Merge branch 'feat/app-user-feedback' of https://github.com/HeyPuter/puter into feat/app-user-feedback

Add `requireVerified: true` to both app feedback routes (`GET /target` and `POST /`) in `AppFeedbackController`. This tightens access control so only verified user accounts can fetch feedback targets or submit app feedback.

.github/workflows/backend-tests.yaml

@@ -13,6 +13,9 @@ on:
       - 'tools/**'
       - 'package.json'
       - 'package-lock.json'
+      # The type-check job below reads these.
+      - 'tsconfig.json'
+      - 'tsconfig.build.json'
       - '.github/workflows/backend-tests.yaml'
 
 permissions:
@@ -20,6 +23,35 @@ permissions:
   pull-requests: write
 
 jobs:
+  # `tsconfig.build.json` builds with `noCheck: true`, so nothing else in CI
+  # runs the type checker and a missing export compiles to `undefined`,
+  # surfacing only when the call is finally reached at runtime. This diffs
+  # against tools/typecheck-baseline.json and fails only on *new* errors.
+  #
+  # The consuming repo runs an equivalent gate, but only once someone bumps the
+  # submodule pointer — too late to keep the error off this repo's default
+  # branch. Hence a gate here, on this repo's own pull requests.
+  #
+  # Its own job rather than a step in `test`: that one runs a base/PR matrix for
+  # coverage comparison, and the check only needs the PR ref, once.
+  typecheck:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+
+      - name: Set up Node.js
+        uses: actions/setup-node@v4
+        with:
+          node-version: '24'
+          cache: 'npm'
+
+      - name: Install dependencies
+        run: npm ci
+
+      - name: Type check (new errors only)
+        run: npm run typecheck
+
   test:
     runs-on: ubuntu-latest
     strategy:

.github/workflows/puterjs-tests.yaml

@@ -23,6 +23,33 @@ permissions:
   pull-requests: write
 
 jobs:
+  # The JSDoc in src/puter-js/src is the source of truth for the SDK's public
+  # types. The declarations shipped to npm are generated from it at build time
+  # and never committed, so what needs guarding is the JSDoc itself: this
+  # generates the declarations and type-checks the published surface without
+  # skipLibCheck, which is how broken re-exports used to go unnoticed.
+  #
+  # Its own job rather than a step in `test`: that one builds bundles and
+  # installs a browser, and this needs neither.
+  types:
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Set up Node.js
+        uses: actions/setup-node@v4
+        with:
+          node-version: '24'
+          cache: 'npm'
+
+      - name: Install dependencies
+        run: npm ci
+
+      - name: Check the puter.js JSDoc produces declarations that type-check
+        run: npm run check:puterjs:types
+
   test:
     runs-on: ubuntu-latest
 

AGENTS.md

@@ -98,24 +98,23 @@ Follow the same layered structure inside an extension — unless it only needs a
 
 [src/puter-js/](src/puter-js/) is the public SDK. It ships live from `https://js.puter.com/v2/` with no version pinning — every existing app picks up changes immediately. Treat every observable behavior (signatures, response fields, error codes) as something a production app depends on.
 
-Layout: SDK modules in [src/puter-js/src/modules/](src/puter-js/src/modules/) (one file or directory per area — `FileSystem/`, `KV.js`, `ai/`, …), shared plumbing in [src/puter-js/src/lib/](src/puter-js/src/lib/), hand-maintained type declarations in [src/puter-js/types/](src/puter-js/types/), API tests in [src/puter-js/tests/api/](src/puter-js/tests/api/), UI e2e tests in [src/puter-js/tests/e2e/](src/puter-js/tests/e2e/), developer docs in [src/docs/](src/docs/).
+Layout: SDK modules in [src/puter-js/src/modules/](src/puter-js/src/modules/) (one file or directory per area — `FileSystem/`, `kv/`, `ai/`, …), shared plumbing in [src/puter-js/src/lib/](src/puter-js/src/lib/), generated (gitignored) type declarations in `src/puter-js/types/`, API tests in [src/puter-js/tests/api/](src/puter-js/tests/api/), UI e2e tests in [src/puter-js/tests/e2e/](src/puter-js/tests/e2e/), developer docs in [src/docs/](src/docs/).
 
-Language & typing: puter.js source is plain JavaScript — never TypeScript files — typed via JSDoc. Reference the hand-maintained declarations in [src/puter-js/types/](src/puter-js/types/) with `import(...)` types rather than re-declaring shapes:
+Language & typing: puter.js source is plain JavaScript — never TypeScript files — typed via JSDoc. **The JSDoc is the source of truth for the SDK's types.** `src/puter-js/types/` is `tsc --emitDeclarationOnly` output: the SDK build (`npm run build` in `src/puter-js`) generates it, the npm tarball ships it so TypeScript consumers get declarations, and git ignores it — it is never committed and never edited by hand. `npm run check:puterjs:types` generates it in CI and type-checks the result without `skipLibCheck`. The one hand-written declaration file is [src/puter-js/index.d.ts](src/puter-js/index.d.ts), which decides what is public and re-exports the generated names.
+
+Declare a shape where it belongs, and reference it with an `import(...)` type from elsewhere. A shape more than one file needs lives in the module's `types.js`; a shape shared across modules lives in [src/puter-js/src/lib/types.js](src/puter-js/src/lib/types.js); a shape with one consumer can stay next to it:
 
 ```js
-/** @typedef {import('../../types/modules/ai').ChatOptions} ChatOptions */
-
-/** @type {ChatOptions} */
-const options = { model: 'gpt-5-nano' };
+/** @typedef {import('./types.js').KVOptConfig} KVOptConfig */
 ```
 
-When a shape only exists locally, declare it as an inline object literal — not the `@typedef {Object}` + `@property` list form — and use `unknown` over `*`:
+Use `@typedef {Object}` + `@property` for any shape whose fields need documenting — it is the only JSDoc form that carries a doc comment per field into the generated declaration. Keep the inline object-literal form for small internal shapes with nothing to say about each field, and prefer `unknown` over `*`:
 
 ```js
 /** @typedef {{ key: string, value: unknown }} KVEntry */
 ```
 
-Public (exposed) methods must carry JSDoc types — parameters and return value, matching the `.d.ts` declarations exactly. Unexposed/private helpers are typed at the contributor's discretion: annotate where it helps the next reader, and either way keep them clean.
+Public (exposed) methods must carry JSDoc types — parameters and return value — with one `@overload` block per accepted call form, since those overloads *are* the published signature. Unexposed/private helpers are typed at the contributor's discretion: annotate where it helps the next reader, and either way keep them clean. Members tagged `@internal` are stripped from the generated declarations, so use that tag rather than `@private` to keep something off the public surface.
 
 Typing in JS files is encouraged: annotate with JSDoc `@type`/`@param`/`@returns` using the TypeScript type system, and define shared shapes with `@typedef`. API types must not be `unknown` or untyped `...args` — spell out the real parameter and return shapes; the only exception is values passed through transparently to an upstream layer that owns their type. For example:
 
@@ -133,7 +132,7 @@ Every SDK change carries all five of the following — a puter.js PR missing one
 1. **Backward compatibility.** Mandatory unless a maintainer explicitly signs off on a break. Existing call signatures keep working (including both positional and options-object forms where a method supports them); new parameters are optional with defaults that preserve old behavior; never rename or repurpose existing params, response fields, or error codes. New parameter names are `camelCase` (existing `snake_case` stays for compatibility). Say in the PR how existing callers are unaffected.
 2. **Tests.** Add or extend a suite in [tests/api/suites/](src/puter-js/tests/api/suites/) (`<area>.suite.ts`; register new suites in `suites/index.ts` — no globbing). One suite runs unchanged on node, browser, and workerd via `npm run test:puterjs`; never write a per-platform test. The runners execute the **built** bundle — run `npm run build:workerLib` after SDK changes or the suite silently tests stale code. For `puter.ui.*` methods rendered by the desktop, use the Playwright e2e harness instead — see [src/puter-js/TESTING.md](src/puter-js/TESTING.md).
 3. **Docs.** New or changed APIs update [src/docs/src/](src/docs/src/): the method page (`<Area>/<method>.md`, with frontmatter and a runnable example) and the area overview when the surface changes. Docs are the contract users code against — signatures, defaults, and return shapes must match the implementation exactly.
-4. **Types.** Update [src/puter-js/types/modules/](src/puter-js/types/modules/)`<module>.d.ts` and re-export new types through `index.d.ts` / `types/puter.d.ts`. Declarations must match runtime behavior exactly — a wrong type is worse than a missing one.
+4. **Types.** Type the change in JSDoc on the implementation, then run `npm run check:puterjs:types` to confirm it still produces declarations that type-check. Name any new type in [src/puter-js/index.d.ts](src/puter-js/index.d.ts) if consumers should be able to import it. Declarations must match runtime behavior exactly — a wrong type is worse than a missing one.
 5. **Error handling.** Reject/throw `{ message, code }` objects with stable `snake_case` codes, matching the existing modules (see `KV.js`). Validate cheap preconditions client-side before making the network call; pass backend errors through unchanged rather than swallowing or re-wrapping them. Error codes are API surface — changing one is a breaking change.
 
 [doc/contributing-apis.md](doc/contributing-apis.md) walks the full lifecycle of adding an API across backend + SDK.

config.template.jsonc

@@ -437,6 +437,24 @@
     ],
 
     // ── Metering ────────────────────────────────────────────────────────
-    // When true, all metering checks pass — no per-actor limits enforced.
-    "unlimitedMetering": false
+    // When true, every account resolves to an unlimited policy: usage is still
+    // recorded, but nothing is ever refused for lack of budget. This is the
+    // setting for a deployment with no way to buy more — without it, accounts
+    // are held to the free monthly allowance and start getting 402s from the AI
+    // surfaces, file transfers and KV once they pass it.
+    "unlimitedMetering": false,
+
+    // Whether an account that has spent its whole allowance is refused the
+    // operations that spend it — file transfers, KV calls. Recording is
+    // unaffected either way. `workers` extends the same refusal to
+    // worker-driven calls, which are exempt by default because a deployed
+    // worker has nowhere to surface a payment prompt.
+    // "meteringEnforcement": { "enabled": true, "workers": false },
+
+    // Fleet-wide spend rate, in micro-cents per minute, past which metering
+    // raises the `metering:excessiveGlobalUsageRate` alarm. Omit it (the
+    // default) to leave the check off: the only useful value is a multiple of
+    // what this deployment's normal traffic costs, so it has to be measured
+    // rather than guessed, and a stale number here alarms on healthy growth.
+    // "maxGlobalUsagePerMinute": 200000000
 }

doc/alarms.md

@@ -6,9 +6,9 @@ transports by **severity**.
 
 ```ts
 this.clients.alarm.create(
-    `driver_rate_limit_hit:${iface}:${method}`, // de-dupe key
-    `Driver rate limit hit on ${iface}:${method}`, // what a human reads
-    { iface, method, userUuid }, // context fields
+    `metering_write_failed:${userUuid}`, // de-dupe key
+    'Metering write failed', // what a human reads
+    { userUuid, appId, error }, // context fields
     'info', // severity
 );
 ```
@@ -37,8 +37,14 @@ pass one explicitly unless you really mean "page someone".
 
 - Did the server fail to do its job in a way nobody expected? → `critical`
 - Is a background job, rate, or dependency degraded? → `warning`
-- Is this a user doing something notable (hitting a limit, tripping an abuse
-  heuristic, overspending)? → `info`
+- Is this a user doing something notable (tripping an abuse heuristic,
+  overspending)? → `info`
+- Did one of *our own* limits reject a caller — a rate limit, a concurrency
+  cap, a quota? → don't alarm at all. The limit doing its job is not an
+  event; the 429 is the whole signal, and alarming on it only produces noise
+  proportional to traffic. An *upstream provider* rate-limiting us is the
+  opposite case and still alarms (`upstream_rate_limited`, `info`) — that one
+  is not something we chose.
 
 An extension whose signals are all one tier can default its own local
 `raiseAlarm` helper to that tier instead of repeating it at every call site —

doc/contributing-apis.md

@@ -54,6 +54,7 @@ Both are supported ways to define an API. **Prefer a controller when you need fi
 
 - **Controller routes** (and extension routes — same options) take [`RouteOptions`](../src/backend/core/http/types.ts): auth gates (`requireAuth`, `requireUserActor`, `noUserSession`, `adminOnly`, `allowedAppIds`, and the access-token controls), `subdomain` routing, per-route `rateLimit`, body parsers, and arbitrary extra `middleware`. The auth flavors are subtle and default-deny — read the JSDoc on each field before picking.
 - **Driver methods** get their policies from the `@Driver` options: per-method `rateLimit` (limit/window/backend), `concurrent` in-flight caps (optionally `bySubscription`), and `noUserSession`. The `/drivers/call` surface enforces them.
+- **An endpoint that spends metered resources** on the caller's behalf — moving file content, making object-store requests, anything else the account is billed for — also declares `requireCredits: true`, which turns an account with nothing left of its budget away with a 402 before the handler runs. Endpoints that only describe or delete things deliberately don't: an account that has run out still has to be able to see what it has, clear it, and reach its billing pages. Drivers have no route options to declare this on, so they call `assertActorHasCredits` themselves ([src/backend/services/metering/enforcement.ts](../src/backend/services/metering/enforcement.ts)) — see `KVStoreDriver`, which does it once for every method.
 
 ### 3. puter.js
 
@@ -62,7 +63,10 @@ Both are supported ways to define an API. **Prefer a controller when you need fi
 
 ### 4. Types
 
-- Add or extend the declaration in [src/puter-js/types/modules/](../src/puter-js/types/modules/)`<module>.d.ts` and re-export new types through `index.d.ts`. Declarations must match the runtime exactly — optionality, defaults, and return types included.
+- Type the method where you wrote it, in JSDoc: `@param`/`@returns` on the implementation, one `@overload` block per accepted call form, and `@typedef {Object}` + `@property` for any new shape. The JSDoc is the source of truth — declarations are generated from it, so there is nothing to keep in sync by hand.
+- Put a shape more than one file needs in the module's `types.js` (e.g. [src/puter-js/src/modules/kv/types.js](../src/puter-js/src/modules/kv/types.js)); anything shared across modules goes in [src/puter-js/src/lib/types.js](../src/puter-js/src/lib/types.js). A shape with one consumer can stay next to it.
+- Run `npm run check:puterjs:types` — it generates the declarations and type-checks the published surface without `skipLibCheck`. **Never edit anything under `src/puter-js/types/`**: it is gitignored build output, produced by the SDK build and shipped in the npm tarball, and the next build overwrites it.
+- Name the new type in [src/puter-js/index.d.ts](../src/puter-js/index.d.ts) if consumers should be able to import it. That file is the one hand-written declaration in the package: it decides what is public and re-exports nothing else.
 
 ### 5. Docs
 

doc/self-hosting.md

@@ -370,6 +370,37 @@ Default is 100 MB per user.
 
 Set `is_storage_limited: false` for unlimited (bounded by host disk).
 
+### Usage metering and budgets
+
+Puter meters what an account costs to serve — bytes sent back, object-store
+requests, KV capacity, AI tokens — against a monthly budget, and refuses the
+operations that spend it once that budget is gone. The refusal is a `402` with
+code `insufficient_funds`; reads that only describe things, and every kind of
+deletion, stay available so an account can always see what it has and clear it.
+
+On a self-hosted install this is almost certainly not what you want. There is
+nowhere to buy more, so accounts are held to the free monthly allowance
+(US$0.25 of measured cost — roughly 2 GiB of downloads) and start being refused
+after that. Turn it off:
+
+```json
+"unlimitedMetering": true
+```
+
+Every account then resolves to an unlimited policy. Usage is still recorded, so
+the dashboard still shows what is being consumed; nothing is ever refused for
+lack of budget.
+
+To keep the budgets but stop them blocking anything — recording only:
+
+```json
+"meteringEnforcement": { "enabled": false }
+```
+
+Calls driven by a deployed worker are exempt from enforcement by default,
+because a worker has no prompt to show and nobody watching it fail. Set
+`"meteringEnforcement": { "workers": true }` to include them.
+
 ### Captcha on signup / login
 
 Built-in proof-of-work captcha — no external service needed.

extensions/appTelemetry.ts

@@ -2,6 +2,14 @@ import { Context } from '@heyputer/backend/src/core';
 import type { Actor } from '@heyputer/backend/src/core/actor';
 import { HttpError } from '@heyputer/backend/src/core/http';
 import { PuterDriver } from '@heyputer/backend/src/drivers/types';
+import type {
+    DriverConcurrentConfig,
+    DriverRateLimitConfig,
+} from '@heyputer/backend/src/drivers/meta';
+import {
+    DEFAULT_FREE_SUBSCRIPTION,
+    DEFAULT_TEMP_SUBSCRIPTION,
+} from '@heyputer/backend/src/services/metering/consts';
 import { extension } from '@heyputer/backend/src/extensions';
 
 // App-telemetry lets an app owner enumerate the users who have
@@ -50,17 +58,42 @@ const parseIntParam = (
  * Driver exposing the `app-telemetry` interface.
  *
  * The `/drivers/call` permission gate checks
- * `service:app-telemetry:ii:app-telemetry`, which every actor already holds
- * via the blanket `service` grant (hardcoded-permissions.js +
+ * `service:app-telemetry:ii:app-telemetry`, which every actor already holds via
+ * the blanket `service` grant (hardcoded-permissions.js +
  * `default_implicit_user_app_permissions`). The real authorization — "is the
- * caller the app owner?" — is enforced inside `get_users` below, exactly as
- * v1 did.
+ * caller the app owner?" — is enforced inside `get_users` below, exactly as v1
+ * did.
  */
 export class AppTelemetryDriver extends PuterDriver {
     readonly driverInterface = 'app-telemetry';
     readonly driverName = 'app-telemetry';
     readonly isDefault = true;
 
+    // Declaring nothing here would leave both methods on the generic
+    // 600/minute driver default, which does not fit a paginated scan that
+    // can ask for MAX_LIMIT rows at MAX_OFFSET. This is a dashboard read —
+    // nobody calls it in a loop.
+    readonly rateLimit: DriverRateLimitConfig = {
+        default: {
+            limit: 60,
+            window: 60_000,
+            bySubscription: {
+                [DEFAULT_FREE_SUBSCRIPTION]: 30,
+                [DEFAULT_TEMP_SUBSCRIPTION]: 10,
+            },
+        },
+    };
+
+    readonly concurrent: DriverConcurrentConfig = {
+        default: {
+            limit: 5,
+            bySubscription: {
+                [DEFAULT_FREE_SUBSCRIPTION]: 2,
+                [DEFAULT_TEMP_SUBSCRIPTION]: 2,
+            },
+        },
+    };
+
     /** Users who have authenticated into the given app (owner-only). */
     async get_users({
         app_uuid,

extensions/installedApps.ts

@@ -81,6 +81,16 @@ export const handleInstalledApps = async (
 
 extension.get(
     '/installedApps',
-    { subdomain: 'api', requireUserActor: true, allowFullAccessToken: true },
+    {
+        subdomain: 'api',
+        requireUserActor: true,
+        allowFullAccessToken: true,
+        rateLimit: {
+            scope: 'installed-apps',
+            limit: 120,
+            window: 60_000,
+            key: 'user',
+        },
+    },
     handleInstalledApps,
 );

extensions/metering.ts

@@ -3,6 +3,7 @@ import { HttpError } from '@heyputer/backend/src/core/http';
 import {
     controllersContainers,
     driversContainers,
+    servicesContainers,
 } from '@heyputer/backend/src/exports';
 import { extension } from '@heyputer/backend/src/extensions';
 import type { Request, Response } from 'express';
@@ -18,7 +19,7 @@ function collectAllCosts(): Record<string, unknown>[] {
     const all: Record<string, unknown>[] = [];
     const collect = (
         source: Record<string, unknown>,
-        kind: 'driver' | 'controller',
+        kind: 'driver' | 'controller' | 'service',
     ) => {
         for (const [name, instance] of Object.entries(source)) {
             const fn = (
@@ -43,6 +44,9 @@ function collectAllCosts(): Record<string, unknown>[] {
     };
     collect(driversContainers as Record<string, unknown>, 'driver');
     collect(controllersContainers as Record<string, unknown>, 'controller');
+    // Services report the costs that aren't tied to one endpoint — egress,
+    // which is metered for every response there is.
+    collect(servicesContainers as Record<string, unknown>, 'service');
     return all;
 }
 
@@ -110,26 +114,46 @@ export const handleMeteringAllCosts = async (
     res.json({ costs: cachedAllCosts });
 };
 
+/** Dashboard reads over the per-actor KV aggregates. */
+const USAGE_READ_LIMIT = {
+    scope: 'metering-usage',
+    limit: 120,
+    window: 60_000,
+    key: 'user' as const,
+};
+
 extension.get(
     '/metering/usage',
-    { subdomain: 'api', requireAuth: true },
+    { subdomain: 'api', requireAuth: true, rateLimit: USAGE_READ_LIMIT },
     handleMeteringUsage,
 );
 
 extension.get(
     '/metering/usage/:appIdOrName',
-    { subdomain: 'api', requireAuth: true },
+    { subdomain: 'api', requireAuth: true, rateLimit: USAGE_READ_LIMIT },
     handleMeteringUsageForApp,
 );
 
 extension.get(
     '/metering/globalUsage',
-    { subdomain: 'api', adminOnly: true },
+    {
+        subdomain: 'api',
+        adminOnly: true,
+        // Sums across every shard of the global aggregate. Admin-gated, so
+        // this is loop protection — but one accidental poll is an
+        // expensive minute.
+        rateLimit: {
+            scope: 'metering-global-usage',
+            limit: 10,
+            window: 60_000,
+            key: 'user',
+        },
+    },
     handleMeteringGlobalUsage,
 );
 
 extension.get(
     '/metering/allCosts',
-    { subdomain: 'api', requireAuth: true },
+    { subdomain: 'api', requireAuth: true, rateLimit: USAGE_READ_LIMIT },
     handleMeteringAllCosts,
 );

extensions/whoami.test.ts

@@ -109,6 +109,9 @@ describe('whoami extension — handleWhoami', () => {
         expect(body.oidc_only).toBe(false);
         // is_user_token is present (true) for user actors.
         expect(body.is_user_token).toBe(true);
+        // Account creation time, in unix seconds.
+        expect(typeof body.created_ts).toBe('number');
+        expect(body.created_ts).toBeGreaterThan(0);
         // `directories` is only sent to user actors — confirm it’s present.
         expect(body.directories).toBeDefined();
         // taskbar_items is only sent to user actors.
@@ -132,7 +135,7 @@ describe('whoami extension — handleWhoami', () => {
         expect(flags.payment_bypass).toBeUndefined();
     });
 
-    it('strips desktop_bg_* and human_readable_age fields for app actors', async () => {
+    it('strips desktop_bg_*, created_ts and human_readable_age fields for app actors', async () => {
         const user = await seedUser();
         const { res, captured } = makeRes();
 
@@ -154,6 +157,8 @@ describe('whoami extension — handleWhoami', () => {
         expect(body.desktop_bg_color).toBeUndefined();
         expect(body.desktop_bg_fit).toBeUndefined();
         expect(body.human_readable_age).toBeUndefined();
+        // Account age, in either form, is not exposed to apps.
+        expect(body.created_ts).toBeUndefined();
         // Directories are user-only.
         expect(body.directories).toBeUndefined();
     });

extensions/whoami.ts

@@ -14,6 +14,15 @@ const timeago = (() => {
     return new TimeAgo('en-US');
 })();
 
+// User timestamps come off the DB as SQL datetime strings; the wire format
+// for all of them is unix seconds. Unparseable values are dropped rather
+// than sent as NaN.
+const toUnixSeconds = (value: unknown): number | undefined => {
+    if (!value) return undefined;
+    const ms = new Date(value as string | number | Date).getTime();
+    return Number.isNaN(ms) ? undefined : Math.round(ms / 1000);
+};
+
 // Allowlist of `config.feature_flags` keys safe to surface via /whoami.
 // Anything not listed here stays server-side, so internal flags
 // (payment_bypass, staff_only_*, etc.) cannot leak by accident. Add a
@@ -98,6 +107,7 @@ export const handleWhoami = async (
             : undefined,
         otp: !!user.otp_enabled,
         feature_flags,
+        created_ts: toUnixSeconds(user.timestamp),
         human_readable_age: user.timestamp
             ? timeago.format(new Date(user.timestamp as string))
             : null,
@@ -141,14 +151,9 @@ export const handleWhoami = async (
     }
 
     // Last activity
-    if (user.last_activity_ts) {
-        try {
-            details.last_activity_ts = Math.round(
-                new Date(user.last_activity_ts as string).getTime() / 1000,
-            );
-        } catch {
-            /* ignore parse error */
-        }
+    const lastActivityTs = toUnixSeconds(user.last_activity_ts);
+    if (lastActivityTs !== undefined) {
+        details.last_activity_ts = lastActivityTs;
     }
 
     // Strip sensitive fields for app actors
@@ -164,6 +169,7 @@ export const handleWhoami = async (
         delete details.desktop_bg_color;
         delete details.desktop_bg_fit;
         delete details.human_readable_age;
+        delete details.created_ts;
         delete details.is_user_token;
         delete details.metadata;
     }
@@ -183,8 +189,7 @@ export const handleWhoami = async (
     }
 
     const subscription = details.subscription as
-        | { offering?: Record<string, unknown> }
-        | undefined;
+        { offering?: Record<string, unknown> } | undefined;
     if (subscription?.offering) {
         delete subscription.offering.group;
         delete subscription.offering.benefits;
@@ -196,6 +201,23 @@ export const handleWhoami = async (
 
 extension.get(
     '/whoami',
-    { subdomain: 'api', requireAuth: true, allowUnconfirmed: true },
+    {
+        subdomain: 'api',
+        requireAuth: true,
+        allowUnconfirmed: true,
+        // The GUI polls this, and each call fans out to every `whoami`
+        // event listener — so it costs more than the response suggests.
+        //
+        // It is also the call everything else leans on to find out who it is
+        // talking to, so it rides along with unrelated work rather than
+        // arriving at its own pace: the ceiling has to clear whatever the
+        // busiest session is doing, not what a person clicks.
+        rateLimit: {
+            scope: 'whoami',
+            limit: 1_800,
+            window: 60_000,
+            key: 'user',
+        },
+    },
     handleWhoami,
 );

package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "puter.com",
-  "version": "2.5.1",
+  "version": "26.07",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "puter.com",
-      "version": "2.5.1",
+      "version": "26.07",
       "license": "AGPL-3.0-only",
       "workspaces": [
         "src/*",
@@ -1304,6 +1304,7 @@
     },
     "node_modules/@clack/prompts/node_modules/is-unicode-supported": {
       "version": "1.3.0",
+      "extraneous": true,
       "inBundle": true,
       "license": "MIT",
       "engines": {
@@ -20252,7 +20253,7 @@
     },
     "src/puter-js": {
       "name": "@heyputer/puter.js",
-      "version": "2.5.4",
+      "version": "2.6.1",
       "license": "Apache-2.0",
       "dependencies": {
         "@heyputer/kv.js": "^0.2.1",

package.json

@@ -67,6 +67,9 @@
     "check-translations": "node tools/check-translations.js",
     "prepare": "husky",
     "build:ts": "tsc -p tsconfig.build.json && node ./tools/write-dist-package-json.mjs",
+    "check:puterjs:types": "node tools/checkPuterjsTypes.mjs",
+    "typecheck": "node tools/typecheck.mjs",
+    "typecheck:update": "node tools/typecheck.mjs --update",
     "setupExtensions": "node ./tools/extensionSetup.mjs"
   },
   "workspaces": [

src/backend/clients/alarm/AlarmClient.test.ts

@@ -371,6 +371,24 @@ describe('AlarmClient alarm registry', () => {
         expect(client.get('flap')?.occurrences).toHaveLength(2);
     });
 
+    it('caps retained occurrences while still counting every repeat', () => {
+        const client = makeClient();
+        const seen = capture(client);
+
+        for (let i = 0; i < 50; i++) {
+            client.create('hot', `occurrence ${i}`, { i });
+        }
+
+        const alarm = client.get('hot')!;
+        expect(alarm.count).toBe(50);
+        expect(alarm.occurrences).toHaveLength(20);
+        expect(alarm.timestamps).toHaveLength(20);
+        // The window kept is the most recent one, not the oldest.
+        expect(alarm.occurrences[19].message).toBe('occurrence 49');
+        // Trimming history must not rewind what the transports are told.
+        expect(seen[49]).toMatchObject({ repeatCount: 50, isRepeat: true });
+    });
+
     it('names anonymous handlers by their registration order', () => {
         const client = makeClient();
         client.addAlertHandler(async () => {

src/backend/clients/alarm/AlarmClient.ts

@@ -60,6 +60,15 @@ interface RegisteredHandler {
 
 /** Severity used when neither the call site nor config picks one. */
 const FALLBACK_SEVERITY: PagerSeverity = 'critical';
+/**
+ * How many recent occurrences an alarm keeps. An alarm is never cleared unless
+ * something calls `clear`, and a hot one repeats for as long as the fault lasts
+ * — so retaining every occurrence means retaining every message and field set
+ * it was ever raised with, request bodies and actors included, for the life of
+ * the process. The last few are what a human reads; the rest is only a count,
+ * and `count` keeps that.
+ */
+const OCCURRENCE_HISTORY_LIMIT = 20;
 /** Keeps `info` alarms out of the paging system unless config says otherwise. */
 const DEFAULT_PAGERDUTY_MIN_SEVERITY: PagerSeverity = 'warning';
 /** Slack's ceiling once a pager exists: chat gets what doesn't page. */
@@ -321,6 +330,7 @@ export class AlarmClient extends PuterClient {
             started: Date.now(),
             // `recordOccurrence` below stamps the first occurrence; seeding one
             // here too would report every alarm as one occurrence ahead.
+            count: 0,
             timestamps: [],
             occurrences: [],
         };
@@ -381,16 +391,25 @@ export class AlarmClient extends PuterClient {
         message: string,
         fields: AlarmFields,
     ): void {
+        const now = Date.now();
         alarm.message = message;
         alarm.fields = { ...alarm.fields, ...fields };
-        alarm.timestamps.push(Date.now());
+        alarm.count++;
         if (fields.error) alarm.error = fields.error;
 
-        alarm.occurrences.push({
-            message,
-            fields,
-            timestamp: Date.now(),
-        });
+        alarm.timestamps.push(now);
+        alarm.occurrences.push({ message, fields, timestamp: now });
+
+        if (alarm.timestamps.length > OCCURRENCE_HISTORY_LIMIT) {
+            alarm.timestamps.splice(
+                0,
+                alarm.timestamps.length - OCCURRENCE_HISTORY_LIMIT,
+            );
+            alarm.occurrences.splice(
+                0,
+                alarm.occurrences.length - OCCURRENCE_HISTORY_LIMIT,
+            );
+        }
     }
 
     private applyKnownErrors(alarm: Alarm): void {
@@ -438,7 +457,7 @@ export class AlarmClient extends PuterClient {
         this.applyKnownErrors(alarm);
 
         console.warn(
-            `[alarm] REPEAT ${displayId(alarm)} :: ${alarm.message} (${alarm.timestamps.length})`,
+            `[alarm] REPEAT ${displayId(alarm)} :: ${alarm.message} (${alarm.count})`,
         );
 
         if (alarm.noAlert) return;
@@ -476,7 +495,7 @@ export class AlarmClient extends PuterClient {
         alarm.severity = resolved;
 
         const fieldsClean = cleanFields(alarm.fields);
-        const repeatCount = alarm.timestamps.length;
+        const repeatCount = alarm.count;
 
         const id = alarm.id || 'something-bad';
 

src/backend/clients/alarm/types.ts

@@ -48,7 +48,17 @@ export interface Alarm extends AlarmOptions {
     fields: AlarmFields;
     error?: Error;
     started: number;
+    /**
+     * Every occurrence ever counted, including those aged out of the two lists
+     * below.
+     */
+    count: number;
+    /**
+     * Timestamps of the most recent occurrences only — see
+     * `OCCURRENCE_HISTORY_LIMIT`.
+     */
     timestamps: number[];
+    /** The most recent occurrences only — see `OCCURRENCE_HISTORY_LIMIT`. */
     occurrences: AlarmOccurrence[];
     severity?: PagerSeverity;
     noAlert?: boolean;

src/backend/clients/database/SQLBatcher.js

@@ -22,6 +22,7 @@ import {
     POOL_ACQUIRE_TIMEOUT,
     isNeverSentError,
     isRetriableError,
+    isRolledBackError,
 } from './retriableErrors.js';
 
 const DEFAULT_MAX_QUEUE_SIZE = 1000;
@@ -361,8 +362,14 @@ export class SQLBatcher {
     // Run one fallback item, retrying transient failures with backoff.
     // A read-only batcher may retry anything transient; a batcher that
     // carries writes only retries failures where the statement provably
-    // never reached the server — a write that died mid-flight may have
-    // committed, and re-running it would double-apply.
+    // did not apply — either it never reached the server, or the server
+    // rolled it back itself. A write that died mid-flight may have
+    // committed, and re-running that one would double-apply.
+    //
+    // Lock contention lands in the second group and is worth retrying rather
+    // than surfacing: an item is a single statement, so a deadlock victim has
+    // been fully undone, and the caller sees an unhandled 500 for what the
+    // database is telling us to just run again.
     async #runFallbackItem(b) {
         let attempt = 0;
         while (true) {
@@ -379,8 +386,8 @@ export class SQLBatcher {
                 };
             } catch (error) {
                 const canRetry = this.readOnly
-                    ? isRetriableError(error)
-                    : isNeverSentError(error);
+                    ? isRetriableError(error) || isRolledBackError(error)
+                    : isNeverSentError(error) || isRolledBackError(error);
                 if (!canRetry || attempt >= ITEM_RETRY_ATTEMPTS) {
                     return { ok: false, error };
                 }

src/backend/clients/database/SQLBatcher.test.ts

@@ -208,6 +208,36 @@ describe('SQLBatcher', () => {
         expect(fallbackAttempts).toBe(2);
     });
 
+    it('retries a deadlocked write instead of surfacing it', async () => {
+        let fallbackAttempts = 0;
+        const conn = makeConnection((sql) => {
+            if (isBatchQuery(sql)) throw makeError('ER_LOCK_DEADLOCK');
+            fallbackAttempts++;
+            if (fallbackAttempts === 1) throw makeError('ER_LOCK_DEADLOCK');
+            return [[{ ok: 1 }], undefined];
+        });
+        const { pool } = makePool(conn);
+        const batcher = new SQLBatcher(pool, { maxTimeInQueue: 5 });
+
+        // The victim statement was rolled back by the server, so re-running it
+        // can't double-apply — the caller should never see the deadlock.
+        const result = await batcher.query('UPDATE notification SET x', []);
+        expect(result[0]).toEqual([{ ok: 1 }]);
+        expect(fallbackAttempts).toBe(2);
+    });
+
+    it('gives up on a write that deadlocks past the retry budget', async () => {
+        const conn = makeConnection(() => {
+            throw makeError('ER_LOCK_DEADLOCK');
+        });
+        const { pool } = makePool(conn);
+        const batcher = new SQLBatcher(pool, { maxTimeInQueue: 5 });
+
+        await expect(batcher.query('UPDATE hot SET x', [])).rejects.toMatchObject(
+            { code: 'ER_LOCK_DEADLOCK' },
+        );
+    });
+
     it('never retries deterministic row-level errors and does not escalate the breaker', async () => {
         let fallbackAttempts = 0;
         const conn = makeConnection((sql) => {

src/backend/clients/database/SqliteDatabaseClient.test.ts

@@ -27,7 +27,7 @@ import { DatabaseClientFactory } from './index.js';
 import { SqliteDatabaseClient } from './SqliteDatabaseClient.js';
 
 /** Highest schema version the migration table can reach. */
-const CURRENT_SCHEMA_VERSION = 60;
+const CURRENT_SCHEMA_VERSION = 61;
 const SYSTEM_USER_UUID = '5d4adce0-a381-4982-9c02-6e2540026238';
 
 const sqliteConfig = (

src/backend/clients/database/SqliteDatabaseClient.ts

@@ -94,6 +94,7 @@ const AVAILABLE_MIGRATIONS: [number, string[]][] = [
     [57, ['0062_blocked-app-origins.sql']],
     [58, ['0063_add_suspended_reason.sql']],
     [59, ['0064_abuse-moderation-events.sql']],
+    [60, ['0065_app-feedback.sql']],
 ];
 
 export class SqliteDatabaseClient extends AbstractDatabaseClient {

src/backend/clients/database/migrations/mysql/mysql_mig_20.sql

@@ -0,0 +1,48 @@
+-- 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/>.
+
+-- User-to-developer feedback for apps that opt in. Mirrors SQLite migration
+-- 0065. Opt-in is the new `apps.feedback_enabled` column (developer-writable
+-- through the regular `puter.apps.update` path). Each row of `app_feedback`
+-- is one message a signed-in user submitted through the GUI feedback dialog;
+-- a copy is emailed to the app owner unless the per-app daily email cap
+-- suppressed it (`email_sent` records which). `app_uid` is denormalized
+-- alongside `app_id` so rows stay attributable after an app is deleted.
+-- `created_at` is unix seconds.
+--
+-- Idempotent: the column add uses _puter_add_col (defined in mig_1, which
+-- always runs first) and the table uses `CREATE TABLE IF NOT EXISTS`, so the
+-- directory can replay safely.
+
+CALL _puter_add_col('apps', 'feedback_enabled', '`feedback_enabled` tinyint(1) DEFAULT ''0''');
+
+CREATE TABLE IF NOT EXISTS `app_feedback` (
+    `id` INT NOT NULL AUTO_INCREMENT,
+    `uid` CHAR(36) NOT NULL,
+    `app_id` INT NOT NULL,
+    `app_uid` CHAR(40) NOT NULL,
+    `user_id` INT NOT NULL,
+    `message` TEXT NOT NULL,
+    `source_env` VARCHAR(16) DEFAULT NULL,
+    `source_origin` VARCHAR(2048) DEFAULT NULL,
+    `email_sent` TINYINT(1) NOT NULL DEFAULT 0,
+    `created_at` BIGINT NOT NULL,
+    PRIMARY KEY (`id`),
+    UNIQUE KEY `uq_app_feedback_uid` (`uid`),
+    KEY `idx_app_feedback_app_created` (`app_id`, `created_at`),
+    KEY `idx_app_feedback_user_created` (`user_id`, `created_at`)
+) ENGINE=InnoDB DEFAULT CHARSET=utf8mb4 COLLATE=utf8mb4_unicode_ci;

src/backend/clients/database/migrations/postgres/postgres_mig_9.sql

@@ -0,0 +1,47 @@
+-- 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/>.
+
+-- User-to-developer feedback for apps that opt in. Mirrors SQLite migration
+-- 0065 / MySQL mysql_mig_20. Opt-in is the new `apps.feedback_enabled` column
+-- (developer-writable through the regular `puter.apps.update` path). Each row
+-- of `app_feedback` is one message a signed-in user submitted through the GUI
+-- feedback dialog; a copy is emailed to the app owner unless the per-app
+-- daily email cap suppressed it (`email_sent` records which). `app_uid` is
+-- denormalized alongside `app_id` so rows stay attributable after an app is
+-- deleted. `created_at` is unix seconds.
+--
+-- Idempotent via IF NOT EXISTS.
+
+ALTER TABLE apps ADD COLUMN IF NOT EXISTS feedback_enabled boolean NOT NULL DEFAULT FALSE;
+
+CREATE TABLE IF NOT EXISTS app_feedback (
+    id BIGSERIAL PRIMARY KEY,
+    uid CHAR(36) NOT NULL UNIQUE,
+    app_id BIGINT NOT NULL,
+    app_uid VARCHAR(40) NOT NULL,
+    user_id BIGINT NOT NULL,
+    message TEXT NOT NULL,
+    source_env VARCHAR(16) DEFAULT NULL,
+    source_origin VARCHAR(2048) DEFAULT NULL,
+    email_sent BOOLEAN NOT NULL DEFAULT FALSE,
+    created_at BIGINT NOT NULL
+);
+
+CREATE INDEX IF NOT EXISTS idx_app_feedback_app_created
+    ON app_feedback (app_id, created_at);
+CREATE INDEX IF NOT EXISTS idx_app_feedback_user_created
+    ON app_feedback (user_id, created_at);

src/backend/clients/database/migrations/sqlite/0065_app-feedback.sql

@@ -0,0 +1,51 @@
+-- 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/>.
+
+-- User-to-developer feedback for apps that opt in. Opt-in is the new
+-- `apps.feedback_enabled` column (developer-writable through the regular
+-- `puter.apps.update` path; a dedicated column rather than a `metadata` key
+-- because Dev Center saves replace the whole metadata blob and would erase
+-- it). Each row of `app_feedback` is one message a signed-in user submitted
+-- through the GUI feedback dialog; a copy is emailed to the app owner unless
+-- the per-app daily email cap suppressed it (`email_sent` records which).
+-- `app_uid` is denormalized alongside `app_id` so rows stay attributable
+-- after an app is deleted (abuse forensics). `source_env` is 'app' (desktop
+-- dialog) or 'web' (puter.com popup opened from an external site);
+-- `source_origin` is the popup opener's browser-attested origin, null for
+-- desktop submissions. `created_at` is unix seconds. The (user_id,
+-- created_at) and (app_id, created_at) indexes serve the sliding-window
+-- rate-limit counts in AppFeedbackStore.
+
+ALTER TABLE apps ADD COLUMN "feedback_enabled" tinyint(1) DEFAULT '0';
+
+CREATE TABLE IF NOT EXISTS `app_feedback` (
+    "id" INTEGER PRIMARY KEY AUTOINCREMENT,
+    "uid" TEXT NOT NULL UNIQUE,
+    "app_id" INTEGER NOT NULL,
+    "app_uid" TEXT NOT NULL,
+    "user_id" INTEGER NOT NULL,
+    "message" TEXT NOT NULL,
+    "source_env" TEXT DEFAULT NULL,     -- 'app' | 'web'
+    "source_origin" TEXT DEFAULT NULL,  -- attested opener origin (web popups)
+    "email_sent" INTEGER NOT NULL DEFAULT 0,
+    "created_at" INTEGER NOT NULL       -- unix seconds
+);
+
+CREATE INDEX IF NOT EXISTS idx_app_feedback_app_created
+    ON `app_feedback` (`app_id`, `created_at`);
+CREATE INDEX IF NOT EXISTS idx_app_feedback_user_created
+    ON `app_feedback` (`user_id`, `created_at`);

src/backend/clients/database/retriableErrors.ts

@@ -54,6 +54,26 @@ const NEVER_SENT_ERROR_CODES = new Set([
     POOL_ACQUIRE_TIMEOUT,
 ]);
 
+/**
+ * Failures the server itself rolled back before returning. The statement did
+ * reach the database, so this is not `NEVER_SENT`, but InnoDB guarantees it
+ * left no effect — which makes a retry just as safe for writes.
+ *
+ * Only true for a _single_ statement: under autocommit each statement is its
+ * own transaction, so "the transaction was rolled back" means "this statement
+ * was rolled back". Retrying a multi-statement string on one of these would
+ * re-run the statements that already committed ahead of the failure.
+ */
+const ROLLED_BACK_ERROR_CODES = new Set([
+    // Deadlock — InnoDB picked this transaction as the victim and undid it.
+    // MySQL's own message for it is "try restarting transaction".
+    'ER_LOCK_DEADLOCK',
+    // Lock wait timeout. Rolls back the statement rather than the transaction
+    // unless innodb_rollback_on_timeout is set — the same thing when the
+    // transaction is one statement.
+    'ER_LOCK_WAIT_TIMEOUT',
+]);
+
 const errorCode = (error: unknown): string | undefined =>
     (error as { code?: string } | null)?.code;
 
@@ -74,3 +94,13 @@ export const isNeverSentError = (error: unknown): boolean => {
     const code = errorCode(error);
     return Boolean(code && NEVER_SENT_ERROR_CODES.has(code));
 };
+
+/**
+ * Lock-contention failures the server rolled back on its own. Safe to retry a
+ * single statement on, writes included — see `ROLLED_BACK_ERROR_CODES` for the
+ * one-statement precondition.
+ */
+export const isRolledBackError = (error: unknown): boolean => {
+    const code = errorCode(error);
+    return Boolean(code && ROLLED_BACK_ERROR_CODES.has(code));
+};

src/backend/clients/email/EmailClient.test.ts

@@ -179,6 +179,29 @@ describe('EmailClient — template rendering', () => {
         expect(captured.html).toContain('<strong>424242</strong>');
     });
 
+    it('does not HTML-escape values in the plain-text subject header', async () => {
+        const client = startClient();
+        const captured: { subject?: string } = {};
+        vi.spyOn(client, 'sendRaw').mockImplementation(async (options) => {
+            captured.subject = options.subject;
+            return null;
+        });
+
+        await client.send('[email protected]', 'app-user-feedback', {
+            owner_username: 'dev',
+            sender_username: 'user',
+            sender_email: null,
+            app_title: "Bob's App & Games",
+            app_name: 'bobs-app',
+            app_link: 'https://puter.example/app/bobs-app',
+            message: 'hi',
+        });
+
+        // A subject is not HTML — entities would render literally in the
+        // recipient's mail client.
+        expect(captured.subject).toBe("New user feedback for Bob's App & Games");
+    });
+
     it('escapes html and converts newlines in nl2br values', async () => {
         const client = startClient();
         let html = '';

src/backend/clients/email/EmailClient.ts

@@ -152,11 +152,16 @@ export class EmailClient extends PuterClient {
 
     // -- Public API: sending ------------------------------------------
 
-    /** Render a template and send it to `to`. */
+    /**
+     * Render a template and send it to `to`. `options.replyTo` sets the
+     * Reply-To header (e.g. so a recipient can respond to the originator of the
+     * message rather than the no-reply From address).
+     */
     async send<T extends EmailTemplateName>(
         to: string,
         template: T,
         values: Record<string, unknown> = {},
+        options: { replyTo?: string } = {},
     ): Promise<void> {
         const compiled = this.compiledTemplates[template];
         if (!compiled) {
@@ -168,6 +173,7 @@ export class EmailClient extends PuterClient {
             to,
             subject: compiled.subject(values),
             html: compiled.html(values),
+            ...(options.replyTo ? { replyTo: options.replyTo } : {}),
         });
     }
 
@@ -285,7 +291,14 @@ export class EmailClient extends PuterClient {
     private compileTemplates(): void {
         for (const [name, template] of Object.entries(EMAIL_TEMPLATES)) {
             this.compiledTemplates[name as EmailTemplateName] = {
-                subject: handlebars.compile(template.subject),
+                // Subjects are plain-text headers: HTML-escaping would put
+                // literal entities in front of the recipient (&amp; etc.).
+                // Header safety is handled elsewhere — the transport encodes
+                // newlines, and free-form values (e.g. app_title) collapse
+                // whitespace upstream.
+                subject: handlebars.compile(template.subject, {
+                    noEscape: true,
+                }),
                 html: handlebars.compile(dedent(template.html)),
             };
         }

src/backend/clients/email/templates.ts

@@ -75,6 +75,25 @@ Please update <a href="https://puter.com/app/{{app_name}}">{{app_title}}</a>.
 <blockquote>{{nl2br message}}</blockquote>
 <p>Best,<br />
 The Puter Team
+</p>
+        `,
+    },
+    'app-user-feedback': {
+        subject: 'New user feedback for {{app_title}}',
+        html: `
+<p>Hi{{#if owner_username}} {{owner_username}}{{/if}},</p>
+<p>
+<strong>{{sender_username}}</strong>{{#if sender_email}} (<a href="mailto:{{sender_email}}">{{sender_email}}</a>){{/if}} sent feedback about <a href="{{app_link}}">{{app_title}}</a>:
+</p>
+<blockquote>{{{nl2br message}}}</blockquote>
+{{#if sender_email}}<p>Just reply to this email to respond to them directly.</p>{{/if}}
+<p>
+You're receiving this because feedback is enabled for
+<a href="{{app_link}}">{{app_title}}</a> — manage it in the
+<a href="{{dev_center_link}}">Dev Center</a> under your app's settings.
+</p>
+<p>Best,<br />
+The Puter Team
 </p>
         `,
     },

src/backend/clients/event/EventClient.test.ts

@@ -214,6 +214,38 @@ describe('EventClient', () => {
         });
     });
 
+    describe('hasListeners', () => {
+        it('is false for a key nobody subscribed to', () => {
+            expect(target.hasListeners(key)).toBe(false);
+        });
+
+        it('is true for an exact subscriber', () => {
+            target.on(key, vi.fn());
+            expect(target.hasListeners(key)).toBe(true);
+        });
+
+        it('is true when only a wildcard prefix matches', () => {
+            target.on(`${key}.*`, vi.fn());
+            expect(target.hasListeners(`${key}.child.deep`)).toBe(true);
+        });
+
+        it('is true for an extension-registered listener', () => {
+            extensionStore.events[key] = [vi.fn()];
+            try {
+                expect(target.hasListeners(key)).toBe(true);
+            } finally {
+                delete extensionStore.events[key];
+            }
+        });
+
+        it('goes back to false once the last listener is removed', () => {
+            const listener = vi.fn();
+            target.on(key, listener);
+            target.off(key, listener);
+            expect(target.hasListeners(key)).toBe(false);
+        });
+    });
+
     describe('extension-registered listeners', () => {
         afterEach(() => {
             delete extensionStore.events[key];

src/backend/clients/event/EventClient.ts

@@ -129,6 +129,27 @@ export class EventClient extends PuterClient {
         });
     }
 
