Skip to content

Open source CapCut alternative.

TypeScript

fix(site): per-user pacing, dev bypass as a server switch, card route misses

David Lebce70c044bdf
diff --git a/site/.env.example b/site/.env.example
index a11b8a11..10668667 100644
--- a/site/.env.example
+++ b/site/.env.example
@@ -9,9 +9,6 @@ CRON_SECRET="replace-with-a-long-random-secret"
GOOGLE_CLIENT_ID="replace-with-google-web-client-id"
GOOGLE_CLIENT_SECRET="replace-with-google-web-client-secret"
-# Required for macOS desktop computer-use through OpenAI.
-OPENAI_API_KEY="replace-with-openai-api-key"
-
# Optional hosted inference providers.
ELEVENLABS_API_KEY="replace-with-elevenlabs-api-key"
@@ -22,10 +19,13 @@ FAL_KEY="replace-with-fal-api-key"
# uses the global endpoint and derives the project from project_id.
GOOGLE_APPLICATION_CREDENTIALS_JSON='{"type":"service_account",...}'
-# Dev-only: a Gemini Developer API key. When set, the Live client connects
-# directly to the Developer API (using GEMINI_LIVE_MODEL; the client default is an
-# audio-output-only model) instead of minting a Vertex token. Production uses the
-# keyless Vertex path above.
+# Dev-only: lets the eval scripts and a hand-run worker call the dev server
+# with the x-donkey-dev-auth-bypass header (no session, no credits). Only read
+# when NODE_ENV is not production; a hosted deployment never sets it.
+DONKEY_DEV_AUTH_BYPASS=1
+
+# Dev-only: a Gemini Developer API key. Production uses the keyless Vertex
+# path above.
GEMINI_API_KEY="replace-with-gemini-api-key"
# Stripe. The Pro subscription plus one-time credit top-ups run through
@@ -39,10 +39,6 @@ STRIPE_WEBHOOK_SECRET="replace-with-stripe-webhook-signing-secret"
STRIPE_PRO_PRICE_ID="replace-with-stripe-pro-price-id"
STRIPE_PORTAL_CONFIGURATION_ID="replace-with-stripe-portal-configuration-id"
-# Browser Use Cloud (agentic web automation via /api/browser/run). Backend-only
-# secret; the Mac app never carries it. Key starts with "bu_".
-BROWSER_USE_API_KEY="replace-with-browser-use-api-key"
-
# Resend (transactional email + contact sync). Unset in dev skips email work:
# signups log the skip and provision everything else. The from address is code
# (src/lib/email/resend.ts). RESEND_WEBHOOK_SECRET signs the contact.updated
diff --git a/site/scripts/lib/cut-eval/README.md b/site/scripts/lib/cut-eval/README.md
index 0172dfae..bae543b6 100644
--- a/site/scripts/lib/cut-eval/README.md
+++ b/site/scripts/lib/cut-eval/README.md
@@ -9,7 +9,8 @@ npm run eval:cut-chat -- [--runs N] [--only <case>] [--bucket chat|single-tool|m
```
It replays real composer turns against the live chat model through the dev
-server's hosted Responses route. Start `next dev` on :3000 first. Auth is the
+server's hosted Responses route. Start `next dev` on :3000 first, with
+`DONKEY_DEV_AUTH_BYPASS=1` in its env (`.env.example` carries it). Auth is the
dev bypass header, so runs spend no credits. The spoken fixture is synthesized
with macOS `say`, so runs need a Mac.
