Release: v0.2.4

apps/api/src/plugins/rate-limit.ts

@@ -2,18 +2,17 @@
  * In-memory rate-limit plugin.
  *
  * Enforces per-IP and per-account caps per specs/api/conventions.md#rate-limiting:
- *   - Unauthenticated reads:  1200 req / min / IP
+ *   - Unauthenticated reads:   300 req / min / IP
  *   - Authenticated reads:     300 req / min / account
- *   - Writes:                   30 req / min / account (120 / min / IP anonymous)
- *   - Credential endpoints:    120 req / min / IP
+ *   - Writes:                   30 req / min / account (30 / min / IP anonymous)
+ *   - Credential endpoints:     20 req / min / IP
  *
  * Only `/api/**` is counted. The SPA shell, its assets, and thumbnails come out
  * of the same process and a single page load fetches dozens of them; counting
  * those against the read cap is what produced a site-wide 429 storm at cutover.
  *
- * Per-IP caps are generous on purpose: production's load balancer does not yet
- * preserve client addresses (cfp-live-cluster #201), so every visitor shares
- * one "IP" until that lands.
+ * Per-IP caps assume the first X-Forwarded-For hop is the real client — the
+ * NodeBalancer speaks PROXY protocol to Envoy (cfp-live-cluster #203).
  *
  * Counters are reset on restart (intentional — single replica, civic scale).
  * Exceeded limit → RateLimitedError(retryAfterSeconds).
@@ -32,11 +31,11 @@ interface BucketEntry {
 const WINDOW_MS = 60_000; // 1 minute
 
 export const RATE_LIMITS = {
-  unauthenticatedReadsPerIp: 1200,
+  unauthenticatedReadsPerIp: 300,
   authenticatedReadsPerAccount: 300,
   writesPerAccount: 30,
-  anonymousWritesPerIp: 120,
-  credentialPerIp: 120,
+  anonymousWritesPerIp: 30,
+  credentialPerIp: 20,
 } as const;
 
 function getOrCreate(map: Map<string, BucketEntry>, key: string): BucketEntry {

plans/rate-limit-scope.md

@@ -53,4 +53,6 @@ cap back down once real client IPs arrive.
 
 ## Follow-ups
 
-- Tracked as: cfp-live-cluster #201 (PROXY protocol so per-IP is per visitor).
+- Tracked as: cfp-live-cluster #203 landed PROXY protocol the same night; the
+  generous caps were then tightened to real per-visitor numbers (reads 300,
+  credential 20, anonymous writes 30 per minute per IP).

specs/api/conventions.md

@@ -161,12 +161,12 @@ Single-replica means rate-limit state is in-memory. Counters reset on restart; a
 
 Only `/api/**` requests are counted. The SPA shell, its assets, and image thumbnails are served by the same process but are never rate limited — a single page load fetches dozens of them.
 
-- Unauthenticated reads: 1200 requests / minute / IP
+- Unauthenticated reads: 300 requests / minute / IP
 - Authenticated reads: 300 requests / minute / account
-- Writes: 30 requests / minute / account (120 / minute / IP when anonymous)
-- Credential endpoints: 120 requests / minute / IP. These are the routes that accept or mint a credential: `POST /api/auth/login`, `/api/auth/github/start` and `/callback`, `/api/auth/link-github`, `/api/auth/password-reset/*`, and `POST /api/account-claim/by-password`. Session reads such as `GET /api/auth/me` (called on every page load) and `/api/auth/refresh` are ordinary reads.
+- Writes: 30 requests / minute / account (30 / minute / IP when anonymous)
+- Credential endpoints: 20 requests / minute / IP. These are the routes that accept or mint a credential: `POST /api/auth/login`, `/api/auth/github/start` and `/callback`, `/api/auth/link-github`, `/api/auth/password-reset/*`, and `POST /api/account-claim/by-password`. Session reads such as `GET /api/auth/me` (called on every page load) and `/api/auth/refresh` are ordinary reads.
 
-The per-IP caps are deliberately generous: in production the gateway sits behind a load balancer that does not yet preserve client addresses, so every visitor currently shares one "IP" (cfp-live-cluster #201). Tighten the credential cap once that lands.
+Per-IP caps assume the gateway delivers the real client address in the first `X-Forwarded-For` hop (the NodeBalancer speaks PROXY protocol to Envoy — cfp-live-cluster #203).
 
 Exceeded → `429 rate_limited`, `Retry-After` header in seconds.