+    /**
+     * Whether anything would run if `key` were emitted, wildcards included.
+     *
+     * For emitters whose work is only worth doing when someone is listening —
+     * gathering the payload costs a round trip, say. Emitting into the void is
+     * otherwise perfectly cheap and doesn't need this.
+     */
+    hasListeners<T extends keyof EventMap>(key: T): boolean {
+        const parts = key.split('.');
+        for (let i = 0; i < parts.length; i++) {
+            const matchKey = (
+                i === parts.length - 1
+                    ? key
+                    : `${parts.slice(0, i + 1).join('.')}.*`
+            ) as ListenKey;
+            if (this.#eventListeners[matchKey]?.length) return true;
+            if (extensionStore.events[matchKey]?.length) return true;
+        }
+        return false;
+    }
+
     /**
      * Subscribe to an event by exact key OR a wildcard prefix.
      *

src/backend/clients/event/types.ts

@@ -20,6 +20,7 @@ import type {
     Response as ExpressResponse,
 } from 'express';
 import { Actor } from '../../core';
+import type { UsageInput } from '../../services/metering/types';
 import { FSEntry } from '../../stores/fs/FSEntry';
 
 // GUI write events spread an entry plus per-event metadata into `response`.
@@ -305,6 +306,38 @@ export type EventMap = {
     };
     'storage.quota.bonus': { userId: number; extra: number };
 
+    // ---- Metering ----
+    // Recurring charges are pure mechanism here: the metering service knows
+    // when to ask (once per user per month, the first time that month's usage
+    // record is touched) and how to record the answer, and an extension
+    // decides what the account owes. Listeners push onto `charges`; the
+    // service applies them as ordinary usage once every listener has run.
+    // Nothing listening → no charges and no claim is ever taken.
+    //
+    // Emitted after the claim is settled, so a listener that calls back into
+    // metering can't re-trigger it.
+    'metering.monthly.charges': {
+        /**
+         * User-scoped, with no app on it: the account is what recurs, and the
+         * app that happened to trigger the month's first call has nothing to do
+         * with what is owed. Price against what the user owns.
+         */
+        actor: Actor;
+        /** The month being charged for, `YYYY-MM` in UTC. */
+        month: string;
+        /**
+         * Push what the user owes here. Every listener's charges are merged
+         * into one amount map and recorded as a single increment.
+         */
+        charges: UsageInput[];
+    };
+
+    // ---- Workers ----
+    // Only a genuinely new worker. Redeploying an existing name updates its
+    // row instead, and never reaches here — so a listener that prices this
+    // is pricing workers that came into existence, not deploys.
+    'worker.create': { actor: Actor; workerName: string };
+
     // ---- Outer / GUI broadcast ----
     'outer.cacheUpdate': {
         cacheKey: string[];
@@ -312,6 +345,15 @@ export type EventMap = {
         ttlSeconds?: number;
     };
     'outer.fs.write-hash': { hash: string; uuid: string };
+    /**
+     * Cache keys the KV read cache must stop serving, because the entries
+     * behind them were just written somewhere else.
+     *
+     * `outer.*` rather than `outer.pubsub.*` on purpose: the cache lives in the
+     * Redis a cluster shares, so one node applying the invalidation covers the
+     * whole cluster — fanning it out to siblings would just repeat the write.
+     */
+    'outer.kv.cacheInvalidated': { cacheKeys: string[] };
     'outer.gui.item.added': GuiEvent;
     'outer.gui.item.updated': GuiEvent;
     'outer.gui.item.moved': GuiEvent;
@@ -402,6 +444,22 @@ export type EventMap = {
     [K in `route.${string}`]: RouteLifecycleEvent;
 } & {
     [K in `pubsub.login.${string}`]: { authtoken: string };
+} & {
+    /**
+     * A user's subscription now resolves to a different policy. Carried on the
+     * `outer.pubsub.*` channel so it reaches sibling nodes and peer clusters,
+     * not just the one that handled the change — every node caches the resolved
+     * policy, so a purchase is only live once they have all dropped theirs.
+     */
+    'outer.pubsub.metering.subscription-changed': { userUuid: string };
+    /**
+     * A user's purchased credit balance changed. Separate from a policy change
+     * because it moves the other half of the same budget, and carried on the
+     * `outer.pubsub.*` channel for the same reason: every node caches whether
+     * an account has budget left, so a top-up only lifts enforcement once they
+     * have all dropped that answer.
+     */
+    'outer.pubsub.metering.credits-changed': { userUuid: string };
 };
 
 /**

src/backend/clients/s3/S3Client.ts

@@ -21,6 +21,7 @@ import {
     AbortMultipartUploadCommand,
     CompleteMultipartUploadCommand,
     CreateMultipartUploadCommand,
+    HeadBucketCommand,
     PutObjectCommand,
     S3Client as AwsS3Client,
     type S3ClientConfig,
@@ -244,6 +245,19 @@ export class S3Client extends PuterClient {
         return client;
     }
 
+    /**
+     * Cheapest round-trip that proves the object store answers for a bucket: no
+     * object data, no listing, just a HEAD. Throws on any failure (missing
+     * bucket, bad credentials, unreachable endpoint) so callers can treat it as
+     * a liveness probe. Reuses the pooled per-region client.
+     */
+    async headBucket(
+        bucket = this.config.s3_bucket || LEGACY_STORAGE_BUCKET,
+        region?: string,
+    ): Promise<void> {
+        await this.get(region).send(new HeadBucketCommand({ Bucket: bucket }));
+    }
+
     // ------------------------------------------------------------------
     // Legacy storage migration
     // ------------------------------------------------------------------

src/backend/controllers/apps/AppController.js

@@ -35,6 +35,39 @@ import DEFAULT_APP_ICON from './default-app-icon.js';
  * are just thin shape adapters that translate REST conventions into driver
  * calls.
  */
+/**
+ * Desktop boot reads the app list and individual app records repeatedly, so the
+ * ceiling is set well above normal boot traffic and exists to catch a runaway
+ * client rather than to pace one.
+ *
+ * "Repeatedly" is the operative word: an app record is read on launch, on
+ * permission checks, and again by anything resolving an app by name, so these
+ * accumulate against whatever else a session is doing rather than arriving on
+ * their own. Sized for a session working hard, not for a person clicking.
+ */
+const APP_READ_LIMIT = {
+    scope: 'app-read',
+    limit: 1_800,
+    window: 60_000,
+    key: 'user',
+};
+
+/**
+ * Unauthenticated icon serving; no actor to key on, so the bucket is the
+ * address — and an address is a NAT, a campus or a carrier gateway that can
+ * hold hundreds of desktops. Each desktop boot pulls the taskbar's icons at
+ * once, so a single burst from one network is already thousands of requests.
+ * Responses are publicly cacheable and usually a redirect, so the ceiling is
+ * not protecting bandwidth; it is there so a client looping on a broken icon
+ * can't spin unbounded.
+ */
+const APP_ICON_LIMIT = {
+    scope: 'app-icon',
+    limit: 12_000,
+    window: 60_000,
+    key: 'ip',
+};
+
 export class AppController extends PuterController {
     get appStore() {
         return this.stores.app;
@@ -100,7 +133,9 @@ export class AppController extends PuterController {
         return d;
     }
 
-    registerRoutes(router) {
+    registerRoutes(
+        /** @type {import('../../core/http/PuterRouter').PuterRouter} */ router,
+    ) {
         // GET /apps — list apps owned by the current user
         router.get(
             '/apps',
@@ -108,6 +143,7 @@ export class AppController extends PuterController {
                 subdomain: 'api',
                 requireUserActor: true,
                 allowFullAccessToken: true,
+                rateLimit: APP_READ_LIMIT,
             },
             async (req, res) => {
                 const apps = await this.appDriver.select({
@@ -122,8 +158,16 @@ export class AppController extends PuterController {
             '/apps/nameAvailable',
             {
                 subdomain: 'api',
-                requireUserActor: true,
-                allowFullAccessToken: true,
+                requireAuth: true,
+                // Answers "does this name exist?" for any name, so it is a
+                // name-enumeration oracle however cheap it is to serve.
+                // Mirrors the `isNameAvailable` budget on AppDriver.
+                rateLimit: {
+                    scope: 'app-name-available',
+                    limit: 60,
+                    window: 60_000,
+                    key: 'user',
+                },
             },
             async (req, res) => {
                 const name = req.query?.name;
@@ -158,6 +202,7 @@ export class AppController extends PuterController {
             {
                 subdomain: 'api',
                 requireAuth: true,
+                rateLimit: APP_READ_LIMIT,
             },
             async (req, res) => {
                 const actor = req.actor;
@@ -216,6 +261,7 @@ export class AppController extends PuterController {
                 subdomain: 'api',
                 requireUserActor: true,
                 allowFullAccessToken: true,
+                rateLimit: APP_READ_LIMIT,
             },
             async (req, res) => {
                 const raw = req.params.name;
@@ -289,6 +335,7 @@ export class AppController extends PuterController {
             {
                 subdomain: 'api',
                 requireAuth: true,
+                rateLimit: APP_READ_LIMIT,
             },
             async (req, res) => {
                 const appList = Array.isArray(req.body) ? req.body : [];
@@ -497,10 +544,14 @@ export class AppController extends PuterController {
         // Icons are <img src> targets from the GUI (root) AND resolved via
         // api_base_url in taskbar payloads. Register on both so either origin
         // works without a cross-subdomain redirect.
-        router.get('/app-icon/:app_uid', { subdomain: ['api', ''] }, serveIcon);
+        router.get(
+            '/app-icon/:app_uid',
+            { subdomain: ['api', ''], rateLimit: APP_ICON_LIMIT },
+            serveIcon,
+        );
         router.get(
             '/app-icon/:app_uid/:size',
-            { subdomain: ['api', ''] },
+            { subdomain: ['api', ''], rateLimit: APP_ICON_LIMIT },
             serveIcon,
         );
     }

src/backend/controllers/apps/AppController.test.ts

@@ -1018,3 +1018,25 @@ describe('AppController GET /app-icon remote icons', () => {
         expect(captured.headers['content-type']).toContain('image/svg+xml');
     });
 });
+
+// ── app-icon rate limit ─────────────────────────────────────────────
+
+describe('AppController app-icon rate limit', () => {
+    it('sizes the icon bucket for a whole network rather than one desktop', () => {
+        // Icons are unauthenticated `<img src>` targets, so the only key is
+        // the address — which one NAT shares across every desktop behind it,
+        // and each boot pulls the taskbar's icons in a burst.
+        for (const path of ['/app-icon/:app_uid', '/app-icon/:app_uid/:size']) {
+            const route = router.routes.find(
+                (r) => r.method === 'get' && r.path === path,
+            );
+            if (!route) throw new Error(`No GET ${path} route`);
+            expect(route.options.rateLimit).toEqual({
+                scope: 'app-icon',
+                limit: 12_000,
+                window: 60_000,
+                key: 'ip',
+            });
+        }
+    });
+});

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

@@ -42,6 +42,7 @@ import type { TokenSource } from '../../core/http/types.js';
 import { PuterServer } from '../../server.js';
 import { FULL_API_ACCESS } from '../../services/permission/consts.js';
 import { setupTestServer } from '../../testUtil.js';
+import { FS_READ_LIMIT } from '../fs/limits.js';
 
 // ── Test harness ────────────────────────────────────────────────────
 
@@ -1390,6 +1391,26 @@ describe('AuthController account-lifecycle route gating', () => {
         expect(opts.requireUserActor).toBe(true);
     });
 
+    // A fresh token is minted per protected mutation and nothing caches them,
+    // so issuance has to outrun the COMBINED rate of everything that spends
+    // one. The session-authenticated download path dominates — a multi-select
+    // download spends a token per file — with logout and the handful of
+    // session-management writes behind it.
+    it('GET /get-anticsrf-token clears the budgets that consume tokens', () => {
+        type Window = { limit: number; window: number };
+        const issuance = routeOptions('get', '/get-anticsrf-token')
+            .rateLimit as Window;
+        const logout = routeOptions('post', '/logout').rateLimit as Window;
+
+        expect(issuance.window).toBe(60_000);
+        expect(logout.window).toBe(60_000);
+        expect(FS_READ_LIMIT.window).toBe(60_000);
+
+        expect(issuance.limit).toBeGreaterThan(
+            FS_READ_LIMIT.limit + logout.limit,
+        );
+    });
+
     it('requireUserActorGate rejects app-under-user and access-token actors', () => {
         const gate = requireUserActorGate();
         const run = (actor: Partial<Actor>) =>
@@ -6907,3 +6928,253 @@ describe('AuthController.loginWait audience binding', () => {
         expect(res.body).toBeUndefined();
     });
 });
+
+// -- Batch grant / revoke + cross-app data grants -----------------------
+
+describe('AuthController — app-data grants', () => {
+    let issuer: { id: number; username: string };
+    let issuerActor: Actor;
+
+    beforeAll(async () => {
+        const name = `ad_${Math.random().toString(36).slice(2, 10)}`;
+        await controller.handleSignup(
+            makeReq({
+                username: name,
+                email: `${name}@test.local`,
+                password: 'correct-horse-battery',
+            }),
+            makeRes(),
+        );
+        const row = await server.stores.user.getByUsername(name);
+        await server.stores.user.update(row!.id, { email_confirmed: 1 });
+        issuer = { id: row!.id, username: row!.username };
+        issuerActor = {
+            user: {
+                id: row!.id,
+                uuid: row!.uuid,
+                username: row!.username,
+                email: row!.email,
+                email_confirmed: true,
+            },
+        } as Actor;
+    });
+    const makeAppRow = async (fields: Record<string, unknown> = {}) =>
+        (await server.stores.app.create(
+            {
+                name: `ad-${uuidv4()}`,
+                title: 'AppDataTest',
+                index_url: 'https://example.test/index.html',
+                ...fields,
+            },
+            { ownerUserId: issuer.id },
+        )) as { id: number; uid: string };
+
+    const grantedPermissions = async (appUid: string): Promise<string[]> => {
+        const rows = (await server.clients.db.read(
+            'SELECT p.`permission` FROM `user_to_app_permissions` p ' +
+                'JOIN `apps` a ON a.`id` = p.`app_id` ' +
+                'WHERE p.`user_id` = ? AND a.`uid` = ?',
+            [issuer.id, appUid],
+        )) as Array<{ permission: string }>;
+        return rows.map((r) => r.permission);
+    };
+
+    const post = (
+        handler: 'handleGrantUserApp' | 'handleRevokeUserApp',
+        body: Record<string, unknown>,
+    ) => {
+        const res = makeRes();
+        return inCtx(issuerActor, () =>
+            controller[handler](makeReq(body, { actor: issuerActor }), res),
+        ).then(() => res);
+    };
+
+    it('grants every permission in a list in one request', async () => {
+        const grantee = await makeAppRow();
+        const target = await makeAppRow();
+        const permissions = [
+            `app-data:${target.uid}:kv:read`,
+            `app-data:${target.uid}:kv:delete`,
+        ];
+
+        await post('handleGrantUserApp', {
+            app_uid: grantee.uid,
+            permissions,
+        });
+
+        expect(await grantedPermissions(grantee.uid)).toEqual(
+            expect.arrayContaining(permissions),
+        );
+    });
+
+    it('revokes every permission in a list in one request', async () => {
+        const grantee = await makeAppRow();
+        const target = await makeAppRow();
+        const permissions = [
+            `app-data:${target.uid}:kv:read`,
+            `app-data:${target.uid}:fs:read`,
+        ];
+        await post('handleGrantUserApp', {
+            app_uid: grantee.uid,
+            permissions,
+        });
+
+        await post('handleRevokeUserApp', {
+            app_uid: grantee.uid,
+            permissions,
+        });
+
+        const remaining = await grantedPermissions(grantee.uid);
+        for (const permission of permissions) {
+            expect(remaining).not.toContain(permission);
+        }
+    });
+
+    it('rejects the scalar and list forms together', async () => {
+        const grantee = await makeAppRow();
+        const target = await makeAppRow();
+        await expect(
+            post('handleGrantUserApp', {
+                app_uid: grantee.uid,
+                permission: `app-data:${target.uid}:kv:read`,
+                permissions: [`app-data:${target.uid}:kv:write`],
+            }),
+        ).rejects.toMatchObject({ statusCode: 400 });
+    });
+
+    it('rejects an empty, oversized, or `*`-bearing list', async () => {
+        const grantee = await makeAppRow();
+        const target = await makeAppRow();
+        for (const permissions of [
+            [],
+            Array.from(
+                { length: 17 },
+                (_x, i) => `app-data:${target.uid}:kv:read${i}`,
+            ),
+            [`app-data:${target.uid}:kv:read`, '*'],
+        ]) {
+            await expect(
+                post('handleGrantUserApp', {
+                    app_uid: grantee.uid,
+                    permissions,
+                }),
+            ).rejects.toMatchObject({ statusCode: 400 });
+        }
+    });
+
+    it('writes nothing when one entry in the list is invalid', async () => {
+        const grantee = await makeAppRow();
+        const target = await makeAppRow();
+        await expect(
+            post('handleGrantUserApp', {
+                app_uid: grantee.uid,
+                permissions: [
+                    `app-data:${target.uid}:kv:read`,
+                    `app-data:app-does-not-exist:kv:read`,
+                ],
+            }),
+        ).rejects.toMatchObject({ statusCode: 404 });
+        // The valid entry must not have landed: validation runs over the whole
+        // list before any row is written.
+        expect(await grantedPermissions(grantee.uid)).not.toContain(
+            `app-data:${target.uid}:kv:read`,
+        );
+    });
+
+    it('writes nothing when an entry is too wide for the column it lands in', async () => {
+        // `#validateAppPermissionParams` allows 4096 chars but the column is 255,
+        // so this passes shape validation and fails inside the grant. Before the
+        // pre-flight, the first entry committed and the caller still got a 400 —
+        // and the dialog reads a 4xx as "nothing was written".
+        const grantee = await makeAppRow();
+        const target = await makeAppRow();
+        const good = `app-data:${target.uid}:kv:read`;
+        await expect(
+            post('handleGrantUserApp', {
+                app_uid: grantee.uid,
+                permissions: [good, 'x'.repeat(300)],
+            }),
+        ).rejects.toMatchObject({ statusCode: 400 });
+        expect(await grantedPermissions(grantee.uid)).not.toContain(good);
+    });
+
+    it('404s when the target app does not exist', async () => {
+        const grantee = await makeAppRow();
+        await expect(
+            post('handleGrantUserApp', {
+                app_uid: grantee.uid,
+                permission: 'app-data:app-nope/:kv:read',
+            }),
+        ).rejects.toMatchObject({ statusCode: 404 });
+    });
+
+    it('rejects a bare `app-data` with no target app', async () => {
+        const grantee = await makeAppRow();
+        for (const permission of ['app-data', 'app-data:', 'app-data::kv']) {
+            await expect(
+                post('handleGrantUserApp', {
+                    app_uid: grantee.uid,
+                    permission,
+                }),
+            ).rejects.toMatchObject({ statusCode: 400 });
+        }
+    });
+
+    it('refuses a grant when the target opted out of sharing', async () => {
+        const grantee = await makeAppRow();
+        const closed = await makeAppRow({
+            metadata: JSON.stringify({ share_app_data: false }),
+        });
+        await expect(
+            post('handleGrantUserApp', {
+                app_uid: grantee.uid,
+                permission: `app-data:${closed.uid}:kv:read`,
+            }),
+        ).rejects.toMatchObject({ statusCode: 403 });
+    });
+
+    it('still allows revoking a grant after the target opts out', async () => {
+        const grantee = await makeAppRow();
+        const target = await makeAppRow();
+        const permission = `app-data:${target.uid}:kv:read`;
+        await post('handleGrantUserApp', {
+            app_uid: grantee.uid,
+            permission,
+        });
+
+        await server.stores.app.update(target.id, {
+            metadata: JSON.stringify({ share_app_data: false }),
+        });
+
+        // A user must always be able to withdraw consent, whatever the target
+        // now says about sharing.
+        await post('handleRevokeUserApp', { app_uid: grantee.uid, permission });
+        expect(await grantedPermissions(grantee.uid)).not.toContain(permission);
+    });
+
+    it("creates the target's AppData directory for an fs grant", async () => {
+        const grantee = await makeAppRow();
+        const target = await makeAppRow();
+        const path = `/${issuer.username}/AppData/${target.uid}`;
+        expect(await server.stores.fsEntry.getEntryByPath(path)).toBeFalsy();
+
+        await post('handleGrantUserApp', {
+            app_uid: grantee.uid,
+            permission: `app-data:${target.uid}:fs:read`,
+        });
+
+        // Without this the grant is valid but every read 404s until the target
+        // app happens to run for the first time.
+        expect(await server.stores.fsEntry.getEntryByPath(path)).toBeTruthy();
+    });
+
+    it('leaves unrelated permissions untouched by the new validation', async () => {
+        const grantee = await makeAppRow();
+        const permission = 'service:unrelated:ii:read';
+        await post('handleGrantUserApp', {
+            app_uid: grantee.uid,
+            permission,
+        });
+        expect(await grantedPermissions(grantee.uid)).toContain(permission);
+    });
+});

src/backend/controllers/auth/AuthController.ts

@@ -27,6 +27,7 @@ import type { HttpErrorOptions } from '../../core/http/HttpError.js';
 import { HttpError } from '../../core/http/HttpError.js';
 import { antiCsrf } from '../../core/http/middleware/antiCsrf.js';
 import { generateCaptcha } from '../../core/http/middleware/captcha.js';
+import type { Actor } from '../../core/actor.js';
 import { checkRateLimit } from '../../core/http/middleware/rateLimit.js';
 import {
     signStepUpToken,
@@ -60,11 +61,19 @@ import {
     generateDefaultFsentries,
     promoteToVerifiedGroup,
 } from '../../util/userProvisioning.js';
+import {
+    APP_DATA_PERMISSION_PREFIX,
+    appDataSharingAllowed,
+    parseAppDataPermission,
+} from '../../services/permission/appDataScopes.js';
 import { PuterController } from '../types.js';
 
 const USERNAME_REGEX = /^\w{1,}$/;
 const USERNAME_MAX_LENGTH = 45;
 const FINGERPRINT_MAX_LENGTH = 128;
+// One consent prompt covers a handful of scopes at most. The cap keeps a
+// crafted request from turning a single grant call into a bulk write.
+const MAX_PERMISSIONS_PER_REQUEST = 16;
 const DISPATCH_ID_MAX_LENGTH = 128;
 // Default SMS send attempts before the card fallback opens.
 const DEFAULT_CARD_FALLBACK_ATTEMPTS = 2;
@@ -74,6 +83,94 @@ const DEFAULT_CARD_FALLBACK_ATTEMPTS = 2;
 // crossed.
 const SEND_PHONE_RATE_LIMIT = 10;
 const SEND_PHONE_RATE_WINDOW_MS = 60 * 60_000;
+
+// -- Post-login route limits -----------------------------------------
+//
+// The credential legs above (login, signup, recovery, confirmation) each
+// carry their own limit. Everything a session can reach *after* signing
+// in shares the four shapes below, keyed on the actor rather than the
+// network — a per-account ceiling is the meaningful one once we know who
+// is calling.
+
+/**
+ * Mints or reconfigures a credential. Deliberately an hour-scale window: these
+ * are human actions taken a handful of times, and an unbounded rate turns one
+ * compromised session into a durable foothold.
+ */
+const CREDENTIAL_MINT_LIMIT = {
+    scope: 'auth-credential-mint',
+    limit: 20,
+    window: 60 * 60_000,
+    key: 'user',
+} as const;
+
+/**
+ * Second-factor configuration, including the verify leg. Shorter window than
+ * the mint limit because enabling 2FA legitimately involves a few attempts in a
+ * row, but unbounded verification is a TOTP brute force.
+ */
+const TWO_FACTOR_LIMIT = {
+    scope: 'auth-2fa-configure',
+    limit: 30,
+    window: 15 * 60_000,
+    key: 'user',
+} as const;
+
+/** Permission and membership writes. Never called in a loop by a client. */
+const GRANT_LIMIT = {
+    scope: 'auth-grant',
+    limit: 60,
+    window: 60_000,
+    key: 'user',
+} as const;
+
+/**
+ * Read-only checks the GUI makes on nearly every interaction. The ceiling is
+ * high enough that only a runaway loop reaches it.
+ */
+const AUTH_CHECK_LIMIT = {
+    scope: 'auth-check',
+    limit: 300,
+    window: 60_000,
+    key: 'user',
+} as const;
+
+/**
+ * Anti-CSRF token issuance. Clients mint a fresh token per protected mutation
+ * and cache nothing, so this ceiling has to clear the SUM of the budgets that
+ * spend tokens — matching any single one of them guarantees the gate fires
+ * before the mutation it guards does.
+ *
+ * What spends them: the session-authenticated download path, one token per
+ * file, at the read budget of 600/min — a multi-selection download burns tokens
+ * exactly the way a bulk delete burns its own budget, so this tracks the bulk
+ * figure used for filesystem mutations; logout at 60/min; and the
+ * session-management writes (revoke, rename), which a person triggers a handful
+ * of times. Call it ~700/min of real demand, and leave enough on top that a
+ * bulk operation runs out of files before it runs out of tokens.
+ */
+const ANTI_CSRF_MINT_LIMIT = {
+    scope: 'anticsrf',
+    limit: 1200,
+    window: 60_000,
+    key: 'user',
+} as const;
+
+/** Settings-page reads — enumerating sessions, permissions, groups. */
+const AUTH_LIST_LIMIT = {
+    scope: 'auth-list',
+    limit: 120,
+    window: 60_000,
+    key: 'user',
+} as const;
+
+/** Session plumbing: logout, GUI token, cookie sync. */
+const SESSION_LIMIT = {
+    scope: 'auth-session',
+    limit: 60,
+    window: 60_000,
+    key: 'user',
+} as const;
 // Once the threshold is crossed the fallback stays open this long, so the
 // user can finish the card flow without racing the attempt counter's expiry.
 const CARD_FALLBACK_OPEN_TTL_SECONDS = 24 * 60 * 60;
@@ -1004,7 +1101,8 @@ export class AuthController extends PuterController {
                     is_temp: user!.password === null && user!.email === null,
                     ip:
                         (req?.headers?.['x-forwarded-for'] as
-                            string | undefined) ||
+                            | string
+                            | undefined) ||
                         (
                             req as unknown as {
                                 connection?: { remoteAddress?: string };
@@ -1040,6 +1138,7 @@ export class AuthController extends PuterController {
         requireUserActor: true,
         allowUnconfirmed: true,
         antiCsrf: true,
+        rateLimit: SESSION_LIMIT,
     })
     async handleLogout(req: Request, res: Response): Promise<void> {
         // Clear the session cookie + `puter_token_v2`. Nothing issues the
@@ -2703,7 +2802,25 @@ export class AuthController extends PuterController {
 
     // -- Captcha generation -------------------------------------------
 
-    @Get('/api/captcha/generate', { subdomain: '*' })
+    @Get('/api/captcha/generate', {
+        subdomain: '*',
+        // Unauthenticated, renders an image per call, and is the gate
+        // protecting /login and /signup — so bulk pre-generation is
+        // directly useful to an attacker. Per-fingerprint for fairness on
+        // shared IPs, plus a per-IP backstop against header rotation.
+        //
+        // The fingerprint bucket is the one sized for a person: a handful of
+        // refreshes while getting a captcha right. The IP bucket is not — one
+        // address is a whole office, campus or carrier gateway, and everyone
+        // behind it is signing in through the same counter, so sizing it for
+        // a browser would deny the captcha to a network rather than to an
+        // attacker. It stays wide enough for that population and narrow
+        // enough that header rotation still runs out.
+        rateLimit: [
+            { scope: 'captcha', limit: 30, window: 60_000 },
+            { scope: 'captcha-ip', limit: 3_000, window: 60_000, key: 'ip' },
+        ],
+    })
     async handleCaptchaGenerate(_req: Request, res: Response): Promise<void> {
         const difficulty =
             (this.config as { captcha?: { difficulty?: string } }).captcha
@@ -2715,6 +2832,7 @@ export class AuthController extends PuterController {
     // -- Anti-CSRF token generation ----------------------------------
 
     @Get('/get-anticsrf-token', {
+        rateLimit: ANTI_CSRF_MINT_LIMIT,
         // Anti-CSRF tokens are only consumed by `requireUserActor` routes,
         // so issuance is scoped to the same actor kind for consistency.
         requireUserActor: true,
@@ -2735,6 +2853,7 @@ export class AuthController extends PuterController {
     @Post('/auth/grant-user-user', {
         subdomain: 'api',
         requireUserActor: true,
+        rateLimit: GRANT_LIMIT,
     })
     async handleGrantUserUser(req: Request, res: Response): Promise<void> {
         const { target_username, permission, extra, meta } = req.body;
@@ -2827,13 +2946,114 @@ export class AuthController extends PuterController {
         return app.uid;
     }
 
+    /**
+     * Resolve the `permission` / `permissions` pair into the list to act on.
+     *
+     * One consent prompt can cover several scopes (read a store, write
+     * another), and a client looping the single form would have to invent its
+     * own partial-failure and rollback handling. Accepting the array keeps that
+     * in one request.
+     */
+    #appPermissionList(body: {
+        permission?: unknown;
+        permissions?: unknown;
+    }): string[] {
+        const { permission, permissions } = body;
+        if (permissions !== undefined && permissions !== null) {
+            if (permission !== undefined && permission !== null) {
+                throw new HttpError(
+                    400,
+                    'Pass `permission` or `permissions`, not both',
+                    { legacyCode: 'bad_request' },
+                );
+            }
+            if (!Array.isArray(permissions) || permissions.length === 0) {
+                throw new HttpError(400, 'Invalid `permissions`', {
+                    legacyCode: 'bad_request',
+                });
+            }
+            if (permissions.length > MAX_PERMISSIONS_PER_REQUEST) {
+                throw new HttpError(400, 'Too many `permissions`', {
+                    legacyCode: 'bad_request',
+                });
+            }
+            for (const entry of permissions) {
+                this.#validateAppPermissionParams({ permission: entry });
+                // `*` means "revoke everything" in the scalar form only —
+                // inside a list it would silently widen a targeted request.
+                if (!entry || entry === '*') {
+                    throw new HttpError(400, 'Invalid `permissions`', {
+                        legacyCode: 'bad_request',
+                    });
+                }
+            }
+            return [...new Set(permissions as string[])];
+        }
+        return typeof permission === 'string' && permission ? [permission] : [];
+    }
+
+    /**
+     * Gate a cross-app data grant: the target must exist, must not have opted
+     * out of sharing, and must be named. Also creates the target's AppData
+     * directory for an `fs` scope, since it is only created lazily when the app
+     * first runs — without this a valid grant would 404 until then.
+     */
+    async #prepareAppDataGrant(
+        actor: Actor,
+        permission: string,
+    ): Promise<void> {
+        const parsed = parseAppDataPermission(permission);
+        if (!parsed) {
+            // A bare `app-data` (or one with an empty target) would cover every
+            // app the user has by prefix implication, which no prompt can
+            // describe. Reject rather than treat it as an unrelated permission.
+            if (
+                permission === APP_DATA_PERMISSION_PREFIX ||
+                permission.startsWith(`${APP_DATA_PERMISSION_PREFIX}:`)
+            ) {
+                throw new HttpError(
+                    400,
+                    'Invalid `app-data` permission: missing target app',
+                    { legacyCode: 'bad_request' },
+                );
+            }
+            return;
+        }
+
+        const target = await this.stores.app.getByUid(parsed.targetAppUid);
+        if (!target) {
+            throw new HttpError(
+                404,
+                `entity_not_found: app:${parsed.targetAppUid}`,
+                { legacyCode: 'subject_does_not_exist' },
+            );
+        }
+        if (!appDataSharingAllowed(target)) {
+            throw new HttpError(
+                403,
+                'This app does not share its data with other apps',
+                { legacyCode: 'forbidden' },
+            );
+        }
+
+        const username = actor.user?.username;
+        const userId = actor.user?.id;
+        if ((parsed.store === 'fs' || !parsed.store) && username && userId) {
+            await this.services.fs.mkdir(userId, {
+                path: `/${username}/AppData/${parsed.targetAppUid}`,
+                createMissingParents: true,
+            } as never);
+        }
+    }
+
     @Post('/auth/grant-user-app', {
         subdomain: 'api',
         requireUserActor: true,
+        rateLimit: GRANT_LIMIT,
     })
     async handleGrantUserApp(req: Request, res: Response): Promise<void> {
         let { app_uid } = req.body;
-        const { origin, permission, extra, meta } = req.body;
+        const { origin, permission, permissions, extra, meta } = req.body;
         this.#validateAppPermissionParams({
             app_uid,
             origin,
@@ -2841,27 +3061,43 @@ export class AuthController extends PuterController {
             extra,
             meta,
         });
+        const list = this.#appPermissionList({ permission, permissions });
         if (origin) {
             app_uid = await this.#registeredAppUidFromOrigin(origin);
         }
-        if (!app_uid || !permission) {
+        if (!app_uid || list.length === 0) {
             throw new HttpError(400, 'Missing `app_uid` or `permission`', {
                 legacyCode: 'bad_request',
             });
         }
-        await this.services.permission.grantUserAppPermission(
-            req.actor!,
-            app_uid,
-            permission,
-            extra ?? undefined,
-            meta ?? undefined,
-        );
+
+        // Validate every entry before writing any, so a bad one in the list
+        // cannot leave a partially-granted set behind: the dialog reads a 4xx as
+        // "nothing was written" and skips its withdrawal, so a partial commit
+        // leaves live access the user was told they refused. The rewrite running
+        // twice is cheaper than splitting the grant into prepare/commit.
+        for (const entry of list) {
+            await this.services.permission.assertUserAppPermissionWritable(
+                entry,
+            );
+            await this.#prepareAppDataGrant(req.actor!, entry);
+        }
+        for (const entry of list) {
+            await this.services.permission.grantUserAppPermission(
+                req.actor!,
+                app_uid,
+                entry,
+                extra ?? undefined,
+                meta ?? undefined,
+            );
+        }
         res.json({});
     }
 
     @Post('/auth/grant-user-group', {
         subdomain: 'api',
         requireUserActor: true,
+        rateLimit: GRANT_LIMIT,
     })
     async handleGrantUserGroup(req: Request, res: Response): Promise<void> {
         const { group_uid, permission, extra, meta } = req.body;
@@ -2890,6 +3126,7 @@ export class AuthController extends PuterController {
     @Post('/auth/revoke-user-user', {
         subdomain: 'api',
         requireUserActor: true,
+        rateLimit: GRANT_LIMIT,
     })
     async handleRevokeUserUser(req: Request, res: Response): Promise<void> {
         const { target_username, permission, meta } = req.body;
@@ -2912,24 +3149,28 @@ export class AuthController extends PuterController {
     @Post('/auth/revoke-user-app', {
         subdomain: 'api',
         requireUserActor: true,
+        rateLimit: GRANT_LIMIT,
     })
     async handleRevokeUserApp(req: Request, res: Response): Promise<void> {
         let { app_uid } = req.body;
-        const { origin, permission, meta } = req.body;
+        const { origin, permission, permissions, meta } = req.body;
         this.#validateAppPermissionParams({
             app_uid,
             origin,
             permission,
             meta,
         });
+        const list = this.#appPermissionList({ permission, permissions });
         if (origin) {
             app_uid = await this.#registeredAppUidFromOrigin(origin);
         }
-        if (!app_uid || !permission) {
+        if (!app_uid || list.length === 0) {
             throw new HttpError(400, 'Missing `app_uid` or `permission`', {
                 legacyCode: 'bad_request',
             });
         }
+        // Deliberately not gated by the target's sharing flag: a user must
+        // always be able to withdraw a grant, whatever the target now says.
         if (permission === '*') {
             await this.services.permission.revokeUserAppAll(
                 req.actor!,
@@ -2937,12 +3178,14 @@ export class AuthController extends PuterController {
                 meta ?? undefined,
             );
         } else {
-            await this.services.permission.revokeUserAppPermission(
-                req.actor!,
-                app_uid,
-                permission,
-                meta ?? undefined,
-            );
+            for (const entry of list) {
+                await this.services.permission.revokeUserAppPermission(
+                    req.actor!,
+                    app_uid,
+                    entry,
+                    meta ?? undefined,
+                );
+            }
         }
         res.json({});
     }
@@ -2950,6 +3193,7 @@ export class AuthController extends PuterController {
     @Post('/auth/revoke-user-group', {
         subdomain: 'api',
         requireUserActor: true,
+        rateLimit: GRANT_LIMIT,
     })
     async handleRevokeUserGroup(req: Request, res: Response): Promise<void> {
         const { group_uid, permission, meta } = req.body;
@@ -2969,7 +3213,11 @@ export class AuthController extends PuterController {
 
     // -- Permission checks -------------------------------------------
 
-    @Post('/auth/check-permissions', { subdomain: 'api', requireAuth: true })
+    @Post('/auth/check-permissions', {
+        subdomain: 'api',
+        requireAuth: true,
+        rateLimit: AUTH_CHECK_LIMIT,
+    })
     async handleCheckPermissions(req: Request, res: Response): Promise<void> {
         const { permissions } = req.body;
         if (!Array.isArray(permissions)) {
@@ -2997,7 +3245,11 @@ export class AuthController extends PuterController {
 
     // -- Session management ------------------------------------------
 
-    @Get('/auth/list-sessions', { subdomain: 'api', requireUserActor: true })
+    @Get('/auth/list-sessions', {
+        subdomain: 'api',
+        requireUserActor: true,
+        rateLimit: AUTH_LIST_LIMIT,
+    })
     async handleListSessions(req: Request, res: Response): Promise<void> {
         const sessions = await this.services.auth.listSessions(req.actor!);
         res.json(sessions);
@@ -3079,7 +3331,11 @@ export class AuthController extends PuterController {
 
     // -- Dev app permissions -----------------------------------------
 
-    @Post('/auth/grant-dev-app', { subdomain: 'api', requireUserActor: true })
+    @Post('/auth/grant-dev-app', {
+        subdomain: 'api',
+        requireUserActor: true,
+        rateLimit: GRANT_LIMIT,
+    })
     async handleGrantDevApp(req: Request, res: Response): Promise<void> {
         let { app_uid } = req.body;
         const { origin, permission, extra, meta } = req.body;
@@ -3112,6 +3368,7 @@ export class AuthController extends PuterController {
     @Post('/auth/revoke-dev-app', {
         subdomain: 'api',
         requireUserActor: true,
+        rateLimit: GRANT_LIMIT,
     })
     async handleRevokeDevApp(req: Request, res: Response): Promise<void> {
         let { app_uid } = req.body;
@@ -3147,6 +3404,7 @@ export class AuthController extends PuterController {
     @Get('/auth/list-permissions', {
         subdomain: 'api',
         requireUserActor: true,
+        rateLimit: AUTH_LIST_LIMIT,
     })
     async handleListPermissions(req: Request, res: Response): Promise<void> {
         const userId = req.actor!.user.id;
@@ -3211,7 +3469,11 @@ export class AuthController extends PuterController {
 
     // -- App origin resolution ---------------------------------------
 
-    @Post('/auth/app-uid-from-origin', { subdomain: 'api', requireAuth: true })
+    @Post('/auth/app-uid-from-origin', {
+        subdomain: 'api',
+        requireAuth: true,
+        rateLimit: AUTH_CHECK_LIMIT,
+    })
     async handleAppUidFromOrigin(req: Request, res: Response): Promise<void> {
         const origin = req.body?.origin || req.query?.origin;
         if (!origin)
@@ -3227,6 +3489,9 @@ export class AuthController extends PuterController {
     @Post('/auth/get-user-app-token', {
         subdomain: 'api',
         requireUserActor: true,
+        // Called once per app launch, and the GUI can legitimately launch
+        // several in quick succession.
+        rateLimit: { ...AUTH_CHECK_LIMIT, scope: 'app-token', limit: 120 },
     })
     async handleGetUserAppToken(req: Request, res: Response): Promise<void> {
         let { app_uid } = req.body;
@@ -3250,6 +3515,22 @@ export class AuthController extends PuterController {
             app = await this.stores.app.createFromOrigin(app_uid, origin, {
                 ownerUserId,
             });
+            // An origin's uid is a deterministic uuidv5, so a deleted app
+            // reappears here under the identical uid. Withdraw any cross-app
+            // data grants left pointing at it before this new row can inherit
+            // consent the user gave its predecessor. Only *this* path can reuse
+            // a uid: `AppStore.create` mints a random uuid4, which no deleted
+            // app can ever hold again.
+            //
+            // Called directly rather than through `app.changed`: the token is
+            // issued below, so this has to be able to stop that, and
+            // `emitAndWait` swallows listener errors. Letting it throw is the
+            // point — a sweep that failed leaves the old grants live against an
+            // app whoever controls the origin now has just claimed.
+            await this.services.appPermission.withdrawAppDataGrants(
+                app_uid,
+                'uid reused by a new app',
+            );
         }
         if (!app) {
             throw new HttpError(404, `App ${app_uid} does not exist`, {
@@ -3329,7 +3610,11 @@ export class AuthController extends PuterController {
         res.json({ token, app_uid });
     }
 
-    @Post('/auth/check-app', { subdomain: 'api', requireUserActor: true })
+    @Post('/auth/check-app', {
+        subdomain: 'api',
+        requireUserActor: true,
+        rateLimit: AUTH_CHECK_LIMIT,
+    })
     async handleCheckApp(req: Request, res: Response): Promise<void> {
         let { app_uid } = req.body;
         const { origin } = req.body;
@@ -3368,6 +3653,7 @@ export class AuthController extends PuterController {
     @Post('/auth/create-access-token', {
         subdomain: 'api',
         requireAuth: true,
+        rateLimit: CREDENTIAL_MINT_LIMIT,
     })
     async handleCreateAccessToken(req: Request, res: Response): Promise<void> {
         const { permissions, expiresIn, label } = req.body;
@@ -3436,6 +3722,7 @@ export class AuthController extends PuterController {
     @Post('/auth/configure-2fa/:action', {
         subdomain: 'api',
         requireUserActor: true,
+        rateLimit: TWO_FACTOR_LIMIT,
     })
     async handleConfigure2fa(req: Request, res: Response): Promise<void> {
         const action = req.params.action;
@@ -3566,7 +3853,11 @@ export class AuthController extends PuterController {
 
     // -- Developer profile -------------------------------------------
 
-    @Get('/get-dev-profile', { subdomain: 'api', requireUserActor: true })
+    @Get('/get-dev-profile', {
+        subdomain: 'api',
+        requireUserActor: true,
+        rateLimit: AUTH_LIST_LIMIT,
+    })
     async handleGetDevProfile(req: Request, res: Response): Promise<void> {
         const user = await this.stores.user.getById(req.actor!.user.id!, {
             force: true,
@@ -3596,7 +3887,13 @@ export class AuthController extends PuterController {
 
     // -- Group management --------------------------------------------
 
-    @Post('/group/create', { subdomain: 'api', requireUserActor: true })
+    @Post('/group/create', {
+        subdomain: 'api',
+        requireUserActor: true,
+        // Creates a persistent row per call with no quota behind it, so it
+        // sits on the hour-scale budget rather than the grant one.
+        rateLimit: { ...CREDENTIAL_MINT_LIMIT, scope: 'group-create' },
+    })
     async handleGroupCreate(req: Request, res: Response): Promise<void> {
         const extra = req.body.extra ?? {};
         const metadata = req.body.metadata ?? {};
@@ -3617,7 +3914,11 @@ export class AuthController extends PuterController {
         res.json({ uid });
     }
 
-    @Post('/group/add-users', { subdomain: 'api', requireUserActor: true })
+    @Post('/group/add-users', {
+        subdomain: 'api',
+        requireUserActor: true,
+        rateLimit: GRANT_LIMIT,
+    })
     async handleGroupAddUsers(req: Request, res: Response): Promise<void> {
         const { uid, users } = req.body ?? {};
         if (!uid)
@@ -3649,7 +3950,11 @@ export class AuthController extends PuterController {
         res.json({});
     }
 
-    @Post('/group/remove-users', { subdomain: 'api', requireUserActor: true })
+    @Post('/group/remove-users', {
+        subdomain: 'api',
+        requireUserActor: true,
+        rateLimit: GRANT_LIMIT,
+    })
     async handleGroupRemoveUsers(req: Request, res: Response): Promise<void> {
         const { uid, users } = req.body ?? {};
         if (!uid)
@@ -3681,7 +3986,11 @@ export class AuthController extends PuterController {
         res.json({});
     }
 
-    @Get('/group/list', { subdomain: 'api', requireUserActor: true })
+    @Get('/group/list', {
+        subdomain: 'api',
+        requireUserActor: true,
+        rateLimit: AUTH_LIST_LIMIT,
+    })
     async handleGroupList(req: Request, res: Response): Promise<void> {
         const userId = req.actor!.user.id!;
         const [owned, member] = await Promise.all([
@@ -3694,7 +4003,22 @@ export class AuthController extends PuterController {
         });
     }
 
-    @Get('/group/public-groups', { subdomain: 'api' })
+    @Get('/group/public-groups', {
+        subdomain: 'api',
+        // The only unauthenticated route in the group set, so IP is the
+        // only key available — and that makes the bucket an aggregate:
+        // one office, campus or carrier gateway is a single key for
+        // everybody behind it, and each of them reads this once while
+        // bootstrapping. Sized for that population of real people rather
+        // than one browser, and no wider: this sits next to the sign-in
+        // surface, so it stays a real bound on enumeration.
+        rateLimit: {
+            scope: 'public-groups',
+            limit: 1_200,
+            window: 60_000,
+            key: 'ip',
+        },
+    })
     async handleGroupPublicGroups(_req: Request, res: Response): Promise<void> {
         res.json({
             user: this.config.default_user_group ?? null,
@@ -3707,6 +4031,7 @@ export class AuthController extends PuterController {
     @Get('/get-gui-token', {
         requireUserActor: true,
         allowUnconfirmed: true,
+        rateLimit: SESSION_LIMIT,
     })
     async handleGetGuiToken(req: Request, res: Response): Promise<void> {
         if (!req.actor?.session?.uid)
@@ -3726,6 +4051,7 @@ export class AuthController extends PuterController {
     }
 
     @Get('/session/sync-cookie', {
+        rateLimit: SESSION_LIMIT,
         // Installs the session cookie. Only page script on our own origin
         // should be able to ask for that (the `tokenSource` check below is the
         // companion rule: the token has to come from an Authorization header,

src/backend/controllers/desktop/DesktopController.js

@@ -32,6 +32,19 @@ const ALLOWED_SORT_ORDER = ['asc', 'desc'];
  * - User-level: desktop background, taskbar items (UserStore)
  * - Folder-level: layout, sort_by/sort_order (fsentries table)
  */
+/**
+ * Desktop preference writes — background, taskbar, layout, sort order. All four
+ * persist to the user row on every call, and the GUI fires them on direct user
+ * action, so a per-minute ceiling well above human speed is enough to catch a
+ * stuck client.
+ */
+const PREFERENCE_WRITE_LIMIT = {
+    scope: 'desktop-preference',
+    limit: 120,
+    window: 60_000,
+    key: 'user',
+};
+
 export class DesktopController extends PuterController {
     constructor(config, clients, stores, services) {
         super(config, clients, stores, services);
@@ -53,6 +66,7 @@ export class DesktopController extends PuterController {
                 subdomain: 'api',
                 requireUserActor: true,
                 allowFullAccessToken: true,
+                rateLimit: PREFERENCE_WRITE_LIMIT,
             },
             async (req, res) => {
                 const { url, color, fit } = req.body ?? {};
@@ -108,6 +122,7 @@ export class DesktopController extends PuterController {
                 subdomain: 'api',
                 requireUserActor: true,
                 allowFullAccessToken: true,
+                rateLimit: PREFERENCE_WRITE_LIMIT,
             },
             async (req, res) => {
                 const { items } = req.body ?? {};
@@ -134,6 +149,7 @@ export class DesktopController extends PuterController {
                 subdomain: 'api',
                 requireUserActor: true,
                 allowFullAccessToken: true,
+                rateLimit: PREFERENCE_WRITE_LIMIT,
             },
             async (req, res) => {
                 const { item_uid, item_path, layout } = req.body ?? {};
@@ -161,6 +177,7 @@ export class DesktopController extends PuterController {
                 subdomain: 'api',
                 requireUserActor: true,
                 allowFullAccessToken: true,
+                rateLimit: PREFERENCE_WRITE_LIMIT,
             },
             async (req, res) => {
                 const { item_uid, item_path, sort_by, sort_order } =

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

@@ -124,9 +124,9 @@ const makeSyntheticDriver = () => ({
 
 const buildController = (driver: ReturnType<typeof makeSyntheticDriver>) => {
     // The controller reads `this.services?.permission` only when an actor
-    // is on the request; otherwise the services bag is unused. The
-    // alarm client is invoked on the rate-limit / concurrency rejection
-    // paths, so stub a no-op `create`.
+    // is on the request; otherwise the services bag is unused. The rejection
+    // paths no longer alarm, but other paths still reach the client, so keep
+    // a no-op `create` stubbed.
     const clients = { alarm: { create: () => {} } };
     return new DriverController(
         {} as any,

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

@@ -409,7 +409,7 @@ describe('DriverController per-method rate limiting', () => {
         configureRateLimit({ disabled: false } as never);
     });
 
-    it('answers 429 and raises a de-duped info alarm once the per-method budget is spent', async () => {
+    it('answers 429 without alarming once the per-method budget is spent', async () => {
         const { handler, alarms, iface } = build({
             run: () => ({ ok: true }),
             rateLimit: { default: { limit: 1, window: 60_000 } },
@@ -430,9 +430,8 @@ describe('DriverController per-method rate limiting', () => {
             legacyCode: 'too_many_requests',
         });
 
-        expect(alarms).toContainEqual({
-            id: 'driver_rate_limit_hit:rate-limited-iface:run',
-            severity: 'info',
-        });
+        // Spending your own budget is the limit working as designed, so it
+        // must not raise anything — the 429 is the whole signal.
+        expect(alarms).toEqual([]);
     });
 });

