) {
+            throw new HttpError(
+                400,
+                'serviceSpeed must be standard, fast, or ultrafast',
+                { legacyCode: 'invalid_service_speed' },
+            );
+        }
         const abortSignal = Context.get('abortSignal');
         abortSignal?.throwIfAborted();
         const actor = Context.get('actor');
@@ -700,15 +715,16 @@ export class ChatCompletionDriver extends PuterDriver {
             args.model = this.#providers[intendedProvider].getDefaultModel();
         }
 
-        let model = await this.#resolveLoadedModel(
+        const initialModel = await this.#resolveLoadedModel(
             args.model,
             intendedProvider,
         );
-        if (!model) {
+        if (!initialModel) {
             throw new HttpError(400, `Model not found: ${args.model}`, {
                 legacyCode: 'bad_request',
             });
         }
+        let model: IChatModel = initialModel;
         // A pinned provider without a route for the model normally resolves
         // to another vendor; with `fallback: false` that is the one thing the
         // caller asked us not to do.
@@ -840,18 +856,6 @@ export class ChatCompletionDriver extends PuterDriver {
         // The instant the current attempt is priced at, from its hold to its
         // charge, so a rate change mid-request can't split the two.
         let pricingTime = new Date();
-        if (!useFakeProvider) {
-            ({
-                hold,
-                fundsCap,
-                costFactor,
-                promptTokenEstimate: attemptPromptTokenEstimate,
-            } = await this.#applyCreditGate(actor, model, args, {
-                promptTokenEstimate,
-                requestedMaxTokens,
-                pricingTime,
-            }));
-        }
 
         // First attempt
         const provider = this.#providers[model.provider!];
@@ -899,24 +903,134 @@ export class ChatCompletionDriver extends PuterDriver {
             return attempt;
         };
 
+        let variantSpeed: ICompleteArguments['serviceSpeed'];
+        let creditGateError: unknown;
         const runAttempt = async (
             attemptProvider: IChatProvider,
-            attemptModel: IChatModel,
+            baseModel: IChatModel,
         ) => {
-            if (!useFakeProvider) {
-                await this.#resolvePuterPaths(attemptProvider, args, actor);
-            }
-            // Last point where a caller who left costs nothing.
-            abortSignal?.throwIfAborted();
-            return attemptProvider.complete(
-                {
-                    ...args,
+            const offers =
+                args.serviceSpeed !== undefined &&
+                args.serviceSpeed !== 'standard'
+                    ? ((await attemptProvider.serviceSpeedModels?.(
+                          baseModel,
+                      )) ?? {})
+                    : {};
+            let requestedSpeed =
+                args.serviceSpeed ??
+                (baseModel.provider === 'claude' && args.speed === 'fast'
+                    ? 'fast'
+                    : undefined);
+            for (;;) {
+                const selection = selectServiceSpeed(
+                    baseModel,
+                    { ...args, serviceSpeed: requestedSpeed },
+                    (id) => {
+                        const variant = this.#resolveModel(
+                            id,
+                            baseModel.provider!,
+                        );
+                        return variant &&
+                            variant.provider === baseModel.provider
+                            ? variant
+                            : undefined;
+                    },
+                    offers,
+                );
+                const attemptModel = selection.model;
+                const requestArgs = {
+                    ...selection.args,
                     model: attemptModel.id,
                     provider: attemptModel.provider,
-                },
-                this.#catalogEntries.get(attemptModel),
-                pricingTime,
-            );
+                };
+                const catalogModel = this.#resolveModel(
+                    attemptModel.id,
+                    attemptModel.provider!,
+                );
+                const rawModel = this.#catalogEntries.get(
+                    catalogModel ?? baseModel,
+                );
+                const snapshot = rawModel
+                    ? {
+                          ...rawModel,
+                          costs: attemptModel.costs,
+                          pricingOverrides: attemptModel.pricingOverrides,
+                          serviceSpeedPricing: attemptModel.serviceSpeedPricing,
+                      }
+                    : attemptModel;
+                if (!useFakeProvider) {
+                    try {
+                        ({
+                            hold,
+                            fundsCap,
+                            costFactor,
+                            promptTokenEstimate: attemptPromptTokenEstimate,
+                        } = await this.#applyCreditGate(
+                            actor,
+                            attemptModel,
+                            requestArgs,
+                            {
+                                promptTokenEstimate,
+                                requestedMaxTokens,
+                                pricingTime,
+                            },
+                        ));
+                    } catch (error) {
+                        creditGateError = error;
+                        throw error;
+                    }
+                    await this.#resolvePuterPaths(
+                        attemptProvider,
+                        requestArgs,
+                        actor,
+                    );
+                }
+                abortSignal?.throwIfAborted();
+                try {
+                    const result = await attemptProvider.complete(
+                        requestArgs,
+                        snapshot,
+                        pricingTime,
+                    );
+                    model = attemptModel;
+                    variantSpeed =
+                        attemptModel.id !== baseModel.id
+                            ? selection.speed
+                            : undefined;
+                    if (
+                        args.serviceSpeed !== undefined &&
+                        'message' in result
+                    ) {
+                        const messageResult = result as IChatMessageResult;
+                        const details =
+                            messageResult.usageDetails ??
+                            usageDetailsFromUsage(messageResult.usage, model);
+                        messageResult.usageDetails = {
+                            ...details,
+                            serviceSpeed:
+                                variantSpeed ??
+                                details?.serviceSpeed ??
+                                details?.speed ??
+                                'standard',
+                        };
+                    }
+                    return result;
+                } catch (error) {
+                    await hold.release();
+                    hold = NO_CREDIT_HOLD;
+                    if (
+                        args.serviceSpeed === undefined ||
+                        selection.speed === 'standard' ||
+                        !isServiceSpeedUnavailable(
+                            error,
+                            attemptModel.id !== baseModel.id,
+                        )
+                    )
+                        throw error;
+                    requestedSpeed = lowerServiceSpeed(selection.speed);
+                    pricingTime = new Date();
+                }
+            }
         };
 
         try {
@@ -984,29 +1098,17 @@ export class ChatCompletionDriver extends PuterDriver {
                 // The previous attempt released its hold when it failed, so
                 // this one starts from nothing held.
                 pricingTime = new Date();
-                if (!useFakeProvider) {
-                    ({
-                        hold,
-                        fundsCap,
-                        costFactor,
-                        promptTokenEstimate: attemptPromptTokenEstimate,
-                    } = await this.#applyCreditGate(actor, fallback, args, {
-                        promptTokenEstimate,
-                        requestedMaxTokens,
-                        pricingTime,
-                    }));
-                }
 
                 tried.add(routeId(fallback.provider!, fallback.id));
 
                 try {
                     res = await runAttempt(fbProvider, fallback);
-                    model = fallback;
                     lastError = null;
                     console.warn(
                         `[ai-chat] fallback served (${completionId}, ${first.provider}:${first.model} -> ${fallback.provider}:${fallback.id}, first ${first.status ?? 'none'}${first.upstreamType ? `/${first.upstreamType}` : ''})`,
                     );
                 } catch (fbErr) {
+                    if (fbErr === creditGateError) throw fbErr;
                     await hold.release();
                     hold = NO_CREDIT_HOLD;
                     abortSignal?.throwIfAborted();
@@ -1162,6 +1264,20 @@ export class ChatCompletionDriver extends PuterDriver {
                         usageDetailsFromUsage(usage, model),
                     );
                 }
+                if (
+                    args.serviceSpeed !== undefined &&
+                    chatStream.usageDetails
+                ) {
+                    const details = chatStream.usageDetails;
+                    chatStream.setUsageDetails({
+                        ...details,
+                        serviceSpeed:
+                            variantSpeed ??
+                            details.serviceSpeed ??
+                            details.speed ??
+                            'standard',
+                    });
+                }
                 withClaudeThinkingTokens(
                     enrichedUsage,
                     model,
@@ -1474,12 +1590,7 @@ export class ChatCompletionDriver extends PuterDriver {
         const metering = this.services.metering;
         const { requestedMaxTokens, pricingTime } = estimates;
         let { promptTokenEstimate } = estimates;
-        // A provider can name its own cost keys for this request (Claude's
-        // fast-mode rate card), cap its output, and size an extra reservation
-        // the per-token math can't see (a web-search/advisor tool that could
-        // run several times before the model answers) — the same clamps it
-        // applies when building the request, so the hold and the request
-        // agree.
+        // Providers reserve tool calls using the same caps they send upstream.
         const pricing =
             this.#providers[model.provider!]?.requestPricing?.(args, model, {
                 promptTokenEstimate,
@@ -2019,7 +2130,9 @@ export class ChatCompletionDriver extends PuterDriver {
     }
 
     #resolveModel(modelId: string, provider?: string): IChatModel | null {
-        const models = this.#modelIdMap[normalizeModelKey(modelId ?? '')];
+        // `model` comes straight from the caller and may not be a string.
+        if (typeof modelId !== 'string') return null;
+        const models = this.#modelIdMap[normalizeModelKey(modelId)];
         if (!models || models.length === 0) return null;
         // An explicitly requested provider is honoured even if its route is
         // marked — the caller asked for that one, not for the cheapest hop.

src/backend/drivers/ai-chat/providers/OpenAICompatProvider.ts

@@ -41,6 +41,7 @@ import {
     meterChatUsage,
     type MeterChatUsageOptions,
 } from '../utils/meterChatUsage.js';
+import { servedServiceSpeed } from '../utils/serviceSpeed.js';
 import { catalogPrice, costKeys } from '../utils/pricing.js';
 import { normalizeModelKey } from '../utils/modelRouting.js';
 import * as OpenAIUtil from '../utils/OpenAIUtil.js';
@@ -232,6 +233,8 @@ export class OpenAICompatProvider implements IChatProvider {
                   )
                 : {}),
         };
+        if (args.serviceSpeed !== undefined && args.service_tier !== undefined)
+            shared.service_tier = args.service_tier;
         const params = this.vendorParams(shared, args, model, actor);
 
         let pricingTime = attemptPricingTime ?? new Date();
@@ -271,15 +274,29 @@ export class OpenAICompatProvider implements IChatProvider {
                     source,
                     model,
                 );
-                if (usageDetails) source.setUsageDetails?.(usageDetails);
+                const serviceSpeed = servedServiceSpeed(source);
+                if (usageDetails)
+                    source.setUsageDetails?.({
+                        ...usageDetails,
+                        ...(args.serviceSpeed !== undefined
+                            ? { serviceSpeed }
+                            : {}),
+                    });
                 const metered = meterChatUsage(
                     this.#metering,
                     actor,
                     this.meteringModelKey(model.id),
                     model,
                     usage,
-                    { ...meterOptions, pricingTime },
+                    { ...meterOptions, pricingTime, serviceSpeed },
                 );
+                if (args.serviceSpeed !== undefined && !usageDetails) {
+                    const details = OpenAIUtil.usageDetailsFromTrackedUsage(
+                        metered.usage,
+                    );
+                    if (details)
+                        source.setUsageDetails?.({ ...details, serviceSpeed });
+                }
                 source.setUsageCosts(metered.costs);
                 return this.reportedUsage(metered.usage);
             },

src/backend/drivers/ai-chat/providers/alibaba/catalog.ts

@@ -174,6 +174,9 @@ export const mapAlibabaCatalogModel = (
             : undefined;
     return {
         ...existing,
+        serviceSpeedVariants: {
+            fast: id === 'glm-5.2' ? 'glm-5.2-fast-preview' : `${id}-prime`,
+        },
         id,
         puterId,
         name: raw.name ?? existing?.name ?? id,

src/backend/drivers/ai-chat/providers/alibaba/models.ts

@@ -37,6 +37,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen-max-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 160,
@@ -60,6 +61,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen-plus-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 40,
@@ -81,6 +83,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen3.5-plus-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 40,
@@ -102,6 +105,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen3.6-plus-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 50,
@@ -123,6 +127,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen3.6-flash-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 19,
@@ -143,6 +148,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen3.7-plus-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 50,
@@ -164,6 +170,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen3.7-max-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 250,
@@ -188,6 +195,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen3.8-max-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 200,
@@ -210,6 +218,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen3.8-max-0902-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 200,
@@ -234,6 +243,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen-flash-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 5,
@@ -257,6 +267,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen3-coder-flash-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 30,
@@ -282,6 +293,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen3.5-397b-a17b-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 60,
@@ -305,6 +317,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen3.5-122b-a10b-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 40,
@@ -328,6 +341,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen3.5-35b-a3b-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 25,
@@ -351,6 +365,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen3.5-27b-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 30,
@@ -376,6 +391,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen3.6-35b-a3b-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 24.8,
@@ -399,6 +415,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen3.6-27b-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 60,
@@ -422,6 +439,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen3-vl-plus-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 20,
@@ -447,6 +465,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen3-omni-flash-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 43,
@@ -469,6 +488,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen-mt-plus-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 246,
@@ -491,6 +511,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen-plus-2025-01-25-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 40,
@@ -511,6 +532,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen-plus-2025-04-28-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 40,
@@ -531,6 +553,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen-plus-2025-07-14-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 40,
@@ -553,6 +576,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen-plus-2025-12-01-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 40,
@@ -574,6 +598,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen3.7-max-2026-05-20-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 250,
@@ -594,6 +619,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen3.7-max-2026-06-08-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 250,
@@ -614,6 +640,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen3.7-max-2026-05-17-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 250,
@@ -635,6 +662,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen3.7-max-preview-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 250,
@@ -657,6 +685,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen3.5-plus-2026-02-15-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 40,
@@ -678,6 +707,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen3.5-plus-2026-04-20-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 40,
@@ -700,6 +730,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen3.5-flash-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 10,
@@ -722,6 +753,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen3.7-flash-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 3,
@@ -743,6 +775,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen3.8-flash-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 15,
@@ -765,6 +798,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen3.8-27b-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 50,
@@ -785,6 +819,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen3.8-2.4t-a95b-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 200,
@@ -809,6 +844,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen3.5-omni-flash-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 40,
@@ -832,6 +868,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen3.5-omni-plus-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 140,
@@ -855,6 +892,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen3.8-omni-flash-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 15,
@@ -877,6 +915,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen-mt-flash-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 16,
@@ -897,6 +936,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'qwen-mt-lite-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 12,
@@ -919,6 +959,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'glm-5.1-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 140,
@@ -938,6 +979,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'glm-5.2-fast-preview' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 140,
@@ -958,6 +1000,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'glm-5.2-fast-preview-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 280,
@@ -981,6 +1024,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'deepseek-v4-flash-0731-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 44,
@@ -1003,6 +1047,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'deepseek-v4.1-flash-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 30,
@@ -1023,6 +1068,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'deepseek-v4-pro-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 240,
@@ -1045,6 +1091,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'deepseek-v4-pro-0813-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 132,
@@ -1065,6 +1112,7 @@ export const ALIBABA_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'kimi-k2.7-code-prime' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 95,

src/backend/drivers/ai-chat/providers/azure/AzureChatProvider.ts

@@ -20,6 +20,10 @@
 import { HttpError } from '@heyputer/backend/src/core/http/HttpError.js';
 import { OpenAI } from 'openai';
 import { ChatCompletionCreateParams } from 'openai/resources/index.js';
+import {
+    serviceSpeedBillingModel,
+    servedServiceSpeed,
+} from '../../utils/serviceSpeed.js';
 import { Context } from '../../../../core/context.js';
 import type { FSService } from '../../../../services/fs/FSService.js';
 import type { MeteringService } from '../../../../services/metering/MeteringService.js';
@@ -153,6 +157,7 @@ export class AzureChatProvider implements IChatProvider {
 
     async complete(
         params: ICompleteArguments,
+        resolved?: IChatModel,
     ): ReturnType<IChatProvider['complete']> {
         const {
             max_tokens,
@@ -180,7 +185,9 @@ export class AzureChatProvider implements IChatProvider {
                     { legacyCode: 'bad_request' },
                 );
             }
-            return await this.#responsesProvider.complete(params);
+            return await (resolved
+                ? this.#responsesProvider.complete(params, resolved)
+                : this.#responsesProvider.complete(params));
         }
         // Inline compaction is a Responses-API feature; chat.completions can't
         // express `context_management` or a `compaction` content block.
@@ -194,7 +201,9 @@ export class AzureChatProvider implements IChatProvider {
                     { legacyCode: 'bad_request' },
                 );
             }
-            return await this.#responsesProvider.complete(params);
+            return await (resolved
+                ? this.#responsesProvider.complete(params, resolved)
+                : this.#responsesProvider.complete(params));
         }
         // Validate messages
         if (!Array.isArray(messages)) {
@@ -207,9 +216,11 @@ export class AzureChatProvider implements IChatProvider {
         model = model ?? this.#defaultModel;
 
         const modelUsed =
-            this.models().find((m) =>
+            resolved ??
+            (this.models().find((m) =>
                 [m.id, ...(m.aliases || [])].includes(model),
-            ) || this.models().find((m) => m.id === this.getDefaultModel())!;
+            ) ||
+                this.models().find((m) => m.id === this.getDefaultModel())!);
 
         // messages.unshift({
         //     role: 'system',
@@ -279,6 +290,9 @@ export class AzureChatProvider implements IChatProvider {
                   }),
             messages: messages,
             model: modelUsed.id,
+            ...(params.service_tier !== undefined
+                ? { service_tier: params.service_tier }
+                : {}),
             ...(mappedTools?.length ? { tools: mappedTools } : {}),
             ...(max_tokens !== undefined
                 ? { max_completion_tokens: max_tokens }
@@ -315,7 +329,12 @@ export class AzureChatProvider implements IChatProvider {
         );
 
         return OpenAiUtil.handle_completion_output({
-            usage_calculator: ({ usage }) => {
+            usage_calculator: ({
+                service_tier,
+                usage,
+                setUsageCosts,
+                setUsageDetails,
+            }) => {
                 const cachedTokens =
                     usage.prompt_tokens_details?.cached_tokens ?? 0;
                 const trackedUsage = {
@@ -335,9 +354,20 @@ export class AzureChatProvider implements IChatProvider {
 
                 const costsOverrideFromModel = buildCostsOverride(
                     trackedUsage,
-                    modelUsed,
+                    serviceSpeedBillingModel(
+                        modelUsed,
+                        servedServiceSpeed({ service_tier }),
+                    ),
                 );
 
+                setUsageCosts(costsOverrideFromModel);
+                if (params.serviceSpeed !== undefined)
+                    setUsageDetails({
+                        ...OpenAiUtil.usageDetailsFromTrackedUsage(
+                            trackedUsage,
+                        )!,
+                        serviceSpeed: servedServiceSpeed({ service_tier }),
+                    });
                 this.#meteringService.utilRecordUsageObject(
                     trackedUsage,
                     actor!,

src/backend/drivers/ai-chat/providers/azure/AzureResponsesProvider.ts

@@ -19,6 +19,10 @@
 
 import { OpenAI } from 'openai';
 import { ResponseCreateParams } from 'openai/resources/responses/responses.mjs';
+import {
+    serviceSpeedBillingModel,
+    servedServiceSpeed,
+} from '../../utils/serviceSpeed.js';
 import { Context } from '../../../../core/context.js';
 import type { FSService } from '../../../../services/fs/FSService.js';
 import type { MeteringService } from '../../../../services/metering/MeteringService.js';
@@ -152,37 +156,41 @@ export class AzureResponsesProvider implements IChatProvider {
             : {};
     }
 
-    async complete({
-        messages,
-        model,
-        max_tokens,
-        moderation,
-        tools,
-        tool_choice,
-        parallel_tool_calls,
-        include,
-        conversation,
-        compaction,
-        context_management,
-        previous_response_id,
-        instructions,
-        metadata,
-        prompt,
-        prompt_cache_key,
-        prompt_cache_retention,
-        store,
-        top_p,
-        truncation,
-        background,
-        service_tier,
-        verbosity,
-        stream,
-        reasoning,
-        reasoning_effort,
-        temperature,
-        text,
-        outputFormat,
-    }: ICompleteArguments): ReturnType<IChatProvider['complete']> {
+    async complete(
+        {
+            messages,
+            model,
+            max_tokens,
+            moderation,
+            tools,
+            tool_choice,
+            parallel_tool_calls,
+            include,
+            conversation,
+            compaction,
+            context_management,
+            previous_response_id,
+            instructions,
+            metadata,
+            prompt,
+            prompt_cache_key,
+            prompt_cache_retention,
+            store,
+            top_p,
+            truncation,
+            background,
+            service_tier,
+            serviceSpeed: requestedServiceSpeed,
+            verbosity,
+            stream,
+            reasoning,
+            reasoning_effort,
+            temperature,
+            text,
+            outputFormat,
+        }: ICompleteArguments,
+        resolved?: IChatModel,
+    ): ReturnType<IChatProvider['complete']> {
         // Validate messages
         if (!Array.isArray(messages)) {
             throw new HttpError(400, '`messages` must be an array', {
@@ -200,12 +208,13 @@ export class AzureResponsesProvider implements IChatProvider {
         model = model ?? this.#defaultModel;
 
         const modelUsed =
-            this.models({ no_restrictions: true }).find((m) =>
+            resolved ??
+            (this.models({ no_restrictions: true }).find((m) =>
                 [m.id, ...(m.aliases || [])].includes(model),
             ) ||
-            this.models({ no_restrictions: true }).find(
-                (m) => m.id === this.getDefaultModel(),
-            )!;
+                this.models({ no_restrictions: true }).find(
+                    (m) => m.id === this.getDefaultModel(),
+                )!);
 
         const userIdentifier = upstreamUserIdentifier(actor);
         // Cache key defaults to the actor identifier; see upstreamUserIdentifier.
@@ -335,9 +344,11 @@ export class AzureResponsesProvider implements IChatProvider {
         return OpenAiUtil.handle_completion_output_responses_api({
             usage_calculator: ({
                 usage,
+                service_tier,
                 webSearchCalls,
                 tool_usage,
                 setUsageCosts,
+                setUsageDetails,
             }) => {
                 const numWebSearches: number =
                     // eslint-disable-next-line @typescript-eslint/no-explicit-any
@@ -363,7 +374,10 @@ export class AzureResponsesProvider implements IChatProvider {
 
                 const costsOverrideFromModel = buildCostsOverride(
                     trackedUsage,
-                    modelUsed,
+                    serviceSpeedBillingModel(
+                        modelUsed,
+                        servedServiceSpeed({ service_tier }),
+                    ),
                 );
                 if (numWebSearches) {
                     costsOverrideFromModel.web_search_calls =
@@ -374,6 +388,14 @@ export class AzureResponsesProvider implements IChatProvider {
                     });
                 }
 
+                setUsageCosts?.(costsOverrideFromModel);
+                if (requestedServiceSpeed !== undefined)
+                    setUsageDetails?.({
+                        ...OpenAiUtil.usageDetailsFromTrackedUsage(
+                            trackedUsage,
+                        )!,
+                        serviceSpeed: servedServiceSpeed({ service_tier }),
+                    });
                 this.#meteringService.utilRecordUsageObject(
                     trackedUsage,
                     actor,

src/backend/drivers/ai-chat/providers/azure/models.test.ts

@@ -40,7 +40,11 @@ const STANDALONE = [
 ];
 
 // Fields an Azure deployment genuinely differs from its source on, by puterId.
-const OVERRIDES: Record<string, string[]> = {};
+const OVERRIDES: Record<string, string[]> = {
+    'azure:x-ai/grok-4.3': ['serviceSpeedRates'],
+    'azure:x-ai/grok-4-20-non-reasoning': ['serviceSpeedRates'],
+    'azure:x-ai/grok-4-20-reasoning': ['serviceSpeedRates'],
+};
 
 const IDENTITY = ['id', 'puterId', 'aliases'];
 
@@ -65,6 +69,18 @@ const mirrored = AZURE_MODELS.filter(
 ).map((m) => [m.puterId!, m] as const);
 
 describe('AZURE_MODELS mirroring', () => {
+    it('inherits OpenAI fast rates without enabling direct-only speed tiers', () => {
+        for (const model of AZURE_MODELS) {
+            expect(model.serviceSpeedRates?.ultrafast).toBeUndefined();
+            if (model.puterId?.startsWith('azure:x-ai/')) {
+                expect(model.serviceSpeedRates?.fast).toBeUndefined();
+            }
+        }
+        const azure = AZURE_MODELS.find((model) => model.id === 'gpt-5.4')!;
+        expect(azure.serviceSpeedRates).toEqual({
+            fast: sourceOf(azure)!.serviceSpeedRates!.fast,
+        });
+    });
     it('has a direct-provider source for every entry not listed as standalone', () => {
         const orphans = mirrored
             .filter(([, m]) => !sourceOf(m))

src/backend/drivers/ai-chat/providers/azure/models.ts

@@ -22,10 +22,8 @@ import { OPEN_AI_MODELS } from '../openai/models.js';
 import { XAI_MODELS } from '../xai/models.js';
 
 /**
- * An Azure deployment of a model we also serve direct. Every field but the
- * identity comes from the source entry, so prices, long-context tiers and
- * limits can't drift from it. `name` is the vendor-qualified name both catalogs
- * share (`openai/gpt-5.4`); extra fields override the source.
+ * Azure shares its source model's list prices, context tiers, and limits.
+ * Identity and available service speeds belong to the deployment.
  */
 const mirror = (
     source: readonly IChatModel[],
@@ -41,6 +39,11 @@ const mirror = (
         puterId: `azure:${name}`,
         id: azure.id,
         ...shared,
+        serviceSpeedRates:
+            source === OPEN_AI_MODELS &&
+            shared.serviceSpeedRates?.fast !== undefined
+                ? { fast: shared.serviceSpeedRates.fast }
+                : undefined,
         aliases: [name],
         ...azure,
     };
@@ -127,6 +130,13 @@ export const AZURE_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: {
+            fast: {
+                prompt_tokens: 250,
+                cached_tokens: 25,
+                completion_tokens: 2000,
+            },
+        },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 125,
@@ -171,6 +181,13 @@ export const AZURE_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: {
+            fast: {
+                prompt_tokens: 45,
+                cached_tokens: 4.5,
+                completion_tokens: 360,
+            },
+        },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 25,

src/backend/drivers/ai-chat/providers/claude/ClaudeProvider.integration.test.ts

@@ -147,6 +147,22 @@ describe.skipIf(skipUnlessEnv(ENV_VAR))('ClaudeProvider (integration)', () => {
         },
     );
 
+    it('returns a completion from claude-fable-5-2', { timeout: INTEGRATION_TEST_TIMEOUT_MS }, async () => {
+        const provider = buildProvider();
+        const result = await withTestActor(() =>
+            provider.complete({
+                model: 'claude-fable-5-2',
+                messages: [{ role: 'user', content: 'Reply with OK.' }],
+                reasoning_effort: 'low',
+                max_tokens: 128,
+            }),
+        );
+        const content = (result as {
+            message: { content: Array<{ type: string; text?: string }> };
+        }).message.content;
+        expect(content.find((block) => block.type === 'text')?.text?.length).toBeGreaterThan(0);
+    });
+
     // (a) Claude Code-shaped system blocks, attribution header first — the
     // pipeline must never merge or reorder system blocks: the header is
     // only recognized by Anthropic when it is the sole content of an

src/backend/drivers/ai-chat/providers/claude/ClaudeProvider.test.ts

@@ -1173,6 +1173,7 @@ describe('ClaudeProvider.complete request shape', () => {
     });
 
     it.each([
+        'claude-fable-5-2',
         'claude-fable-5-1',
         'claude-haiku-5-5',
         'claude-opus-5-5',
@@ -1198,6 +1199,7 @@ describe('ClaudeProvider.complete request shape', () => {
     );
 
     it.each([
+        'claude-fable-5-2',
         'claude-fable-5-1',
         'claude-haiku-5-5',
         'claude-opus-5-5',
@@ -1303,6 +1305,73 @@ describe('ClaudeProvider model resolution', () => {
         );
     });
 
+    it.each([
+        'claude-fable-5-2',
+        'claude-fable-5-2-latest',
+        'claude-fable-5.2',
+        'anthropic/claude-fable-5-2',
+    ])(
+        'routes %s to Fable 5.2 with published rates',
+        async (model) => {
+            const { provider } = makeProvider();
+            messagesCreateMock.mockResolvedValueOnce({
+                content: [{ type: 'text', text: 'ok' }],
+                usage: {
+                    input_tokens: 100,
+                    output_tokens: 50,
+                    cache_read_input_tokens: 1000,
+                    cache_creation_input_tokens: 200,
+                    cache_creation: {
+                        ephemeral_5m_input_tokens: 150,
+                        ephemeral_1h_input_tokens: 50,
+                    },
+                },
+            });
+
+            await withTestActor(() =>
+                provider.complete({
+                    model,
+                    messages: [{ role: 'user', content: 'hi' }],
+                }),
+            );
+
+            expect(messagesCreateMock.mock.calls[0]![0].model).toBe(
+                'claude-fable-5-2',
+            );
+            const fable52 = CLAUDE_MODELS.find(
+                (m) => m.id === 'claude-fable-5-2',
+            )!;
+            expect(fable52.costs).toMatchObject({
+                input_tokens: 800,
+                output_tokens: 4000,
+                cache_read_input_tokens: 20,
+                ephemeral_5m_input_tokens: 1000,
+                ephemeral_1h_input_tokens: 1600,
+            });
+            expect(fable52.serviceSpeedRates?.fast).toMatchObject({
+                input_tokens: 1600,
+                output_tokens: 8000,
+                cache_read_input_tokens: 40,
+                ephemeral_5m_input_tokens: 2000,
+                ephemeral_1h_input_tokens: 3200,
+            });
+            expect(fable52.context).toBe(1_000_000);
+            expect(fable52.max_tokens).toBe(128_000);
+            const [, , prefix, overrides] = recordSpy.mock.calls[0]!;
+            expect(prefix).toBe('claude:claude-fable-5-2');
+            expect(overrides).toMatchObject({
+                input_tokens: 100 * Number(fable52.costs.input_tokens),
+                output_tokens: 50 * Number(fable52.costs.output_tokens),
+                cache_read_input_tokens:
+                    1000 * Number(fable52.costs.cache_read_input_tokens),
+                ephemeral_5m_input_tokens:
+                    150 * Number(fable52.costs.ephemeral_5m_input_tokens),
+                ephemeral_1h_input_tokens:
+                    50 * Number(fable52.costs.ephemeral_1h_input_tokens),
+            });
+        },
+    );
+
     it('routes the bare claude-fable alias to fable 5.1 rather than fable 5', async () => {
         const { provider } = makeProvider();
         messagesCreateMock.mockResolvedValueOnce(baseResponse);
@@ -2311,16 +2380,20 @@ describe('ClaudeProvider metering', () => {
         expect('thinking_tokens' in overrides).toBe(false);
     });
 
-    it('bills at the fast_* rates when the upstream reports usage.speed:"fast"', async () => {
+    it.each(['claude-opus-5-5', 'claude-fable-5-2'])('bills at fast rates for %s when the upstream reports fast usage', async (model) => {
         const { provider } = makeProvider();
         messagesCreateMock.mockResolvedValueOnce({
             content: [{ type: 'text', text: 'hi' }],
-            usage: { input_tokens: 10, output_tokens: 5, speed: 'fast' },
+            usage: {
+                input_tokens: 10, output_tokens: 5, speed: 'fast',
+                cache_creation: { ephemeral_5m_input_tokens: 7, ephemeral_1h_input_tokens: 3 },
+                cache_read_input_tokens: 20,
+            },
         });
 
         const result = (await withTestActor(() =>
             provider.complete({
-                model: 'claude-opus-5-5',
+                model,
                 messages: [{ role: 'user', content: 'hi' }],
                 speed: 'fast',
             }),
@@ -2331,11 +2404,15 @@ describe('ClaudeProvider metering', () => {
         expect('input_tokens' in result.usage).toBe(false);
         expect(result.usageDetails.speed).toBe('fast');
 
-        const opus = CLAUDE_MODELS.find((m) => m.id === 'claude-opus-5-5')!;
+        const selectedModel = CLAUDE_MODELS.find((m) => m.id === model)!;
         const [, , prefix, overrides] = recordSpy.mock.calls[0]!;
-        expect(prefix).toBe('claude:claude-opus-5-5');
-        expect(overrides.fast_input_tokens).toBe(10 * Number(opus.costs.fast_input_tokens));
-        expect(overrides.fast_output_tokens).toBe(5 * Number(opus.costs.fast_output_tokens));
+        expect(prefix).toBe(`claude:${model}`);
+        const fastRates = selectedModel.serviceSpeedRates!.fast as Record<string, number>;
+        expect(overrides.fast_input_tokens).toBe(10 * fastRates.input_tokens);
+        expect(overrides.fast_output_tokens).toBe(5 * fastRates.output_tokens);
+        expect(overrides.fast_ephemeral_5m_input_tokens).toBe(7 * fastRates.ephemeral_5m_input_tokens);
+        expect(overrides.fast_ephemeral_1h_input_tokens).toBe(3 * fastRates.ephemeral_1h_input_tokens);
+        expect(overrides.fast_cache_read_input_tokens).toBe(20 * fastRates.cache_read_input_tokens);
     });
 
     it('bills at standard rates when fast mode was requested but not granted by the upstream', async () => {
@@ -2357,13 +2434,15 @@ describe('ClaudeProvider metering', () => {
         expect('fast_input_tokens' in result.usage).toBe(false);
     });
 
-    it('retries at standard speed on Claude when fast mode is rate limited', async () => {
+    it.each([
+        [429, "This request would exceed your organization's rate limit of 0 fast mode input tokens per minute"],
+        [529, 'Overloaded'],
+        [529, 'Fast mode capacity unavailable'],
+    ])('retries at standard speed on Claude after fast capacity failure %s', async (status, message) => {
         const { provider } = makeProvider();
         const fastLimited = Object.assign(
-            new Error(
-                "This request would exceed your organization's rate limit of 0 fast mode input tokens per minute",
-            ),
-            { status: 429 },
+            new Error(message),
+            { status },
         );
         messagesCreateMock.mockRejectedValueOnce(fastLimited).mockResolvedValueOnce({
             content: [{ type: 'text', text: 'hi' }],
@@ -2587,13 +2666,13 @@ describe('ClaudeProvider.requestPricing', () => {
     const haiku = CLAUDE_MODELS.find((m) => m.id === 'claude-haiku-4-5-20251001')!;
     const opus55 = CLAUDE_MODELS.find((m) => m.id === 'claude-opus-5-5')!;
 
-    it('returns fast cost keys only when the model has fast rates and fast was requested', () => {
+    it('leaves token reservation keys to the selected catalog rate card', () => {
         const { provider } = makeProvider();
         expect(
             provider.requestPricing({ speed: 'fast', tools: [] } as never, opus55, {
                 promptTokenEstimate: 10,
             }),
-        ).toMatchObject({ inputKey: 'fast_input_tokens', outputKey: 'fast_output_tokens' });
+        ).toEqual({});
         expect(
             provider.requestPricing({ speed: 'fast', tools: [] } as never, haiku, {
                 promptTokenEstimate: 10,

src/backend/drivers/ai-chat/providers/claude/ClaudeProvider.ts

@@ -43,6 +43,7 @@ import {
 } from '../../utils/mediaParts.js';
 import { isToolChoice, toolChoiceFromWire } from '../../utils/openaiParams.js';
 import { buildCostsOverride } from '../../utils/pricing.js';
+import { serviceSpeedCosts } from '../../utils/serviceSpeed.js';
 import type {
     AIChatStream,
     AIChatTextStream,
@@ -107,20 +108,21 @@ const toAnthropicMediaPart = (part: any): any => {
     return part;
 };
 
-/** A 429 from fast mode's separate quota, not the model's normal limits. */
-const isFastModeRateLimit = (e: unknown): boolean => {
+/** Fast quota or overload failures can be retried at standard speed. */
+const isFastModeUnavailable = (e: unknown): boolean => {
     const err = e as { status?: unknown; message?: unknown };
     return (
-        err?.status === 429 &&
-        typeof err.message === 'string' &&
-        /fast mode/i.test(err.message)
+        typeof err?.message === 'string' &&
+        ((err?.status === 429 && /fast mode/i.test(err.message)) ||
+            (err?.status === 529 && /overload|fast mode/i.test(err.message)))
     );
 };
 
 // Models whose current turn rejects a non-default temperature/top_p/top_k
-// outright. Fable 5/5.1, Sonnet 5/5.5, Haiku 5.5, and Opus 4.7+.
+// outright. Fable 5/5.1/5.2, Sonnet 5/5.5, Haiku 5.5, and Opus 4.7+.
 const OMITS_SAMPLING_PARAMS = new Set([
     'claude-haiku-5-5',
+    'claude-fable-5-2',
     'claude-fable-5-1',
     'claude-fable-5',
     'claude-sonnet-5-5',
@@ -133,6 +135,7 @@ const OMITS_SAMPLING_PARAMS = new Set([
 
 const SUPPORTS_EFFORT = new Set([
     'claude-haiku-5-5',
+    'claude-fable-5-2',
     'claude-fable-5-1',
     'claude-fable-5',
     'claude-sonnet-5-5',
@@ -298,20 +301,12 @@ export class ClaudeProvider implements IChatProvider {
         return `claude:${modelId}`;
     }
 