diff --git a/site/src/app/api/inference/assets/refresh/route.ts b/site/src/app/api/inference/assets/refresh/route.ts
index dd6e5769..67bb9683 100644
--- a/site/src/app/api/inference/assets/refresh/route.ts
+++ b/site/src/app/api/inference/assets/refresh/route.ts
@@ -13,7 +13,10 @@ import {
validationErrorResponse,
} from "@/lib/inference/responses";
import { storedGenerationForProviderSchema } from "@/lib/inference/schemas";
-import { withDonkeyAuth } from "@/lib/donkey-api-auth";
+import {
+ INFERENCE_RATE_LIMIT,
+ withDonkeyAuth,
+} from "@/lib/donkey-api-auth";
import { InferenceProviderError } from "@/lib/inference/providers";
export const POST = withDonkeyAuth(async (request) => {
@@ -82,4 +85,4 @@ export const POST = withDonkeyAuth(async (request) => {
throw error;
}
-}, { allowRunner: true });
+}, { allowRunner: true, rateLimit: INFERENCE_RATE_LIMIT });
diff --git a/site/src/app/api/inference/assets/route.ts b/site/src/app/api/inference/assets/route.ts
index 3cd21a47..07634631 100644
--- a/site/src/app/api/inference/assets/route.ts
+++ b/site/src/app/api/inference/assets/route.ts
@@ -20,7 +20,10 @@ import {
validationErrorResponse,
} from "@/lib/inference/responses";
import { assetGenerationRequestSchema } from "@/lib/inference/schemas";
-import { withDonkeyAuth } from "@/lib/donkey-api-auth";
+import {
+ INFERENCE_RATE_LIMIT,
+ withDonkeyAuth,
+} from "@/lib/donkey-api-auth";
import { resolveInferenceBlobs } from "@/lib/inference/blobs";
import { InferenceProviderError, type JsonObject } from "@/lib/inference/providers";
import { toJsonObject, toJsonValue } from "@/lib/inference/json";
@@ -223,4 +226,4 @@ export const POST = withDonkeyAuth(async (request) => {
throw error;
}
-}, { allowRunner: true });
+}, { allowRunner: true, rateLimit: INFERENCE_RATE_LIMIT });
diff --git a/site/src/app/api/inference/models/route.ts b/site/src/app/api/inference/models/route.ts
index 375c26bb..b1039d07 100644
--- a/site/src/app/api/inference/models/route.ts
+++ b/site/src/app/api/inference/models/route.ts
@@ -3,7 +3,10 @@ import { NextResponse } from "next/server";
import { createProviderRegistry } from "@/lib/inference/router";
import { requireInferenceClientId } from "@/lib/inference/responses";
import { parseRequestedModalities } from "@/lib/inference/schemas";
-import { withDonkeyAuth } from "@/lib/donkey-api-auth";
+import {
+ INFERENCE_RATE_LIMIT,
+ withDonkeyAuth,
+} from "@/lib/donkey-api-auth";
export const GET = withDonkeyAuth(async (request) => {
const client = requireInferenceClientId(request.donkey.clientId);
@@ -20,4 +23,4 @@ export const GET = withDonkeyAuth(async (request) => {
return NextResponse.json({
data: models,
});
-});
+}, { rateLimit: INFERENCE_RATE_LIMIT });
diff --git a/site/src/app/api/inference/responses/route.ts b/site/src/app/api/inference/responses/route.ts
index e95623d7..5d918741 100644
--- a/site/src/app/api/inference/responses/route.ts
+++ b/site/src/app/api/inference/responses/route.ts
@@ -22,6 +22,7 @@ import {
import { InferenceProviderError } from "@/lib/inference/providers";
import { responseCreateRequestSchema } from "@/lib/inference/schemas";
import {
+ INFERENCE_RATE_LIMIT,
shouldBypassDonkeyInferenceCredits,
withDonkeyAuth,
} from "@/lib/donkey-api-auth";
@@ -244,4 +245,4 @@ export const POST = withDonkeyAuth(async (request) => {
throw error;
}
-}, { allowRunner: true });
+}, { allowRunner: true, rateLimit: INFERENCE_RATE_LIMIT });
diff --git a/site/src/app/api/inference/uploads/route.ts b/site/src/app/api/inference/uploads/route.ts
index 9d98447c..a2ab7d78 100644
--- a/site/src/app/api/inference/uploads/route.ts
+++ b/site/src/app/api/inference/uploads/route.ts
@@ -11,7 +11,10 @@ import {
validationErrorResponse,
} from "@/lib/inference/responses";
import { inferenceUploadRequestSchema } from "@/lib/inference/schemas";
-import { withDonkeyAuth } from "@/lib/donkey-api-auth";
+import {
+ INFERENCE_RATE_LIMIT,
+ withDonkeyAuth,
+} from "@/lib/donkey-api-auth";
// Media a hosted inference call carries rides object storage rather than the
// request body (see lib/inference/blobs.ts). This mints the upload URLs: one
@@ -80,4 +83,4 @@ export const POST = withDonkeyAuth(async (request) => {
}
throw error;
}
-}, { allowRunner: true });
+}, { allowRunner: true, rateLimit: INFERENCE_RATE_LIMIT });
diff --git a/site/src/app/cut/s/[token]/card/[ext]/route.ts b/site/src/app/cut/s/[token]/card/[ext]/route.ts
index af8a0575..2d258da1 100644
--- a/site/src/app/cut/s/[token]/card/[ext]/route.ts
+++ b/site/src/app/cut/s/[token]/card/[ext]/route.ts
@@ -23,6 +23,9 @@ export async function GET(
{ params }: { params: Promise<{ token: string; ext: string }> }
) {
const { token, ext } = await params;
+ // Two card kinds exist. Anything else is a miss, answered before any
+ // lookup or render is spent on it.