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

@@ -158,11 +158,14 @@ interface MockRes {
     sentBody: string | undefined;
     contentType: string | undefined;
     pipedFrom: Readable | undefined;
+    listeners: Record<string, Array<() => void>>;
     status(code: number): MockRes;
     json(body: unknown): MockRes;
     setHeader(key: string, value: string): MockRes;
     type(t: string): MockRes;
     send(body: string): MockRes;
+    once(event: string, fn: () => void): MockRes;
+    emit(event: string): void;
 }
 const makeRes = (): MockRes => {
     const res: MockRes = {
@@ -172,6 +175,19 @@ const makeRes = (): MockRes => {
         sentBody: undefined,
         contentType: undefined,
         pipedFrom: undefined,
+        // `#handleCall` releases a driver's concurrency slot on `finish` /
+        // `close`, so the stub has to behave like an emitter for any driver
+        // that declares a `concurrent` policy.
+        listeners: {},
+        once(event: string, fn: () => void) {
+            (this.listeners[event] ??= []).push(fn);
+            return this;
+        },
+        emit(event: string) {
+            const fns = this.listeners[event] ?? [];
+            this.listeners[event] = [];
+            for (const fn of fns) fn();
+        },
         status(code: number) {
             this.statusCode = code;
             return this;

src/backend/controllers/drivers/DriverController.ts

@@ -45,6 +45,27 @@ import { PuterController } from '../types.js';
 
 type DriverInstance = WithLifecycle & Record<string, unknown>;
 
+/**
+ * Coarse envelope over the whole `/call` surface, so that spreading calls
+ * across many interfaces can't dodge every individual bucket. Per-driver limits
+ * are what actually shape traffic.
+ *
+ * "Coarse" is a constraint, not a description: for this to be an envelope it
+ * has to sit _above_ every per-driver budget, or it silently becomes the real
+ * limit for the widest ones and overrides the tier policy they declare.
+ * `driverPolicies.test.ts` asserts that ordering against every registered
+ * driver, so raising a driver's budget past this number fails there rather than
+ * in production. The headroom above the widest driver (notifications, at
+ * 3000/30s) is what leaves room for one caller to be busy on two interfaces at
+ * once.
+ */
+export const DRIVERS_CALL_LIMIT = {
+    scope: 'drivers-call',
+    limit: 8000,
+    window: 60_000,
+    key: 'user' as const,
+};
+
 // Every driver call is timed here already, for the lifecycle events below.
 // Recording the same number as a histogram makes the per-interface latency
 // distribution available downstream; which interfaces are worth keeping is a
@@ -180,12 +201,26 @@ export class DriverController extends PuterController {
     registerRoutes(router: PuterRouter): void {
         router.post(
             '/call',
-            { subdomain: 'api', requireAuth: true },
+            {
+                subdomain: 'api',
+                requireAuth: true,
+                rateLimit: DRIVERS_CALL_LIMIT,
+            },
             this.#handleCall,
         );
         router.get(
             '/list-interfaces',
-            { subdomain: 'api', requireAuth: true },
+            {
+                subdomain: 'api',
+                requireAuth: true,
+                // Static introspection output, read once at boot.
+                rateLimit: {
+                    scope: 'drivers-list-interfaces',
+                    limit: 60,
+                    window: 60_000,
+                    key: 'user',
+                },
+            },
             this.#handleListInterfaces,
         );
     }
@@ -305,19 +340,8 @@ export class DriverController extends PuterController {
         if (
             !(await checkDriverRateLimit(req, ifaceName, method, rateLimitSpec))
         ) {
-            // De-dupe on (iface, method) so a hot loop across many users
-            // aggregates as occurrences on a single low-severity alarm
-            // instead of fanning out one per user.
-            this.clients.alarm.create(
-                `driver_rate_limit_hit:${ifaceName}:${method}`,
-                `Driver rate limit hit on ${ifaceName}:${method}`,
-                {
-                    iface: ifaceName,
-                    method,
-                    userUuid: req.actor?.user?.uuid,
-                },
-                'info',
-            );
+            // Deliberately unalarmed: a caller spending its own budget is
+            // the limit working, not an incident. The 429 is the signal.
             throw new HttpError(429, 'Too many requests.', {
                 legacyCode: 'too_many_requests',
             });
@@ -340,16 +364,8 @@ export class DriverController extends PuterController {
                 concurrentSpec,
             );
             if (!handle.ok) {
-                this.clients.alarm.create(
-                    `driver_concurrent_limit_hit:${ifaceName}:${method}`,
-                    `Driver concurrency limit hit on ${ifaceName}:${method}`,
-                    {
-                        iface: ifaceName,
-                        method,
-                        userUuid: req.actor?.user?.uuid,
-                    },
-                    'info',
-                );
+                // Unalarmed for the same reason as the rate-limit rejection
+                // above: hitting a declared cap is the cap doing its job.
                 throw new HttpError(429, 'Too many concurrent requests.', {
                     legacyCode: 'too_many_requests',
                 });

src/backend/controllers/feedback/AppFeedbackController.test.ts

@@ -0,0 +1,594 @@
+/*
+ * 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 { Request, RequestHandler, Response } from 'express';
+import {
+    afterAll,
+    afterEach,
+    beforeAll,
+    describe,
+    expect,
+    it,
+    vi,
+} from 'vitest';
+import { v4 as uuidv4 } from 'uuid';
+import type { Actor } from '../../core/actor.js';
+import { PuterRouter } from '../../core/http/PuterRouter.js';
+import { PuterServer } from '../../server.js';
+import { AppFeedbackService } from '../../services/feedback/AppFeedbackService.js';
+import { setupTestServer } from '../../testUtil.js';
+
+// Boots one real PuterServer (in-memory sqlite + mocked externals) and
+// registers AppFeedbackController's decorated routes onto a fresh
+// PuterRouter. Tests drive the captured handlers with stub req/res; the
+// stores/services underneath are the live wired ones, so rows land in the
+// real `app_feedback` table.
+
+let server: PuterServer;
+let router: PuterRouter;
+
+beforeAll(async () => {
+    server = await setupTestServer();
+    router = new PuterRouter();
+    server.controllers.appFeedback.registerRoutes(router);
+});
+
+afterAll(async () => {
+    await server?.shutdown();
+});
+
+afterEach(() => {
+    vi.restoreAllMocks();
+});
+
+const makeUser = async (): Promise<{ actor: Actor; userId: number }> => {
+    const username = `fdbk-${Math.random().toString(36).slice(2, 10)}`;
+    const created = await server.stores.user.create({
+        username,
+        uuid: uuidv4(),
+        password: null,
+        email: `${username}@test.local`,
+        free_storage: 100 * 1024 * 1024,
+        requires_email_confirmation: false,
+    });
+    const refreshed = (await server.stores.user.getById(created.id))!;
+    return {
+        userId: refreshed.id,
+        actor: {
+            user: {
+                id: refreshed.id,
+                uuid: refreshed.uuid,
+                username: refreshed.username,
+                email: refreshed.email ?? null,
+                email_confirmed: true,
+            } as Actor['user'],
+        },
+    };
+};
+
+const makeApp = async (
+    ownerUserId: number,
+    opts: { feedbackEnabled?: boolean; indexUrl?: string; name?: string } = {},
+) => {
+    const name =
+        opts.name ?? `fdbk-app-${Math.random().toString(36).slice(2, 10)}`;
+    return await server.stores.app.create(
+        {
+            name,
+            title: `Feedback Test ${name}`,
+            index_url: opts.indexUrl ?? `https://${name}.example.com`,
+            ...(opts.feedbackEnabled ? { feedback_enabled: 1 } : {}),
+        },
+        { ownerUserId },
+    );
+};
+
+// Feedback is only offered when the deployment can deliver it (email
+// transport configured); most tests want that baseline without asserting
+// anything about the mail itself.
+const mockEmailConfigured = () =>
+    vi.spyOn(server.clients.email, 'isConfigured', 'get').mockReturnValue(
+        true,
+    );
+
+const confirmOwnerEmail = async (userId: number) => {
+    await server.clients.db.write(
+        'UPDATE `user` SET `email_confirmed` = ? WHERE `id` = ?',
+        [server.clients.db.booleanValue(true), userId],
+    );
+    const user = await server.stores.user.getById(userId);
+    if (user) await server.stores.user.invalidate(user);
+};
+
+const makeReq = (init: {
+    body?: unknown;
+    actor?: Actor;
+    query?: Record<string, unknown>;
+}): Request => {
+    return {
+        body: init.body ?? {},
+        query: init.query ?? {},
+        headers: {},
+        actor: init.actor,
+    } as unknown as Request;
+};
+
+const makeRes = () => {
+    const captured: { statusCode: number; body: unknown } = {
+        statusCode: 200,
+        body: undefined,
+    };
+    const res = {
+        json: vi.fn((value: unknown) => {
+            captured.body = value;
+            return res;
+        }),
+        status: vi.fn((code: number) => {
+            captured.statusCode = code;
+            return res;
+        }),
+    };
+    return { res: res as unknown as Response, captured };
+};
+
+const findRoute = (method: string, path: string) => {
+    const route = router.routes.find(
+        (r) => r.method === method && r.path === path,
+    );
+    if (!route) throw new Error(`No ${method.toUpperCase()} ${path} route`);
+    return route;
+};
+
+const callRoute = async (
+    method: string,
+    path: string,
+    req: Request,
+    res: Response,
+) => {
+    const handler: RequestHandler = findRoute(method, path).handler;
+    await handler(req, res, () => {
+        throw new Error('handler called next() unexpectedly');
+    });
+};
+
+const submit = (actor: Actor, body: unknown) => {
+    const { res, captured } = makeRes();
+    return callRoute('post', '/', makeReq({ body, actor }), res).then(
+        () => captured,
+    );
+};
+
+// ── Route gates ─────────────────────────────────────────────────────
+
+describe('AppFeedbackController route options', () => {
+    it('rejects app actors and cross-origin pages on submit', () => {
+        const { options } = findRoute('post', '/');
+        // requireUserActor is what makes feedback impossible to submit
+        // programmatically with an app token; guiOriginOnly keeps
+        // cross-origin browser pages out even with a leaked user token.
+        expect(options.requireUserActor).toBe(true);
+        expect(options.guiOriginOnly).toBe(true);
+    });
+
+    it('stacks a per-user budget with a per-IP backstop', () => {
+        const { options } = findRoute('post', '/');
+        const limits = options.rateLimit;
+        expect(Array.isArray(limits)).toBe(true);
+        const keys = (limits as Array<{ key?: unknown }>).map((l) => l.key);
+        expect(keys).toContain('user');
+        expect(keys).toContain('ip');
+    });
+
+    it('requires a user actor on the target pre-flight too', () => {
+        const { options } = findRoute('get', '/target');
+        expect(options.requireUserActor).toBe(true);
+    });
+});
+
+// ── GET /app-feedback/target ────────────────────────────────────────
+
+describe('AppFeedbackController GET /target', () => {
+    it('throws 400 when neither or both of app/origin are given', async () => {
+        const { actor } = await makeUser();
+        for (const query of [
+            {},
+            { app: 'x', origin: 'https://x.example.com' },
+        ]) {
+            const { res } = makeRes();
+            await expect(
+                callRoute('get', '/target', makeReq({ query, actor }), res),
+            ).rejects.toMatchObject({ statusCode: 400 });
+        }
+    });
+
+    it('reports enabled:false for an unknown app', async () => {
+        const { actor } = await makeUser();
+        const { res, captured } = makeRes();
+        await callRoute(
+            'get',
+            '/target',
+            makeReq({ query: { app: 'no-such-app-xyz' }, actor }),
+            res,
+        );
+        expect(captured.body).toEqual({ enabled: false, app: null });
+    });
+
+    it('reports enabled:false for an app that has not opted in', async () => {
+        mockEmailConfigured();
+        const { userId: ownerId } = await makeUser();
+        const app = await makeApp(ownerId);
+        const { actor } = await makeUser();
+        const { res, captured } = makeRes();
+        await callRoute(
+            'get',
+            '/target',
+            makeReq({ query: { app: app.name }, actor }),
+            res,
+        );
+        expect(captured.body).toMatchObject({ enabled: false });
+    });
+
+    it('reports enabled:true with canonical title/name for an opted-in app', async () => {
+        mockEmailConfigured();
+        const { userId: ownerId } = await makeUser();
+        const app = await makeApp(ownerId, { feedbackEnabled: true });
+        const { actor } = await makeUser();
+        const { res, captured } = makeRes();
+        await callRoute(
+            'get',
+            '/target',
+            makeReq({ query: { app: app.uid }, actor }),
+            res,
+        );
+        expect(captured.body).toEqual({
+            enabled: true,
+            app: { name: app.name, title: app.title },
+        });
+    });
+
+    it('resolves an opted-in app whose name starts with "app-"', async () => {
+        mockEmailConfigured();
+        const { userId: ownerId } = await makeUser();
+        const name = `app-fdbk-${Math.random().toString(36).slice(2, 10)}`;
+        const app = await makeApp(ownerId, { feedbackEnabled: true, name });
+        const { actor } = await makeUser();
+        const { res, captured } = makeRes();
+        await callRoute(
+            'get',
+            '/target',
+            makeReq({ query: { app: name }, actor }),
+            res,
+        );
+        expect(captured.body).toEqual({
+            enabled: true,
+            app: { name: app.name, title: app.title },
+        });
+    });
+
+    it('resolves an origin to the app whose index_url it matches', async () => {
+        mockEmailConfigured();
+        const { userId: ownerId } = await makeUser();
+        const app = await makeApp(ownerId, { feedbackEnabled: true });
+        const origin = new URL(app.index_url).origin;
+        const { actor } = await makeUser();
+        const { res, captured } = makeRes();
+        await callRoute(
+            'get',
+            '/target',
+            makeReq({ query: { origin }, actor }),
+            res,
+        );
+        expect(captured.body).toEqual({
+            enabled: true,
+            app: { name: app.name, title: app.title },
+        });
+    });
+
+    it('reports enabled:false when the email transport is unconfigured', async () => {
+        // No mockEmailConfigured(): this is the self-hosted no-SMTP default.
+        // Feedback that can never be delivered must not be solicited.
+        const { userId: ownerId } = await makeUser();
+        const app = await makeApp(ownerId, { feedbackEnabled: true });
+        const { actor } = await makeUser();
+        const { res, captured } = makeRes();
+        await callRoute(
+            'get',
+            '/target',
+            makeReq({ query: { app: app.name }, actor }),
+            res,
+        );
+        expect(captured.body).toMatchObject({ enabled: false });
+    });
+
+    it('reports enabled:false for an origin with no registered app', async () => {
+        const { actor } = await makeUser();
+        const { res, captured } = makeRes();
+        await callRoute(
+            'get',
+            '/target',
+            makeReq({
+                query: { origin: 'https://nobody-registered.example.com' },
+                actor,
+            }),
+            res,
+        );
+        expect(captured.body).toEqual({ enabled: false, app: null });
+    });
+});
+
+// ── POST /app-feedback ──────────────────────────────────────────────
+
+describe('AppFeedbackController POST /', () => {
+    it('throws 400 when message is missing or not a string', async () => {
+        const { userId: ownerId } = await makeUser();
+        const app = await makeApp(ownerId, { feedbackEnabled: true });
+        const { actor } = await makeUser();
+        for (const message of [undefined, 12345, '']) {
+            await expect(
+                submit(actor, { app: app.name, message }),
+            ).rejects.toMatchObject({ statusCode: 400 });
+        }
+    });
+
+    it('throws 400 when both app and origin are given', async () => {
+        const { actor } = await makeUser();
+        await expect(
+            submit(actor, {
+                app: 'x',
+                origin: 'https://x.example.com',
+                message: 'hi',
+            }),
+        ).rejects.toMatchObject({ statusCode: 400 });
+    });
+
+    it('throws 400 when the origin exceeds the stored column size', async () => {
+        // source_origin is VARCHAR(2048) on MySQL/Postgres; a longer origin
+        // must be rejected up front, not fail (or silently truncate) at the
+        // INSERT after passing every other validation.
+        mockEmailConfigured();
+        const { userId: ownerId } = await makeUser();
+        const app = await makeApp(ownerId, { feedbackEnabled: true });
+        const origin = `${new URL(app.index_url).origin}/${'x'.repeat(2500)}`;
+        const { actor } = await makeUser();
+        await expect(
+            submit(actor, { origin, message: 'hi' }),
+        ).rejects.toMatchObject({ statusCode: 400 });
+    });
+
+    it('throws 403 feedback_not_enabled when the app has not opted in', async () => {
+        mockEmailConfigured();
+        const { userId: ownerId } = await makeUser();
+        const app = await makeApp(ownerId);
+        const { actor } = await makeUser();
+        await expect(
+            submit(actor, { app: app.name, message: 'hi there' }),
+        ).rejects.toMatchObject({
+            statusCode: 403,
+            legacyCode: 'feedback_not_enabled',
+        });
+    });
+
+    it('throws 403 for an unknown app and an unknown origin alike', async () => {
+        const { actor } = await makeUser();
+        await expect(
+            submit(actor, { app: 'no-such-app-xyz', message: 'hi' }),
+        ).rejects.toMatchObject({ statusCode: 403 });
+        await expect(
+            submit(actor, {
+                origin: 'https://nobody-registered.example.com',
+                message: 'hi',
+            }),
+        ).rejects.toMatchObject({ statusCode: 403 });
+    });
+
+    it('throws 403 feedback_not_enabled when the email transport is unconfigured', async () => {
+        // No mockEmailConfigured(): the opted-in app must still refuse — a
+        // stored row nothing can read, sold to the sender as delivered, is
+        // worse than an honest refusal.
+        const { userId: ownerId } = await makeUser();
+        const app = await makeApp(ownerId, { feedbackEnabled: true });
+        const { actor } = await makeUser();
+        await expect(
+            submit(actor, { app: app.name, message: 'into the void' }),
+        ).rejects.toMatchObject({
+            statusCode: 403,
+            legacyCode: 'feedback_not_enabled',
+        });
+    });
+
+    it('throws 400 when the message exceeds the length limit', async () => {
+        mockEmailConfigured();
+        const { userId: ownerId } = await makeUser();
+        const app = await makeApp(ownerId, { feedbackEnabled: true });
+        const { actor } = await makeUser();
+        await expect(
+            submit(actor, {
+                app: app.name,
+                message: 'x'.repeat(
+                    AppFeedbackService.MESSAGE_MAX_LENGTH + 1,
+                ),
+            }),
+        ).rejects.toMatchObject({ statusCode: 400 });
+    });
+
+    it('stores a normalized row and responds with an empty object', async () => {
+        mockEmailConfigured();
+        const { userId: ownerId } = await makeUser();
+        const app = await makeApp(ownerId, { feedbackEnabled: true });
+        const { actor, userId } = await makeUser();
+        const captured = await submit(actor, {
+            app: app.uid,
+            message: '  Great\r\napp!  ',
+            context: 'app',
+        });
+        expect(captured.body).toEqual({});
+
+        const rows = (await server.clients.db.read(
+            'SELECT * FROM `app_feedback` WHERE `user_id` = ?',
+            [userId],
+        )) as Array<Record<string, unknown>>;
+        expect(rows).toHaveLength(1);
+        expect(rows[0]).toMatchObject({
+            app_uid: app.uid,
+            message: 'Great\napp!',
+            source_env: 'app',
+            source_origin: null,
+        });
+        // Engine-agnostic reads: pg returns BIGINT as string and BOOLEAN as
+        // boolean, sqlite returns numbers for both.
+        expect(Number(rows[0].app_id)).toBe(app.id);
+        expect(Boolean(rows[0].email_sent)).toBe(false);
+        expect(typeof rows[0].uid).toBe('string');
+        expect(Number.isFinite(Number(rows[0].created_at))).toBe(true);
+    });
+
+    it('records the attested origin on web submissions', async () => {
+        mockEmailConfigured();
+        const { userId: ownerId } = await makeUser();
+        const app = await makeApp(ownerId, { feedbackEnabled: true });
+        const origin = new URL(app.index_url).origin;
+        const { actor, userId } = await makeUser();
+        await submit(actor, { origin, message: 'from the web', context: 'web' });
+
+        const rows = (await server.clients.db.read(
+            'SELECT `source_env`, `source_origin` FROM `app_feedback` WHERE `user_id` = ?',
+            [userId],
+        )) as Array<Record<string, unknown>>;
+        expect(rows).toHaveLength(1);
+        expect(rows[0]).toMatchObject({
+            source_env: 'web',
+            source_origin: origin,
+        });
+    });
+
+    it('enforces the per-user-per-app daily cap with 429', async () => {
+        mockEmailConfigured();
+        const { userId: ownerId } = await makeUser();
+        const app = await makeApp(ownerId, { feedbackEnabled: true });
+        const { actor, userId } = await makeUser();
+        const since = Math.floor(Date.now() / 1000);
+        for (let i = 0; i < AppFeedbackService.PER_USER_APP_DAILY_LIMIT; i++) {
+            await server.stores.appFeedback.create({
+                appId: app.id,
+                appUid: app.uid,
+                userId,
+                message: `seed ${i}`,
+            });
+        }
+        expect(
+            await server.stores.appFeedback.countByUserAndAppSince(
+                userId,
+                app.id,
+                since - 60,
+            ),
+        ).toBe(AppFeedbackService.PER_USER_APP_DAILY_LIMIT);
+        await expect(
+            submit(actor, { app: app.name, message: 'one too many' }),
+        ).rejects.toMatchObject({
+            statusCode: 429,
+            legacyCode: 'too_many_requests',
+        });
+    });
+
+    it('rolls back the stored row when a concurrent burst breaches the cap', async () => {
+        mockEmailConfigured();
+        const { userId: ownerId } = await makeUser();
+        const app = await makeApp(ownerId, { feedbackEnabled: true });
+        const { actor, userId } = await makeUser();
+        for (let i = 0; i < AppFeedbackService.PER_USER_APP_DAILY_LIMIT; i++) {
+            await server.stores.appFeedback.create({
+                appId: app.id,
+                appUid: app.uid,
+                userId,
+                message: `seed ${i}`,
+            });
+        }
+        // Simulate the losing side of the check-then-insert race: the
+        // pre-insert check reads a stale under-cap count; the post-insert
+        // recount (real implementation) sees the truth.
+        vi.spyOn(
+            server.stores.appFeedback,
+            'countByUserAndAppSince',
+        ).mockResolvedValueOnce(0);
+
+        await expect(
+            submit(actor, { app: app.name, message: 'raced past the cap' }),
+        ).rejects.toMatchObject({
+            statusCode: 429,
+            legacyCode: 'too_many_requests',
+        });
+
+        const rows = (await server.clients.db.read(
+            'SELECT COUNT(*) AS n FROM `app_feedback` WHERE `user_id` = ?',
+            [userId],
+        )) as Array<{ n: unknown }>;
+        expect(Number(rows[0]?.n)).toBe(
+            AppFeedbackService.PER_USER_APP_DAILY_LIMIT,
+        );
+    });
+
+    it('enforces the per-user daily cap across apps with 429', async () => {
+        mockEmailConfigured();
+        const { userId: ownerId } = await makeUser();
+        const target = await makeApp(ownerId, { feedbackEnabled: true });
+        const other = await makeApp(ownerId, { feedbackEnabled: true });
+        const { actor, userId } = await makeUser();
+        for (let i = 0; i < AppFeedbackService.PER_USER_DAILY_LIMIT; i++) {
+            await server.stores.appFeedback.create({
+                appId: other.id,
+                appUid: other.uid,
+                userId,
+                message: `seed ${i}`,
+            });
+        }
+        await expect(
+            submit(actor, { app: target.name, message: 'over the limit' }),
+        ).rejects.toMatchObject({ statusCode: 429 });
+    });
+});
+
+// ── Owner email delivery ────────────────────────────────────────────
+
+// Which submissions get emailed, and what the mail contains, is service
+// logic covered in AppFeedbackService.test.ts. What the controller owes the
+// caller is that mail trouble never becomes the sender's problem.
+describe('AppFeedbackController owner email', () => {
+    it('a failing email send never fails the request', async () => {
+        mockEmailConfigured();
+        vi.spyOn(server.clients.email, 'send').mockRejectedValue(
+            new Error('smtp down'),
+        );
+        const { userId: ownerId } = await makeUser();
+        await confirmOwnerEmail(ownerId);
+        const app = await makeApp(ownerId, { feedbackEnabled: true });
+        const { actor, userId } = await makeUser();
+
+        const captured = await submit(actor, {
+            app: app.name,
+            message: 'still stored',
+        });
+        expect(captured.body).toEqual({});
+        const rows = (await server.clients.db.read(
+            'SELECT `email_sent` FROM `app_feedback` WHERE `user_id` = ?',
+            [userId],
+        )) as Array<{ email_sent: unknown }>;
+        expect(rows).toHaveLength(1);
+        expect(Boolean(rows[0]?.email_sent)).toBe(false);
+    });
+});

src/backend/controllers/feedback/AppFeedbackController.ts

@@ -0,0 +1,169 @@
+/*
+ * 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 { Request, Response } from 'express';
+import { Controller, Get, Post } from '../../core/http/decorators.js';
+import { HttpError } from '../../core/http/HttpError.js';
+import { AppFeedbackService } from '../../services/feedback/AppFeedbackService.js';
+import { PuterController } from '../types.js';
+
+/**
+ * Endpoints behind the "send feedback to this app's developer" dialog
+ * (`puter.ui.showFeedbackDialog()`). Only the GUI (desktop dialog or the
+ * puter.com popup) calls these; apps cannot — both routes reject app actors,
+ * which is what makes feedback impossible to submit programmatically on a
+ * user's behalf.
+ *
+ * The target may be named either by `app` (uid or name — the desktop knows
+ * which app asked) or by `origin` (external site — the popup passes its
+ * browser-attested opener origin). Exactly one must be provided.
+ */
+
+/** Sanity cap on the raw body field; the service enforces the real limit. */
+const RAW_MESSAGE_CAP = 50_000;
+
+// Upper bound on the `app`/`origin` target params. Must not exceed the
+// `source_origin` column (VARCHAR(2048) on MySQL/Postgres): the raw origin is
+// stored verbatim, and a longer value would fail the INSERT after passing
+// every validation — or be silently truncated on non-strict MySQL.
+const TARGET_PARAM_MAX_LENGTH = 2048;
+
+const readTargetParam = (value: unknown): string | undefined => {
+    return typeof value === 'string' &&
+        value.length > 0 &&
+        value.length <= TARGET_PARAM_MAX_LENGTH
+        ? value
+        : undefined;
+};
+
+@Controller('/app-feedback')
+export class AppFeedbackController extends PuterController {
+    /**
+     * GET /app-feedback/target — pre-flight for the dialog: whether the target
+     * app accepts feedback, plus its canonical title/name for display. Reveals
+     * nothing that `puter.apps.get` doesn't already.
+     */
+    @Get('/target', {
+        subdomain: 'api',
+        requireUserActor: true,
+        requireVerified: true,
+        rateLimit: {
+            scope: 'app-feedback-target',
+            limit: 60,
+            window: 60_000,
+            key: 'user',
+        },
+    })
+    async target(req: Request, res: Response): Promise<void> {
+        const app = readTargetParam(req.query.app);
+        const origin = readTargetParam(req.query.origin);
+        if (!app === !origin) {
+            throw new HttpError(
+                400,
+                'Exactly one of `app` and `origin` is required',
+                { legacyCode: 'bad_request' },
+            );
+        }
+
+        const service = this.services.appFeedback as AppFeedbackService;
+        res.json(await service.getTarget({ app, origin }));
+    }
+
+    /**
+     * POST /app-feedback — store one feedback message and email the app's
+     * developer. Strict limits: the route limits below are the cheap first
+     * line; AppFeedbackService enforces durable per-user/per-app daily caps
+     * from the database (the route limiter fails open, the DB caps don't).
+     */
+    @Post('/', {
+        subdomain: 'api',
+        requireUserActor: true,
+        requireVerified: true,
+        // Submissions only ever originate from our own GUI pages (desktop
+        // dialog / popup). Cross-origin browser pages get stopped here even
+        // if they somehow hold a user token; non-browser clients still pass
+        // and are handled by the caps.
+        guiOriginOnly: true,
+        rateLimit: [
+            {
+                scope: 'app-feedback-user',
+                limit: 5,
+                window: 30 * 60_000,
+                key: 'user',
+            },
+            // IP backstop so freshly minted accounts can't stack per-user
+            // budgets from one machine.
+            {
+                scope: 'app-feedback-ip',
+                limit: 30,
+                window: 24 * 60 * 60_000,
+                key: 'ip',
+            },
+        ],
+    })
+    async submit(req: Request, res: Response): Promise<void> {
+        const body = req.body ?? {};
+        const app = readTargetParam(body.app);
+        const origin = readTargetParam(body.origin);
+        if (!app === !origin) {
+            throw new HttpError(
+                400,
+                'Exactly one of `app` and `origin` is required',
+                { legacyCode: 'bad_request' },
+            );
+        }
+
+        const message = body.message;
+        if (typeof message !== 'string' || message.length === 0) {
+            throw new HttpError(400, '`message` is required', {
+                legacyCode: 'bad_request',
+            });
+        }
+        if (message.length > RAW_MESSAGE_CAP) {
+            throw new HttpError(
+                400,
+                `\`message\` is too long (max ${AppFeedbackService.MESSAGE_MAX_LENGTH} characters)`,
+                { legacyCode: 'bad_request' },
+            );
+        }
+
+        const sourceEnv =
+            body.context === 'app' || body.context === 'web'
+                ? body.context
+                : undefined;
+
+        const userId = req.actor?.user?.id;
+        if (!userId) {
+            throw new HttpError(401, 'Unauthorized', {
+                legacyCode: 'unauthorized',
+            });
+        }
+
+        const service = this.services.appFeedback as AppFeedbackService;
+        await service.submit({
+            userId,
+            app,
+            origin,
+            message,
+            sourceEnv,
+            sourceOrigin: origin ?? null,
+        });
+        res.json({});
+    }
+}

src/backend/controllers/fs/FSController.test.ts

@@ -21,7 +21,7 @@ import type { Request, Response } from 'express';
 import type { Readable } from 'node:stream';
 import { v4 as uuidv4 } from 'uuid';
 import { afterAll, beforeAll, describe, expect, it, vi } from 'vitest';
-import type { Actor } from '../../core/actor.js';
+import { makeActor, type Actor } from '../../core/actor.js';
 import { runWithContext } from '../../core/context.js';
 import { PuterServer } from '../../server.js';
 import { setupTestServer } from '../../testUtil.js';
@@ -927,7 +927,7 @@ describe('FSController.searchEntries', () => {
         const { actor: userActor } = await makeUser();
         const username = userActor.user!.username!;
         const appUid = `app-search-${uuidv4()}`;
-        const appActor: Actor = { ...userActor, app: { uid: appUid } };
+        const appActor = makeActor({ ...userActor, app: { uid: appUid } });
         const needle = `appneedle-${Math.random().toString(36).slice(2, 8)}`;
 
         // User-owned entry outside AppData — must NOT appear for the app.
@@ -978,7 +978,7 @@ describe('FSController.searchEntries', () => {
         const { actor: userActor } = await makeUser();
         const username = userActor.user!.username!;
         const appUid = `app-search-${uuidv4()}`;
-        const appActor: Actor = { ...userActor, app: { uid: appUid } };
+        const appActor = makeActor({ ...userActor, app: { uid: appUid } });
         const needle = `appneedle-${Math.random().toString(36).slice(2, 8)}`;
 
         // Only seed outside AppData — the app must not be able to find it.
@@ -1127,7 +1127,7 @@ describe('FSController.mkdirEntry', () => {
         const { actor: userActor } = await makeUser();
         const username = userActor.user!.username!;
         const appUid = `app-mkdir-${uuidv4()}`;
-        const appActor: Actor = { ...userActor, app: { uid: appUid } };
+        const appActor = makeActor({ ...userActor, app: { uid: appUid } });
         const target = `/${username}/AppData/${appUid}`;
 
         await withActor(userActor, () =>
@@ -1889,10 +1889,10 @@ describe('FSController.readdirEntries recursive', () => {
     it('masks denials for app-under-user actors as a 404 (legacy parity)', async () => {
         const { actor: userActor } = await makeUser();
         const username = userActor.user!.username!;
-        const appActor: Actor = {
+        const appActor = makeActor({
             ...userActor,
             app: { uid: `app-readdir-${uuidv4()}` },
-        };
+        });
         // The user's Documents is outside the app's AppData subtree, so the
         // app can't list it. Legacy `/readdir` masks this as a 404
         // subject_does_not_exist rather than leaking a 403.
@@ -2372,15 +2372,18 @@ describe('FSController.statEntry additional branches', () => {
 // ── #getReportedCosts ───────────────────────────────────────────────
 
 describe('FSController.getReportedCosts', () => {
-    it('mirrors every FS_COSTS entry as a per-byte line item', async () => {
-        const { FS_COSTS } = await import('./costs.js');
+    it('mirrors every storage-operation price as a per-operation line item', async () => {
+        const { STORAGE_OP_COSTS } =
+            await import('../../services/metering/costs.js');
         const reported = controller.getReportedCosts();
-        expect(reported.length).toBe(Object.keys(FS_COSTS).length);
-        for (const [usageType, ucentsPerUnit] of Object.entries(FS_COSTS)) {
+        expect(reported.length).toBe(Object.keys(STORAGE_OP_COSTS).length);
+        for (const [usageType, ucentsPerUnit] of Object.entries(
+            STORAGE_OP_COSTS,
+        )) {
             expect(reported).toContainEqual({
                 usageType,
                 ucentsPerUnit,
-                unit: 'byte',
+                unit: 'operation',
                 source: 'controller:fs',
             });
         }

src/backend/controllers/fs/FSController.ts

@@ -22,7 +22,6 @@ import type { Request, Response } from 'express';
 import { posix as pathPosix } from 'node:path';
 import { pipeline } from 'node:stream/promises';
 import type { Actor } from '../../core/actor.js';
-import { effectiveActorApp } from '../../core/actor.js';
 import { Context } from '../../core/context.js';
 import { HttpError } from '../../core/http/HttpError.js';
 import { Controller, Get, Post } from '../../core/http/decorators.js';
@@ -39,7 +38,19 @@ import {
 } from '../../util/concurrency.js';
 import { applyInlineContentSecurity } from '../../util/inlineContentSecurity.js';
 import { PuterController } from '../types.js';
-import { FS_COSTS } from './costs.js';
+import { STORAGE_OP_COSTS } from '../../services/metering/costs.js';
+import {
+    FS_MULTIPART_LIMIT,
+    FS_MUTATE_LIMIT,
+    FS_READ_CONCURRENT,
+    FS_READ_LIMIT,
+    FS_READDIR_LIMIT,
+    FS_SEARCH_CONCURRENT,
+    FS_SEARCH_LIMIT,
+    FS_STAT_LIMIT,
+    FS_WRITE_CONCURRENT,
+    FS_WRITE_LIMIT,
+} from './limits.js';
 import {
     assertAccess as assertLegacyAccess,
     fsEntryMimeType,
@@ -106,16 +117,26 @@ const DEFAULT_BATCH_WRITE_SIDE_EFFECT_CONCURRENCY = 8;
 
 @Controller('/fs')
 export class FSController extends PuterController {
+    // Object-store requests are reported here because the filesystem is what
+    // makes them. Bytes leaving the server are not: they are metered for every
+    // response, so `MeteringService` prices them.
     override getReportedCosts() {
-        return Object.entries(FS_COSTS).map(([usageType, ucentsPerUnit]) => ({
-            usageType,
-            ucentsPerUnit,
-            unit: 'byte',
-            source: 'controller:fs',
-        }));
+        return Object.entries(STORAGE_OP_COSTS).map(
+            ([usageType, ucentsPerUnit]) => ({
+                usageType,
+                ucentsPerUnit,
+                unit: 'operation',
+                source: 'controller:fs',
+            }),
+        );
     }
 
-    @Post('/startWrite', { subdomain: 'api', requireVerified: true })
+    @Post('/startWrite', {
+        subdomain: 'api',
+        requireVerified: true,
+        requireCredits: true,
+        rateLimit: FS_MULTIPART_LIMIT,
+    })
     async startWrite(
         req: Request<RouteParams, null, SignedWriteRequest>,
         res: Response<ClientSignedWriteResponse>,
@@ -173,7 +194,12 @@ export class FSController extends PuterController {
         );
     }
 
-    @Post('/startBatchWrite', { subdomain: 'api', requireVerified: true })
+    @Post('/startBatchWrite', {
+        subdomain: 'api',
+        requireVerified: true,
+        requireCredits: true,
+        rateLimit: FS_MULTIPART_LIMIT,
+    })
     async startBatchWrites(
         req: Request<RouteParams, null, SignedWriteRequest[]>,
         res: Response<ClientSignedWriteResponse[]>,
@@ -273,7 +299,12 @@ export class FSController extends PuterController {
         );
     }
 
-    @Post('/completeWrite', { subdomain: 'api', requireVerified: true })
+    @Post('/completeWrite', {
+        subdomain: 'api',
+        requireVerified: true,
+        requireCredits: true,
+        rateLimit: FS_MULTIPART_LIMIT,
+    })
     async completeWrite(
         req: Request<RouteParams, null, CompleteWriteRequest>,
         res: Response<ClientCompleteWriteResponse>,
@@ -304,7 +335,12 @@ export class FSController extends PuterController {
         );
     }
 
-    @Post('/completeBatchWrite', { subdomain: 'api', requireVerified: true })
+    @Post('/completeBatchWrite', {
+        subdomain: 'api',
+        requireVerified: true,
+        requireCredits: true,
+        rateLimit: FS_MULTIPART_LIMIT,
+    })
     async completeBatchWrites(
         req: Request<RouteParams, null, CompleteWriteRequest[]>,
         res: Response<ClientCompleteWriteResponse[]>,
@@ -347,7 +383,11 @@ export class FSController extends PuterController {
         );
     }
 
-    @Post('/abortWrite', { subdomain: 'api', requireVerified: true })
+    @Post('/abortWrite', {
+        subdomain: 'api',
+        requireVerified: true,
+        rateLimit: FS_MULTIPART_LIMIT,
+    })
     async abortWrite(
         req: Request<RouteParams, null, AbortWriteRequest>,
         res: Response<{ ok: true }>,
@@ -363,7 +403,12 @@ export class FSController extends PuterController {
         res.json({ ok: true });
     }
 
-    @Post('/signMultipartParts', { subdomain: 'api', requireVerified: true })
+    @Post('/signMultipartParts', {
+        subdomain: 'api',
+        requireVerified: true,
+        requireCredits: true,
+        rateLimit: FS_MULTIPART_LIMIT,
+    })
     async signMultipartParts(
         req: Request<RouteParams, null, SignMultipartPartsRequest>,
         res: Response<ClientSignMultipartPartsResponse>,
@@ -376,7 +421,13 @@ export class FSController extends PuterController {
         res.json(this.#withoutStorageInternals(response));
     }
 
-    @Post('/write', { subdomain: 'api', requireVerified: true })
+    @Post('/write', {
+        subdomain: 'api',
+        requireVerified: true,
+        requireCredits: true,
+        rateLimit: FS_WRITE_LIMIT,
+        concurrent: FS_WRITE_CONCURRENT,
+    })
     async write(
         req: Request<RouteParams, null, WriteRequest>,
         res: Response<ClientWriteResponse>,
@@ -422,7 +473,13 @@ export class FSController extends PuterController {
         res.json(this.#withRequiredClientFsEntry(updatedResponse));
     }
 
-    @Post('/batchWrite', { subdomain: 'api', requireVerified: true })
+    @Post('/batchWrite', {
+        subdomain: 'api',
+        requireVerified: true,
+        requireCredits: true,
+        rateLimit: FS_WRITE_LIMIT,
+        concurrent: FS_WRITE_CONCURRENT,
+    })
     async batchWrites(
         req: Request<RouteParams, null, WriteRequest[]>,
         res: Response<ClientWriteResponse[]>,
@@ -889,7 +946,11 @@ export class FSController extends PuterController {
 
     // -- Read-side routes ------------------------------------------------
 
-    @Post('/stat', { subdomain: 'api', requireVerified: true })
+    @Post('/stat', {
+        subdomain: 'api',
+        requireVerified: true,
+        rateLimit: FS_STAT_LIMIT,
+    })
     async statEntry(req: Request, res: Response) {
         const actor = this.#requireActor(req);
         const userId = this.#getActorUserId(req);
@@ -1003,12 +1064,20 @@ export class FSController extends PuterController {
      * without a JSON body. Every value arrives as a string, so parameter
      * parsing goes through the same coercion helpers the POST path uses.
      */
-    @Get('/readdir', { subdomain: 'api', requireVerified: true })
+    @Get('/readdir', {
+        subdomain: 'api',
+        requireVerified: true,
+        rateLimit: FS_READDIR_LIMIT,
+    })
     async readdirEntriesViaGet(req: Request, res: Response) {
         return this.readdirEntries(req, res);
     }
 
-    @Post('/readdir', { subdomain: 'api', requireVerified: true })
+    @Post('/readdir', {
+        subdomain: 'api',
+        requireVerified: true,
+        rateLimit: FS_READDIR_LIMIT,
+    })
     async readdirEntries(req: Request, res: Response) {
         const actor = this.#requireActor(req);
         // GET carries its parameters in the query string; POST in the body.
@@ -1204,7 +1273,12 @@ export class FSController extends PuterController {
         }
     }
 
-    @Post('/search', { subdomain: 'api', requireVerified: true })
+    @Post('/search', {
+        subdomain: 'api',
+        requireVerified: true,
+        rateLimit: FS_SEARCH_LIMIT,
+        concurrent: FS_SEARCH_CONCURRENT,
+    })
     async searchEntries(req: Request, res: Response) {
         const actor = this.#requireActor(req);
         const userId = this.#getActorUserId(req);
@@ -1230,7 +1304,13 @@ export class FSController extends PuterController {
         res.json(results);
     }
 
-    @Get('/read', { subdomain: 'api', requireVerified: true })
+    @Get('/read', {
+        subdomain: 'api',
+        requireVerified: true,
+        requireCredits: true,
+        rateLimit: FS_READ_LIMIT,
+        concurrent: FS_READ_CONCURRENT,
+    })
     async readEntry(req: Request, res: Response) {
         const actor = this.#requireActor(req);
         const query = this.#toObjectRecord(req.query);
@@ -1274,15 +1354,6 @@ export class FSController extends PuterController {
         );
         res.status(range ? 206 : 200);
 
-        const metering = this.services.metering as
-            | {
-                  batchIncrementUsages?: (
-                      actor: unknown,
-                      entries: unknown[],
-                  ) => void;
-              }
-            | undefined;
-
         try {
             await pipeline(download.body, res);
         } catch {
@@ -1290,28 +1361,15 @@ export class FSController extends PuterController {
             // tore down both ends. Response is partially sent; nothing to do.
             return;
         }
-
-        // Meter egress only on successful completion.
-        if (metering?.batchIncrementUsages && download.contentLength) {
-            try {
-                const bytes = download.contentLength;
-                metering.batchIncrementUsages(actor, [
-                    {
-                        usageType: 'filesystem:egress:bytes',
-                        usageAmount: bytes,
-                        costOverride:
-                            FS_COSTS['filesystem:egress:bytes'] * bytes,
-                    },
-                ]);
-            } catch {
-                // ignore — metering is non-critical.
-            }
-        }
     }
 
     // -- Mutation routes ------------------------------------------------
 
-    @Post('/mkdir', { subdomain: 'api', requireVerified: true })
+    @Post('/mkdir', {
+        subdomain: 'api',
+        requireVerified: true,
+        rateLimit: FS_MUTATE_LIMIT,
+    })
     async mkdirEntry(req: Request, res: Response) {
         const actor = this.#requireActor(req);
         const userId = this.#getActorUserId(req);
@@ -1352,7 +1410,11 @@ export class FSController extends PuterController {
         res.json(this.#toClientEntry(entry));
     }
 
-    @Post('/touch', { subdomain: 'api', requireVerified: true })
+    @Post('/touch', {
+        subdomain: 'api',
+        requireVerified: true,
+        rateLimit: FS_MUTATE_LIMIT,
+    })
     async touchEntry(req: Request, res: Response) {
         const actor = this.#requireActor(req);
         const userId = this.#getActorUserId(req);
@@ -1388,7 +1450,11 @@ export class FSController extends PuterController {
         res.json(this.#toClientEntry(entry));
     }
 
-    @Post('/rename', { subdomain: 'api', requireVerified: true })
+    @Post('/rename', {
+        subdomain: 'api',
+        requireVerified: true,
+        rateLimit: FS_MUTATE_LIMIT,
+    })
     async renameEntry(req: Request, res: Response) {
         const actor = this.#requireActor(req);
         const body = this.#toObjectRecord(req.body);
@@ -1406,7 +1472,11 @@ export class FSController extends PuterController {
         res.json(this.#toClientEntry(renamed));
     }
 
-    @Post('/delete', { subdomain: 'api', requireVerified: true })
+    @Post('/delete', {
+        subdomain: 'api',
+        requireVerified: true,
+        rateLimit: FS_MUTATE_LIMIT,
+    })
     async deleteEntry(req: Request, res: Response) {
         const actor = this.#requireActor(req);
         const userId = this.#getActorUserId(req);
@@ -1424,7 +1494,11 @@ export class FSController extends PuterController {
         res.json({ ok: true });
     }
 
-    @Post('/move', { subdomain: 'api', requireVerified: true })
+    @Post('/move', {
+        subdomain: 'api',
+        requireVerified: true,
+        rateLimit: FS_MUTATE_LIMIT,
+    })
     async moveEntry(req: Request, res: Response) {
         const actor = this.#requireActor(req);
         const userId = this.#getActorUserId(req);
@@ -1452,7 +1526,12 @@ export class FSController extends PuterController {
         res.json(this.#toClientEntry(moved));
     }
 
-    @Post('/copy', { subdomain: 'api', requireVerified: true })
+    @Post('/copy', {
+        subdomain: 'api',
+        requireVerified: true,
+        requireCredits: true,
+        rateLimit: FS_MUTATE_LIMIT,
+    })
     async copyEntry(req: Request, res: Response) {
         const actor = this.#requireActor(req);
         const userId = this.#getActorUserId(req);
@@ -1467,6 +1546,7 @@ export class FSController extends PuterController {
         await this.#assertAccess(actor, source.path, 'read');
         await this.#assertAccess(actor, destinationParent.path, 'write');
 
+        const storageAllowanceMax = this.#getStorageAllowanceMaxOverride(req);
         const copy = await this.services.fs.copy(userId, {
             source,
             destinationParent,
@@ -1475,12 +1555,19 @@ export class FSController extends PuterController {
             overwrite: this.#toBoolean(body.overwrite) ?? false,
             dedupeName:
                 this.#toBoolean(body.dedupe_name ?? body.change_name) ?? true,
+            ...(storageAllowanceMax !== undefined
+                ? { storageAllowanceMax }
+                : {}),
         });
         this.#emitGuiItemAdded(copy);
         res.json(this.#toClientEntry(copy));
     }
 
-    @Post('/mkshortcut', { subdomain: 'api', requireVerified: true })
+    @Post('/mkshortcut', {
+        subdomain: 'api',
+        requireVerified: true,
+        rateLimit: FS_MUTATE_LIMIT,
+    })
     async mkshortcutEntry(req: Request, res: Response) {
         const actor = this.#requireActor(req);
         const userId = this.#getActorUserId(req);
@@ -2091,8 +2178,7 @@ export class FSController extends PuterController {
         // the ActorUser type. Access via the escape hatch until a proper
         // storage-quota mechanism is in place.
         const actorUser = req.actor?.user as
-            | Record<string, unknown>
-            | undefined;
+            Record<string, unknown> | undefined;
 
         const candidates = [
             this.#toStorageCapacityCandidate(actorUser?.free_storage),
@@ -2437,7 +2523,7 @@ export class FSController extends PuterController {
     // search results to this path so app actors can't see entries outside
     // their AppData via `/fs/search`.
     #appDataScopeForActor(actor: Actor): string | undefined {
-        const app = effectiveActorApp(actor);
+        const app = actor.effectiveApp;
         if (!app) return undefined;
         const username = actor.user?.username;
         if (typeof username !== 'string' || username.length === 0)

src/backend/controllers/fs/LegacyFSController.routes.test.ts

@@ -21,7 +21,7 @@ import type { Request, RequestHandler, Response } from 'express';
 import { Readable, Writable } from 'node:stream';
 import { v4 as uuidv4 } from 'uuid';
 import { afterAll, beforeAll, describe, expect, it, vi } from 'vitest';
-import type { Actor } from '../../core/actor.js';
+import { makeActor, type Actor } from '../../core/actor.js';
 import { runWithContext } from '../../core/context.js';
 import { PuterRouter } from '../../core/http/PuterRouter.js';
 import { PuterServer } from '../../server.js';
@@ -377,7 +377,7 @@ describe('LegacyFSController.readdirSubdomains', () => {
             'INSERT INTO `subdomains` (`uuid`, `subdomain`, `user_id`) VALUES (?, ?, ?)',
             [uuidv4(), `sd-${Math.random().toString(36).slice(2, 8)}`, userId],
         );
-        const appActor: Actor = { ...actor, app: { uid: `app-${uuidv4()}` } };
+        const appActor = makeActor({ ...actor, app: { uid: `app-${uuidv4()}` } });
         const { res, captured } = makeRes();
         await withActor(appActor, () =>
             controller.readdirSubdomains(makeReq({ actor: appActor }), res),

src/backend/controllers/fs/LegacyFSController.test.ts

@@ -21,7 +21,7 @@ import type { Request, Response } from 'express';
 import { Readable } from 'node:stream';
 import { afterAll, beforeAll, describe, expect, it, vi } from 'vitest';
 import { v4 as uuidv4 } from 'uuid';
-import type { Actor } from '../../core/actor.js';
+import { makeActor, type Actor } from '../../core/actor.js';
 import { runWithContext } from '../../core/context.js';
 import { PuterRouter } from '../../core/http/PuterRouter.js';
 import { PuterServer } from '../../server.js';
@@ -250,7 +250,7 @@ describe('LegacyFSController.mkdir', () => {
         const { actor: userActor } = await makeUser();
         const username = userActor.user!.username!;
         const appUid = `app-legacy-mkdir-${uuidv4()}`;
-        const appActor: Actor = { ...userActor, app: { uid: appUid } };
+        const appActor = makeActor({ ...userActor, app: { uid: appUid } });
         const parent = `/${username}/AppData`;
 
         await withActor(userActor, () =>
@@ -934,7 +934,7 @@ describe('LegacyFSController.copy', () => {
         const username = actor.user!.username!;
         const src = `/${username}/Documents/dup.txt`;
         const existing = `/${username}/Pictures/dup.txt`;
-        for ( const p of [src, existing] ) {
+        for (const p of [src, existing]) {
             await withActor(actor, () =>
                 controller.touch(
                     makeReq({ body: { path: p }, actor }),
@@ -964,9 +964,8 @@ describe('LegacyFSController.copy', () => {
             fields: { entry_name: 'dup.txt' },
         });
 