-    /**
-     * Sizes the credit hold for Claude-only extras the generic per-token math
-     * can't see: fast mode's own rate card, and a web-search/advisor tool that
-     * could run several times before the model answers.
-     */
+    /** Reserves tools that can run several times before the model answers. */
     requestPricing(
         args: ICompleteArguments,
         model: IChatModel,
         est: { promptTokenEstimate: number },
-    ): { inputKey?: string; outputKey?: string; extraCost?: number } {
-        const fast =
-            args.speed === 'fast' &&
-            model.costs?.fast_output_tokens !== undefined;
-
+    ): { extraCost?: number } {
         let extraCost = 0;
         for (const raw of args.tools ?? []) {
             const tool = raw as Record<string, unknown>;
@@ -350,12 +345,6 @@ export class ClaudeProvider implements IChatProvider {
         }
 
         return {
-            ...(fast
-                ? {
-                      inputKey: 'fast_input_tokens',
-                      outputKey: 'fast_output_tokens',
-                  }
-                : {}),
             ...(extraCost > 0 ? { extraCost } : {}),
         };
     }
@@ -364,11 +353,9 @@ export class ClaudeProvider implements IChatProvider {
         try {
             return await this.#completeOnce(args);
         } catch (e) {
-            // Fast mode has its own quota. Anthropic's advice on that 429 is to
-            // drop `speed` and run at standard speed; doing it here keeps the
-            // request on Claude instead of marking the route unhealthy and
-            // replaying it on a reseller.
-            if (args.speed === 'fast' && isFastModeRateLimit(e)) {
+            // Retry fast-capacity failures on Claude at standard speed before
+            // considering a different provider.
+            if (args.speed === 'fast' && isFastModeUnavailable(e)) {
                 return await this.#completeOnce({ ...args, speed: undefined });
             }
             throw e;
@@ -637,7 +624,7 @@ export class ClaudeProvider implements IChatProvider {
             : undefined;
         const fast =
             args.speed === 'fast' &&
-            modelUsed.costs?.fast_output_tokens !== undefined;
+            modelUsed.serviceSpeedRates?.fast !== undefined;
         const safeguardsPolicy = applySafeguardsPolicy(
             args.safeguards,
             args.anthropicBetas,
@@ -1071,6 +1058,7 @@ export class ClaudeProvider implements IChatProvider {
         // keeps reasoning visible in the stream.
         if (
             modelId === 'claude-haiku-5-5' ||
+            modelId === 'claude-fable-5-2' ||
             modelId === 'claude-fable-5-1' ||
             modelId === 'claude-fable-5' ||
             modelId === 'claude-sonnet-5-5' ||
@@ -1505,7 +1493,8 @@ export class ClaudeProvider implements IChatProvider {
 
         const fast =
             raw.speed === 'fast' &&
-            modelUsed.costs?.fast_output_tokens !== undefined;
+            modelUsed.serviceSpeedRates?.fast !== undefined;
+        // Keep published usage keys while resolving rates from the common catalog shape.
         const executorUsage: Record<string, number> = fast
             ? Object.fromEntries(
                   Object.entries(exec).map(([k, v]) => [`fast_${k}`, v]),
@@ -1517,7 +1506,24 @@ export class ClaudeProvider implements IChatProvider {
         const fetches = raw.server_tool_use?.web_fetch_requests ?? 0;
         if (fetches) executorUsage.web_fetch_requests = fetches;
 
-        const executorCosts = buildCostsOverride(executorUsage, modelUsed);
+        const fastRates = fast
+            ? serviceSpeedCosts(modelUsed, 'fast')
+            : undefined;
+        const billingModel = fastRates
+            ? {
+                  ...modelUsed,
+                  costs: {
+                      ...modelUsed.costs,
+                      ...Object.fromEntries(
+                          Object.keys(exec).map((key) => [
+                              `fast_${key}`,
+                              fastRates[key],
+                          ]),
+                      ),
+                  },
+              }
+            : modelUsed;
+        const executorCosts = buildCostsOverride(executorUsage, billingModel);
 
         let advisorUsage: Record<string, number> | undefined;
         let advisorCosts: Record<string, number> | undefined;

src/backend/drivers/ai-chat/providers/claude/models.ts

@@ -55,6 +55,46 @@ export const CLAUDE_MODELS: IChatModel[] = [
         context: 1_000_000,
         max_tokens: 128_000,
     },
+    {
+        puterId: 'anthropic:anthropic/claude-fable-5-2',
+        id: 'claude-fable-5-2',
+        modalities: { input: ['text', 'image', 'pdf'], output: ['text'] },
+        open_weights: false,
+        tool_call: true,
+        release_date: '2026-11-09',
+        aliases: [
+            'claude-fable-5-2-latest',
+            'claude-fable-5.2',
+            'anthropic/claude-fable-5-2',
+        ],
+        name: 'Claude Fable 5.2',
+        midConversationSystem: true,
+        costs_currency: 'usd-cents',
+        input_cost_key: 'input_tokens',
+        output_cost_key: 'output_tokens',
+        // Source: https://www.anthropic.com/claude-fable-and-mythos-5-2
+        serviceSpeedRates: {
+            fast: {
+                input_tokens: 1600,
+                output_tokens: 8000,
+                ephemeral_5m_input_tokens: 1600 * 1.25,
+                ephemeral_1h_input_tokens: 1600 * 2,
+                cache_read_input_tokens: 1600 * 0.025,
+            },
+        },
+        costs: {
+            tokens: 1_000_000,
+            input_tokens: 800,
+            ephemeral_5m_input_tokens: 800 * 1.25,
+            ephemeral_1h_input_tokens: 800 * 2,
+            // 0.025x input, not the usual 0.1x.
+            cache_read_input_tokens: 800 * 0.025,
+            output_tokens: 4000,
+            web_search_requests: 1_000_000, // $10/1k requests
+        },
+        context: 1000000,
+        max_tokens: 128000,
+    },
     {
         puterId: 'anthropic:anthropic/claude-fable-5-1',
         id: 'claude-fable-5-1',
@@ -190,6 +230,15 @@ export const CLAUDE_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'input_tokens',
         output_cost_key: 'output_tokens',
+        serviceSpeedRates: {
+            fast: {
+                input_tokens: 800,
+                output_tokens: 4000,
+                ephemeral_5m_input_tokens: 1000,
+                ephemeral_1h_input_tokens: 1600,
+                cache_read_input_tokens: 40,
+            },
+        },
         costs: {
             tokens: 1_000_000,
             input_tokens: 400,
@@ -199,12 +248,6 @@ export const CLAUDE_MODELS: IChatModel[] = [
             cache_read_input_tokens: 400 * 0.05,
             output_tokens: 2000,
             web_search_requests: 1_000_000,
-            // Fast mode (`speed:'fast'`), live-verified.
-            fast_input_tokens: 800,
-            fast_output_tokens: 4000,
-            fast_ephemeral_5m_input_tokens: 1000,
-            fast_ephemeral_1h_input_tokens: 1600,
-            fast_cache_read_input_tokens: 40, // 0.05x fast_input_tokens
         },
         context: 1000000,
         max_tokens: 128000,
@@ -223,6 +266,15 @@ export const CLAUDE_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'input_tokens',
         output_cost_key: 'output_tokens',
+        serviceSpeedRates: {
+            fast: {
+                input_tokens: 1000,
+                output_tokens: 5000,
+                ephemeral_5m_input_tokens: 1250,
+                ephemeral_1h_input_tokens: 2000,
+                cache_read_input_tokens: 100,
+            },
+        },
         costs: {
             tokens: 1_000_000,
             input_tokens: 500,
@@ -231,11 +283,6 @@ export const CLAUDE_MODELS: IChatModel[] = [
             cache_read_input_tokens: 500 * 0.1,
             output_tokens: 2500,
             web_search_requests: 1_000_000,
-            fast_input_tokens: 1000,
-            fast_output_tokens: 5000,
-            fast_ephemeral_5m_input_tokens: 1250,
-            fast_ephemeral_1h_input_tokens: 2000,
-            fast_cache_read_input_tokens: 100,
         },
         context: 1000000,
         max_tokens: 128000,
@@ -258,6 +305,15 @@ export const CLAUDE_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'input_tokens',
         output_cost_key: 'output_tokens',
+        serviceSpeedRates: {
+            fast: {
+                input_tokens: 1000,
+                output_tokens: 5000,
+                ephemeral_5m_input_tokens: 1250,
+                ephemeral_1h_input_tokens: 2000,
+                cache_read_input_tokens: 100,
+            },
+        },
         costs: {
             tokens: 1_000_000,
             input_tokens: 500,
@@ -266,11 +322,6 @@ export const CLAUDE_MODELS: IChatModel[] = [
             cache_read_input_tokens: 500 * 0.1,
             output_tokens: 2500,
             web_search_requests: 1_000_000,
-            fast_input_tokens: 1000,
-            fast_output_tokens: 5000,
-            fast_ephemeral_5m_input_tokens: 1250,
-            fast_ephemeral_1h_input_tokens: 2000,
-            fast_cache_read_input_tokens: 100,
         },
         context: 1000000,
         max_tokens: 128000,

src/backend/drivers/ai-chat/providers/gemini/GeminiChatProvider.ts

@@ -84,8 +84,14 @@ export class GeminiChatProvider extends OpenAICompatProvider {
             chat: {
                 completions: {
                     create: (params, options) => {
-                        const { model } = params as { model: string };
-                        if (!GEMINI_NATIVE_MODELS.has(model)) {
+                        const { model, service_tier } = params as {
+                            model: string;
+                            service_tier?: string;
+                        };
+                        if (
+                            !GEMINI_NATIVE_MODELS.has(model) &&
+                            (service_tier === undefined || !native)
+                        ) {
                             return compat.chat.completions.create(
                                 params,
                                 options,
@@ -103,7 +109,12 @@ export class GeminiChatProvider extends OpenAICompatProvider {
                 },
             },
         };
-        const models = native ? GEMINI_MODELS : COMPAT_ONLY_MODELS;
+        const models: IChatModel[] = native
+            ? GEMINI_MODELS
+            : COMPAT_ONLY_MODELS.map((model) => ({
+                  ...model,
+                  serviceSpeeds: ['standard'],
+              }));
         super(meteringService, {
             client,
             defaultModel: 'gemini-2.5-flash',
@@ -183,7 +194,10 @@ export class GeminiChatProvider extends OpenAICompatProvider {
     }
 
     protected override meteredUsage(source: UsageSource, model: IChatModel) {
-        if (GEMINI_NATIVE_MODELS.has(model.id)) {
+        if (
+            GEMINI_NATIVE_MODELS.has(model.id) ||
+            'promptTokenCount' in (source.usage ?? {})
+        ) {
             return nativeMeteredUsage(
                 source.usage as unknown as
                     GenerateContentResponseUsageMetadata | undefined,

src/backend/drivers/ai-chat/providers/gemini/geminiNative.ts

@@ -80,6 +80,7 @@ interface ChatParams {
     temperature?: number;
     reasoning_effort?: 'minimal' | 'low' | 'medium' | 'high';
     stream?: boolean;
+    service_tier?: string;
 }
 
 /** Ids made up for calls Gemini returned without one; never sent back. */
@@ -240,6 +241,12 @@ const toGenerateContentRequest = (
         ? toolConfigOf(params.tool_choice)
         : undefined;
     const config: GenerateContentConfig = {
+        ...(params.service_tier !== undefined
+            ? {
+                  serviceTier:
+                      params.service_tier as GenerateContentConfig['serviceTier'],
+              }
+            : {}),
         ...(signal ? { abortSignal: signal } : {}),
         ...(systemInstruction ? { systemInstruction } : {}),
         ...(params.max_completion_tokens !== undefined

src/backend/drivers/ai-chat/providers/gemini/models.ts

@@ -142,6 +142,7 @@ export const GEMINI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 1.8 },
         costs: ARGON_LAUNCH_COSTS,
         pricingOverrides: listPricingFrom(
             '2026-12-01T00:00:00Z',
@@ -174,6 +175,7 @@ export const GEMINI_MODELS: IChatModel[] = [
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
         // Google serves this endpoint with gemini-3.6-flash and bills its rates.
+        serviceSpeedRates: { fast: 1.8 },
         costs: FLASH_LAUNCH_COSTS,
         pricingOverrides: FLASH_LIST_PRICING,
     },
@@ -201,6 +203,7 @@ export const GEMINI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 1.8 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 30,
@@ -228,6 +231,7 @@ export const GEMINI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 1.8 },
         costs: FLASH_LAUNCH_COSTS,
         pricingOverrides: FLASH_LIST_PRICING,
     },
@@ -250,6 +254,7 @@ export const GEMINI_MODELS: IChatModel[] = [
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
         // Google serves this endpoint with gemini-3.8-flash and bills its rates.
+        serviceSpeedRates: { fast: 1.8 },
         costs: FLASH_LAUNCH_COSTS,
         pricingOverrides: FLASH_LIST_PRICING,
     },
@@ -270,6 +275,7 @@ export const GEMINI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 1.8 },
         costs: FLASH_LAUNCH_COSTS,
         pricingOverrides: FLASH_LIST_PRICING,
     },
@@ -291,6 +297,7 @@ export const GEMINI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 1.8 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 25,
@@ -318,6 +325,7 @@ export const GEMINI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 1.8 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 50,
@@ -344,6 +352,7 @@ export const GEMINI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 1.8 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 30,
@@ -374,6 +383,7 @@ export const GEMINI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 1.8 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 10,
@@ -402,6 +412,7 @@ export const GEMINI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 1.8 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 125,
@@ -448,6 +459,7 @@ export const GEMINI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 1.8 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 200,

src/backend/drivers/ai-chat/providers/infron/InfronProvider.ts

@@ -266,6 +266,15 @@ export class InfronProvider extends OpenAICompatProvider {
         return coerced_models;
     }
 
+    async serviceSpeedModels(model: IChatModel) {
+        const variant = (await this.models()).find(
+            (entry) => entry.id === `${model.id}:priority`,
+        );
+        return variant
+            ? { fast: { ...variant, provider: model.provider } }
+            : {};
+    }
+
     protected override vendorParams(
         params: Record<string, unknown>,
         _args: unknown,

src/backend/drivers/ai-chat/providers/minimax/models.ts

@@ -43,6 +43,8 @@ export const MINIMAX_MODELS: MiniMaxChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 1.5 },
+        serviceSpeedVariants: { fast: 'minimax-m3-highspeed' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 30, // $0.30 per 1M (input <= 512k tokens)
@@ -70,6 +72,7 @@ export const MINIMAX_MODELS: MiniMaxChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'minimax-m2.7-highspeed' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 30, // $0.30 per 1M
@@ -122,6 +125,7 @@ export const MINIMAX_MODELS: MiniMaxChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'minimax-m2.5-highspeed' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 30, // $0.30 per 1M
@@ -174,6 +178,7 @@ export const MINIMAX_MODELS: MiniMaxChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'minimax-m2.1-highspeed' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 30, // $0.30 per 1M
@@ -222,6 +227,7 @@ export const MINIMAX_MODELS: MiniMaxChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'minimax-m2-highspeed' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 30, // $0.30 per 1M

src/backend/drivers/ai-chat/providers/mistral/MistralAiProvider.test.ts

@@ -20,13 +20,12 @@
 /**
  * Offline unit tests for MistralAIProvider.
  *
- * Boots a real PuterServer (in-memory sqlite + dynamo + s3 + mock
- * redis) and constructs MistralAIProvider directly against the live
- * wired `MeteringService` so the recording side is exercised end-to-
- * end. The Mistral SDK is mocked at the module boundary (the real
- * network egress point) so the provider never reaches the network.
- * The companion integration test (MistralAiProvider.integration.test.ts)
- * exercises the real Mistral endpoint.
+ * Boots a real PuterServer (in-memory sqlite + dynamo + s3 + mock redis) and
+ * constructs MistralAIProvider directly against the live wired
+ * `MeteringService` so the recording side is exercised end-to- end. The Mistral
+ * SDK is mocked at the module boundary (the real network egress point) so the
+ * provider never reaches the network. The companion integration test
+ * (MistralAiProvider.integration.test.ts) exercises the real Mistral endpoint.
  */
 
 import { Context } from '../../../../core/context.js';
@@ -147,7 +146,9 @@ describe('MistralAIProvider construction', () => {
     it('constructs the Mistral SDK with the configured API key', () => {
         makeProvider();
         expect(mistralCtor).toHaveBeenCalledTimes(1);
-        expect(mistralCtor).toHaveBeenCalledWith({ apiKey: 'test-key' });
+        expect(mistralCtor).toHaveBeenCalledWith(
+            expect.objectContaining({ apiKey: 'test-key' }),
+        );
     });
 });
 
@@ -606,7 +607,11 @@ describe('MistralAIProvider.complete non-stream output', () => {
             if (stream) {
                 streamMock.mockResolvedValueOnce(
                     asAsyncIterable([
-                        { data: { choices: [{ delta: { content: 'hello' } }] } },
+                        {
+                            data: {
+                                choices: [{ delta: { content: 'hello' } }],
+                            },
+                        },
                         { data: { choices: [], usage } },
                     ]),
                 );
@@ -643,7 +648,11 @@ describe('MistralAIProvider.complete non-stream output', () => {
                 expect(result).toMatchObject({ message: { content: 'hello' } });
             }
             expect(recordSpy).toHaveBeenCalledWith(
-                { prompt_tokens: 5, cached_tokens: 1008, completion_tokens: 30 },
+                {
+                    prompt_tokens: 5,
+                    cached_tokens: 1008,
+                    completion_tokens: 30,
+                },
                 SYSTEM_ACTOR,
                 `mistral:${model}`,
                 {
@@ -688,8 +697,7 @@ describe('MistralAIProvider.complete non-stream output', () => {
         });
 
         expect(recordSpy).toHaveBeenCalledTimes(1);
-        const [usage, actor, prefix, overrides] =
-            recordSpy.mock.calls[0]!;
+        const [usage, actor, prefix, overrides] = recordSpy.mock.calls[0]!;
         expect(usage).toEqual({
             prompt_tokens: 100,
             completion_tokens: 50,
@@ -718,15 +726,11 @@ describe('MistralAIProvider.complete non-stream output', () => {
                         content: [
                             {
                                 type: 'thinking',
-                                thinking: [
-                                    { type: 'text', text: 'step one.' },
-                                ],
+                                thinking: [{ type: 'text', text: 'step one.' }],
                             },
                             {
                                 type: 'thinking',
-                                thinking: [
-                                    { type: 'text', text: 'step two.' },
-                                ],
+                                thinking: [{ type: 'text', text: 'step two.' }],
                             },
                             { type: 'text', text: 'the answer' },
                         ],
@@ -916,8 +920,7 @@ describe('MistralAIProvider.complete streaming', () => {
 
         // mistral-small-2603: prompt=15, completion=60.
         expect(recordSpy).toHaveBeenCalledTimes(1);
-        const [, , prefix, overrides] =
-            recordSpy.mock.calls[0]!;
+        const [, , prefix, overrides] = recordSpy.mock.calls[0]!;
         expect(prefix).toBe('mistral:mistral-small-2603');
         expect(overrides).toMatchObject({
             prompt_tokens: 4 * 15,
@@ -1115,7 +1118,7 @@ describe('MistralAIProvider image_url coercion', () => {
         usage: { promptTokens: 1, completionTokens: 1 },
     };
 
-    it('rewrites canonical image_url parts to the SDK\'s camelCase imageUrl', async () => {
+    it("rewrites canonical image_url parts to the SDK's camelCase imageUrl", async () => {
         const { provider } = makeProvider();
         completeMock.mockResolvedValueOnce(baseCompletion);
 
@@ -1175,7 +1178,8 @@ describe('MistralAIProvider image_url coercion', () => {
                             },
                             {
                                 type: 'image_url',
-                                image_url: 'https://example.com/already-flat.png',
+                                image_url:
+                                    'https://example.com/already-flat.png',
                             },
                         ],
                     },
@@ -1251,7 +1255,9 @@ describe('MistralAIProvider image_url coercion', () => {
                             { type: 'text', text: 'what is in this image?' },
                             {
                                 type: 'image_url',
-                                image_url: { url: 'https://example.com/photo.jpg' },
+                                image_url: {
+                                    url: 'https://example.com/photo.jpg',
+                                },
                             },
                         ],
                     },
@@ -1260,7 +1266,11 @@ describe('MistralAIProvider image_url coercion', () => {
         );
 
         const [args] = completeMock.mock.calls[0]!;
-        const parts = args.messages[0].content as { type: string; text?: string; imageUrl?: unknown }[];
+        const parts = args.messages[0].content as {
+            type: string;
+            text?: string;
+            imageUrl?: unknown;
+        }[];
         const textPart = parts.find((p) => p.type === 'text');
         const imgPart = parts.find((p) => p.type === 'image_url');
         expect(textPart?.text).toBe('what is in this image?');
@@ -1300,3 +1310,41 @@ describe('MistralAIProvider.checkModeration', () => {
         );
     });
 });
+
+describe('Mistral service speed wire mapping', () => {
+    it('adds the tier after SDK serialization and removes its internal header', async () => {
+        makeProvider();
+        const options = mistralCtor.mock.calls[0][0] as {
+            httpClient: import('@mistralai/mistralai/lib/http.js').HTTPClient;
+        };
+        let sent: Request | undefined;
+        vi.stubGlobal('fetch', async (input: Request) => {
+            sent = input;
+            return new Response('{}', {
+                headers: { 'content-type': 'application/json' },
+            });
+        });
+        try {
+            await options.httpClient.request(
+                new Request('https://api.mistral.ai/v1/chat/completions', {
+                    method: 'POST',
+                    headers: {
+                        'content-type': 'application/json',
+                        'x-puter-service-tier': 'auto',
+                    },
+                    body: JSON.stringify({
+                        model: 'mistral-medium-latest',
+                        messages: [],
+                    }),
+                }),
+            );
+            expect(sent?.headers.has('x-puter-service-tier')).toBe(false);
+            expect(await sent?.json()).toMatchObject({
+                model: 'mistral-medium-latest',
+                service_tier: 'auto',
+            });
+        } finally {
+            vi.unstubAllGlobals();
+        }
+    });
+});

src/backend/drivers/ai-chat/providers/mistral/MistralAiProvider.ts

@@ -17,6 +17,8 @@
  * along with this program.  If not, see <https://www.gnu.org/licenses/>.
  */
 
+import { HTTPClient } from '@mistralai/mistralai/lib/http.js';
+import { servedServiceSpeed } from '../../utils/serviceSpeed.js';
 import { Mistral } from '@mistralai/mistralai';
 import { ChatCompletionResponse } from '@mistralai/mistralai/models/components/chatcompletionresponse.js';
 import { MistralError } from '@mistralai/mistralai/models/errors/mistralerror.js';
@@ -26,6 +28,7 @@ import { Context } from '../../../../core/context.js';
 import { HttpError } from '../../../../core/http/HttpError.js';
 import type { MeteringService } from '../../../../services/metering/MeteringService.js';
 import type {
+    IChatModel,
     IChatCompleteResult,
     IChatProvider,
     ICompleteArguments,
@@ -106,6 +109,7 @@ const coerceMistralUsage = (usage: UsageInfo) => {
         prompt_tokens: (usage.promptTokens ?? 0) - cachedTokens,
         completion_tokens: usage.completionTokens ?? 0,
         prompt_tokens_details: { cached_tokens: cachedTokens },
+        service_tier: usage.additionalProperties?.service_tier,
     };
 };
 
@@ -115,7 +119,24 @@ export class MistralAIProvider implements IChatProvider {
     #meteringService: MeteringService;
 
     constructor(config: { apiKey: string }, meteringService: MeteringService) {
+        const httpClient = new HTTPClient();
+        // The SDK strips service_tier during body serialization.
+        httpClient.addHook('beforeRequest', async (request) => {
+            const tier = request.headers.get('x-puter-service-tier');
+            if (!tier) return;
+            const headers = new Headers(request.headers);
+            headers.delete('x-puter-service-tier');
+            const body = await request.json();
+            return new Request(request, {
+                headers,
+                body: JSON.stringify({
+                    ...body,
+                    service_tier: tier === 'auto' ? 'auto' : 'standard_only',
+                }),
+            });
+        });
         this.#client = new Mistral({
+            httpClient,
             apiKey: config.apiKey,
         });
         this.#meteringService = meteringService;
@@ -181,17 +202,22 @@ export class MistralAIProvider implements IChatProvider {
         return `mistral:${modelId}`;
     }
 
-    async complete({
-        messages,
-        stream,
-        model,
-        tools,
-        max_tokens,
-        temperature,
-        normalize,
-        response,
-        custom,
-    }: ICompleteArguments): Promise<IChatCompleteResult> {
+    async complete(
+        {
+            messages,
+            stream,
+            model,
+            tools,
+            max_tokens,
+            temperature,
+            normalize,
+            response,
+            custom,
+            service_tier,
+            serviceSpeed: requestedServiceSpeed,
+        }: ICompleteArguments,
+        resolved?: IChatModel,
+    ): Promise<IChatCompleteResult> {
         // Mistral's reasoning prompt mode: with `prompt_mode: 'reasoning'`,
         // reasoning-capable models return their thinking as structured ThinkChunk
         // content (which the splitter below separates into `reasoning`)
@@ -217,10 +243,13 @@ export class MistralAIProvider implements IChatProvider {
         }
 
         const selectedModel =
-            (await this.models()).find((m) =>
+            resolved ??
+            ((await this.models()).find((m) =>
                 [m.id, ...(m.aliases || [])].includes(model),
             ) ||
-            (await this.models()).find((m) => m.id === this.getDefaultModel())!;
+                (await this.models()).find(
+                    (m) => m.id === this.getDefaultModel(),
+                )!);
         const actor = Context.get('actor');
         let completion: Awaited<
             ReturnType<Mistral['chat']['complete' | 'stream']>
@@ -242,6 +271,9 @@ export class MistralAIProvider implements IChatProvider {
                 // The SDK retries nothing by default and has no timeout once a
                 // signal is passed; a stream is bounded by its caller instead.
                 {
+                    ...(service_tier !== undefined
+                        ? { headers: { 'x-puter-service-tier': service_tier } }
+                        : {}),
                     signal: stream
                         ? Context.get('upstreamSignal')
                         : withSdkTimeout(Context.get('upstreamSignal')),
@@ -398,7 +430,8 @@ export class MistralAIProvider implements IChatProvider {
             },
             completion: completion as ChatCompletionResponse,
             stream,
-            usage_calculator: ({ usage, setUsageCosts }) => {
+            usage_calculator: ({ usage, setUsageCosts, setUsageDetails }) => {
+                const serviceSpeed = servedServiceSpeed({ usage });
                 // `coerceMistralUsage` has already taken cached reads out of
                 // `prompt_tokens`.
                 const metered = meterChatUsage(
@@ -407,8 +440,16 @@ export class MistralAIProvider implements IChatProvider {
                     this.meteringModelKey(selectedModel.id),
                     selectedModel,
                     OpenAIUtil.extractMeteredUsage(usage),
+                    { serviceSpeed },
                 );
                 setUsageCosts(metered.costs);
+                if (requestedServiceSpeed !== undefined)
+                    setUsageDetails({
+                        ...OpenAIUtil.usageDetailsFromTrackedUsage(
+                            metered.usage,
+                        )!,
+                        serviceSpeed,
+                    });
                 return metered.usage;
             },
         });

src/backend/drivers/ai-chat/providers/mistral/models.ts

@@ -42,6 +42,7 @@ export const MISTRAL_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 1.75 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 136,
@@ -75,6 +76,7 @@ export const MISTRAL_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 1.75 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 150,
@@ -105,6 +107,7 @@ export const MISTRAL_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 1.75 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 50,
@@ -135,6 +138,7 @@ export const MISTRAL_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 1.75 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 15,
@@ -159,6 +163,7 @@ export const MISTRAL_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 1.75 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 140,
@@ -188,6 +193,7 @@ export const MISTRAL_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 1.75 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 30,
@@ -213,6 +219,7 @@ export const MISTRAL_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 1.75 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 20,
@@ -238,6 +245,7 @@ export const MISTRAL_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 1.75 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 15,
@@ -262,6 +270,7 @@ export const MISTRAL_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 1.75 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 10,

src/backend/drivers/ai-chat/providers/moonshot/models.ts

@@ -30,6 +30,7 @@ export const MOONSHOT_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'kimi-k3.1-pro-highspeed' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 200, // $2.00 per 1M
@@ -77,6 +78,7 @@ export const MOONSHOT_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'kimi-k3-highspeed' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 300, // $3.00 per 1M
@@ -104,6 +106,7 @@ export const MOONSHOT_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'kimi-k2.6-highspeed' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 95, // $0.95 per 1M
@@ -126,6 +129,7 @@ export const MOONSHOT_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedVariants: { fast: 'kimi-k2.7-code-highspeed' },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 95,

src/backend/drivers/ai-chat/providers/openai/OpenAiChatCompletionsProvider.ts

@@ -20,6 +20,7 @@
 import { HttpError } from '@heyputer/backend/src/core/http/HttpError.js';
 import { OpenAI } from 'openai';
 import { ChatCompletionCreateParams } from 'openai/resources/index.js';
+import { servedServiceSpeed } from '../../utils/serviceSpeed.js';
 import { Context } from '../../../../core/context.js';
 import type { MeteringService } from '../../../../services/metering/MeteringService.js';
 import type {
@@ -117,6 +118,7 @@ export class OpenAiChatProvider implements IChatProvider {
 
     async complete(
         params: ICompleteArguments,
+        resolved?: IChatModel,
     ): ReturnType<IChatProvider['complete']> {
         const {
             max_tokens,
@@ -134,6 +136,11 @@ export class OpenAiChatProvider implements IChatProvider {
         } = params;
         let { messages, model } = params;
         const cacheOptions = promptCacheOptionsToWire(prompt_cache_options);
+        if (params.service_tier === 'ultrafast' && this.#responsesProvider) {
+            return resolved
+                ? this.#responsesProvider.complete(params, resolved)
+                : this.#responsesProvider.complete(params);
+        }
         if (
             tools?.filter((e) => isWebSearchTool(e as Record<string, unknown>))
                 .length
@@ -147,7 +154,9 @@ export class OpenAiChatProvider implements IChatProvider {
                     { legacyCode: 'bad_request' },
                 );
             }
-            return await this.#responsesProvider.complete(params);
+            return await (resolved
+                ? this.#responsesProvider.complete(params, resolved)
+                : this.#responsesProvider.complete(params));
         }
         // Inline compaction is a Responses-API feature; chat.completions can't
         // express `context_management` or a `compaction` content block.
@@ -161,7 +170,9 @@ export class OpenAiChatProvider implements IChatProvider {
                     { legacyCode: 'bad_request' },
                 );
             }
-            return await this.#responsesProvider.complete(params);
+            return await (resolved
+                ? this.#responsesProvider.complete(params, resolved)
+                : this.#responsesProvider.complete(params));
         }
         // Only the Responses API writes 12h cache entries or takes a lifetime
         // on a `prompt_cache_breakpoint`; Chat Completions rejects both.
@@ -176,7 +187,9 @@ export class OpenAiChatProvider implements IChatProvider {
                     { legacyCode: 'bad_request' },
                 );
             }
-            return await this.#responsesProvider.complete(params);
+            return await (resolved
+                ? this.#responsesProvider.complete(params, resolved)
+                : this.#responsesProvider.complete(params));
         }
         // Validate messages
         if (!Array.isArray(messages)) {
@@ -189,9 +202,11 @@ export class OpenAiChatProvider implements IChatProvider {
         model = model ?? this.#defaultModel;
 
         const modelUsed =
-            this.models().find((m) =>
+            resolved ??
+            (this.models().find((m) =>
                 [m.id, ...(m.aliases || [])].includes(model),
-            ) || this.models().find((m) => m.id === this.getDefaultModel())!;
+            ) ||
+                this.models().find((m) => m.id === this.getDefaultModel())!);
 
         // messages.unshift({
         //     role: 'system',
@@ -248,6 +263,9 @@ export class OpenAiChatProvider implements IChatProvider {
                 : {}),
             messages: messages,
             model: modelUsed.id,
+            ...(params.service_tier !== undefined
+                ? { service_tier: params.service_tier }
+                : {}),
             ...(mappedTools?.length ? { tools: mappedTools } : {}),
             ...(max_tokens !== undefined
                 ? { max_completion_tokens: max_tokens }
@@ -284,7 +302,13 @@ export class OpenAiChatProvider implements IChatProvider {
         );
 
         return OpenAiUtil.handle_completion_output({
-            usage_calculator: ({ usage, setUsageCosts, setUsageDetails }) => {
+            usage_calculator: ({
+                usage,
+                service_tier,
+                setUsageCosts,
+                setUsageDetails,
+            }) => {
+                const serviceSpeed = servedServiceSpeed({ service_tier });
                 // GPT-5.6 and later report cache writes inside
                 // `prompt_tokens`, like cached reads; Chat Completions only
                 // writes the default 30m entries, billed at 1.25x input.
@@ -296,13 +320,17 @@ export class OpenAiChatProvider implements IChatProvider {
                     this.meteringModelKey(modelUsed.id),
                     modelUsed,
                     trackedUsage,
+                    { serviceSpeed },
                 );
                 setUsageCosts(metered.costs);
-                setUsageDetails(
-                    OpenAiUtil.usageDetailsFromTrackedUsage(metered.usage, {
+                setUsageDetails({
+                    ...OpenAiUtil.usageDetailsFromTrackedUsage(metered.usage, {
                         writeTtl: '30m',
                     })!,
-                );
+                    ...(params.serviceSpeed !== undefined
+                        ? { serviceSpeed }
+                        : {}),
+                });
                 return metered.usage;
             },
             stream,

src/backend/drivers/ai-chat/providers/openai/OpenAiChatResponsesProvider.ts

@@ -19,6 +19,7 @@
 
 import { OpenAI } from 'openai';
 import { ResponseCreateParams } from 'openai/resources/responses/responses.mjs';
+import { servedServiceSpeed } from '../../utils/serviceSpeed.js';
 import { Context } from '../../../../core/context.js';
 import type { MeteringService } from '../../../../services/metering/MeteringService.js';
 import type {
@@ -134,38 +135,42 @@ export class OpenAiResponsesChatProvider implements IChatProvider {
         return rate ? { extraCost: rate } : {};
     }
 
-    async complete({
-        messages,
-        model,
-        max_tokens,
-        moderation,
-        tools,
-        tool_choice,
-        parallel_tool_calls,
-        include,
-        conversation,
-        compaction,
-        context_management,
-        previous_response_id,
-        instructions,
-        metadata,
-        prompt,
-        prompt_cache_key,
-        prompt_cache_retention,
-        prompt_cache_options,
-        store,
-        top_p,
-        truncation,
-        background,
-        service_tier,
-        verbosity,
-        stream,
-        reasoning,
-        reasoning_effort,
-        temperature,
-        text,
-        outputFormat,
-    }: ICompleteArguments): ReturnType<IChatProvider['complete']> {
+    async complete(
+        {
+            messages,
+            model,
+            max_tokens,
+            moderation,
+            tools,
+            tool_choice,
+            parallel_tool_calls,
+            include,
+            conversation,
+            compaction,
+            context_management,
+            previous_response_id,
+            instructions,
+            metadata,
+            prompt,
+            prompt_cache_key,
+            prompt_cache_retention,
+            prompt_cache_options,
+            store,
+            top_p,
+            truncation,
+            background,
+            service_tier,
+            serviceSpeed: requestedServiceSpeed,
+            verbosity,
+            stream,
+            reasoning,
+            reasoning_effort,
+            temperature,
+            text,
+            outputFormat,
+        }: ICompleteArguments,
+        resolved?: IChatModel,
+    ): ReturnType<IChatProvider['complete']> {
         // Validate messages
         if (!Array.isArray(messages)) {
             throw new HttpError(400, '`messages` must be an array', {
@@ -183,12 +188,13 @@ export class OpenAiResponsesChatProvider implements IChatProvider {
         model = model ?? this.#defaultModel;
 
         const modelUsed =
-            this.models({ no_restrictions: true }).find((m) =>
+            resolved ??
+            (this.models({ no_restrictions: true }).find((m) =>
                 [m.id, ...(m.aliases || [])].includes(model),
             ) ||
-            this.models({ no_restrictions: true }).find(
-                (m) => m.id === this.getDefaultModel(),
-            )!;
+                this.models({ no_restrictions: true }).find(
+                    (m) => m.id === this.getDefaultModel(),
+                )!);
 