+ if (ext !== "gif" && ext !== "jpg") return new Response("Not found.", { status: 404 });
const view = await resolveShare(token, request);
if (view instanceof Response) return view;
const target = { userId: view.share.userId, projectId: view.share.projectId };
@@ -33,10 +36,8 @@ export async function GET(
}
: { "Cache-Control": "private, no-store" };
- if (ext === "gif" || ext === "jpg") {
- const url = await cardUrl(target, ext as CardKind);
- if (url) return new Response(null, { status: 302, headers: { ...headers, Location: url } });
- }
+ const url = await cardUrl(target, ext as CardKind);
+ if (url) return new Response(null, { status: 302, headers: { ...headers, Location: url } });
// No card yet — draw one so the link never unfurls blank. Kept on a short
// leash so the rendered card takes over as soon as it exists.
diff --git a/site/src/lib/donkey-api-auth.ts b/site/src/lib/donkey-api-auth.ts
index 53b3d479..a60106c4 100644
--- a/site/src/lib/donkey-api-auth.ts
+++ b/site/src/lib/donkey-api-auth.ts
@@ -1,3 +1,4 @@
+import { checkInMemoryRateLimit, rateLimitResponse } from "@/lib/inference/rate-limit";
import { createHash, timingSafeEqual } from "node:crypto";
import { type NextRequest, NextResponse } from "next/server";
@@ -26,6 +27,10 @@ export type DonkeyAuthOptions = {
// Accept a `dk_live_` bearer key as the user. Off by default: only routes
// that opt in (the Cut API surface) take machine callers.
allowApiKey?: boolean;
+ // Requests one user may make through this handler per window. The default
+ // fits an editor polling several jobs at once; routes that run a model per
+ // call set a tighter one.
+ rateLimit?: { limit: number; windowMs: number };
// Accept the Cut runner's shared-secret grant. Off by default: only the
// routes the worker actually calls back (the Cut API surface and hosted
// inference) opt in, so a leaked secret stays contained to them.
@@ -79,6 +84,9 @@ export type DonkeyAuthHandler<
const clientIdHeader = "x-donkey-client-id";
const conversationIdHeader = "x-donkey-conversation-id";
const devAuthBypassHeader = "x-donkey-dev-auth-bypass";
+const DEFAULT_RATE_LIMIT = { limit: 900, windowMs: 60_000 };
+/** For handlers that run a model per request. */
+export const INFERENCE_RATE_LIMIT = { limit: 300, windowMs: 60_000 };
const devAuthBypassUserID = "donkey-dev-auth-bypass";
function conversationIdFromHeaders(headers: Headers): string | null {
@@ -197,10 +205,16 @@ function runnerAuthContext(headers: Headers): DonkeyAuthContext | null {
};
}
+// The bypass is a dev-server switch: it has to be turned on in the server's
+// own environment as well as asked for by the caller, so a build that merely
+// forgot NODE_ENV never answers the header. Production refuses it outright.
function devAuthBypassContext(headers: Headers): DonkeyAuthContext | null {
if (process.env.NODE_ENV === "production") {
return null;
}
+ if (process.env.DONKEY_DEV_AUTH_BYPASS !== "1") {
+ return null;
+ }
if (headers.get(devAuthBypassHeader)?.trim() !== "1") {
return null;
}
@@ -246,6 +260,19 @@ export function withDonkeyAuth<
return forbiddenResponse();
}
+ // Every authenticated handler is paced per user. Buckets live in this
+ // process, so the ceiling is per instance; it still turns a hot loop
+ // against one route into 429s instead of a bill.
+ const pace = options.rateLimit ?? DEFAULT_RATE_LIMIT;
+ const limited = checkInMemoryRateLimit({
+ key: `${authContext.userId}:${pace.limit}:${pace.windowMs}`,
+ limit: pace.limit,
+ windowMs: pace.windowMs,
+ });
+ if (!limited.ok) {
+ return rateLimitResponse(limited.retryAfterSeconds);
+ }
+
const authenticatedRequest = Object.assign(request, {
donkey: authContext,
}) as TReq;
diff --git a/site/src/lib/inference/rate-limit.ts b/site/src/lib/inference/rate-limit.ts
index 37a93433..b6a13d98 100644
--- a/site/src/lib/inference/rate-limit.ts
+++ b/site/src/lib/inference/rate-limit.ts
@@ -55,7 +55,7 @@ export function rateLimitResponse(retryAfterSeconds: number) {
return NextResponse.json(
{
error: "rate_limited",
- message: "Too many screenshot parse requests. Please retry shortly.",
+ message: "Too many requests. Please retry shortly.",
},
{
headers: {