-        const replaced = (await server.stores.fsEntry.getEntryByPath(
-            existing,
-        ))!;
+        const replaced =
+            (await server.stores.fsEntry.getEntryByPath(existing))!;
 
         // With overwrite: the replaced entry rides along in the response
         // (so the caller can drop its row) and item.removed tells every
@@ -1091,7 +1090,7 @@ describe('LegacyFSController.move', () => {
         const username = actor.user!.username!;
         const src = `/${username}/Documents/clash.txt`;
         const existing = `/${username}/Pictures/clash.txt`;
-        for ( const p of [src, existing] ) {
+        for (const p of [src, existing]) {
             await withActor(actor, () =>
                 controller.touch(
                     makeReq({ body: { path: p }, actor }),
@@ -1121,9 +1120,8 @@ describe('LegacyFSController.move', () => {
             fields: { entry_name: 'clash.txt' },
         });
 
-        const replaced = (await server.stores.fsEntry.getEntryByPath(
-            existing,
-        ))!;
+        const replaced =
+            (await server.stores.fsEntry.getEntryByPath(existing))!;
 
         // With overwrite: the replaced entry rides along in the response
         // (so the caller can drop its row) and item.removed tells every
@@ -1214,7 +1212,7 @@ describe('LegacyFSController.search', () => {
         const { actor: userActor } = await makeUser();
         const username = userActor.user!.username!;
         const appUid = `app-legacy-search-${uuidv4()}`;
-        const appActor: Actor = { ...userActor, app: { uid: appUid } };
+        const appActor = makeActor({ ...userActor, app: { uid: appUid } });
         const needle = `appneedle-${Math.random().toString(36).slice(2, 8)}`;
 
         await withActor(userActor, () =>
@@ -1419,10 +1417,10 @@ describe('LegacyFSController.sign', () => {
             },
             { ownerUserId: userActor.user!.id! },
         );
-        const attackerActor: Actor = {
+        const attackerActor = makeActor({
             ...userActor,
             app: { uid: `attacker-${uuidv4()}` },
-        };
+        });
 
         const { res } = makeRes();
         await expect(
@@ -1456,7 +1454,7 @@ describe('LegacyFSController.sign', () => {
             },
             { ownerUserId: userActor.user!.id! },
         );
-        const appActor: Actor = { ...userActor, app: { uid: ownApp.uid } };
+        const appActor = makeActor({ ...userActor, app: { uid: ownApp.uid } });
 
         const { res, captured } = makeRes();
         await withActor(appActor, () =>
@@ -1736,6 +1734,185 @@ describe('LegacyFSController.openItem (write_url stripping)', () => {
     });
 });
 
+// ── openItem: the grant is a user-authority action ──────────────────
+//
+// /open_item writes a user→app ACL row. An app actor reaching it could hand
+// itself write on any file it can read — read-only access to a folder would
+// silently widen into persistent write on the files inside it, with no second
+// consent prompt. The user has to be the one opening the item, and the row it
+// leaves can never exceed what that user proved on the entry.
+
+describe('LegacyFSController.openItem (grant scope)', () => {
+    const makeApp = async (ownerUserId: number) => {
+        const name = `opener-${uuidv4()}`;
+        return (await server.stores.app.create(
+            {
+                name,
+                title: 'Opener test app',
+                index_url: `https://${name}.test/`,
+            },
+            { ownerUserId },
+        )) as { id: number; uid: string };
+    };
+
+    it('is registered behind the user-session gate', () => {
+        const router = new PuterRouter();
+        (
+            controller as unknown as {
+                registerRoutes: (r: PuterRouter) => void;
+            }
+        ).registerRoutes(router);
+        const route = router.routes.find(
+            (r) => r.method === 'post' && r.path === '/open_item',
+        );
+        expect(route?.options).toMatchObject({
+            requireUserActor: true,
+            allowFullAccessToken: true,
+        });
+    });
+
+    it('rejects an app actor, leaving no grant behind', async () => {
+        const { actor: userActor, userId } = await makeUser();
+        const username = userActor.user!.username!;
+        const target = `/${username}/Documents/app-opened.txt`;
+        await withActor(userActor, () =>
+            controller.touch(
+                makeReq({ body: { path: target }, actor: userActor }),
+                makeRes().res,
+            ),
+        );
+        const entry = await server.stores.fsEntry.getEntryByPath(target);
+        // The app holds the read the user approved, and is its own default
+        // opener — the shape that would let it widen that read into write.
+        const app = await makeApp(userId);
+        await withActor(userActor, () =>
+            server.services.permission.grantUserAppPermission(
+                userActor,
+                app.uid,
+                `fs:${entry!.uuid}:read`,
+            ),
+        );
+        const appActor = makeActor({
+            ...userActor,
+            app: { uid: app.uid, id: app.id },
+        });
+
+        const spy = vi
+            .spyOn(server.services.suggestedApps, 'getSuggestedApps')
+            .mockResolvedValue([{ uuid: app.uid }] as never);
+        try {
+            await expect(
+                withActor(appActor, () =>
+                    controller.openItem(
+                        makeReq({
+                            body: { uid: entry!.uuid },
+                            actor: appActor,
+                        }),
+                        makeRes().res,
+                    ),
+                ),
+            ).rejects.toMatchObject({ statusCode: 403 });
+        } finally {
+            spy.mockRestore();
+        }
+
+        expect(
+            await server.stores.permission.hasUserAppPerm(
+                userId,
+                app.id,
+                `fs:${entry!.uuid}:write`,
+            ),
+        ).toBe(false);
+    });
+
+    it('grants the suggested app write when the caller has write', async () => {
+        const { actor, userId } = await makeUser();
+        const target = `/${actor.user!.username}/Documents/owned-open.txt`;
+        await withActor(actor, () =>
+            controller.touch(
+                makeReq({ body: { path: target }, actor }),
+                makeRes().res,
+            ),
+        );
+        const entry = await server.stores.fsEntry.getEntryByPath(target);
+        const app = await makeApp(userId);
+
+        const spy = vi
+            .spyOn(server.services.suggestedApps, 'getSuggestedApps')
+            .mockResolvedValue([{ uuid: app.uid }] as never);
+        try {
+            await withActor(actor, () =>
+                controller.openItem(
+                    makeReq({ body: { path: target }, actor }),
+                    makeRes().res,
+                ),
+            );
+        } finally {
+            spy.mockRestore();
+        }
+
+        expect(
+            await server.stores.permission.hasUserAppPerm(
+                userId,
+                app.id,
+                `fs:${entry!.uuid}:write`,
+            ),
+        ).toBe(true);
+    });
+
+    it('grants only read when the caller is a read-only sharee', async () => {
+        const victim = await makeUser();
+        const sharee = await makeUser();
+        const target = `/${victim.actor.user!.username}/Documents/ro-open.txt`;
+        await withActor(victim.actor, () =>
+            controller.touch(
+                makeReq({ body: { path: target }, actor: victim.actor }),
+                makeRes().res,
+            ),
+        );
+        const entry = await server.stores.fsEntry.getEntryByPath(target);
+        await server.services.permission.grantUserUserPermission(
+            victim.actor,
+            sharee.actor.user!.username!,
+            `fs:${entry!.uuid}:read`,
+            {},
+        );
+        const app = await makeApp(sharee.userId);
+
+        const spy = vi
+            .spyOn(server.services.suggestedApps, 'getSuggestedApps')
+            .mockResolvedValue([{ uuid: app.uid }] as never);
+        try {
+            await withActor(sharee.actor, () =>
+                controller.openItem(
+                    makeReq({
+                        body: { uid: entry!.uuid },
+                        actor: sharee.actor,
+                    }),
+                    makeRes().res,
+                ),
+            );
+        } finally {
+            spy.mockRestore();
+        }
+
+        expect(
+            await server.stores.permission.hasUserAppPerm(
+                sharee.userId,
+                app.id,
+                `fs:${entry!.uuid}:write`,
+            ),
+        ).toBe(false);
+        expect(
+            await server.stores.permission.hasUserAppPerm(
+                sharee.userId,
+                app.id,
+                `fs:${entry!.uuid}:read`,
+            ),
+        ).toBe(true);
+    });
+});
+
 // ── requestAppRootDir ───────────────────────────────────────────────
 
 describe('LegacyFSController.requestAppRootDir', () => {
@@ -1766,10 +1943,10 @@ describe('LegacyFSController.requestAppRootDir', () => {
     it('rejects with 403 when the actor.app.uid differs from the requested app_uid', async () => {
         const { actor } = await makeUser();
         const { res } = makeRes();
-        const appActor = {
+        const appActor = makeActor({
             ...actor,
             app: { uid: 'app-mismatch' },
-        } as unknown as Actor;
+        });
         await expect(
             withActor(appActor, () =>
                 controller.requestAppRootDir(
@@ -1787,10 +1964,10 @@ describe('LegacyFSController.requestAppRootDir', () => {
         const { actor } = await makeUser();
         const username = actor.user!.username!;
         const appUid = 'app-self';
-        const appActor = {
+        const appActor = makeActor({
             ...actor,
             app: { uid: appUid },
-        } as unknown as Actor;
+        });
         const { res, captured } = makeRes();
         await withActor(appActor, () =>
             controller.requestAppRootDir(
@@ -2463,10 +2640,10 @@ describe('LegacyFSController.sign app sandbox + write downgrade', () => {
         const { actor } = await makeUser();
         const username = actor.user!.username!;
         // Build an app-under-user actor whose AppData root is the test app.
-        const appActor = {
+        const appActor = makeActor({
             ...actor,
             app: { uid: 'sandbox-app' },
-        } as unknown as Actor;
+        });
 
         // Create a file *outside* /Documents (anywhere outside AppData/<uid>).
         const target = `/${username}/Documents/forbidden.txt`;
@@ -2730,10 +2907,10 @@ describe('LegacyFSController.suggestApps', () => {
     it('refuses to look up entries an app actor cannot see', async () => {
         const { actor: userActor } = await makeUser();
         const username = userActor.user!.username!;
-        const appActor: Actor = {
+        const appActor = makeActor({
             ...userActor,
             app: { uid: `app-suggest-${uuidv4()}` },
-        };
+        });
         const target = `/${username}/Documents/probe.txt`;
         await withActor(userActor, () =>
             controller.touch(
@@ -2801,10 +2978,10 @@ describe('LegacyFSController.suggestApps', () => {
 describe('LegacyFSController.readdirSubdomains', () => {
     it('returns an empty array for app-under-user actors', async () => {
         const { actor: userActor, userId } = await makeUser();
-        const appActor: Actor = {
+        const appActor = makeActor({
             ...userActor,
             app: { uid: `app-subd-${uuidv4()}` },
-        };
+        });
         await server.clients.db.write(
             'INSERT INTO `subdomains` (`uuid`, `subdomain`, `user_id`) VALUES (?, ?, ?)',
             [uuidv4(), `sd-${uuidv4().slice(0, 8)}`, userId],
@@ -2843,10 +3020,10 @@ describe('LegacyFSController.updateFsentryThumbnail', () => {
     it('rejects an app actor probing entries outside AppData with 404', async () => {
         const { actor: userActor } = await makeUser();
         const username = userActor.user!.username!;
-        const appActor: Actor = {
+        const appActor = makeActor({
             ...userActor,
             app: { uid: `app-thumb-${uuidv4()}` },
-        };
+        });
         const target = `/${username}/Documents/thumbme.txt`;
         await withActor(userActor, () =>
             controller.touch(
@@ -2915,40 +3092,48 @@ describe('LegacyFSController.updateFsentryThumbnail', () => {
     // object — an fs object's key is its fsentry uuid — and have the server
     // read or destroy it. Only inline image data is accepted.
     it.each([
-        ['an s3:// pointer', 's3://puter-local/00000000-0000-4000-8000-000000000000'],
+        [
+            'an s3:// pointer',
+            's3://puter-local/00000000-0000-4000-8000-000000000000',
+        ],
         ['an https URL', 'https://cdn.example.com/x.png'],
         ['a bare object key', 'thumbnails/whatever'],
-    ])('rejects %s instead of storing it verbatim', async (_label, thumbnail) => {
-        const { actor } = await makeUser();
-        const username = actor.user!.username!;
-        const target = `/${username}/Documents/thumbme3-${uuidv4()}.txt`;
-        await withActor(actor, () =>
-            controller.touch(
-                makeReq({
-                    body: { path: target, set_modified_to_now: true },
-                    actor,
-                }),
-                makeRes().res,
-            ),
-        );
-        const entry = await server.stores.fsEntry.getEntryByPath(target);
-
-        const { res } = makeRes();
-        await expect(
-            withActor(actor, () =>
-                controller.updateFsentryThumbnail(
+    ])(
+        'rejects %s instead of storing it verbatim',
+        async (_label, thumbnail) => {
+            const { actor } = await makeUser();
+            const username = actor.user!.username!;
+            const target = `/${username}/Documents/thumbme3-${uuidv4()}.txt`;
+            await withActor(actor, () =>
+                controller.touch(
                     makeReq({
-                        body: { uid: entry!.uuid, thumbnail },
+                        body: { path: target, set_modified_to_now: true },
                         actor,
                     }),
-                    res,
+                    makeRes().res,
                 ),
-            ),
-        ).rejects.toMatchObject({ statusCode: 400 });
+            );
+            const entry = await server.stores.fsEntry.getEntryByPath(target);
 
-        const after = await server.stores.fsEntry.getEntryByUuid(entry!.uuid);
-        expect(after?.thumbnail ?? null).toBeNull();
-    });
+            const { res } = makeRes();
+            await expect(
+                withActor(actor, () =>
+                    controller.updateFsentryThumbnail(
+                        makeReq({
+                            body: { uid: entry!.uuid, thumbnail },
+                            actor,
+                        }),
+                        res,
+                    ),
+                ),
+            ).rejects.toMatchObject({ statusCode: 400 });
+
+            const after = await server.stores.fsEntry.getEntryByUuid(
+                entry!.uuid,
+            );
+            expect(after?.thumbnail ?? null).toBeNull();
+        },
+    );
 });
 
 // ── GET /get-launch-apps ────────────────────────────────────────────

src/backend/controllers/fs/LegacyFSController.ts

@@ -21,17 +21,21 @@ import Busboy from 'busboy';
 import type { NextFunction, Request, RequestHandler, Response } from 'express';
 import { contentType as contentTypeFromMime } from 'mime-types';
 import { posix as pathPosix } from 'node:path';
-import type { Actor } from '../../core/actor.js';
-import { effectiveActorApp, isAccessTokenActor } from '../../core/actor.js';
+import {
+    assertResolvedActor,
+    isAccessTokenActor,
+    makeActor,
+} from '../../core/actor.js';
 import { Context } from '../../core/context.js';
 import { HttpError } from '../../core/http/HttpError.js';
+import { RouteOptions } from '../../core/http/index.js';
 import {
     assertNotSuspended,
     assertVerifiedAccount,
 } from '../../core/http/middleware/gates.js';
-import { RouteOptions } from '../../core/http/index.js';
 import type { PuterRouter } from '../../core/http/PuterRouter.js';
 import type { ACLService } from '../../services/acl/ACLService.js';
+import { assertActorHasCredits } from '../../services/metering/enforcement.js';
 import type { SignedFile } from '../../util/fileSigning.js';
 import { verifySignature } from '../../util/fileSigning.js';
 import {
@@ -40,7 +44,24 @@ import {
 } from '../../util/hostedAppBacking.js';
 import { applyInlineContentSecurity } from '../../util/inlineContentSecurity.js';
 import { PuterController } from '../types.js';
-import { FS_COSTS } from './costs.js';
+import {
+    FS_BATCH_CONCURRENT,
+    FS_BATCH_LIMIT,
+    FS_DF_LIMIT,
+    FS_HELPER_LIMIT,
+    FS_MUTATE_LIMIT,
+    FS_POLL_LIMIT,
+    FS_READ_CONCURRENT,
+    FS_READ_LIMIT,
+    FS_READDIR_LIMIT,
+    FS_SEARCH_CONCURRENT,
+    FS_SEARCH_LIMIT,
+    FS_SIGN_LIMIT,
+    FS_SIGNED_CONCURRENT,
+    FS_SIGNED_READ_LIMIT,
+    FS_SIGNED_WRITE_LIMIT,
+    FS_STAT_LIMIT,
+} from './limits.js';
 import {
     asRecord,
     assertAccess,
@@ -95,6 +116,14 @@ export class LegacyFSController extends PuterController {
             subdomain: 'api',
             requireVerified: true,
         } as RouteOptions;
+        // Operations that move file content or make object-store requests on
+        // the caller's behalf, and so are refused to an account with nothing
+        // left of its budget. The metadata routes above deliberately aren't:
+        // an account that has run out still has to be able to look at what it
+        // has and delete it. The signature-authorised routes aren't either —
+        // they carry no session to answer for, and `/sign` (which does) is
+        // where the URL that reaches them is minted.
+        const spends = { ...apiOptions, requireCredits: true } as RouteOptions;
         // Signed-URL routes: the handler validates the URL signature itself,
         // so no auth gate is applied (matches v1, which mounted these routers
         // with no middleware).
@@ -103,40 +132,118 @@ export class LegacyFSController extends PuterController {
         } as RouteOptions;
 
         // Core filesystem_api routes — direct handlers over the FS service.
-        router.post('/stat', apiOptions, this.stat);
-        router.post('/readdir', apiOptions, this.readdir);
-        router.post('/mkdir', apiOptions, this.mkdir);
-        router.post('/copy', apiOptions, this.copy);
-        router.post('/move', apiOptions, this.move);
-        router.post('/delete', apiOptions, this.delete);
-        router.post('/rename', apiOptions, this.rename);
-        router.post('/touch', apiOptions, this.touch);
-        router.post('/search', apiOptions, this.search);
-        router.get('/read', apiOptions, this.read);
+        // Limits come from `./limits` and carry an explicit `scope`, so these
+        // draw from the same per-user budget as their v2 counterparts rather
+        // than handing a caller a second allowance for the same operation.
+        const mutate = { ...apiOptions, rateLimit: FS_MUTATE_LIMIT };
+        router.post(
+            '/stat',
+            { ...apiOptions, rateLimit: FS_STAT_LIMIT },
+            this.stat,
+        );
+        router.post(
+            '/readdir',
+            { ...apiOptions, rateLimit: FS_READDIR_LIMIT },
+            this.readdir,
+        );
+        router.post('/mkdir', mutate, this.mkdir);
+        router.post('/copy', { ...mutate, requireCredits: true }, this.copy);
+        router.post('/move', mutate, this.move);
+        router.post('/delete', mutate, this.delete);
+        router.post('/rename', mutate, this.rename);
+        router.post('/touch', mutate, this.touch);
+        router.post(
+            '/search',
+            {
+                ...apiOptions,
+                rateLimit: FS_SEARCH_LIMIT,
+                concurrent: FS_SEARCH_CONCURRENT,
+            },
+            this.search,
+        );
+        router.get(
+            '/read',
+            {
+                ...spends,
+                rateLimit: FS_READ_LIMIT,
+                concurrent: FS_READ_CONCURRENT,
+            },
+            this.read,
+        );
         router.get(
             '/token-read',
             {
                 subdomain: 'api',
                 requireVerified: false,
                 allowAccessToken: true,
+                // An access token may or may not carry a user, so this shares
+                // the network-keyed budget the other signed routes use.
+                rateLimit: FS_SIGNED_READ_LIMIT,
             },
             this.tokenRead,
         );
 
-        router.post('/batch', apiOptions, this.batch);
+        router.post(
+            '/batch',
+            {
+                ...spends,
+                rateLimit: FS_BATCH_LIMIT,
+                concurrent: FS_BATCH_CONCURRENT,
+            },
+            this.batch,
+        );
 
         // Signed-URL + meta routes.
-        router.post('/sign', apiOptions, this.sign);
-        router.post('/writeFile', signedOptions, this.writeFile);
-        router.get('/file', signedOptions, this.file);
-        router.all('/df', apiOptions, this.df);
-        router.post('/open_item', apiOptions, this.openItem);
+        router.post(
+            '/sign',
+            // Gated even though it moves nothing itself: the URL it returns
+            // outlives the request and is served by a route with no session to
+            // check, so this is the last point at which the account is known.
+            { ...spends, rateLimit: FS_SIGN_LIMIT },
+            this.sign,
+        );
+        router.post(
+            '/writeFile',
+            {
+                ...signedOptions,
+                rateLimit: FS_SIGNED_WRITE_LIMIT,
+                concurrent: FS_SIGNED_CONCURRENT,
+            },
+            this.writeFile,
+        );
+        router.get(
+            '/file',
+            {
+                ...signedOptions,
+                rateLimit: FS_SIGNED_READ_LIMIT,
+                concurrent: FS_SIGNED_CONCURRENT,
+            },
+            this.file,
+        );
+        router.all('/df', { ...apiOptions, rateLimit: FS_DF_LIMIT }, this.df);
+        router.post(
+            '/open_item',
+            {
+                ...apiOptions,
+                // Opening an item grants an app access to a file on the user's
+                // behalf, so only the user's own credential may drive it — an
+                // app calling it for itself would be widening its own ACL.
+                requireUserActor: true,
+                allowFullAccessToken: true,
+                rateLimit: FS_HELPER_LIMIT,
+            },
+            this.openItem,
+        );
         router.post(
             '/auth/request-app-root-dir',
-            apiOptions,
+            { ...apiOptions, rateLimit: FS_SIGN_LIMIT },
             this.requestAppRootDir,
         );
-        router.post('/auth/check-app-acl', apiOptions, this.checkAppAcl);
+        router.post(
+            '/auth/check-app-acl',
+            { ...apiOptions, rateLimit: FS_SIGN_LIMIT },
+            this.checkAppAcl,
+        );
 
         // `/down` — session-auth'd file download. Unlike `/file` (signed URL)
         // this accepts a path on the user's behalf and streams as attachment.
@@ -153,7 +260,10 @@ export class LegacyFSController extends PuterController {
                 // (CSRF can't forge a header-credentialed request).
                 allowFullAccessToken: true,
                 requireVerified: true,
+                requireCredits: true,
                 antiCsrf: true,
+                rateLimit: FS_READ_LIMIT,
+                concurrent: FS_READ_CONCURRENT,
             },
             this.down,
         );
@@ -164,92 +274,102 @@ export class LegacyFSController extends PuterController {
             });
         });
 
-        router.get('/get-launch-apps', apiOptions, async (req, res) => {
-            const recommendedSvc = this.services.recommendedApps as unknown as
-                { getRecommendedApps?: () => Promise<unknown[]> } | undefined;
-            const recommended = recommendedSvc?.getRecommendedApps
-                ? await recommendedSvc.getRecommendedApps()
-                : [];
+        router.get(
+            '/get-launch-apps',
+            { ...apiOptions, rateLimit: FS_HELPER_LIMIT },
+            async (req, res) => {
+                const recommendedSvc = this.services
+                    .recommendedApps as unknown as
+                    | { getRecommendedApps?: () => Promise<unknown[]> }
+                    | undefined;
+                const recommended = recommendedSvc?.getRecommendedApps
+                    ? await recommendedSvc.getRecommendedApps()
+                    : [];
 
-            let recent: unknown[] = [];
-            const userId = req.actor?.user?.id;
-            if (userId) {
-                const recentUids =
-                    (await (
+                let recent: unknown[] = [];
+                const userId = req.actor?.user?.id;
+                if (userId) {
+                    const recentUids =
+                        (await (
+                            this.stores.app as unknown as {
+                                getRecentAppOpens?: (
+                                    id: number,
+                                    opts?: { limit?: number },
+                                ) => Promise<string[]>;
+                            }
+                        ).getRecentAppOpens?.(userId, { limit: 10 })) ?? [];
+                    // One batched read for the rows, then the backing checks
+                    // concurrently. Serially awaiting a lookup per uid put ~2
+                    // round trips of latency on every desktop boot.
+                    const appsByUid = await (
                         this.stores.app as unknown as {
-                            getRecentAppOpens?: (
-                                id: number,
-                                opts?: { limit?: number },
-                            ) => Promise<string[]>;
+                            getByUids: (
+                                uids: string[],
+                            ) => Promise<Map<string, Record<string, unknown>>>;
                         }
-                    ).getRecentAppOpens?.(userId, { limit: 10 })) ?? [];
-                // One batched read for the rows, then the backing checks
-                // concurrently. Serially awaiting a lookup per uid put ~2
-                // round trips of latency on every desktop boot.
-                const appsByUid = await (
-                    this.stores.app as unknown as {
-                        getByUids: (
-                            uids: string[],
-                        ) => Promise<Map<string, Record<string, unknown>>>;
-                    }
-                ).getByUids(recentUids);
+                    ).getByUids(recentUids);
 
-                // `recentUids` is ordered most-recent-first; preserve it.
-                const orderedApps = recentUids
-                    .map((uid) => appsByUid.get(uid))
-                    .filter((app): app is Record<string, unknown> =>
-                        Boolean(app),
+                    // `recentUids` is ordered most-recent-first; preserve it.
+                    const orderedApps = recentUids
+                        .map((uid) => appsByUid.get(uid))
+                        .filter((app): app is Record<string, unknown> =>
+                            Boolean(app),
+                        );
+
+                    // Don't hand out an index_url whose puter-hosted backing is
+                    // gone or reclaimed. The taskbar launches recents by name (so
+                    // AppDriver's guard applies), but this list is a
+                    // launch-metadata producer like any other — a future consumer
+                    // reading index_url straight off it shouldn't inherit a stale
+                    // origin.
+                    const backingGoneFlags = await Promise.all(
+                        orderedApps.map((app) =>
+                            hostedIndexUrlBackingIsUnavailable({
+                                app,
+                                subdomainStore: this.stores.subdomain,
+                                config: this.config,
+                            }).catch(() => true),
+                        ),
                     );
 
-                // Don't hand out an index_url whose puter-hosted backing is
-                // gone or reclaimed. The taskbar launches recents by name (so
-                // AppDriver's guard applies), but this list is a
-                // launch-metadata producer like any other — a future consumer
-                // reading index_url straight off it shouldn't inherit a stale
-                // origin.
-                const backingGoneFlags = await Promise.all(
-                    orderedApps.map((app) =>
-                        hostedIndexUrlBackingIsUnavailable({
-                            app,
-                            subdomainStore: this.stores.subdomain,
-                            config: this.config,
-                        }).catch(() => true),
-                    ),
-                );
+                    recent = orderedApps.map((app, index) => {
+                        const backingGone = backingGoneFlags[index];
+                        return {
+                            uuid: app.uid,
+                            name: app.name,
+                            title: app.title,
+                            icon: app.icon ?? null,
+                            godmode: Boolean(app.godmode),
+                            maximize_on_start: Boolean(app.maximize_on_start),
+                            index_url: backingGone ? null : app.index_url,
+                            ...(backingGone
+                                ? { privateAccess: buildHostedBackingDenial() }
+                                : {}),
+                            // An app with no owner isn't owned by a Puter user —
+                            // it's an "external" (origin-bootstrapped) app.
+                            external:
+                                app.owner_user_id == null ||
+                                app.owner_user_id === '',
+                        };
+                    });
+                }
 
-                recent = orderedApps.map((app, index) => {
-                    const backingGone = backingGoneFlags[index];
-                    return {
-                        uuid: app.uid,
-                        name: app.name,
-                        title: app.title,
-                        icon: app.icon ?? null,
-                        godmode: Boolean(app.godmode),
-                        maximize_on_start: Boolean(app.maximize_on_start),
-                        index_url: backingGone ? null : app.index_url,
-                        ...(backingGone
-                            ? { privateAccess: buildHostedBackingDenial() }
-                            : {}),
-                        // An app with no owner isn't owned by a Puter user —
-                        // it's an "external" (origin-bootstrapped) app.
-                        external:
-                            app.owner_user_id == null ||
-                            app.owner_user_id === '',
-                    };
-                });
-            }
+                res.json({ recommended, recent });
+            },
+        );
 
-            res.json({ recommended, recent });
-        });
-
-        router.post('/suggest_apps', apiOptions, this.suggestApps);
+        router.post(
+            '/suggest_apps',
+            { ...apiOptions, rateLimit: FS_HELPER_LIMIT },
+            this.suggestApps,
+        );
 
         // puter-js polls this to decide whether to purge its in-memory FS
         // cache. SocketService bumps a per-user Redis key on every
         // `outer.gui.item.*` mutation — read it back here.
         router.get(
             '/cache/last-change-timestamp',
-            apiOptions,
+            { ...apiOptions, rateLimit: FS_POLL_LIMIT },
             async (req, res) => {
                 const userId = req.actor?.user?.id;
                 if (!userId) {
@@ -270,10 +390,14 @@ export class LegacyFSController extends PuterController {
             },
         );
 
-        router.post('/readdir-subdomains', apiOptions, this.readdirSubdomains);
+        router.post(
+            '/readdir-subdomains',
+            { ...apiOptions, rateLimit: FS_HELPER_LIMIT },
+            this.readdirSubdomains,
+        );
         router.post(
             '/update-fsentry-thumbnail',
-            apiOptions,
+            { ...apiOptions, rateLimit: FS_HELPER_LIMIT },
             this.updateFsentryThumbnail,
         );
 
@@ -908,7 +1032,7 @@ export class LegacyFSController extends PuterController {
         const actor = this.#requireActor(req);
         // Subdomain enumeration is a user-level concern; app actors would
         // otherwise see root_dir uids pointing outside their AppData scope.
-        if (effectiveActorApp(actor)) {
+        if (actor.effectiveApp) {
             res.json([]);
             return;
         }
@@ -985,7 +1109,7 @@ export class LegacyFSController extends PuterController {
         // App-under-user actors only see entries within their AppData root;
         // user actors are unscoped. Mirrors the ACL short-circuit in
         // ACLService.check.
-        const app = effectiveActorApp(actor);
+        const app = actor.effectiveApp;
         const username = actor.user?.username;
         const pathScope =
             app && typeof username === 'string' && username.length > 0
@@ -1070,32 +1194,6 @@ export class LegacyFSController extends PuterController {
         );
         res.status(range ? 206 : 200);
 
-        // Best-effort egress metering.
-        const metering = this.services.metering as
-            | {
-                  batchIncrementUsages?: (
-                      actor: unknown,
-                      entries: unknown[],
-                  ) => void;
-              }
-            | undefined;
-        if (metering?.batchIncrementUsages && download.contentLength) {
-            download.body.once('end', () => {
-                try {
-                    const bytes = download.contentLength!;
-                    metering.batchIncrementUsages!(actor, [
-                        {
-                            usageType: 'filesystem:egress:bytes',
-                            usageAmount: bytes,
-                            costOverride:
-                                FS_COSTS['filesystem:egress:bytes'] * bytes,
-                        },
-                    ]);
-                } catch {
-                    // ignore — non-critical.
-                }
-            });
-        }
         download.body.on('error', (err) => {
             res.destroy(err);
         });
@@ -1125,9 +1223,20 @@ export class LegacyFSController extends PuterController {
         assertNotSuspended(actor!.user);
         assertVerifiedAccount(actor!.user);
 
-        req.actor = actor!;
+        req.actor = assertResolvedActor(actor!);
         Context.set('actor', actor);
 
+        // And the budget gate the other read routes declare with
+        // `requireCredits`. The global auth probe only looks for `auth_token`,
+        // so `?token=` leaves `req.actor` unset for the whole gate chain and
+        // the declarative form would wave every request through — this streams
+        // file content like `/read` does, so it is refused on the same terms.
+        await assertActorHasCredits(
+            this.services.metering,
+            req.actor,
+            this.config,
+        );
+
         // Forward back to regular read after setting actor
         return this.read(req, res, undefined, { realMime: true });
     };
@@ -1514,6 +1623,23 @@ export class LegacyFSController extends PuterController {
             });
         }
 
+        // Name who this response's bytes are billed to. A signature authorises
+        // access to a file; it says nothing about who is asking, so an
+        // unidentified caller is billed to the account whose file it is —
+        // otherwise a signed URL is a way to serve content for free. A caller
+        // who did identify themselves pays for what they fetch, as on a hosted
+        // site.
+        if (owner?.uuid) {
+            req.egressActor = req.actor ?? {
+                user: {
+                    uuid: owner.uuid,
+                    id: owner.id,
+                    username: owner.username,
+                    suspended: !!(owner as { suspended?: unknown }).suspended,
+                },
+            };
+        }
+
         // Directory: return a signed listing of direct children.
         // The caller only proved read access, so strip write_url from
         // each child to prevent privilege escalation via /writeFile.
@@ -1581,16 +1707,25 @@ export class LegacyFSController extends PuterController {
     };
 
     /**
-     * POST /open_item — resolve an entry, grant the default suggested app write
+     * POST /open_item — resolve an entry, grant the default suggested app
      * access to it, and return a signed URL + user-app token so the launched
      * app can read/write the file via its app-under-user token.
      *
-     * Matches v1 semantics: permission is always granted as `write` — the
-     * underlying user's permission check still caps the effective access
-     * (grantUserAppPermission doesn't escalate user privileges).
+     * This is a user-authority action: it writes a user→app ACL row. Two things
+     * keep an app from opening an item to widen its own access — the caller
+     * must be the user themselves (the route gate, re-checked here because
+     * extensions can reach handlers directly), and the grant is capped at the
+     * access the caller actually proved on the entry.
      */
     openItem = async (req: Request, res: Response): Promise<void> => {
         const actor = this.#requireActor(req);
+        if ((actor as { app?: unknown }).app) {
+            throw new HttpError(
+                403,
+                'This endpoint is only available to user sessions',
+                { legacyCode: 'forbidden' },
+            );
+        }
         const body = asRecord(req.body);
         const entry = await resolveV1Selector(this.stores.fsEntry, body);
 
@@ -1602,8 +1737,8 @@ export class LegacyFSController extends PuterController {
             'read',
         );
 
-        // Downgrade the envelope when the caller only proved read.
-        // `/writeFile`'s ACL re-check would still block the write, but
+        // Downgrade the envelope and the grant when the caller only proved
+        // read. `/writeFile`'s ACL re-check would still block the write, but
         // returning `write_url` to a read-only caller is the same
         // privilege-leak shape that `/sign` and `/readdir` strip.
         const writeOk = await this.services.acl.check(
@@ -1631,7 +1766,7 @@ export class LegacyFSController extends PuterController {
             await this.services.permission.grantUserAppPermission(
                 actor,
                 defaultAppUid,
-                `fs:${entry.uuid}:write`,
+                `fs:${entry.uuid}:${writeOk ? 'write' : 'read'}`,
                 {},
                 { reason: 'open_item' },
             );
@@ -1727,10 +1862,10 @@ export class LegacyFSController extends PuterController {
             });
 
         // Build an actor-under-user shape for the check.
-        const actorForApp = {
-            user: (req.actor as { user?: unknown }).user,
+        const actorForApp = makeActor({
+            user: req.actor!.user,
             app: { uid: (app as { uid: string }).uid },
-        } as unknown as Actor;
+        });
         const descriptor = {
             path: subject.path,
             resolveAncestors: () =>

src/backend/controllers/fs/costs.ts

@@ -1,30 +0,0 @@
-/*
- * 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 { toMicroCents } from '../../services/metering/utils.js';
-
-// Microcents per byte. Egress roughly matches S3 data-transfer-out
-// (~$0.12/GiB); cached egress is CloudFront-backed (~$0.10/GiB).
-// Ingress and deletes are currently free.
-export const FS_COSTS = {
-    'filesystem:ingress:bytes': 0,
-    'filesystem:delete:bytes': 0,
-    'filesystem:egress:bytes': toMicroCents(0.12 / 1024 / 1024 / 1024),
-    'filesystem:cached-egress:bytes': toMicroCents(0.1 / 1024 / 1024 / 1024),
-} as const;

src/backend/controllers/fs/limits.test.ts

@@ -0,0 +1,163 @@
+/*
+ * 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 * as limits from './limits.js';
+import {
+    DEFAULT_FREE_SUBSCRIPTION,
+    DEFAULT_TEMP_SUBSCRIPTION,
+} from '../../services/metering/consts.js';
+import type { RouteOptions, RouteRateLimit } from '../../core/http/types';
+
+type Spec = RouteRateLimit | NonNullable<RouteOptions['concurrent']>;
+
+const all = Object.entries(limits) as Array<[string, Spec | RouteRateLimit[]]>;
+const flat: Array<[string, Spec]> = all.flatMap(([name, value]) =>
+    Array.isArray(value)
+        ? value.map((v, i): [string, Spec] => [`${name}[${i}]`, v])
+        : [[name, value] as [string, Spec]],
+);
+
+const windows = flat.filter(([, s]) => 'window' in s) as Array<
+    [string, RouteRateLimit]
+>;
+const concurrents = flat.filter(([, s]) => !('window' in s));
+
+describe('filesystem limit specs', () => {
+    // The whole point of this module is that the legacy and v2 controllers
+    // import the same specs. That only ties the counters together if every
+    // spec carries an explicit scope — without one the gate falls back to
+    // the route path, and the two controllers get separate budgets.
+    it.each(flat)('%s pins an explicit scope', (_name, spec) => {
+        expect(spec.scope).toBeTruthy();
+    });
+
+    it.each(windows)('%s has a positive limit and window', (_name, spec) => {
+        expect(spec.limit).toBeGreaterThan(0);
+        expect(spec.window).toBeGreaterThan(0);
+    });
+
+    // Base is the paid value; the free tiers are carved out beneath it.
+    // A free tier above the base would mean paying made you worse off.
+    it.each(windows)('%s never lets a free tier exceed paid', (_name, spec) => {
+        for (const n of Object.values(spec.bySubscription ?? {})) {
+            expect(n).toBeLessThanOrEqual(spec.limit);
+        }
+    });
+
+    it.each(windows)(
+        '%s caps temp at or below registered-free',
+        (_name, spec) => {
+            const free = spec.bySubscription?.[DEFAULT_FREE_SUBSCRIPTION];
+            const temp = spec.bySubscription?.[DEFAULT_TEMP_SUBSCRIPTION];
+            if (free === undefined || temp === undefined) return;
+            expect(temp).toBeLessThanOrEqual(free);
+        },
+    );
+
+    // A single in-flight slot turns incidental client parallelism into a
+    // spurious 429; paid tiers keep room to actually parallelise.
+    it.each(concurrents)(
+        '%s keeps concurrency at 5+ paid and 2+ for every tier',
+        (_name, spec) => {
+            expect(spec.limit).toBeGreaterThanOrEqual(5);
+            for (const n of Object.values(spec.bySubscription ?? {})) {
+                expect(n).toBeGreaterThanOrEqual(2);
+            }
+        },
+    );
+
+    it('gives search the tightest window of the read paths', () => {
+        expect(limits.FS_SEARCH_LIMIT.limit).toBeLessThan(
+            limits.FS_STAT_LIMIT.limit,
+        );
+        expect(limits.FS_SEARCH_LIMIT.limit).toBeLessThan(
+            limits.FS_READ_LIMIT.limit,
+        );
+    });
+
+    // The desktop deletes/moves one item per request with no batching and no
+    // pacing, so a single "empty trash" or "select all, delete" has to fit
+    // inside the minute window on every tier. The hour window is the abuse
+    // ceiling that a minute window this wide can no longer be.
+    it('gives mutations a bulk-sized minute window plus an hourly backstop', () => {
+        const [minute, hourly] = limits.FS_MUTATE_LIMIT;
+        expect(limits.FS_MUTATE_LIMIT).toHaveLength(2);
+        expect(minute.window).toBe(60_000);
+        expect(hourly.window).toBe(60 * 60_000);
+
+        const tiers = (spec: RouteRateLimit) => [
+            spec.limit,
+            spec.bySubscription![DEFAULT_FREE_SUBSCRIPTION],
+            spec.bySubscription![DEFAULT_TEMP_SUBSCRIPTION],
+        ];
+
+        // A few hundred items clears in one pass, anonymous included.
+        for (const n of tiers(minute)) expect(n).toBeGreaterThanOrEqual(500);
+        // Anonymous stays meaningfully tighter than paid on both windows.
+        expect(tiers(minute)[2]).toBeLessThanOrEqual(minute.limit / 2);
+        expect(tiers(hourly)[2]).toBeLessThanOrEqual(hourly.limit / 2);
+
+        // The hour window has to bind rather than decorate: it allows a
+        // handful of bulk passes an hour, not sixty minutes' worth.
+        for (const [perMinute, perHour] of tiers(minute).map(
+            (n, i): [number, number] => [n, tiers(hourly)[i]],
+        )) {
+            expect(perHour).toBeGreaterThan(perMinute);
+            expect(perHour).toBeLessThanOrEqual(perMinute * 10);
+        }
+    });
+
+    // The DAV gate is consumed imperatively, before the request has an actor:
+    // it keys on the network fingerprint and reads only `limit` / `window`. A
+    // `key: 'user'` or a `bySubscription` map here would describe tiering that
+    // never happens.
+    it('shapes the DAV specs the way the DAV gate consumes them', () => {
+        expect(limits.DAV_LIMIT.key).toBe('fingerprint');
+        expect(limits.DAV_CONCURRENT.key).toBe('fingerprint');
+        expect(limits.DAV_LIMIT.bySubscription).toBeUndefined();
+        expect(limits.DAV_CONCURRENT.bySubscription).toBeUndefined();
+    });
+
+    // With no session to key on, the alternative is the bare address — and an
+    // address is a household, an office or a carrier gateway, so keying there
+    // makes one bucket serve everyone behind it and tighten as more real users
+    // arrive. The fingerprint separates clients within a network while still
+    // being something one client can't vary per request.
+    it('keys the signed-URL routes on the network fingerprint', () => {
+        expect(limits.FS_SIGNED_READ_LIMIT.key).toBe('fingerprint');
+        expect(limits.FS_SIGNED_WRITE_LIMIT.key).toBe('fingerprint');
+        expect(limits.FS_SIGNED_CONCURRENT.key).toBe('fingerprint');
+    });
+
+    // These serve page subresources — a gallery, an app's own assets — so the
+    // in-flight cap has to sit above what a browser opens to one origin at
+    // once, or a normal page load is what trips it.
+    it('leaves the signed-URL in-flight cap above a browser`s own parallelism', () => {
+        expect(limits.FS_SIGNED_CONCURRENT.limit).toBeGreaterThan(30);
+    });
+
+    it('uses distinct scopes so counters cannot collide', () => {
+        const scopes = flat
+            .filter(([, s]) => 'window' in s)
+            .map(([, s]) => s.scope);
+        expect(new Set(scopes).size).toBe(scopes.length);
+    });
+});

src/backend/controllers/fs/limits.ts

@@ -0,0 +1,245 @@
+/*
+ * 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 {
+    DEFAULT_FREE_SUBSCRIPTION,
+    DEFAULT_TEMP_SUBSCRIPTION,
+} from '../../services/metering/consts.js';
+import type { RouteOptions, RouteRateLimit } from '../../core/http/types';
+
+// -- Shared filesystem limits ----------------------------------------
+//
+// The v2 controller and the legacy controller expose the same operations
+// on separate route tables. Both import the specs below so a caller can't
+// get two budgets for one operation by switching endpoints — the `scope`
+// is what ties the counters together, and it only works if both sides
+// pass the same one.
+//
+// The base `limit` is what subscribed tiers see; `bySubscription` carves
+// out the free tiers beneath it. Any plan id not enumerated (a Stripe
+// plan, the dev-only `unlimited`) falls through to the base, so new plans
+// are generous by default rather than accidentally throttled.
+//
+// Storage quota already bounds total bytes, and egress and object-store
+// requests are metered. These bound request *count*, which neither does:
+// metadata reads cost database time while costing the caller almost
+// nothing, and metering settles seconds behind the traffic, so a limit is
+// what actually stops a runaway loop in the moment.
+
+/** Per-user sliding window. Free tiers are carved out of the paid base. */
+const userWindow = (
+    scope: string,
+    paid: number,
+    free: number,
+    temp: number,
+    window = 60_000,
+): RouteRateLimit => ({
+    scope,
+    limit: paid,
+    window,
+    key: 'user',
+    bySubscription: {
+        [DEFAULT_FREE_SUBSCRIPTION]: free,
+        [DEFAULT_TEMP_SUBSCRIPTION]: temp,
+    },
+});
+
+/**
+ * In-flight cap. Nothing drops below 2 — a single slot turns any incidental
+ * parallelism in a client (two tabs, a prefetch alongside a user action) into a
+ * spurious 429 — and subscribed tiers keep enough headroom to actually
+ * parallelise.
+ */
+const userConcurrent = (
+    scope: string,
+    paid: number,
+    free: number,
+    temp: number,
+): NonNullable<RouteOptions['concurrent']> => ({
+    scope,
+    limit: paid,
+    key: 'user',
+    bySubscription: {
+        [DEFAULT_FREE_SUBSCRIPTION]: free,
+        [DEFAULT_TEMP_SUBSCRIPTION]: temp,
+    },
+});
+
+/**
+ * Per-network window, for the signature-authenticated routes with no session.
+ *
+ * `fingerprint` rather than `ip`: without an actor the only alternative is the
+ * bare address, and an address is an aggregate — a household, an office, a
+ * mobile carrier's gateway. Keying on it means one bucket for everyone behind
+ * it, which is a shared cap that gets tighter the more real users are present.
+ * The fingerprint folds in the headers a client can't vary per request without
+ * also looking like a different client, so separate browsers behind one NAT get
+ * separate buckets while a single client still can't mint fresh ones per call.
+ */
+const networkWindow = (
+    scope: string,
+    limit: number,
+    window = 60_000,
+): RouteRateLimit => ({ scope, limit, window, key: 'fingerprint' });
+
+// -- Metadata reads --------------------------------------------------
+
+/** Chattiest call in the GUI; high enough to only ever catch a loop. */
+export const FS_STAT_LIMIT = userWindow('fs:stat', 1200, 600, 300);
+
+/**
+ * Desktop boot fans out hard, so the minute budget is generous — the short
+ * second window is what actually catches a runaway loop before it has spent the
+ * whole minute's allowance.
+ */
+export const FS_READDIR_LIMIT: RouteRateLimit[] = [
+    userWindow('fs:readdir', 600, 300, 120),
+    userWindow('fs:readdir-burst', 120, 60, 30, 10_000),
+];
+
+/**
+ * Unindexed scan across the user's tree, and close to unmetered — a result set
+ * is a few hundred bytes of egress against an arbitrary amount of database
+ * work, so this is the cheapest way to occupy a connection. Tightest limit in
+ * the file.
+ */
+export const FS_SEARCH_LIMIT = userWindow('fs:search', 60, 30, 10);
+export const FS_SEARCH_CONCURRENT = userConcurrent('fs:search', 5, 2, 2);
+
+/** Aggregation over the whole tree; the GUI needs it rarely. */
+export const FS_DF_LIMIT = userWindow('fs:df', 60, 30, 15);
+
+// -- Content transfer ------------------------------------------------
+
+/**
+ * Egress is billed, so cost is already bounded — the cap here is against
+ * connection exhaustion, which is why the concurrency slot matters more than
+ * the window.
+ */
+export const FS_READ_LIMIT = userWindow('fs:read', 600, 300, 120);
+export const FS_READ_CONCURRENT = userConcurrent('fs:read', 10, 5, 3);
+
+export const FS_WRITE_LIMIT = userWindow('fs:write', 300, 120, 30);
+export const FS_WRITE_CONCURRENT = userConcurrent('fs:write', 15, 6, 3);
+
+/**
+ * Multipart handshake — several calls per upload, so a multiple of the write
+ * budget rather than a peer of it. Signing is the half that costs us
+ * object-store calls.
+ *
+ * The multiple is what matters: a per-minute ceiling here divides down into a
+ * much smaller number of files, and selecting a folder's worth of them is one
+ * gesture. `FS_WRITE_LIMIT` and storage quota already bound what actually gets
+ * written, so this only has to stay out of the way of a real upload.
+ */
+export const FS_MULTIPART_LIMIT = userWindow('fs:multipart', 2400, 1200, 600);
+
+// -- Metadata mutations ----------------------------------------------
+
+/**
+ * Mkdir / touch / rename / delete / move / copy / mkshortcut.
+ *
+ * The desktop issues one call per item and does not pace them: emptying the
+ * trash, deleting a multi-selection, or dragging a folder's worth of files
+ * fires the whole set back to back. A minute budget in the low hundreds turns
+ * an ordinary "select all, delete" into a partial failure, so the minute window
+ * is sized to clear a bulk pass over a few hundred items on every tier — an
+ * anonymous session gets a real desktop, so it needs a real bulk allowance too,
+ * just a smaller one.
+ *
+ * The hour window is where the abuse ceiling actually lives. It allows a few of
+ * those bulk passes per hour, which is more than a person driving a file
+ * manager will ever need and well under what a script grinding metadata writes
+ * would want.
+ */
+export const FS_MUTATE_LIMIT: RouteRateLimit[] = [
+    userWindow('fs:mutate', 1200, 900, 600),
+    userWindow('fs:mutate-sustained', 6000, 3000, 1800, 60 * 60_000),
+];
+
+/** Mints a URL that outlives the request, so worth its own budget. */
+export const FS_SIGN_LIMIT = userWindow('fs:sign', 300, 150, 60);
+
+/** Low-frequency GUI helpers. */
+export const FS_HELPER_LIMIT = userWindow('fs:helper', 120, 60, 30);
+
+/**
+ * Puter.js polls this on a timer to decide whether to purge its FS cache. Same
+ * ceiling for free and paid — it is a single cache read.
+ */
+export const FS_POLL_LIMIT = userWindow('fs:poll', 240, 240, 120);
+
+// -- Legacy multipart upload -----------------------------------------
+
+/**
+ * `/batch` buffers every file fully into memory before any quota or storage
+ * check runs, up to BATCH_MAX_FILES × BATCH_MAX_FILE_SIZE. The concurrency slot
+ * is doing the real work here; the window is secondary, and sized to say so —
+ * one upload is one call, so a per-minute ceiling in the tens is a cap on how
+ * many files someone may upload rather than a bound on cost. What actually
+ * bounds the memory this route can tie up is how many run at once.
+ */
+export const FS_BATCH_LIMIT = userWindow('fs:batch', 600, 300, 300);
+export const FS_BATCH_CONCURRENT = userConcurrent('fs:batch', 5, 2, 2);
+
+// -- Signed-URL routes (no session to key on) ------------------------
+//
+// The URL's own signature is what authorizes these; the limits below only
+// bound what an unsigned flood can cost. Sized for what the routes are
+// actually used for, which is content the browser fetches as a subresource:
+// a gallery, a document's images, an app loading its own assets. A page
+// opening a few dozen of those at once is ordinary, and every rejection here
+// is a broken image rather than a slow one — so the ceiling clears a burst of
+// real page loads and only catches something looping.
+
+export const FS_SIGNED_READ_LIMIT = networkWindow('fs:signed-read', 3_000);
+export const FS_SIGNED_WRITE_LIMIT = networkWindow('fs:signed-write', 600);
+export const FS_SIGNED_CONCURRENT: NonNullable<RouteOptions['concurrent']> = {
+    scope: 'fs:signed',
+    // Above what a browser will open to one origin at once, so the in-flight
+    // cap never decides the outcome for a single client — it is there for a
+    // client that opens connections without closing them.
+    limit: 60,
+    key: 'fingerprint',
+};
+
+// -- WebDAV ----------------------------------------------------------
+
+/**
+ * One `router.use` fronts the whole DAV surface, so a single gate there covers
+ * every verb. Desktop DAV clients are bursty — a lower ceiling shows up as
+ * spurious failures in Finder / Explorer.
+ *
+ * Unlike everything above, the DAV gate runs before the request is
+ * authenticated: there is no actor yet, so there is no subscription to resolve
+ * a tier against and one ceiling applies to every caller. These are shaped to
+ * say that — keyed on the network fingerprint the gate buckets on, with no
+ * `bySubscription` map to imply a tiering that never happens.
+ */
+export const DAV_LIMIT: RouteRateLimit = {
+    scope: 'dav',
+    limit: 600,
+    window: 60_000,
+    key: 'fingerprint',
+};
+export const DAV_CONCURRENT: NonNullable<RouteOptions['concurrent']> = {
+    scope: 'dav',
+    limit: 10,
+    key: 'fingerprint',
+};

src/backend/controllers/hosting/HostingController.js

@@ -19,6 +19,10 @@
 
 import { HttpError } from '../../core/http/HttpError.js';
 import { PuterController } from '../types.js';
+import {
+    DEFAULT_FREE_SUBSCRIPTION,
+    DEFAULT_TEMP_SUBSCRIPTION,
+} from '../../services/metering/consts.js';
 