         // messages.unshift({
         //     role: 'system',
@@ -322,6 +328,7 @@ export class OpenAiResponsesChatProvider implements IChatProvider {
         return OpenAiUtil.handle_completion_output_responses_api({
             usage_calculator: ({
                 usage,
+                service_tier,
                 webSearchCalls,
                 setUsageCosts,
                 setUsageDetails,
@@ -347,23 +354,31 @@ export class OpenAiResponsesChatProvider implements IChatProvider {
                     // model is a reasoning model, not a static per-model
                     // cost-table entry — the model's rates would otherwise
                     // price it (wrongly) at the input/output token rate.
-                    webSearchCalls
-                        ? {
-                              costOverrides: {
-                                  web_search_calls:
-                                      webSearchCalls *
-                                      (webSearchCallRate(tools, modelUsed.id) ??
-                                          0),
-                              },
-                          }
-                        : {},
+                    {
+                        serviceSpeed: servedServiceSpeed({ service_tier }),
+                        ...(webSearchCalls
+                            ? {
+                                  costOverrides: {
+                                      web_search_calls:
+                                          webSearchCalls *
+                                          (webSearchCallRate(
+                                              tools,
+                                              modelUsed.id,
+                                          ) ?? 0),
+                                  },
+                              }
+                            : {}),
+                    },
                 );
                 setUsageCosts?.(metered.costs);
-                setUsageDetails?.(
-                    OpenAiUtil.usageDetailsFromTrackedUsage(metered.usage, {
+                setUsageDetails?.({
+                    ...(requestedServiceSpeed !== undefined
+                        ? { serviceSpeed: servedServiceSpeed({ service_tier }) }
+                        : {}),
+                    ...OpenAiUtil.usageDetailsFromTrackedUsage(metered.usage, {
                         writeTtl: '30m',
                     })!,
-                );
+                });
                 return metered.usage;
             },
             stream,

src/backend/drivers/ai-chat/providers/openai/models.ts

@@ -46,6 +46,7 @@ export const OPEN_AI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 2 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 100,
@@ -72,6 +73,7 @@ export const OPEN_AI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 2, ultrafast: 6 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 200,
@@ -98,6 +100,7 @@ export const OPEN_AI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 2 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 200,
@@ -124,6 +127,7 @@ export const OPEN_AI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 2 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 10,
@@ -150,6 +154,7 @@ export const OPEN_AI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 2, ultrafast: 6 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 1000,
@@ -176,6 +181,7 @@ export const OPEN_AI_MODELS: IChatModel[] = [
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
         // OpenAI's promotional pricing, guaranteed only through 2026-11-21.
+        serviceSpeedRates: { fast: 2 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 400,
@@ -201,6 +207,7 @@ export const OPEN_AI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 2 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 200,
@@ -226,6 +233,7 @@ export const OPEN_AI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 2 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 20,
@@ -251,6 +259,13 @@ export const OPEN_AI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: {
+            fast: {
+                prompt_tokens: 1250,
+                cached_tokens: 125,
+                completion_tokens: 7500,
+            },
+        },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 500,
@@ -295,6 +310,13 @@ export const OPEN_AI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: {
+            fast: {
+                prompt_tokens: 500,
+                cached_tokens: 50,
+                completion_tokens: 3000,
+            },
+        },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 250,
@@ -338,6 +360,13 @@ export const OPEN_AI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: {
+            fast: {
+                prompt_tokens: 150,
+                cached_tokens: 15,
+                completion_tokens: 900,
+            },
+        },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 75,
@@ -425,6 +454,13 @@ export const OPEN_AI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: {
+            fast: {
+                prompt_tokens: 350,
+                cached_tokens: 35,
+                completion_tokens: 2800,
+            },
+        },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 175,
@@ -446,6 +482,13 @@ export const OPEN_AI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: {
+            fast: {
+                prompt_tokens: 250,
+                cached_tokens: 25,
+                completion_tokens: 2000,
+            },
+        },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 125,
@@ -467,6 +510,13 @@ export const OPEN_AI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: {
+            fast: {
+                prompt_tokens: 425,
+                cached_tokens: 212.5,
+                completion_tokens: 1700,
+            },
+        },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 250,
@@ -490,6 +540,13 @@ export const OPEN_AI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: {
+            fast: {
+                prompt_tokens: 25,
+                cached_tokens: 12.5,
+                completion_tokens: 100,
+            },
+        },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 15,
@@ -509,6 +566,13 @@ export const OPEN_AI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: {
+            fast: {
+                prompt_tokens: 350,
+                cached_tokens: 87.5,
+                completion_tokens: 1400,
+            },
+        },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 200,
@@ -530,6 +594,13 @@ export const OPEN_AI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: {
+            fast: {
+                prompt_tokens: 70,
+                cached_tokens: 17.5,
+                completion_tokens: 280,
+            },
+        },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 40,
@@ -571,6 +642,13 @@ export const OPEN_AI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: {
+            fast: {
+                prompt_tokens: 425,
+                cached_tokens: 212.5,
+                completion_tokens: 1700,
+            },
+        },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 250,

src/backend/drivers/ai-chat/providers/openrouter/OpenRouterProvider.test.ts

@@ -20,15 +20,14 @@
 /**
  * Offline unit tests for OpenRouterProvider.
  *
- * Boots a real PuterServer (in-memory sqlite + dynamo + s3 + mock
- * redis) and constructs OpenRouterProvider directly against the live
- * wired `MeteringService` so the recording side is exercised end-to-
- * end. OpenRouter is OpenAI-compatible, so the OpenAI SDK is mocked
- * at the module boundary; the model catalog is fetched via `axios`
- * which is mocked at its module boundary too. Both are the real
- * network egress points.
- * The companion integration test (OpenRouterProvider.integration.test.ts)
- * exercises the real OpenRouter endpoint.
+ * Boots a real PuterServer (in-memory sqlite + dynamo + s3 + mock redis) and
+ * constructs OpenRouterProvider directly against the live wired
+ * `MeteringService` so the recording side is exercised end-to- end. OpenRouter
+ * is OpenAI-compatible, so the OpenAI SDK is mocked at the module boundary; the
+ * model catalog is fetched via `axios` which is mocked at its module boundary
+ * too. Both are the real network egress points. The companion integration test
+ * (OpenRouterProvider.integration.test.ts) exercises the real OpenRouter
+ * endpoint.
  */
 
 import { Writable } from 'node:stream';
@@ -300,16 +299,20 @@ describe('OpenRouterProvider model catalog', () => {
     it('coerces string prices and normalizes conditional tiers', async () => {
         const { provider } = makeProvider();
         const models = await provider.models();
-        const sol = models.find((m) => m.id === 'openrouter:openai/gpt-6.1-sol')!;
+        const sol = models.find(
+            (m) => m.id === 'openrouter:openai/gpt-6.1-sol',
+        )!;
         expect(sol.costs).toEqual({
             tokens: 1_000_000,
             prompt: 200,
             completion: 1000,
         });
-        expect(sol.pricingOverrides).toEqual([{
-            minPromptTokens: 272000,
-            costs: { prompt: 400, completion: 1500 },
-        }]);
+        expect(sol.pricingOverrides).toEqual([
+            {
+                minPromptTokens: 272000,
+                costs: { prompt: 400, completion: 1500 },
+            },
+        ]);
     });
 
     it('caches the coerced model list after the first axios round-trip', async () => {
@@ -368,9 +371,12 @@ describe('OpenRouterProvider model catalog', () => {
 
     it('does not reinterpret catalog-added timestamps as release dates', async () => {
         const { provider } = makeProvider();
-        expect((await provider.models()).every(model => model.release_date === undefined)).toBe(true);
+        expect(
+            (await provider.models()).every(
+                (model) => model.release_date === undefined,
+            ),
+        ).toBe(true);
     });
-
 });
 
 // ── Disallowed model gate ───────────────────────────────────────────
@@ -868,11 +874,14 @@ describe('OpenRouterProvider billed cost through the driver', () => {
             event.factor = factor;
         };
         server.clients.event.on('ai.cost.factor.*', listener as never);
-        return () => server.clients.event.off('ai.cost.factor.*', listener as never);
+        return () =>
+            server.clients.event.off('ai.cost.factor.*', listener as never);
     };
 
-    /** µ¢ the ledger was handed for every usage recorded so far. */
-    const ledgerTotal = (spy: MockInstance<MeteringService['batchIncrementUsages']>) =>
+    /** Μ¢ the ledger was handed for every usage recorded so far. */
+    const ledgerTotal = (
+        spy: MockInstance<MeteringService['batchIncrementUsages']>,
+    ) =>
         spy.mock.calls
             .flatMap(([, usages]) => usages)
             .reduce((total, usage) => total + (usage.costOverride ?? 0), 0);
@@ -928,7 +937,10 @@ describe('OpenRouterProvider billed cost through the driver', () => {
     it('reports what a billed completion was charged, after the cost factor', async () => {
         const driver = await makeDriver();
         const stopFactor = withCostFactor(1.6);
-        const recorded = vi.spyOn(server.services.metering, 'batchIncrementUsages');
+        const recorded = vi.spyOn(
+            server.services.metering,
+            'batchIncrementUsages',
+        );
         createMock.mockResolvedValueOnce({
             choices: [
                 {
@@ -1023,7 +1035,10 @@ describe('OpenRouterProvider billed cost through the driver', () => {
     it('ends a stream with what it was charged at the catalog rates, after the cost factor', async () => {
         const driver = await makeDriver();
         const stopFactor = withCostFactor(1.6);
-        const recorded = vi.spyOn(server.services.metering, 'batchIncrementUsages');
+        const recorded = vi.spyOn(
+            server.services.metering,
+            'batchIncrementUsages',
+        );
         createMock.mockReturnValueOnce(
             asAsyncIterable([
                 { choices: [{ delta: { content: 'hello' } }] },
@@ -1126,31 +1141,180 @@ describe('OpenRouterProvider.checkModeration', () => {
     });
 });
 
-
 describe('OpenRouter catalog refresh validation', () => {
     it('keeps missing model fields and prices unknown and blocks unpriced inference', async () => {
-        axiosRequestMock.mockResolvedValue({ data: { data: [{ id: 'fixture/model', name: 'Fixture' }] } });
+        axiosRequestMock.mockResolvedValue({
+            data: { data: [{ id: 'fixture/model', name: 'Fixture' }] },
+        });
         const { provider } = makeProvider();
         const [model] = await provider.models();
         expect(model.costs).toEqual({ tokens: 1_000_000 });
         expect(model.context).toBeUndefined();
         expect(model.max_tokens).toBe(0);
-        await expect(withTestActor(() => provider.complete({ model: 'openrouter:fixture/model', messages: [] })))
-            .rejects.toMatchObject({ legacyCode: 'model_pricing_unavailable' });
+        await expect(
+            withTestActor(() =>
+                provider.complete({
+                    model: 'openrouter:fixture/model',
+                    messages: [],
+                }),
+            ),
+        ).rejects.toMatchObject({ legacyCode: 'model_pricing_unavailable' });
         expect(createMock).not.toHaveBeenCalled();
     });
 
-    it.each(['malformed', 'failed', 'ambiguous conditional price'])('retains last-good data after a %s refresh', async failure => {
-        vi.useFakeTimers({ toFake: ['Date'] });
-        vi.spyOn(console, 'warn').mockImplementation(() => {});
+    it.each(['malformed', 'failed', 'ambiguous conditional price'])(
+        'retains last-good data after a %s refresh',
+        async (failure) => {
+            vi.useFakeTimers({ toFake: ['Date'] });
+            vi.spyOn(console, 'warn').mockImplementation(() => {});
+            const { provider } = makeProvider();
+            const good = await provider.models();
+            vi.advanceTimersByTime(15 * 60 * 1000 + 1);
+            if (failure === 'failed')
+                axiosRequestMock.mockRejectedValueOnce(
+                    new Error('unavailable'),
+                );
+            else if (failure === 'malformed')
+                axiosRequestMock.mockResolvedValueOnce({ data: {} });
+            else
+                axiosRequestMock.mockResolvedValueOnce({
+                    data: {
+                        data: [
+                            {
+                                ...SAMPLE_API_MODELS[0],
+                                pricing: {
+                                    prompt: '0.000002',
+                                    overrides: [
+                                        { min_prompt_tokens: 10, prompt: '' },
+                                    ],
+                                },
+                            },
+                        ],
+                    },
+                });
+            expect(await provider.models()).toBe(good);
+        },
+    );
+});
+
+describe('OpenRouter service speed offerings', () => {
+    it('loads and caches tier prices while reserving for endpoint and context differences', async () => {
         const { provider } = makeProvider();
-        const good = await provider.models();
-        vi.advanceTimersByTime(15 * 60 * 1000 + 1);
-        if (failure === 'failed') axiosRequestMock.mockRejectedValueOnce(new Error('unavailable'));
-        else if (failure === 'malformed') axiosRequestMock.mockResolvedValueOnce({ data: {} });
-        else axiosRequestMock.mockResolvedValueOnce({ data: { data: [{ ...SAMPLE_API_MODELS[0], pricing: {
-            prompt: '0.000002', overrides: [{ min_prompt_tokens: 10, prompt: '' }],
-        } }] } });
-        expect(await provider.models()).toBe(good);
+        const base = {
+            ...(await provider.models())[0],
+            provider: 'openrouter',
+        };
+        axiosRequestMock.mockClear();
+        axiosRequestMock.mockResolvedValueOnce({
+            data: {
+                data: {
+                    endpoints: [
+                        {
+                            service_tier: 'priority',
+                            pricing: {
+                                prompt: '0.000004',
+                                completion: '0.00002',
+                                input_cache_read: '0.0000002',
+                            },
+                        },
+                        {
+                            service_tier: 'priority',
+                            pricing: {
+                                prompt: '0.000005',
+                                completion: '0.00001',
+                            },
+                        },
+                        {
+                            service_tier: 'ultrafast',
+                            pricing: {
+                                prompt: '0.000012',
+                                completion: '0.00006',
+                                overrides: [
+                                    {
+                                        min_prompt_tokens: 272000,
+                                        prompt: '0.000024',
+                                        completion: '0.00009',
+                                    },
+                                ],
+                            },
+                        },
+                        {
+                            service_tier: 'priority',
+                            pricing: { prompt: 'invalid' },
+                        },
+                    ],
+                },
+            },
+        });
+        const offers = await provider.serviceSpeedModels(base);
+        expect(offers.fast?.costs).toMatchObject({
+            prompt: 500,
+            completion: 2000,
+            input_cache_read: 20,
+        });
+        expect(offers.ultrafast?.costs).toMatchObject({
+            prompt: 2400,
+            completion: 9000,
+        });
+        expect(await provider.serviceSpeedModels(base)).toBe(offers);
+        expect(axiosRequestMock).toHaveBeenCalledTimes(1);
+    });
+
+    it('uses standard when the tier catalog cannot be fetched', async () => {
+        const { provider } = makeProvider();
+        const base = {
+            ...(await provider.models())[0],
+            provider: 'openrouter',
+        };
+        axiosRequestMock.mockRejectedValueOnce(new Error('Unavailable'));
+        expect(await provider.serviceSpeedModels(base)).toEqual({});
+    });
+});
+
+describe('OpenRouter speed billing', () => {
+    it('requires an authoritative bill instead of charging the reservation ceiling', async () => {
+        const { selectServiceSpeed } =
+            await import('../../utils/serviceSpeed.js');
+        const { provider } = makeProvider();
+        const base = {
+            ...(await provider.models())[0],
+            provider: 'openrouter',
+        };
+        const selected = selectServiceSpeed(
+            base,
+            {
+                model: base.id,
+                messages: [{ role: 'user', content: 'Hello' }],
+                serviceSpeed: 'fast',
+            },
+            () => undefined,
+            {
+                fast: {
+                    ...base,
+                    costs: { ...base.costs, prompt: 500, completion: 2000 },
+                },
+            },
+        );
+        const completion = {
+            choices: [
+                {
+                    index: 0,
+                    message: { role: 'assistant', content: 'OK' },
+                    finish_reason: 'stop',
+                },
+            ],
+            usage: { prompt_tokens: 10, completion_tokens: 1 },
+        };
+        createMock.mockResolvedValueOnce({
+            ...completion,
+            service_tier: 'priority',
+            usage: { ...completion.usage, cost: undefined },
+        });
+        await expect(
+            withTestActor(() =>
+                provider.complete(selected.args, selected.model),
+            ),
+        ).rejects.toMatchObject({ legacyCode: 'model_pricing_unavailable' });
+        expect(recordSpy).not.toHaveBeenCalled();
     });
 });

src/backend/drivers/ai-chat/providers/openrouter/OpenRouterProvider.ts

@@ -22,8 +22,10 @@ import { OpenAI } from 'openai';
 import { HttpError, isHttpError } from '../../../../core/http/HttpError.js';
 import { UpstreamResponseError } from '../../../util/upstreamErrors.js';
 import type { MeteringService } from '../../../../services/metering/MeteringService.js';
-import type { IChatModel, ICompleteArguments } from '../../types.js';
+import type { IChatModel, ICompleteArguments, ModelCost } from '../../types.js';
+import { BoundedTtlMap } from '../../../../util/boundedTtlMap.js';
 import { cachedRemoteCatalog } from '../../utils/cachedRemoteCatalog.js';
+import { servedServiceSpeed } from '../../utils/serviceSpeed.js';
 import { catalogPrice } from '../../utils/pricing.js';
 import { mapOpenRouterPricing } from './pricing.js';
 import { sdkClientOptions } from '../../utils/sdkClient.js';
@@ -64,6 +66,75 @@ const isPolicyBlock = (err: unknown): boolean => {
 export class OpenRouterProvider extends OpenAICompatProvider {
     #apiBaseUrl: string;
 
+    #speedCatalogs = new BoundedTtlMap<
+        string,
+        () => Promise<Partial<Record<'fast' | 'ultrafast', IChatModel>>>
+    >({ maxEntries: 256 });
+
+    async serviceSpeedModels(model: IChatModel) {
+        let load = this.#speedCatalogs.get(model.id);
+        if (!load) {
+            load = cachedRemoteCatalog({
+                name: 'OpenRouter speed catalog',
+                fallback: {} as Partial<
+                    Record<'fast' | 'ultrafast', IChatModel>
+                >,
+                fetch: async (signal) => {
+                    const id = model.id.replace(/^openrouter:/, '');
+                    const response = await axios.request({
+                        url: `${this.#apiBaseUrl}/models/${id}/endpoints`,
+                        method: 'GET',
+                        signal,
+                        maxContentLength: 2 * 1024 * 1024,
+                    });
+                    const offers: Partial<
+                        Record<'fast' | 'ultrafast', IChatModel>
+                    > = {};
+                    for (const endpoint of response.data?.data?.endpoints ??
+                        []) {
+                        const speed =
+                            endpoint.service_tier === 'ultrafast'
+                                ? 'ultrafast'
+                                : ['priority', 'fast'].includes(
+                                        endpoint.service_tier,
+                                    )
+                                  ? 'fast'
+                                  : undefined;
+                        if (!speed || !endpoint.pricing) continue;
+                        const pricing = mapOpenRouterPricing(endpoint.pricing);
+                        if (
+                            pricing.costs.prompt === undefined ||
+                            pricing.costs.completion === undefined
+                        )
+                            continue;
+                        // Reserve for the most expensive eligible endpoint and context tier.
+                        const costs: ModelCost = {
+                            ...offers[speed]?.costs,
+                            tokens: model.costs.tokens,
+                        };
+                        for (const rates of [
+                            pricing.costs,
+                            ...pricing.overrides.map(
+                                (override) => override.costs,
+                            ),
+                        ]) {
+                            for (const [key, value] of Object.entries(rates))
+                                costs[key] = Math.max(costs[key] ?? 0, value);
+                        }
+                        offers[speed] = {
+                            ...model,
+                            costs,
+                            pricingOverrides: [],
+                        };
+                    }
+                    return offers;
+                },
+            });
+            this.#speedCatalogs.set(model.id, load);
+        }
+        return load();
+    }
+
     constructor(config: ChatProviderConfig, meteringService: MeteringService) {
         const apiBaseUrl = config.apiBaseUrl || 'https://openrouter.ai/api/v1';
         super(meteringService, {
@@ -202,8 +273,17 @@ export class OpenRouterProvider extends OpenAICompatProvider {
         return { ...params, usage: { include: true } };
     }
 
-    protected override meteredUsage(source: UsageSource) {
+    protected override meteredUsage(source: UsageSource, model: IChatModel) {
         const usage = source.usage as OpenrouterUsage;
+        if (
+            model.serviceSpeedPricing &&
+            servedServiceSpeed(source) !== 'standard' &&
+            catalogPrice(usage.cost) === undefined
+        ) {
+            throw new HttpError(503, 'Service tier billing is unavailable', {
+                legacyCode: 'model_pricing_unavailable',
+            });
+        }
         return {
             usage: {
                 prompt:

src/backend/drivers/ai-chat/providers/xai/XAIProvider.test.ts

@@ -20,13 +20,12 @@
 /**
  * Offline unit tests for XAIProvider.
  *
- * Boots a real PuterServer (in-memory sqlite + dynamo + s3 + mock
- * redis) and constructs XAIProvider directly against the live wired
- * `MeteringService` so the recording side is exercised end-to-end.
- * xAI is OpenAI-compatible so the OpenAI SDK is mocked at the module
- * boundary; that's the real network egress point. The companion
- * integration test (XAIProvider.integration.test.ts) exercises the
- * real xAI endpoint.
+ * Boots a real PuterServer (in-memory sqlite + dynamo + s3 + mock redis) and
+ * constructs XAIProvider directly against the live wired `MeteringService` so
+ * the recording side is exercised end-to-end. xAI is OpenAI-compatible so the
+ * OpenAI SDK is mocked at the module boundary; that's the real network egress
+ * point. The companion integration test (XAIProvider.integration.test.ts)
+ * exercises the real xAI endpoint.
  */
 
 import { Writable } from 'node:stream';
@@ -933,8 +932,7 @@ describe('XAIProvider.complete non-stream output', () => {
         // namespace, actor, and a costsOverride priced from
         // models.ts (grok-4.6: prompt=200, completion=600, cached=50).
         expect(recordSpy).toHaveBeenCalledTimes(1);
-        const [usage, actor, prefix, overrides] =
-            recordSpy.mock.calls[0]!;
+        const [usage, actor, prefix, overrides] = recordSpy.mock.calls[0]!;
         expect(usage).toEqual({
             prompt_tokens: 90,
             completion_tokens: 50,
@@ -1022,8 +1020,7 @@ describe('XAIProvider.complete non-stream output', () => {
             }),
         );
 
-        const [usage, , , overrides] =
-            recordSpy.mock.calls[0]!;
+        const [usage, , , overrides] = recordSpy.mock.calls[0]!;
         expect(usage.cached_tokens).toBe(0);
         // 0 cached tokens × any rate = 0 metered cost.
         expect(overrides).toMatchObject({ cached_tokens: 0 });
@@ -1081,8 +1078,7 @@ describe('XAIProvider.complete streaming', () => {
 
         // grok-build-0.1 costs: prompt=100, completion=200, cached=20.
         expect(recordSpy).toHaveBeenCalledTimes(1);
-        const [, , prefix, overrides] =
-            recordSpy.mock.calls[0]!;
+        const [, , prefix, overrides] = recordSpy.mock.calls[0]!;
         expect(prefix).toBe('xai:grok-build-0.1');
         expect(overrides).toEqual({
             prompt_tokens: 3 * 100,
@@ -1200,3 +1196,60 @@ describe('XAIProvider.checkModeration', () => {
         );
     });
 });
+
+describe('xAI service speed billing', () => {
+    it('records the reported Priority charge with cached input and reasoning', async () => {
+        const { selectServiceSpeed } =
+            await import('../../utils/serviceSpeed.js');
+        const { provider } = makeProvider();
+        const base = {
+            ...(await provider.models()).find(
+                (entry) => entry.id === 'grok-4.7',
+            )!,
+            provider: 'xai',
+        };
+        const selected = selectServiceSpeed(
+            base,
+            {
+                model: base.id,
+                messages: [{ role: 'user', content: 'Hello' }],
+                serviceSpeed: 'ultrafast',
+            },
+            () => undefined,
+        );
+        createMock.mockResolvedValueOnce({
+            choices: [
+                {
+                    index: 0,
+                    message: { role: 'assistant', content: 'OK' },
+                    finish_reason: 'stop',
+                },
+            ],
+            service_tier: 'priority',
+            usage: {
+                prompt_tokens: 1247,
+                completion_tokens: 1,
+                total_tokens: 1282,
+                prompt_tokens_details: { cached_tokens: 1152 },
+                completion_tokens_details: { reasoning_tokens: 34 },
+                cost_in_usd_ticks: 19520000,
+            },
+        });
+        const result = await withTestActor(() =>
+            provider.complete(selected.args, selected.model),
+        );
+        expect(result).toMatchObject({
+            usageDetails: {
+                serviceSpeed: 'fast',
+                outputTokens: 35,
+                reasoningTokens: 34,
+            },
+        });
+        expect(recordSpy).toHaveBeenCalledWith(
+            expect.objectContaining({ billedUsage: 1, completion_tokens: 35 }),
+            expect.anything(),
+            'xai:grok-4.7',
+            expect.objectContaining({ billedUsage: 195200 }),
+        );
+    });
+});

src/backend/drivers/ai-chat/providers/xai/XAIProvider.ts

@@ -19,6 +19,8 @@
 
 import { OpenAI } from 'openai';
 import { ChatCompletionCreateParams } from 'openai/resources/index.js';
+import { catalogPrice } from '../../utils/pricing.js';
+import { servedServiceSpeed } from '../../utils/serviceSpeed.js';
 import { Context } from '../../../../core/context.js';
 import type { MeteringService } from '../../../../services/metering/MeteringService.js';
 import * as OpenAIUtil from '../../utils/OpenAIUtil.js';
@@ -74,7 +76,15 @@ export class XAIProvider implements IChatProvider {
     }
 
     async complete(
-        { messages, stream, model, tools, max_tokens }: ICompleteArguments,
+        {
+            messages,
+            stream,
+            model,
+            tools,
+            max_tokens,
+            service_tier,
+            serviceSpeed: requestedServiceSpeed,
+        }: ICompleteArguments,
         resolved?: IChatModel,
     ): Promise<IChatCompleteResult> {
         const actor = Context.get('actor');
@@ -88,6 +98,7 @@ export class XAIProvider implements IChatProvider {
                 {
                     messages,
                     model: modelUsed.id,
+                    ...(service_tier !== undefined ? { service_tier } : {}),
                     ...(tools ? { tools } : {}),
                     ...(max_tokens !== undefined ? { max_tokens } : {}),
                     stream,
@@ -107,18 +118,49 @@ export class XAIProvider implements IChatProvider {
         }
 
         return OpenAIUtil.handle_completion_output({
-            usage_calculator: ({ usage, setUsageCosts }) => {
+            usage_calculator: ({
+                usage,
+                service_tier: servedTier,
+                setUsageCosts,
+                setUsageDetails,
+            }) => {
+                const serviceSpeed = servedServiceSpeed({
+                    service_tier: servedTier,
+                });
                 // xAI reports cached reads inside `prompt_tokens`; bill them
                 // only at the cached rate.
                 const metered = OpenAIUtil.extractMeteredUsage(usage);
+                const reasoningTokens =
+                    usage.completion_tokens_details?.reasoning_tokens ?? 0;
+                const trackedUsage = OpenAIUtil.splitCachedPrompt(usage);
+                trackedUsage.completion_tokens += reasoningTokens;
+                const reportedTicks = catalogPrice(
+                    (usage as typeof usage & { cost_in_usd_ticks?: number })
+                        .cost_in_usd_ticks,
+                );
                 const { costs } = meterChatUsage(
                     this.#meteringService,
                     actor,
                     this.meteringModelKey(modelUsed.id),
                     modelUsed,
-                    OpenAIUtil.splitCachedPrompt(usage),
+                    trackedUsage,
+                    {
+                        serviceSpeed,
+                        authoritativeUsd:
+                            reportedTicks === undefined
+                                ? undefined
+                                : reportedTicks / 10_000_000_000,
+                    },
                 );
                 setUsageCosts(costs);
+                if (requestedServiceSpeed !== undefined)
+                    setUsageDetails?.({
+                        ...OpenAIUtil.usageDetailsFromTrackedUsage(
+                            trackedUsage,
+                        )!,
+                        serviceSpeed,
+                        ...(reasoningTokens ? { reasoningTokens } : {}),
+                    });
                 return metered;
             },
             stream,

src/backend/drivers/ai-chat/providers/xai/models.ts

@@ -42,6 +42,7 @@ export const XAI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 2 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 200,
@@ -65,6 +66,7 @@ export const XAI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 2 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 200,
@@ -87,6 +89,7 @@ export const XAI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 2 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 200,
@@ -109,6 +112,7 @@ export const XAI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 2 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 125,
@@ -151,6 +155,7 @@ export const XAI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 2 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 125,
@@ -186,6 +191,7 @@ export const XAI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 2 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 125,
@@ -217,6 +223,7 @@ export const XAI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 2 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 125,
@@ -246,6 +253,7 @@ export const XAI_MODELS: IChatModel[] = [
         costs_currency: 'usd-cents',
         input_cost_key: 'prompt_tokens',
         output_cost_key: 'completion_tokens',
+        serviceSpeedRates: { fast: 2 },
         costs: {
             tokens: 1_000_000,
             prompt_tokens: 100,

src/backend/drivers/ai-chat/types.ts

@@ -26,6 +26,8 @@
 
 export type ModelCost = Record<string, number>;
 
+export type ServiceSpeed = 'standard' | 'fast' | 'ultrafast';
+
 export interface ModelModalities {
     input: string[];
     output: string[];
@@ -67,6 +69,22 @@ export interface IChatModel<T extends ModelCost = ModelCost> extends Record<
         output_multiplier: number;
     };
     pricingOverrides?: ModelPricingOverride[];
+    /**
+     * A token-rate multiplier or explicit per-key rates merged with standard
+     * costs.
+     */
+    serviceSpeedRates?: Partial<
+        Record<'fast' | 'ultrafast', number | ModelCost>
+    >;
+    serviceSpeedVariants?: { fast?: string };
+    /** Request-local pricing snapshot; never part of the model catalog. */
+    serviceSpeedPricing?: {
+        speed: ServiceSpeed;
+        standardCosts: ModelCost;
+        standardOverrides?: ModelPricingOverride[];
+        fastCosts?: ModelCost;
+        ultrafastCosts?: ModelCost;
+    };
     context?: number;
     max_tokens: number;
     subscriberOnly?: boolean;
@@ -76,6 +94,8 @@ export interface IChatModel<T extends ModelCost = ModelCost> extends Record<
     tool_call?: boolean;
     responses_api?: boolean;
     responses_api_only?: boolean;
+    /** Restricts service speeds for deployments with a limited API dialect. */
+    serviceSpeeds?: ServiceSpeed[];
     /**
      * Omitted preserves sampling; reasoningDisabled requires explicit effort
      * none.
@@ -163,6 +183,7 @@ export interface UsageDetails {
     webFetchRequests?: number;
     speed?: 'fast' | 'standard';
     serviceTier?: string;
+    serviceSpeed?: ServiceSpeed;
     /**
      * Per-pass breakdown when the upstream ran several passes (compaction,
      * advisor).
@@ -198,6 +219,7 @@ export interface ICompleteArguments {
      * request on another provider that serves the same model.
      */
     fallback?: boolean;
+    serviceSpeed?: ServiceSpeed;
     stream?: boolean;
     model: string;
     test_mode?: boolean;
@@ -217,10 +239,7 @@ export interface ICompleteArguments {
     taskBudget?: { total: number; remaining?: number };
     /** Claude only: top-level automatic `cache_control`. */
     cacheControl?: CacheControl;
-    /**
-     * Claude only; forwarded only when the model's catalog entry has
-     * `fast_output_tokens`.
-     */
+    /** Claude only; fast is forwarded when the catalog declares fast rates. */
     speed?: 'fast' | 'standard';
     /** Claude only. */
     safeguards?: SafeguardRequest[];
@@ -260,7 +279,15 @@ export interface ICompleteArguments {
     truncation?: 'auto' | 'disabled' | undefined;
     background?: boolean;
     service_tier?:
-        'auto' | 'default' | 'flex' | 'scale' | 'priority' | undefined;
+        | 'auto'
+        | 'default'
+        | 'flex'
+        | 'scale'
+        | 'priority'
+        | 'fast'
+        | 'ultrafast'
+        | 'standard'
+        | undefined;
     max_tokens?: number;
     temperature?: number;
     reasoning?: { effort: ReasoningEffort } | undefined;
@@ -328,6 +355,10 @@ export interface IChatMessageResult {
 export type IChatCompleteResult = IChatStreamResult | IChatMessageResult;
 
 export interface IChatProvider {
+    /** Account/catalog-specific speed offerings, fetched before the credit hold. */
+    serviceSpeedModels?(
+        model: IChatModel,
+    ): Promise<Partial<Record<'fast' | 'ultrafast', IChatModel>>>;
     models(extra_params?: unknown): IChatModel[] | Promise<IChatModel[]>;
     getDefaultModel(): string;
     /**

src/backend/drivers/ai-chat/utils/OpenAIUtil.js

@@ -650,12 +650,15 @@ export const create_chat_stream_handler =
             block.end();
         };
 
+        let serviceTier;
         let last_usage = null;
         let last_extra_content = null;
         let finish_reason = null;
         for await (let chunk of completion) {
             await chatStream.waitForDrain();
             chunk = deviations.chunk_but_like_actually(chunk);
+            if (chunk.service_tier !== undefined)
+                serviceTier = chunk.service_tier;
             const chunk_usage = deviations.index_usage_from_stream_chunk(chunk);
             if (chunk_usage) last_usage = chunk_usage;
             if (chunk.choices.length < 1) continue;
@@ -746,6 +749,7 @@ export const create_chat_stream_handler =
         const usage = last_usage
             ? usage_calculator({
                   usage: last_usage,
+                  service_tier: serviceTier,
                   extra_content: last_extra_content,
                   setUsageCosts: (c) => chatStream.setUsageCosts(c),
                   setUsageDetails: (d) => {
@@ -779,6 +783,7 @@ export const create_chat_stream_handler_responses_api =
         const message = chatStream.message();
         const textblock = message.contentBlock({ type: 'text' });
 
+        let serviceTier;
         let last_usage = null;
         let completed = false;
         let sawFunctionCall = false;
@@ -787,6 +792,8 @@ export const create_chat_stream_handler_responses_api =
         let toolUsage = null;
         for await (const chunk of completion) {
             await chatStream.waitForDrain();
+            if (chunk.response?.service_tier !== undefined)
+                serviceTier = chunk.response.service_tier;
             if (chunk.type === 'response.failed' || chunk.type === 'error') {
                 throw new Error(
                     chunk.response?.error?.message ??
@@ -918,6 +925,7 @@ export const create_chat_stream_handler_responses_api =
         const usage = last_usage
             ? usage_calculator({
                   usage: last_usage,
+                  service_tier: serviceTier,
                   webSearchCalls,
                   tool_usage: toolUsage,
                   setUsageCosts: (c) => chatStream.setUsageCosts(c),
@@ -972,6 +980,7 @@ export const handle_completion_output = async (
     /**
      * @type {Record<string, unknown> & {
      *     usage_calculator: (args: {
+     *         service_tier?: string;
      *         usage: import('openai/resources/completions.mjs').CompletionUsage;
      *         setUsageCosts: (costs: Record<string, number>) => void;
      *         setUsageDetails: (
@@ -1066,6 +1075,7 @@ export const handle_completion_output = async (
  * @param {((text: string) => Promise<{ flagged: boolean }>) | undefined} [params.moderate]
  * @param {(args: {
  *     usage: import('openai/resources/responses/responses.mjs').ResponseUsage;
+ *     service_tier?: string;
  *     webSearchCalls?: number;
  *     tool_usage?: Record<string, unknown>;
  *     setUsageCosts?: (costs: Record<string, number>) => void;

src/backend/drivers/ai-chat/utils/meterChatUsage.ts

@@ -19,13 +19,15 @@
 
 import type { Actor } from '../../../core/actor.js';
 import type { MeteringService } from '../../../services/metering/MeteringService.js';
-import type { IChatModel } from '../types.js';
+import type { IChatModel, ServiceSpeed } from '../types.js';
+import { serviceSpeedBillingModel } from './serviceSpeed.js';
 import { buildCostsOverride } from './pricing.js';
 
 /** Microcents per US dollar. */
 const MICROCENTS_PER_USD = 100_000_000;
 
 export interface MeterChatUsageOptions {
+    serviceSpeed?: ServiceSpeed;
     /**
      * What the upstream says it billed, in USD. Recorded as one `billedUsage`
      * line instead of pricing the tokens, which are still recorded at zero.
@@ -77,10 +79,15 @@ export const meterChatUsage = (
         authoritativeUsd,
         costOverrides,
         pricingTime,
+        serviceSpeed = 'standard',
     }: MeterChatUsageOptions = {},
 ): { usage: Record<string, number>; costs: Record<string, number> } => {
     const priced = {
-        ...buildCostsOverride(usage, model, pricingTime),
+        ...buildCostsOverride(
+            usage,
+            serviceSpeedBillingModel(model, serviceSpeed),
+            pricingTime,
+        ),
         ...costOverrides,
     };
     if (authoritativeUsd !== undefined) {

src/backend/drivers/ai-chat/utils/serviceSpeed.fuzz.test.ts

@@ -0,0 +1,150 @@
+/*
+ * Copyright (C) 2024-present Puter Technologies Inc.
+ *
+ * This file is part of Puter.
+ *
+ * Puter is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU Affero General Public License as published
+ * by the Free Software Foundation, either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * Puter is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+ * GNU Affero General Public License for more details.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with Puter. If not, see <https://www.gnu.org/licenses/>.
+ */
+
+import { describe, expect, it } from 'vitest';
+import type { IChatModel, ServiceSpeed } from '../types.js';
+import { ALIBABA_MODELS } from '../providers/alibaba/models.js';
+import { AZURE_MODELS } from '../providers/azure/models.js';
+import { BYTEPLUS_MODELS } from '../providers/byteplus/models.js';
+import { CLAUDE_MODELS } from '../providers/claude/models.js';
+import { DEEPSEEK_MODELS } from '../providers/deepseek/models.js';
+import { GEMINI_MODELS } from '../providers/gemini/models.js';
+import { GROQ_MODELS } from '../providers/groq/models.js';
+import { HOONIFY_MODELS } from '../providers/hoonify/models.js';
+import { META_MODELS } from '../providers/meta/models.js';
+import { MINIMAX_MODELS } from '../providers/minimax/models.js';
+import { MISTRAL_MODELS } from '../providers/mistral/models.js';
+import { MOONSHOT_MODELS } from '../providers/moonshot/models.js';
+import { OPEN_AI_MODELS } from '../providers/openai/models.js';
+import { XAI_MODELS } from '../providers/xai/models.js';
+import { ZAI_MODELS } from '../providers/zai/models.js';
+import { shouldPresentAsOpenAI } from './normalizeToOpenAI.js';
+import {
+    isServiceSpeedUnavailable,
+    lowerServiceSpeed,
+    selectServiceSpeed,
+    servedServiceSpeed,
+    serviceSpeedBillingModel,
+} from './serviceSpeed.js';
+
+const catalogs: Array<[string, IChatModel[]]> = [
+    ['alibaba', ALIBABA_MODELS],
+    ['azure-openai', AZURE_MODELS],
+    ['byteplus', BYTEPLUS_MODELS],
+    ['claude', CLAUDE_MODELS],
+    ['deepseek', DEEPSEEK_MODELS],
+    ['gemini', GEMINI_MODELS],
+    ['groq', GROQ_MODELS],
+    ['hoonify', HOONIFY_MODELS],
+    ['meta', META_MODELS],
+    ['minimax', MINIMAX_MODELS],
+    ['mistral', MISTRAL_MODELS],
+    ['moonshotai', MOONSHOT_MODELS],
+    ['openai-completion', OPEN_AI_MODELS],
+    ['openai-responses', OPEN_AI_MODELS],
+    ['xai', XAI_MODELS],
+    ['zai', ZAI_MODELS],
+];
+const speeds: ServiceSpeed[] = ['standard', 'fast', 'ultrafast'];
+
+describe('service speed seeded fuzzing', () => {
+    it('preserves catalog, request, normalization, and billing invariants across model changes', () => {
+        let seed = 0x51eed;
+        const pick = <T>(values: T[]): T => {
+            seed = (Math.imul(seed, 1664525) + 1013904223) >>> 0;
+            return values[seed % values.length]!;
+        };
+        const seenProviders = new Set<string>();
+        for (let index = 0; index < 10_000; index++) {
+            const [provider, catalog] = pick(catalogs);
+            seenProviders.add(provider);
+            const base = { ...structuredClone(pick(catalog)), provider };
+            const before = structuredClone(base);
+            Object.freeze(base.costs);
+            Object.freeze(base);
+            const args = Object.freeze({
+                model: base.id,
+                messages: [{ role: 'user', content: `Case ${index}` }],
+                serviceSpeed: pick([...speeds, undefined]),
+                normalize: pick([true, false, undefined]),
+                temperature: pick([0, 0.5, 1]),
+                service_tier: pick(['flex' as const, 'ultrafast' as const, undefined]),
+                speed: pick(['standard' as const, 'fast' as const, undefined]),
+            });
+            const selected = selectServiceSpeed(base, args, (id) => {
+                const variant = catalog.find((entry) => entry.id === id);
+                return variant && { ...variant, provider };
+            });
+            expect(base).toEqual(before);
+            expect(selected.args.messages).toBe(args.messages);
+            expect(selected.args.temperature).toBe(args.temperature);
+            expect(selected.model.provider).toBe(provider);
+            expect(selected.model.release_date).toBe(base.release_date);
+            expect(
+                shouldPresentAsOpenAI(selected.args, selected.model.release_date),
+            ).toBe(shouldPresentAsOpenAI(args, base.release_date));
+            if (args.serviceSpeed === undefined) {
+                expect(selected.args).toBe(args);
+                expect(selected.model).toBe(base);
+                continue;
+            }
+            expect(speeds.indexOf(selected.speed)).toBeLessThanOrEqual(
+                speeds.indexOf(args.serviceSpeed),
+            );
+            expect(selected.args.serviceSpeed).toBe(selected.speed);
+            if (args.serviceSpeed === 'standard') {
+                expect(selected.args.speed).not.toBe('fast');
+                expect(selected.args.service_tier).not.toBe('ultrafast');
+            }
+            for (const cost of Object.values(selected.model.costs)) {
+                expect(Number.isFinite(cost)).toBe(true);
+                expect(cost).toBeGreaterThanOrEqual(0);
+            }
+            if (selected.model.serviceSpeedPricing) {
+                const settled = serviceSpeedBillingModel(selected.model, 'standard');
+                expect(settled.costs).toEqual(base.costs);
+                expect(settled.pricingOverrides).toEqual(base.pricingOverrides);
+            }
+            let speed = selected.speed;
+            for (let attempts = 0; speed !== 'standard'; attempts++) {
+                expect(attempts).toBeLessThan(2);
+                speed = lowerServiceSpeed(speed);
+            }
+        }
+        expect([...seenProviders].sort()).toEqual(catalogs.map(([provider]) => provider).sort());
+    });
+
+    it('handles malformed upstream tier and admission-error metadata without coercion or retries', () => {
+        const values = [null, false, 0, 429, '', 'unrelated', [], {},
+            ['service_tier capacity unavailable'],
+            { toString: null }, { toString: 'priority capacity unavailable' }];
+        for (const value of values) {
+            expect(isServiceSpeedUnavailable(value)).toBe(false);
+            for (const status of [undefined, 400, 401, 403, 404, 429, 500, 503]) {
+                for (const variant of [false, true]) {
+                    expect(isServiceSpeedUnavailable({ status, message: value }, variant)).toBe(false);
+                    expect(isServiceSpeedUnavailable({ status, error: { message: value } }, variant)).toBe(false);
+                }
+            }
+            expect(servedServiceSpeed({ usage: value })).toBe('standard');
+            expect(servedServiceSpeed({ service_tier: value })).toBe('standard');
+            expect(servedServiceSpeed({ usage: { speed: value, service_tier: value } })).toBe('standard');
+        }
+    });
+});

src/backend/drivers/ai-chat/utils/serviceSpeed.integration.test.ts

@@ -0,0 +1,148 @@
+/*
+ * Copyright (C) 2024-present Puter Technologies Inc.
+ *
+ * This file is part of Puter.
+ *
+ * Puter is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU Affero General Public License as published
+ * by the Free Software Foundation, either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU Affero General Public License for more details.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program.  If not, see <https://www.gnu.org/licenses/>.
+ */
+
+import { describe, expect, it } from 'vitest';
+import {
+    makeMeteringStub,
+    optionalEnv,
+    withTestActor,
+} from '../../integrationTestUtil.js';
+import type {
+    IChatMessageResult,
+    IChatProvider,
+    ICompleteArguments,
+} from '../types.js';
+import { OpenAiResponsesChatProvider } from '../providers/openai/OpenAiChatResponsesProvider.js';
+import { GeminiChatProvider } from '../providers/gemini/GeminiChatProvider.js';
+import { MiniMaxProvider } from '../providers/minimax/MiniMaxProvider.js';
+import { MoonshotProvider } from '../providers/moonshot/MoonshotProvider.js';
+import { MistralAIProvider } from '../providers/mistral/MistralAiProvider.js';
+import { selectServiceSpeed } from './serviceSpeed.js';
+
+type Case = {
+    name: string;
+    env: string;
+    provider: string;
+    model: string;
+    build: (key: string) => IChatProvider;
+};
+const cases: Case[] = [
+    {
+        name: 'OpenAI Ultrafast',
+        env: 'PUTER_TEST_AI_OPENAI_API_KEY',
+        provider: 'openai-responses',
+        model: 'gpt-6.1-sol',
+        build: (key) =>
+            new OpenAiResponsesChatProvider(
+                { apiKey: key },
+                makeMeteringStub(),
+            ),
+    },
+    {
+        name: 'Gemini Priority',
+        env: 'PUTER_TEST_AI_GEMINI_API_KEY',
+        provider: 'gemini',
+        model: 'gemini-3.8-flash',
+        build: (key) =>
+            new GeminiChatProvider({ apiKey: key }, makeMeteringStub()),
+    },
+    {
+        name: 'MiniMax Priority',
+        env: 'PUTER_TEST_AI_MINIMAX_API_KEY',
+        provider: 'minimax',
+        model: 'minimax-m3',
+        build: (key) =>
+            new MiniMaxProvider({ apiKey: key }, makeMeteringStub()),
+    },
+    {
+        name: 'Moonshot Highspeed',
+        env: 'PUTER_TEST_AI_MOONSHOT_API_KEY',
+        provider: 'moonshotai',
+        model: 'kimi-k3.1-pro',
+        build: (key) =>
+            new MoonshotProvider({ apiKey: key }, makeMeteringStub()),
+    },
+    {
+        name: 'Mistral served tier',
+        env: 'PUTER_TEST_AI_MISTRAL_API_KEY',
+        provider: 'mistral',
+        model: 'mistral-medium-latest',
+        build: (key) =>
+            new MistralAIProvider({ apiKey: key }, makeMeteringStub()),
+    },
+];
+
+for (const testCase of cases) {
+    describe.skipIf(!optionalEnv(testCase.env))(
+        `serviceSpeed ${testCase.name} (integration)`,
+        () => {
+            it(
+                'completes with a selected speed and normalized text',
+                { timeout: 90_000 },
+                async () => {
+                    const provider = testCase.build(optionalEnv(testCase.env)!);
+                    const catalog = (await provider.models()).map((model) => ({
+                        ...model,
+                        provider: testCase.provider,
+                    }));
+                    const base = catalog.find(
+                        (model) =>
+                            model.id === testCase.model ||
+                            model.aliases?.includes(testCase.model),
+                    )!;
+                    expect(base).toBeDefined();
+                    const args: ICompleteArguments = {
+                        model: base.id,
+                        messages: [
+                            { role: 'user', content: 'Reply with exactly OK.' },
+                        ],
+                        serviceSpeed: 'ultrafast',
+                        max_tokens: 256,
+                        reasoning_effort: 'low',
+                        normalize: true,
+                    };
+                    const selection = selectServiceSpeed(base, args, (id) =>
+                        catalog.find((model) => model.id === id),
+                    );
+                    expect(selection.speed).toBe(
+                        testCase.provider === 'openai-responses'
+                            ? 'ultrafast'
+                            : 'fast',
+                    );
+                    const result = (await withTestActor(() =>
+                        provider.complete(
+                            { ...selection.args, model: selection.model.id },
+                            selection.model,
+                        ),
+                    )) as IChatMessageResult;
+                    expect(typeof result.message.content).toBe('string');
+                    expect(result.message.content.length).toBeGreaterThan(0);
+                    if (
+                        ['openai-responses', 'gemini', 'minimax'].includes(
+                            testCase.provider,
+                        )
+                    )
+                        expect(result.usageDetails?.serviceSpeed).toBe(
+                            selection.speed,
+                        );
+                },
+            );
+        },
+    );
+}

src/backend/drivers/ai-chat/utils/serviceSpeed.test.ts

@@ -0,0 +1,368 @@
+/*
+ * Copyright (C) 2024-present Puter Technologies Inc.
+ *
+ * This file is part of Puter.
+ *
+ * Puter is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU Affero General Public License as published
+ * by the Free Software Foundation, either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU Affero General Public License for more details.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program.  If not, see <https://www.gnu.org/licenses/>.
+ */
+
+import { describe, expect, it } from 'vitest';
+import { Writable } from 'node:stream';
+import { AIChatStream } from './Streaming.js';
+import {
+    handle_completion_output,
+    handle_completion_output_responses_api,
+} from './OpenAIUtil.js';
+import type { IChatModel, ServiceSpeed } from '../types.js';
+import {
+    selectServiceSpeed,
+    servedServiceSpeed,
+    serviceSpeedBillingModel,
+    isServiceSpeedUnavailable,
+} from './serviceSpeed.js';
+import { meterChatUsage } from './meterChatUsage.js';
+import { OPEN_AI_MODELS } from '../providers/openai/models.js';
+import { XAI_MODELS } from '../providers/xai/models.js';
+import { GEMINI_MODELS } from '../providers/gemini/models.js';
+import { CLAUDE_MODELS } from '../providers/claude/models.js';
+
+const model = (provider: string, id: string): IChatModel => ({
+    provider,
+    id,
+    serviceSpeedRates: [...OPEN_AI_MODELS, ...XAI_MODELS, ...GEMINI_MODELS, ...CLAUDE_MODELS]
+        .find((entry) => entry.id === id)?.serviceSpeedRates,
+    costs_currency: 'usd-cents',
+    input_cost_key: 'prompt_tokens',
+    output_cost_key: 'completion_tokens',
+    max_tokens: 1000,
+    costs: {
+        tokens: 1_000_000,
+        prompt_tokens: 200,
+        completion_tokens: 1000,
+        cached_tokens: 10,
+        web_search_requests: 1_000_000,
+    },
+});
+const request = (serviceSpeed?: ServiceSpeed) => ({
+    model: 'test',
+    messages: [{ role: 'user', content: 'Hello' }],
+    serviceSpeed,
+});
+const noVariant = () => undefined;
+
+describe('service speed selection', () => {
+    it.each([2, { prompt_tokens: 400, completion_tokens: 2000 }])(
+        'uses catalog rates independently of provider and model names: %j',
+        (rates) => {
+            const entry = {
+                ...model('catalog-provider', 'renamed-model'),
+                serviceSpeedRates: { fast: rates },
+            };
+            const selected = selectServiceSpeed(entry, request('ultrafast'), noVariant);
+            expect(selected.speed).toBe('fast');
+            expect(selected.model.costs.prompt_tokens).toBe(400);
+            expect(selected.model.costs.completion_tokens).toBe(2000);
+            expect(entry.costs.prompt_tokens).toBe(200);
+        },
+    );
+    it('preserves an omitted option and native parameters', () => {
+        const entry = model('openai-responses', 'gpt-6.1-sol');
+        const args = { ...request(), service_tier: 'flex' as const };
+        const selected = selectServiceSpeed(entry, args, noVariant);
+        expect(selected.model).toBe(entry);
+        expect(selected.args).toBe(args);
+    });
+    it.each([
+        'deepseek',
+        'zai',
+        'groq',
+        'byteplus',
+        'together-ai',
+        'meta',
+        'ollama',
+        'hoonify',
+        'neuralwatt',
+        'fake-chat',
+    ])('uses standard when %s has no verified faster offering', (provider) => {
+        const selected = selectServiceSpeed(
+            model(provider, 'some-model'),
+            request('ultrafast'),
+            noVariant,
+        );
+        expect(selected.speed).toBe('standard');
+        expect(selected.args.service_tier).toBeUndefined();
+    });
+    it.each(['standard', 'fast', 'ultrafast'] as const)(
+        'selects %s without changing the model or catalog prices',
+        (speed) => {
+            const entry = model('openai-responses', 'gpt-6.1-sol');
+            const selected = selectServiceSpeed(
+                entry,
+                request(speed),
+                noVariant,
+            );
+            expect(selected.speed).toBe(speed);
+            expect(selected.model.id).toBe(entry.id);
+            expect(selected.model.costs.prompt_tokens).toBe(
+                200 * { standard: 1, fast: 2, ultrafast: 6 }[speed],
+            );
+            expect(selected.model.costs.web_search_requests).toBe(1_000_000);
+            expect(entry.costs.prompt_tokens).toBe(200);
+        },
+    );
+    it('downgrades ultrafast to fast on a model without ultrafast', () => {
+        const selected = selectServiceSpeed(
+            model('xai', 'grok-4.7'),
+            request('ultrafast'),
+            noVariant,
+        );
+        expect(selected.speed).toBe('fast');
+        expect(selected.args.service_tier).toBe('priority');
+    });
+    it('keeps a compatibility-only Gemini deployment at standard speed', () => {
+        const entry = {
+            ...model('gemini', 'gemini-3.8-flash'),
+            serviceSpeeds: ['standard'] as ServiceSpeed[],
+        };
+        expect(
+            selectServiceSpeed(entry, request('ultrafast'), noVariant).speed,
+        ).toBe('standard');
+    });
+    it('uses Claude speed independently of its service tier', () => {
+        const entry = model('claude', 'claude-opus-5-5');
+        expect(
+            selectServiceSpeed(entry, request('ultrafast'), noVariant).args
+                .speed,
+        ).toBe('fast');
+        expect(
+            selectServiceSpeed(entry, request('ultrafast'), noVariant).args
+                .service_tier,
+        ).toBeUndefined();
+        expect(
+            servedServiceSpeed({
+                usage: { speed: 'fast', service_tier: 'standard' },
+            }),
+        ).toBe('fast');
+    });
+    it.each([
+        ['minimax', 'minimax-m2.7', 'minimax-m2.7-highspeed'],
+        ['moonshotai', 'kimi-k3.1-pro', 'kimi-k3.1-pro-highspeed'],
+        ['alibaba', 'glm-5.3', 'glm-5.3-prime'],
+        ['alibaba', 'glm-5.2', 'glm-5.2-fast-preview'],
+        ['infron', 'infron:vendor/model', 'infron:vendor/model:priority'],
+    ])(
+        'selects an equivalent %s variant and preserves normalization policy',
+        (provider, id, variantId) => {
+            const base = {
+                ...model(provider, id),
+                serviceSpeedVariants: { fast: variantId },
+                release_date: '2026-08-01',
+            };
+            const variant = {
+                ...model(provider, variantId),
+                release_date: '2026-10-01',
+            };
+            const selected = selectServiceSpeed(
+                base,
+                request('ultrafast'),
+                (candidate) => (candidate === variantId ? variant : undefined),
+            );
+            expect(selected.model.id).toBe(variantId);
+            expect(selected.model.release_date).toBe(base.release_date);
+            expect(selected.speed).toBe('fast');
+            expect(
+                selectServiceSpeed(base, request('fast'), noVariant).speed,
+            ).toBe('standard');
+        },
+    );
+    it('keeps caller media on Gemini compatibility when native speed cannot carry it', () => {
+        const args = {
+            ...request('fast'),
+            messages: [
+                {
+                    role: 'user',
+                    content: [
+                        {
+                            type: 'video_url',
+                            video_url: { url: 'data:video/mp4;base64,AA==' },
+                        },
+                    ],
+                },
+            ],
+        };
+        expect(
+            selectServiceSpeed(
+                model('gemini', 'gemini-3.8-flash'),
+                args,
+                noVariant,
+            ).speed,
+        ).toBe('standard');
+    });
+});
+
+describe('served speed billing', () => {
+    it.each(['standard', 'fast', 'ultrafast'] as const)(
+        'settles an ultrafast hold at the served %s rate',
+        (speed) => {
+            const selected = selectServiceSpeed(
+                model('openai-responses', 'gpt-6.1-sol'),
+                request('ultrafast'),
+                noVariant,
+            ).model;
+            const recorded: Array<Record<string, number> | undefined> = [];
+            const meter = {
+                utilRecordUsageObject: (
+                    _usage: Record<string, number>,
+                    _actor: unknown,
+                    _key: string,
+                    costs?: Record<string, number>,
+                ) => {
+                    recorded.push(costs);
+                },
+            };
+            const result = meterChatUsage(
+                meter as never,
+                undefined,
+                'openai:gpt-6.1-sol',
+                selected,
+                { prompt_tokens: 10, completion_tokens: 5, cached_tokens: 20 },
+                { serviceSpeed: speed },
+            );
+            expect(result.costs.prompt_tokens).toBe(
+                2000 * { standard: 1, fast: 2, ultrafast: 6 }[speed],
+            );
+            expect(result.costs.cached_tokens).toBe(
+                200 * { standard: 1, fast: 2, ultrafast: 6 }[speed],
+            );
+            expect(recorded).toEqual([result.costs]);
+        },
+    );
+    it('restores standard scheduled and long-context rates after a downgrade', () => {
+        const entry = model('openai-responses', 'gpt-6.1-sol');
+        entry.pricingOverrides = [
+            {
+                minPromptTokens: 100,
+                costs: { prompt_tokens: 400, completion_tokens: 1500 },
+            },
+        ];
+        const selected = selectServiceSpeed(
+            entry,
+            request('ultrafast'),
+            noVariant,
+        ).model;
+        expect(
+            serviceSpeedBillingModel(selected, 'standard').pricingOverrides,
+        ).toEqual(entry.pricingOverrides);
+        expect(
+            serviceSpeedBillingModel(selected, 'fast').pricingOverrides?.[0]
+                .costs.prompt_tokens,
+        ).toBe(800);
+    });
+    it.each([
+        [{ service_tier: 'priority' }, 'fast'],
+        [{ usage: { serviceTier: 'priority' } }, 'fast'],
+        [{ usage: { service_tier: 'standard' } }, 'standard'],
+        [{ service_tier: 'ultrafast' }, 'ultrafast'],
+        [{}, 'standard'],
+    ])('reads actual speed from %j', (source, speed) => {
+        expect(servedServiceSpeed(source)).toBe(speed);
+    });
+});
+
+describe('speed admission failures', () => {
+    it.each([
+        [{ status: 429, message: 'Fast mode rate limit exceeded' }, true],
+        [{ status: 403, message: 'Ultrafast access unavailable' }, true],
+        [{ status: 400, message: 'service_tier is not supported' }, true],
+        [{ status: 401, message: 'Invalid API key' }, false],
+        [{ status: 429, message: 'Rate limit exceeded' }, false],
+        [{ status: 400, message: 'Invalid messages' }, false],
+    ])('classifies %j without swallowing unrelated errors', (error, retry) => {
+        expect(isServiceSpeedUnavailable(error)).toBe(retry);
+    });
+    it('downgrades a missing speed variant without retrying an invalid base model', () => {
+        const error = { status: 404, message: 'Model not found' };
+        expect(isServiceSpeedUnavailable(error, true)).toBe(true);
+        expect(isServiceSpeedUnavailable(error)).toBe(false);
+    });
+});
+
+describe('streamed service tier metadata', () => {
+    it('retains a tier from an earlier Chat Completions chunk until usage arrives', async () => {
+        let servedTier: string | undefined;
+        const result = await handle_completion_output({
+            stream: true,
+            completion: (async function* () {
+                yield {
+                    service_tier: 'priority',
+                    choices: [{ index: 0, delta: { content: 'OK' } }],
+                };
+                yield {
+                    choices: [{ index: 0, delta: {}, finish_reason: 'stop' }],
+                };
+                yield {
+                    choices: [],
+                    usage: { prompt_tokens: 10, completion_tokens: 1 },
+                };
+            })(),
+            usage_calculator: (source) => {
+                servedTier = source.service_tier;
+                return source.usage;
+            },
+        });
+        const chatStream = new AIChatStream({
+            stream: new Writable({
+                write(_chunk, _encoding, callback) {
+                    callback();
+                },
+            }),
+        });
+        await result.init_chat_stream!({ chatStream });
+        expect(servedTier).toBe('priority');
+    });
+
+    it('uses the final Responses tier when the provider downgrades a stream', async () => {
+        let servedTier: string | undefined;
+        const result = await handle_completion_output_responses_api({
+            stream: true,
+            completion: (async function* () {
+                yield {
+                    type: 'response.created',
+                    response: { service_tier: 'ultrafast' },
+                };
+                yield { type: 'response.output_text.delta', delta: 'OK' };
+                yield {
+                    type: 'response.completed',
+                    response: {
+                        service_tier: 'default',
+                        usage: { input_tokens: 10, output_tokens: 1 },
+                    },
+                };
+            })(),
+            usage_calculator: (source) => {
+                servedTier = source.service_tier;
+                return source.usage;
+            },
+        });
+        const chatStream = new AIChatStream({
+            stream: new Writable({
+                write(_chunk, _encoding, callback) {
+                    callback();
+                },
+            }),
+        });
+        await result.init_chat_stream!({ chatStream });
+        expect(servedTier).toBe('default');
+    });
+});

src/backend/drivers/ai-chat/utils/serviceSpeed.ts

@@ -0,0 +1,261 @@
+/*
+ * Copyright (C) 2024-present Puter Technologies Inc.
+ *
+ * This file is part of Puter.
+ *
+ * Puter is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU Affero General Public License as published
+ * by the Free Software Foundation, either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU Affero General Public License for more details.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program.  If not, see <https://www.gnu.org/licenses/>.
+ */
+
+import { requiredInputModalities } from './mediaParts.js';
+import type {
+    IChatModel,
+    ICompleteArguments,
+    ModelCost,
+    ServiceSpeed,
+} from '../types.js';
+
+export const lowerServiceSpeed = (speed: ServiceSpeed): ServiceSpeed =>
+    speed === 'ultrafast' ? 'fast' : 'standard';
+
+const tokenRates = (costs: ModelCost, multiplier: number): ModelCost =>
+    Object.fromEntries(
+        Object.entries(costs).map(([key, value]) => [
+            key,
+            key !== 'tokens' &&
+            /token|^prompt$|^completion$|^input_cache_/.test(key)
+                ? value * multiplier
+                : value,
+        ]),
+    );
+
+export const serviceSpeedCosts = (
+    model: IChatModel,
+    speed: 'fast' | 'ultrafast',
+): ModelCost | undefined => {
+    if (model.serviceSpeeds && !model.serviceSpeeds.includes(speed))
+        return undefined;
+    const rates = model.serviceSpeedRates?.[speed];
+    return typeof rates === 'number'
+        ? tokenRates(model.costs, rates)
+        : rates
+          ? { ...model.costs, ...rates }
+          : undefined;
+};
+
+export interface ServiceSpeedSelection {
+    model: IChatModel;
+    args: ICompleteArguments;
+    speed: ServiceSpeed;
+}
+
+/** Select only an equivalent model from the same provider's loaded catalog. */
+export const selectServiceSpeed = (
+    model: IChatModel,
+    args: ICompleteArguments,
+    findVariant: (id: string) => IChatModel | undefined,
+    offers: Partial<Record<'fast' | 'ultrafast', IChatModel>> = {},
+): ServiceSpeedSelection => {
+    if (args.serviceSpeed === undefined)
+        return { model, args, speed: 'standard' };
+    const provider = model.provider;
+    const fast =
+        offers.fast?.costs ??
+        (provider === 'gemini' &&
+        (requiredInputModalities(args.messages).some(
+            (modality) => !['text', 'image'].includes(modality),
+        ) ||
+            args.messages.some(
+                (message) =>
+                    Array.isArray(message.content) &&
+                    message.content.some(
+                        (part: Record<string, unknown>) =>
+                            part &&
+                            [
+                                'input_audio',
+                                'audio_url',
+                                'document',
+                                'file',
+                                'input_file',
+                            ].includes(String(part.type)),
+                    ),
+            ))
+            ? undefined
+            : serviceSpeedCosts(model, 'fast'));
+    const ultra =
+        offers.ultrafast?.costs ?? serviceSpeedCosts(model, 'ultrafast');
+    let speed = args.serviceSpeed;
+    const variantId = model.serviceSpeedVariants?.fast;
+    const variant =
+        offers.fast && offers.fast.id !== model.id
+            ? offers.fast
+            : variantId
+              ? findVariant(variantId)
+              : undefined;
+    while (speed !== 'standard') {
+        if (speed === 'ultrafast' && ultra) break;
+        if (speed === 'fast' && (fast || variant)) break;
+        speed = lowerServiceSpeed(speed);
+    }
+    const request: ICompleteArguments = {
+        ...args,
+        serviceSpeed: speed,
+        speed: undefined,
+        service_tier: undefined,
+    };
+    if (speed === 'fast' && variant) {
+        return {
+            model: { ...variant, release_date: model.release_date },
+            args: request,
+            speed,
+        };
+    }
+    if (provider === 'gemini' && fast)
+        request.service_tier = speed === 'fast' ? 'priority' : 'standard';
+    else if (provider === 'claude')
+        request.speed = speed === 'fast' ? 'fast' : 'standard';
+    else if (provider === 'mistral')
+        request.service_tier = speed === 'fast' ? 'auto' : 'default';
+    else if (
+        provider === 'minimax' &&
+        model.serviceSpeedRates?.fast !== undefined
+    )
+        request.service_tier = speed === 'fast' ? 'priority' : 'standard';
+    else if (fast || ultra)
+        request.service_tier =
+            speed === 'ultrafast'
+                ? 'ultrafast'
+                : speed === 'fast'
+                  ? 'priority'
+                  : 'default';
+    if (
+        speed === 'standard' &&
+        [
+            'openai-completion',
+            'openai-responses',
+            'azure-openai',
+            'azure-openai-responses',
+            'openrouter',
+        ].includes(provider ?? '')
+    )
+        request.service_tier = 'default';
+    const selectedCosts =
+        speed === 'ultrafast' ? ultra : speed === 'fast' ? fast : undefined;
+    return {
+        args: request,
+        speed,
+        model: selectedCosts
+            ? {
+                  ...model,
+                  costs: selectedCosts,
+                  pricingOverrides: model.pricingOverrides?.map((override) => ({
+                      ...override,
+                      costs: Object.fromEntries(
+                          Object.entries(override.costs).map(([key, value]) => [
+                              key,
+                              model.costs[key] > 0
+                                  ? (value *
+                                        (selectedCosts[key] ??
+                                            model.costs[key])) /
+                                    model.costs[key]
+                                  : value,
+                          ]),
+                      ),
+                  })),
+                  serviceSpeedPricing: {
+                      speed,
+                      standardCosts: model.costs,
+                      standardOverrides: model.pricingOverrides,
+                      fastCosts: fast,
+                      ultrafastCosts: ultra,
+                  },
+              }
+            : model,
+    };
+};
+
+/** The response is authoritative; an absent tier never earns a premium. */
+export const servedServiceSpeed = (
+    source: Record<string, unknown>,
+): ServiceSpeed => {
+    const usage = source.usage as Record<string, unknown> | undefined;
+    const tier =
+        usage?.speed ??
+        usage?.serviceTier ??
+        usage?.service_tier ??
+        source.service_tier;
+    return tier === 'ultrafast'
+        ? 'ultrafast'
+        : tier === 'fast' || tier === 'priority'
+          ? 'fast'
+          : 'standard';
+};
+
+export const serviceSpeedBillingModel = (
+    model: IChatModel,
+    speed: ServiceSpeed,
+): IChatModel => {
+    const pricing = model.serviceSpeedPricing;
+    if (!pricing || pricing.speed === speed) return model;
+    const costs =
+        speed === 'fast'
+            ? (pricing.fastCosts ?? pricing.standardCosts)
+            : pricing.standardCosts;
+    return {
+        ...model,
+        costs,
+        pricingOverrides: pricing.standardOverrides?.map((override) => ({
+            ...override,
+            costs: Object.fromEntries(
+                Object.entries(override.costs).map(([key, value]) => [
+                    key,
+                    pricing.standardCosts[key] > 0
+                        ? (value * (costs[key] ?? pricing.standardCosts[key])) /
+                          pricing.standardCosts[key]
+                        : value,
+                ]),
+            ),
+        })),
+    };
+};
+
+/** Retry only failures that explicitly concern admission to the faster tier. */
+export const isServiceSpeedUnavailable = (
+    error: unknown,
+    variant = false,
+): boolean => {
+    const e = error as {
+        status?: number;
+        statusCode?: number;
+        message?: unknown;
+        error?: { message?: unknown };
+    };
+    const status = e?.status ?? e?.statusCode;
+    const message = e?.error?.message ?? e?.message ?? '';
+    if (typeof message !== 'string') return false;
+    if (
+        variant &&
+        [400, 404].includes(status ?? 0) &&
+        /model.*(?:not found|not exist|unavailable)/i.test(message)
+    )
+        return true;
+    return (
+        [400, 403, 429, 503].includes(status ?? 0) &&
+        /fast mode|ultrafast|service[_ ]?tier|priority (?:tier|processing|capacity)|speed parameter/i.test(
+            message,
+        ) &&
+        /unavailable|unsupported|not support|rate.limit|quota|capacity|not enabled|not allowed|access|exceed|invalid value/i.test(
+            message,
+        )
+    );
+};

src/backend/drivers/ai-tts/providers/gemini/GeminiTTSProvider.ts

@@ -37,7 +37,7 @@ import {
     postGeminiJson,
     geminiTokenCount as tokenCount,
 } from '../../../util/geminiRest.js';
-import { decodeStrictBase64 } from '../../../../util/mediaSniff.js';
+import { decodeStrictBase64InSlices } from '../../../../util/mediaSniff.js';
 import { isUpstreamTimeoutError } from '../../../util/upstreamErrors.js';
 import { SAMPLE_AUDIO_URL } from '../../../util/testMode.js';
 import type {
@@ -456,7 +456,8 @@ export class GeminiTTSProvider implements ITTSProvider {
                             },
                         );
                     }
-                    const { body, contentType } = this.#decodeAudio(audio);
+                    const { body, contentType } =
+                        await this.#decodeAudio(audio);
 
                     outputTokens = usageTokens(
                         usage?.output_tokens_by_modality,
@@ -474,7 +475,7 @@ export class GeminiTTSProvider implements ITTSProvider {
                         dataType: 'stream',
                         content_type: contentType,
                         chunked: true,
-                        stream: Readable.from(body),
+                        stream: Readable.from(body, { objectMode: false }),
                     };
                 } finally {
                     // Every outcome pays once for the counts it reports.
@@ -519,13 +520,15 @@ export class GeminiTTSProvider implements ITTSProvider {
     }
 
     /** WAV passes through byte for byte; raw PCM is wrapped in a WAV header. */
-    #decodeAudio(audio: InteractionContent): {
-        body: Buffer;
+    async #decodeAudio(audio: InteractionContent): Promise<{
+        body: Buffer[];
         contentType: string;
-    } {
+    }> {
         const { data, mime_type: mimeType } = audio;
         const bytes =
-            typeof data === 'string' ? decodeStrictBase64(data) : null;
+            typeof data === 'string'
+                ? await decodeStrictBase64InSlices(data)
+                : null;
         if (!bytes || typeof mimeType !== 'string') {
             throw upstreamFailure('Gemini TTS returned invalid audio data');
         }
@@ -541,7 +544,7 @@ export class GeminiTTSProvider implements ITTSProvider {
             ) {
                 throw upstreamFailure('Gemini TTS returned invalid audio data');
             }
-            return { body: bytes, contentType: 'audio/wav' };
+            return { body: [bytes], contentType: 'audio/wav' };
         }
         if (PCM_MIME_TYPES.has(essence) && bytes.length > 0) {
             const rate = Number(
@@ -551,8 +554,9 @@ export class GeminiTTSProvider implements ITTSProvider {
                 Number.isSafeInteger(rate) && rate > 0
                     ? rate
                     : DEFAULT_PCM_SAMPLE_RATE;
+            // Sent ahead of the samples, which are never copied behind it.
             return {
-                body: this.#wrapPcmInWav(bytes, sampleRate, 1, 16),
+                body: [this.#wavHeader(bytes.length, sampleRate, 1, 16), bytes],
                 contentType: 'audio/wav',
             };
         }
@@ -560,17 +564,15 @@ export class GeminiTTSProvider implements ITTSProvider {
     }
 
     /** Wrap raw PCM samples in a WAV container so browsers can play it. */
-    #wrapPcmInWav(
-        pcm: Buffer,
+    #wavHeader(
+        dataSize: number,
         sampleRate: number,
         channels: number,
         bitsPerSample: number,
     ): Buffer {
         const byteRate = (sampleRate * channels * bitsPerSample) / 8;
         const blockAlign = (channels * bitsPerSample) / 8;
-        const dataSize = pcm.length;
-        const headerSize = 44;
-        const buffer = Buffer.alloc(headerSize + dataSize);
+        const buffer = Buffer.alloc(44);
 
         // RIFF header
         buffer.write('RIFF', 0);
@@ -590,7 +592,6 @@ export class GeminiTTSProvider implements ITTSProvider {
         // data sub-chunk
         buffer.write('data', 36);
         buffer.writeUInt32LE(dataSize, 40);
-        pcm.copy(buffer, headerSize);
 
         return buffer;
     }

src/backend/drivers/ai-video/providers/gemini/GeminiVideoProvider.ts

@@ -25,7 +25,7 @@ import { insufficientCreditsError } from '../../../../services/metering/enforcem
 import type { UsageInput } from '../../../../services/metering/types.js';
 import { normalizeRawBase64ImageString } from '../../../../util/appIcon.js';
 import {
-    decodeStrictBase64,
+    decodeStrictBase64InSlices,
     sniffImageMime,
     sniffVideoMime,
 } from '../../../../util/mediaSniff.js';
@@ -405,7 +405,7 @@ export class GeminiVideoProvider extends VideoProvider {
                     'Gemini video generation did not complete the request',
                 );
             }
-            const video = this.#readVideo(videoBlock);
+            const video = await this.#readVideo(videoBlock);
             if (!billed) {
                 throw geminiUpstreamFailure(
                     'Gemini video generation returned unusable usage',
@@ -722,7 +722,7 @@ export class GeminiVideoProvider extends VideoProvider {
     }
 
     /** The base64 MP4 held by the video block. */
-    #readVideo(video: InteractionContent | undefined): string {
+    async #readVideo(video: InteractionContent | undefined): Promise<string> {
         if (!video) {
             throw new HttpError(400, 'Gemini did not return a video', {
                 legacyCode: 'upstream_bad_request',
@@ -731,7 +731,7 @@ export class GeminiVideoProvider extends VideoProvider {
         }
         const bytes =
             video.mime_type === 'video/mp4' && typeof video.data === 'string'
-                ? decodeStrictBase64(video.data)
+                ? await decodeStrictBase64InSlices(video.data)
                 : null;
         if (!bytes || sniffVideoMime(bytes) !== 'video/mp4') {
             throw geminiUpstreamFailure(

src/backend/drivers/apps/AppDriver.js

@@ -330,14 +330,11 @@ export class AppDriver extends PuterDriver {
     }
 
     /**
-     * Paged app listing. Unlike `/query/app`, which answers only for apps named
-     * by the caller and so must not confirm ones it may not see, this is the
-     * listing surface a launcher reads: `protected` apps are filtered by the
-     * permission check below, and everything else is listable by design. Narrow
-     * it with `predicate: ['user-can-edit']` to the caller's own.
+     * Paged listing of the caller's own apps. Apps owned by anyone else are
+     * reachable only by name or uid, through `read` and `/query/app`.
      */
     async select(args = {}) {
-        const { predicate, params = {} } = args;
+        const { params = {} } = args;
         const actor = requireContextActor();
         this.#requireUserOrAppActor(actor);
 