 /**
  * Site hosting endpoints. Listing and create/update are not exposed as
@@ -46,6 +50,18 @@ export class HostingController extends PuterController {
                 requireUserActor: true,
                 allowFullAccessToken: true,
                 requireVerified: true,
+                // Destructive, and pairs with the `subdomains:create`
+                // budget on the driver side.
+                rateLimit: {
+                    scope: 'delete-site',
+                    limit: 60,
+                    window: 60_000,
+                    key: 'user',
+                    bySubscription: {
+                        [DEFAULT_FREE_SUBSCRIPTION]: 30,
+                        [DEFAULT_TEMP_SUBSCRIPTION]: 10,
+                    },
+                },
             },
             async (req, res) => {
                 const { site_uuid } = req.body ?? {};

src/backend/controllers/index.ts

@@ -18,6 +18,7 @@
  */
 
 import { AppController } from './apps/AppController.js';
+import { AppFeedbackController } from './feedback/AppFeedbackController.js';
 import { AuthController } from './auth/AuthController.js';
 import { BroadcastController } from './broadcast/BroadcastController.js';
 import { DesktopController } from './desktop/DesktopController.js';
@@ -43,6 +44,7 @@ export const puterControllers = {
     staticPages: StaticPagesController,
     auth: AuthController,
     apps: AppController,
+    appFeedback: AppFeedbackController,
     desktop: DesktopController,
     hosting: HostingController,
     system: SystemController,

src/backend/controllers/notification/NotificationController.ts

@@ -41,6 +41,14 @@ export class NotificationController extends PuterController {
         subdomain: 'api',
         requireUserActor: true,
         allowFullAccessToken: true,
+        // Fires per notification interaction, so the ceiling stays
+        // generous — it is here to catch a loop, not to pace a user.
+        rateLimit: {
+            scope: 'notification-mark',
+            limit: 300,
+            window: 60_000,
+            key: 'user',
+        },
     })
     async markAck(req: Request, res: Response): Promise<void> {
         const uid = req.body?.uid;
@@ -84,6 +92,14 @@ export class NotificationController extends PuterController {
         subdomain: 'api',
         requireUserActor: true,
         allowFullAccessToken: true,
+        // Fires per notification interaction, so the ceiling stays
+        // generous — it is here to catch a loop, not to pace a user.
+        rateLimit: {
+            scope: 'notification-mark',
+            limit: 300,
+            window: 60_000,
+            key: 'user',
+        },
     })
     async markRead(req: Request, res: Response): Promise<void> {
         const uid = req.body?.uid;

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

@@ -1746,3 +1746,35 @@ describe('OIDCController POST /auth/oidc/verify-popup-return', () => {
         expect((await redeem(proof)).body).toMatchObject({ oidc_login: false });
     });
 });
+
+// ── rate-limit scopes ───────────────────────────────────────────────
+
+describe('OIDCController rate limits', () => {
+    const rateLimitOf = (method: string, path: string) => {
+        const route = router.routes.find(
+            (r) => r.method === method && r.path === path,
+        );
+        if (!route) throw new Error(`No ${method.toUpperCase()} ${path} route`);
+        return route.options.rateLimit as { scope: string; limit: number };
+    };
+
+    it('keeps the revalidate landing page off the identity-provider bucket', () => {
+        // `/auth/revalidate-done` serves a constant HTML page; the start and
+        // callback routes exchange codes with an identity provider. Sharing
+        // a scope made the cheap page inherit the flow routes' tight ceiling,
+        // which one shared address can exhaust on its own.
+        const done = rateLimitOf('get', '/auth/revalidate-done');
+        const start = rateLimitOf('get', '/auth/oidc/:provider/start');
+        expect(done.scope).not.toBe(start.scope);
+        expect(done.scope).toBe('oidc-revalidate-done');
+        expect(done.limit).toBeGreaterThan(start.limit);
+    });
+
+    it('sizes the public provider list for a shared address, not a fleet', () => {
+        // Unauthenticated and keyed on IP, so the bucket covers every client
+        // behind one NAT or campus, not one browser. It is still login
+        // surface, so it stays bounded well below the ceilings given to
+        // static reads like icons or version info.
+        expect(rateLimitOf('get', '/auth/oidc/providers').limit).toBe(1_200);
+    });
+});

src/backend/controllers/oidc/OIDCController.ts

@@ -228,10 +228,26 @@ export class OIDCController extends PuterController {
 
         // -- GET /auth/oidc/providers --------------------------------
         // Public — list enabled provider IDs for the frontend.
+        //
+        // Every render of a login form reads this, and with no actor the
+        // bucket is the address: one corporate egress, campus or carrier
+        // gateway stands for every person behind it. Size it for that — a
+        // large shared address is thousands of people, and a shift or class
+        // change bunches their sign-ins into the same minute — but no
+        // further. This is login surface, so the ceiling should still bound
+        // someone enumerating which identity providers a deployment accepts.
 
         router.get(
             '/auth/oidc/providers',
-            { subdomain: 'api' },
+            {
+                subdomain: 'api',
+                rateLimit: {
+                    scope: 'oidc-providers',
+                    limit: 1_200,
+                    window: 60_000,
+                    key: 'ip',
+                },
+            },
             async (_req: Request, res: Response) => {
                 const providers =
                     await this.services.oidc.getEnabledProviderIds();
@@ -584,10 +600,26 @@ export class OIDCController extends PuterController {
 
         // -- GET /auth/revalidate-done -------------------------------
         // Landing page after revalidation; posts to opener for popup flow.
+        //
+        // Deliberately not on the `oidc-general` bucket the start/callback
+        // routes share: those exchange codes with an identity provider, this
+        // one is a constant HTML page that touches nothing. Sharing a bucket
+        // meant everyone reachable through one address — an office, a school,
+        // a carrier gateway — competed for the same 30 popup closes a minute.
+        // The replacement is sized for the humans behind one such address
+        // re-validating at once, not for a fleet: it is still auth surface,
+        // and one page view per revalidation is a low-volume event.
 
         router.get(
             '/auth/revalidate-done',
-            { subdomain: '' },
+            {
+                subdomain: '',
+                rateLimit: {
+                    scope: 'oidc-revalidate-done',
+                    limit: 600,
+                    window: 60_000,
+                },
+            },
             (_req: Request, res: Response) => {
                 const origin = this.config.origin ?? '';
                 res.set('Content-Type', 'text/html; charset=utf-8');

src/backend/controllers/peer/PeerController.test.ts

@@ -438,6 +438,7 @@ describe('PeerController TURN', () => {
                     id: created.id,
                     username: created.username,
                 },
+                effectiveApp: null,
             });
         } finally {
             meterSpy.mockRestore();

src/backend/controllers/peer/PeerController.ts

@@ -19,11 +19,15 @@
 
 import { createHmac, randomBytes, timingSafeEqual } from 'node:crypto';
 import type { Request, Response } from 'express';
-import type { Actor } from '../../core/actor.js';
+import { makeActor, type Actor } from '../../core/actor.js';
 import { HttpError } from '../../core/http/HttpError.js';
 import type { PuterRouter } from '../../core/http/PuterRouter.js';
 import { PuterController } from '../types.js';
 import { PEER_COSTS } from './costs.js';
+import {
+    DEFAULT_FREE_SUBSCRIPTION,
+    DEFAULT_TEMP_SUBSCRIPTION,
+} from '../../services/metering/consts.js';
 
 /**
  * Constant-time secret comparison for the internal-auth header. HMAC both sides
@@ -103,17 +107,58 @@ export class PeerController extends PuterController {
     registerRoutes(router: PuterRouter): void {
         router.get(
             '/peer/signaller-info',
-            { subdomain: 'api' },
+            {
+                subdomain: 'api',
+                // Public config read — the signaller URL and the fallback
+                // ICE list, both deploy constants. Unauthenticated, so the
+                // key is the address, and one address covers every client
+                // on that network; a peer session starts with this call, so
+                // the bucket has to hold a whole network's sessions. Nothing
+                // here is secret or expensive, so the ceiling only bounds a
+                // client stuck re-reading it.
+                rateLimit: {
+                    scope: 'peer-signaller-info',
+                    limit: 3_000,
+                    window: 60_000,
+                    key: 'ip',
+                },
+            },
             this.#signallerInfo,
         );
         router.post(
             '/peer/generate-turn',
-            { subdomain: 'api', requireAuth: true },
+            {
+                subdomain: 'api',
+                requireAuth: true,
+                // Every call reaches the upstream TURN API and mints
+                // credentials against a paid allocation, so this is a
+                // spend limit as much as an abuse limit.
+                rateLimit: {
+                    scope: 'peer-generate-turn',
+                    limit: 30,
+                    window: 60_000,
+                    key: 'user',
+                    bySubscription: {
+                        [DEFAULT_FREE_SUBSCRIPTION]: 10,
+                        [DEFAULT_TEMP_SUBSCRIPTION]: 5,
+                    },
+                },
+            },
             this.#generateTurn,
         );
         router.post(
             '/turn/ingest-usage',
-            { subdomain: 'api' },
+            {
+                subdomain: 'api',
+                // Shared-secret authenticated, so this only bounds how
+                // fast someone can guess the secret.
+                rateLimit: {
+                    scope: 'turn-ingest',
+                    limit: 60,
+                    window: 60_000,
+                    key: 'ip',
+                },
+            },
             this.#ingestUsage,
         );
     }
@@ -217,13 +262,13 @@ export class PeerController extends PuterController {
                 if (!user) continue;
                 const costInMicrocents =
                     egressBytes * PEER_COSTS['turn:egress-bytes'];
-                const actor = {
+                const actor = makeActor({
                     user: {
                         uuid: user.uuid,
                         id: user.id,
                         username: user.username,
                     },
-                };
+                });
                 await this.services.metering.incrementUsage(
                     actor,
                     'turn:egress-bytes',

src/backend/controllers/puterai/PuterAIController.test.ts

@@ -229,9 +229,18 @@ describe('PuterAIController.registerRoutes', () => {
         const modelsRoute = calls.find(
             (c) => c.path === '/puterai/chat/models',
         );
+        // Unauthenticated, so the limit keys on IP rather than an actor —
+        // which makes the bucket an aggregate over every client behind that
+        // address, hence a ceiling sized for a network rather than a browser.
         expect(modelsRoute?.opts).toEqual({
             subdomain: 'api',
             requireAuth: false,
+            rateLimit: {
+                scope: 'puterai-models',
+                limit: 3_000,
+                window: 60_000,
+                key: 'ip',
+            },
         });
     });
 

src/backend/controllers/puterai/PuterAIController.ts

@@ -90,7 +90,23 @@ export class PuterAIController extends PuterController {
                 key: aiPolicyKey,
             },
         } as RouteOptions;
-        const publicOpts = { subdomain: 'api', requireAuth: false } as const;
+        // Model listings are unauthenticated, so the only key available is
+        // the address — which is an aggregate, not a user: a NAT, a school,
+        // a mobile carrier gateway or a server-side renderer all arrive as
+        // one address, and the SDK and GUI both fetch the catalogue on
+        // startup. The ceiling therefore has to cover a whole network's page
+        // loads. What it still protects is the serialisation cost of the
+        // catalogue under a client stuck in a fetch loop.
+        const publicOpts = {
+            subdomain: 'api',
+            requireAuth: false,
+            rateLimit: {
+                scope: 'puterai-models',
+                limit: 3_000,
+                window: 60_000,
+                key: 'ip',
+            },
+        } as RouteOptions;
 
         // Every route below carries the `/puterai` prefix for wire
         // compatibility with puter-js and existing API tests.
@@ -153,7 +169,23 @@ export class PuterAIController extends PuterController {
         // URL itself is HMAC-signed, so no additional auth gate.
         router.get(
             '/puterai/video/proxy',
-            { subdomain: 'api' },
+            {
+                subdomain: 'api',
+                // HMAC-signed but unauthenticated, and it streams provider
+                // bandwidth through us — so the in-flight cap matters as
+                // much as the window.
+                rateLimit: {
+                    scope: 'puterai-video-proxy',
+                    limit: 60,
+                    window: 60_000,
+                    key: 'ip',
+                },
+                concurrent: {
+                    scope: 'puterai-video-proxy',
+                    limit: 5,
+                    key: 'ip',
+                },
+            },
             this.#videoProxy,
         );
     }

src/backend/controllers/static/StaticPagesController.ts

@@ -32,6 +32,17 @@ import { promoteToVerifiedGroup } from '../../util/userProvisioning.js';
  * All root-subdomain-only, all unauthenticated (the confirm/unsubscribe tokens
  * in the query string are the auth).
  */
+/**
+ * Unauthenticated pages reached from a token-bearing email link. Matches what
+ * the emailSend extension already applies to its own link pages.
+ */
+const TOKEN_LINK_LIMIT = {
+    scope: 'token-link-page',
+    limit: 30,
+    window: 60_000,
+    key: 'ip' as const,
+};
+
 export class StaticPagesController extends PuterController {
     registerRoutes(router: PuterRouter) {
         const origin = this.config.origin ?? '';
@@ -159,151 +170,178 @@ export class StaticPagesController extends PuterController {
         });
 
         // -- /sitemap.xml --------------------------------------------
-        router.get('/sitemap.xml', {}, async (req, res) => {
-            const domain = this.config.domain ?? req.hostname;
-            const origin = `${req.protocol}://${domain}`;
-            const apps = (await this.clients.db.read(
-                `SELECT \`name\` FROM \`apps\` WHERE \`approved_for_listing\` = ${this.clients.db.booleanLiteral(true)}`,
-            )) as Array<{ name: string }>;
-            const urls = [
-                `<url><loc>${req.protocol}://docs.${domain}/</loc></url>`,
-                ...apps.map(
-                    (a) => `<url><loc>${origin}/app/${a.name}</loc></url>`,
-                ),
-            ];
-            const body =
-                '<?xml version="1.0" encoding="UTF-8"?>' +
-                '<urlset xmlns="http://www.sitemaps.org/schemas/sitemap/0.9">' +
-                urls.join('') +
-                '</urlset>';
-            res.type('application/xml').send(body);
-        });
+        router.get(
+            '/sitemap.xml',
+            {
+                // Unauthenticated and runs a full-table scan over approved
+                // apps on every request, with no response cache in front.
+                rateLimit: {
+                    scope: 'sitemap',
+                    limit: 10,
+                    window: 60_000,
+                    key: 'ip',
+                },
+            },
+            async (req, res) => {
+                const domain = this.config.domain ?? req.hostname;
+                const origin = `${req.protocol}://${domain}`;
+                const apps = (await this.clients.db.read(
+                    `SELECT \`name\` FROM \`apps\` WHERE \`approved_for_listing\` = ${this.clients.db.booleanLiteral(true)}`,
+                )) as Array<{ name: string }>;
+                const urls = [
+                    `<url><loc>${req.protocol}://docs.${domain}/</loc></url>`,
+                    ...apps.map(
+                        (a) => `<url><loc>${origin}/app/${a.name}</loc></url>`,
+                    ),
+                ];
+                const body =
+                    '<?xml version="1.0" encoding="UTF-8"?>' +
+                    '<urlset xmlns="http://www.sitemaps.org/schemas/sitemap/0.9">' +
+                    urls.join('') +
+                    '</urlset>';
+                res.type('application/xml').send(body);
+            },
+        );
 
         // -- /unsubscribe --------------------------------------------
-        router.get('/unsubscribe', {}, async (req, res) => {
-            const userUuid =
-                typeof req.query.user_uuid === 'string'
-                    ? req.query.user_uuid
-                    : undefined;
-            if (!userUuid) {
-                res.send(err('user_uuid is required'));
-                return;
-            }
+        router.get(
+            '/unsubscribe',
+            { rateLimit: TOKEN_LINK_LIMIT },
+            async (req, res) => {
+                const userUuid =
+                    typeof req.query.user_uuid === 'string'
+                        ? req.query.user_uuid
+                        : undefined;
+                if (!userUuid) {
+                    res.send(err('user_uuid is required'));
+                    return;
+                }
 
-            const user = await this.stores.user.getByUuid(userUuid);
-            if (!user) {
-                res.send(err('User not found.'));
-                return;
-            }
-            if (user.unsubscribed) {
-                res.send(ok('You are already unsubscribed.'));
-                return;
-            }
+                const user = await this.stores.user.getByUuid(userUuid);
+                if (!user) {
+                    res.send(err('User not found.'));
+                    return;
+                }
+                if (user.unsubscribed) {
+                    res.send(ok('You are already unsubscribed.'));
+                    return;
+                }
 
-            await this.stores.user.update(user.id, { unsubscribed: 1 });
-            res.send(ok('You have successfully unsubscribed from all emails.'));
-        });
+                await this.stores.user.update(user.id, { unsubscribed: 1 });
+                res.send(
+                    ok('You have successfully unsubscribed from all emails.'),
+                );
+            },
+        );
 
         // -- /confirm-email-by-token ---------------------------------
-        router.get('/confirm-email-by-token', {}, async (req, res) => {
-            const userUuid =
-                typeof req.query.user_uuid === 'string'
-                    ? req.query.user_uuid
-                    : undefined;
-            const token =
-                typeof req.query.token === 'string'
-                    ? req.query.token
-                    : undefined;
-            if (!userUuid) {
-                res.send(err('user_uuid is required'));
-                return;
-            }
-            if (!token) {
-                res.send(err('token is required'));
-                return;
-            }
+        router.get(
+            '/confirm-email-by-token',
+            { rateLimit: TOKEN_LINK_LIMIT },
+            async (req, res) => {
+                const userUuid =
+                    typeof req.query.user_uuid === 'string'
+                        ? req.query.user_uuid
+                        : undefined;
+                const token =
+                    typeof req.query.token === 'string'
+                        ? req.query.token
+                        : undefined;
+                if (!userUuid) {
+                    res.send(err('user_uuid is required'));
+                    return;
+                }
+                if (!token) {
+                    res.send(err('token is required'));
+                    return;
+                }
 
-            const user = await this.stores.user.getByProperty(
-                'uuid',
-                userUuid,
-                { force: true },
-            );
-            if (!user) {
-                res.send(err('user not found.'));
-                return;
-            }
-            if (user.email_confirmed) {
-                res.send(ok('Email already confirmed.'));
-                return;
-            }
-            if (user.email_confirm_token !== token) {
-                res.send(err('invalid token.'));
-                return;
-            }
+                const user = await this.stores.user.getByProperty(
+                    'uuid',
+                    userUuid,
+                    { force: true },
+                );
+                if (!user) {
+                    res.send(err('user not found.'));
+                    return;
+                }
+                if (user.email_confirmed) {
+                    res.send(ok('Email already confirmed.'));
+                    return;
+                }
+                if (user.email_confirm_token !== token) {
+                    res.send(err('invalid token.'));
+                    return;
+                }
 
-            // v2 writes `clean_email` at signup (lowercased email). Older rows
-            // that predate that may be null — fall back to email.lower().
-            const cleanEmail =
-                (user.clean_email as string | null | undefined) ??
-                String(user.email ?? '').toLowerCase();
+                // v2 writes `clean_email` at signup (lowercased email). Older rows
+                // that predate that may be null — fall back to email.lower().
+                const cleanEmail =
+                    (user.clean_email as string | null | undefined) ??
+                    String(user.email ?? '').toLowerCase();
 
-            const [dupe] = (await this.clients.db.read(
-                `SELECT EXISTS(
+                const [dupe] = (await this.clients.db.read(
+                    `SELECT EXISTS(
                     SELECT 1 FROM \`user\` WHERE (\`email\` = ? OR \`clean_email\` = ?)
                     AND \`email_confirmed\` = ${this.clients.db.booleanLiteral(true)}
                     AND \`password\` IS NOT NULL
                 ) AS email_exists`,
-                [user.email, cleanEmail],
-            )) as Array<{ email_exists: number }>;
-            if (dupe?.email_exists) {
-                res.send(
-                    err('This email was confirmed on a different account.'),
+                    [user.email, cleanEmail],
+                )) as Array<{ email_exists: number }>;
+                if (dupe?.email_exists) {
+                    res.send(
+                        err('This email was confirmed on a different account.'),
+                    );
+                    return;
+                }
+
+                // Revoke any other accounts' pending change-email slots targeting
+                // this address — they're no longer valid once someone confirms it.
+                await this.clients.db.write(
+                    'UPDATE `user` SET `unconfirmed_change_email` = NULL, `change_email_confirm_token` = NULL WHERE `unconfirmed_change_email` = ?',
+                    [user.email],
                 );
-                return;
-            }
 
-            // Revoke any other accounts' pending change-email slots targeting
-            // this address — they're no longer valid once someone confirms it.
-            await this.clients.db.write(
-                'UPDATE `user` SET `unconfirmed_change_email` = NULL, `change_email_confirm_token` = NULL WHERE `unconfirmed_change_email` = ?',
-                [user.email],
-            );
+                await this.stores.user.update(user.id, {
+                    email_confirmed: 1,
+                    requires_email_confirmation: 0,
+                    email_confirm_code: null,
+                    email_confirm_token: null,
+                });
 
-            await this.stores.user.update(user.id, {
-                email_confirmed: 1,
-                requires_email_confirmation: 0,
-                email_confirm_code: null,
-                email_confirm_token: null,
-            });
-
-            await promoteToVerifiedGroup(this.stores.group, this.config, user);
-
-            // Best-effort side-channels — don't fail the user-visible response
-            // if sockets or the event bus are unavailable.
-            try {
-                await this.services.socket.send(
-                    { room: user.id },
-                    'user.email_confirmed',
-                    {},
+                await promoteToVerifiedGroup(
+                    this.stores.group,
+                    this.config,
+                    user,
                 );
-            } catch {
-                /* ignore */
-            }
-            try {
-                this.clients.event?.emit(
-                    'user.email-confirmed',
-                    {
-                        user_id: user.id,
-                        user_uid: user.uuid,
-                        email: user.email,
-                    },
-                    {},
-                );
-            } catch {
-                /* ignore */
-            }
 
-            res.send(ok('Your email has been successfully confirmed.'));
-        });
+                // Best-effort side-channels — don't fail the user-visible response
+                // if sockets or the event bus are unavailable.
+                try {
+                    await this.services.socket.send(
+                        { room: user.id },
+                        'user.email_confirmed',
+                        {},
+                    );
+                } catch {
+                    /* ignore */
+                }
+                try {
+                    this.clients.event?.emit(
+                        'user.email-confirmed',
+                        {
+                            user_id: user.id,
+                            user_uid: user.uuid,
+                            email: user.email,
+                        },
+                        {},
+                    );
+                } catch {
+                    /* ignore */
+                }
+
+                res.send(ok('Your email has been successfully confirmed.'));
+            },
+        );
     }
 }

src/backend/controllers/system/SystemController.js

@@ -25,6 +25,75 @@ import { PuterController } from '../types.js';
  *
  * These are all low-risk, authenticated or not, and mostly stateless.
  */
+/**
+ * Liveness polling. The callers here are infrastructure, not people: a load
+ * balancer, an orchestrator and any external uptime prober all poll this, and
+ * they typically egress from a small set of addresses. A 429 here is read as an
+ * unhealthy node and takes the node out of rotation, so the ceiling is set
+ * where only a runaway loop can reach it. The handler itself reads a status
+ * snapshot refreshed on a background timer, so the per-request cost is close to
+ * nil.
+ *
+ * `memory` rather than the shared default, and that choice is load-bearing:
+ *
+ * - This route decides whether a node stays in rotation, so it must not depend on
+ *   anything it isn't already reporting on. The default backend is redis, and
+ *   the cluster is configured with an offline queue and no per-command timeout
+ *   — so while redis is unreachable a gated request waits on it rather than
+ *   failing fast. The gate does fail open, but only once the call rejects, and
+ *   the ALB gives a target 4s per probe and evicts after two. A redis
+ *   degradation could therefore empty every target group in every region, which
+ *   is the outcome the `@dependencies` degrade rules in the health check query
+ *   exist to prevent. Keeping the counter in-process removes redis from the
+ *   liveness path entirely.
+ * - Per-node counting is also the more honest bucket here. The ceiling only ever
+ *   needs to cover the pollers hitting _this_ node, not (pollers x fleet size)
+ *   as a shared counter does.
+ */
+const HEALTHCHECK_LIMIT = {
+    scope: 'healthcheck',
+    limit: 30_000,
+    window: 60_000,
+    key: 'ip',
+    backend: 'memory',
+};
+
+/**
+ * Deploy-constant build info, polled by clients. One address is a NAT, a
+ * campus, a proxy or a server-side renderer, so this bucket aggregates every
+ * client behind it — sizing it for a single browser would throttle a whole
+ * office. The response is cached per-client for a minute, which bounds each
+ * client to roughly one hit per window; the ceiling is what is left to catch a
+ * client that ignores the cache.
+ */
+const VERSION_LIMIT = {
+    scope: 'version',
+    limit: 6_000,
+    window: 60_000,
+    key: 'ip',
+};
+
+/**
+ * Deploy-constant deployment identity, read once per page load to decide
+ * whether to offer signup. Same aggregation as `/version` — the bucket is a
+ * whole network's worth of clients — and the payload is four constants, so the
+ * limit only guards against an unbounded client loop.
+ */
+const WHOAREWE_LIMIT = {
+    scope: 'whoarewe',
+    limit: 6_000,
+    window: 60_000,
+    key: 'ip',
+};
+
+/** Static introspection output, read once at boot rather than in a loop. */
+const LSMOD_LIMIT = {
+    scope: 'lsmod',
+    limit: 60,
+    window: 60_000,
+    key: 'user',
+};
+
 export class SystemController extends PuterController {
     constructor(config, clients, stores, services, drivers) {
         super(config, clients, stores, services, drivers);
@@ -44,7 +113,10 @@ export class SystemController extends PuterController {
         // `?ignore=a,b` disregards the named checks for this request only.
         // `?marked-degraded=a,b` demotes the named checks to a non-fatal
         // `degraded` list: `ok` stays true but the response is 207 so the
-        // caller can tell the node is running in a degraded state.
+        // caller can tell the node is running in a degraded state. Either list
+        // accepts `@<group>` to stand for every check in a group — notably
+        // `@dependencies` for the backing-service probes — so a caller polling
+        // this route doesn't have to enumerate them.
         const parseNames = (value) =>
             typeof value === 'string'
                 ? value
@@ -52,44 +124,53 @@ export class SystemController extends PuterController {
                       .map((name) => name.trim())
                       .filter(Boolean)
                 : [];
-        router.get('/healthcheck', { subdomain: '*' }, async (req, res) => {
-            const health = this.services.health;
-            if (!health || typeof health.getStatus !== 'function') {
-                // Fallback for boot ordering / missing service.
-                return res.send('ok');
-            }
-            const status = await health.getStatus({
-                ignore: parseNames(req.query.ignore),
-                degrade: parseNames(req.query['marked-degraded']),
-            });
-            if (!status.ok) return res.status(503).json(status);
-            if (status.degraded?.length) return res.status(207).json(status);
-            return res.json(status);
-        });
+        router.get(
+            '/healthcheck',
+            { subdomain: '*', rateLimit: HEALTHCHECK_LIMIT },
+            async (req, res) => {
+                const health = this.services.health;
+                if (!health || typeof health.getStatus !== 'function') {
+                    // Fallback for boot ordering / missing service.
+                    return res.send('ok');
+                }
+                const status = await health.getStatus({
+                    ignore: parseNames(req.query.ignore),
+                    degrade: parseNames(req.query['marked-degraded']),
+                });
+                if (!status.ok) return res.status(503).json(status);
+                if (status.degraded?.length)
+                    return res.status(207).json(status);
+                return res.json(status);
+            },
+        );
 
         // -- Version -------------------------------------------------
 
-        router.get('/version', { subdomain: '*' }, (_req, res) => {
-            const version =
-                this.config.version ??
-                process.env.npm_package_version ??
-                'unknown';
-            const parts = String(version).split('.');
-            // Deploy-constant, and callers poll it. Cache per-client only:
-            // a shared cache could pin one region's `location` for everyone,
-            // and the short window still bounds how long a client can miss a
-            // new deploy.
-            res.setHeader('Cache-Control', 'private, max-age=60');
-            res.json({
-                version,
-                major: parts[0] ? Number(parts[0]) : null,
-                minor: parts[1] ? Number(parts[1]) : null,
-                patch: parts[2] ? Number(parts[2]) : null,
-                environment: this.config.env ?? 'prod',
-                location: this.config.serverId ?? null,
-                deploy_timestamp: this.bootTime,
-            });
-        });
+        router.get(
+            '/version',
+            { subdomain: '*', rateLimit: VERSION_LIMIT },
+            (_req, res) => {
+                const version =
+                    this.config.version ??
+                    process.env.npm_package_version ??
+                    'unknown';
+                const parts = String(version).split('.');
+                // Deploy-constant, and callers poll it. Cache per-client only:
+                // a shared cache could pin one region's `location` for everyone,
+                // and the short window still bounds how long a client can miss a
+                // new deploy.
+                res.setHeader('Cache-Control', 'private, max-age=60');
+                res.json({
+                    version,
+                    major: parts[0] ? Number(parts[0]) : null,
+                    minor: parts[1] ? Number(parts[1]) : null,
+                    patch: parts[2] ? Number(parts[2]) : null,
+                    environment: this.config.env ?? 'prod',
+                    location: this.config.serverId ?? null,
+                    deploy_timestamp: this.bootTime,
+                });
+            },
+        );
 
         // -- Contact us ----------------------------------------------
 
@@ -153,7 +234,7 @@ export class SystemController extends PuterController {
 
         // -- GET /whoarewe -------------------------------------------
 
-        router.get('/whoarewe', {}, (_req, res) => {
+        router.get('/whoarewe', { rateLimit: WHOAREWE_LIMIT }, (_req, res) => {
             res.json({
                 name: 'Puter',
                 version: this.config.version ?? null,
@@ -181,8 +262,16 @@ export class SystemController extends PuterController {
             }
             res.json({ interfaces });
         };
-        router.get('/lsmod', { subdomain: 'api', requireAuth: true }, lsmod);
-        router.post('/lsmod', { subdomain: 'api', requireAuth: true }, lsmod);
+        router.get(
+            '/lsmod',
+            { subdomain: 'api', requireAuth: true, rateLimit: LSMOD_LIMIT },
+            lsmod,
+        );
+        router.post(
+            '/lsmod',
+            { subdomain: 'api', requireAuth: true, rateLimit: LSMOD_LIMIT },
+            lsmod,
+        );
     }
 
     onServerStart() {}

src/backend/controllers/system/SystemController.test.ts

@@ -141,6 +141,19 @@ const callRoute = async (
 // ── /healthcheck ────────────────────────────────────────────────────
 
 describe('SystemController GET /healthcheck', () => {
+    // This route decides whether a node stays in rotation, so its rate limit
+    // must not reach for a backing service. The default backend is redis,
+    // whose client queues rather than fails fast while it is unreachable —
+    // enough to push a probe past the 4s the load balancer allows and evict
+    // every target during a redis degradation. In-process counting keeps the
+    // liveness path free of anything it is itself reporting on.
+    it('counts in-process, so liveness never waits on redis', () => {
+        const route = router.routes.find(
+            (r) => r.method === 'get' && r.path === '/healthcheck',
+        );
+        expect(route?.options.rateLimit?.backend).toBe('memory');
+    });
+
     it('returns the live ServerHealthService status payload', async () => {
         const { res, captured } = makeRes();
         await callRoute('get', '/healthcheck', makeReq({}), res);
@@ -461,6 +474,52 @@ describe('SystemController GET /lsmod', () => {
     });
 });
 
+// ── rate-limit scopes ───────────────────────────────────────────────
+
+describe('SystemController public route rate limits', () => {
+    const rateLimitOf = (path: string) => {
+        const route = router.routes.find(
+            (r) => r.method === 'get' && r.path === path,
+        );
+        if (!route) throw new Error(`No GET ${path} route`);
+        return route.options.rateLimit as {
+            scope: string;
+            limit: number;
+            window: number;
+            key: string;
+        };
+    };
+
+    it('gives /healthcheck, /version and /whoarewe separate buckets', () => {
+        // These once shared one scope, which meant clients polling /version
+        // could exhaust the budget that liveness probes depend on. Keep them
+        // apart: a 429 on /healthcheck is read as an unhealthy node.
+        const scopes = [
+            rateLimitOf('/healthcheck').scope,
+            rateLimitOf('/version').scope,
+            rateLimitOf('/whoarewe').scope,
+        ];
+        expect(new Set(scopes).size).toBe(3);
+        expect(scopes).toEqual(['healthcheck', 'version', 'whoarewe']);
+    });
+
+    it('sizes the unauthenticated buckets for a shared address, not one client', () => {
+        // All three key on IP, and an IP is a NAT, a campus or a carrier
+        // gateway — the bucket aggregates everyone behind it, and the
+        // limiter counts region-wide rather than per process.
+        for (const path of ['/healthcheck', '/version', '/whoarewe']) {
+            const limit = rateLimitOf(path);
+            expect(limit.key).toBe('ip');
+            expect(limit.window).toBe(60_000);
+            expect(limit.limit).toBeGreaterThanOrEqual(6_000);
+        }
+        // Liveness polling is the most generous of the three by design.
+        expect(rateLimitOf('/healthcheck').limit).toBe(30_000);
+        expect(rateLimitOf('/version').limit).toBe(6_000);
+        expect(rateLimitOf('/whoarewe').limit).toBe(6_000);
+    });
+});
+
 // ── lifecycle ───────────────────────────────────────────────────────
 
 describe('SystemController.onServerPrepareShutdown', () => {

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

@@ -64,7 +64,20 @@ const makeRes = () => {
         headers: {},
         ended: false,
     };
+    const listeners: Record<string, Array<() => void>> = {};
     const res = {
+        // The DAV mount holds a concurrency slot for the life of the request
+        // and releases it on `finish` / `close`, so the stub has to behave
+        // like an emitter or every dispatch throws.
+        once: vi.fn((event: string, fn: () => void) => {
+            (listeners[event] ??= []).push(fn);
+            return res;
+        }),
+        emit: vi.fn((event: string) => {
+            const fns = listeners[event] ?? [];
+            listeners[event] = [];
+            for (const fn of fns) fn();
+        }),
         json: vi.fn((value: unknown) => {
             captured.body = value;
             return res;
@@ -89,10 +102,12 @@ const makeRes = () => {
         }),
         send: vi.fn((value: unknown) => {
             captured.body = value;
+            res.emit('finish');
             return res;
         }),
         end: vi.fn(() => {
             captured.ended = true;
+            res.emit('finish');
             return res;
         }),
         headersSent: false,
@@ -962,6 +977,11 @@ describe('WebDAVController verbs', () => {
             send: (value: unknown) => {
                 captured.body = value;
                 captured.ended = true;
+                // End the underlying Writable so `finish` fires, as it does
+                // on a real response. The DAV mount releases its concurrency
+                // slot on that event — without it every `send()` path would
+                // leak a slot and later requests would 429.
+                if (!sink.writableEnded) sink.end();
                 return res;
             },
             headersSent: false,

src/backend/controllers/webdav/WebDAVController.ts

@@ -21,7 +21,7 @@ import { compare as bcryptCompare } from 'bcrypt';
 import type { Request, Response } from 'express';
 import { posix as pathPosix } from 'node:path';
 import { EventMap } from '../../clients/event/types.js';
-import type { Actor } from '../../core/actor.js';
+import { makeActor, type Actor } from '../../core/actor.js';
 import { HttpError } from '../../core/http/HttpError.js';
 import {
     assertNotSuspended,
@@ -42,6 +42,13 @@ import {
     hasWritePermission,
     refreshLock,
 } from './locks.js';
+import { DAV_CONCURRENT, DAV_LIMIT } from '../fs/limits.js';
+import {
+    acquireConcurrent,
+    checkRateLimit,
+    computeNetworkFingerprint,
+} from '../../core/http/middleware/rateLimit.js';
+import { assertActorHasCredits } from '../../services/metering/enforcement.js';
 
 const DAV_HEADERS = {
     DAV: '1, 2, ordered-collections',
@@ -54,6 +61,13 @@ const ALLOW_METHODS =
 // macOS creates these files; reject them to keep the FS clean.
 const MACOS_JUNK_REGEX = /(?:^\.DS_Store$|^\._)/;
 
+/**
+ * Verbs that move file content or duplicate it in the object store, and so are
+ * refused to an account with nothing left of its budget. HEAD is here with GET
+ * because a client asking for headers is a client about to fetch the body.
+ */
+const CREDIT_GATED_DAV_METHODS = new Set(['GET', 'HEAD', 'PUT', 'COPY']);
+
 /**
  * WebDAV controller — full RFC 4918 surface on the `dav.*` subdomain.
  *
@@ -70,10 +84,17 @@ export class WebDAVController extends PuterController {
         // Single catch-all on the `dav` subdomain. We dispatch by req.method
         // inside the handler because WebDAV uses non-standard HTTP verbs that
         // Express doesn't have first-class router methods for in all versions.
+        //
+        // The rate limit is applied inside the handler rather than through
+        // `RouteOptions`. For a `use` mount the subdomain check lives in the
+        // handler wrapper, not in the middleware chain — so a `rateLimit`
+        // here would run for every request on every subdomain and count
+        // non-DAV traffic against the DAV budget.
         router.use(
             { subdomain: 'dav' },
             async (req: Request, res: Response, _next) => {
                 try {
+                    if (!(await this.#admit(req, res))) return;
                     await this.#dispatch(req, res);
                 } catch (err) {
                     if (err instanceof HttpError) {
@@ -88,6 +109,41 @@ export class WebDAVController extends PuterController {
         );
     }
 
+    /**
+     * Rate + concurrency gate for the whole DAV surface. Returns false when the
+     * request was rejected (429 already sent).
+     *
+     * Runs before `#dispatch` authenticates, so it keys on the network
+     * fingerprint rather than an actor. That is the coarser bucket, but a DAV
+     * client sends credentials on every request anyway — there is no
+     * unauthenticated browsing phase to protect a per-user key from.
+     */
+    async #admit(req: Request, res: Response): Promise<boolean> {
+        const key = computeNetworkFingerprint(req);
+        if (
+            !(await checkRateLimit(
+                `${DAV_LIMIT.scope}:${key}`,
+                DAV_LIMIT.limit,
+                DAV_LIMIT.window,
+            ))
+        ) {
+            res.status(429).send('Too many requests.');
+            return false;
+        }
+        const slot = await acquireConcurrent(
+            `${DAV_CONCURRENT.scope}:${key}`,
+            DAV_CONCURRENT.limit,
+        );
+        if (!slot.ok) {
+            res.status(429).send('Too many concurrent requests.');
+            return false;
+        }
+        // `finish` and `close` can both fire; release is once-only.
+        res.once('finish', () => void slot.release());
+        res.once('close', () => void slot.release());
+        return true;
+    }
+
     async #dispatch(req: Request, res: Response): Promise<void> {
         // Authenticate
         const actor = await this.#resolveActor(req, res);
@@ -104,6 +160,19 @@ export class WebDAVController extends PuterController {
         assertNotSuspended(actor.user);
         assertVerifiedAccount(actor.user);
 
+        // And the same budget gate the FS routes declare with
+        // `requireCredits`, for the verbs that move content — DAV serves the
+        // same files over a metered host, so leaving it out would make mounting
+        // the drive the way around enforcement. The verbs that only describe or
+        // remove things stay open, as they do over HTTP.
+        if (CREDIT_GATED_DAV_METHODS.has(req.method.toUpperCase())) {
+            await assertActorHasCredits(
+                this.services.metering,
+                actor,
+                this.config,
+            );
+        }
+
         // Expand `~`/`~/...` against the authenticated actor's username.
         // WebDAV doesn't standardize `~`, but some clients do — and the
         // pre-existing behaviour silently expanded it via the FS store.
@@ -238,7 +307,7 @@ export class WebDAVController extends PuterController {
         }
 
         // Build a session-less actor for the user
-        return {
+        return makeActor({
             user: {
                 id: user.id,
                 uuid: user.uuid,
@@ -258,7 +327,7 @@ export class WebDAVController extends PuterController {
                 requires_card_verification:
                     user.requires_card_verification ?? false,
             },
-        };
+        });
     }
 
     // -- OPTIONS ------------------------------------------------------

src/backend/controllers/wisp/WispController.ts

@@ -21,6 +21,10 @@ import type { Request, Response } from 'express';
 import { HttpError } from '../../core/http/HttpError.js';
 import type { PuterRouter } from '../../core/http/PuterRouter.js';
 import { PuterController } from '../types.js';
+import {
+    DEFAULT_FREE_SUBSCRIPTION,
+    DEFAULT_TEMP_SUBSCRIPTION,
+} from '../../services/metering/consts.js';
 
 /**
  * WISP relay token controller — create and verify short-lived JWT tokens for
@@ -32,12 +36,39 @@ export class WispController extends PuterController {
     registerRoutes(router: PuterRouter): void {
         router.post(
             '/wisp/relay-token/create',
-            { subdomain: 'api', requireAuth: true },
+            {
+                subdomain: 'api',
+                requireAuth: true,
+                // Auth is optional in practice (the handler tolerates an
+                // anonymous actor), so the key falls back to a fingerprint
+                // when there is no user to key on.
+                rateLimit: {
+                    scope: 'wisp-token-create',
+                    limit: 60,
+                    window: 60_000,
+                    key: 'user',
+                    bySubscription: {
+                        [DEFAULT_FREE_SUBSCRIPTION]: 30,
+                        [DEFAULT_TEMP_SUBSCRIPTION]: 10,
+                    },
+                },
+            },
             this.#create,
         );
         router.post(
             '/wisp/relay-token/verify',
-            { subdomain: 'api', requireAuth: false },
+            {
+                subdomain: 'api',
+                requireAuth: false,
+                // Unauthenticated by design, which makes it a token-guessing
+                // oracle without a ceiling.
+                rateLimit: {
+                    scope: 'wisp-token-verify',
+                    limit: 300,
+                    window: 60_000,
+                    key: 'ip',
+                },
+            },
             this.#verify,
         );
     }

src/backend/core/actor.test.ts

@@ -0,0 +1,104 @@
+/*
+ * 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 {
+    assertResolvedActor,
+    makeActor,
+    SYSTEM_ACTOR,
+    userRelatedActor,
+    type Actor,
+} from './actor';
+
+describe('makeActor / effectiveApp', () => {
+    const user = { uuid: 'u-1', id: 1, username: 'u' };
+
+    it('resolves a plain user actor to no app', () => {
+        expect(makeActor({ user }).effectiveApp).toBeNull();
+    });
+
+    it("resolves an app-under-user actor to its own app", () => {
+        expect(makeActor({ user, app: { uid: 'app-1' } }).effectiveApp).toEqual({
+            uid: 'app-1',
+        });
+    });
+
+    it("resolves a token to the app that issued it", () => {
+        // The whole point: a token actor carries no `app` of its own, so
+        // anything reading `app` sees a bare user token and skips app gating.
+        const issuer = makeActor({ user, app: { uid: 'app-1' } });
+        const token = makeActor({
+            user,
+            accessToken: { uid: 'tok-1', issuer },
+        });
+        expect(token.app).toBeUndefined();
+        expect(token.effectiveApp).toEqual({ uid: 'app-1' });
+    });
+
+    it('collapses a chain of tokens in one hop', () => {
+        const issuer = makeActor({ user, app: { uid: 'app-1' } });
+        const inner = makeActor({ user, accessToken: { uid: 't1', issuer } });
+        const outer = makeActor({
+            user,
+            accessToken: { uid: 't2', issuer: inner },
+        });
+        expect(outer.effectiveApp).toEqual({ uid: 'app-1' });
+    });
+
+    it('resolves a user-issued token to no app', () => {
+        const issuer = makeActor({ user });
+        const token = makeActor({
+            user,
+            accessToken: { uid: 'tok-1', issuer, fullAccess: true },
+        });
+        expect(token.effectiveApp).toBeNull();
+    });
+
+    it('drops the app when narrowing to the underlying user', () => {
+        const app = makeActor({ user, app: { uid: 'app-1' } });
+        expect(userRelatedActor(app).effectiveApp).toBeNull();
+    });
+
+    it('resolves the system actor', () => {
+        expect(SYSTEM_ACTOR.effectiveApp).toBeNull();
+    });
+});
+
+describe('assertResolvedActor', () => {
+    it('passes a resolved actor through unchanged', () => {
+        const actor = makeActor({ user: { uuid: 'u-1' } });
+        expect(assertResolvedActor(actor)).toBe(actor);
+        // `null` is a resolved answer, not a missing one.
+        expect(assertResolvedActor({ user: {}, effectiveApp: null })).toEqual({
+            user: {},
+            effectiveApp: null,
+        });
+    });
+
+    it('throws on an actor that skipped makeActor', () => {
+        // The field is optional so pre-existing literals still compile, which
+        // means an unresolved one can reach a gate. Fail loudly at the edge
+        // rather than let a gate read `undefined` as "no app" and wave it
+        // through — an app-under-user actor is the dangerous case.
+        const unresolved = { user: { uuid: 'u-1' }, app: { uid: 'app-1' } };
+        expect(() => assertResolvedActor(unresolved as Actor)).toThrow(
+            /effectiveApp/,
+        );
+    });
+});

src/backend/core/actor.ts

@@ -47,6 +47,26 @@ export interface ActorAccessToken {
 export interface Actor {
     user: Partial<UserRow>;
     app?: ActorApp | null;
+    /**
+     * The app this actor ultimately acts as: its own `app`, or failing that the
+     * app of whoever issued its access token.
+     *
+     * Read this — not `app` — in any gate asking "which app is doing this?". An
+     * access-token actor carries no `app` of its own, so `app` alone reads as
+     * "no app" even for a token an app minted, and a gate keyed off it fails
+     * open exactly where it must not.
+     *
+     * `null` and `undefined` are not the same thing here:
+     *
+     * - `null` — resolved, and this actor is not acting as any app.
+     * - Absent — never resolved, because the actor skipped `makeActor`.
+     *
+     * A gate must not read the second as the first: that is the fail-open this
+     * field exists to prevent. Optional only so an actor literal that predates
+     * the field still compiles; `assertResolvedActor` is what keeps the request
+     * path honest, and `makeActor` is the one place the derivation lives.
+     */
+    effectiveApp?: ActorApp | null;
     /** True for the system actor; skips metering / quota tracking. */
     system?: boolean;
     accessToken?: ActorAccessToken | null;
@@ -67,9 +87,41 @@ export const SYSTEM_ACTOR_UUID = '5d4adce0-a381-4982-9c02-6e2540026238';
 /** The default system actor used when no actor is supplied. */
 export const SYSTEM_ACTOR: Actor = {
     user: { uuid: SYSTEM_ACTOR_UUID, username: 'system' },
+    effectiveApp: null,
     system: true,
 };
 
+/**
+ * Build an actor, deriving `effectiveApp` from its own app and, failing that,
+ * from the app of whoever issued its access token.
+ *
+ * The issuer was itself built here, so its chain is already collapsed — one hop
+ * is enough, and no caller has to walk anything.
+ */
+export const makeActor = (actor: Omit<Actor, 'effectiveApp'>): Actor => ({
+    ...actor,
+    effectiveApp: actor.app ?? actor.accessToken?.issuer.effectiveApp ?? null,
+});
+
+/**
+ * Fail closed on an actor whose `effectiveApp` was never derived.
+ *
+ * Every actor on the request path is built by `AuthService` through
+ * `makeActor`, so this cannot fire in production — which is the point. It turns
+ * a future actor literal that skips the builder into a loud 500 at the edge
+ * rather than a silent bypass deep inside a gate that read `undefined` as "no
+ * app". Call it once, where the request actor is established.
+ */
+export const assertResolvedActor = (actor: Actor): Actor => {
+    if (actor.effectiveApp === undefined) {
+        throw new Error(
+            'actor was built without `makeActor`: `effectiveApp` is unresolved, ' +
+                'and app-scoped gates would read that as "no app"',
+        );
+    }
+    return actor;
+};
+
 export const isSystemActor = (actor: Actor | undefined | null): boolean => {
     return !!actor?.system || actor?.user?.uuid === SYSTEM_ACTOR_UUID;
 };
@@ -106,15 +158,5 @@ export const actorUid = (actor: Actor): string => {
  */
 export const userRelatedActor = (actor: Actor): Actor => {
     if (!actor.app && !actor.accessToken) return actor;
-    return { user: actor.user };
-};
-
-/**
- * Walk the access-token issuer chain and return the first app identity found,
- * or null if no app appears anywhere in the chain.
- */
-export const effectiveActorApp = (actor: Actor): ActorApp | null => {
-    if (actor.app) return actor.app;
-    if (actor.accessToken) return effectiveActorApp(actor.accessToken.issuer);
-    return null;
+    return { user: actor.user, effectiveApp: null };
 };

src/backend/core/http/expressAugmentation.ts

@@ -18,6 +18,7 @@
  */
 
 import type { Actor } from '../actor';
+import type { StorageOpCounts } from '../storageOps';
 import type { TokenSource } from './types';
 
 /**
@@ -103,6 +104,22 @@ declare global {
              * middleware.
              */
             cookies?: Record<string, string>;
+
+            /**
+             * Who this response's egress is billed to, when that isn't the
+             * actor who made the request. Set by handlers that serve one
+             * account's bytes to an unidentified caller — a hosted site's
+             * visitor being the case that matters. Takes precedence over
+             * `actor` in the egress middleware.
+             */
+            egressActor?: Actor;
+
+            /**
+             * Object-store requests made while serving this request, by class.
+             * Tallied through `recordStorageOps` and billed when the response
+             * ends.
+             */
+            storageOps?: StorageOpCounts;
         }
     }
 }

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

@@ -21,7 +21,7 @@ import type { Request, Response } from 'express';
 import jwt from 'jsonwebtoken';
 import { afterAll, beforeAll, describe, expect, it, vi } from 'vitest';
 import { v4 as uuidv4 } from 'uuid';
-import type { Actor } from '../../actor';
+import { makeActor, type Actor } from '../../actor';
 import type { AuthService } from '../../../services/auth/AuthService';
 import { PuterServer } from '../../../server';
 import { setupTestServer } from '../../../testUtil';
@@ -59,7 +59,16 @@ const makeStubAuth = (defaultActor: Actor | null = null): StubAuth => {
         authenticate: async (token: string) => {
             seenTokens.push(token);
             if (nextResult === 'throw') throw new Error('verify failed');
-            return nextResult;
+            // The real service builds every actor through `makeActor`, and the
+            // probe asserts that contract — so the stub has to honour it too,
+            // rather than handing back a literal the probe rejects.
+            if (!('actor' in nextResult)) return nextResult;
+            // Resolve only if the fixture didn't: the probe asserts the
+            // `makeActor` contract the real service satisfies, but tests that
+            // check actor identity need the same object back.
+            return nextResult.actor.effectiveApp === undefined
+                ? { ...nextResult, actor: makeActor(nextResult.actor) }
+                : nextResult;
         },
         // Back-compat wrapper for callers that still want Actor | null.
         authenticateFromToken: async (token: string) => {
@@ -273,7 +282,7 @@ describe('createAuthProbe — tokenSource', () => {
         ['handshake', { handshakeQuery: { auth_token: 'tok' } }],
     ];
 
-    const actor: Actor = { user: { uuid: 'u-1' } };
+    const actor: Actor = makeActor({ user: { uuid: 'u-1' } });
 
     it.each(cases)('records %s', async (expected, init) => {
         const stub = makeStubAuth(actor);
@@ -523,7 +532,7 @@ describe('createAuthProbe — cookie reading', () => {
 
 describe('createAuthProbe — actor attachment + failure tracking', () => {
     it('attaches actor + token on a successful authenticate', async () => {
-        const actor: Actor = { user: { uuid: 'u-1' } };
+        const actor: Actor = makeActor({ user: { uuid: 'u-1' } });
         const stub = makeStubAuth(actor);
         const probe = createAuthProbe({ authService: stub.service });
         const { req } = await runProbe(

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

@@ -22,6 +22,7 @@ import type {
     AuthService,
     ReauthReason,
 } from '../../../services/auth/AuthService';
+import { assertResolvedActor } from '../../actor';
 import type { TokenSource } from '../types';
 
 // Ensure the `Request.actor` / `Request.token` augmentation is in scope
@@ -155,7 +156,7 @@ export const createAuthProbe = (opts: AuthProbeOptions): RequestHandler => {
             }
 
             if (result.actor) {
-                req.actor = result.actor;
+                req.actor = assertResolvedActor(result.actor);
                 req.token = token;
                 req.tokenSource = source;
             } else if (result.invalid) {

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

@@ -0,0 +1,50 @@
+/*
+ * 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 { RequestHandler } from 'express';
+import type { IConfig } from '../../../types';
+import {
+    assertActorHasCredits,
+    type CreditMeteringLike,
+} from '../../../services/metering/enforcement.js';
+import '../expressAugmentation';
+
+/**
+ * Reject an authenticated caller with nothing left of their budget, for routes
+ * that opt in with `requireCredits`.
+ *
+ * Anonymous callers pass: the signed-URL routes authorize on the URL rather
+ * than a session, and there is no account to charge or turn away. So do worker
+ * sessions, unless configured otherwise — see `creditEnforcementExempt`.
+ *
+ * The answer comes from a short-lived per-actor cache in the metering service,
+ * so this normally costs nothing beyond a map lookup. That is what makes it
+ * affordable on routes that are called hundreds of times a minute.
+ */
+export const requireCreditsGate = (
+    metering: CreditMeteringLike | undefined,
+    config: IConfig,
+): RequestHandler => {
+    return (req, _res, next) => {
+        assertActorHasCredits(metering, req.actor, config).then(
+            () => next(),
+            (err) => next(err),
+        );
+    };
+};

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

@@ -0,0 +1,176 @@
+/*
+ * 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 { afterAll, beforeAll, describe, expect, it } from 'vitest';
+import type { Actor } from '../../actor';
+import { PERIOD_ESCAPE } from '../../../services/metering/consts.js';
+import type { UsageByType } from '../../../services/metering/types';
+import { setupPuterTestEnv, type PuterTestEnv } from '../../../testUtil.js';
+
+/**
+ * Egress metering over real HTTP. The unit tests drive the middleware with a
+ * response double; only a listening server proves the byte counter survives the
+ * middleware stack it is installed under (compression included) and that the
+ * actor is resolvable by the time the response closes.
+ */
+describe('egress metering over HTTP', () => {
+    let env: PuterTestEnv;
+
+    beforeAll(async () => {
+        env = await setupPuterTestEnv();
+    }, 120_000);
+
+    afterAll(async () => {
+        await env?.shutdown();
+    });
+
+    const escape = (usageType: string) =>
+        usageType.replace(/\./g, PERIOD_ESCAPE);
+
+    const usageFor = async (actor: Actor): Promise<UsageByType> => {
+        await env.server.services.metering.flushBufferedUsages();
+        const { usage } =
+            await env.server.services.metering.getActorCurrentMonthUsageDetails(
+                actor,
+            );
+        return usage;
+    };
+
+    const actorFor = async (username: string): Promise<Actor> => {
+        const user = await env.server.stores.user.getByUsername(username);
+        return { user: user! } as Actor;
+    };
+
+    it('bills a file read to the reader, bytes and object-store request alike', async () => {
+        const { username, token } = env.users.user;
+        const actor = await actorFor(username);
+
+        const body = Buffer.from('x'.repeat(4096));
+        await env.server.services.fs.write(actor.user.id!, {
+            fileMetadata: {
+                path: `/${username}/Desktop/egress.txt`,
+                size: body.byteLength,
+                contentType: 'text/plain',
+            },
+            fileContent: body,
+        });
+
+        const before = await usageFor(actor);
+        const beforeEgress =
+            (before[escape('egress:bytes')] as { units?: number } | undefined)
+                ?.units ?? 0;
+
+        const readUrl = new URL('/fs/read', env.apiOrigin);
+        readUrl.searchParams.set('path', `/${username}/Desktop/egress.txt`);
+        const read = await fetch(readUrl, {
+            headers: { Authorization: `Bearer ${token}` },
+        });
+        expect(read.status).toBe(200);
+        expect(await read.text()).toHaveLength(body.byteLength);
+
+        const after = await usageFor(actor);
+        const egress = after[escape('egress:bytes')] as {
+            units: number;
+            cost: number;
+        };
+        // Compression may shrink the payload on the wire, so the floor is the
+        // headers rather than the file — what matters is that the read was
+        // counted at all, and that it cost something.
+        expect(egress.units).toBeGreaterThan(beforeEgress);
+        expect(egress.cost).toBeGreaterThan(0);
+
+        const reads = after[escape('storage:read:ops')] as {
+            units: number;
+        };
+        expect(reads.units).toBeGreaterThanOrEqual(1);
+    });
+
+    it('bills a signed-URL read to the account whose file it is', async () => {
+        const { username, token } = env.users.other;
+        const actor = await actorFor(username);
+
+        const body = Buffer.from('y'.repeat(4096));
+        const path = `/${username}/Desktop/signed-egress.txt`;
+        await env.server.services.fs.write(actor.user.id!, {
+            fileMetadata: {
+                path,
+                size: body.byteLength,
+                contentType: 'text/plain',
+            },
+            fileContent: body,
+        });
+
+        const signed = await fetch(new URL('/sign', env.apiOrigin), {
+            method: 'POST',
+            headers: {
+                Authorization: `Bearer ${token}`,
+                'Content-Type': 'application/json',
+            },
+            body: JSON.stringify({ items: [{ path, action: 'read' }] }),
+        });
+        expect(signed.status).toBe(200);
+        const readUrl = new URL(
+            (await signed.json()).signatures[0].read_url as string,
+        );
+
+        const before = await usageFor(actor);
+        const beforeEgress =
+            (before[escape('egress:bytes')] as { units?: number } | undefined)
+                ?.units ?? 0;
+
+        // A signature proves access to the file, not who is asking — so this
+        // fetch carries no credential, and the owner is the only account there
+        // is to bill.
+        const fetched = await fetch(
+            new URL(`${readUrl.pathname}${readUrl.search}`, env.apiOrigin),
+        );
+        expect(fetched.status).toBe(200);
+        expect(await fetched.text()).toHaveLength(body.byteLength);
+
+        const after = await usageFor(actor);
+        const egress = after[escape('egress:bytes')] as {
+            units: number;
+            cost: number;
+        };
+        expect(egress.units).toBeGreaterThan(beforeEgress);
+        expect(egress.cost).toBeGreaterThan(0);
+    });
+
+    it('leaves root-origin asset traffic out of the actor’s usage', async () => {
+        const { username, token } = env.users.user;
+        const actor = await actorFor(username);
+
+        const before = await usageFor(actor);
+        const beforeEgress =
+            (before[escape('egress:bytes')] as { units?: number } | undefined)
+                ?.units ?? 0;
+
+        const sdk = await fetch(new URL('/puter.js/v2', env.origin), {
+            headers: { Authorization: `Bearer ${token}` },
+        });
+        expect(sdk.status).toBe(200);
+        expect((await sdk.text()).length).toBeGreaterThan(1000);
+
+        const after = await usageFor(actor);
+        const afterEgress =
+            (after[escape('egress:bytes')] as { units?: number } | undefined)
+                ?.units ?? 0;
+        expect(afterEgress).toBe(beforeEgress);
+    });
+});

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

@@ -0,0 +1,254 @@
+/*
+ * 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 { EventEmitter } from 'node:events';
+import type { Request, Response } from 'express';
+import { describe, expect, it, vi } from 'vitest';
+import type { Actor } from '../../actor';
+import { SYSTEM_ACTOR } from '../../actor';
+import {
+    EGRESS_COSTS,
+    STORAGE_OP_COSTS,
+} from '../../../services/metering/costs';
+import type { UsageInput } from '../../../services/metering/types';
+import { createEgressMeteringMiddleware } from './egressMetering';
+
+const actor = (uuid = 'user-1'): Actor =>
+    ({ user: { uuid, username: 'u' } }) as Actor;
+
+/**
+ * Response double that behaves like the parts of `http.ServerResponse` the
+ * middleware touches: writes go somewhere, `close` is emitted once the response
+ * is over, and headers are readable at that point.
+ */
+const makeRes = (headers: Record<string, string> = {}) => {
+    const emitter = new EventEmitter();
+    const written: unknown[] = [];
+    const res = Object.assign(emitter, {
+        write: vi.fn((chunk: unknown) => {
+            written.push(chunk);
+            return true;
+        }),
+        end: vi.fn((chunk?: unknown) => {
+            if (chunk !== undefined && typeof chunk !== 'function')
+                written.push(chunk);
+            return res;
+        }),
+        getHeaders: () => headers,
+    }) as unknown as Response & { written: unknown[] };
+    return Object.assign(res, { written });
+};
+
+const run = (
+    reqPartial: Partial<Request> & Record<string, unknown> = {},
+    headers: Record<string, string> = {},
+) => {
+    const buffered: Array<{ actor: Actor; usages: UsageInput[] }> = [];
+    const bufferIncrementUsages = vi.fn((a: Actor, usages: UsageInput[]) => {
+        buffered.push({ actor: a, usages });
+    });
+    const middleware = createEgressMeteringMiddleware({
+        services: { metering: { bufferIncrementUsages } },
+    });
+
+    const req = {
+        subdomains: ['api'],
+        actor: actor(),
+        ...reqPartial,
+    } as unknown as Request;
+    const res = makeRes(headers);
+    const next = vi.fn();
+    middleware(req, res, next);
+
+    return { req, res, next, buffered, bufferIncrementUsages };
+};
+
+const finish = (res: Response) => res.emit('close');
+
+const usageOf = (usages: UsageInput[], usageType: string) =>
+    usages.find((u) => u.usageType === usageType);
+
+describe('createEgressMeteringMiddleware', () => {
+    it('passes the request straight through', () => {
+        const { next, res } = run();
+        expect(next).toHaveBeenCalledOnce();
+        // The write hooks must not swallow the payload.
+        res.write(Buffer.from('abc'));
+        res.end('de');
+        expect((res as Response & { written: unknown[] }).written).toEqual([
+            Buffer.from('abc'),
+            'de',
+        ]);
+    });
+
+    it('bills every byte written, plus the headers, at the response cost', () => {
+        const { res, buffered } = run({}, { 'content-type': 'text/plain' });
+
+        res.write(Buffer.alloc(1000));
+        res.end(Buffer.alloc(24));
+        finish(res);
+
+        expect(buffered).toHaveLength(1);
+        const egress = usageOf(buffered[0]!.usages, 'egress:bytes')!;
+        // Header estimate is small but non-zero, so the total is a floor.
+        expect(egress.usageAmount).toBeGreaterThan(1024);
+        expect(egress.usageAmount).toBeLessThan(1100);
+        expect(egress.costOverride).toBeCloseTo(
+            EGRESS_COSTS['egress:bytes'] * egress.usageAmount,
+            10,
+        );
+    });
+
+    it('counts string chunks by their encoded length, not their character count', () => {
+        const empty = run();
+        finish(empty.res);
+        const headerBytes = usageOf(
+            empty.buffered[0]!.usages,
+            'egress:bytes',
+        )!.usageAmount;
+
+        const { res, buffered } = run();
+        res.end('déjà');
+        finish(res);
+        const withBody = usageOf(
+            buffered[0]!.usages,
+            'egress:bytes',
+        )!.usageAmount;
+
+        expect(withBody - headerBytes).toBe(Buffer.byteLength('déjà'));
+    });
+
+    it('meters a response that died mid-stream for what it managed to send', () => {
+        const { res, buffered } = run();
+        res.write(Buffer.alloc(500));
+        // No end() — the connection dropped.
+        finish(res);
+
+        expect(
+            usageOf(buffered[0]!.usages, 'egress:bytes')!.usageAmount,
+        ).toBeGreaterThan(500);
+    });
+
+    it('bills the object-store requests made while serving the response', () => {
+        const { req, res, buffered } = run();
+        (req as Request).storageOps = { write: 3, read: 2, delete: 5 };
+        res.end('x');
+        finish(res);
+
+        const usages = buffered[0]!.usages;
+        expect(usageOf(usages, 'storage:write:ops')).toMatchObject({
+            usageAmount: 3,
+            costOverride: STORAGE_OP_COSTS['storage:write:ops'] * 3,
+        });
+        expect(usageOf(usages, 'storage:read:ops')).toMatchObject({
+            usageAmount: 2,
+        });
+        // Removals are counted but free.
+        expect(usageOf(usages, 'storage:delete:ops')).toMatchObject({
+            usageAmount: 5,
+            costOverride: 0,
+        });
+    });
+
+    it('bills `egressActor` ahead of the requesting actor', () => {
+        const owner = actor('owner-1');
+        const { res, buffered } = run({
+            subdomains: ['some-site'],
+            actor: undefined,
+            egressActor: owner,
+        });
+        res.end('hello');
+        finish(res);
+
+        expect(buffered[0]!.actor).toBe(owner);
+    });
+
+    it('meters a host it otherwise ignores once a billing target is named', () => {
+        const { res, buffered } = run({
+            subdomains: [],
+            egressActor: actor('owner-1'),
+        });
+        res.end('hello');
+        finish(res);
+
+        expect(buffered).toHaveLength(1);
+    });
+
+    it('leaves first-party asset traffic unmetered', () => {
+        for (const subdomains of [[], ['js'], ['docs']]) {
+            const { res, bufferIncrementUsages } = run({ subdomains });
+            res.end(Buffer.alloc(5_000_000));
+            finish(res);
+            expect(bufferIncrementUsages).not.toHaveBeenCalled();
+        }
+    });
+
+    it('meters the dav surface alongside the api', () => {
+        const { res, bufferIncrementUsages } = run({ subdomains: ['dav'] });
+        res.end('hello');
+        finish(res);
+        expect(bufferIncrementUsages).toHaveBeenCalledOnce();
+    });
+
+    it('skips requests with no actor and the system actor', () => {
+        for (const req of [
+            { actor: undefined },
+            { actor: { user: {} } as Actor },
+            { actor: SYSTEM_ACTOR },
+        ]) {
+            const { res, bufferIncrementUsages } = run(req);
+            res.end('hello');
+            finish(res);
+            expect(bufferIncrementUsages).not.toHaveBeenCalled();
+        }
+    });
+
+    it('records nothing when metering is not installed', () => {
+        const middleware = createEgressMeteringMiddleware({ services: {} });
+        const req = {
+            subdomains: ['api'],
+            actor: actor(),
+        } as unknown as Request;
+        const res = makeRes();
+        const next = vi.fn();
+
+        middleware(req, res, next);
+        res.end('hello');
+        expect(() => finish(res)).not.toThrow();
+        expect(next).toHaveBeenCalledOnce();
+    });
+
+    it('never lets a metering failure escape into the response path', () => {
+        const bufferIncrementUsages = vi.fn(() => {
+            throw new Error('metering down');
+        });
+        const middleware = createEgressMeteringMiddleware({
+            services: { metering: { bufferIncrementUsages } },
+        });
+        const req = {
+            subdomains: ['api'],
+            actor: actor(),
+        } as unknown as Request;
+        const res = makeRes();
+
+        middleware(req, res, vi.fn());
+        res.end('hello');
+        expect(() => finish(res)).not.toThrow();
+    });
+});

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

@@ -0,0 +1,227 @@
+/*
+ * 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 { Request, RequestHandler, Response } from 'express';
+import type { Actor } from '../../actor';
+import { isSystemActor } from '../../actor';
+import type { StorageOpClass } from '../../storageOps';
+import {
+    EGRESS_COSTS,
+    STORAGE_OP_COSTS,
+    STORAGE_OP_USAGE_TYPES,
+} from '../../../services/metering/costs.js';
+import type { UsageInput } from '../../../services/metering/types';
+import '../expressAugmentation';
+
+/**
+ * Subset of the metering service this middleware needs. Metering is optional
+ * from here — a deployment without it still serves traffic.
+ */
+interface MeteringLike {
+    bufferIncrementUsages?: (actor: Actor, usages: UsageInput[]) => void;
+}
+
+interface Layers {
+    services: { metering?: MeteringLike };
+}
+
+/**
+ * Hosts whose responses are a user's own data or an account's app traffic, and
+ * so are billed to that account. Everything else the origin serves — the
+ * desktop shell, the SDK, the homepage, static assets — is Puter's own cost of
+ * being reachable and is deliberately not charged to whoever happens to be
+ * signed in while loading it.
+ *
+ * Hosted sites are not in here: they arrive on a per-site subdomain and opt in
+ * by naming who to bill (`req.egressActor`).
+ */
+const METERED_SUBDOMAINS = new Set(['api', 'dav']);
+
+/**
+ * Rough size of the status line and headers, which never reach `res.write` and
+ * so can only be estimated. Sized as `NAME: value\r\n` per header plus the
+ * status line and the blank line that ends the block. Small next to any real
+ * payload, but responses that are almost all headers (a 204, a redirect) are
+ * common enough that ignoring it would under-count a chatty client by a wide
+ * margin.
+ */
+const estimateHeaderBytes = (res: Response): number => {
+    // "HTTP/1.1 200 OK\r\n" plus the "\r\n" that terminates the block.
+    let bytes = 19;
+    let headers: ReturnType<Response['getHeaders']>;
+    try {
+        headers = res.getHeaders();
+    } catch {
+        return bytes;
+    }
+    for (const [name, value] of Object.entries(headers)) {
+        if (value === undefined) continue;
+        const rendered = Array.isArray(value)
+            ? value.join(', ')
+            : String(value);
+        bytes += name.length + rendered.length + 4;
+    }
+    return bytes;
+};
+
+const chunkBytes = (chunk: unknown, encoding: unknown): number => {
+    if (typeof chunk === 'string') {
+        return Buffer.byteLength(
+            chunk,
+            typeof encoding === 'string'
+                ? (encoding as BufferEncoding)
+                : 'utf8',
+        );
+    }
+    if (chunk instanceof Uint8Array || Buffer.isBuffer(chunk)) {
+        return chunk.byteLength;
+    }
+    return 0;
+};
+
+/** Who the response's bytes are billed to, or undefined if nobody. */
+const resolveEgressActor = (req: Request): Actor | undefined => {
+    const actor = req.egressActor ?? req.actor;
+    if (!actor?.user?.uuid) return undefined;
+    if (isSystemActor(actor)) return undefined;
+    return actor;
+};
+
+const isMeteredHost = (req: Request): boolean => {
+    // An explicit billing target is the opt-in for hosts that aren't metered
+    // by default, so honour it whatever the subdomain says.
+    if (req.egressActor) return true;
+    const subdomain = req.subdomains?.[req.subdomains.length - 1];
+    return !!subdomain && METERED_SUBDOMAINS.has(subdomain);
+};
+
+const storageOpUsages = (req: Request): UsageInput[] => {
+    const ops = req.storageOps;
+    if (!ops) return [];
+    const usages: UsageInput[] = [];
+    for (const [opClass, count] of Object.entries(ops)) {
+        if (!count || count <= 0) continue;
+        const usageType = STORAGE_OP_USAGE_TYPES[opClass as StorageOpClass];
+        if (!usageType) continue;
+        usages.push({
+            usageType,
+            usageAmount: count,
+            costOverride: STORAGE_OP_COSTS[usageType] * count,
+        });
+    }
+    return usages;
+};
+
+/**
+ * Meters what a request actually costs to serve: every byte written back to the
+ * client, plus the object-store requests made on its behalf.
+ *
+ * This is the only place egress is counted. Handlers that stream file content
+ * used to meter their own bytes, which measured the payload they handed to
+ * express rather than what left the process, missed every other response, and
+ * charged an increment per download. Counting here instead means one rule for
+ * all traffic, and bytes counted after compression has had its say.
+ *
+ * Install FIRST, ahead of the compression middleware: middleware that wraps
+ * `res.write` later ends up wrapping this one, so anything installed after
+ * compression sees the payload before it is compressed. The actor is read at
+ * the end of the response rather than here, by which time the auth probe has
+ * run.
+ *
+ * Never rejects, never delays the response: increments are handed to metering
+ * once the response is over, and metering buffers them rather than writing per
+ * request.
+ */
+export const createEgressMeteringMiddleware = (
+    layers: Layers,
+): RequestHandler => {
+    return (req, res, next) => {
+        const metering = layers.services.metering;
+        if (!metering?.bufferIncrementUsages) {
+            next();
+            return;
+        }
+
+        let bodyBytes = 0;
+
+        const write = res.write.bind(res);
+        const end = res.end.bind(res);
+
+        res.write = ((
+            chunk: unknown,
+            encoding?: unknown,
+            callback?: unknown,
+        ) => {
+            bodyBytes += chunkBytes(chunk, encoding);
+            return (write as (...args: unknown[]) => boolean)(
+                chunk,
+                encoding,
+                callback,
+            );
+        }) as Response['write'];
+
+        res.end = ((
+            chunk?: unknown,
+            encoding?: unknown,
+            callback?: unknown,
+        ) => {
+            // `end()` also takes a callback in the first or second slot.
+            if (typeof chunk !== 'function') {
+                bodyBytes += chunkBytes(chunk, encoding);
+            }
+            return (end as (...args: unknown[]) => Response)(
+                chunk,
+                encoding,
+                callback,
+            );
+        }) as Response['end'];
+
+        // 'close' rather than 'finish': it fires for a response that completed
+        // AND for one whose connection died mid-stream, and an aborted download
+        // still sent whatever it sent.
+        res.once('close', () => {
+            try {
+                if (!isMeteredHost(req)) return;
+                const actor = resolveEgressActor(req);
+                if (!actor) return;
+
+                const usages = storageOpUsages(req);
+                const bytes = bodyBytes + estimateHeaderBytes(res);
+                if (bytes > 0) {
+                    usages.push({
+                        usageType: 'egress:bytes',
+                        usageAmount: bytes,
+                        costOverride: EGRESS_COSTS['egress:bytes'] * bytes,
+                    });
+                }
+                if (usages.length === 0) return;
+
+                metering.bufferIncrementUsages!(actor, usages);
+            } catch (e) {
+                // Metering is never worth failing a request that already
+                // succeeded over.
+                console.warn(
+                    `[metering] egress metering failed: ${(e as Error).message}`,
+                );
+            }
+        });
+
+        next();
+    };
+};

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

@@ -19,6 +19,7 @@
 
 import type { Request, Response } from 'express';
 import { describe, expect, it } from 'vitest';
+import { makeActor, type Actor } from '../../actor';
 import { HttpError, isHttpError } from '../HttpError';
 import {
     DEFAULT_ADMIN_USERNAMES,
@@ -42,6 +43,23 @@ import {
 
 type NextArg = undefined | 'route' | HttpError | unknown;
 
+/**
+ * Rebuild an actor literal through `makeActor`, issuer-first, so the derived
+ * `effectiveApp` is present at every level of the token chain.
+ */
+const reviveActor = (actor: Actor): Actor =>
+    makeActor({
+        ...actor,
+        ...(actor.accessToken
+            ? {
+                  accessToken: {
+                      ...actor.accessToken,
+                      issuer: reviveActor(actor.accessToken.issuer),
+                  },
+              }
+            : {}),
+    });
+
 const runGate = (
     gate: (
         req: Request,
@@ -50,6 +68,12 @@ const runGate = (
     ) => unknown,
     req: Partial<Request>,
 ): NextArg => {
+    // Normalise the actor the way AuthService does before any gate sees one,
+    // so `effectiveApp` is derived here rather than spelled out on every
+    // literal below. Issuers too: a gate reading the chain reads the derived
+    // field, not `issuer.app`.
+    if (req.actor) req = { ...req, actor: reviveActor(req.actor) };
+
     let captured: NextArg = undefined;
     let called = false;
     gate(req as Request, {} as Response, (arg?: unknown) => {

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

@@ -18,7 +18,6 @@
  */
 
 import type { Request, RequestHandler } from 'express';
-import { effectiveActorApp } from '../../actor';
 import type { Actor } from '../../actor';
 import { HttpError } from '../HttpError';
 import { assertVerifiedEmail } from '../verifiedEmail';
@@ -238,7 +237,7 @@ export const DEFAULT_ADMIN_USERNAMES = ['admin', 'system'] as const;
  * pair, not a replacement for it.
  *
  * Also requires a _root token_ — an actor with no app anywhere in its token
- * chain (see `effectiveActorApp`) — so a third-party app an admin has
+ * chain (see `Actor.effectiveApp`) — so a third-party app an admin has
  * authorized can't reach admin endpoints on the admin's behalf. The one
  * exception is `appGated`: on a route that is also appId-gated
  * (`allowedAppIds`), a direct app-under-user actor is deferred to
@@ -278,7 +277,7 @@ export const adminOnlyGate = (
         // through an app. A direct app-under-user actor is deferred to
         // `allowedAppIdsGate` when the route is appId-gated; chain-only apps
         // are rejected even then, since that gate can't see them.
-        const chainApp = req.actor ? effectiveActorApp(req.actor) : null;
+        const chainApp = req.actor?.effectiveApp ?? null;
         if (chainApp && !(opts.appGated && req.actor?.app?.uid)) {
             next(
                 new HttpError(403, 'Only admins may request this resource', {

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

@@ -21,7 +21,6 @@ import type { RequestHandler } from 'express';
 import { contentType as contentTypeFromMime } from 'mime-types';
 import { posix as pathPosix } from 'node:path';
 import type { puterClients } from '../../../clients';
-import { FS_COSTS } from '../../../controllers/fs/costs';
 import type { puterServices } from '../../../services';
 import type { puterStores } from '../../../stores';
 import type { IConfig, LayerInstances } from '../../../types';
@@ -592,45 +591,20 @@ export const createPuterSiteMiddleware = (
         res.setHeader('Access-Control-Allow-Origin', '*');
         res.status(statusOverride ?? (range ? 206 : 200));
 
-        // Best-effort egress metering against the site owner. The request
-        // itself is unauthenticated (public site visitor), so we can't use
-        // req.actor — charge the account that hosts the file. Same cost
-        // key as FS read egress (`filesystem:egress:bytes`). Fires once
-        // the body stream ends so we only meter bytes actually delivered
-        // (not aborted mid-stream).
-        const metering = layers.services.metering as unknown as
-            | {
-                  batchIncrementUsages?: (
-                      actor: unknown,
-                      entries: unknown[],
-                  ) => void;
-              }
-            | undefined;
-        if (metering?.batchIncrementUsages && download.contentLength) {
-            const ownerActor = {
-                user: {
-                    uuid: owner.uuid,
-                    id: owner.id,
-                    username: owner.username,
-                    suspended: !!owner.suspended,
-                },
-            };
-            download.body.once('end', () => {
-                try {
-                    const bytes = download.contentLength!;
-                    metering.batchIncrementUsages!(ownerActor, [
-                        {
-                            usageType: 'filesystem:egress:bytes',
-                            usageAmount: bytes,
-                            costOverride:
-                                FS_COSTS['filesystem:egress:bytes'] * bytes,
-                        },
-                    ]);
-                } catch {
-                    // ignore — non-critical.
-                }
-            });
-        }
+        // Name who this response's bytes are billed to; the egress middleware
+        // does the metering. A visitor carrying a token pays for what they
+        // fetch, and the account hosting the site covers everyone else —
+        // hosting is unauthenticated by design, so most visitors are nobody in
+        // particular. Set unconditionally because hosting subdomains are not
+        // metered by default: this is what opts the response in.
+        req.egressActor = req.actor ?? {
+            user: {
+                uuid: owner.uuid,
+                id: owner.id,
+                username: owner.username,
+                suspended: !!owner.suspended,
+            },
+        };
 
         req.on('close', () => download.body.destroy());
         download.body.on('error', (err) => res.destroy(err));

src/backend/core/http/middleware/rateLimit.js

@@ -57,32 +57,57 @@ export const RATE_LIMIT_BACKENDS = ['memory', 'redis', 'kv'];
 const MEMORY_MAX_KEYS = 10_000;
 const MEMORY_MAX_RETAIN_MS = 60 * 60_000;
 
+/**
+ * Key → `{ ts, windowMs }`. The window is kept alongside the timestamps because
+ * the sweep has to know it: collecting on `MEMORY_MAX_RETAIN_MS` alone would
+ * drop the state of any limit whose window is longer than the retention floor,
+ * which silently shortens that limit to the floor. Day-scale windows are a real
+ * shape — a "few per day" grant, for one — and under `memory` they were being
+ * reset every hour. Retention is therefore whichever is longer, and the key cap
+ * below is what actually bounds memory.
+ */
 const memoryWindows = new Map();
+
+/**
+ * Drop memory buckets that can no longer affect a decision. Runs on a timer
+ * below; exported as a test seam because the timer isn't drivable from a test
+ * (it's created at module load, before any fake clock is installed).
+ */
+export function sweepMemoryWindows() {
+    const now = Date.now();
+    for (const [k, entry] of memoryWindows) {
+        const retainMs = Math.max(MEMORY_MAX_RETAIN_MS, entry.windowMs);
+        if (
+            entry.ts.length === 0 ||
+            entry.ts[entry.ts.length - 1] < now - retainMs
+        )
+            memoryWindows.delete(k);
+    }
+}
+
 {
-    const sweep = setInterval(() => {
-        const cutoff = Date.now() - MEMORY_MAX_RETAIN_MS;
-        for (const [k, ts] of memoryWindows) {
-            if (ts.length === 0 || ts[ts.length - 1] < cutoff)
-                memoryWindows.delete(k);
-        }
-    }, 60_000);
+    const sweep = setInterval(sweepMemoryWindows, 60_000);
     sweep.unref?.();
 }
 
 async function checkMemory(key, limit, windowMs) {
     const now = Date.now();
     const cutoff = now - windowMs;
-    let timestamps = memoryWindows.get(key);
-    if (!timestamps) {
+    let entry = memoryWindows.get(key);
+    if (!entry) {
         // Map preserves insertion order; FIFO-evict before adding so a
         // unique-key flood between sweep ticks can't blow up memory.
         if (memoryWindows.size >= MEMORY_MAX_KEYS) {
             const oldest = memoryWindows.keys().next().value;
             memoryWindows.delete(oldest);
         }
-        timestamps = [];
-        memoryWindows.set(key, timestamps);
+        entry = { ts: [], windowMs };
+        memoryWindows.set(key, entry);
+    } else {
+        // A scope's window can change across a deploy; the live value wins.
+        entry.windowMs = windowMs;
     }
+    const timestamps = entry.ts;
     while (timestamps.length > 0 && timestamps[0] < cutoff) timestamps.shift();
     if (timestamps.length >= limit) return false;
     timestamps.push(now);
@@ -185,37 +210,57 @@ async function acquireMemoryConcurrent(key, limit) {
             if (c <= 1) memoryConcurrentCounts.delete(key);
             else memoryConcurrentCounts.set(key, c - 1);
         },
+        // Nothing expires a memory slot but the process holding it, so there
+        // is no staleness to renew away.
+        renew: async () => {},
     };
 }
 
 async function acquireRedisConcurrent(redis, key, limit) {
     const redisKey = `concurrent:${key}`;
-    // Atomic INCR + EXPIRE. If the new count exceeds the limit we DECR
-    // ourselves back out; the brief over-count is invisible to other
-    // callers because INCR is atomic per-key. EXPIRE is a safety net for
-    // process death between acquire and release — slots eventually clear
-    // on their own so a crashed worker can't pin the bucket.
+    const member = `${Date.now()}-${crypto.randomUUID()}`;
+    // One sorted-set member per held slot, scored by acquire time — the same
+    // shape `checkRedis` uses for windows, and for the same reason: expiry has
+    // to be per-slot, not per-key.
+    //
+    // A counter with a key-wide TTL cannot express that. Whoever touches the
+    // key last decides when *every* slot on it expires, so a rejected acquire
+    // extends the life of the slots that rejected it — and a client that
+    // retries on rejection (a websocket reconnect loop is the pointed case)
+    // holds a leaked bucket open forever, locking its owner out of a resource
+    // nobody is actually using. Here the sweep below drops each slot on its own
+    // age, so a leak drains on schedule no matter how hard anyone retries.
+    const now = Date.now();
     const results = await redis
         .multi()
-        .incr(redisKey)
+        // Slots older than the orphan window belonged to a process that died
+        // before releasing; drop them before counting.
+        .zremrangebyscore(redisKey, 0, now - ORPHAN_SAFETY_TTL_MS)
+        .zadd(redisKey, now, member)
+        .zcard(redisKey)
+        // Key-level TTL is only garbage collection for a bucket that goes
+        // quiet — the per-member sweep above is what bounds a live one.
         .expire(redisKey, ORPHAN_SAFETY_TTL_SEC)
         .exec();
     const count = Number(
-        Array.isArray(results[0]) ? results[0][1] : results[0],
+        Array.isArray(results[2]) ? results[2][1] : results[2],
     );
     if (count > limit) {
-        await redis.decr(redisKey);
+        await redis.zrem(redisKey, member);
         return { ok: false };
     }
     return {
         ok: true,
         release: async () => {
-            // DECR can race below zero if the TTL fired between acquire
-            // and release (slot cleared, counter resets, we DECR to -1).
-            // Clamp on the next observation; it costs an extra read only
-            // on the rare TTL race.
-            const after = await redis.decr(redisKey);
-            if (after < 0) await redis.set(redisKey, 0);
+            await redis.zrem(redisKey, member);
+        },
+        renew: async () => {
+            // Re-score in place so a slot held longer than the orphan window
+            // isn't mistaken for one whose owner died. `ZADD XX` only touches
+            // a member that's still there, so renewing after release (or after
+            // a sweep) can't resurrect the slot.
+            await redis.zadd(redisKey, 'XX', Date.now(), member);
+            await redis.expire(redisKey, ORPHAN_SAFETY_TTL_SEC);
         },
     };
 }
@@ -246,6 +291,15 @@ async function acquireKvConcurrent(kv, key, limit) {
         release: async () => {
             await kv.del({ key: slotKey });
         },
+        renew: async () => {
+            // Push the row's own TTL out; a slot that outlives the orphan
+            // window is held, not abandoned.
+            await kv.set({
+                key: slotKey,
+                value: 1,
+                expireAt: Math.ceil((Date.now() + ORPHAN_SAFETY_TTL_MS) / 1000),
+            });
+        },
     };
 }
 
@@ -559,6 +613,75 @@ export async function checkRateLimit(key, limit, windowMs, backend) {
     }
 }
 
+/**
+ * Imperative concurrency acquire — the `acquire` twin to `checkRateLimit`, for
+ * long-lived things that aren't a request/response pair and so can't use
+ * `concurrencyGate`. The websocket handshake is the motivating case: the slot
+ * has to be held for the life of the connection, not the life of a response.
+ *
+ * Caller MUST invoke `release()` exactly once when the thing being counted ends
+ * (`ok: false` still returns a no-op `release`, so callers can release
+ * unconditionally). Fails open on backend error.
+ *
+ * `release()` returns a promise that settles once the slot is actually back —
+ * await it when the next observation has to see the freed slot. Fire-and-forget
+ * is fine for the usual case (an event handler on connection close), which is
+ * why it never rejects.
+ *
+ * A holder that can outlive `ORPHAN_SAFETY_TTL_MS` must call `renew()` on a
+ * timer, or the orphan sweep will reclaim its slot as abandoned and the cap
+ * stops counting it. Anything that finishes in seconds can ignore it.
+ */
+export async function acquireConcurrent(key, limit, backend) {
+    const bk = resolveBackend(backend);
+    try {
+        const result = await bk.acquire(key, limit);
+        if (!result.ok)
+            return {
+                ok: false,
+                release: async () => {},
+                renew: async () => {},
+            };
+        let released = false;
+        return {
+            ok: true,
+            release: async () => {
+                if (released) return;
+                released = true;
+                try {
+                    await result.release();
+                } catch (err) {
+                    console.error(
+                        '[concurrent] imperative release failed:',
+                        err,
+                    );
+                }
+            },
+            renew: async () => {
+                if (released) return;
+                try {
+                    await result.renew?.();
+                } catch (err) {
+                    console.error('[concurrent] imperative renew failed:', err);
+                }
+            },
+        };
+    } catch (err) {
+        console.error(
+            '[concurrent] imperative acquire failed, failing open:',
+            err,
+        );
+        return { ok: true, release: async () => {}, renew: async () => {} };
+    }
+}
+
+/**
+ * How long a held slot stays valid without a `renew()`. Exported so a
+ * long-lived holder can pick a renewal cadence from it rather than hardcoding
+ * one that drifts out of step.
+ */
+export const CONCURRENT_SLOT_TTL_MS = ORPHAN_SAFETY_TTL_MS;
+
 // -- Subscription-aware limit resolution -----------------------------
 
 /**

src/backend/core/http/middleware/rateLimit.test.js

@@ -32,6 +32,8 @@ import { EventEmitter } from 'node:events';
 import { isHttpError } from '../HttpError.js';
 import { setupTestServer } from '../../../testUtil.ts';
 import {
+    CONCURRENT_SLOT_TTL_MS,
+    acquireConcurrent,
     acquireDriverConcurrent,
     checkDriverRateLimit,
     checkRateLimit,
@@ -39,6 +41,7 @@ import {
     configureRateLimit,
     listConfiguredRateLimitBackends,
     rateLimitGate,
+    sweepMemoryWindows,
 } from './rateLimit.js';
 
 // The rate-limit module is configured once at boot in production. In tests
@@ -848,7 +851,7 @@ describe('concurrencyGate — redis backend', () => {
         await redis.flushall();
     });
 
-    it('uses INCR/EXPIRE to coordinate slots and releases on finish', async () => {
+    it('coordinates slots across callers and releases on finish', async () => {
         const opts = {
             limit: 1,
             key: 'ip',
@@ -879,6 +882,85 @@ describe('concurrencyGate — redis backend', () => {
     });
 });
 
+// ── concurrency: orphaned slots ─────────────────────────────────────
+//
+// A slot whose holder died without releasing has to age out on its own. The
+// pointed case is a caller that retries on rejection — a reconnecting socket —
+// where a per-key expiry gets refreshed by the very attempts it is rejecting
+// and the bucket never drains.
+
+describe('acquireConcurrent — orphan recovery (redis)', () => {
+    let redis;
+    beforeAll(() => {
+        redis = new RedisMock();
+        configureRateLimit({ default: 'redis', redis });
+    });
+    afterAll(async () => {
+        vi.useRealTimers();
+        await redis?.quit?.();
+        configureRateLimit();
+    });
+    beforeEach(async () => {
+        vi.useFakeTimers();
+        vi.setSystemTime(new Date('2026-01-01T00:00:00Z'));
+        await redis.flushall();
+    });
+
+    it('frees a leaked slot after the orphan window even while callers retry', async () => {
+        const key = 'orphan-retry';
+        const start = Date.now();
+        const step = 5 * 60_000;
+
+        // Holder takes the only slot and dies — never releases.
+        expect((await acquireConcurrent(key, 1)).ok).toBe(true);
+        expect((await acquireConcurrent(key, 1)).ok).toBe(false);
+
+        // Retry throughout the orphan window, the way a reconnecting client
+        // does. Each rejection must leave the leaked slot's own age untouched.
+        for (
+            let elapsed = step;
+            elapsed < CONCURRENT_SLOT_TTL_MS;
+            elapsed += step
+        ) {
+            vi.setSystemTime(new Date(start + elapsed));
+            expect((await acquireConcurrent(key, 1)).ok).toBe(false);
+        }
+
+        vi.setSystemTime(new Date(start + CONCURRENT_SLOT_TTL_MS + step));
+        expect((await acquireConcurrent(key, 1)).ok).toBe(true);
+    });
+
+    it('keeps a renewed slot held past the orphan window', async () => {
+        const key = 'orphan-renew';
+        const slot = await acquireConcurrent(key, 1);
+        expect(slot.ok).toBe(true);
+
+        // A live long-lived holder renews rather than letting the sweep
+        // mistake it for an abandoned one.
+        for (
+            let elapsed = 0;
+            elapsed < 2 * CONCURRENT_SLOT_TTL_MS;
+            elapsed += 10 * 60_000
+        ) {
+            vi.setSystemTime(new Date(Date.now() + 10 * 60_000));
+            await slot.renew();
+        }
+
+        expect((await acquireConcurrent(key, 1)).ok).toBe(false);
+        await slot.release();
+        expect((await acquireConcurrent(key, 1)).ok).toBe(true);
+    });
+
+    it('does not resurrect a slot renewed after release', async () => {
+        const key = 'orphan-renew-after-release';
+        const slot = await acquireConcurrent(key, 1);
+        await slot.release();
+        await slot.renew();
+
+        expect((await acquireConcurrent(key, 1)).ok).toBe(true);
+    });
+});
+
 // ── concurrency: subscription-based limits ──────────────────────────
 
 describe('concurrencyGate — bySubscription overrides', () => {
@@ -1184,6 +1266,31 @@ describe('driver helpers — backend failures', () => {
 
     const req = { ip: '7.7.7.7', headers: {}, socket: {} };
 
+    // Minimal redis whose concurrent-acquire always reports one held slot, so
+    // a test can substitute its own failing `zrem` and exercise release alone.
+    const stubConcurrentRedis = () => ({
+        multi: () => ({
+            zremrangebyscore: function () {
+                return this;
+            },
+            zadd: function () {
+                return this;
+            },
+            zcard: function () {
+                return this;
+            },
+            expire: function () {
+                return this;
+            },
+            exec: async () => [
+                [null, 0],
+                [null, 1],
+                [null, 1],
+                [null, 1],
+            ],
+        }),
+    });
+
     it('checkDriverRateLimit fails open when the backend throws', async () => {
         configureRateLimit({
             default: 'redis',
@@ -1225,20 +1332,11 @@ describe('driver helpers — backend failures', () => {
         configureRateLimit({
             default: 'redis',
             redis: {
-                multi: () => ({
-                    incr: function () {
-                        return this;
-                    },
-                    expire: function () {
-                        return this;
-                    },
-                    exec: async () => [[null, 1]],
-                }),
-                decr: async () => {
-                    if (failRelease) throw new Error('decr exploded');
-                    return 0;
+                ...stubConcurrentRedis(),
+                zrem: async () => {
+                    if (failRelease) throw new Error('zrem exploded');
+                    return 1;
                 },
-                set: async () => 'OK',
             },
         });
         const handle = await acquireDriverConcurrent(req, 'iface', 'm', {
@@ -1261,19 +1359,10 @@ describe('driver helpers — backend failures', () => {
         configureRateLimit({
             default: 'redis',
             redis: {
-                multi: () => ({
-                    incr: function () {
-                        return this;
-                    },
-                    expire: function () {
-                        return this;
-                    },
-                    exec: async () => [[null, 1]],
-                }),
-                decr: async () => {
-                    throw new Error('decr exploded');
+                ...stubConcurrentRedis(),
+                zrem: async () => {
+                    throw new Error('zrem exploded');
                 },
-                set: async () => 'OK',
             },
         });
         const res = new EventEmitter();
@@ -1295,25 +1384,26 @@ describe('driver helpers — backend failures', () => {
         spy.mockRestore();
     });
 
-    it('clamps a redis concurrent counter that went negative after a TTL race', async () => {
+    it('absorbs a release whose slot was already swept away', async () => {
         const redis = new RedisMock();
         await redis.flushall();
         configureRateLimit({ default: 'redis', redis });
 
-        const handle = await acquireDriverConcurrent(req, 'iface', 'clamp', {
-            limit: 2,
+        const handle = await acquireDriverConcurrent(req, 'iface', 'sweep', {
+            limit: 1,
         });
         expect(handle.ok).toBe(true);
 
-        // Simulate the orphan TTL firing between acquire and release.
-        const keys = await redis.keys('concurrent:*');
-        for (const k of keys) await redis.del(k);
+        // The orphan sweep collected the whole bucket before release ran.
+        for (const k of await redis.keys('concurrent:*')) await redis.del(k);
+        await expect(handle.release()).resolves.toBeUndefined();
 