@@ -353,16 +350,12 @@ export class AppDriver extends PuterDriver {
         const includeTotal = args.includeTotal === true;
         const paginated = hasCursor || offset !== undefined || includeTotal;
 
-        const filters = {};
-        // predicate: ['user-can-edit'] → scope to owner
-        const ownerScoped =
-            Array.isArray(predicate) && predicate[0] === 'user-can-edit';
-        if (ownerScoped) {
-            filters.ownerUserId = actor.user.id;
-        }
+        // Always the caller's own. `predicate: ['user-can-edit']` asks for the
+        // same thing and is accepted for older clients.
+        const ownerUserId = actor.user.id;
 
         let apps = await this.appStore.list({
-            ...filters,
+            ownerUserId,
             limit: paginated ? limit + 1 : limit,
             offset,
             afterId: payload?.id !== undefined ? Number(payload.id) : undefined,
@@ -371,23 +364,18 @@ export class AppDriver extends PuterDriver {
         let cursor;
         if (paginated && apps.length > limit) {
             apps = apps.slice(0, limit);
-            // The cursor tracks the last fetched row, not the last visible
-            // one, so rows hidden by the permission filter below aren't
-            // re-scanned on the next page.
             cursor = sealCursor(
                 { id: Number(apps[apps.length - 1].id) },
                 this.config.jwt_secret_v2,
             );
         }
 
-        // Protected apps are filtered by one batched grant check; stats and
-        // the shared views are each one batched lookup across the page.
-        const visible = await this.services.app.filterReadable(apps, actor);
+        // Stats and the shared views are each one batched lookup per page.
         const [statsByUid, views] = await Promise.all([
-            this.appStore.getAppsStats(visible.map((a) => a.uid)),
-            this.services.app.views(visible, actor),
+            this.appStore.getAppsStats(apps.map((a) => a.uid)),
+            this.services.app.views(apps, actor),
         ]);
-        const items = visible.map((app, i) =>
+        const items = apps.map((app, i) =>
             toAppReadView(app, views[i], {
                 viewer: actor.user,
                 stats: statsByUid.get(app.uid),
@@ -396,14 +384,9 @@ export class AppDriver extends PuterDriver {
         );
         if (!paginated) return items;
 
-        let total;
-        if (includeTotal) {
-            total = ownerScoped
-                ? await this.appStore.count({ ownerUserId: actor.user.id })
-                : await this.appStore.count({
-                      visibleToUserId: actor.user?.id ?? null,
-                  });
-        }
+        const total = includeTotal
+            ? await this.appStore.count({ ownerUserId })
+            : undefined;
 
         return {
             items,

src/backend/drivers/apps/AppDriver.test.ts

@@ -509,7 +509,7 @@ describe('AppDriver.read', () => {
 // ── select ──────────────────────────────────────────────────────────
 
 describe('AppDriver.select', () => {
-    it('returns visible apps including those owned by other users', async () => {
+    it("returns only the caller's apps when no predicate is sent", async () => {
         const a = await makeUser();
         const b = await makeUser();
         const aName = uniqueName('a');
@@ -538,7 +538,7 @@ describe('AppDriver.select', () => {
         )) as Array<Record<string, unknown>>;
         const names = result.map((r) => r.name);
         expect(names).toContain(aName);
-        expect(names).toContain(bName);
+        expect(names).not.toContain(bName);
     });
 
     it('predicate `user-can-edit` filters to actor-owned apps only', async () => {
@@ -1670,11 +1670,10 @@ describe('AppDriver.select pagination', () => {
         expect(page.total).toBe(names.length);
     });
 
-    it("hides other users' protected apps from paginated catalog listings", async () => {
+    it("pages through only the caller's apps, protected ones included", async () => {
         const a = await makeApps(3);
-        const [visible1, hidden, visible2] = a.names;
         const created = await withActor(a.actor, () =>
-            driver.read({ id: { name: hidden } }),
+            driver.read({ id: { name: a.names[1] } }),
         );
         const row = await server.stores.app.getByUid(
             (created as Record<string, unknown>).uid as string,
@@ -1685,20 +1684,29 @@ describe('AppDriver.select pagination', () => {
         );
         await server.stores.app.invalidateByUid(row!.uid as string);
 
-        const b = await makeUser();
-        const seen: string[] = [];
-        let cursor: string | null | undefined = null;
-        do {
-            const page = (await withActor(b.actor, () =>
-                driver.select({ limit: 50, cursor }),
-            )) as { items: Array<{ name: string }>; cursor?: string };
-            seen.push(...page.items.map((r) => r.name));
-            cursor = page.cursor;
-        } while (cursor);
+        const listAll = async (actor: Actor) => {
+            const seen: string[] = [];
+            let total: number | undefined;
+            let cursor: string | null | undefined = null;
+            do {
+                const page = (await withActor(actor, () =>
+                    driver.select({ limit: 2, cursor, includeTotal: true }),
+                )) as {
+                    items: Array<{ name: string }>;
+                    cursor?: string;
+                    total?: number;
+                };
+                seen.push(...page.items.map((r) => r.name));
+                total ??= page.total;
+                cursor = page.cursor;
+            } while (cursor);
+            return { seen, total };
+        };
 
-        expect(seen).toContain(visible1);
-        expect(seen).toContain(visible2);
-        expect(seen).not.toContain(hidden);
+        expect(await listAll(a.actor)).toEqual({ seen: a.names, total: 3 });
+
+        const b = await makeUser();
+        expect(await listAll(b.actor)).toEqual({ seen: [], total: 0 });
     });
 });
 
@@ -2517,22 +2525,6 @@ describe('AppDriver.select query-count regression', () => {
 // -- select / read parity --
 
 describe('AppDriver.select / read parity', () => {
-    const findViaBroadSelect = async (
-        actor: Actor,
-        uid: string,
-    ): Promise<Record<string, unknown> | undefined> => {
-        let cursor: string | null | undefined = null;
-        do {
-            const page = (await withActor(actor, () =>
-                driver.select({ limit: 50, cursor }),
-            )) as { items: Array<Record<string, unknown>>; cursor?: string };
-            const found = page.items.find((r) => r.uid === uid);
-            if (found) return found;
-            cursor = page.cursor;
-        } while (cursor);
-        return undefined;
-    };
-
     it('every select item (minus stats) deep-equals the corresponding read', async () => {
         const { actor } = await makeUser();
         const names: string[] = [];
@@ -2640,7 +2632,7 @@ describe('AppDriver.select / read parity', () => {
         ).toBe(false);
     });
 
-    it('dangling hosted app via select: owner keeps index_url, others do not', async () => {
+    it('dangling hosted app: owner keeps index_url in select, others do not in read', async () => {
         const owner = await makeUser();
         const other = await makeUser();
         const sub = uniqueName('qcdangsel');
@@ -2672,13 +2664,12 @@ describe('AppDriver.select / read parity', () => {
                 ?.reason,
         ).toBe('hosted_backing_unavailable');
 
-        const otherItem = await findViaBroadSelect(
-            other.actor,
-            created.uid as string,
-        );
-        expect(otherItem?.index_url).toBeUndefined();
+        const otherItem = (await withActor(other.actor, () =>
+            driver.read({ uid: created.uid }),
+        )) as Record<string, unknown>;
+        expect(otherItem.index_url).toBeUndefined();
         expect(
-            (otherItem?.privateAccess as { hasAccess?: boolean } | undefined)
+            (otherItem.privateAccess as { hasAccess?: boolean } | undefined)
                 ?.hasAccess,
         ).toBe(false);
     });

src/backend/drivers/util/geminiRest.ts

@@ -18,6 +18,7 @@
  */
 
 import { HttpError } from '../../core/http/HttpError.js';
+import { parseJsonInSlices } from '../../util/slicedJson.js';
 import { readBoundedBody } from './fileInput.js';
 import { upstreamFetch } from './upstreamErrors.js';
 
@@ -39,8 +40,8 @@ export const geminiUpstreamFailure = (message: string): HttpError =>
 /**
  * POSTs `body` to a Gemini REST `path` and parses the JSON object it answers
  * with. `timeoutMs` covers the body read, and a reply over `maxBytes` fails as
- * an upstream error unless `tooLarge` supplies another. `T` is the documented reply shape; callers check every
- * field before use.
+ * an upstream error unless `tooLarge` supplies another. `T` is the documented
+ * reply shape; callers check every field before use.
  */
 export async function postGeminiJson<T extends object>(
     apiKey: string,
@@ -87,7 +88,8 @@ export async function postGeminiJson<T extends object>(
     const raw = await readBoundedBody(response, maxBytes, tooLarge);
     let parsed: unknown = null;
     try {
-        parsed = JSON.parse(raw.toString('utf8'));
+        // Media replies run to tens of megabytes of base64.
+        parsed = await parseJsonInSlices(raw);
     } catch {
         // Handled with the non-object case below.
     }

src/backend/server.test.ts

@@ -368,12 +368,117 @@ describe('PuterServer host header validation', () => {
         expect(api.status).toBe(404);
     });
 
+    it('lets a deployment subdomain frame the feedback embed and names it', async () => {
+        const parent = `http://spreadsheet.puter.localhost:${port}`;
+        const res = await request('/embed/feedback', {
+            host: `puter.localhost:${port}`,
+            referer: `${parent}/sheet/1`,
+        });
+        expect(res.status).toBe(200);
+        expect(res.headers['x-frame-options']).toBeUndefined();
+        expect(res.headers['content-security-policy']).toBe(
+            `frame-ancestors 'self' ${parent}`,
+        );
+        expect(res.headers['cache-control']).toBe('no-store');
+        expect(res.body).toContain('/dist/feedback-embed.min.js');
+        expect(res.body).not.toContain('/dist/bundle.min.js');
+        const config = JSON.parse(
+            res.body.match(/id="feedback-embed-config">(.*?)<\/script>/)[1],
+        );
+        expect(config).toEqual({
+            apiOrigin: `http://api.puter.localhost:${port}`,
+            hostOrigin: parent,
+        });
+
+        const hosted = await request('/embed/feedback', {
+            host: `puter.localhost:${port}`,
+            referer: `http://someone.site.puter.localhost:${port}/`,
+        });
+        expect(hosted.headers['content-security-policy']).toBe("frame-ancestors 'self'");
+        expect(
+            JSON.parse(
+                hosted.body.match(/id="feedback-embed-config">(.*?)<\/script>/)[1],
+            ).hostOrigin,
+        ).toBeNull();
+
+        const redirected = await request('/embed/feedback?puter.api_origin=https://other.test', {
+            host: `puter.localhost:${port}`,
+        });
+        expect(redirected.status).toBe(302);
+        expect(redirected.headers.location).toBe('/embed/feedback');
+    });
+
+    it.each([
+        'sign-in',
+        'show-open-file-picker',
+        'show-save-file-picker',
+        'show-directory-picker',
+        'upgrade',
+        'top-up',
+        'request-email-confirmation',
+        'attach-custom-domain',
+    ])('lets a deployment subdomain frame the %s popup and names it to the GUI', async (action) => {
+        const parent = `http://spreadsheet.puter.localhost:${port}`;
+        const guiParams = (body: string) =>
+            JSON.parse(body.match(/gui\((\{.*\})\);/)[1]);
+        const res = await request(`/embed/${action}`, {
+            host: `puter.localhost:${port}`,
+            referer: `${parent}/sheet/1`,
+        });
+        expect(res.status).toBe(200);
+        expect(res.headers['x-frame-options']).toBeUndefined();
+        expect(res.headers['content-security-policy']).toBe(
+            `frame-ancestors 'self' ${parent}`,
+        );
+        expect(res.headers['cache-control']).toBe('no-store');
+        expect(guiParams(res.body).embed_host_origin).toBe(parent);
+
+        const hosted = await request(`/embed/${action}`, {
+            host: `puter.localhost:${port}`,
+            referer: `http://someone.site.puter.localhost:${port}/`,
+        });
+        expect(hosted.headers['content-security-policy']).toBe("frame-ancestors 'self'");
+        expect(guiParams(hosted.body).embed_host_origin).toBeNull();
+
+        const redirected = await request(`/embed/${action}?auth_token=x&embedded_in_popup=true`, {
+            host: `puter.localhost:${port}`,
+            referer: `${parent}/sheet/1`,
+        });
+        expect(redirected.status).toBe(302);
+        expect(redirected.headers.location).toBe(`/embed/${action}`);
+        expect(redirected.headers['x-frame-options']).toBe('SAMEORIGIN');
+    });
+
+    it.each(['/embed/SIGN-IN', '/embed/Show-Open-File-Picker', '/embed/sign-in/', '/embed/Upgrade', '/embed/upgrade/', '/embed/Top-Up', '/embed/top-up/', '/embed/Request-Email-Confirmation'])(
+        'lets a first-party page frame only the exact spelling of %s',
+        async (spelling) => {
+            const res = await request(spelling, {
+                host: `puter.localhost:${port}`,
+                referer: `http://spreadsheet.puter.localhost:${port}/`,
+            });
+            expect(res.status).toBe(302);
+            expect(res.headers.location).toBe(spelling.toLowerCase().replace(/\/$/, ''));
+            expect(res.headers['x-frame-options']).toBe('SAMEORIGIN');
+            expect(res.headers['content-security-policy']).toBeUndefined();
+        },
+    );
+
+    it('names no framing page to the popups it serves in a window', async () => {
+        const res = await request('/action/sign-in?embedded_in_popup=true', {
+            host: `puter.localhost:${port}`,
+            referer: `http://spreadsheet.puter.localhost:${port}/`,
+        });
+        expect(res.headers['x-frame-options']).toBe('SAMEORIGIN');
+        expect(res.headers['content-security-policy']).toBeUndefined();
+        expect(res.body).not.toContain('embed_host_origin');
+    });
+
     it.each([
         ['a user-hosted site', 'http://someone.site.puter.localhost:PORT/'],
         ['a private app host', 'http://someone.app.puter.localhost:PORT/'],
         ['the API host', 'http://api.puter.localhost:PORT/'],
         ['the WebDAV host', 'http://dav.puter.localhost:PORT/'],
-        ['a native app host', 'http://editor.puter.localhost:PORT/'],
+        ['a native app host', 'http://about.puter.localhost:PORT/'],
         ['a nested subdomain', 'http://a.spreadsheet.puter.localhost:PORT/'],
         ['another site', 'http://evil.test:PORT/'],
         ['another port', 'http://spreadsheet.puter.localhost:1/'],
@@ -689,6 +794,26 @@ describe('PuterServer request parsing', () => {
         }
     });
 
+    it('parses a chat call above 50 MiB whose body opens with its token', async () => {
+        const fields = `"interface":"puter-chat-completion","method":"no-such-method","pad":"${'x'.repeat(51 * MiB)}"`;
+        const asText = { 'content-type': 'text/plain;actually=json' };
+        const first = await postJson(
+            '/drivers/call',
+            `{"auth_token":"${token}",${fields}}`,
+            asText,
+        );
+        expect(first.status).toBe(404);
+        expect(first.body).toContain('no-such-method');
+
+        // A token after other fields can't be read ahead of the body.
+        const last = await postJson(
+            '/drivers/call',
+            `{${fields},"auth_token":"${token}"}`,
+            asText,
+        );
+        expect(last.status).toBe(413);
+    });
+
     it('keeps the default JSON limit for other drivers', async () => {
         const payload = JSON.stringify({
             interface: 'puter-kvstore',

src/backend/server.ts

@@ -36,6 +36,7 @@ import { createAuthProbe } from './core/http/middleware/authProbe';
 import { createRequestContextMiddleware } from './core/http/middleware/requestContext';
 import { createFingerprintMiddleware } from './core/http/middleware/fingerprint';
 import { createErrorHandler } from './core/http/middleware/errorHandler';
+import { createJsonBodyParser } from './core/http/middleware/jsonBody';
 import {
     createLargeJsonBodyParser,
     DEFAULT_JSON_LIMIT_BYTES,
@@ -144,6 +145,8 @@ const LARGE_JSON_ROUTES = [
 const LARGE_JSON_LIMIT_BYTES = 128 * 1024 * 1024;
 /** Two max-size bodies per process. */
 const LARGE_JSON_BUDGET_BYTES = 2 * LARGE_JSON_LIMIT_BYTES;
+/** These parse in one go, so they stay small enough for that to be quick. */
+const NON_UTF8_JSON_LIMIT_BYTES = 1024 * 1024;
 
 /**
  * Through `/drivers/call`, only chat takes a large body; other drivers keep the
@@ -575,26 +578,42 @@ export class PuterServer {
                 }),
             );
         }
+        // A WebDAV body is the file being stored, whatever content type the
+        // client labels it with, so the controller reads it off the stream.
+        const exceptDav =
+            (parser: RequestHandler): RequestHandler =>
+            (req, res, next) => {
+                if (activeSubdomain(req) === 'dav') {
+                    next();
+                    return;
+                }
+                return parser(req, res, next);
+            };
         this.#app.use(
-            express.json({
-                limit: DEFAULT_JSON_LIMIT_BYTES,
-                verify: captureRawBody,
-            }),
+            exceptDav(
+                createJsonBodyParser({
+                    limit: DEFAULT_JSON_LIMIT_BYTES,
+                    verify: captureRawBody,
+                }),
+            ),
         );
+        // JSON in another charset, which the parser above leaves alone.
         this.#app.use(
-            express.json({
-                limit: DEFAULT_JSON_LIMIT_BYTES,
-                type: (req) =>
-                    req.headers['content-type'] === 'text/plain;actually=json',
-                verify: captureRawBody,
-            }),
+            exceptDav(
+                express.json({
+                    limit: NON_UTF8_JSON_LIMIT_BYTES,
+                    verify: captureRawBody,
+                }),
+            ),
         );
         // Form-encoded bodies (e.g. `/down` from the GUI's iframe-triggered
         // download form). Needs to run before the auth probe so
         // `req.body.auth_token` is populated for urlencoded POSTs the same
         // way it is for JSON POSTs. Small cap — this parser is only here to
         // cover the auth_token / anti_csrf field shape, not file uploads.
-        this.#app.use(express.urlencoded({ extended: true, limit: '100kb' }));
+        this.#app.use(
+            exceptDav(express.urlencoded({ extended: true, limit: '100kb' })),
+        );
 
         // -- Auth probe ----------------------------------------------
         if (authProbe) {

src/backend/services/apps/RecommendedAppsService.ts

@@ -31,24 +31,13 @@ export const RECOMMENDED_APP_NAMES = [
     'spreadsheet',
     'presentation',
     'whiteboard',
+    'notes',
     'pdf-editor',
     'contacts',
     'calendar',
     'meetings',
     'teamchat',
     'email',
-    'dev-center',
-    'cap-table',
-    'invoices',
-    'crm',
-    'signatures',
-    'projects',
-    'point-of-sale',
-    'inventory',
-    'hiring',
-    'time-tracking',
-    'shift-planner',
-    'equipment',
     'chess',
     'checkers',
     'backgammon',

src/backend/services/auth/TokenService.test.ts

@@ -20,6 +20,7 @@
 import { createHmac } from 'node:crypto';
 import jwt from 'jsonwebtoken';
 import { describe, expect, it } from 'vitest';
+import { isTokenForAnotherUser } from '../../../gui/src/helpers/tokenUser.js';
 import { TokenService } from './TokenService.js';
 
 const V2_SECRET = 'test-v2-secret';
@@ -384,3 +385,21 @@ describe('TokenService payload key handling', () => {
         },
     );
 });
+
+describe('session token read by the desktop', () => {
+    // The desktop reads the user out of a token another tab stored, to tell
+    // an account switch from a renewed token for the same account.
+    it('names its user in the form the desktop decodes', () => {
+        const svc = createTokenService();
+        const userUid = '5d4adce0-a381-4982-9c02-6e2540026238';
+        const otherUid = 'a1b2c3d4-0000-4000-8000-00000000beef';
+        const token = svc.sign('auth', {
+            type: 'gui',
+            version: '2',
+            user_uid: userUid,
+        });
+
+        expect(isTokenForAnotherUser(token, { uuid: userUid })).toBe(false);
+        expect(isTokenForAnotherUser(token, { uuid: otherUid })).toBe(true);
+    });
+});

src/backend/services/fs/FSService.test.ts

@@ -647,6 +647,22 @@ describe('FSService overwrite and dedupe resolution', () => {
         expect(await readBack(second)).toBe('second');
     });
 
+    it('leaves one readable file when two overwrites create the same path at once', async () => {
+        const path = `${user.home}/Documents/raced.txt`;
+        const written = await Promise.all([
+            writeFile(user, path, 'short', { overwrite: true }),
+            writeFile(user, path, 'a much longer body', { overwrite: true }),
+        ]);
+
+        // Both started before the row existed, so each uploaded under its own
+        // key; the row has to end up describing bytes that are really there.
+        expect(written[0].uuid).toBe(written[1].uuid);
+        const stored = (await entryAt(user, '/Documents/raced.txt'))!;
+        const body = await readBack(stored);
+        expect(['short', 'a much longer body']).toContain(body);
+        expect(stored.size).toBe(Buffer.byteLength(body));
+    });
+
     it('dedupes into an unused " (n)" name, skipping names already taken', async () => {
         await writeFile(user, `${user.home}/Documents/d.txt`, 'a');
         await writeFile(user, `${user.home}/Documents/d (1).txt`, 'a');

src/backend/services/fs/FSService.ts

@@ -2073,10 +2073,7 @@ export class FSService extends PuterService {
                 );
             });
 
-            const fsEntries = await this.stores.fsEntry.batchCreateEntries(
-                createInputs,
-                true,
-            );
+            const fsEntries = await this.#createUploadedEntries(createInputs);
             return preparedBatch.items.map((item, index) => {
                 const fsEntry = fsEntries[index];
                 if (!fsEntry) {
@@ -3001,30 +2998,97 @@ export class FSService extends PuterService {
                 legacyCode: 'conflict',
             });
         }
+        await this.#moveObjectOntoKey(
+            session.objectKey,
+            liveKey,
+            { bucket, region, size },
+            session.contentType,
+        );
+    }
+
+    async #moveObjectOntoKey(
+        fromKey: string,
+        toKey: string,
+        at: { bucket: string; region: string; size: number },
+        contentType?: string | null,
+    ): Promise<void> {
         await this.stores.s3Object.copyObjectOfAnySize(
             {
-                sourceBucket: bucket,
-                sourceKey: session.objectKey,
-                destinationBucket: bucket,
-                destinationKey: liveKey,
-                contentType: session.contentType,
+                sourceBucket: at.bucket,
+                sourceKey: fromKey,
+                destinationBucket: at.bucket,
+                destinationKey: toKey,
+                ...(contentType ? { contentType } : {}),
                 metadataDirective: 'REPLACE',
             },
-            region,
-            size,
+            at.region,
+            at.size,
         );
         try {
             await this.stores.s3Object.deleteObject(
-                bucket,
-                session.objectKey,
-                region,
+                at.bucket,
+                fromKey,
+                at.region,
             );
         } catch (e) {
-            // The live object is already right; the staged one is only litter.
-            console.warn('[fs] staged overwrite left behind:', e);
+            // The destination is already right; the source is only litter.
+            console.warn('[fs] moved object left behind:', e);
         }
     }
 
+    /**
+     * Creates the rows for objects already uploaded. A row another writer
+     * created first keeps its uuid, so those bytes move onto that key.
+     */
+    async #createUploadedEntries(
+        inputs: FSEntryCreateInput[],
+    ): Promise<FSEntry[]> {
+        const entries = await this.stores.fsEntry.batchCreateEntries(
+            inputs,
+            true,
+        );
+        const raced = inputs.flatMap((input, index) => {
+            const liveKey = entries[index]?.uuid;
+            return liveKey && liveKey !== input.uuid
+                ? [{ input, index, liveKey }]
+                : [];
+        });
+        if (raced.length === 0) return entries;
+
+        await Promise.all(
+            raced.map(({ input, liveKey }) =>
+                this.#moveObjectOntoKey(
+                    input.uuid,
+                    liveKey,
+                    {
+                        bucket: this.stores.s3Object.resolveBucket(
+                            input.bucket,
+                        ),
+                        region: this.stores.s3Object.resolveRegion(
+                            input.bucketRegion,
+                        ),
+                        size: input.size,
+                    },
+                    input.contentType,
+                ),
+            ),
+        );
+        const adopted = await this.stores.fsEntry.batchCreateEntries(
+            raced.map(({ input, liveKey }) => ({ ...input, uuid: liveKey })),
+            true,
+        );
+        raced.forEach(({ index, liveKey }, i) => {
+            const entry = adopted[i];
+            if (entry?.uuid !== liveKey) {
+                throw new HttpError(409, 'Write target changed mid-write', {
+                    legacyCode: 'conflict',
+                });
+            }
+            entries[index] = entry;
+        });
+        return entries;
+    }
+
     async completeUrlWrite(
         userId: number,
         completeWriteRequest: CompleteWriteRequest,
@@ -3581,10 +3645,10 @@ export class FSService extends PuterService {
             : uploadBody.contentHashSha256;
 
         const createInput = this.#toCreateInput(normalizedInput, objectKey);
-        const fsEntry = await this.stores.fsEntry.createEntry(
-            createInput,
-            normalizedInput.createMissingParents,
-        );
+        const [fsEntry] = await this.#createUploadedEntries([createInput]);
+        if (!fsEntry) {
+            throw new Error('Failed to create entry');
+        }
 
         this.#emitFsEvent(
             existingEntry ? 'fs.write.file' : 'fs.create.file',
@@ -3682,17 +3746,18 @@ export class FSService extends PuterService {
                 updates.filter(isValid),
             );
         let next = 0;
-        return updates.map((update): PromiseSettledResult<FSEntry> =>
-            isValid(update)
-                ? stored[next++]
-                : {
-                      status: 'rejected',
-                      reason: new HttpError(
-                          400,
-                          'Invalid file entry identifier for thumbnail update',
-                          { legacyCode: 'bad_request' },
-                      ),
-                  },
+        return updates.map(
+            (update): PromiseSettledResult<FSEntry> =>
+                isValid(update)
+                    ? stored[next++]
+                    : {
+                          status: 'rejected',
+                          reason: new HttpError(
+                              400,
+                              'Invalid file entry identifier for thumbnail update',
+                              { legacyCode: 'bad_request' },
+                          ),
+                      },
         );
     }
 

src/backend/services/homepage/PuterHomepageService.test.ts

@@ -107,6 +107,31 @@ describe('app browser embed shell', () => {
     });
 });
 
+describe('feedback embed shell', () => {
+    it('carries the framing page and API origin, and loads only its own script', () => {
+        const html = makeService({
+            api_base_url: 'https://api.deployment.test',
+            gui_bundle: '/static/bundle.min.js?v=42',
+        }).renderFeedbackEmbed({ hostOrigin: 'https://notes.deployment.test' });
+        const configJson = html.match(/id="feedback-embed-config">(.*?)<\/script>/)![1];
+        expect(JSON.parse(configJson)).toEqual({
+            apiOrigin: 'https://api.deployment.test',
+            hostOrigin: 'https://notes.deployment.test',
+        });
+        expect(html).toContain('<script defer src="/static/feedback-embed.min.js?v=42"></script>');
+        expect(html).toContain('<link rel="stylesheet" href="/css/feedback-embed.css">');
+        expect(html).toContain('<main id="feedback-embed"></main>');
+        expect(html).not.toContain('bundle.min.js');
+        expect(html).not.toContain('apps-embed');
+    });
+
+    it('carries no framing page when none is allowed', () => {
+        const html = makeService({}).renderFeedbackEmbed({ hostOrigin: null });
+        const configJson = html.match(/id="feedback-embed-config">(.*?)<\/script>/)![1];
+        expect(JSON.parse(configJson).hostOrigin).toBeNull();
+    });
+});
+
 /** Capture what the service sends, and return the rendered HTML. */
 const render = async (
     service: PuterHomepageService,

src/backend/services/homepage/PuterHomepageService.ts

@@ -48,6 +48,15 @@ export interface LaunchOptions {
     on_initialized?: Array<Record<string, unknown>>;
 }
 
+/**
+ * A GUI popup a first-party page frames at `/embed/<action>`. `hostOrigin` is
+ * that page, read from the referrer and pinned with frame-ancestors; null when
+ * no allowed page frames it.
+ */
+export interface PopupFrame {
+    hostOrigin: string | null;
+}
+
 interface PuterGuiAddonsEvent {
     req: Request;
     path: string;
@@ -171,15 +180,50 @@ export class PuterHomepageService extends PuterService {
     renderAppBrowserEmbed({
         guestApps = [],
     }: { guestApps?: Array<Record<string, unknown>> } = {}): string {
-        const configJson = JSON.stringify({
-            apiOrigin: this.config.api_base_url,
-            sdkUrl:
-                this.config.env === 'dev'
-                    ? '/sdk/puter.dev.js'
-                    : (this.config.gui_puterjs_bundle ??
-                      'https://js.puter.com/v2/'),
-            guestApps,
-        }).replace(/</g, '\\u003c');
+        return this.#renderEmbed({
+            rootId: 'app-browser',
+            rootClass: 'dashboard',
+            stylesheet: 'apps-embed.css',
+            script: 'apps-embed.min.js',
+            config: {
+                apiOrigin: this.config.api_base_url,
+                sdkUrl:
+                    this.config.env === 'dev'
+                        ? '/sdk/puter.dev.js'
+                        : (this.config.gui_puterjs_bundle ??
+                          'https://js.puter.com/v2/'),
+                guestApps,
+            },
+        });
+    }
+
+    /**
+     * `hostOrigin` is the first-party page framing the dialog, which the
+     * feedback is sent for; null when no allowed page frames it.
+     */
+    renderFeedbackEmbed({ hostOrigin }: { hostOrigin: string | null }): string {
+        return this.#renderEmbed({
+            rootId: 'feedback-embed',
+            stylesheet: 'feedback-embed.css',
+            script: 'feedback-embed.min.js',
+            config: { apiOrigin: this.config.api_base_url, hostOrigin },
+        });
+    }
+
+    #renderEmbed({
+        rootId,
+        rootClass,
+        stylesheet,
+        script,
+        config,
+    }: {
+        rootId: string;
+        rootClass?: string;
+        stylesheet: string;
+        script: string;
+        config: Record<string, unknown>;
+    }): string {
+        const configJson = JSON.stringify(config).replace(/</g, '\\u003c');
 
         return `<!DOCTYPE html>
 <html lang="en">
@@ -190,27 +234,27 @@ export class PuterHomepageService extends PuterService {
     <title>${encode(String(this.config.gui_params?.title ?? 'Puter'))}</title>
     <link rel="stylesheet" href="/css/normalize.css">
     <link rel="stylesheet" href="/css/dashboard.css">
-    <link rel="stylesheet" href="/css/apps-embed.css">
+    <link rel="stylesheet" href="/css/${stylesheet}">
 </head>
 <body>
-    <main id="app-browser" class="dashboard"></main>
-    <script type="application/json" id="app-browser-config">${configJson}</script>
-    <script defer src="${encode(this.#appBrowserEmbedScript())}"></script>
+    <main id="${rootId}"${rootClass ? ` class="${rootClass}"` : ''}></main>
+    <script type="application/json" id="${rootId}-config">${configJson}</script>
+    <script defer src="${encode(this.#embedScript(script))}"></script>
 </body>
 </html>`;
     }
 
     /**
-     * The embed's script is built beside the GUI bundle, so it is served from
+     * Embed scripts are built beside the GUI bundle, so they are served from
      * wherever `gui_bundle` is (a CDN in production), with the same query.
      */
-    #appBrowserEmbedScript(): string {
+    #embedScript(name: string): string {
         const base = 'http://relative.invalid';
         const bundle = new URL(
             this.config.gui_bundle ?? '/dist/bundle.min.js',
             base,
         );
-        const script = new URL('apps-embed.min.js', bundle);
+        const script = new URL(name, bundle);
         script.search = bundle.search;
         return script.origin === base
             ? `${script.pathname}${script.search}`
@@ -268,6 +312,7 @@ export class PuterHomepageService extends PuterService {
         ctx: { req: Request; res: Response; actor?: Actor | null },
         meta: PageMeta,
         launchOptions: LaunchOptions = {},
+        popupFrame?: PopupFrame,
     ): Promise<void> {
         const { req, res, actor } = ctx;
 
@@ -292,6 +337,7 @@ export class PuterHomepageService extends PuterService {
             { req, actor: actor ?? null },
             meta,
             launchOptions,
+            popupFrame,
         );
         res.send(html);
     }
@@ -302,6 +348,7 @@ export class PuterHomepageService extends PuterService {
         ctx: { req: Request; actor: Actor | null },
         meta: PageMeta,
         launchOptions: LaunchOptions,
+        popupFrame?: PopupFrame,
     ): Promise<string> {
         const { req, actor } = ctx;
         const env = this.config.env ?? 'prod';
@@ -347,6 +394,8 @@ export class PuterHomepageService extends PuterService {
             eventsNotifications: notificationsFoldInEnabled(this.config),
             ...meta,
             launch_options: launchOptions,
+            // After `meta`, so nothing page-specific can name the opener.
+            ...(popupFrame ? { embed_host_origin: popupFrame.hostOrigin } : {}),
         };
 
         const event: PuterGuiAddonsEvent = {

src/backend/services/profile/ProfileService.test.ts

@@ -18,6 +18,7 @@
  */
 
 import { Writable } from 'node:stream';
+import v8 from 'node:v8';
 import type { Request, Response } from 'express';
 import { afterAll, beforeAll, describe, expect, it, vi } from 'vitest';
 import { v4 as uuidv4 } from 'uuid';
@@ -229,6 +230,37 @@ describe('ProfileService', () => {
         ).toBeNull();
     });
 
+    // V8 stops optimizing regexps once a process holds enough compiled code,
+    // as long-running servers do, and an unoptimized one overflows on this.
+    it('answers a picture of several megabytes with 413', () => {
+        const picture = `data:image/png;base64,${'A'.repeat(6 * 1024 * 1024)}`;
+        v8.setFlagsFromString('--no-regexp-optimization');
+        try {
+            expect(() => service.validatePatch({ picture })).toThrow(
+                expect.objectContaining({
+                    statusCode: 413,
+                    code: 'profile_picture_too_large',
+                }),
+            );
+        } finally {
+            v8.setFlagsFromString('--regexp-optimization');
+        }
+    });
+
+    it('accepts a picture of the largest size, with or without padding', () => {
+        const prefix = 'data:image/png;base64,';
+        const full = `${prefix}${'A'.repeat(PROFILE_PICTURE_MAX_BYTES - prefix.length)}`;
+        expect(service.validatePatch({ picture: full })).toEqual({
+            picture: full,
+        });
+        expect(service.validatePatch({ picture: `${prefix}QQ` })).toEqual({
+            picture: `${prefix}QQ`,
+        });
+        expect(() => service.validatePatch({ picture: `${full}A` })).toThrow(
+            expect.objectContaining({ code: 'profile_picture_too_large' }),
+        );
+    });
+
     it('keeps what validates from a legacy profile and names the rest', () => {
         expect(
             service.normalizeLegacyProfile({

src/backend/services/profile/ProfileService.ts

@@ -83,6 +83,14 @@ const validateField = (field: ProfileField, value: unknown): string | null => {
         );
     }
     if (field === 'picture') {
+        // Sized before the pattern runs, which fails outright on megabytes of
+        // text. Only ASCII passes the pattern, so the length is the byte count.
+        if (value.length > PROFILE_PICTURE_MAX_BYTES) {
+            throw new HttpError(413, 'picture is too large', {
+                code: 'profile_picture_too_large',
+                fields: { field, maxBytes: PROFILE_PICTURE_MAX_BYTES },
+            });
+        }
         if (!PICTURE_DATA_URL.test(value)) {
             throw invalid(
                 field,
@@ -90,12 +98,6 @@ const validateField = (field: ProfileField, value: unknown): string | null => {
                 'profile_picture_invalid',
             );
         }
-        if (Buffer.byteLength(value) > PROFILE_PICTURE_MAX_BYTES) {
-            throw new HttpError(413, 'picture is too large', {
-                code: 'profile_picture_too_large',
-                fields: { field, maxBytes: PROFILE_PICTURE_MAX_BYTES },
-            });
-        }
         return value;
     }
     const maxLength =

src/backend/services/realtime/README.md

@@ -1,11 +1,11 @@
 # OpenAI Realtime relay
 
 The relay serves `/puterai/openai/v1/realtime` on the API host. The SDK exposes
-it through [`puter.ai.convo()`](../../../docs/src/AI/convo.md).
+it through [`puter.ai.realtime()`](../../../docs/src/AI/realtime.md).
 
 **Beta:** This API is incomplete. Its methods, options, events, and behavior may change abruptly, including breaking changes without notice.
 
-Direct clients are documented in the [Realtime WebSocket API reference](../../../docs/src/AI/realtime.md).
+Direct clients are documented in the [Realtime WebSocket API reference](../../../docs/src/AI/realtime-websocket.md).
 