-        await handle.release();
-        const remaining = await redis.keys('concurrent:*');
-        for (const k of remaining) {
-            expect(Number(await redis.get(k))).toBe(0);
-        }
+        // Releasing a slot that is already gone must not leave the bucket
+        // owing anything — the next caller gets a clean one.
+        const next = await acquireDriverConcurrent(req, 'iface', 'sweep', {
+            limit: 1,
+        });
+        expect(next.ok).toBe(true);
         await redis.quit?.();
     });
 });
@@ -1342,3 +1432,112 @@ describe('memory backend key cap', () => {
         expect(await checkRateLimit(victim, 1, 60_000, 'memory')).toBe(true);
     });
 });
+
+describe('memory backend sweep retention', () => {
+    beforeAll(() => configureRateLimit());
+    afterAll(() => {
+        vi.useRealTimers();
+        configureRateLimit();
+    });
+
+    // The sweep's retention floor is an hour. A window longer than that has
+    // to survive it, or the limit is silently shortened to the floor — a
+    // day-scale "few per day" grant would reset hourly.
+    it('keeps a bucket whose window outlives the retention floor', async () => {
+        vi.useFakeTimers();
+        const day = 24 * 60 * 60_000;
+        const key = `long-window-${Math.random()}`;
+
+        vi.setSystemTime(new Date('2026-01-01T00:00:00Z'));
+        expect(await checkRateLimit(key, 1, day, 'memory')).toBe(true);
+        expect(await checkRateLimit(key, 1, day, 'memory')).toBe(false);
+
+        // Two hours on: past the retention floor, far short of the window.
+        vi.setSystemTime(new Date('2026-01-01T02:00:00Z'));
+        sweepMemoryWindows();
+        expect(await checkRateLimit(key, 1, day, 'memory')).toBe(false);
+
+        // Past the window itself, the bucket is collectable again.
+        vi.setSystemTime(new Date('2026-01-02T01:00:00Z'));
+        sweepMemoryWindows();
+        expect(await checkRateLimit(key, 1, day, 'memory')).toBe(true);
+    });
+
+    it('still collects a short-window bucket at the retention floor', async () => {
+        vi.useFakeTimers();
+        const key = `short-window-${Math.random()}`;
+
+        vi.setSystemTime(new Date('2026-01-01T00:00:00Z'));
+        expect(await checkRateLimit(key, 1, 60_000, 'memory')).toBe(true);
+
+        vi.setSystemTime(new Date('2026-01-01T02:00:00Z'));
+        sweepMemoryWindows();
+        // Nothing asserts the delete directly; a fresh bucket is the
+        // observable consequence of having been collected.
+        expect(await checkRateLimit(key, 1, 60_000, 'memory')).toBe(true);
+    });
+});
+
+describe('acquireConcurrent (imperative)', () => {
+    beforeAll(() => configureRateLimit());
+    afterAll(() => configureRateLimit());
+
+    // The websocket handshake is the motivating case: the slot is held for
+    // the life of the connection, not the life of a response, so it cannot
+    // go through `concurrencyGate`.
+    it('admits up to the limit and rejects past it', async () => {
+        const key = `imperative-${Math.random()}`;
+        const a = await acquireConcurrent(key, 2, 'memory');
+        const b = await acquireConcurrent(key, 2, 'memory');
+        const c = await acquireConcurrent(key, 2, 'memory');
+
+        expect(a.ok).toBe(true);
+        expect(b.ok).toBe(true);
+        expect(c.ok).toBe(false);
+
+        await a.release();
+        expect((await acquireConcurrent(key, 2, 'memory')).ok).toBe(true);
+    });
+
+    it('returns a no-op release on rejection so callers can release blindly', async () => {
+        const key = `imperative-noop-${Math.random()}`;
+        const held = await acquireConcurrent(key, 1, 'memory');
+        const denied = await acquireConcurrent(key, 1, 'memory');
+
+        expect(denied.ok).toBe(false);
+        // Releasing a slot we never got must not free the one we did.
+        await denied.release();
+        expect((await acquireConcurrent(key, 1, 'memory')).ok).toBe(false);
+
+        await held.release();
+        expect((await acquireConcurrent(key, 1, 'memory')).ok).toBe(true);
+    });
+
+    it('releases at most once even if called repeatedly', async () => {
+        const key = `imperative-once-${Math.random()}`;
+        const a = await acquireConcurrent(key, 1, 'memory');
+        await a.release();
+        await a.release();
+        await a.release();
+
+        // A double release would have driven the counter negative and
+        // handed out more slots than the limit allows.
+        expect((await acquireConcurrent(key, 1, 'memory')).ok).toBe(true);
+        expect((await acquireConcurrent(key, 1, 'memory')).ok).toBe(false);
+    });
+
+    it('fails open when the backend throws', async () => {
+        const err = new Error('backend down');
+        const redis = {
+            multi: () => ({
+                incr: () => {
+                    throw err;
+                },
+            }),
+        };
+        configureRateLimit({ redis });
+        const result = await acquireConcurrent('any', 1, 'redis');
+        expect(result.ok).toBe(true);
+        configureRateLimit();
+    });
+});

src/backend/core/http/types.ts

@@ -26,12 +26,7 @@ import type { Actor } from '../actor';
  * it instead of accepting any token that authenticates.
  */
 export type TokenSource =
-    | 'body'
-    | 'header'
-    | 'x-api-key'
-    | 'cookie'
-    | 'query'
-    | 'handshake';
+    'body' | 'header' | 'x-api-key' | 'cookie' | 'query' | 'handshake';
 
 /**
  * Every route method PuterRouter exposes. Mirrors the express router surface
@@ -72,8 +67,8 @@ export type RoutePath = string | RegExp | Array<string | RegExp>;
  * middleware chain in this order:
  *
  *     subdomain → requireAuth (+ suspended) → emailConfirmed →
- *     requireUserActor → adminOnly → allowedAppIds →
- *     caller `middleware: []` → handler
+ *     requireUserActor → adminOnly → allowedAppIds → rateLimit →
+ *     requireCredits → concurrent → caller `middleware: []` → handler
  *
  * `requireUserActor`, `adminOnly`, and `allowedAppIds` all imply `requireAuth`;
  * the materializer dedupes so only one auth gate ends up in the chain.
@@ -316,6 +311,23 @@ export interface RouteOptions {
         backend?: 'memory' | 'redis' | 'kv';
     };
 
+    /**
+     * Reject an account with nothing left of its usage budget with 402
+     * `insufficient_funds`.
+     *
+     * For routes that spend metered resources on the caller's behalf — moving
+     * file content, making object-store requests. Not for the routes that show
+     * an account what it has or let it delete things: an account that has run
+     * out still needs to be able to see its files, clear space, and reach its
+     * billing pages, and blocking that leaves no way out other than paying.
+     *
+     * Anonymous callers and worker sessions pass; see `requireCreditsGate`.
+     * Rate limits remain the bound on request _count_ — this bounds spend, and
+     * lags the traffic that produced it by the metering buffer window, so it
+     * stops sustained usage rather than a burst.
+     */
+    requireCredits?: boolean;
+
     // Reserved — wire as the corresponding features/services land:
     // bodyFiles?: string[];      // multer-style multipart fields
     // responseTimeout?: number;

src/backend/core/index.ts

@@ -28,5 +28,7 @@ export {
     isAppActor,
     isAccessTokenActor,
     actorUid,
+    assertResolvedActor,
+    makeActor,
     userRelatedActor,
 } from './actor';

src/backend/core/storageOps.test.ts

@@ -0,0 +1,62 @@
+/*
+ * 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 { Request } from 'express';
+import { describe, expect, it } from 'vitest';
+import { runWithContext } from './context';
+import { recordStorageOps } from './storageOps';
+
+const withRequest = (fn: () => void): Request => {
+    const req = {} as Request;
+    runWithContext({ req }, fn);
+    return req;
+};
+
+describe('recordStorageOps', () => {
+    it('tallies each class on the request in scope', () => {
+        const req = withRequest(() => {
+            recordStorageOps('write');
+            recordStorageOps('write', 3);
+            recordStorageOps('read');
+            recordStorageOps('delete', 2);
+        });
+
+        expect(req.storageOps).toEqual({ write: 4, read: 1, delete: 2 });
+    });
+
+    it('ignores counts that are not a positive number', () => {
+        const req = withRequest(() => {
+            recordStorageOps('write', 0);
+            recordStorageOps('write', -1);
+            recordStorageOps('write', Number.NaN);
+        });
+
+        expect(req.storageOps).toBeUndefined();
+    });
+
+    it('does nothing outside a request', () => {
+        expect(() => recordStorageOps('write')).not.toThrow();
+    });
+
+    it('does nothing when the context carries no request', () => {
+        expect(() =>
+            runWithContext({}, () => recordStorageOps('write')),
+        ).not.toThrow();
+    });
+});

src/backend/core/storageOps.ts

@@ -0,0 +1,51 @@
+/*
+ * 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 { Context } from './context';
+import './http/expressAugmentation';
+
+/**
+ * Object-store request classes. They are priced apart because they cost
+ * different amounts per request: uploads, copies and multipart parts are the
+ * expensive class, fetches and metadata lookups an order of magnitude cheaper,
+ * and removals are not charged for at all.
+ */
+export type StorageOpClass = 'write' | 'read' | 'delete';
+
+export type StorageOpCounts = Partial<Record<StorageOpClass, number>>;
+
+/**
+ * Tally object-store requests against the request that caused them.
+ *
+ * The counts ride on the express request rather than going straight to
+ * metering, so that the whole cost of serving a request — its response bytes
+ * and the object-store calls behind them — settles as one write when the
+ * response ends. Work with no request in scope (boot, background sweeps)
+ * tallies nothing, which is why this never throws when called outside one.
+ */
+export const recordStorageOps = (
+    opClass: StorageOpClass,
+    count: number = 1,
+): void => {
+    if (!Number.isFinite(count) || count <= 0) return;
+    const req = Context.get('req');
+    if (!req) return;
+    const ops = (req.storageOps ??= {});
+    ops[opClass] = (ops[opClass] ?? 0) + count;
+};

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

@@ -48,6 +48,7 @@ import { ChatCompletionDriver } from './ChatCompletionDriver.js';
 import { AzureChatProvider } from './providers/azure/AzureChatProvider.js';
 import { FakeChatProvider } from './providers/FakeChatProvider.js';
 import { InfronProvider } from './providers/infron/InfronProvider.js';
+import { NeuralwattProvider } from './providers/neuralwatt/NeuralwattProvider.js';
 import { OpenAiChatProvider } from './providers/openai/OpenAiChatCompletionsProvider.js';
 import { OpenRouterProvider } from './providers/openrouter/OpenRouterProvider.js';
 import { TogetherAIProvider } from './providers/together/TogetherAIProvider.js';
@@ -72,6 +73,7 @@ const FULL_PROVIDER_CONFIG = {
         'together-ai': { apiKey: 'k' },
         openrouter: { apiKey: 'k', apiBaseUrl: 'https://openrouter.test' },
         infron: { apiKey: 'k' },
+        neuralwatt: { apiKey: 'k' },
         // Suppress auto-discovery of a developer's local Ollama.
         ollama: { enabled: false },
     },
@@ -120,6 +122,9 @@ beforeAll(async () => {
     vi.spyOn(InfronProvider.prototype, 'models').mockResolvedValue(
         aggregatorCatalog('infron-only-model') as never,
     );
+    vi.spyOn(NeuralwattProvider.prototype, 'models').mockResolvedValue(
+        aggregatorCatalog('neuralwatt-only-model') as never,
+    );
     fullDriver = await makeDriver(FULL_PROVIDER_CONFIG);
     fakeOnlyDriver = await makeDriver({
         providers: { ollama: { enabled: false } },
@@ -176,6 +181,7 @@ describe('ChatCompletionDriver provider registration', () => {
             'together-ai',
             'openrouter',
             'infron',
+            'neuralwatt',
             'fake-chat',
         ]) {
             expect(providers).toContain(expected);

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

@@ -0,0 +1,329 @@
+/**
+ * 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/](https://www.gnu.org/licenses/).
+ */
+
+/**
+ * Which provider actually serves a model that several providers advertise.
+ *
+ * Lives apart from ChatCompletionDriver.test.ts because `vi.mock` is
+ * file-scoped and hoisted — mocking the OpenAI SDK and axios here would
+ * otherwise leak into every test in that file. Both mocks sit at the real
+ * network egress points, so the driver's registration, model-map build and
+ * resolution all run for real.
+ */
+
+import {
+    afterAll,
+    beforeAll,
+    beforeEach,
+    describe,
+    expect,
+    it,
+    vi,
+    type MockInstance,
+} from 'vitest';
+
+import { HttpError } from '../../core/http/HttpError.js';
+import { PuterServer } from '../../server.js';
+import { setupTestServer } from '../../testUtil.js';
+import { kv } from '../../util/kvSingleton.js';
+import { withTestActor } from '../integrationTestUtil.js';
+import { ChatCompletionDriver } from './ChatCompletionDriver.js';
+import {
+    clearUnhealthyRoutes,
+    markRouteUnhealthy,
+} from './utils/providerHealth.js';
+
+// -- OpenAI SDK mock ------------------------------------------------
+// Gemini reaches Google through `new openai.OpenAI()` (default export) and
+// the gateway through the named one, so both must resolve to the same ctor.
+
+const { createMock } = vi.hoisted(() => ({ createMock: vi.fn() }));
+
+vi.mock('openai', () => {
+    const OpenAICtor = vi.fn().mockImplementation(function (
+        this: Record<string, unknown>,
+    ) {
+        this.chat = { completions: { create: createMock } };
+    });
+    return { OpenAI: OpenAICtor, default: { OpenAI: OpenAICtor } };
+});
+
+// -- axios mock (gateway model catalog) -----------------------------
+
+const { axiosRequestMock } = vi.hoisted(() => ({ axiosRequestMock: vi.fn() }));
+
+vi.mock('axios', () => ({
+    default: { request: axiosRequestMock },
+    request: axiosRequestMock,
+}));
+
+// -- Harness --------------------------------------------------------
+
+let server: PuterServer;
+let driver: ChatCompletionDriver;
+
+const INFRON_KV_KEY = 'infronChat:models';
+const OPENROUTER_KV_KEY = 'openrouterChat:models';
+
+// Google lists gemini-2.5-flash input at $0.30/MTok. The gateway quotes a
+// floor price across its upstream routes, so it undercuts — which is exactly
+// the condition that used to hand it the traffic.
+const GATEWAY_CATALOG = [
+    {
+        id: 'google/gemini-2.5-flash',
+        display_name: 'Google: Gemini 2.5 Flash',
+        category_type: 'LLM',
+        supported_endpoint_types: ['openai'],
+        context_length: 1_048_576,
+        max_output_tokens: 65_536,
+        min_prompt_price: 0.15,
+        min_completion_price: 1.0,
+    },
+    {
+        // Only the gateway carries this one — no first-party counterpart.
+        id: 'google/gemini-2.5-flash-image-preview',
+        display_name: 'Google: Gemini 2.5 Flash Image Preview',
+        category_type: 'LLM',
+        supported_endpoint_types: ['openai'],
+        context_length: 32_768,
+        max_output_tokens: 8_192,
+        min_prompt_price: 0.3,
+        min_completion_price: 2.5,
+    },
+    {
+        // A vendor we integrate with directly that isn't Google — the case
+        // resold duplicates used to be dropped for.
+        id: 'deepseek/deepseek-v4-pro',
+        display_name: 'DeepSeek: V4 Pro',
+        category_type: 'LLM',
+        supported_endpoint_types: ['openai'],
+        context_length: 1_000_000,
+        max_output_tokens: 65_536,
+        min_prompt_price: 0.1,
+        min_completion_price: 0.5,
+    },
+];
+
+// OpenRouter's catalog is shaped differently, and it carries the same model
+// under two upstream orgs — four routes total for deepseek-v4-pro once the
+// vendor and Infron are counted.
+const OPENROUTER_CATALOG = [
+    'deepseek/deepseek-v4-pro',
+    'deepseek-ai/deepseek-v4-pro',
+    'google/gemini-2.5-flash',
+].map((id) => ({
+    id,
+    name: `${id} (via OpenRouter)`,
+    pricing: { prompt: '0.0000001', completion: '0.0000005' },
+    context_length: 1_000_000,
+    top_provider: { max_completion_tokens: 65_536 },
+    created: 1_700_000_000,
+}));
+
+beforeAll(async () => {
+    server = await setupTestServer();
+    kv.del?.(INFRON_KV_KEY);
+    kv.del?.(OPENROUTER_KV_KEY);
+    axiosRequestMock.mockImplementation(({ url }: { url: string }) => ({
+        data: {
+            data: url.includes('openrouter')
+                ? OPENROUTER_CATALOG
+                : GATEWAY_CATALOG,
+        },
+    }));
+
+    // Built once, not per-test: `#buildModelMap` mutates the catalogs
+    // providers hand back (lowercasing ids, pushing `puterId` onto the
+    // shared `aliases` array), and GeminiChatProvider returns its
+    // module-level GEMINI_MODELS by reference.
+    driver = new ChatCompletionDriver(
+        {
+            providers: {
+                gemini: { apiKey: 'test-key' },
+                deepseek: { apiKey: 'test-key' },
+                infron: { apiKey: 'test-key' },
+                openrouter: { apiKey: 'test-key' },
+                ollama: { enabled: false },
+            },
+        } as never,
+        server.clients,
+        server.stores,
+        server.services,
+    );
+    driver.onServerStart();
+    // `onServerStart` doesn't await `#buildModelMap`, and the gateway
+    // catalogs resolve on a microtask — poll until both gateways land.
+    for (let i = 0; i < 200; i++) {
+        const ids = await driver.list();
+        if (
+            ids.some((id) => id.startsWith('infron:')) &&
+            ids.some((id) => id.startsWith('openrouter:'))
+        ) {
+            break;
+        }
+        await new Promise((r) => setTimeout(r, 5));
+    }
+});
+
+// Every failure in this file marks the route it hit. Without this the first
+// test would decide where the second one starts.
+beforeEach(() => clearUnhealthyRoutes());
+
+afterAll(async () => {
+    await server?.shutdown();
+    clearUnhealthyRoutes();
+});
+
+/**
+ * Route a request and report who was tried, in order. Forcing the upstream to
+ * reject is what makes the whole chain observable: the driver records every
+ * attempt on the thrown error, and `attempts[0]` is who it chose first.
+ */
+const attemptsFor = async (model: string) => {
+    createMock.mockRejectedValue(new Error('upstream down'));
+    let caught: HttpError | undefined;
+    try {
+        await withTestActor(() =>
+            driver.complete({
+                model,
+                messages: [{ role: 'user', content: 'hi' }],
+            }),
+        );
+    } catch (e) {
+        caught = e as HttpError;
+    }
+    expect(caught).toBeInstanceOf(HttpError);
+    return (
+        caught as unknown as {
+            fields: { attempts: { model: string; provider: string }[] };
+        }
+    ).fields.attempts;
+};
+
+describe('ChatCompletionDriver gemini routing', () => {
+    it('serves gemini models from Google, with the gateway only as fallback', async () => {
+        const attempts = await attemptsFor('gemini-2.5-flash');
+
+        expect(attempts[0]).toMatchObject({
+            provider: 'gemini',
+            model: 'gemini-2.5-flash',
+        });
+        expect(attempts[1]).toMatchObject({
+            provider: 'infron',
+            model: 'infron:google/gemini-2.5-flash',
+        });
+    });
+
+    it('routes the prefixed and puterId forms to Google too', async () => {
+        for (const alias of [
+            'google/gemini-2.5-flash',
+            'google:google/gemini-2.5-flash',
+        ]) {
+            const attempts = await attemptsFor(alias);
+            expect(attempts[0].provider).toBe('gemini');
+        }
+    });
+
+    it('still routes models only the gateway carries to the gateway', async () => {
+        const attempts = await attemptsFor(
+            'google/gemini-2.5-flash-image-preview',
+        );
+
+        expect(attempts[0]).toMatchObject({ provider: 'infron' });
+    });
+});
+
+describe('ChatCompletionDriver duplicate-model fallback', () => {
+    // deepseek-v4-pro is served directly by DeepSeek, by Infron, and twice by
+    // OpenRouter (two upstream orgs) — four routes in one bucket.
+    const SHARED = 'deepseek-v4-pro';
+
+    it('keeps a reseller duplicate of any vendor, not just Google', async () => {
+        const attempts = await attemptsFor(SHARED);
+
+        expect(attempts[0]).toMatchObject({ provider: 'deepseek' });
+        expect(attempts.map((a) => a.provider)).toContain('infron');
+    });
+
+    it('leaves openrouter below the other resellers in the chain', async () => {
+        const attempts = await attemptsFor(SHARED);
+        const providers = attempts.map((a) => a.provider);
+
+        expect(providers.indexOf('infron')).toBeLessThan(
+            providers.indexOf('openrouter'),
+        );
+    });
+
+    it('stops after three attempts even with a fourth route available', async () => {
+        const attempts = await attemptsFor(SHARED);
+        expect(attempts).toHaveLength(3);
+
+        // Proof the cap is what stopped the chain rather than the bucket
+        // running dry: take the three just tried out of contention and a
+        // fourth route is still there to be served.
+        const burned = attempts.map((a) => `${a.provider}:${a.model}`);
+        for (const a of attempts) markRouteUnhealthy(a.provider, a.model);
+
+        const next = await attemptsFor(SHARED);
+        expect(burned).not.toContain(`${next[0].provider}:${next[0].model}`);
+    });
+
+    it('never tries the same provider-and-model pair twice', async () => {
+        const attempts = await attemptsFor(SHARED);
+        const routes = attempts.map((a) => `${a.provider}:${a.model}`);
+
+        expect(new Set(routes).size).toBe(routes.length);
+    });
+});
+
+describe('ChatCompletionDriver unhealthy-route skipping', () => {
+    it('skips a route marked by an earlier failure and serves the next one', async () => {
+        markRouteUnhealthy('deepseek', 'deepseek-v4-pro');
+
+        const attempts = await attemptsFor('deepseek-v4-pro');
+
+        expect(attempts[0].provider).toBe('infron');
+        expect(attempts.map((a) => a.provider)).not.toContain('deepseek');
+    });
+
+    it('marks the routes a failing request burned through', async () => {
+        await attemptsFor('deepseek-v4-pro');
+
+        // The marks the first request left behind push the second one past
+        // everything that just failed.
+        const next = await attemptsFor('deepseek-v4-pro');
+        expect(next[0].provider).not.toBe('deepseek');
+    });
+
+    it('still serves a marked route when it is the only one left', async () => {
+        // gemini-2.5-flash-image-preview has a single route; marking it must
+        // degrade to trying it anyway rather than failing with no attempt.
+        markRouteUnhealthy(
+            'infron',
+            'infron:google/gemini-2.5-flash-image-preview',
+        );
+
+        const attempts = await attemptsFor(
+            'google/gemini-2.5-flash-image-preview',
+        );
+
+        expect(attempts[0]).toMatchObject({ provider: 'infron' });
+    });
+});

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

@@ -587,6 +587,64 @@ describe('ChatCompletionDriver.complete credit gate and max_tokens cap', () => {
         expect(completeSpy).not.toHaveBeenCalled();
     });
 
+    // A provider that can't report a model's output ceiling used to make the
+    // cap arithmetic go negative — `null - approxTokens` is negative, not NaN
+    // — so a funded account was told it had insufficient funds.
+    for (const [label, ceiling] of [
+        ['null', null],
+        ['zero', 0],
+        ['undefined', undefined],
+    ] as const) {
+        it(`serves models whose output ceiling is ${label}`, async () => {
+            vi.spyOn(
+                FakeChatProvider.prototype,
+                'models',
+            ).mockResolvedValueOnce([
+                {
+                    id: 'nocap',
+                    aliases: [],
+                    costs_currency: 'usd-cents',
+                    costs: { input_tokens: 1000, output_tokens: 2000 },
+                    max_tokens: ceiling,
+                },
+            ] as never);
+            const d = await makeDriver();
+
+            vi.spyOn(
+                server.services.metering,
+                'hasEnoughCredits',
+            ).mockResolvedValue(true);
+            vi.spyOn(
+                server.services.metering,
+                'getRemainingUsage',
+            ).mockResolvedValue(100_000);
+
+            const completeSpy = vi
+                .spyOn(FakeChatProvider.prototype, 'complete')
+                .mockResolvedValueOnce({
+                    message: {
+                        role: 'assistant',
+                        content: [{ type: 'text', text: 'ok' }],
+                    },
+                    usage: { input_tokens: 1, output_tokens: 1 },
+                    finish_reason: 'stop',
+                } as never);
+
+            await withTestActor(() =>
+                d.complete({
+                    model: 'nocap',
+                    messages: [{ role: 'user', content: 'hi' }],
+                }),
+            );
+
+            // Credits still bound the request; only the unknown model
+            // ceiling is ignored.
+            const passed = completeSpy.mock.calls[0]![0] as ICompleteArguments;
+            expect(passed.max_tokens!).toBeGreaterThan(0);
+            expect(passed.max_tokens!).toBeLessThanOrEqual(50);
+        });
+    }
+
     it('rejects subscriber-only models for the default free subscription', async () => {
         vi.spyOn(FakeChatProvider.prototype, 'models').mockResolvedValueOnce([
             {

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

@@ -41,6 +41,7 @@ import { InfronProvider } from './providers/infron/InfronProvider.js';
 import { MiniMaxProvider } from './providers/minimax/MiniMaxProvider.js';
 import { MistralAIProvider } from './providers/mistral/MistralAiProvider.js';
 import { MoonshotProvider } from './providers/moonshot/MoonshotProvider.js';
+import { NeuralwattProvider } from './providers/neuralwatt/NeuralwattProvider.js';
 import { OllamaChatProvider } from './providers/ollama/OllamaProvider.js';
 import { OpenAiChatProvider } from './providers/openai/OpenAiChatCompletionsProvider.js';
 import { OpenAiResponsesChatProvider } from './providers/openai/OpenAiChatResponsesProvider.js';
@@ -60,9 +61,18 @@ import {
     normalize_messages,
     normalize_single_message,
 } from './utils/Messages.js';
+import {
+    compareModelPreference,
+    isIdentityKey,
+    normalizeModelKey,
+} from './utils/modelRouting.js';
+import {
+    isRouteUnhealthy,
+    markRouteUnhealthy,
+} from './utils/providerHealth.js';
 import { AIChatStream } from './utils/Streaming.js';
 
-const MAX_FALLBACKS = 4; // includes first attempt
+const MAX_ATTEMPTS = 3; // the first attempt plus two fallbacks
 
 type ProviderAttempt = {
     model: string;
@@ -123,6 +133,25 @@ const isUpstream5xx = (a: ProviderAttempt) =>
         a.error,
     );
 
+/**
+ * Whether a failure indicts the route rather than the request.
+ *
+ * Outages, rate limits and bad credentials will hit the next caller too, so the
+ * route is worth marking. A 4xx the upstream returned on the request's own
+ * merits (malformed tools, oversized prompt) says nothing about the route and
+ * must not take it out of rotation for everyone else. Attempts with no status
+ * at all are transport failures — treat them as route problems.
+ */
+const isRouteLevelFailure = (a: ProviderAttempt) =>
+    a.status === undefined ||
+    isRateLimit(a) ||
+    isAuthFailure(a) ||
+    isUpstream5xx(a);
+
+// One bucket can hold the same model id under several providers *and* several
+// ids under one provider, so only the pair identifies an attempt.
+const routeId = (provider: string, modelId: string) => `${provider}:${modelId}`;
+
 /**
  * Map an exhausted fallback chain to a single user-facing HttpError.
  *
@@ -402,11 +431,19 @@ export class ChatCompletionDriver extends PuterDriver {
                     remainingCredits - approximateInputCost;
                 const maxAllowedOutputTokens =
                     maxAllowedOutputUcents / outputTokenCost;
+                // A provider may not know a model's output ceiling. Drop the
+                // term rather than let a missing value drive the cap: `null`
+                // coerces to 0, so the subtraction goes negative instead of
+                // NaN and the user is told they're out of credits.
+                const modelOutputCeiling =
+                    Number.isFinite(model.max_tokens) && model.max_tokens > 0
+                        ? model.max_tokens - approximateTokenCount
+                        : Number.POSITIVE_INFINITY;
                 const cap = Math.floor(
                     Math.min(
                         args.max_tokens ?? Number.POSITIVE_INFINITY,
                         maxAllowedOutputTokens,
-                        model.max_tokens - approximateTokenCount,
+                        modelOutputCeiling,
                     ),
                 );
                 // `cap` is the credit-bounded ceiling on output tokens. When
@@ -437,6 +474,20 @@ export class ChatCompletionDriver extends PuterDriver {
         const attempts: ProviderAttempt[] = [];
         let res: IChatCompleteResult | undefined;
 
+        // A failed route is remembered briefly so the next request skips it
+        // rather than paying its timeout again.
+        const recordFailure = (
+            modelId: string,
+            providerId: string,
+            err: unknown,
+        ) => {
+            const attempt = toAttempt(modelId, providerId, err);
+            attempts.push(attempt);
+            if (isRouteLevelFailure(attempt)) {
+                markRouteUnhealthy(providerId, modelId);
+            }
+        };
+
         try {
             res = await provider.complete({
                 ...args,
@@ -444,19 +495,17 @@ export class ChatCompletionDriver extends PuterDriver {
                 provider: model.provider,
             });
         } catch (e) {
-            attempts.push(toAttempt(model.id, model.provider!, e));
+            recordFailure(model.id, model.provider!, e);
 
-            // Fallback loop
-            const tried = [model.id];
-            const triedProviders = [model.provider!];
+            // Fallback loop — the bucket holds every provider that serves this
+            // model, ranked by `compareModelPreference`, so each miss walks one
+            // step down that order.
+            const bucketKey = model.id;
+            const tried = new Set([routeId(model.provider!, model.id)]);
             let lastError: Error | null = e as Error;
 
-            while (lastError && tried.length < MAX_FALLBACKS) {
-                const fallback = this.#findFallback(
-                    model.id,
-                    tried,
-                    triedProviders,
-                );
+            while (lastError && attempts.length < MAX_ATTEMPTS) {
+                const fallback = this.#findFallback(bucketKey, tried);
                 if (!fallback) break;
 
                 const fbProvider = this.#providers[fallback.provider!];
@@ -473,8 +522,7 @@ export class ChatCompletionDriver extends PuterDriver {
                     });
                 }
 
-                tried.push(fallback.id);
-                triedProviders.push(fallback.provider!);
+                tried.add(routeId(fallback.provider!, fallback.id));
 
                 try {
                     res = await fbProvider.complete({
@@ -486,9 +534,7 @@ export class ChatCompletionDriver extends PuterDriver {
                     lastError = null;
                 } catch (fbErr) {
                     lastError = fbErr as Error;
-                    attempts.push(
-                        toAttempt(fallback.id, fallback.provider!, fbErr),
-                    );
+                    recordFailure(fallback.id, fallback.provider!, fbErr);
                 }
             }
         }
@@ -999,127 +1045,97 @@ export class ChatCompletionDriver extends PuterDriver {
             );
         }
 
+        const neuralwatt = providers['neuralwatt'];
+        const neuralwattKey = readKey(neuralwatt);
+        if (neuralwattKey) {
+            this.#providers['neuralwatt'] = new NeuralwattProvider(
+                {
+                    apiKey: neuralwattKey,
+                    apiBaseUrl: neuralwatt?.apiBaseUrl as string | undefined,
+                },
+                metering,
+            );
+        }
+
         // Fake provider — always available for testing
         this.#providers['fake-chat'] = new FakeChatProvider();
     }
 
     // -- Model map ---------------------------------------------------
 
+    /**
+     * Group every provider's catalog into per-model buckets.
+     *
+     * A bucket is the set of routes to one model: the vendor we integrate with
+     * directly plus every reseller carrying it. They are deliberately _not_
+     * deduplicated — the duplicates are what the fallback loop walks when a
+     * route fails. `compareModelPreference` decides who serves first, so a
+     * reseller only takes traffic once the vendor has actually failed.
+     *
+     * Entries join a bucket by identity key (see `isIdentityKey`); display
+     * names remain addressable but never merge two providers' entries.
+     */
     async #buildModelMap() {
-        const AGGREGATORS = new Set(['together-ai', 'openrouter', 'infron']);
-
         for (const providerName in this.#providers) {
             const provider = this.#providers[providerName];
-            const isAggregator = AGGREGATORS.has(providerName);
 
             for (const model of await provider.models()) {
-                model.id = model.id.trim().toLowerCase();
-                if (!this.#modelIdMap[model.id]) {
-                    this.#modelIdMap[model.id] = [];
-                }
-                this.#modelIdMap[model.id].push({
-                    ...model,
-                    provider: providerName,
-                });
-
+                model.id = normalizeModelKey(model.id);
                 if (model.puterId) {
-                    if (model.aliases) {
-                        model.aliases.push(model.puterId);
-                    } else {
-                        model.aliases = [model.puterId];
-                    }
+                    model.aliases = model.aliases
+                        ? [...model.aliases, model.puterId]
+                        : [model.puterId];
                 }
 
-                if (isAggregator && model.aliases) {
-                    let skip = false;
-                    for (const rawAlias of model.aliases) {
-                        const alias = rawAlias.trim().toLowerCase();
-                        const existing = this.#modelIdMap[alias];
-                        if (
-                            existing &&
-                            existing !== this.#modelIdMap[model.id]
-                        ) {
-                            if (existing.some((m) => m.provider === 'gemini')) {
-                                // Gemini exception — let the aggregator
-                                // entry through.
-                                continue;
-                            }
-                            skip = true;
-                            break;
-                        }
-                    }
-                    if (skip) {
-                        // Remove the entry we just pushed; leave the bucket
-                        // intact for other providers.
-                        const bucket = this.#modelIdMap[model.id];
-                        bucket.pop();
-                        if (bucket.length === 0) {
-                            delete this.#modelIdMap[model.id];
-                        }
-                        continue;
-                    }
+                // Catalogs derive an alias by stripping the vendor org off the
+                // id, which yields '' for ids that carry no org. Drop those —
+                // an empty key would pool unrelated models together.
+                const keys = [model.id, ...(model.aliases ?? [])]
+                    .map(normalizeModelKey)
+                    .filter((key) => key.length > 0);
+
+                const bucket =
+                    keys
+                        .filter(isIdentityKey)
+                        .map((key) => this.#modelIdMap[key])
+                        .find(Boolean) ?? [];
+                bucket.push({ ...model, provider: providerName });
+
+                // First registration owns a key: a name already claimed by
+                // another model keeps pointing where it did.
+                for (const key of keys) {
+                    this.#modelIdMap[key] ??= bucket;
                 }
 
-                if (model.aliases) {
-                    for (let alias of model.aliases) {
-                        alias = alias.trim().toLowerCase();
-                        if (!this.#modelIdMap[alias]) {
-                            this.#modelIdMap[alias] =
-                                this.#modelIdMap[model.id];
-                        } else if (
-                            this.#modelIdMap[alias] !==
-                            this.#modelIdMap[model.id]
-                        ) {
-                            this.#modelIdMap[alias].push({
-                                ...model,
-                                provider: providerName,
-                            });
-                            this.#modelIdMap[model.id] =
-                                this.#modelIdMap[alias];
-                        }
-                    }
-                }
-
-                // Sort: together-ai always last; then cheapest input-cost
-                // first; ties break by shorter id (usually the official
-                // name over a long aggregator-qualified one).
-                this.#modelIdMap[model.id].sort((a, b) => {
-                    const aAgg = a.provider === 'together-ai';
-                    const bAgg = b.provider === 'together-ai';
-                    if (aAgg !== bAgg) return aAgg ? 1 : -1;
-                    const aCost = a.costs[
-                        (a.input_cost_key as string) || 'input_tokens'
-                    ] as number;
-                    const bCost = b.costs[
-                        (b.input_cost_key as string) || 'input_tokens'
-                    ] as number;
-                    if (aCost === bCost) return a.id.length - b.id.length;
-                    return aCost - bCost;
-                });
+                bucket.sort(compareModelPreference);
             }
         }
     }
 
     #resolveModel(modelId: string, provider?: string): IChatModel | null {
-        const models = this.#modelIdMap[modelId?.trim().toLowerCase()];
+        const models = this.#modelIdMap[normalizeModelKey(modelId ?? '')];
         if (!models || models.length === 0) return null;
-        if (!provider) return models[0];
-        return models.find((m) => m.provider === provider) ?? models[0];
+        // An explicitly requested provider is honoured even if its route is
+        // marked — the caller asked for that one, not for the cheapest hop.
+        if (provider) {
+            const pinned = models.find((m) => m.provider === provider);
+            if (pinned) return pinned;
+        }
+        return this.#preferHealthy(models) ?? models[0];
     }
 
-    #findFallback(
-        modelId: string,
-        tried: string[],
-        triedProviders: string[],
-    ): IChatModel | null {
+    #findFallback(modelId: string, tried: Set<string>): IChatModel | null {
         const models = this.#modelIdMap[modelId];
         if (!models) return null;
-        return (
-            models.find(
-                (m) =>
-                    !tried.includes(m.id) ||
-                    !triedProviders.includes(m.provider!),
-            ) ?? null
+        const untried = models.filter(
+            (m) => !tried.has(routeId(m.provider!, m.id)),
         );
+        // Degrade to a marked route rather than to no route at all: the marks
+        // are a hint about recent failures, not a quota.
+        return this.#preferHealthy(untried) ?? untried[0] ?? null;
+    }
+
+    #preferHealthy(models: IChatModel[]): IChatModel | undefined {
+        return models.find((m) => !isRouteUnhealthy(m.provider!, m.id));
     }
 }

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

@@ -314,28 +314,4 @@ export const CLAUDE_MODELS: IChatModel[] = [
         context: 200000,
         max_tokens: 64000,
     },
-    {
-        puterId: 'anthropic:anthropic/claude-opus-4-1',
-        id: 'claude-opus-4-1-20250805',
-        modalities: { input: ['text', 'image', 'pdf'], output: ['text'] },
-        open_weights: false,
-        tool_call: true,
-        knowledge: '2025-03-31',
-        release_date: '2025-08-05',
-        aliases: ['claude-opus-4-1', 'anthropic/claude-opus-4-1'],
-        name: 'Claude Opus 4.1',
-        costs_currency: 'usd-cents',
-        input_cost_key: 'input_tokens',
-        output_cost_key: 'output_tokens',
-        costs: {
-            tokens: 1_000_000,
-            input_tokens: 1500,
-            ephemeral_5m_input_tokens: 1500 * 1.25,
-            ephemeral_1h_input_tokens: 1500 * 2,
-            cache_read_input_tokens: 1500 * 0.1,
-            output_tokens: 7500,
-        },
-        context: 200000,
-        max_tokens: 32000,
-    },
 ];

src/backend/drivers/ai-chat/providers/neuralwatt/NeuralwattProvider.test.ts

@@ -0,0 +1,699 @@
+/*
+ * 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/>.
+ */
+
+/**
+ * Offline unit tests for NeuralwattProvider.
+ *
+ * Boots a real PuterServer and constructs NeuralwattProvider against the
+ * live MeteringService. The OpenAI SDK and axios (catalog + quota) are
+ * mocked at their module boundaries.
+ */
+
+import { Writable } from 'node:stream';
+import {
+    afterAll,
+    afterEach,
+    beforeAll,
+    beforeEach,
+    describe,
+    expect,
+    it,
+    vi,
+    type MockInstance,
+} from 'vitest';
+
+import type { MeteringService } from '../../../../services/metering/MeteringService.js';
+import { PuterServer } from '../../../../server.js';
+import { setupTestServer } from '../../../../testUtil.js';
+import { kv } from '../../../../util/kvSingleton.js';
+import { withTestActor } from '../../../integrationTestUtil.js';
+import { AIChatStream } from '../../utils/Streaming.js';
+import {
+    mapNeuralwattApiModel,
+    NEURALWATT_DEFAULT_MODEL,
+    stripNeuralwattPrefix,
+} from './models.js';
+import { NeuralwattProvider } from './NeuralwattProvider.js';
+
+// ── OpenAI SDK mock ─────────────────────────────────────────────────
+
+const { createMock, openAICtor } = vi.hoisted(() => ({
+    createMock: vi.fn(),
+    openAICtor: vi.fn(),
+}));
+
+vi.mock('openai', () => {
+    const OpenAICtor = vi.fn().mockImplementation(function (
+        this: Record<string, unknown>,
+        opts: unknown,
+    ) {
+        openAICtor(opts);
+        this.chat = { completions: { create: createMock } };
+    });
+    return { OpenAI: OpenAICtor, default: { OpenAI: OpenAICtor } };
+});
+
+// ── axios mock (models + quota) ─────────────────────────────────────
+
+const { axiosRequestMock } = vi.hoisted(() => ({
+    axiosRequestMock: vi.fn(),
+}));
+
+vi.mock('axios', () => ({
+    default: { request: axiosRequestMock },
+    request: axiosRequestMock,
+}));
+
+// ── Test harness ────────────────────────────────────────────────────
+
+let server: PuterServer;
+let recordSpy: MockInstance<MeteringService['utilRecordUsageObject']>;
+
+const KV_MODELS_KEY = 'neuralwattChat:models';
+const KV_QUOTA_KEY = 'neuralwattChat:quota';
+
+const SAMPLE_API_MODELS = [
+    {
+        id: 'deepseek-v4-flash',
+        created: 1_700_000_000,
+        max_model_len: 1_000_000,
+        metadata: {
+            display_name: 'DeepSeek V4 Flash',
+            description: 'Fast tool-calling model',
+            pricing: {
+                input_per_million: 0.14,
+                output_per_million: 0.28,
+                cached_input_per_million: 0.014,
+                pricing_tbd: false,
+            },
+            capabilities: {
+                tools: true,
+                vision: false,
+                streaming: true,
+            },
+            limits: {
+                max_context_length: 1_000_000,
+                max_output_tokens: 384_000,
+            },
+        },
+    },
+    {
+        id: 'zai-org/GLM-5.1-FP8',
+        metadata: {
+            display_name: 'GLM 5.1',
+            pricing: {
+                input_per_million: 0.35,
+                output_per_million: 1.38,
+                cached_input_per_million: 0.035,
+                pricing_tbd: false,
+            },
+            capabilities: {
+                tools: true,
+                vision: false,
+                reasoning: true,
+            },
+            limits: {
+                max_context_length: 202_752,
+                max_output_tokens: 16_384,
+            },
+        },
+    },
+    {
+        id: 'coming-soon-model',
+        metadata: {
+            display_name: 'Coming Soon',
+            pricing: {
+                input_per_million: 0,
+                output_per_million: 0,
+                pricing_tbd: true,
+            },
+            capabilities: { tools: true },
+            limits: { max_context_length: 8_000, max_output_tokens: 1_000 },
+        },
+    },
+    {
+        id: 'deprecated-model',
+        metadata: {
+            display_name: 'Old',
+            deprecated: true,
+            pricing: {
+                input_per_million: 1,
+                output_per_million: 2,
+                pricing_tbd: false,
+            },
+            capabilities: { tools: true },
+            limits: { max_context_length: 8_000, max_output_tokens: 1_000 },
+        },
+    },
+];
+
+const mockCatalogAndQuota = (
+    accountingMethod: 'energy' | 'token' = 'energy',
+) => {
+    axiosRequestMock.mockImplementation(async (opts: { url?: string }) => {
+        if (opts.url?.endsWith('/quota')) {
+            return {
+                data: {
+                    balance: { accounting_method: accountingMethod },
+                },
+            };
+        }
+        return { data: { data: SAMPLE_API_MODELS } };
+    });
+};
+
+beforeAll(async () => {
+    server = await setupTestServer();
+});
+
+afterAll(async () => {
+    await server?.shutdown();
+});
+
+const makeProvider = (config?: { apiBaseUrl?: string }) => {
+    const provider = new NeuralwattProvider(
+        { apiKey: 'test-key', ...config },
+        server.services.metering,
+    );
+    return { provider };
+};
+
+const asAsyncIterable = <T>(items: T[]): AsyncIterable<T> => ({
+    async *[Symbol.asyncIterator]() {
+        for (const item of items) {
+            yield item;
+        }
+    },
+});
+
+const makeCapturingChatStream = () => {
+    const chunks: string[] = [];
+    const sink = new Writable({
+        write(chunk, _enc, cb) {
+            chunks.push(chunk.toString('utf8'));
+            cb();
+        },
+    });
+    const chatStream = new AIChatStream({ stream: sink });
+    return {
+        chatStream,
+        events: () =>
+            chunks
+                .join('')
+                .split('\n')
+                .filter(Boolean)
+                .map((line) => JSON.parse(line)),
+    };
+};
+
+beforeEach(() => {
+    createMock.mockReset();
+    openAICtor.mockReset();
+    axiosRequestMock.mockReset();
+    mockCatalogAndQuota('energy');
+    kv.del(KV_MODELS_KEY);
+    kv.del(KV_QUOTA_KEY);
+    recordSpy = vi.spyOn(server.services.metering, 'utilRecordUsageObject');
+});
+
+afterEach(() => {
+    vi.restoreAllMocks();
+    kv.del(KV_MODELS_KEY);
+    kv.del(KV_QUOTA_KEY);
+});
+
+// ── Mapping helpers ─────────────────────────────────────────────────
+
+describe('Neuralwatt model mapping helpers', () => {
+    it('strips the neuralwatt: prefix for upstream ids', () => {
+        expect(stripNeuralwattPrefix('neuralwatt:deepseek-v4-flash')).toBe(
+            'deepseek-v4-flash',
+        );
+        expect(stripNeuralwattPrefix('deepseek-v4-flash')).toBe(
+            'deepseek-v4-flash',
+        );
+    });
+
+    it('maps catalog pricing into usd-cents cost keys and skips pricing_tbd', () => {
+        const mapped = mapNeuralwattApiModel(SAMPLE_API_MODELS[0]!);
+        expect(mapped).toMatchObject({
+            id: 'neuralwatt:deepseek-v4-flash',
+            costs_currency: 'usd-cents',
+            input_cost_key: 'prompt_tokens',
+            output_cost_key: 'completion_tokens',
+            costs: {
+                tokens: 1_000_000,
+                prompt_tokens: 0.14 * 100,
+                completion_tokens: 0.28 * 100,
+                cached_tokens: 0.014 * 100,
+            },
+            tool_call: true,
+            max_tokens: 384_000,
+            modalities: { input: ['text'], output: ['text'] },
+        });
+        expect(mapNeuralwattApiModel(SAMPLE_API_MODELS[2]!)).toBeNull();
+    });
+
+    it('marks vision models from capabilities.vision', () => {
+        const mapped = mapNeuralwattApiModel({
+            id: 'gemma-4-31b',
+            metadata: {
+                display_name: 'Gemma 4 31B',
+                pricing: {
+                    input_per_million: 0.1,
+                    output_per_million: 0.2,
+                    pricing_tbd: false,
+                },
+                capabilities: { tools: true, vision: true },
+                limits: {
+                    max_context_length: 256_000,
+                    max_output_tokens: 16_384,
+                    max_images: 8,
+                },
+            },
+        });
+        expect(mapped).toMatchObject({
+            id: 'neuralwatt:gemma-4-31b',
+            modalities: { input: ['text', 'image'], output: ['text'] },
+            max_images: 8,
+        });
+    });
+});
+
+// ── Construction ────────────────────────────────────────────────────
+
+describe('NeuralwattProvider construction', () => {
+    it('points the OpenAI SDK at the Neuralwatt base URL with the configured key', () => {
+        makeProvider();
+        expect(openAICtor).toHaveBeenCalledTimes(1);
+        expect(openAICtor).toHaveBeenCalledWith({
+            apiKey: 'test-key',
+            baseURL: 'https://api.neuralwatt.com/v1',
+        });
+    });
+
+    it('honours an apiBaseUrl override', () => {
+        makeProvider({ apiBaseUrl: 'https://custom.neuralwatt.example/v1' });
+        expect(openAICtor).toHaveBeenLastCalledWith({
+            apiKey: 'test-key',
+            baseURL: 'https://custom.neuralwatt.example/v1',
+        });
+    });
+});
+
+// ── Model catalog + quota ───────────────────────────────────────────
+
+describe('NeuralwattProvider model catalog', () => {
+    it('returns the neuralwatt-prefixed default model id', () => {
+        const { provider } = makeProvider();
+        expect(provider.getDefaultModel()).toBe(NEURALWATT_DEFAULT_MODEL);
+    });
+
+    it('sends the API key as a bearer token on the catalog fetch', async () => {
+        const { provider } = makeProvider();
+        await provider.models();
+        const modelsCall = axiosRequestMock.mock.calls.find(
+            ([args]) =>
+                typeof args?.url === 'string' && args.url.endsWith('/models'),
+        );
+        expect(modelsCall?.[0]).toMatchObject({
+            url: 'https://api.neuralwatt.com/v1/models',
+            headers: { Authorization: 'Bearer test-key' },
+        });
+    });
+
+    it('list() prefixes ids and skips deprecated / pricing_tbd entries', async () => {
+        const { provider } = makeProvider();
+        const ids = await provider.list();
+        expect(ids).toContain('neuralwatt:deepseek-v4-flash');
+        expect(ids).toContain('neuralwatt:zai-org/GLM-5.1-FP8');
+        expect(ids).toContain('GLM-5.1-FP8');
+        expect(ids).not.toContain('neuralwatt:coming-soon-model');
+        expect(ids).not.toContain('neuralwatt:deprecated-model');
+    });
+
+    it('caches the model list in kv after the first axios round-trip', async () => {
+        const { provider } = makeProvider();
+        await provider.models();
+        await provider.models();
+        const modelsCalls = axiosRequestMock.mock.calls.filter(
+            ([args]) =>
+                typeof args?.url === 'string' && args.url.endsWith('/models'),
+        );
+        expect(modelsCalls).toHaveLength(1);
+    });
+
+    it('caches accounting_method from /quota', async () => {
+        const { provider } = makeProvider();
+        await expect(provider.getAccountingMethod()).resolves.toBe('energy');
+        await expect(provider.getAccountingMethod()).resolves.toBe('energy');
+        const quotaCalls = axiosRequestMock.mock.calls.filter(
+            ([args]) =>
+                typeof args?.url === 'string' && args.url.endsWith('/quota'),
+        );
+        expect(quotaCalls).toHaveLength(1);
+    });
+});
+
+// ── Request shape ───────────────────────────────────────────────────
+
+describe('NeuralwattProvider.complete request shape', () => {
+    const baseCompletion = {
+        choices: [
+            {
+                message: { content: 'hi', role: 'assistant' },
+                finish_reason: 'stop',
+            },
+        ],
+        usage: { prompt_tokens: 1, completion_tokens: 1 },
+        cost: { request_cost_usd: 0 },
+        energy: {
+            energy_kwh: 0.000001,
+            energy_joules: 3.6,
+            measurement_available: true,
+        },
+    };
+
+    it('strips the neuralwatt: prefix from the wire model id', async () => {
+        const { provider } = makeProvider();
+        createMock.mockResolvedValueOnce(baseCompletion);
+
+        await withTestActor(() =>
+            provider.complete({
+                model: 'neuralwatt:deepseek-v4-flash',
+                messages: [{ role: 'user', content: 'hello' }],
+            }),
+        );
+
+        const [args] = createMock.mock.calls[0]!;
+        expect(args.model).toBe('deepseek-v4-flash');
+    });
+
+    it('only sets stream_options.include_usage when streaming', async () => {
+        const { provider } = makeProvider();
+
+        createMock.mockResolvedValueOnce(baseCompletion);
+        await withTestActor(() =>
+            provider.complete({
+                model: 'neuralwatt:deepseek-v4-flash',
+                messages: [{ role: 'user', content: 'hi' }],
+                stream: false,
+            }),
+        );
+        expect(createMock.mock.calls[0]![0].stream).toBe(false);
+        expect('stream_options' in createMock.mock.calls[0]![0]).toBe(false);
+
+        createMock.mockReturnValueOnce(asAsyncIterable([]));
+        await withTestActor(() =>
+            provider.complete({
+                model: 'neuralwatt:deepseek-v4-flash',
+                messages: [{ role: 'user', content: 'hi' }],
+                stream: true,
+            }),
+        );
+        expect(createMock.mock.calls[1]![0].stream_options).toEqual({
+            include_usage: true,
+        });
+    });
+
+    it('rejects image content on a text-only catalog model', async () => {
+        const { provider } = makeProvider();
+        await expect(
+            withTestActor(() =>
+                provider.complete({
+                    model: 'neuralwatt:deepseek-v4-flash',
+                    messages: [
+                        {
+                            role: 'user',
+                            content: [
+                                {
+                                    type: 'image_url',
+                                    image_url: {
+                                        url: 'https://example.com/cat.png',
+                                    },
+                                },
+                            ],
+                        },
+                    ],
+                }),
+            ),
+        ).rejects.toMatchObject({
+            statusCode: 400,
+            message: expect.stringContaining('does not support image input'),
+        });
+        expect(createMock).not.toHaveBeenCalled();
+    });
+
+    it('prefers a vision catalog model when the prompt has images and no model was named', async () => {
+        // Seed a vision model into the catalog payload.
+        axiosRequestMock.mockImplementation(async (opts: { url?: string }) => {
+            if (opts.url?.endsWith('/quota')) {
+                return {
+                    data: { balance: { accounting_method: 'energy' } },
+                };
+            }
+            return {
+                data: {
+                    data: [
+                        ...SAMPLE_API_MODELS,
+                        {
+                            id: 'gemma-4-31b',
+                            metadata: {
+                                display_name: 'Gemma 4 31B',
+                                pricing: {
+                                    input_per_million: 0.1,
+                                    output_per_million: 0.2,
+                                    pricing_tbd: false,
+                                },
+                                capabilities: {
+                                    tools: true,
+                                    vision: true,
+                                },
+                                limits: {
+                                    max_context_length: 256_000,
+                                    max_output_tokens: 16_384,
+                                },
+                            },
+                        },
+                    ],
+                },
+            };
+        });
+
+        const { provider } = makeProvider();
+        createMock.mockResolvedValueOnce({
+            choices: [
+                {
+                    message: { content: 'a cat', role: 'assistant' },
+                    finish_reason: 'stop',
+                },
+            ],
+            usage: { prompt_tokens: 1, completion_tokens: 1 },
+            cost: { request_cost_usd: 0 },
+        });
+
+        await withTestActor(() =>
+            provider.complete({
+                model: '',
+                messages: [
+                    {
+                        role: 'user',
+                        content: [
+                            {
+                                type: 'image_url',
+                                image_url: {
+                                    url: 'data:image/png;base64,abc',
+                                },
+                            },
+                        ],
+                    },
+                ],
+            }),
+        );
+
+        expect(createMock.mock.calls[0]![0].model).toBe('gemma-4-31b');
+    });
+});
+
+// ── Non-stream metering ─────────────────────────────────────────────
+
+describe('NeuralwattProvider.complete non-stream output', () => {
+    it('bills from cost.request_cost_usd and records energy units at zero cost', async () => {
+        const { provider } = makeProvider();
+        createMock.mockResolvedValueOnce({
+            choices: [
+                {
+                    message: { content: 'hi there', role: 'assistant' },
+                    finish_reason: 'stop',
+                },
+            ],
+            usage: {
+                prompt_tokens: 100,
+                completion_tokens: 50,
+                prompt_tokens_details: { cached_tokens: 10 },
+            },
+            cost: { request_cost_usd: 0.0001 },
+            energy: {
+                energy_kwh: 0.00000145,
+                energy_joules: 5.23,
+                measurement_available: true,
+            },
+        });
+
+        const result = (await withTestActor(() =>
+            provider.complete({
+                model: 'neuralwatt:deepseek-v4-flash',
+                messages: [{ role: 'user', content: 'hi' }],
+            }),
+        )) as { usage: Record<string, number | string> };
+
+        expect(recordSpy).toHaveBeenCalledTimes(1);
+        const [usage, , prefix, overrides] = recordSpy.mock.calls[0]!;
+        expect(prefix).toBe('neuralwatt:deepseek-v4-flash');
+        expect(usage).toMatchObject({
+            prompt_tokens: 100,
+            completion_tokens: 50,
+            cached_tokens: 10,
+            energy_kwh: 0.00000145,
+            energy_joules: 5.23,
+            billedUsage: 1,
+        });
+        expect(overrides).toMatchObject({
+            prompt_tokens: 0,
+            completion_tokens: 0,
+            cached_tokens: 0,
+            energy_kwh: 0,
+            energy_joules: 0,
+            billedUsage: 0.0001 * 100_000_000,
+        });
+        expect(result.usage.usd_cents).toBe(0.0001 * 100);
+        expect(result.usage.accounting_method).toBe('energy');
+    });
+
+    it('falls back to catalog token pricing when request_cost_usd is absent', async () => {
+        const { provider } = makeProvider();
+        createMock.mockResolvedValueOnce({
+            choices: [
+                {
+                    message: { content: 'ok', role: 'assistant' },
+                    finish_reason: 'stop',
+                },
+            ],
+            usage: {
+                prompt_tokens: 100,
+                completion_tokens: 50,
+                prompt_tokens_details: { cached_tokens: 10 },
+            },
+        });
+
+        await withTestActor(() =>
+            provider.complete({
+                model: 'neuralwatt:deepseek-v4-flash',
+                messages: [{ role: 'user', content: 'hi' }],
+            }),
+        );
+
+        // usdPerMToken: $0.14/M → 14 cents per MTok unit in costs map
+        const promptRate = 0.14 * 100;
+        const completionRate = 0.28 * 100;
+        const cachedRate = 0.014 * 100;
+        const [usage, , , overrides] = recordSpy.mock.calls[0]!;
+        expect(usage).toMatchObject({
+            prompt_tokens: 100,
+            completion_tokens: 50,
+            cached_tokens: 10,
+        });
+        expect(overrides).toMatchObject({
+            prompt_tokens: 100 * promptRate,
+            completion_tokens: 50 * completionRate,
+            cached_tokens: 10 * cachedRate,
+        });
+    });
+});
+
+// ── Streaming ───────────────────────────────────────────────────────
+
+describe('NeuralwattProvider.complete streaming', () => {
+    it('streams text deltas and meters final-chunk cost + energy', async () => {
+        const { provider } = makeProvider();
+        createMock.mockReturnValueOnce(
+            asAsyncIterable([
+                {
+                    choices: [
+                        {
+                            delta: { content: 'Hello' },
+                            finish_reason: null,
+                        },
+                    ],
+                },
+                {
+                    choices: [
+                        {
+                            delta: { content: '!' },
+                            finish_reason: 'stop',
+                        },
+                    ],
+                    usage: {
+                        prompt_tokens: 10,
+                        completion_tokens: 2,
+                    },
+                    cost: { request_cost_usd: 0.00005 },
+                    energy: {
+                        energy_kwh: 0.000002,
+                        energy_joules: 7.2,
+                        measurement_available: true,
+                    },
+                },
+            ]),
+        );
+
+        const result = (await withTestActor(() =>
+            provider.complete({
+                model: 'neuralwatt:deepseek-v4-flash',
+                messages: [{ role: 'user', content: 'hi' }],
+                stream: true,
+            }),
+        )) as {
+            stream: true;
+            init_chat_stream: (p: {
+                chatStream: AIChatStream;
+            }) => Promise<void>;
+        };
+
+        const { chatStream, events } = makeCapturingChatStream();
+        await result.init_chat_stream({ chatStream });
+
+        const textEvents = events().filter((e) => e.type === 'text');
+        expect(textEvents.map((e) => e.text).join('')).toBe('Hello!');
+
+        expect(recordSpy).toHaveBeenCalledTimes(1);
+        const [usage, , , overrides] = recordSpy.mock.calls[0]!;
+        expect(usage).toMatchObject({
+            prompt_tokens: 10,
+            completion_tokens: 2,
+            energy_kwh: 0.000002,
+            energy_joules: 7.2,
+            billedUsage: 1,
+        });
+        expect(overrides).toMatchObject({
+            billedUsage: 0.00005 * 100_000_000,
+            energy_kwh: 0,
+        });
+    });
+});