 ## Transport
 
@@ -27,7 +27,7 @@ The socket opens before authentication completes. Authentication and model
 failures arrive as OpenAI-shaped error events followed by a close frame.
 
 ```js
-const conversation = await puter.ai.convo({ output: ['text'] });
+const conversation = await puter.ai.realtime({ output: ['text'] });
 conversation.on('error', (event) => console.error(event));
 conversation.on('text', ({ delta }) => console.log(delta));
 conversation.on('turn.done', () => conversation.close());

src/backend/stores/app/AppStore.js

@@ -26,6 +26,7 @@ import {
     PuterStore,
 } from '../types';
 import { HttpError } from '../../core/http/HttpError.js';
+import { runWithConcurrencyLimit } from '../../util/concurrency.js';
 import { isUniqueViolation } from '../../util/dbError.js';
 import {
     validateUrl,
@@ -89,6 +90,8 @@ const STATS_CACHE_TTL_SECONDS = 30 * 60;
 // `app-` + a UUID. ClickHouse's `app_opens.app_uid` is FixedString(40) and
 // fails the whole query on a longer value.
 const APP_UID_MAX_BYTES = 40;
+// ClickHouse caps the size of one query parameter, and the uid list is one.
+const CLICKHOUSE_UIDS_PER_QUERY = 1000;
 
 // Period helpers for detailed/grouped stats. Ported from v1
 // AppInformationService — queries go straight to ClickHouse/MySQL on demand
@@ -686,12 +689,7 @@ export class AppStore extends PuterStore {
         return apps;
     }
 
-    /**
-     * Count apps matching the same equality filters as `list`.
-     * `visibleToUserId` counts the public catalog for one caller —
-     * non-protected apps plus their own; apps visible only via explicit
-     * permission grants are not included.
-     */
+    /** Count apps matching the same equality filters as `list`. */
     async count(filters = {}) {
         const where = [];
         const params = [];
@@ -704,12 +702,6 @@ export class AppStore extends PuterStore {
             where.push('`app_owner` = ?');
             params.push(filters.appOwner);
         }
-        if (filters.visibleToUserId !== undefined) {
-            where.push(
-                '(`protected` IS NULL OR `protected` = false OR `owner_user_id` = ?)',
-            );
-            params.push(filters.visibleToUserId);
-        }
 
         const whereClause = where.length ? `WHERE ${where.join(' AND ')}` : '';
         const rows = await this.clients.db.read(
@@ -1738,20 +1730,30 @@ export class AppStore extends PuterStore {
 
         const clickhouse = this.clients.clickhouse;
         if (clickhouse) {
-            const res = await clickhouse.query({
-                query: `
-                    SELECT app_uid,
-                           uniqExact((user_id, intDiv(ts, ${APP_OPEN_COUNT_PERIOD_SECONDS}))) AS open_count,
-                           count(DISTINCT user_id) AS user_count
-                    FROM app_opens
-                    WHERE app_uid IN {uids:Array(String)}
-                    GROUP BY app_uid
-                `,
-                query_params: { uids: uids.map((u) => String(u)) },
-                format: 'JSONEachRow',
-            });
-            const rows = await res.json();
-            for (const row of rows) {
+            const chunks = [];
+            for (let i = 0; i < uids.length; i += CLICKHOUSE_UIDS_PER_QUERY) {
+                chunks.push(uids.slice(i, i + CLICKHOUSE_UIDS_PER_QUERY));
+            }
+            const results = await runWithConcurrencyLimit(
+                chunks,
+                4,
+                async (chunk) => {
+                    const res = await clickhouse.query({
+                        query: `
+                            SELECT app_uid,
+                                   uniqExact((user_id, intDiv(ts, ${APP_OPEN_COUNT_PERIOD_SECONDS}))) AS open_count,
+                                   count(DISTINCT user_id) AS user_count
+                            FROM app_opens
+                            WHERE app_uid IN {uids:Array(String)}
+                            GROUP BY app_uid
+                        `,
+                        query_params: { uids: chunk.map((u) => String(u)) },
+                        format: 'JSONEachRow',
+                    });
+                    return res.json();
+                },
+            );
+            for (const row of results.flat()) {
                 out.set(row.app_uid, {
                     open_count: parseInt(row.open_count, 10) || 0,
                     user_count: parseInt(row.user_count, 10) || 0,

src/backend/stores/app/AppStore.test.js

@@ -200,6 +200,37 @@ describe('AppStore app stats (cache-on-read)', () => {
         }
     });
 
+    it('splits a long uid list across ClickHouse queries', async () => {
+        const uids = Array.from({ length: 2500 }, () => freshUid());
+        const queried = [];
+
+        server.clients.clickhouse = {
+            query: async ({ query_params }) => {
+                queried.push(query_params.uids.length);
+                return {
+                    json: async () =>
+                        query_params.uids.map((u) => ({
+                            app_uid: u,
+                            open_count: '2',
+                            user_count: '1',
+                        })),
+                };
+            },
+        };
+
+        try {
+            const stats = await appStore.getAppsStats(uids);
+            expect(queried.sort((a, b) => b - a)).toEqual([1000, 1000, 500]);
+            expect(stats.size).toBe(2500);
+            expect(stats.get(uids[2499])).toMatchObject({
+                open_count: 2,
+                user_count: 1,
+            });
+        } finally {
+            delete server.clients.clickhouse;
+        }
+    });
+
     it('keeps ids longer than an app uid out of the ClickHouse query', async () => {
         const uid = freshUid();
         const longId = `iap-${'a'.repeat(36)}-extra`;
@@ -1292,31 +1323,14 @@ describe('AppStore CRUD and cache invalidation', () => {
         ).toBe('Bumped');
     });
 
-    it('counts apps by owner, creator and per-caller visibility', async () => {
+    it('counts apps by owner and creator', async () => {
         const countOwner = 4600;
-        const viewer = 4601;
         const open = await createApp({}, { ownerUserId: countOwner });
         await createApp({}, { ownerUserId: countOwner, appOwner: open.id });
-        const hidden = await createApp({}, { ownerUserId: countOwner });
-        await db.write('UPDATE `apps` SET `protected` = 1 WHERE `id` = ?', [
-            hidden.id,
-        ]);
+        await createApp({}, { ownerUserId: countOwner });
 
         expect(await appStore.count({ ownerUserId: countOwner })).toBe(3);
         expect(await appStore.count({ appOwner: open.id })).toBe(1);
-
-        // The protected app is invisible to a stranger but visible to its
-        // owner.
-        const strangerView = await appStore.count({
-            ownerUserId: countOwner,
-            visibleToUserId: viewer,
-        });
-        expect(strangerView).toBe(2);
-        const ownerView = await appStore.count({
-            ownerUserId: countOwner,
-            visibleToUserId: countOwner,
-        });
-        expect(ownerView).toBe(3);
     });
 
     // -- filetype association batch + recent opens ----------------------

src/backend/stores/fs/FSEntryStore.test.ts

@@ -444,7 +444,7 @@ describe('FSEntryStore cache invalidation', () => {
             store.batchCreateEntries([
                 {
                     userId: user.userId,
-                    uuid: uuidv4(),
+                    uuid: file.uuid,
                     path: file.path,
                     size: 7,
                     overwrite: true,
@@ -1185,7 +1185,7 @@ describe('FSEntryStore entry creation', () => {
         const [updated] = await store.batchCreateEntries([
             {
                 userId: user.userId,
-                uuid: uuidv4(),
+                uuid: original.uuid,
                 path: `${user.home}/Documents/ow.txt`,
                 size: 42,
                 overwrite: true,
@@ -1204,6 +1204,33 @@ describe('FSEntryStore entry creation', () => {
         expect(updated?.thumbnail).toBe('data:image/png;base64,AA');
     });
 
+    it('leaves the row alone when an overwrite brings a different object key', async () => {
+        const user = await makeUser();
+        const original = await createFile(
+            user,
+            `${user.home}/Documents/otherkey.txt`,
+            { size: 1 },
+        );
+
+        // The bytes for this write sit under a key the row doesn't name, so
+        // repointing bucket or size would describe an object that isn't there.
+        const returned = await createFile(user, original.path, {
+            size: 42,
+            overwrite: true,
+            bucket: 'elsewhere',
+            bucketRegion: 'eu-central-1',
+        });
+
+        expect(returned.uuid).toBe(original.uuid);
+        await expect(
+            store.getEntryByPath(original.path),
+        ).resolves.toMatchObject({
+            uuid: original.uuid,
+            size: 1,
+            bucket: 'puter-local',
+        });
+    });
+
     it('refuses to clobber an existing entry without overwrite, or a directory with it', async () => {
         const user = await makeUser();
         await createFile(user, `${user.home}/Documents/keep.txt`);
@@ -1227,7 +1254,10 @@ describe('FSEntryStore entry creation', () => {
 
     it('propagates the underlying failure when an overwrite update fails', async () => {
         const user = await makeUser();
-        await createFile(user, `${user.home}/Documents/upfail.txt`);
+        const existing = await createFile(
+            user,
+            `${user.home}/Documents/upfail.txt`,
+        );
 
         const db = server.clients.db;
         const originalWrite = (Object.getPrototypeOf(db) as typeof db).write;
@@ -1243,6 +1273,7 @@ describe('FSEntryStore entry creation', () => {
         await expect(
             createFile(user, `${user.home}/Documents/upfail.txt`, {
                 overwrite: true,
+                uuid: existing.uuid,
             }),
         ).rejects.toThrow('update rejected');
 
@@ -3801,7 +3832,7 @@ describe('FSEntryStore lost-insert recovery', () => {
         spy.mockRestore();
     });
 
-    it('updates the race winner in place when overwrite was requested', async () => {
+    it('hands back the race winner untouched when overwrite was requested', async () => {
         const user = await makeUser();
         const uuid = uuidv4();
         const spy = loseInsertRace(uuid, () =>
@@ -3827,13 +3858,14 @@ describe('FSEntryStore lost-insert recovery', () => {
         ]);
         spy.mockRestore();
 
-        // The winner's row survived; our payload was written onto it.
+        // Our object is under `uuid`, not the winner's key, so the winner's
+        // row is returned as it stands for the caller to move the bytes onto.
         expect(entry?.uuid).not.toBe(uuid);
-        expect(entry?.size).toBe(77);
-        expect(entry?.thumbnail).toBe('thumb');
+        expect(entry?.size).not.toBe(77);
+        expect(entry?.thumbnail).not.toBe('thumb');
         await expect(
             store.getEntryByPath(`${user.home}/Documents/recovered.txt`),
-        ).resolves.toMatchObject({ size: 77, uuid: entry?.uuid });
+        ).resolves.toMatchObject({ size: entry?.size, uuid: entry?.uuid });
     });
 });
 

src/backend/stores/fs/FSEntryStore.ts

@@ -1941,6 +1941,11 @@ export class FSEntryStore extends PuterStore {
         return entry;
     }
 
+    /**
+     * An overwrite whose `uuid` is not the existing row's leaves that row
+     * untouched and returns it: the caller's object is under another key, so
+     * the row can't describe it until the caller moves the bytes.
+     */
     async batchCreateEntries(
         entries: FSEntryCreateInput[],
         createPaths = true,
@@ -2074,6 +2079,10 @@ export class FSEntryStore extends PuterStore {
                             { legacyCode: 'cannot_overwrite_a_directory' },
                         );
                     }
+                    if (existingEntry.uuid !== entry.input.uuid) {
+                        updatedResultsByIndex.set(entry.index, existingEntry);
+                        continue;
+                    }
 
                     const updatedEntry = {
                         ...existingEntry,
@@ -2398,6 +2407,10 @@ export class FSEntryStore extends PuterStore {
                         );
                         continue;
                     }
+                    if (existing.uuid !== entry.input.uuid) {
+                        updatedResultsByIndex.set(entry.index, existing);
+                        continue;
+                    }
 
                     const updatedEntry: FSEntry = {
                         ...existing,

src/backend/types.ts

@@ -1102,8 +1102,9 @@ interface IConfigOptional {
     /** Force the bundled GUI even in dev. Default: false. */
     use_bundled_gui: boolean;
     /**
-     * Override the GUI bundle JS path. Default: `/dist/bundle.min.js`. The app
-     * browser embed loads `apps-embed.min.js` from the same directory.
+     * Override the GUI bundle JS path. Default: `/dist/bundle.min.js`. The
+     * `/embed/*` pages load `apps-embed.min.js` and `feedback-embed.min.js`
+     * from the same directory.
      */
     gui_bundle: string;
     /** Override the GUI CSS path when bundled. Default: `/dist/bundle.min.css`. */

src/backend/util/mediaSniff.test.ts

@@ -18,7 +18,10 @@
  */
 
 import { describe, expect, it } from 'vitest';
-import { decodeStrictBase64 } from './mediaSniff.ts';
+import {
+    decodeStrictBase64,
+    decodeStrictBase64InSlices,
+} from './mediaSniff.ts';
 
 describe('decodeStrictBase64', () => {
     it.each([
@@ -50,3 +53,52 @@ describe('decodeStrictBase64', () => {
         }
     });
 });
+
+describe('decodeStrictBase64InSlices', () => {
+    // Several slices, with a last one that is partial and padded.
+    const bytes = Buffer.alloc(5 * 1024 * 1024 + 2);
+    for (let i = 0; i < bytes.length; i += 4099) bytes[i] = i % 251;
+    const valid = bytes.toString('base64');
+
+    it('decodes a multi-MB payload to the same bytes', async () => {
+        expect(valid.endsWith('==')).toBe(true);
+        expect((await decodeStrictBase64InSlices(valid))?.equals(bytes)).toBe(
+            true,
+        );
+        const short = Buffer.from('abcd').toString('base64');
+        expect(await decodeStrictBase64InSlices(short)).toEqual(
+            Buffer.from('abcd'),
+        );
+    });
+
+    it.each([
+        ['empty input', ''],
+        ['unpadded input', valid.slice(0, -2)],
+        [
+            'a character outside the alphabet in the first slice',
+            `*${valid.slice(1)}`,
+        ],
+        [
+            'a character outside the alphabet in the last slice',
+            `${valid.slice(0, -5)}*${valid.slice(-4)}`,
+        ],
+        [
+            'the URL-safe alphabet',
+            `${valid.slice(0, 3_000_000)}-${valid.slice(3_000_001)}`,
+        ],
+        [
+            'a non-ASCII character',
+            `${valid.slice(0, 3_000_000)}ī${valid.slice(3_000_001)}`,
+        ],
+        [
+            'padding where a slice ends',
+            `${valid.slice(0, 1024 * 1024 - 1)}=${valid.slice(1024 * 1024)}`,
+        ],
+        [
+            'padding inside a slice',
+            `${valid.slice(0, 2_000_002)}==${valid.slice(2_000_004)}`,
+        ],
+    ])('rejects %s', async (_label, value) => {
+        expect(await decodeStrictBase64InSlices(value)).toBeNull();
+    });
+});

src/backend/util/mediaSniff.ts

@@ -17,6 +17,8 @@
  * along with this program.  If not, see <https://www.gnu.org/licenses/>.
  */
 
+import { setImmediate as nextTurn } from 'node:timers/promises';
+
 // Identify uploaded media by its bytes. A caller-supplied MIME type or file
 // extension is never trusted.
 
@@ -69,6 +71,33 @@ export function decodeStrictBase64(value: string): Buffer | null {
     return bytes.length === (value.length / 4) * 3 - padding ? bytes : null;
 }
 
+/** Base64 characters decoded per turn; a multiple of four. */
+const BASE64_SLICE_CHARS = 1024 * 1024;
+
+/**
+ * `decodeStrictBase64` for payloads of many megabytes: decodes a slice per turn
+ * of the event loop, so the whole string never holds it.
+ */
+export async function decodeStrictBase64InSlices(
+    value: string,
+): Promise<Buffer | null> {
+    if (value.length <= BASE64_SLICE_CHARS) return decodeStrictBase64(value);
+    if (value.length % 4 !== 0) return null;
+    const padding = value.endsWith('==') ? 2 : value.endsWith('=') ? 1 : 0;
+    const bytes = Buffer.allocUnsafe((value.length / 4) * 3 - padding);
+    let written = 0;
+    for (let at = 0; at < value.length; at += BASE64_SLICE_CHARS) {
+        const slice = decodeStrictBase64(
+            value.slice(at, at + BASE64_SLICE_CHARS),
+        );
+        // Padding inside the string leaves a slice short, and the total with it.
+        if (!slice || written + slice.length > bytes.length) return null;
+        written += slice.copy(bytes, written);
+        await nextTurn();
+    }
+    return written === bytes.length ? bytes : null;
+}
+
 function looksLikeSvg(bytes: Buffer): boolean {
     let head = bytes.subarray(0, SVG_SNIFF_WINDOW).toString('utf8');
     if (head.charCodeAt(0) === 0xfeff) head = head.slice(1);

src/backend/util/slicedJson.test.ts

@@ -0,0 +1,374 @@
+/*
+ * Copyright (C) 2024-present Puter Technologies Inc.
+ *
+ * This file is part of Puter.
+ *
+ * Puter is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU Affero General Public License as published
+ * by the Free Software Foundation, either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU Affero General Public License for more details.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program.  If not, see <https://www.gnu.org/licenses/>.
+ */
+
+import { afterEach, describe, expect, it, vi } from 'vitest';
+import {
+    MAX_DEPTH,
+    parseJsonInSlices,
+    SLICE_BYTES,
+    STRING_PIECE_BYTES,
+} from './slicedJson.js';
+
+type Outcome = { value: unknown } | { error: string };
+
+const native = (body: Buffer): Outcome => {
+    try {
+        return { value: JSON.parse(body.toString('utf8')) };
+    } catch (e) {
+        return { error: (e as Error).name };
+    }
+};
+
+const sliced = async (body: Buffer): Promise<Outcome> => {
+    try {
+        return { value: await parseJsonInSlices(body) };
+    } catch (e) {
+        return { error: (e as Error).name };
+    }
+};
+
+/** Key order, own `__proto__` keys and lone surrogates all show in the text. */
+const shown = (outcome: Outcome): string =>
+    'error' in outcome ? outcome.error : JSON.stringify(outcome.value);
+
+const expectSameAsNative = async (body: Buffer | string): Promise<void> => {
+    const buffer = typeof body === 'string' ? Buffer.from(body) : body;
+    expect(shown(await sliced(buffer))).toBe(shown(native(buffer)));
+};
+
+/** Whitespace that pushes `text` onto the sliced path without changing it. */
+const padded = (text: string): string => `${' '.repeat(SLICE_BYTES)}${text}`;
+
+// mulberry32: the cases must be the same on every run.
+const seeded = (seed: number) => () => {
+    seed = (seed + 0x6d2b79f5) | 0;
+    let t = Math.imul(seed ^ (seed >>> 15), 1 | seed);
+    t = (t + Math.imul(t ^ (t >>> 7), 61 | t)) ^ t;
+    return ((t ^ (t >>> 14)) >>> 0) / 4294967296;
+};
+
+const STRING_PARTS = [
+    'a',
+    'word ',
+    '"',
+    '\\',
+    '\\\\"',
+    '\n',
+    '\u0001',
+    'é',
+    '日本',
+    '😀',
+    '\ud83d',
+    '\u2028',
+    '/',
+    '\\u00e9',
+];
+const KEYS = ['id', 'name', '__proto__', '7', '3', 'constructor', 'a b', ''];
+// No key holds a backslash: with one in a key seen earlier, some V8 versions
+// misread a later escaped key, in `JSON.parse` itself.
+const KEY_PARTS = STRING_PARTS.filter((part) => !part.includes('\\'));
+const SCALARS = [
+    '0',
+    '-0',
+    '1.5e300',
+    '-12.25',
+    '1e-7',
+    '9007199254740993',
+    'true',
+    'false',
+    'null',
+];
+const GAPS = ['', '', '', ' ', '\n\t', '\r\n  '];
+
+/** A random JSON document, written as text so spacing and escapes vary. */
+const randomDocument = (
+    random: () => number,
+    targetBytes: number,
+    { plainKeys = false } = {},
+): string => {
+    const pick = <T>(items: readonly T[]): T =>
+        items[Math.floor(random() * items.length)]!;
+    const gap = () => pick(GAPS);
+    const string = (maxParts: number, from = STRING_PARTS): string => {
+        let text = '';
+        const parts = Math.floor(random() * maxParts);
+        for (let i = 0; i < parts; i++) text += pick(from);
+        return JSON.stringify(text);
+    };
+    let written = 0;
+    const value = (depth: number): string => {
+        const roll = random();
+        let text: string;
+        if (depth > 0 && roll < 0.25) {
+            const count = Math.floor(random() * 12);
+            const items: string[] = [];
+            for (let i = 0; i < count; i++) items.push(value(depth - 1));
+            text = `[${gap()}${items.join(`${gap()},${gap()}`)}${gap()}]`;
+        } else if (depth > 0 && roll < 0.5) {
+            const count = Math.floor(random() * 12);
+            const members: string[] = [];
+            for (let i = 0; i < count; i++) {
+                const key =
+                    plainKeys || random() < 0.7
+                        ? JSON.stringify(pick(KEYS))
+                        : string(6, KEY_PARTS);
+                members.push(`${key}${gap()}:${gap()}${value(depth - 1)}`);
+            }
+            text = `{${gap()}${members.join(`${gap()},${gap()}`)}${gap()}}`;
+        } else if (roll < 0.75) {
+            // Now and then one long enough to be read in pieces.
+            text = string(random() < 0.002 ? 3 * SLICE_BYTES : 20);
+        } else {
+            text = pick(SCALARS);
+        }
+        written += text.length;
+        return text;
+    };
+    const items: string[] = [];
+    while (written < targetBytes) items.push(value(6));
+    return random() < 0.5
+        ? `[${items.join(',')}]`
+        : `{${items.map((item, i) => `"k${i % 50}":${item}`).join(',')}}`;
+};
+
+describe('parseJsonInSlices', () => {
+    afterEach(() => {
+        vi.restoreAllMocks();
+    });
+
+    it('parses what JSON.parse parses, the same way', async () => {
+        const random = seeded(1);
+        for (let i = 0; i < 60; i++) {
+            const size = SLICE_BYTES * (1 + Math.floor(random() * 12));
+            await expectSameAsNative(randomDocument(random, size));
+        }
+    }, 60_000);
+
+    it('parses the same when every clock read ends the turn', async () => {
+        let now = 0;
+        vi.spyOn(performance, 'now').mockImplementation(() => (now += 1000));
+        const random = seeded(3);
+        for (let i = 0; i < 30; i++) {
+            await expectSameAsNative(randomDocument(random, 6 * SLICE_BYTES));
+        }
+        // Strings long enough for a turn to end inside them.
+        await expectSameAsNative(
+            `["${'ab\\n'.repeat(SLICE_BYTES)}",{"k":"${'\\\\'.repeat(SLICE_BYTES)}"},${'1234567890'.repeat(SLICE_BYTES / 2)}]`,
+        );
+    }, 60_000);
+
+    it('rejects what JSON.parse rejects', async () => {
+        const random = seeded(2);
+        const replacements = Buffer.from('[]{},:"\\ 0a\néx-.e');
+        for (let i = 0; i < 40; i++) {
+            // Plain keys, so no edit can leave a backslash in one.
+            const valid = Buffer.from(
+                randomDocument(random, 2 * SLICE_BYTES, { plainKeys: true }),
+            );
+            for (let edit = 0; edit < 12; edit++) {
+                const body = Buffer.from(valid);
+                const at = Math.floor(random() * body.length);
+                const roll = random();
+                if (roll < 0.4) {
+                    body[at] =
+                        replacements[
+                            Math.floor(random() * replacements.length)
+                        ]!;
+                    await expectSameAsNative(body);
+                } else if (roll < 0.7) {
+                    await expectSameAsNative(
+                        Buffer.concat([
+                            body.subarray(0, at),
+                            body.subarray(at + 1),
+                        ]),
+                    );
+                } else {
+                    await expectSameAsNative(body.subarray(0, at));
+                }
+            }
+        }
+    }, 120_000);
+
+    it.each([
+        ['an empty body', ''],
+        ['only whitespace', ' \n '],
+        ['two values', '[1] [2]'],
+        ['a trailing comma in an array', '[1,2,]'],
+        ['a trailing comma in an object', '{"a":1,}'],
+        ['a leading comma', '[,1]'],
+        ['a doubled comma', '[1,,2]'],
+        ['a missing colon', '{"a" 1}'],
+        ['a missing value', '{"a":}'],
+        ['a bare key', '{a:1}'],
+        ['a colon in an array', '[1:2]'],
+        ['mismatched brackets', '[1}'],
+        ['an extra close', '[1]]'],
+        ['an unterminated string', '["abc'],
+        ['an unterminated array', '[1,2'],
+        ['a byte order mark', '\ufeff[1]'],
+        ['a comment', '[1/* x */]'],
+        ['a bad literal', '[tru]'],
+        ['a bad number', '[01]'],
+        ['a bad escape', '["\\x"]'],
+        ['a raw control character in a string', '["a\tb"]'],
+        ['a wide space between values', '[1,\u00a02]'],
+    ])('rejects %s', async (_name, text) => {
+        await expectSameAsNative(text);
+        // Short input takes the native path, so check the sliced one too.
+        await expectSameAsNative(padded(text));
+        await expectSameAsNative(
+            `[${'1,'.repeat(SLICE_BYTES)}${text},${'2,'.repeat(SLICE_BYTES)}3]`,
+        );
+    });
+
+    it('reads a string longer than one piece, wherever a piece ends', async () => {
+        const units = [
+            '\\u00e9',
+            '\\n',
+            '\\\\',
+            '\\"',
+            'é',
+            '日',
+            '😀',
+            '\\ud83d\\ude00',
+        ];
+        for (const unit of units) {
+            for (let shift = 0; shift < 9; shift++) {
+                const lead = 'a'.repeat(STRING_PIECE_BYTES - 12 + shift);
+                const text = `{"text":"${lead}${unit.repeat(8)}${lead}","n":1}`;
+                await expectSameAsNative(text);
+            }
+        }
+    }, 60_000);
+
+    it('reads long runs of escapes and stray UTF-8 bytes', async () => {
+        const length = STRING_PIECE_BYTES + SLICE_BYTES;
+        await expectSameAsNative(`["${'\\\\'.repeat(length)}"]`);
+        await expectSameAsNative(`["${'\\"'.repeat(length)}"]`);
+        await expectSameAsNative(`["${'\\'.repeat(2 * length + 1)}"]`);
+        await expectSameAsNative(
+            Buffer.concat([
+                Buffer.from('["'),
+                Buffer.alloc(2 * length, 0x80),
+                Buffer.from([0xe2, 0x82]),
+                Buffer.from('"]'),
+            ]),
+        );
+    }, 60_000);
+
+    it('reads a long key', async () => {
+        const key = 'k'.repeat(2 * SLICE_BYTES);
+        await expectSameAsNative(`{"a":1,"${key}":{"b":[2]},"${key}x":"v"}`);
+        await expectSameAsNative(`{"${key}\\n":2,"z":3}`);
+    });
+
+    it('reads long gaps of whitespace', async () => {
+        const gap = ' '.repeat(2 * SLICE_BYTES);
+        await expectSameAsNative(
+            `${gap}{${gap}"a"${gap}:${gap}1${gap},${gap}"b":[${gap}2${gap},3${gap}]${gap}}${gap}`,
+        );
+        await expectSameAsNative(`[1,${gap}2${gap}]`);
+    });
+
+    it('reads a scalar or a string at the top level', async () => {
+        await expectSameAsNative(padded('12.5'));
+        await expectSameAsNative(padded('null'));
+        await expectSameAsNative(`"${'x'.repeat(4 * SLICE_BYTES)}"`);
+    });
+
+    it('keeps the last of a repeated key, in its first position', async () => {
+        const filler = `"f":"${'x'.repeat(SLICE_BYTES)}"`;
+        await expectSameAsNative(`{"a":1,"b":2,${filler},"a":3,"7":4,"3":5}`);
+        const value = (await parseJsonInSlices(
+            Buffer.from(`{"a":1,"b":2,${filler},"a":3}`),
+        )) as Record<string, unknown>;
+        expect(Object.keys(value)).toEqual(['a', 'b', 'f']);
+        expect(value.a).toBe(3);
+    });
+
+    it('keeps __proto__ as a plain own property', async () => {
+        const filler = `"f":"${'x'.repeat(SLICE_BYTES)}"`;
+        const bodies = [
+            // In a batch of members, as a container of its own, and alone.
+            `{${filler},"__proto__":{"polluted":true},"z":1}`,
+            `{"__proto__":{"polluted":true,${filler}}}`,
+            `{"__proto__":"${'y'.repeat(2 * SLICE_BYTES)}"}`,
+        ];
+        for (const body of bodies) {
+            const value = (await parseJsonInSlices(
+                Buffer.from(body),
+            )) as Record<string, unknown>;
+            expect(Object.getPrototypeOf(value)).toBe(Object.prototype);
+            expect(Object.hasOwn(value, '__proto__')).toBe(true);
+            expect(({} as { polluted?: boolean }).polluted).toBeUndefined();
+        }
+    });
+
+    it('keeps the sign of negative zero', async () => {
+        const value = (await parseJsonInSlices(
+            Buffer.from(`[${'0,'.repeat(SLICE_BYTES)}-0]`),
+        )) as number[];
+        expect(Object.is(value.at(-1), -0)).toBe(true);
+    });
+
+    it('reads nesting up to the depth limit and refuses more', async () => {
+        const nested = (depth: number) =>
+            padded(`${'['.repeat(depth)}${']'.repeat(depth)}`);
+        // The root is not a level of nesting.
+        let value = await parseJsonInSlices(Buffer.from(nested(MAX_DEPTH)));
+        let depth = 0;
+        while (Array.isArray(value)) {
+            depth++;
+            value = value[0];
+        }
+        expect(depth).toBe(MAX_DEPTH);
+
+        await expect(
+            parseJsonInSlices(Buffer.from(nested(MAX_DEPTH + 1))),
+        ).rejects.toThrow(SyntaxError);
+    });
+
+    it('reads members spread thinly across many levels', async () => {
+        // No level holds a slice of members, but together they hold many.
+        const members = '1,'.repeat(SLICE_BYTES / 8);
+        const depth = 200;
+        await expectSameAsNative(
+            `${`[${members}`.repeat(depth)}0${']'.repeat(depth)}`,
+        );
+        await expectSameAsNative(
+            `${`{"a":[${members}0],"b":`.repeat(depth)}0${'}'.repeat(depth)}`,
+        );
+    });
+
+    it('lets the event loop run while it parses', async () => {
+        // Dense input: the slowest kind to parse for its size.
+        const body = Buffer.from(`[${'[],'.repeat(1_000_000)}[]]`);
+        let turns = 0;
+        let running = true;
+        const tick = () => {
+            turns++;
+            if (running) setImmediate(tick);
+        };
+        setImmediate(tick);
+        const value = (await parseJsonInSlices(body)) as unknown[];
+        running = false;
+        expect(value).toHaveLength(1_000_001);
+        expect(turns).toBeGreaterThan(5);
+    });
+});

src/backend/util/slicedJson.ts

@@ -0,0 +1,581 @@
+/*
+ * Copyright (C) 2024-present Puter Technologies Inc.
+ *
+ * This file is part of Puter.
+ *
+ * Puter is free software: you can redistribute it and/or modify
+ * it under the terms of the GNU Affero General Public License as published
+ * by the Free Software Foundation, either version 3 of the License, or
+ * (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+ * GNU Affero General Public License for more details.
+ *
+ * You should have received a copy of the GNU Affero General Public License
+ * along with this program.  If not, see <https://www.gnu.org/licenses/>.
+ */
+
+import { setImmediate as nextTurn } from 'node:timers/promises';
+
+/**
+ * Text handed to one native `JSON.parse` call stays within a few multiples of
+ * this, which bounds how long the call can run. Shorter input parses whole.
+ */
+export const SLICE_BYTES = 32 * 1024;
+/** Bytes of a long string decoded per turn. */
+export const STRING_PIECE_BYTES = 512 * 1024;
+/** Deeper nesting is refused. No mainstream serializer writes this deep. */
+export const MAX_DEPTH = 10_000;
+/** How long a parse runs before the event loop gets a turn. */
+const TURN_BUDGET_MS = 3;
+/** Bytes scanned between clock reads. */
+const CLOCK_INTERVAL_BYTES = 16 * 1024;
+/** Bytes of a string scanned one at a time before a native search takes over. */
+const BYTEWISE_BYTES = 64;
+/** Rounds of a string scan between clock reads. */
+const CLOCK_INTERVAL_ROUNDS = 64;
+
+const QUOTE = 0x22;
+const BACKSLASH = 0x5c;
+
+// Byte classes. Anything unlisted is part of a number or literal.
+const C_SCALAR = 0;
+const C_SPACE = 1;
+const C_QUOTE = 2;
+const C_OPEN_ARRAY = 3;
+const C_OPEN_OBJECT = 4;
+const C_CLOSE_ARRAY = 5;
+const C_CLOSE_OBJECT = 6;
+const C_COMMA = 7;
+const C_COLON = 8;
+const BYTE_CLASS = new Uint8Array(256);
+for (const byte of [0x20, 0x09, 0x0a, 0x0d]) BYTE_CLASS[byte] = C_SPACE;
+BYTE_CLASS[QUOTE] = C_QUOTE;
+BYTE_CLASS[0x5b] = C_OPEN_ARRAY;
+BYTE_CLASS[0x7b] = C_OPEN_OBJECT;
+BYTE_CLASS[0x5d] = C_CLOSE_ARRAY;
+BYTE_CLASS[0x7d] = C_CLOSE_OBJECT;
+BYTE_CLASS[0x2c] = C_COMMA;
+BYTE_CLASS[0x3a] = C_COLON;
+
+const ROOT = 0;
+const ARRAY = 1;
+const OBJECT = 2;
+
+/** Just opened: a member or the close. */
+const S_FIRST = 0;
+/** After a comma: a member. */
+const S_MEMBER = 1;
+/** After an object key. */
+const S_COLON = 2;
+/** After a colon, and the root before its value. */
+const S_VALUE = 3;
+/** After a value: a comma or the close. The root is done. */
+const S_AFTER = 4;
+
+const DONE = 0;
+const TURN = 1;
+const LONG_STRING = 2;
+
+/**
+ * One open container, or the root. Members are scanned, not parsed: `pending`
+ * marks text handed to `JSON.parse` once enough of it has built up, and a
+ * container too large for one call gets its `out` filled a batch at a time.
+ */
+class Frame {
+    kind = ROOT;
+    state = S_VALUE;
+    /** Offset of the opening bracket. */
+    start = 0;
+    /** The container under construction, once it is known to be large. */
+    out: unknown[] | Record<string, unknown> | null = null;
+    /** Where the members not yet parsed start, or -1. */
+    pending = -1;
+    /** Where the last complete member ended. */
+    lastEnd = -1;
+    /** Bounds of the member being scanned. */
+    memberStart = -1;
+    keyStart = -1;
+    keyEnd = -1;
+    valueStart = -1;
+    /** The current member's key, when the member is parsed apart from `pending`. */
+    ownKey: string | null = null;
+    /** Bytes of the parent's `pending` waiting behind this container. */
+    backlog = 0;
+
+    open(kind: number, start: number): void {
+        this.kind = kind;
+        this.state = kind === ROOT ? S_VALUE : S_FIRST;
+        this.start = start;
+        this.out = null;
+        this.pending = -1;
+        this.lastEnd = -1;
+        this.memberStart = -1;
+        this.keyStart = -1;
+        this.keyEnd = -1;
+        this.valueStart = -1;
+        this.ownKey = null;
+        this.backlog = 0;
+    }
+}
+
+const unexpected = (at: number): SyntaxError =>
+    new SyntaxError(`Unexpected token in JSON at position ${at}`);
+
+/** `JSON.parse` keeps `__proto__` as an own property; assignment would not. */
+const setOwn = (
+    target: Record<string, unknown>,
+    key: string,
+    value: unknown,
+): void => {
+    if (key === '__proto__') {
+        Object.defineProperty(target, key, {
+            value,
+            writable: true,
+            enumerable: true,
+            configurable: true,
+        });
+    } else {
+        target[key] = value;
+    }
+};
+
+class SlicedJsonParser {
+    result: unknown;
+    /** Bounds of the string `run` stopped at, quotes included. */
+    longStart = -1;
+    longEnd = -1;
+
+    readonly #body: Buffer;
+    readonly #end: number;
+    #pos = 0;
+    #frames = [new Frame()];
+    #depth = 0;
+    /** Frames below this depth have an `out`; the root counts as one. */
+    #filling = 1;
+    /** Bytes of `pending` in frames below the top one. */
+    #backlog = 0;
+    #inScalar = false;
+    /** The string being scanned, and where its scan resumes. */
+    #stringStart = -1;
+    #stringPos = 0;
+    /** Next quote and backslash at or after the last search origin. */
+    #quoteAt = -1;
+    #backslashAt = -1;
+    #longIsKey = false;
+
+    constructor(body: Buffer) {
+        this.#body = body;
+        this.#end = body.length;
+    }
+
+    /** Scans until the input ends, `deadline` passes, or a long string starts. */
+    run(deadline: number): number {
+        const body = this.#body;
+        const end = this.#end;
+        let i = this.#pos;
+        let frame = this.#frames[this.#depth]!;
+        let clockAt = i + CLOCK_INTERVAL_BYTES;
+
+        for (;;) {
+            if (this.#stringStart !== -1) {
+                const close = this.#stringEnd(deadline);
+                if (close === -1) return TURN;
+                const start = this.#stringStart;
+                this.#stringStart = -1;
+                i = close + 1;
+                if (this.#string(frame, start, close)) {
+                    this.#pos = i;
+                    return LONG_STRING;
+                }
+                continue;
+            }
+            if (i >= clockAt) {
+                clockAt = i + CLOCK_INTERVAL_BYTES;
+                if (performance.now() >= deadline) {
+                    this.#pos = i;
+                    return TURN;
+                }
+            }
+            if (i >= end) break;
+
+            const byteClass = BYTE_CLASS[body[i]!]!;
+            if (byteClass === C_SCALAR) {
+                if (!this.#inScalar) {
+                    this.#valueStart(frame, i);
+                    this.#inScalar = true;
+                }
+                const stop = Math.min(end, clockAt);
+                i++;
+                while (i < stop && BYTE_CLASS[body[i]!] === C_SCALAR) i++;
+                continue;
+            }
+            if (this.#inScalar) {
+                this.#inScalar = false;
+                this.#valueEnd(frame, i);
+            }
+
+            switch (byteClass) {
+                case C_SPACE: {
+                    const stop = Math.min(end, clockAt);
+                    i++;
+                    while (i < stop && BYTE_CLASS[body[i]!] === C_SPACE) i++;
+                    break;
+                }
+                case C_QUOTE:
+                    this.#stringStart = i;
+                    this.#stringPos = i + 1;
+                    break;
+                case C_OPEN_ARRAY:
+                case C_OPEN_OBJECT: {
+                    this.#valueStart(frame, i);
+                    if (this.#depth === MAX_DEPTH) {
+                        throw new SyntaxError('JSON is nested too deeply');
+                    }
+                    const parent = frame;
+                    // Members before this one wait until it closes or grows.
+                    const before =
+                        parent.pending !== -1 &&
+                        parent.pending !== parent.memberStart
+                            ? parent.lastEnd - parent.pending
+                            : 0;
+                    this.#depth++;
+                    frame = this.#frames[this.#depth] ??= new Frame();
+                    frame.open(byteClass === C_OPEN_ARRAY ? ARRAY : OBJECT, i);
+                    frame.backlog = before;
+                    this.#backlog += before;
+                    // Parsed now, so deep nesting can't save it all up.
+                    if (this.#backlog >= SLICE_BYTES) this.#fill();
+                    i++;
+                    break;
+                }
+                case C_CLOSE_ARRAY:
+                case C_CLOSE_OBJECT: {
+                    if (
+                        frame.kind !==
+                            (byteClass === C_CLOSE_ARRAY ? ARRAY : OBJECT) ||
+                        (frame.state !== S_FIRST && frame.state !== S_AFTER)
+                    ) {
+                        throw unexpected(i);
+                    }
+                    let filled = this.#depth < this.#filling;
+                    if (!filled && i + 1 - frame.start > SLICE_BYTES) {
+                        this.#fill();
+                        filled = true;
+                    }
+                    if (filled) {
+                        if (frame.pending !== -1) {
+                            this.#parsePending(frame);
+                        }
+                        frame.out = null;
+                        this.#filling = this.#depth;
+                    }
+                    this.#backlog -= frame.backlog;
+                    this.#depth--;
+                    frame = this.#frames[this.#depth]!;
+                    i++;
+                    if (filled) {
+                        frame.state = S_AFTER;
+                        frame.lastEnd = i;
+                    } else {
+                        this.#valueEnd(frame, i);
+                    }
+                    break;
+                }
+                case C_COMMA:
+                    if (frame.kind === ROOT || frame.state !== S_AFTER) {
+                        throw unexpected(i);
+                    }
+                    frame.state = S_MEMBER;
+                    i++;
+                    break;
+                default:
+                    if (frame.kind !== OBJECT || frame.state !== S_COLON) {
+                        throw unexpected(i);
+                    }
+                    frame.state = S_VALUE;
+                    i++;
+            }
+        }
+
+        if (this.#inScalar) {
+            this.#inScalar = false;
+            this.#valueEnd(frame, end);
+        }
+        if (this.#depth !== 0 || frame.state !== S_AFTER) {
+            throw new SyntaxError('Unexpected end of JSON input');
+        }
+        if (frame.pending !== -1) this.#parsePending(frame);
+        return DONE;
+    }
+
+    /** Supplies the string `run` stopped at. */
+    resolveLongString(value: string): void {
+        const frame = this.#frames[this.#depth]!;
+        if (this.#longIsKey) {
+            frame.ownKey = value;
+            return;
+        }
+        this.#attach(frame, frame.ownKey, value);
+        frame.ownKey = null;
+        frame.lastEnd = this.longEnd + 1;
+        frame.state = S_AFTER;
+    }
+
+    /**
+     * Offset of the quote closing the string being scanned, or -1 when
+     * `deadline` passed first.
+     */
+    #stringEnd(deadline: number): number {
+        const body = this.#body;
+        const end = this.#end;
+        let at = this.#stringPos;
+
+        for (let round = 1; ; round++) {
+            // Bytewise through a short string or a stretch of escapes.
+            const near = Math.min(at + BYTEWISE_BYTES, end);
+            while (at < near) {
+                const byte = body[at]!;
+                if (byte === QUOTE) return at;
+                at += byte === BACKSLASH ? 2 : 1;
+            }
+            if (at >= end) throw new SyntaxError('Unterminated string in JSON');
+
+            // Then a native search across text that holds neither.
+            if (this.#quoteAt < at) {
+                this.#quoteAt = body.indexOf(QUOTE, at);
+                if (this.#quoteAt === -1) {
+                    throw new SyntaxError('Unterminated string in JSON');
+                }
+            }
+            if (this.#backslashAt < at) {
+                const found = body.indexOf(BACKSLASH, at);
+                this.#backslashAt = found === -1 ? end : found;
+            }
+            if (this.#backslashAt > this.#quoteAt) return this.#quoteAt;
+            at = this.#backslashAt;
+
+            if (round % CLOCK_INTERVAL_ROUNDS === 0) {
+                if (performance.now() >= deadline) {
+                    this.#stringPos = at;
+                    return -1;
+                }
+            }
+        }
+    }
+
+    /** Handles a scanned string; true when `run` must stop for a long one. */
+    #string(frame: Frame, start: number, close: number): boolean {
+        const long = close - start > SLICE_BYTES;
+        if (
+            frame.kind === OBJECT &&
+            (frame.state === S_FIRST || frame.state === S_MEMBER)
+        ) {
+            this.#memberStart(frame, start);
+            frame.keyStart = start;
+            frame.keyEnd = close;
+            frame.state = S_COLON;
+            if (!long) return false;
+            this.#detachMember(frame);
+            this.#longIsKey = true;
+        } else {
+            this.#valueStart(frame, start);
+            if (!long) {
+                this.#valueEnd(frame, close + 1);
+                return false;
+            }
+            if (frame.ownKey === null) {
+                this.#detachMember(frame);
+                if (frame.kind === OBJECT) frame.ownKey = this.#key(frame);
+            }
+            this.#longIsKey = false;
+        }
+        this.longStart = start;
+        this.longEnd = close;
+        return true;
+    }
+
+    #memberStart(frame: Frame, at: number): void {
+        if (frame.pending !== -1 && at - frame.pending >= SLICE_BYTES) {
+            this.#fill();
+            this.#parsePending(frame);
+        }
+        frame.memberStart = at;
+        if (frame.pending === -1) frame.pending = at;
+    }
+
+    #valueStart(frame: Frame, at: number): void {
+        if (frame.kind === OBJECT) {
+            if (frame.state !== S_VALUE) throw unexpected(at);
+            // A long gap after the key would reach `JSON.parse` with it.
+            if (frame.ownKey === null && at - frame.memberStart > SLICE_BYTES) {
+                this.#detachMember(frame);
+                frame.ownKey = this.#key(frame);
+            }
+        } else {
+            if (
+                frame.kind === ROOT
+                    ? frame.state !== S_VALUE
+                    : frame.state > S_MEMBER
+            ) {
+                throw unexpected(at);
+            }
+            this.#memberStart(frame, at);
+        }
+        frame.valueStart = at;
+    }
+
+    #valueEnd(frame: Frame, at: number): void {
+        frame.lastEnd = at;
+        frame.state = S_AFTER;
+        if (frame.ownKey !== null) {
+            this.#attach(
+                frame,
+                frame.ownKey,
+                JSON.parse(this.#text(frame.valueStart, at)),
+            );
+            frame.ownKey = null;
+        }
+    }
+
+    /** Takes the current member out of `pending`; its value is attached alone. */
+    #detachMember(frame: Frame): void {
+        this.#fill();
+        if (frame.pending !== -1 && frame.pending !== frame.memberStart) {
+            this.#parsePending(frame);
+        }
+        frame.pending = -1;
+    }
+
+    /**
+     * Gives every open container an `out`, attached to its parent after the
+     * members that precede it.
+     */
+    #fill(): void {
+        const frames = this.#frames;
+        for (let level = this.#filling; level <= this.#depth; level++) {
+            const parent = frames[level - 1]!;
+            const child = frames[level]!;
+            if (
+                parent.pending !== -1 &&
+                parent.pending !== parent.memberStart
+            ) {
+                this.#parsePending(parent);
+            }
+            parent.pending = -1;
+            const key =
+                parent.kind === OBJECT
+                    ? (parent.ownKey ?? this.#key(parent))
+                    : null;
+            parent.ownKey = null;
+            child.out = child.kind === ARRAY ? [] : {};
+            child.backlog = 0;
+            this.#attach(parent, key, child.out);
+        }
+        this.#filling = this.#depth + 1;
+        this.#backlog = 0;
+    }
+
+    /** Parses the complete members in `pending` into the frame's container. */
+    #parsePending(frame: Frame): void {
+        const text = this.#text(frame.pending, frame.lastEnd);
+        frame.pending = -1;
+        if (frame.kind === ROOT) {
+            this.result = JSON.parse(text);
+        } else if (frame.kind === ARRAY) {
+            const out = frame.out as unknown[];
+            const items = JSON.parse(`[${text}]`) as unknown[];
+            for (let k = 0; k < items.length; k++) out.push(items[k]);
+        } else {
+            const out = frame.out as Record<string, unknown>;
+            const members = JSON.parse(`{${text}}`) as Record<string, unknown>;
+            for (const key of Object.keys(members)) {
+                setOwn(out, key, members[key]);
+            }
+        }
+    }
+
+    #attach(frame: Frame, key: string | null, value: unknown): void {
+        if (frame.kind === ROOT) this.result = value;
+        else if (frame.kind === ARRAY) (frame.out as unknown[]).push(value);
+        else setOwn(frame.out as Record<string, unknown>, key!, value);
+    }
+
+    #key(frame: Frame): string {
+        return JSON.parse(
+            this.#text(frame.keyStart, frame.keyEnd + 1),
+        ) as string;
+    }
+
+    #text(from: number, to: number): string {
+        return this.#body.toString('utf8', from, to);
+    }
+}
+
+/**
+ * Where a string's content can be cut at or before `target` without splitting
+ * an escape or a UTF-8 sequence. `from` starts a character or an escape.
+ */
+const pieceEnd = (body: Buffer, from: number, target: number): number => {
+    // Far enough back that a `\uXXXX` reaching `target` starts after it.
+    let at = Math.max(from, target - 8);
+    // An odd run of backslashes before `at` means `at` is an escaped byte.
+    let run = 0;
+    while (at - run > from && body[at - run - 1] === BACKSLASH) run++;
+    at -= run % 2;
+    for (;;) {
+        const length =
+            body[at] !== BACKSLASH ? 1 : body[at + 1] === 0x75 ? 6 : 2;
+        if (at + length > target) break;
+        at += length;
+    }
+    // A sequence has at most three continuation bytes; more are stray.
+    for (let back = 0; back < 3 && (body[at]! & 0xc0) === 0x80; back++) at--;
+    return at;
+};
+
+/** The string literal at `[start, close]`, decoded a piece per turn. */
+const readLongString = async (
+    body: Buffer,
+    start: number,
+    close: number,
+): Promise<string> => {
+    const pieces: string[] = [];
+    let at = start + 1;
+    while (at < close) {
+        const end =
+            close - at <= STRING_PIECE_BYTES
+                ? close
+                : pieceEnd(body, at, at + STRING_PIECE_BYTES);
+        pieces.push(
+            JSON.parse(`"${body.toString('utf8', at, end)}"`) as string,
+        );
+        at = end;
+        await nextTurn();
+    }
+    return pieces.join('');
+};
+
+/**
+ * `JSON.parse` of UTF-8 `body`, done in slices with the event loop running
+ * between them, so no input holds it for long. Throws a `SyntaxError` for
+ * malformed JSON, and for nesting deeper than `MAX_DEPTH`.
+ */
+export async function parseJsonInSlices(body: Buffer): Promise<unknown> {
+    if (body.length <= SLICE_BYTES) return JSON.parse(body.toString('utf8'));
+
+    const parser = new SlicedJsonParser(body);
+    for (;;) {
+        const signal = parser.run(performance.now() + TURN_BUDGET_MS);
+        if (signal === DONE) return parser.result;
+        if (signal === LONG_STRING) {
+            parser.resolveLongString(
+                await readLongString(body, parser.longStart, parser.longEnd),
+            );
+        } else {
+            await nextTurn();
+        }
+    }
+}

src/dev-center/js/apps.js

@@ -1339,7 +1339,7 @@ async function saveEditedApp (mode) {
         // Update title
         $(`.app-title[data-uid="${uid}"]`).html(html_encode(app.title));
         // Update app link
-        $(`.app-url[data-uid="${uid}"]`).html(applink(app));
+        $(`.app-url[data-uid="${uid}"]`).html(html_encode(applink(app)));
         $(`.app-url[data-uid="${uid}"]`).attr('href', applink(app));
         // Update icons
         $(`.app-icon[data-uid="${uid}"]`).attr('src', html_encode(app.icon ? app.icon : './img/app.svg'));

src/dev-center/js/dev-center.js

@@ -38,22 +38,27 @@ window.user = null;
     }
 })();
 
-// domain and APIOrigin
-if ( window.url_params.has('puter.domain') ) {
-    window.domain = window.url_params.get('puter.domain');
+// domain and port are built into links and deploy URLs, so anything other
+// than a bare hostname and a numeric port is ignored
+const domainParam = window.url_params.get('puter.domain');
+if ( domainParam && /^[a-z0-9-]+(\.[a-z0-9-]+)*$/i.test(domainParam) ) {
+    window.domain = domainParam;
 }
 
+const portParam = window.url_params.get('puter.port');
+window.port = portParam && /^\d{1,5}$/.test(portParam) ? portParam : '';
+
 // static hosting domain
 window.static_hosting_domain = 'puter.site';
-   if ( window.domain === 'puter.localhost' ) {
-       window.static_hosting_domain = 'site.puter.localhost';
-   } else if ( window.domain && window.domain !== 'puter.com' ) {
-       window.static_hosting_domain = `site.${window.domain}`;
-   }
+if ( window.domain === 'puter.localhost' ) {
+    window.static_hosting_domain = 'site.puter.localhost';
+} else if ( window.domain && window.domain !== 'puter.com' ) {
+    window.static_hosting_domain = `site.${window.domain}`;
+}
 
 // add port to static_hosting_domain if provided
-if ( window.url_params.has('puter.port') && window.url_params.get('puter.port') ) {
-    window.static_hosting_domain = `${window.static_hosting_domain }:${ html_encode(window.url_params.get('puter.port'))}`;
+if ( window.port ) {
+    window.static_hosting_domain = `${window.static_hosting_domain}:${window.port}`;
 }
 
 // protocol
@@ -63,12 +68,6 @@ if ( window.url_params.has('puter.protocol') && window.url_params.get('puter.pro
     window.protocol = 'http';
 }
 
-// port
-window.port = '';
-if ( window.url_params.has('puter.port') && window.url_params.get('puter.port') ) {
-    window.port = html_encode(window.url_params.get('puter.port'));
-}
-
 // source_path
 if ( window.url_params.has('source_path') ) {
     window.source_path = window.url_params.get('source_path');

src/docs/src/AI.md

@@ -200,8 +200,8 @@ These AI features are supported out of the box when using Puter.js:
 - **[`puter.ai.img2txt()`](/AI/img2txt/)** - Extract text from images (OCR)
 - **[`puter.ai.txt2vid()`](/AI/txt2vid/)** - Generate short video clips from text or a reference image with Wan, Seedance, Veo, Kling, MiniMax H3 and other models, or edit a video you supply with Gemini Omni, FLUX 3 or Kling Omni
 - **[`puter.ai.speech2speech()`](/AI/speech2speech/)** - Convert speech in one voice to another voice
-- **[`puter.ai.convo()`](/AI/convo/)** - Prepare a beta realtime conversation with optional text, microphone, playback, and interruption helpers
-- **[Realtime WebSocket API](/AI/realtime/)** - Connect directly without puter.js. This API is in beta and may change abruptly.
+- **[`puter.ai.realtime()`](/AI/realtime/)** - Prepare a beta realtime conversation with optional text, microphone, playback, and interruption helpers
+- **[Realtime WebSocket API](/AI/realtime-websocket/)** - Connect directly without puter.js. This API is in beta and may change abruptly.
 - **[`puter.ai.listModels()`](/AI/listModels/)** - List available AI chat models (and providers) that Puter currently exposes.
 - **[`puter.ai.listModelProviders()`](/AI/listModelProviders/)** - List the AI providers that Puter currently exposes.
 - **[`puter.ai.txt2speech.listEngines()`](/AI/txt2speech.listEngines/)** - List available TTS engines/models

src/docs/src/AI/chat.md

@@ -30,6 +30,7 @@ An object containing the following properties:
 | Option | Type | Description |
 |--------|------|-------------|
 | `model` | `String` | The model you want to use for the completion. If not specified, defaults to `gpt-5.4-nano`. More than 500 models are available from providers including OpenAI, Anthropic, Google, Alibaba Cloud, xAI, Mistral, OpenRouter, Infron, and others. For a full list, see the [AI models list](https://developer.puter.com/ai/models/) page. |
+| `serviceSpeed` | `String` | Preferred service speed, `"standard"`, `"fast"`, or `"ultrafast"`. Unavailable speeds fall back from ultrafast to fast, then standard. Omitting it preserves the model's default service. Faster service may cost more. See [Service speed](#service-speed) and [Model-specific features](#model-specific-features). |
 | `stream` | `Boolean` | A boolean indicating whether you want to stream the completion. Defaults to `false`. |
 | `max_tokens` | `Number` | The maximum number of tokens to generate in the completion. By default, the specific model's maximum is used. |
 | `temperature` | `Number` | A number between 0 and 2 indicating the randomness of the completion. Lower values make the output more focused and deterministic, while higher values make it more random. By default, the model's own temperature is used. See [Model-specific features](#model-specific-features). |
@@ -138,6 +139,8 @@ by a newline. Use the async iterator to receive every chunk as an object.
 
 In case of an error, the `Promise` will reject with an error message.
 
+An unrecognized `serviceSpeed` value rejects with `invalid_service_speed`.
+
 For streaming responses, an error from the AI provider is returned as a chunk
 with `type: "error"` and a `message` describing the failure. This also happens
 if the provider stops sending the response before it is complete. Check for
@@ -370,10 +373,20 @@ const response = await puter.ai.chat('Hello', {
 
 With `provider` set as well, the request is answered by that provider or fails. If that provider does not serve the model, the request fails with a `400` instead of going to another provider. The error is the same one you would get if every provider had failed, for example `400 upstream_provider_unavailable` or `429 upstream_rate_limited`.
 
+## Service speed
+
+The `serviceSpeed` option selects service capacity independently of reasoning effort. If the requested speed is unavailable for the selected model, the request uses the next supported speed in the order `ultrafast`, `fast`, `standard`.
+
+A faster tier that rejects a request because its capacity or access is unavailable can retry at a lower speed on the same provider. The `fallback` option controls retries on other providers separately. Once streaming output starts, a failed stream ends with an error instead of restarting at another speed.
+
+The `usageDetails.serviceSpeed` field reports the served speed. Billing uses the tier or equivalent model that served the request. A provider downgrade is billed at the lower tier's rates. Response normalization follows the same rule as a request without this option, including when a faster equivalent model is selected.
+
 ## Model-specific features
 
 ### OpenAI models
 
+Fast service is available on models with a published Fast rate card. Ultrafast is available for GPT-6.1 Sol and GPT-6 Astra. Other models fall back to fast or standard. Cache and long-context rates apply to the served tier.
+
 #### `temperature`
 
 Some OpenAI reasoning models ignore `temperature` and `top_p`. GPT-5.6, GPT-6 Sol/Luna, and GPT-6.1 Terra accept them only with `reasoning_effort: 'none'`, and GPT-6 Astra, GPT-6.1 Sol, and GPT-5.3 Codex always ignore them.
@@ -418,6 +431,8 @@ The `usageDetails` object on every result carries the same counts under provider
 
 ### Anthropic models
 
+Fast service is available on Claude models with a Fast rate card. Ultrafast falls back to fast. When the account has no Fast quota, the request retries at standard speed. The upstream service tier and inference speed are separate fields.
+
 #### Prompt Caching
 
 Specific to Anthropic models, you can use the cache control feature, allowing you to optimize costs for repeated prompts.
@@ -442,6 +457,8 @@ You can find the implementation in our [prompt caching example](/playground/ai-c
 
 ### Gemini models
 
+Fast service uses Priority inference on supported text and image requests. Other media use standard service to preserve input handling.
+
 #### `temperature`
 
 `gemini-4-argon` ignores `temperature`. `gemini-3.1-flash-lite` and `gemini-3-flash-preview` use it.
@@ -543,6 +560,8 @@ for await (const part of resp) {
 
 ### Moonshot Kimi models
 
+Models with an equivalent Highspeed variant select that variant for fast service. Other models use standard service.
+
 #### `reasoning_effort`
 
 Kimi K3.1 Pro accepts `low`, `high` and `max`, uses `max` when you leave it unset, and fails with `400 bad_request` on any other value.
@@ -557,6 +576,30 @@ Moonshot's Kimi K3 and Kimi K3.1 Pro models cache prompt prefixes on their own a
 
 Muse Spark always reasons, so `none` is ignored for it.
 
+### xAI models
+
+Fast service requests Priority processing. Token charges, including cached input and reasoning, follow the served tier.
+
+### MiniMax models
+
+MiniMax M3 uses Priority admission for fast service. Earlier models with an equivalent Highspeed variant use that variant. Other models use standard service.
+
+### Mistral models
+
+Fast service requests Priority when the account and model are eligible. The provider can serve the request at standard speed.
+
+### Alibaba Cloud models
+
+Models with an equivalent Prime or Fast variant select that variant for fast service. Other models use standard service.
+
+### OpenRouter models
+
+Faster tiers come from the selected model's endpoint catalog. A tier can fall back upstream. Billing uses the endpoint that serves the request.
+
+### Infron models
+
+Models with a Priority offering select that offering for fast service. Other models use standard service.
+
 ## Examples
 
 ### Ask GPT-6 Luna a question
@@ -1092,3 +1135,23 @@ Policy 8 - Account Management: Each Enterprise and Ultimate customer is assigned
 </body>
 </html>
 ```
+
+### Select service speed
+
+For example, request the fastest supported service:
+
+```html;ai-chat-service-speed
+<html>
+<body>
+    <script src="https://js.puter.com/v2/"></script>
+    <script type="module">
+        const response = await puter.ai.chat('Explain closures in one paragraph.', {
+            model: 'gpt-6.1-sol',
+            serviceSpeed: 'ultrafast',
+        });
+        puter.print(response.message.content);
+        puter.print(response.usageDetails.serviceSpeed);
+    </script>
+</body>
+</html>
+```