src/backend/drivers/ai-chat/providers/neuralwatt/NeuralwattProvider.ts

@@ -0,0 +1,384 @@
+/*
+ * 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 axios from 'axios';
+import { OpenAI } from 'openai';
+import { ChatCompletionCreateParams } from 'openai/resources';
+import { Context } from '../../../../core/context.js';
+import { HttpError } from '../../../../core/http/HttpError.js';
+import type { MeteringService } from '../../../../services/metering/MeteringService.js';
+import { kv } from '../../../../util/kvSingleton.js';
+import * as OpenAIUtil from '../../utils/OpenAIUtil.js';
+import type {
+    IChatModel,
+    IChatProvider,
+    IChatCompleteResult,
+    ICompleteArguments,
+} from '../../types.js';
+import { inlineHttpImageUrls } from '../moonshot/imageHandling.js';
+import {
+    mapNeuralwattApiModel,
+    messagesHaveImageContent,
+    modelSupportsVision,
+    NEURALWATT_DEFAULT_MODEL,
+    NEURALWATT_ID_PREFIX,
+    stripNeuralwattPrefix,
+    type NeuralwattAccountingMethod,
+    type NeuralwattApiModel,
+    type NeuralwattCost,
+    type NeuralwattEnergy,
+} from './models.js';
+
+const DEFAULT_API_BASE_URL = 'https://api.neuralwatt.com/v1';
+const KV_MODELS_KEY = 'neuralwattChat:models';
+const KV_QUOTA_KEY = 'neuralwattChat:quota';
+const CACHE_TTL_SEC = 15 * 60;
+
+type NeuralwattUsage = OpenAI.Completions.CompletionUsage & {
+    request_cost_usd?: number;
+    energy_kwh?: number;
+    energy_joules?: number;
+    measurement_available?: boolean;
+};
+
+export class NeuralwattProvider implements IChatProvider {
+    #meteringService: MeteringService;
+
+    #openai: OpenAI;
+
+    #apiKey: string;
+
+    #apiBaseUrl: string = DEFAULT_API_BASE_URL;
+
+    constructor(
+        config: { apiBaseUrl?: string; apiKey: string },
+        meteringService: MeteringService,
+    ) {
+        this.#apiBaseUrl = config.apiBaseUrl || DEFAULT_API_BASE_URL;
+        this.#apiKey = config.apiKey;
+        this.#openai = new OpenAI({
+            apiKey: config.apiKey,
+            baseURL: this.#apiBaseUrl,
+        });
+        this.#meteringService = meteringService;
+    }
+
+    getDefaultModel() {
+        return NEURALWATT_DEFAULT_MODEL;
+    }
+
+    async list() {
+        const models = await this.models();
+        const modelNames: string[] = [];
+        for (const model of models) {
+            modelNames.push(model.id);
+            if (model.aliases) modelNames.push(...model.aliases);
+        }
+        return modelNames;
+    }
+
+    async models(): Promise<IChatModel[]> {
+        let apiModels = kv.get(KV_MODELS_KEY) as
+            | NeuralwattApiModel[]
+            | undefined;
+        if (!apiModels) {
+            try {
+                const resp = await axios.request({
+                    method: 'GET',
+                    url: `${this.#apiBaseUrl}/models`,
+                    headers: {
+                        Authorization: `Bearer ${this.#apiKey}`,
+                    },
+                });
+                apiModels = resp.data.data ?? [];
+                kv.set(KV_MODELS_KEY, apiModels, { EX: CACHE_TTL_SEC });
+            } catch (e) {
+                console.error(
+                    'Failed to fetch Neuralwatt models:',
+                    (e as Error).message,
+                );
+            }
+        }
+        if (!apiModels) return [];
+
+        const coerced: IChatModel[] = [];
+        for (const model of apiModels) {
+            if (model.metadata?.deprecated) continue;
+            const mapped = mapNeuralwattApiModel(model);
+            if (mapped) coerced.push(mapped);
+        }
+        return coerced;
+    }
+
+    /**
+     * Cached account accounting method (`energy` | `token`) from
+     * `GET /v1/quota`. Used only to annotate returned usage — billing
+     * always prefers `cost.request_cost_usd` on the completion.
+     */
+    async getAccountingMethod(): Promise<
+        NeuralwattAccountingMethod | undefined
+    > {
+        let method = kv.get(KV_QUOTA_KEY) as
+            | NeuralwattAccountingMethod
+            | undefined;
+        if (method === 'energy' || method === 'token') return method;
+
+        try {
+            const resp = await axios.request({
+                method: 'GET',
+                url: `${this.#apiBaseUrl}/quota`,
+                headers: {
+                    Authorization: `Bearer ${this.#apiKey}`,
+                },
+            });
+            const raw = resp.data?.balance?.accounting_method;
+            if (raw === 'energy' || raw === 'token') {
+                method = raw;
+                kv.set(KV_QUOTA_KEY, method, { EX: CACHE_TTL_SEC });
+                return method;
+            }
+        } catch (e) {
+            console.error(
+                'Failed to fetch Neuralwatt quota:',
+                (e as Error).message,
+            );
+        }
+        return undefined;
+    }
+
+    async complete({
+        messages,
+        stream,
+        model,
+        tools,
+        max_tokens,
+        temperature,
+        reasoning_effort,
+        reasoning,
+    }: ICompleteArguments): Promise<IChatCompleteResult> {
+        // Catalog carries per-model vision / reasoning_effort flags from
+        // Neuralwatt `GET /models` — resolve against it before shaping the
+        // upstream request so image-bearing prompts land on a vision model.
+        const availableModels = await this.models();
+        const hasImages = messagesHaveImageContent(messages ?? []);
+        const modelLower = (model ?? '').toLowerCase();
+        let modelUsed =
+            availableModels.find((m) =>
+                [m.id, ...(m.aliases || [])].some(
+                    (id) => id.toLowerCase() === modelLower,
+                ),
+            ) ?? undefined;
+
+        if (!modelUsed) {
+            if (hasImages) {
+                modelUsed = availableModels.find((m) =>
+                    modelSupportsVision(m),
+                );
+            }
+            modelUsed =
+                modelUsed ||
+                availableModels.find((m) => m.id === this.getDefaultModel()) ||
+                availableModels[0];
+        }
+
+        if (!modelUsed) {
+            throw new Error('No Neuralwatt models available');
+        }
+
+        if (hasImages && !modelSupportsVision(modelUsed)) {
+            throw new HttpError(
+                400,
+                `Model ${modelUsed.id} does not support image input`,
+                { legacyCode: 'bad_request' },
+            );
+        }
+
+        const modelIdForParams = stripNeuralwattPrefix(modelUsed.id);
+        const actor = Context.get('actor');
+        const accountingMethod = await this.getAccountingMethod();
+
+        // Vision models: Neuralwatt (like Moonshot) expects inline data URLs
+        // rather than remote http(s) fetches for image_url parts.
+        if (modelSupportsVision(modelUsed)) {
+            await inlineHttpImageUrls(messages);
+        }
+
+        messages = await OpenAIUtil.process_input_messages(messages);
+
+        const requestedReasoningEffort =
+            reasoning_effort ?? reasoning?.effort;
+        const supportsReasoningEffort = modelUsed.reasoning_effort === true;
+
+        const completionParams = {
+            messages,
+            model: modelIdForParams,
+            ...(tools ? { tools } : {}),
+            ...(max_tokens !== undefined ? { max_tokens } : {}),
+            ...(temperature !== undefined ? { temperature } : {}),
+            ...(supportsReasoningEffort && requestedReasoningEffort
+                ? { reasoning_effort: requestedReasoningEffort }
+                : {}),
+            stream,
+            ...(stream
+                ? {
+                      stream_options: { include_usage: true },
+                  }
+                : {}),
+        } as ChatCompletionCreateParams;
+
+        const completion =
+            await this.#openai.chat.completions.create(completionParams);
+
+        const usage_calculator = ({
+            usage,
+            cost,
+            energy,
+        }: {
+            usage: NeuralwattUsage;
+            cost?: NeuralwattCost;
+            energy?: NeuralwattEnergy;
+        }) => {
+            // Non-streaming spreads the full completion into this call;
+            // streaming merges top-level cost/energy onto `usage` via the
+            // index_usage_from_stream_chunk deviation below.
+            const requestCostUsd =
+                typeof cost?.request_cost_usd === 'number'
+                    ? cost.request_cost_usd
+                    : typeof usage.request_cost_usd === 'number'
+                      ? usage.request_cost_usd
+                      : undefined;
+
+            const energyBlock = energy ?? {
+                energy_kwh: usage.energy_kwh,
+                energy_joules: usage.energy_joules,
+                measurement_available: usage.measurement_available,
+            };
+
+            const trackedTokens = OpenAIUtil.extractMeteredUsage(usage);
+            const energyUnits: Record<string, number> = {};
+            if (
+                energyBlock.measurement_available !== false &&
+                typeof energyBlock.energy_kwh === 'number' &&
+                Number.isFinite(energyBlock.energy_kwh) &&
+                energyBlock.energy_kwh > 0
+            ) {
+                energyUnits.energy_kwh = energyBlock.energy_kwh;
+            }
+            if (
+                energyBlock.measurement_available !== false &&
+                typeof energyBlock.energy_joules === 'number' &&
+                Number.isFinite(energyBlock.energy_joules) &&
+                energyBlock.energy_joules > 0
+            ) {
+                energyUnits.energy_joules = energyBlock.energy_joules;
+            }
+
+            const annotate = (tracked: Record<string, number>) => {
+                const out: Record<string, number | string> = { ...tracked };
+                if (accountingMethod) {
+                    out.accounting_method = accountingMethod;
+                }
+                return out;
+            };
+
+            if (
+                typeof requestCostUsd === 'number' &&
+                Number.isFinite(requestCostUsd)
+            ) {
+                const billedTrackedUsage = {
+                    ...trackedTokens,
+                    ...energyUnits,
+                    billedUsage: 1,
+                };
+                const costOverwrites = Object.fromEntries(
+                    Object.keys(billedTrackedUsage).map((k) => [k, 0]),
+                );
+                costOverwrites.billedUsage = requestCostUsd * 100_000_000;
+                this.#meteringService.utilRecordUsageObject(
+                    billedTrackedUsage,
+                    actor!,
+                    modelUsed.id,
+                    costOverwrites,
+                );
+                const result = annotate(billedTrackedUsage);
+                result.usd_cents = requestCostUsd * 100;
+                return result;
+            }
+
+            // Fallback: catalog token rates (preflight / when Neuralwatt
+            // omits request_cost_usd).
+            const trackedUsage = { ...trackedTokens, ...energyUnits };
+            const costOverwrites = Object.fromEntries(
+                Object.entries(trackedUsage).map(([k, v]) => {
+                    if (k === 'energy_kwh' || k === 'energy_joules') {
+                        return [k, 0];
+                    }
+                    return [k, (modelUsed.costs[k] ?? 0) * v];
+                }),
+            );
+            this.#meteringService.utilRecordUsageObject(
+                trackedUsage,
+                actor!,
+                modelUsed.id,
+                costOverwrites,
+            );
+            return annotate(trackedUsage);
+        };
+
+        return OpenAIUtil.handle_completion_output({
+            deviations: {
+                index_usage_from_stream_chunk: (chunk: {
+                    usage?: NeuralwattUsage;
+                    cost?: NeuralwattCost;
+                    energy?: NeuralwattEnergy;
+                }) => {
+                    if (!chunk.usage) return chunk.usage;
+                    return {
+                        ...chunk.usage,
+                        ...(typeof chunk.cost?.request_cost_usd === 'number'
+                            ? {
+                                  request_cost_usd:
+                                      chunk.cost.request_cost_usd,
+                              }
+                            : {}),
+                        ...(chunk.energy
+                            ? {
+                                  energy_kwh: chunk.energy.energy_kwh,
+                                  energy_joules: chunk.energy.energy_joules,
+                                  measurement_available:
+                                      chunk.energy.measurement_available,
+                              }
+                            : {}),
+                    };
+                },
+            },
+            usage_calculator,
+            stream,
+            completion,
+        });
+    }
+
+    checkModeration(
+        _text: string,
+    ): ReturnType<IChatProvider['checkModeration']> {
+        throw new Error('Method not implemented.');
+    }
+}
+
+export { NEURALWATT_ID_PREFIX, NEURALWATT_DEFAULT_MODEL };

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

@@ -0,0 +1,186 @@
+/*
+ * 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 { IChatModel } from '../../types.js';
+import { usdPerMToken } from '../../utils/pricing.js';
+
+export const NEURALWATT_ID_PREFIX = 'neuralwatt:';
+export const NEURALWATT_DEFAULT_MODEL = 'neuralwatt:deepseek-v4-flash';
+
+/**
+ * Shape of one entry in Neuralwatt's `GET /v1/models` catalog.
+ * Pricing is USD per million tokens under `metadata.pricing` and is used
+ * only for Puter preflight estimates / token-cost fallback. Live billing
+ * uses each completion's `cost.request_cost_usd`.
+ */
+
+export type NeuralwattApiModel = {
+    id: string;
+    object?: string;
+    created?: number;
+    owned_by?: string;
+    max_model_len?: number;
+    metadata?: {
+        display_name?: string;
+        description?: string | null;
+        provider?: string;
+        huggingface_id?: string | null;
+        pricing?: {
+            input_per_million?: number;
+            output_per_million?: number;
+            cached_input_per_million?: number | null;
+            cached_output_per_million?: number | null;
+            currency?: string;
+            pricing_tbd?: boolean;
+        };
+        capabilities?: {
+            tools?: boolean;
+            json_mode?: boolean;
+            vision?: boolean;
+            reasoning?: boolean;
+            reasoning_effort?: boolean;
+            streaming?: boolean;
+            system_role?: boolean;
+            developer_role?: boolean;
+        };
+        limits?: {
+            max_context_length?: number | null;
+            max_output_tokens?: number | null;
+            max_images?: number | null;
+        };
+        deprecated?: boolean;
+        deprecated_message?: string | null;
+    };
+};
+
+export type NeuralwattAccountingMethod = 'energy' | 'token';
+
+export type NeuralwattEnergy = {
+    energy_joules?: number;
+    energy_kwh?: number;
+    measurement_available?: boolean;
+    avg_power_watts?: number;
+    duration_seconds?: number;
+    attribution_method?: string;
+    attribution_ratio?: number;
+};
+
+export type NeuralwattCost = {
+    request_cost_usd?: number;
+    cache_savings_usd?: number;
+    allowance_remaining_usd?: number;
+};
+
+/** Strip the Puter-facing `neuralwatt:` prefix for upstream API calls. */
+export const stripNeuralwattPrefix = (modelId: string): string =>
+    modelId.startsWith(NEURALWATT_ID_PREFIX)
+        ? modelId.slice(NEURALWATT_ID_PREFIX.length)
+        : modelId;
+
+/**
+ * Map a Neuralwatt catalog entry to Puter's `IChatModel`. Returns `null`
+ * when pricing is TBD (placeholders) so the model is not offered for
+ * preflight credit checks until Neuralwatt publishes real rates.
+ */
+export const mapNeuralwattApiModel = (
+    model: NeuralwattApiModel,
+): IChatModel | null => {
+    const pricing = model.metadata?.pricing;
+    if (pricing?.pricing_tbd) return null;
+
+    const inputUsd = Number(pricing?.input_per_million ?? 0);
+    const outputUsd = Number(pricing?.output_per_million ?? 0);
+    const cachedUsd =
+        pricing?.cached_input_per_million == null
+            ? 0
+            : Number(pricing.cached_input_per_million);
+
+    const caps = model.metadata?.capabilities;
+    const limits = model.metadata?.limits;
+    const context =
+        limits?.max_context_length ?? model.max_model_len ?? undefined;
+    const maxTokens = limits?.max_output_tokens ?? context ?? 0;
+
+    const inputModalities = ['text'];
+    if (caps?.vision) inputModalities.push('image');
+
+    const displayName = model.metadata?.display_name || model.id;
+    const pathTail = model.id.includes('/')
+        ? model.id.split('/').slice(1).join('/')
+        : undefined;
+
+    return {
+        id: `${NEURALWATT_ID_PREFIX}${model.id}`,
+        name: `${displayName} (Neuralwatt)`,
+        aliases: [
+            model.id,
+            `neuralwatt/${model.id}`,
+            ...(pathTail && pathTail !== model.id ? [pathTail] : []),
+        ],
+        context,
+        max_tokens: maxTokens,
+        costs_currency: 'usd-cents',
+        input_cost_key: 'prompt_tokens',
+        output_cost_key: 'completion_tokens',
+        costs: usdPerMToken(inputUsd, outputUsd, cachedUsd),
+        modalities: { input: inputModalities, output: ['text'] },
+        tool_call: caps?.tools === true,
+        // Surfaced from Neuralwatt `/models` so `complete()` can gate
+        // reasoning_effort without re-fetching the catalog entry.
+        reasoning_effort: caps?.reasoning_effort === true,
+        ...(typeof limits?.max_images === 'number'
+            ? { max_images: limits.max_images }
+            : {}),
+        ...(model.metadata?.description
+            ? { description: model.metadata.description }
+            : {}),
+        ...(model.created
+            ? {
+                  release_date: new Date(model.created * 1000)
+                      .toISOString()
+                      .slice(0, 10),
+              }
+            : {}),
+    };
+};
+
+/** True when the Puter model catalog entry advertises image input. */
+export const modelSupportsVision = (model: IChatModel): boolean =>
+    Array.isArray(model.modalities?.input) &&
+    model.modalities.input.includes('image');
+
+/**
+ * Detect image / puter_path parts so we can prefer a vision-capable model
+ * from the Neuralwatt catalog (or reject a text-only pick).
+ */
+export const messagesHaveImageContent = (
+    messages: Array<{ content?: unknown }>,
+): boolean => {
+    for (const message of messages) {
+        if (!Array.isArray(message.content)) continue;
+        for (const part of message.content as Array<Record<string, unknown>>) {
+            if (!part || typeof part !== 'object') continue;
+            if (part.type === 'image_url' || part.image_url) return true;
+            if (typeof part.puter_path === 'string' && part.puter_path) {
+                return true;
+            }
+        }
+    }
+    return false;
+};

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

@@ -728,19 +728,4 @@ export const OPEN_AI_MODELS: IChatModel[] = [
         context: 1_047_576,
         max_tokens: 32768,
     },
-    {
-        puterId: 'openai:openai/gpt-4.5-preview',
-        id: 'gpt-4.5-preview',
-        aliases: ['openai/gpt-4.5-preview'],
-        costs_currency: 'usd-cents',
-        input_cost_key: 'prompt_tokens',
-        output_cost_key: 'completion_tokens',
-        costs: {
-            tokens: 1_000_000,
-            prompt_tokens: 7500,
-            completion_tokens: 15000,
-        },
-        context: 128_000,
-        max_tokens: 32768,
-    },
 ];

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

@@ -105,6 +105,15 @@ const SAMPLE_API_MODELS = [
         pricing: { prompt: 0.000002, completion: 0.00001 },
         top_provider: { max_completion_tokens: 8192 },
     },
+    {
+        // OpenRouter reports no separate output cap for some models.
+        id: 'meta/muse-spark-1.2',
+        name: 'Muse Spark 1.2',
+        created: 1786032000,
+        context_length: 1048576,
+        pricing: { prompt: 0.00000125, completion: 0.00000425 },
+        top_provider: { max_completion_tokens: null },
+    },
     {
         // 'openrouter/auto' is filtered out — disallowed.
         id: 'openrouter/auto',
@@ -228,6 +237,25 @@ describe('OpenRouterProvider model catalog', () => {
         expect(axiosRequestMock).toHaveBeenCalledTimes(1);
     });
 
+    it('falls back to the context window when no output cap is reported', async () => {
+        const { provider } = makeProvider();
+        const models = await provider.models();
+
+        // A null max_completion_tokens previously landed on the model as-is,
+        // which made the driver's cap arithmetic go negative and reject the
+        // request as insufficient funds.
+        const noCap = models.find(
+            (m) => m.id === 'openrouter:meta/muse-spark-1.2',
+        )!;
+        expect(noCap.max_tokens).toBe(1048576);
+
+        // A model that reports its own cap keeps it.
+        const capped = models.find(
+            (m) => m.id === 'openrouter:openai/gpt-5-nano',
+        )!;
+        expect(capped.max_tokens).toBe(16000);
+    });
+
     it('maps OpenRouter created timestamps to release_date metadata', async () => {
         const { provider } = makeProvider();
         const models = await provider.models();

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

@@ -72,9 +72,11 @@ export class OpenRouterProvider implements IChatProvider {
     }
     /**
      * Returns a list of available model names including their aliases
+     *
+     * Retrieves all available model IDs and their aliases, flattening them into
+     * a single array of strings that can be used for model selection
+     *
      * @returns {Promise<string[]>} Array of model identifiers and their aliases
-     * @description Retrieves all available model IDs and their aliases,
-     * flattening them into a single array of strings that can be used for model selection
      */
     async list() {
         const models = await this.models();
@@ -85,10 +87,7 @@ export class OpenRouterProvider implements IChatProvider {
         return model_names;
     }
 
-    /**
-     * AI Chat completion method.
-     * See AIChatService for more details.
-     */
+    /** AI Chat completion method. See AIChatService for more details. */
     async complete({
         messages,
         stream,
@@ -279,7 +278,11 @@ export class OpenRouterProvider implements IChatProvider {
                     model.id.split('/').slice(1).join('/'),
                 ],
                 context: model.context_length,
-                max_tokens: model.top_provider.max_completion_tokens,
+                // OpenRouter leaves max_completion_tokens null when a model
+                // declares no output cap separate from its context window.
+                max_tokens:
+                    model.top_provider.max_completion_tokens ??
+                    model.context_length,
                 costs_currency: 'usd-cents',
                 input_cost_key: 'prompt',
                 output_cost_key: 'completion',

src/backend/drivers/ai-chat/providers/together/TogetherAIProvider.integration.test.ts

@@ -63,5 +63,33 @@ describe.skipIf(skipUnlessEnv(ENV_VAR))(
                 ?.content;
             expect(typeof text === 'string' && text.length > 0).toBe(true);
         });
+
+        it('recovers when max_tokens leaves no room for the prompt', { timeout: INTEGRATION_TEST_TIMEOUT_MS }, async () => {
+            const provider = new TogetherAIProvider(
+                { apiKey: optionalEnv(ENV_VAR)! },
+                makeMeteringStub(),
+            );
+
+            // Asking for the model's whole context as output leaves no room
+            // for the prompt, which Together rejects outright — the provider
+            // should retry uncapped rather than surface a 400.
+            const model = (await provider.models()).find(
+                (m) => m.id === 'togetherai:Qwen/Qwen2.5-7B-Instruct-Turbo',
+            )!;
+
+            const result = await withTestActor(() =>
+                provider.complete({
+                    model: model.id,
+                    messages: [
+                        { role: 'user', content: 'Say hi in one word.' },
+                    ],
+                    max_tokens: model.context!,
+                }),
+            );
+
+            const text = (result as { message?: { content?: string } }).message
+                ?.content;
+            expect(typeof text === 'string' && text.length > 0).toBe(true);
+        });
     },
 );

src/backend/drivers/ai-chat/providers/together/TogetherAIProvider.test.ts

@@ -209,6 +209,17 @@ describe('TogetherAIProvider model catalog', () => {
         expect(ids).toContain('model-fallback-test-1');
     });
 
+    it('reserves headroom under the context length for the output cap', async () => {
+        const { provider } = makeProvider();
+        const model = (await provider.models()).find(
+            (m) => m.id === 'togetherai:Qwen/Qwen2.5-7B-Instruct-Turbo',
+        )!;
+        // The advertised context window is unchanged; only the output cap
+        // leaves room for the driver's under-counting input estimator.
+        expect(model.context).toBe(32768);
+        expect(model.max_tokens).toBe(Math.floor(32768 * 0.95));
+    });
+
     it('caches the coerced model list in kv after the first call', async () => {
         const { provider } = makeProvider();
         await provider.models();
@@ -357,6 +368,109 @@ describe('TogetherAIProvider.complete request shape', () => {
         expect(createMock).not.toHaveBeenCalled();
         expect(recordSpy).not.toHaveBeenCalled();
     });
+
+    // Together's APIError carries the whole response body on `.error`, so the
+    // provider message sits at `.error.error.message` — one level deeper than
+    // the OpenAI SDK puts it.
+    const contextLengthError = {
+        status: 400,
+        message:
+            '400 {"id":"ovG6YRd-6z2FuN","error":{"message":"Failed to start generation: The input token count (11) plus the requested output count (1048573) exceeds the model\'s maximum context length (1048576)"}}',
+        error: {
+            id: 'ovG6YRd-6z2FuN',
+            error: {
+                message:
+                    "Failed to start generation: The input token count (11) plus the requested output count (1048573) exceeds the model's maximum context length (1048576)",
+                type: 'invalid_request_error',
+            },
+        },
+    };
+
+    it('retries without max_tokens when Together rejects with a context-length error', async () => {
+        const { provider } = makeProvider();
+
+        createMock
+            .mockRejectedValueOnce(contextLengthError)
+            .mockResolvedValueOnce(baseCompletion);
+
+        await withTestActor(() =>
+            provider.complete({
+                model: 'togetherai:Qwen/Qwen2.5-7B-Instruct-Turbo',
+                messages: [{ role: 'user', content: 'hi' }],
+                max_tokens: 1048573,
+            }),
+        );
+
+        // Provider mutates a single completionParams object across both calls
+        // (`delete completionParams.max_tokens` after the first throw), so the
+        // first call's recorded args are retroactively altered — only the
+        // surviving shape is worth asserting on.
+        expect(createMock).toHaveBeenCalledTimes(2);
+        expect('max_tokens' in createMock.mock.calls[1]![0]).toBe(false);
+        // The retry is still metered exactly once.
+        expect(recordSpy).toHaveBeenCalledTimes(1);
+    });
+
+    it('retries a streaming request the same way', async () => {
+        const { provider } = makeProvider();
+
+        createMock
+            .mockRejectedValueOnce(contextLengthError)
+            .mockReturnValueOnce(asAsyncIterable([]));
+
+        await withTestActor(() =>
+            provider.complete({
+                model: 'togetherai:Qwen/Qwen2.5-7B-Instruct-Turbo',
+                messages: [{ role: 'user', content: 'hi' }],
+                max_tokens: 1048573,
+                stream: true,
+            }),
+        );
+
+        expect(createMock).toHaveBeenCalledTimes(2);
+        expect('max_tokens' in createMock.mock.calls[1]![0]).toBe(false);
+        expect(createMock.mock.calls[1]![0].stream).toBe(true);
+    });
+
+    it('rethrows non-context-length errors without retrying', async () => {
+        const { provider } = makeProvider();
+        const apiError = {
+            status: 401,
+            error: { error: { message: 'Invalid API key' } },
+        };
+        createMock.mockRejectedValueOnce(apiError);
+
+        await expect(
+            withTestActor(() =>
+                provider.complete({
+                    model: 'togetherai:Qwen/Qwen2.5-7B-Instruct-Turbo',
+                    messages: [{ role: 'user', content: 'boom' }],
+                    max_tokens: 100,
+                }),
+            ),
+        ).rejects.toBe(apiError);
+
+        expect(createMock).toHaveBeenCalledTimes(1);
+        expect(recordSpy).not.toHaveBeenCalled();
+    });
+
+    it('rethrows a bodyless connection error instead of failing to read it', async () => {
+        const { provider } = makeProvider();
+        const connErr = new Error('Connection error.');
+        createMock.mockRejectedValueOnce(connErr);
+
+        await expect(
+            withTestActor(() =>
+                provider.complete({
+                    model: 'togetherai:Qwen/Qwen2.5-7B-Instruct-Turbo',
+                    messages: [{ role: 'user', content: 'hi' }],
+                    max_tokens: 100,
+                }),
+            ),
+        ).rejects.toBe(connErr);
+
+        expect(createMock).toHaveBeenCalledTimes(1);
+    });
 });
 
 // ── Model resolution ────────────────────────────────────────────────

src/backend/drivers/ai-chat/providers/together/TogetherAIProvider.ts

@@ -29,6 +29,24 @@ const TOGETHER_AI_CHAT_COST_MAP = {
     completion_tokens: 'output',
 };
 
+/**
+ * Whether the SDK rejected a request because the prompt plus the requested
+ * output exceeds the model's context window. Unlike the OpenAI SDK, Together's
+ * `APIError.error` is the whole response body, so the provider message sits one
+ * level deeper; `message` is the stringified body and covers older shapes.
+ */
+const isContextLengthError = (e: unknown) => {
+    const err = e as {
+        error?: { error?: { message?: string } };
+        message?: string;
+    };
+    const message = err?.error?.error?.message ?? err?.message;
+    return (
+        typeof message === 'string' &&
+        message.includes('maximum context length')
+    );
+};
+
 export class TogetherAIProvider implements IChatProvider {
     #together: Together;
 
@@ -83,7 +101,13 @@ export class TogetherAIProvider implements IChatProvider {
                             ),
                         ),
                     },
-                    max_tokens: model.context_length ?? 8000,
+                    // Together only reports a context length. The driver caps
+                    // output at max_tokens minus an estimated input count, and
+                    // that estimate runs low — reserve headroom so a short
+                    // prompt doesn't ask for more than the context allows.
+                    max_tokens: model.context_length
+                        ? Math.floor(model.context_length * 0.95)
+                        : 8000,
                 });
             }
         }
@@ -145,7 +169,7 @@ export class TogetherAIProvider implements IChatProvider {
 
         messages = await OpenAIUtil.process_input_messages(messages);
 
-        const completion = await this.#together.chat.completions.create({
+        const completionParams = {
             model: modelIdForParams,
             messages,
             stream,
@@ -153,7 +177,21 @@ export class TogetherAIProvider implements IChatProvider {
             ...(max_tokens !== undefined ? { max_tokens } : {}),
             ...(temperature !== undefined ? { temperature } : {}),
             ...(stream ? { stream_options: { include_usage: true } } : {}),
-        } as Together.Chat.Completions.CompletionCreateParamsNonStreaming);
+        } as Together.Chat.Completions.CompletionCreateParamsNonStreaming;
+
+        let completion;
+        try {
+            completion =
+                await this.#together.chat.completions.create(completionParams);
+        } catch (e: unknown) {
+            // An overestimated max_tokens makes Together reject the request
+            // outright rather than truncating. The user can afford the query
+            // either way, so retry once without the cap.
+            if (!isContextLengthError(e)) throw e;
+            delete completionParams.max_tokens;
+            completion =
+                await this.#together.chat.completions.create(completionParams);
+        }
 
         return OpenAIUtil.handle_completion_output({
             usage_calculator: ({ usage }) => {

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

@@ -0,0 +1,188 @@
+/**
+ * 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/](https://www.gnu.org/licenses/).
+ */
+
+import { describe, expect, it } from 'vitest';
+import { GEMINI_MODELS } from '../providers/gemini/models.js';
+import type { IChatModel } from '../types.js';
+import {
+    compareModelPreference,
+    isIdentityKey,
+    normalizeModelKey,
+} from './modelRouting.js';
+
+// `#buildModelMap` mutates the catalogs providers hand back, and
+// `GeminiChatProvider.models()` returns the module-level `GEMINI_MODELS` by
+// reference — clone so these fixtures can't be perturbed by another suite.
+const geminiModel = (id: string, provider = 'gemini'): IChatModel => {
+    const found = GEMINI_MODELS.find((m) => m.id === id);
+    if (!found) throw new Error(`no such gemini model: ${id}`);
+    return { ...structuredClone(found), provider };
+};
+
+// Mirrors how the reseller providers coerce a catalog entry: `<gateway>:` on
+// the id, `input_cost_key: 'prompt'`, and prices as microcents per token.
+const resoldModel = (
+    provider: string,
+    catalogId: string,
+    promptCost: number,
+): IChatModel =>
+    ({
+        id: `${provider}:${catalogId}`,
+        name: `${catalogId} (${provider})`,
+        aliases: [catalogId, catalogId.split('/').slice(1).join('/')],
+        costs_currency: 'usd-cents',
+        input_cost_key: 'prompt',
+        output_cost_key: 'completion',
+        costs: { tokens: 1_000_000, prompt: promptCost, completion: 100 },
+        provider,
+    }) as IChatModel;
+
+const winner = (...models: IChatModel[]) =>
+    [...models].sort(compareModelPreference)[0];
+
+describe('compareModelPreference', () => {
+    it('serves the vendor directly even when a reseller quotes a lower price', () => {
+        // Google lists gemini-2.5-flash input at 30 microcents/token; the
+        // gateway advertises a floor price across its upstream routes and
+        // undercuts it. Price must not decide who serves the request.
+        const direct = geminiModel('gemini-2.5-flash');
+        const resold = resoldModel('infron', 'google/gemini-2.5-flash', 15);
+
+        expect(direct.costs.prompt_tokens).toBeGreaterThan(
+            resold.costs.prompt as number,
+        );
+        expect(winner(resold, direct).provider).toBe('gemini');
+        expect(winner(direct, resold).provider).toBe('gemini');
+    });
+
+    it('ranks every reseller behind the vendor, not just one of them', () => {
+        const direct = geminiModel('gemini-3.1-pro-preview');
+        const bucket = [
+            resoldModel('openrouter', 'google/gemini-3.1-pro-preview', 90),
+            resoldModel('infron', 'google/gemini-3.1-pro-preview', 80),
+            resoldModel('together-ai', 'google/gemini-3.1-pro-preview', 70),
+            resoldModel('neuralwatt', 'google/gemini-3.1-pro-preview', 60),
+            direct,
+        ];
+
+        expect(
+            bucket.sort(compareModelPreference).map((m) => m.provider),
+        ).toEqual([
+            'gemini',
+            'neuralwatt',
+            'infron',
+            'openrouter',
+            'together-ai',
+        ]);
+    });
+
+    it('keeps together-ai strictly behind the other resellers', () => {
+        // A flat direct/reseller split would let these tie and re-order by
+        // price; together-ai stays last regardless of how cheap it quotes.
+        const together = resoldModel('together-ai', 'meta/llama-4', 1);
+        const openrouter = resoldModel('openrouter', 'meta/llama-4', 500);
+
+        expect(winner(together, openrouter).provider).toBe('openrouter');
+    });
+
+    it('puts openrouter then together-ai at the very bottom of the chain', () => {
+        // Both quote well under the other resellers — price must not lift
+        // either of them out of the last two slots.
+        const bucket = [
+            resoldModel('together-ai', 'meta/llama-4', 1),
+            resoldModel('openrouter', 'meta/llama-4', 2),
+            resoldModel('neuralwatt', 'meta/llama-4', 400),
+            resoldModel('infron', 'meta/llama-4', 500),
+        ];
+
+        expect(
+            bucket.sort(compareModelPreference).map((m) => m.provider),
+        ).toEqual(['neuralwatt', 'infron', 'openrouter', 'together-ai']);
+    });
+
+    it('still orders two direct providers by cheapest input cost', () => {
+        const cheap = geminiModel('gemini-2.5-flash-lite');
+        const pricey = geminiModel('gemini-2.5-pro');
+
+        expect(cheap.costs.prompt_tokens).toBeLessThan(
+            pricey.costs.prompt_tokens as number,
+        );
+        expect(winner(pricey, cheap).id).toBe('gemini-2.5-flash-lite');
+    });
+
+    it('breaks price ties on the shorter id', () => {
+        const short = { ...geminiModel('gemini-2.5-flash'), id: 'gemini-x' };
+        const long = {
+            ...geminiModel('gemini-2.5-flash'),
+            id: 'some-vendor/gemini-x-2025-preview',
+            provider: 'azure-openai',
+        };
+
+        expect(winner(long, short).id).toBe('gemini-x');
+    });
+
+    it('leaves a reseller serving models no vendor provider carries', () => {
+        // The image-preview models are absent from GEMINI_MODELS, so the
+        // gateway is the only route and must stay the winner.
+        expect(
+            GEMINI_MODELS.some(
+                (m) => m.id === 'gemini-2.5-flash-image-preview',
+            ),
+        ).toBe(false);
+
+        const onlyRoute = resoldModel(
+            'infron',
+            'google/gemini-2.5-flash-image-preview',
+            20,
+        );
+        expect(winner(onlyRoute).provider).toBe('infron');
+    });
+});
+
+describe('isIdentityKey', () => {
+    it('accepts the machine ids a catalog uses to name a model', () => {
+        for (const key of [
+            'claude-sonnet-4',
+            'anthropic/claude-sonnet-4',
+            'openrouter:anthropic/claude-sonnet-4',
+            'gpt-4o',
+        ]) {
+            expect(isIdentityKey(key)).toBe(true);
+        }
+    });
+
+    it('rejects display names, so a shared label cannot merge two providers', () => {
+        // Gateways carry these alongside the machine ids. Merging on one
+        // would be merging two providers on a human-readable string.
+        for (const key of [
+            normalizeModelKey('Google: Gemini 2.5 Flash'),
+            normalizeModelKey('Anthropic: Claude Sonnet 4'),
+            normalizeModelKey('Meta Llama 3.1 8B Instruct Turbo'),
+        ]) {
+            expect(isIdentityKey(key)).toBe(false);
+        }
+    });
+
+    it('rejects the empty key catalogs produce for ids carrying no vendor org', () => {
+        // `'gpt-4o'.split('/').slice(1).join('/')` is '' — pooling models
+        // under that key would put unrelated models in one bucket.
+        expect(isIdentityKey('')).toBe(false);
+    });
+});

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

@@ -0,0 +1,92 @@
+/**
+ * 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/](https://www.gnu.org/licenses/).
+ */
+
+import type { IChatModel } from '../types.js';
+
+/**
+ * Providers that resell other vendors' models rather than serving their own.
+ * Their catalogs duplicate models we already reach directly, so their entries
+ * are kept only as fallback routes.
+ */
+export const AGGREGATOR_PROVIDERS = new Set([
+    'together-ai',
+    'openrouter',
+    'infron',
+    'neuralwatt',
+]);
+
+// Lower rank is served first. `openrouter` and `together-ai` sit at the very
+// bottom, in that order, behind the other resellers.
+const providerRank = (provider?: string): number => {
+    if (provider === 'together-ai') return 3;
+    if (provider === 'openrouter') return 2;
+    if (provider && AGGREGATOR_PROVIDERS.has(provider)) return 1;
+    return 0;
+};
+
+/**
+ * Lookup form for a model id or alias: the model map is keyed case- and
+ * whitespace-insensitively.
+ */
+export const normalizeModelKey = (key: string): string =>
+    key.trim().toLowerCase();
+
+/**
+ * Whether a key asserts _which model this is_, rather than merely being another
+ * way to name it.
+ *
+ * Catalogs mix both into `aliases`: machine ids (`anthropic/claude-sonnet-4`,
+ * `claude-sonnet-4`) alongside human labels (`Anthropic: Claude Sonnet 4`).
+ * Only the former may pull an entry into another provider's bucket — two
+ * gateways agreeing on a display string is not evidence they serve the same
+ * weights, and a label collision would otherwise silently reroute traffic.
+ * Labels stay usable for lookup; they just don't merge anything.
+ *
+ * Vendor model ids never contain whitespace, and every display name in the
+ * catalogs we consume does — that separation is the whole test.
+ */
+export const isIdentityKey = (key: string): boolean =>
+    key.length > 0 && !/\s/.test(key);
+
+/**
+ * Orders the candidates that share a model bucket; the first one gets served.
+ *
+ * Direct vendors outrank resellers regardless of quoted price. Resellers
+ * advertise a floor price across their upstream routes, so on price alone they
+ * undercut the vendor's list price and capture traffic for models we hold a
+ * direct integration for. Within a rank, cheapest input cost wins and ties
+ * break by shorter id — usually the official name over a qualified one.
+ */
+export const compareModelPreference = (
+    a: IChatModel,
+    b: IChatModel,
+): number => {
+    const rankDiff = providerRank(a.provider) - providerRank(b.provider);
+    if (rankDiff !== 0) return rankDiff;
+
+    const aCost = a.costs[
+        (a.input_cost_key as string) || 'input_tokens'
+    ] as number;
+    const bCost = b.costs[
+        (b.input_cost_key as string) || 'input_tokens'
+    ] as number;
+    if (aCost === bCost) return a.id.length - b.id.length;
+    return aCost - bCost;
+};