src/docs/src/AI/convo.md

@@ -1,151 +0,0 @@
----
-title: puter.ai.convo()
-description: Prepare an OpenAI realtime conversation with optional audio and text helpers.
-platforms: [websites, apps, nodejs, workers]
----
-
-Prepares a realtime OpenAI conversation that your app can connect to when it's ready.
-
-Preparing the conversation verifies Puter sign-in. Realtime conversations require a paid Puter plan, and usage is billed to the signed-in user. Native OpenAI events and convenience methods use the same session. Microphone capture and speaker playback start when your app requests them.
-
-In Node.js, the SDK requires a global `WebSocket`, available in Node.js 22 or later.
-
-**Beta:** This API is incomplete. Its methods, options, events, and behavior may change abruptly, including breaking changes without notice.
-
-To connect without puter.js, see the [Realtime WebSocket API](/AI/realtime/).
-
-## Syntax
-
-```js
-const conversation = await puter.ai.convo(options = {});
-await conversation.connect();
-```
-
-## Parameters
-
-#### `options` (Object) (Optional)
-
-The options object accepts these fields.
-
-| Option | Type | Description |
-|--------|------|-------------|
-| `model` | `String` | The realtime model. Defaults to `gpt-realtime-2.1-mini`. Also accepts `gpt-realtime-2.1`, `gpt-realtime-2`, `gpt-realtime-1.5`, `gpt-realtime`, and `gpt-realtime-mini`. |
-| `instructions` | `String` | Session instructions, up to 8192 characters. Defaults to the provider's default. |
-| `output` | `Array<String>` | Accepts `['text']`, `['audio']`, or both. Defaults to `['audio']`. Audio includes a transcript. Requesting both uses audio output and emits transcript deltas through the `text` event. |
-| `voice` | `String` | OpenAI voice identifier, up to 128 characters. Defaults to the provider's default. |
-| `turnDetection` | `Boolean` | Whether the provider detects the end of spoken turns automatically. Defaults to `true`. Set to `false` for manual audio commits. |
-
-## Return value
-
-A Promise resolves to a prepared conversation after options and sign-in are checked. In a browser, preparation opens the Puter sign-in prompt if needed. Preparation opens no realtime connection and requests no microphone access.
-
-The conversation exposes these properties and methods.
-
-| Member | Behavior |
-| --- | --- |
-| `state` | `prepared`, `connecting`, `connected`, or `closed`. |
-| `model` | Resolved model name. |
-| `connect()` | Opens the connection and applies settings. Resolves after the provider acknowledges session configuration. Concurrent calls share one connection Promise, including bounded retries for transient setup failures. Calling it while connected opens no additional connection. During a reconnection, it resolves when the reconnection succeeds and rejects if the conversation closes instead. |
-| `configure(options)` | Updates the given session settings and resolves after the provider acknowledges this update. Settings it isn't given, including ones set through native events, are left alone. Accepts the factory options except `model`. One update can be pending at a time. Provider rules apply to changes such as switching a voice after audio has been generated. |
-| `sendText(text)` | Adds a user message and requests a response. Text must be nonempty and at most 8192 characters. |
-| `appendAudio(pcm)` | Appends an `Int16Array` containing PCM16, 24 kHz mono audio. A chunk contains at most 12000 samples. |
-| `commitAudio()` | Commits buffered audio and requests a response. Requires turn detection to be off, through `turnDetection: false` or a native session update. The provider requires at least 100 ms of buffered audio. |
-| `interrupt(position?)` | Stops built-in playback immediately and waits for cancellation and playback-history synchronization. A reply requested but not yet started is cancelled as soon as it starts. Custom players provide `{ itemId, contentIndex, playedMs }`. An idle call resolves immediately. Readiness failures reject the returned Promise with `session_not_ready` or `session_closed`. |
-| `send(event)` | Sends a native OpenAI JSON event. |
-| `on(type, handler)` | Registers a listener and returns an unsubscribe function. Use `'*'` to receive every native and convenience event. |
-| `off(type, handler)` | Removes a listener. |
-| `close()` | Closes the session, rejects pending operations, and releases SDK-owned capture and playback resources. Repeated calls have no additional effect. |
-
-Attach listeners before connecting to receive initial session events. Closing during connection rejects the pending attempt.
-
-If the connection drops because of network loss, a server restart or a provider error, the conversation reconnects on its own. Normal closes with code 1000 end the conversation. While it reconnects, `state` is `connecting`, a `reconnecting` event fires, and pending operations reject with `connection_lost`. The new session gets every setting the provider acknowledged, from the factory options, `configure()` and native `session.update` events. Then `ready` fires again with `reconnected: true`. Listeners, the microphone and the speaker carry over. Microphone audio captured while reconnecting is dropped.
-
-The conversation also restores recent context. Before `ready` fires, it adds back up to 40 of the most recent conversation items, 48 KiB in total. Restored items are user text, user speech that has a transcript, assistant replies as text, and function calls with their outputs. User speech has a transcript only when input transcription is configured through native session settings. An assistant reply that was interrupted, or still playing when the connection dropped, is cut to what the user saw or heard. The cut is estimated from how much of the audio played, so it can be off by a few words. Images, speech without a transcript, deleted items, and replies still generating when the connection dropped are left out. Restored items get new item IDs, don't fire item events, and count as input tokens on the next response. Item IDs from before the reconnect don't work in later events. Restoration uses a contiguous suffix of restorable items. An oversized newest item leaves that suffix empty. Older items left out to fit the window fire a `context_trimmed` event with the count in `omittedItems`; this is routine and not an error. A provider refusal or a failure to send the restored items fires one `error` event with code `context_restore_failed` per restoration. In both cases the conversation stays connected. Items are preserved without truncating text, function arguments or function results.
-
-When Puter knows ahead of time that the server a session runs on will restart, it sends a `puter.reconnect` event, which also reaches your listeners. Without that notice, a restart drops the connection and the conversation reconnects as above. Within a few seconds, the conversation opens a replacement session in the background while the current one keeps working. It switches once no reply is generating or playing, no user turn is being spoken or transcribed, no function call is waiting for its output, and nothing you sent is waiting on the provider. The switch goes through the same steps as a reconnect, and the `reconnecting` event has code 1012. If the replacement can't open yet, the conversation keeps the current session and tries again later. A refusal that retrying can't fix, such as `insufficient_funds`, stops the attempts. If the server closes the session first, the conversation switches to the replacement, or reconnects if none is open.
-
-The conversation makes up to ten attempts over about a minute. In a browser that reports being offline, it first waits up to two minutes for the connection to come back. If every attempt fails, or the server refuses with an error such as `insufficient_funds`, the conversation closes. A closed conversation cannot reconnect. To start another session, prepare a new one.
-
-Native event definitions are listed in the [OpenAI realtime reference](https://developers.openai.com/api/reference/resources/realtime/server-events). Native events retain their provider fields. Sending native events preserves direct control of the session. Puter refuses `conversation.item.retrieve` events and MCP tools (`"type": "mcp"`) in `session.update` and `response.create` events. Sending one fires an `error` event with code `invalid_event`, and the conversation closes with code 1008. Audio helpers expect the session to use PCM16 at 24 kHz.
-
-## Errors and limits
-
-Failures use objects containing `code` and `message`. Preparation can reject with `invalid_argument`, `model_not_found`, `auth_required`, `auth_canceled`, or `auth_failed`. Connection can reject with `subscription_required` or `insufficient_funds`, which also show the Puter upgrade prompt. Connection and asynchronous operations can reject with provider errors, `session_closed`, `connection_lost`, `operation_timeout`, or `operation_in_progress`. Sending before readiness, or while reconnecting, throws `session_not_ready`.
-
-Invalid input throws `invalid_argument`, `invalid_event`, or `input_too_large`. A full SDK send queue throws `queue_full`. Media helpers report `unsupported_environment`, `permission_denied`, `media_error`, `media_timeout`, or `queue_full`. Runtime media errors and `context_restore_failed` arrive through the `error` event. Provider errors also arrive as native `error` events with a nested `error` object.
-
-Each native event is capped at 64 KiB. The SDK bounds JSON traversal, outgoing queues, acknowledgment waits, microphone work, and playback. All numeric limits are listed on [rate limits and quotas](/rate-limits-and-quotas/).
-
-## Convenience events
-
-The conversation emits these events alongside native events.
-
-| Event | Fields |
-| --- | --- |
-| `ready` | `model`, `provider`, and `reconnected: true` after a reconnection |
-| `reconnecting` | `code`, `reason` of the dropped connection. After a `puter.reconnect` event, `code` is 1012. |
-| `context_trimmed` | `omittedItems`, the number of restorable items left out of the restored context after a reconnection. Fires before `ready`. |
-| `text` | `delta`, `responseId` |
-| `transcript` | `role`, `text`, `final`. Assistant deltas use `final: false`. Completed assistant and user transcripts use `final: true`. User transcripts require transcription to be configured through native session settings. |
-| `audio` | `pcm`, `sampleRate`, `responseId`, `itemId`, `contentIndex`. The PCM chunk is an `Int16Array`. |
-| `interrupted` | `source` (`manual` or `speech`), `responseId` |
-| `turn.done` | `responseId`, native provider `usage`, `status` |
-| `close` | `code`, `reason` |
-
-Each event includes its `type`. When the provider event leaves out an ID, `responseId` and `itemId` are empty strings and `contentIndex` is `0`. The `interrupted` event signals local interruption immediately. Its corresponding operation resolves after the provider acknowledges cancellation and any requested truncation. Late audio from a canceled response remains visible as native events and is excluded from convenience audio and built-in playback.
-
-## Browser media
-
-The optional microphone and speaker helpers require a browser with Web Audio support. Microphone capture also requires media permissions and a secure context. Start audio from a user action so the browser can allow playback.
-
-| Member | Behavior |
-| --- | --- |
-| `microphone.start()` | Requests microphone access and streams PCM16, resampled to 24 kHz, in 20 ms chunks through an AudioWorklet. If the page or network falls behind, chunks are dropped rather than stopping capture. |
-| `microphone.mute()` | Disables capture tracks and stops forwarding captured chunks. |
-| `microphone.unmute()` | Enables capture tracks and forwarding. |
-| `microphone.stop()` | Stops capture tracks and releases capture resources. A `start()` still waiting on the permission prompt rejects with `session_closed`. |
-| `speaker.start({ bargeIn: true })` | Starts playback. Barge-in defaults to `true`. With barge-in enabled, provider speech-start events interrupt playback and synchronize played audio with conversation history. Starting the speaker requests no microphone access. |
-| `speaker.stop()` | Clears playback and releases its audio context. |
-
-Both start methods require a connected session. Capture and playback are independent, so custom audio input can use the built-in speaker, and built-in capture can use custom playback.
-
-## Examples
-
-### Have a text conversation
-
-A text conversation uses convenience methods for input and streamed output.
-
-```js
-const conversation = await puter.ai.convo({
-    output: ['text'],
-    instructions: 'Keep replies short.',
-});
-
-conversation.on('text', ({ delta }) => console.log(delta));
-conversation.on('turn.done', () => conversation.close());
-conversation.on('error', (event) => console.error(event));
-
-await conversation.connect();
-conversation.sendText('Say hello in three words.');
-```
-
-### Have a voice conversation
-
-A voice conversation enables the microphone and speaker after connection.
-
-```js
-const conversation = await puter.ai.convo();
-conversation.on('transcript', ({ role, text }) => console.log(role, text));
-conversation.on('error', (event) => console.error(event));
-
-await conversation.connect();
-await conversation.speaker.start();
-await conversation.microphone.start();
-
-// To interrupt a response and submit a replacement text turn:
-await conversation.interrupt();
-conversation.sendText('Actually, make it Wednesday.');
-
-// When the user ends the conversation:
-conversation.close();
-```

src/docs/src/AI/realtime-websocket.md

@@ -0,0 +1,126 @@
+---
+title: Realtime WebSocket API (Beta)
+description: Connect to OpenAI realtime conversations through Puter without puter.js.
+platforms: [websites, apps, nodejs, workers]
+---
+
+The Realtime WebSocket API lets your client send and receive native OpenAI realtime JSON events through Puter.
+
+Puter authenticates the user and charges their account for usage.
+
+**Beta:** This API is incomplete. Its methods, options, events, and behavior may change abruptly, including breaking changes without notice.
+
+For optional text, microphone, playback, and interruption helpers, use [`puter.ai.realtime()`](/AI/realtime/).
+
+## Connection and authentication
+
+Connect to this URL on the Puter API host:
+
+```text
+wss://api.puter.com/puterai/openai/v1/realtime?model=gpt-realtime-2.1-mini
+```
+
+The `model` query parameter defaults to `gpt-realtime-2.1-mini`. The model is fixed for the connection. A session update that changes it is rejected. It also accepts `gpt-realtime-2.1`, `gpt-realtime-2`, `gpt-realtime-1.5`, `gpt-realtime`, and `gpt-realtime-mini`.
+
+Authenticate with a Puter session, app, or full-access API token. Puter supplies the provider credential. Server clients pass this header:
+
+```text
+Authorization: Bearer <PUTER_TOKEN>
+```
+
+Browser WebSocket clients pass `realtime` and `openai-insecure-api-key.<PUTER_TOKEN>` as subprotocols. The server echoes only `realtime`. The endpoint also accepts an `auth_token` query parameter. Prefer the authorization header where supported because URLs and subprotocols can appear in logs. Authentication takes precedence in this order, header, subprotocol, then query parameter.
+
+The socket can open before authentication and provider setup finish. Listen for errors and closure immediately. Wait for `session.created`, send your session settings, then wait for `session.updated` before submitting a turn.
+
+To open a replacement while another connection remains healthy, set the optional `purpose=handover` query parameter. These attempts share a smaller per-credential budget, checked before the general connection-attempt budget. Initial connections and dropped-connection recovery use the URL without this parameter. The limits are listed in [rate limits and quotas](/rate-limits-and-quotas/).
+
+## Text example
+
+Install the `ws` package and set `PUTER_TOKEN` to your Puter token. Run this example as a Node.js ES module:
+
+```js
+import WebSocket from 'ws';
+
+const socket = new WebSocket(
+    'wss://api.puter.com/puterai/openai/v1/realtime?model=gpt-realtime-2.1-mini',
+    { headers: { Authorization: `Bearer ${process.env.PUTER_TOKEN}` } },
+);
+const send = (event) => socket.send(JSON.stringify(event));
+
+socket.on('message', (data) => {
+    const event = JSON.parse(data.toString());
+
+    if (event.type === 'session.created') {
+        send({
+            type: 'session.update',
+            session: {
+                type: 'realtime',
+                output_modalities: ['text'],
+                instructions: 'Keep replies short.',
+            },
+        });
+    }
+    if (event.type === 'session.updated') {
+        send({
+            type: 'conversation.item.create',
+            item: {
+                type: 'message',
+                role: 'user',
+                content: [{ type: 'input_text', text: 'Hello!' }],
+            },
+        });
+        send({ type: 'response.create' });
+    }
+    if (event.type === 'response.output_text.delta')
+        process.stdout.write(event.delta);
+    if (event.type === 'error')
+        console.error(event.error);
+    if (event.type === 'response.done')
+        socket.close();
+});
+socket.on('error', console.error);
+socket.on('close', (code, reason) => console.log(code, reason.toString()));
+```
+
+## Events, audio, and interruption
+
+Each message contains one JSON object with a nonempty string `type` of at most 128 characters. Binary messages are rejected. Native events retain their provider fields. The endpoint doesn't deliver `rate_limits.updated`, which describes Puter's provider account rather than yours, and it removes provider organization IDs from error messages. The endpoint exposes the WebSocket event protocol. Your client manages capture, playback, acknowledgments, and timeouts.
+
+The endpoint refuses `conversation.item.retrieve` events and MCP tools (`"type": "mcp"`) in `session.update` and `response.create` events. A refused event isn't sent to the provider. The session gets an `invalid_event` error and closes with 1008.
+
+PCM16 audio configured at 24 kHz mono is sent as base64 chunks in `input_audio_buffer.append` events. Output arrives in `response.output_audio.delta` events. With manual turn detection, commit buffered input with `input_audio_buffer.commit`, then request a reply with `response.create`.
+
+To interrupt, stop local playback immediately and send `response.cancel` with the active `response_id`. If some generated audio was not played, send `conversation.item.truncate` with its `item_id`, `content_index`, and `audio_end_ms` set to the amount actually played. Wait for `response.done` and any requested `conversation.item.truncated` acknowledgment before submitting a replacement turn.
+
+Connection loss ends the session. Reconnecting creates a new conversation. Tool execution and automatic reconnection belong to your client.
+
+## Server restarts
+
+When Puter knows ahead of time that the server a session runs on will restart, it sends each open session this event, at most once a minute:
+
+```json
+{
+    "type": "puter.reconnect",
+    "event_id": "event_puter_1767225600000",
+    "reason": "server_restarting",
+    "close_in_ms": 90000
+}
+```
+
+The `reason` field is `server_restarting`. The `close_in_ms` field, when present, is about how long until the server closes the session itself. The session keeps working until then. Open a new connection after a short, randomized delay, and once it gets `session.created` and your client is idle, with no response generating and no user turn in progress, move to it and close the old one. A connection that reaches a server that is also restarting gets a `server_restarting` error and closes with 1012. In that case keep the old session and try again after a short, randomized delay. If the server closes the old session first, the close code is 1012, and your client reconnects the same way. A restart can also drop the connection without any notice, so treat an unexpected close the same way.
+
+A new connection starts an empty conversation. To keep context, add earlier items back with `conversation.item.create` before you request a response. Event types that start with `puter.` come from Puter, not the provider.
+
+## Errors, billing, and limits
+
+API errors arrive as `{ "type": "error", "error": { "code": "...", "message": "..." } }`. Handle socket errors and close frames too. Invalid, missing, or revoked credentials use `invalid_api_key`. If Puter can't check a credential or the account's plan, the error is `internal_error` and the session closes with 1011, so try again after a short, randomized delay. Other relay codes include `permission_denied`, `model_not_found`, `insufficient_funds`, `subscription_required`, `rate_limit_exceeded`, `invalid_event`, `connection_timeout`, `provider_unavailable`, `upstream_error`, and `server_restarting`.
+
+Policy failures generally close with 1008, oversized client messages with 1009, upstream or setup failures with 1011, server restarts with 1012, and temporary server overload with 1013. Reconnect after 1011, 1012, or 1013. Authentication and model failures can arrive after the WebSocket handshake succeeds.
+
+The endpoint requires a paid Puter plan. Usage is charged after completed responses and input transcriptions. While a response generates, part of your balance is held for it, so a session opens only if your balance covers one response. If the provider connection closes before a response reports its usage, the response is charged the full amount held for it. Each account can open 20 sessions at once by default. Each JSON message you send is limited to 64 KiB, including reassembled fragments. Messages from the provider can be up to 4 MiB. Stream audio in small chunks. All message rates, queue caps, timeouts, and other bounds are published in [rate limits and quotas](/rate-limits-and-quotas/).
+
+Input transcription holds a small amount of your balance per transcribed item, priced at the transcription model you select. Two items are reserved ahead of your audio and replenished as the provider commits them, so a session holds only its spare items plus the items still awaiting a transcript. Changing the transcription model reprices the spare items. If your balance can't cover another item, audio waits while earlier items settle; once none are left to settle, the session closes with `insufficient_funds`. Failed transcriptions, items without final usage, and audio whose commit the provider never acknowledges before the connection closes are each charged that model's fixed per-item amount. Unacknowledged audio is charged one item per started 160 seconds. With server turn detection, only detected speech counts, so silence is never charged. The amounts, timeout and audio limits are listed in [rate limits and quotas](/rate-limits-and-quotas/).
+
+Provider-reported realtime models must have a known price. A missing model uses the last known model. An explicitly unpriced model closes the connection with `upstream_error`.
+
+After recording usage, the connection can continue while settled credit is positive, including a fractional balance. Each new billable operation still needs its full reservation after accounting for other outstanding holds.

src/docs/src/AI/realtime.md

@@ -1,126 +1,151 @@
 ---
-title: Realtime WebSocket API (Beta)
-description: Connect to OpenAI realtime conversations through Puter without puter.js.
+title: puter.ai.realtime()
+description: Prepare an OpenAI realtime conversation with optional audio and text helpers.
 platforms: [websites, apps, nodejs, workers]
 ---
 
-The Realtime WebSocket API lets your client send and receive native OpenAI realtime JSON events through Puter.
+Prepares a realtime OpenAI conversation that your app can connect to when it's ready.
 
-Puter authenticates the user and charges their account for usage.
+Preparing the conversation verifies Puter sign-in. Realtime conversations require a paid Puter plan, and usage is billed to the signed-in user. Native OpenAI events and convenience methods use the same session. Microphone capture and speaker playback start when your app requests them.
+
+In Node.js, the SDK requires a global `WebSocket`, available in Node.js 22 or later.
 
 **Beta:** This API is incomplete. Its methods, options, events, and behavior may change abruptly, including breaking changes without notice.
 
-For optional text, microphone, playback, and interruption helpers, use [`puter.ai.convo()`](/AI/convo/).
+To connect without puter.js, see the [Realtime WebSocket API](/AI/realtime-websocket/).
 
-## Connection and authentication
-
-Connect to this URL on the Puter API host:
-
-```text
-wss://api.puter.com/puterai/openai/v1/realtime?model=gpt-realtime-2.1-mini
-```
-
-The `model` query parameter defaults to `gpt-realtime-2.1-mini`. The model is fixed for the connection. A session update that changes it is rejected. It also accepts `gpt-realtime-2.1`, `gpt-realtime-2`, `gpt-realtime-1.5`, `gpt-realtime`, and `gpt-realtime-mini`.
-
-Authenticate with a Puter session, app, or full-access API token. Puter supplies the provider credential. Server clients pass this header:
-
-```text
-Authorization: Bearer <PUTER_TOKEN>
-```
-
-Browser WebSocket clients pass `realtime` and `openai-insecure-api-key.<PUTER_TOKEN>` as subprotocols. The server echoes only `realtime`. The endpoint also accepts an `auth_token` query parameter. Prefer the authorization header where supported because URLs and subprotocols can appear in logs. Authentication takes precedence in this order, header, subprotocol, then query parameter.
-
-The socket can open before authentication and provider setup finish. Listen for errors and closure immediately. Wait for `session.created`, send your session settings, then wait for `session.updated` before submitting a turn.
-
-To open a replacement while another connection remains healthy, set the optional `purpose=handover` query parameter. These attempts share a smaller per-credential budget, checked before the general connection-attempt budget. Initial connections and dropped-connection recovery use the URL without this parameter. The limits are listed in [rate limits and quotas](/rate-limits-and-quotas/).
-
-## Text example
-
-Install the `ws` package and set `PUTER_TOKEN` to your Puter token. Run this example as a Node.js ES module:
+## Syntax
 
 ```js
-import WebSocket from 'ws';
+const conversation = await puter.ai.realtime(options = {});
+await conversation.connect();
+```
 
-const socket = new WebSocket(
-    'wss://api.puter.com/puterai/openai/v1/realtime?model=gpt-realtime-2.1-mini',
-    { headers: { Authorization: `Bearer ${process.env.PUTER_TOKEN}` } },
-);
-const send = (event) => socket.send(JSON.stringify(event));
+## Parameters
 
-socket.on('message', (data) => {
-    const event = JSON.parse(data.toString());
+#### `options` (Object) (Optional)
 
-    if (event.type === 'session.created') {
-        send({
-            type: 'session.update',
-            session: {
-                type: 'realtime',
-                output_modalities: ['text'],
-                instructions: 'Keep replies short.',
-            },
-        });
-    }
-    if (event.type === 'session.updated') {
-        send({
-            type: 'conversation.item.create',
-            item: {
-                type: 'message',
-                role: 'user',
-                content: [{ type: 'input_text', text: 'Hello!' }],
-            },
-        });
-        send({ type: 'response.create' });
-    }
-    if (event.type === 'response.output_text.delta')
-        process.stdout.write(event.delta);
-    if (event.type === 'error')
-        console.error(event.error);
-    if (event.type === 'response.done')
-        socket.close();
+The options object accepts these fields.
+
+| Option | Type | Description |
+|--------|------|-------------|
+| `model` | `String` | The realtime model. Defaults to `gpt-realtime-2.1-mini`. Also accepts `gpt-realtime-2.1`, `gpt-realtime-2`, `gpt-realtime-1.5`, `gpt-realtime`, and `gpt-realtime-mini`. |
+| `instructions` | `String` | Session instructions, up to 8192 characters. Defaults to the provider's default. |
+| `output` | `Array<String>` | Accepts `['text']`, `['audio']`, or both. Defaults to `['audio']`. Audio includes a transcript. Requesting both uses audio output and emits transcript deltas through the `text` event. |
+| `voice` | `String` | OpenAI voice identifier, up to 128 characters. Defaults to the provider's default. |
+| `turnDetection` | `Boolean` | Whether the provider detects the end of spoken turns automatically. Defaults to `true`. Set to `false` for manual audio commits. |
+
+## Return value
+
+A Promise resolves to a prepared conversation after options and sign-in are checked. In a browser, preparation opens the Puter sign-in prompt if needed. Preparation opens no realtime connection and requests no microphone access.
+
+The conversation exposes these properties and methods.
+
+| Member | Behavior |
+| --- | --- |
+| `state` | `prepared`, `connecting`, `connected`, or `closed`. |
+| `model` | Resolved model name. |
+| `connect()` | Opens the connection and applies settings. Resolves after the provider acknowledges session configuration. Concurrent calls share one connection Promise, including bounded retries for transient setup failures. Calling it while connected opens no additional connection. During a reconnection, it resolves when the reconnection succeeds and rejects if the conversation closes instead. |
+| `configure(options)` | Updates the given session settings and resolves after the provider acknowledges this update. Settings it isn't given, including ones set through native events, are left alone. Accepts the factory options except `model`. One update can be pending at a time. Provider rules apply to changes such as switching a voice after audio has been generated. |
+| `sendText(text)` | Adds a user message and requests a response. Text must be nonempty and at most 8192 characters. |
+| `appendAudio(pcm)` | Appends an `Int16Array` containing PCM16, 24 kHz mono audio. A chunk contains at most 12000 samples. |
+| `commitAudio()` | Commits buffered audio and requests a response. Requires turn detection to be off, through `turnDetection: false` or a native session update. The provider requires at least 100 ms of buffered audio. |
+| `interrupt(position?)` | Stops built-in playback immediately and waits for cancellation and playback-history synchronization. A reply requested but not yet started is cancelled as soon as it starts. Custom players provide `{ itemId, contentIndex, playedMs }`. An idle call resolves immediately. Readiness failures reject the returned Promise with `session_not_ready` or `session_closed`. |
+| `send(event)` | Sends a native OpenAI JSON event. |
+| `on(type, handler)` | Registers a listener and returns an unsubscribe function. Use `'*'` to receive every native and convenience event. |
+| `off(type, handler)` | Removes a listener. |
+| `close()` | Closes the session, rejects pending operations, and releases SDK-owned capture and playback resources. Repeated calls have no additional effect. |
+
+Attach listeners before connecting to receive initial session events. Closing during connection rejects the pending attempt.
+
+If the connection drops because of network loss, a server restart or a provider error, the conversation reconnects on its own. Normal closes with code 1000 end the conversation. While it reconnects, `state` is `connecting`, a `reconnecting` event fires, and pending operations reject with `connection_lost`. The new session gets every setting the provider acknowledged, from the factory options, `configure()` and native `session.update` events. Then `ready` fires again with `reconnected: true`. Listeners, the microphone and the speaker carry over. Microphone audio captured while reconnecting is dropped.
+
+The conversation also restores recent context. Before `ready` fires, it adds back up to 40 of the most recent conversation items, 48 KiB in total. Restored items are user text, user speech that has a transcript, assistant replies as text, and function calls with their outputs. User speech has a transcript only when input transcription is configured through native session settings. An assistant reply that was interrupted, or still playing when the connection dropped, is cut to what the user saw or heard. The cut is estimated from how much of the audio played, so it can be off by a few words. Images, speech without a transcript, deleted items, and replies still generating when the connection dropped are left out. Restored items get new item IDs, don't fire item events, and count as input tokens on the next response. Item IDs from before the reconnect don't work in later events. Restoration uses a contiguous suffix of restorable items. An oversized newest item leaves that suffix empty. Older items left out to fit the window fire a `context_trimmed` event with the count in `omittedItems`; this is routine and not an error. A provider refusal or a failure to send the restored items fires one `error` event with code `context_restore_failed` per restoration. In both cases the conversation stays connected. Items are preserved without truncating text, function arguments or function results.
+
+When Puter knows ahead of time that the server a session runs on will restart, it sends a `puter.reconnect` event, which also reaches your listeners. Without that notice, a restart drops the connection and the conversation reconnects as above. Within a few seconds, the conversation opens a replacement session in the background while the current one keeps working. It switches once no reply is generating or playing, no user turn is being spoken or transcribed, no function call is waiting for its output, and nothing you sent is waiting on the provider. The switch goes through the same steps as a reconnect, and the `reconnecting` event has code 1012. If the replacement can't open yet, the conversation keeps the current session and tries again later. A refusal that retrying can't fix, such as `insufficient_funds`, stops the attempts. If the server closes the session first, the conversation switches to the replacement, or reconnects if none is open.
+
+The conversation makes up to ten attempts over about a minute. In a browser that reports being offline, it first waits up to two minutes for the connection to come back. If every attempt fails, or the server refuses with an error such as `insufficient_funds`, the conversation closes. A closed conversation cannot reconnect. To start another session, prepare a new one.
+
+Native event definitions are listed in the [OpenAI realtime reference](https://developers.openai.com/api/reference/resources/realtime/server-events). Native events retain their provider fields. Sending native events preserves direct control of the session. Puter refuses `conversation.item.retrieve` events and MCP tools (`"type": "mcp"`) in `session.update` and `response.create` events. Sending one fires an `error` event with code `invalid_event`, and the conversation closes with code 1008. Audio helpers expect the session to use PCM16 at 24 kHz.
+
+## Errors and limits
+
+Failures use objects containing `code` and `message`. Preparation can reject with `invalid_argument`, `model_not_found`, `auth_required`, `auth_canceled`, or `auth_failed`. Connection can reject with `subscription_required` or `insufficient_funds`, which also show the Puter upgrade prompt. Connection and asynchronous operations can reject with provider errors, `session_closed`, `connection_lost`, `operation_timeout`, or `operation_in_progress`. Sending before readiness, or while reconnecting, throws `session_not_ready`.
+
+Invalid input throws `invalid_argument`, `invalid_event`, or `input_too_large`. A full SDK send queue throws `queue_full`. Media helpers report `unsupported_environment`, `permission_denied`, `media_error`, `media_timeout`, or `queue_full`. Runtime media errors and `context_restore_failed` arrive through the `error` event. Provider errors also arrive as native `error` events with a nested `error` object.
+
+Each native event is capped at 64 KiB. The SDK bounds JSON traversal, outgoing queues, acknowledgment waits, microphone work, and playback. All numeric limits are listed on [rate limits and quotas](/rate-limits-and-quotas/).
+
+## Convenience events
+
+The conversation emits these events alongside native events.
+
+| Event | Fields |
+| --- | --- |
+| `ready` | `model`, `provider`, and `reconnected: true` after a reconnection |
+| `reconnecting` | `code`, `reason` of the dropped connection. After a `puter.reconnect` event, `code` is 1012. |
+| `context_trimmed` | `omittedItems`, the number of restorable items left out of the restored context after a reconnection. Fires before `ready`. |
+| `text` | `delta`, `responseId` |
+| `transcript` | `role`, `text`, `final`. Assistant deltas use `final: false`. Completed assistant and user transcripts use `final: true`. User transcripts require transcription to be configured through native session settings. |
+| `audio` | `pcm`, `sampleRate`, `responseId`, `itemId`, `contentIndex`. The PCM chunk is an `Int16Array`. |
+| `interrupted` | `source` (`manual` or `speech`), `responseId` |
+| `turn.done` | `responseId`, native provider `usage`, `status` |
+| `close` | `code`, `reason` |
+
+Each event includes its `type`. When the provider event leaves out an ID, `responseId` and `itemId` are empty strings and `contentIndex` is `0`. The `interrupted` event signals local interruption immediately. Its corresponding operation resolves after the provider acknowledges cancellation and any requested truncation. Late audio from a canceled response remains visible as native events and is excluded from convenience audio and built-in playback.
+
+## Browser media
+
+The optional microphone and speaker helpers require a browser with Web Audio support. Microphone capture also requires media permissions and a secure context. Start audio from a user action so the browser can allow playback.
+
+| Member | Behavior |
+| --- | --- |
+| `microphone.start()` | Requests microphone access and streams PCM16, resampled to 24 kHz, in 20 ms chunks through an AudioWorklet. If the page or network falls behind, chunks are dropped rather than stopping capture. |
+| `microphone.mute()` | Disables capture tracks and stops forwarding captured chunks. |
+| `microphone.unmute()` | Enables capture tracks and forwarding. |
+| `microphone.stop()` | Stops capture tracks and releases capture resources. A `start()` still waiting on the permission prompt rejects with `session_closed`. |
+| `speaker.start({ bargeIn: true })` | Starts playback. Barge-in defaults to `true`. With barge-in enabled, provider speech-start events interrupt playback and synchronize played audio with conversation history. Starting the speaker requests no microphone access. |
+| `speaker.stop()` | Clears playback and releases its audio context. |
+
+Both start methods require a connected session. Capture and playback are independent, so custom audio input can use the built-in speaker, and built-in capture can use custom playback.
+
+## Examples
+
+### Have a text conversation
+
+A text conversation uses convenience methods for input and streamed output.
+
+```js
+const conversation = await puter.ai.realtime({
+    output: ['text'],
+    instructions: 'Keep replies short.',
 });
-socket.on('error', console.error);
-socket.on('close', (code, reason) => console.log(code, reason.toString()));
+
+conversation.on('text', ({ delta }) => console.log(delta));
+conversation.on('turn.done', () => conversation.close());
+conversation.on('error', (event) => console.error(event));
+
+await conversation.connect();
+conversation.sendText('Say hello in three words.');
 ```
 
-## Events, audio, and interruption
+### Have a voice conversation
 
-Each message contains one JSON object with a nonempty string `type` of at most 128 characters. Binary messages are rejected. Native events retain their provider fields. The endpoint doesn't deliver `rate_limits.updated`, which describes Puter's provider account rather than yours, and it removes provider organization IDs from error messages. The endpoint exposes the WebSocket event protocol. Your client manages capture, playback, acknowledgments, and timeouts.
+A voice conversation enables the microphone and speaker after connection.
 
-The endpoint refuses `conversation.item.retrieve` events and MCP tools (`"type": "mcp"`) in `session.update` and `response.create` events. A refused event isn't sent to the provider. The session gets an `invalid_event` error and closes with 1008.
+```js
+const conversation = await puter.ai.realtime();
+conversation.on('transcript', ({ role, text }) => console.log(role, text));
+conversation.on('error', (event) => console.error(event));
 
-PCM16 audio configured at 24 kHz mono is sent as base64 chunks in `input_audio_buffer.append` events. Output arrives in `response.output_audio.delta` events. With manual turn detection, commit buffered input with `input_audio_buffer.commit`, then request a reply with `response.create`.
+await conversation.connect();
+await conversation.speaker.start();
+await conversation.microphone.start();
 
-To interrupt, stop local playback immediately and send `response.cancel` with the active `response_id`. If some generated audio was not played, send `conversation.item.truncate` with its `item_id`, `content_index`, and `audio_end_ms` set to the amount actually played. Wait for `response.done` and any requested `conversation.item.truncated` acknowledgment before submitting a replacement turn.
+// To interrupt a response and submit a replacement text turn:
+await conversation.interrupt();
+conversation.sendText('Actually, make it Wednesday.');
 
-Connection loss ends the session. Reconnecting creates a new conversation. Tool execution and automatic reconnection belong to your client.
-
-## Server restarts
-
-When Puter knows ahead of time that the server a session runs on will restart, it sends each open session this event, at most once a minute:
-
-```json
-{
-    "type": "puter.reconnect",
-    "event_id": "event_puter_1767225600000",
-    "reason": "server_restarting",
-    "close_in_ms": 90000
-}
+// When the user ends the conversation:
+conversation.close();
 ```
-
-The `reason` field is `server_restarting`. The `close_in_ms` field, when present, is about how long until the server closes the session itself. The session keeps working until then. Open a new connection after a short, randomized delay, and once it gets `session.created` and your client is idle, with no response generating and no user turn in progress, move to it and close the old one. A connection that reaches a server that is also restarting gets a `server_restarting` error and closes with 1012. In that case keep the old session and try again after a short, randomized delay. If the server closes the old session first, the close code is 1012, and your client reconnects the same way. A restart can also drop the connection without any notice, so treat an unexpected close the same way.
-
-A new connection starts an empty conversation. To keep context, add earlier items back with `conversation.item.create` before you request a response. Event types that start with `puter.` come from Puter, not the provider.
-
-## Errors, billing, and limits
-
-API errors arrive as `{ "type": "error", "error": { "code": "...", "message": "..." } }`. Handle socket errors and close frames too. Invalid, missing, or revoked credentials use `invalid_api_key`. If Puter can't check a credential or the account's plan, the error is `internal_error` and the session closes with 1011, so try again after a short, randomized delay. Other relay codes include `permission_denied`, `model_not_found`, `insufficient_funds`, `subscription_required`, `rate_limit_exceeded`, `invalid_event`, `connection_timeout`, `provider_unavailable`, `upstream_error`, and `server_restarting`.
-
-Policy failures generally close with 1008, oversized client messages with 1009, upstream or setup failures with 1011, server restarts with 1012, and temporary server overload with 1013. Reconnect after 1011, 1012, or 1013. Authentication and model failures can arrive after the WebSocket handshake succeeds.
-
-The endpoint requires a paid Puter plan. Usage is charged after completed responses and input transcriptions. While a response generates, part of your balance is held for it, so a session opens only if your balance covers one response. If the provider connection closes before a response reports its usage, the response is charged the full amount held for it. Each account can open 20 sessions at once by default. Each JSON message you send is limited to 64 KiB, including reassembled fragments. Messages from the provider can be up to 4 MiB. Stream audio in small chunks. All message rates, queue caps, timeouts, and other bounds are published in [rate limits and quotas](/rate-limits-and-quotas/).
-
-Input transcription holds a small amount of your balance per transcribed item, priced at the transcription model you select. Two items are reserved ahead of your audio and replenished as the provider commits them, so a session holds only its spare items plus the items still awaiting a transcript. Changing the transcription model reprices the spare items. If your balance can't cover another item, audio waits while earlier items settle; once none are left to settle, the session closes with `insufficient_funds`. Failed transcriptions, items without final usage, and audio whose commit the provider never acknowledges before the connection closes are each charged that model's fixed per-item amount. Unacknowledged audio is charged one item per started 160 seconds. With server turn detection, only detected speech counts, so silence is never charged. The amounts, timeout and audio limits are listed in [rate limits and quotas](/rate-limits-and-quotas/).
-
-Provider-reported realtime models must have a known price. A missing model uses the last known model. An explicitly unpriced model closes the connection with `upstream_error`.
-
-After recording usage, the connection can continue while settled credit is positive, including a fractional balance. Each new billable operation still needs its full reservation after accounting for other outstanding holds.

src/docs/src/Auth/signIn.md

@@ -6,7 +6,7 @@ platforms: [websites]
 
 Initiates the sign in process for the user.
 
-This will open a popup window with the appropriate authentication method. Puter automatically handles the authentication process and will resolve the promise when the user has signed in.
+This will open a popup window with the appropriate authentication method. Puter automatically handles the authentication process and will resolve the promise when the user has signed in. Puter's own apps on puter.com subdomains show the sign-in in a frame over the page instead, and open the popup when the user picks a sign-in provider such as Google.
 
 It is important to note that all essential methods in Puter handle authentication automatically. This method is only necessary if you want to handle authentication manually, for example if you want to build your own custom authentication flow.
 

src/docs/src/Hosting/attach.md

@@ -12,7 +12,7 @@ When the user adds a new domain, the dialog shows the DNS records to set at thei
 
 Your app doesn't need any permission to call this method. The user makes the choice in the dialog, and your app only receives the domain they chose. Custom domains are billed to the user's Puter account, and the dialog shows the price before the user adds one.
 
-Inside Puter, the dialog is rendered by the desktop environment. On a website, a puter.com popup hosts the dialog and signs the user in first if needed.
+Inside Puter, the dialog is rendered by the desktop environment. On a website, a puter.com popup hosts the dialog and signs the user in first if needed. Puter's own apps on puter.com subdomains show it in a frame over the page instead, when the user is signed in to both.
 
 ## Syntax
 

src/docs/src/Objects/chatresponse.md

@@ -64,7 +64,7 @@ Extra detail on why generation stopped (e.g. a refusal classification). Provider
 
 #### `usageDetails` (Object)
 
-A provider-neutral usage breakdown, present alongside `usage` on every result. `inputTokens` excludes cache reads/writes; `outputTokens` includes reasoning tokens. Fields: `inputTokens`, `outputTokens`, `cacheReadTokens`, `cacheWrite5mTokens`, `cacheWrite30mTokens`, `cacheWrite1hTokens`, `cacheWrite12hTokens`, `reasoningTokens`, `webSearchRequests`, `webFetchRequests`, `speed` (`"fast"` or `"standard"`), `serviceTier`, and `iterations` (a per-pass breakdown when the upstream ran several passes, e.g. compaction or an advisor sub-call). Each `cacheWrite*Tokens` field counts the tokens written to the cache with that lifetime: Claude and Moonshot write 5 minute and 1 hour entries, OpenAI writes 30 minute and 12 hour entries.
+A provider-neutral usage breakdown, present alongside `usage` on every result. `inputTokens` excludes cache reads/writes; `outputTokens` includes reasoning tokens. Fields: `inputTokens`, `outputTokens`, `cacheReadTokens`, `cacheWrite5mTokens`, `cacheWrite30mTokens`, `cacheWrite1hTokens`, `cacheWrite12hTokens`, `reasoningTokens`, `webSearchRequests`, `webFetchRequests`, `speed` (`"fast"` or `"standard"`), `serviceTier`, `serviceSpeed` (`"standard"`, `"fast"`, or `"ultrafast"` when requested), and `iterations` (a per-pass breakdown when the upstream ran several passes, e.g. compaction or an advisor sub-call). Each `cacheWrite*Tokens` field counts the tokens written to the cache with that lifetime: Claude and Moonshot write 5 minute and 1 hour entries, OpenAI writes 30 minute and 12 hour entries.
 
 Prefer `usageDetails` over `usage` for anything beyond the raw billing totals. Claude's `usage.thinking_tokens` is informational: it's included in `output_tokens` and not billed separately — `usageDetails.reasoningTokens` carries the same count.
 

src/docs/src/Objects/chatresponsechunk.md

@@ -102,7 +102,7 @@ Extra detail on why generation stopped (e.g. a refusal classification). Provider
 
 #### `usageDetails` (Object)
 
-A provider-neutral usage breakdown, present on the final `usage` chunk. `inputTokens` excludes cache reads/writes; `outputTokens` includes reasoning tokens. Fields: `inputTokens`, `outputTokens`, `cacheReadTokens`, `cacheWrite5mTokens`, `cacheWrite30mTokens`, `cacheWrite1hTokens`, `cacheWrite12hTokens`, `reasoningTokens`, `webSearchRequests`, `webFetchRequests`, `speed` (`"fast"` or `"standard"`), `serviceTier`, and `iterations` (a per-pass breakdown when the upstream ran several passes, e.g. compaction or an advisor sub-call). Each `cacheWrite*Tokens` field counts the tokens written to the cache with that lifetime: Claude and Moonshot write 5 minute and 1 hour entries, OpenAI writes 30 minute and 12 hour entries.
+A provider-neutral usage breakdown, present on the final `usage` chunk. `inputTokens` excludes cache reads/writes; `outputTokens` includes reasoning tokens. Fields: `inputTokens`, `outputTokens`, `cacheReadTokens`, `cacheWrite5mTokens`, `cacheWrite30mTokens`, `cacheWrite1hTokens`, `cacheWrite12hTokens`, `reasoningTokens`, `webSearchRequests`, `webFetchRequests`, `speed` (`"fast"` or `"standard"`), `serviceTier`, `serviceSpeed` (`"standard"`, `"fast"`, or `"ultrafast"` when requested), and `iterations` (a per-pass breakdown when the upstream ran several passes, e.g. compaction or an advisor sub-call). Each `cacheWrite*Tokens` field counts the tokens written to the cache with that lifetime: Claude and Moonshot write 5 minute and 1 hour entries, OpenAI writes 30 minute and 12 hour entries.
 
 #### `message` (String)
 

src/docs/src/UI/showFeedbackDialog.md

@@ -8,7 +8,7 @@ Opens a dialog where the user can send you, the app's developer, feedback about
 
 The message is delivered by Puter: it is stored and emailed to the email address on your Puter account. The feedback never passes through your app's code, and the dialog tells the user what is shared with you: their username, plus their email address (as the email's reply-to, so you can respond) when their email is verified.
 
-Inside Puter, the dialog is rendered by the desktop environment. On a website, a puter.com popup hosts the dialog (signing the user in first if needed).
+Inside Puter, the dialog is rendered by the desktop environment. On a website, a puter.com popup hosts the dialog (signing the user in first if needed). Puter's own apps on puter.com subdomains show it in a frame over the page instead, when the user is signed in to both.
 
 **Feedback is opt-in.** Users can only send feedback if it's enabled for your app. Apps created in the Dev Center have feedback enabled at creation — you can turn it off with the "User Feedback" toggle in the app's settings. Apps created through `puter.apps.create` default to off; enable it by setting `feedbackEnabled`:
 

src/docs/src/examples.js

@@ -49,6 +49,12 @@ const examples = [
                 slug: 'ai-chatgpt',
                 source: '/playground/examples/ai-chatgpt.html',
             },
+            {
+                title: 'Select Service Speed',
+                description: 'Request faster AI service and inspect the speed that served the response.',
+                slug: 'ai-chat-service-speed',
+                source: '/playground/examples/ai-chat-service-speed.html',
+            },
             {
                 title: 'Image Analysis',
                 description: 'Analyze images with AI using Puter.js GPT Vision API. Run and modify this code example instantly in your browser.',

src/docs/src/playground/examples/ai-chat-service-speed.html

@@ -0,0 +1,13 @@
+<html>
+<body>
+    <script src="https://js.puter.com/v2/"></script>
+    <script type="module">
+        const response = await puter.ai.chat('Explain closures in one paragraph.', {
+            model: 'gpt-6.1-sol',
+            serviceSpeed: 'ultrafast',
+        });
+        puter.print(response.message.content);
+        puter.print(response.usageDetails.serviceSpeed);
+    </script>
+</body>
+</html>

src/docs/src/rate-limits-and-quotas.md

@@ -91,7 +91,7 @@ Gemini speech-to-text first counts the audio's tokens, which times out after 30
 
 The OpenAI- and Anthropic-compatible endpoints (`/puterai/openai/v1/*`, `/puterai/anthropic/v1/messages`) require a paid plan; a free account gets `402 subscription_required`. The same models are available to every account through `puter.ai.*` and `/drivers/call`, and the model catalogue endpoints are open to everyone.
 
-Realtime sessions ([`puter.ai.convo()`](/AI/convo/), `/puterai/openai/v1/realtime`) require a paid plan, like the other OpenAI-compatible endpoints, and are capped at 20 open at once per account. An account without a plan gets an `error` event with code `subscription_required`. A session over the cap gets `rate_limit_exceeded`. Both close with 1008.
+Realtime sessions ([`puter.ai.realtime()`](/AI/realtime/), `/puterai/openai/v1/realtime`) require a paid plan, like the other OpenAI-compatible endpoints, and are capped at 20 open at once per account. An account without a plan gets an `error` event with code `subscription_required`. A session over the cap gets `rate_limit_exceeded`. Both close with 1008.
 
 | Realtime transport limit | Value | Behavior at the limit |
 | --- | --- | --- |
@@ -131,7 +131,7 @@ Realtime sessions ([`puter.ai.convo()`](/AI/convo/), `/puterai/openai/v1/realtim
 
 Send audio as a stream of small chunks. The message cap applies to text, configuration, tool output, and audio alike.
 
-The [`puter.ai.convo()`](/AI/convo/) SDK also applies these limits before sending or retaining data.
+The [`puter.ai.realtime()`](/AI/realtime/) SDK also applies these limits before sending or retaining data.
 
 | Conversation SDK limit | Value | Behavior at the limit |
 | --- | --- | --- |
@@ -159,7 +159,7 @@ The [`puter.ai.convo()`](/AI/convo/) SDK also applies these limits before sendin
 
 `/puterai/anthropic/v1/messages/count_tokens` has its own budget of 120 requests per minute per user, with no concurrency limit, and is not charged.
 
-Chat requests accept JSON bodies up to 128 MiB (134,217,728 bytes), including `text/plain;actually=json` bodies. That covers `puter-chat-completion` calls through `/drivers/call`, which `puter.ai.chat()` uses, and POSTs to `/puterai/openai/v1/*`, `/puterai/anthropic/v1/messages` and `/puterai/anthropic/v1/messages/count_tokens`. A body over 50 MiB needs a `Content-Length` header and a token in the `Authorization` or `x-api-key` header (or a session cookie); an `auth_token` inside the body isn't read until the body is, so it only works up to 50 MiB. Larger bodies, and bodies over 50 MiB for other drivers, return HTTP 413. One account can have up to 128 MiB of these large bodies in flight at once; past that, requests get `429` with `Retry-After`. A server handling too many large bodies at once returns `503` with `Retry-After`. Retry either after the given number of seconds. Other routes accept JSON bodies up to 50 MiB.
+Chat requests accept JSON bodies up to 128 MiB (134,217,728 bytes), including `text/plain;actually=json` bodies. That covers `puter-chat-completion` calls through `/drivers/call`, which `puter.ai.chat()` uses, and POSTs to `/puterai/openai/v1/*`, `/puterai/anthropic/v1/messages` and `/puterai/anthropic/v1/messages/count_tokens`. A body over 50 MiB needs a `Content-Length` header and a credential that can be read before the rest of the body. That is a token in the `Authorization` or `x-api-key` header, a session cookie, or an `auth_token` that is the first field of the body, which is where Puter.js sends it. An `auth_token` after other fields isn't read until the body is, so it only works up to 50 MiB. Larger bodies, and bodies over 50 MiB for other drivers, return HTTP 413. One account can have up to 128 MiB of these large bodies in flight at once; past that, requests get `429` with `Retry-After`. A server handling too many large bodies at once returns `503` with `Retry-After`. Retry either after the given number of seconds. Other routes accept JSON bodies up to 50 MiB. A JSON body in a charset other than UTF-8 is accepted up to 1 MiB. A JSON body over 32 KiB can nest arrays and objects up to 10,000 levels deep, and deeper nesting returns HTTP 400.
 
 Claude's server-executed tools are clamped so a request can't reserve an unbounded amount of credit:
 
@@ -226,6 +226,11 @@ Image and video inputs to [`txt2img()`](/AI/txt2img/) and [`txt2vid()`](/AI/txt2
 
 puter.js uploads a `Blob`, a `File` or a `data:` URI over 1 MiB to a temporary Puter file for the call. It counts against the user's storage until puter.js deletes it after the call.
 
+
+### Chat service speed catalogs
+
+OpenRouter speed endpoint catalogs are cached for 15 minutes, with at most 256 model entries per backend process. A catalog fetch times out after 10 seconds and accepts at most 2 MiB. Failed fetches are remembered for 30 seconds. If no usable speed offering is available, the request uses standard service.
+
 ### Remote chat image inlining
 
 For chat providers that require remote image URLs to be inlined (currently Moonshot):