fix(security): reject ':' in tenant ids, count only POSTs for eve rate limit (#5)

* fix(security): reject ':' in tenant ids; count only POSTs for eve rate limit

Harden the per-user key boundary and fix the eve rate-limit double-count,
plus stale-doc cleanup. Folded into the existing tenant-isolation changeset.

Tenant isolation
- AgentMemory/ChatHistory/ToolCache now reject a ':' (the key separator) in
  userId/sessionId/toolName. The whole isolation model rests on the key shape
  `<prefix>:<userId>:<...>`, and only emptiness was validated — a ':' let keys
  collide across users (e.g. user "a"+session "b:c" reached user "a:b"+session
  "c" via the direct-key getChat/saveChat/deleteChat/forget paths).
- Added unit tests for the rejection across all three primitives.

eve rate limiting
- createRateLimitAuth counts only POST requests. eve drives each turn as two
  authenticated requests (the message POST + a follow-up GET .../stream) and
  the auth walk runs on both, so a turn was charged twice. Now one turn = one
  token; session-read GETs fall through unthrottled. Documented in the README
  and docstring; updated tests (a bare Request defaults to GET).

Docs
- Document deriving userId from a verified auth source (Clerk, Auth.js,
  Supabase Auth, Auth0, ...), never a client-supplied value; added DEMO-ONLY
  warnings to the ai-sdk-demo routes that trust an x-user-id header.
- Fixed stale references: search-tools `{@link withIndex}` -> ReactiveSearchIndex;
  ai-sdk memory "scope" -> "userId"; corrected CLAUDE.md's superseded
  namespace key-naming note.

* docs(eve): drop @upstash/agentkit-sdk from install (it's already a dependency)

Author: CahidArda

.changeset/harden-tenant-isolation.md

@@ -8,10 +8,10 @@ Tenant-isolation hardening, a type-safe reactive search index, and a consistent
 
 **Tenant isolation**
 
-- `ChatHistory` is keyed per user (`<prefix>:<userId>:<sessionId>`), so a chat can never be read or overwritten by a different user. Every method takes a single object; `userId`/`sessionId` are required and validated non-empty.
-- `AgentMemory` requires a non-empty `userId` on every call (no silent shared bucket); `add`/`recall` take a single object param.
-- `ToolCache` keys are `<prefix>:<userId>:<toolName>:<hash>` — scoped per user, then per tool.
-- `createRateLimit`/`createRateLimitAuth` require an explicit `limiter` (removed `limit`/`window`); eve's `createRateLimitAuth` requires `identifier` (no implicit global bucket).
+- `ChatHistory` is keyed per user (`<prefix>:<userId>:<sessionId>`), so a chat can never be read or overwritten by a different user. Every method takes a single object; `userId`/`sessionId` are required, validated non-empty, and rejected if they contain the `:` key separator (which would otherwise let keys collide across users).
+- `AgentMemory` requires a non-empty `userId` on every call (no silent shared bucket) and rejects a `:` in `userId`; `add`/`recall` take a single object param.
+- `ToolCache` keys are `<prefix>:<userId>:<toolName>:<hash>` — scoped per user, then per tool; `userId`/`toolName` are rejected if they contain `:`.
+- `createRateLimit`/`createRateLimitAuth` require an explicit `limiter` (removed `limit`/`window`); eve's `createRateLimitAuth` requires `identifier` (no implicit global bucket) and counts only `POST` requests, so a turn (a message `POST` plus its follow-up stream `GET`) is charged once, not twice.
 - The eve sandbox denies network egress by default.
 
 **Reactive search index**

CLAUDE.md

@@ -205,13 +205,17 @@ pnpm -r --filter "./examples/*" build   # build both demo apps
 - Conventional commits; use `!` for breaking changes. Commit at meaningful checkpoints.
 
 ## TODO (current task)
+> **Historical log — superseded naming.** The items below record a completed task and use the
+> intermediate `namespace` name, which was later renamed to **`userId`** (the per-call tenant value)
+> plus **`toolName`** (the cache's tool segment). For the live key naming and conventions, see the
+> "API conventions" section above — not this checklist.
 - [x] Remove model cache (code + examples done; READMEs pending below).
 - [x] ai-sdk: add `cachedTools` (map of `tool()`-built tools, namespace defaults to map key) alongside `cachedTool`; `cachePrefix` → `namespace`; dropped `toolCache` from the config.
 - [x] `cachedTool`/`cachedTools` are fully type-safe (config extends the AI SDK `tool()` type — input/output inference, no `any`).
 - [x] Search tools: ensure the index (create + `waitIndexing`, memoized) before running each tool — a missing Upstash index returns `null`/`-1` rather than throwing, so we ensure up front.
 - [x] `createMemoryTools` (ai-sdk) + eve memory tools: `scope` → `namespace` (string or per-call function). Core `AgentMemory` add/recall/forget use `namespace`.
 - [x] Rate limiting: `namespace` is a plain string; prefix `agentkit:rateLimit`.
-- [x] Key naming: `agentkit:rateLimit:<identifier>`, `agentkit:toolCache:<namespace>:<hash>`, `agentkit:memory:<namespace>:<id>`.
+- [x] Key naming (now `userId`/`toolName`, not `namespace`): `agentkit:rateLimit:<identifier>`, `agentkit:toolCache:<userId>:<toolName>:<hash>`, `agentkit:memory:<userId>:<id>`.
 - [x] Unit/e2e tests use `gpt-4o` (`TEST_MODEL`).
 - [x] eve: dropped the `./model` subpath — model wrappers are exported from the package root.
 - [x] ai-sdk example app fleshed out (memory + search + cached tool + rate limit).

examples/ai-sdk-demo/app/api/chat/route.ts

@@ -27,6 +27,11 @@ export async function POST(req: Request) {
   // as a header. Memory, history, tool cache and rate limit are all scoped to this user; only the
   // shared books index is common to everyone.
   const { id, messages } = (await req.json()) as { id: string; messages: UIMessage[] };
+  // ⚠️ DEMO ONLY: this trusts a client-supplied header to identify the user, so anyone can send
+  // `x-user-id: <someone-else>` and read/overwrite their data. It's only safe here because
+  // `normalizeUser` allow-lists it to two fixed demo users. In production, derive `userId` from a
+  // VERIFIED server-side session (Clerk, Auth.js/NextAuth, Supabase Auth, Auth0, …) — e.g.
+  // `const { userId } = await auth()` with Clerk — and NEVER from a request header/param/body.
   const userId = normalizeUser(req.headers.get(USER_HEADER));
   const redis = getRedis();
 

examples/ai-sdk-demo/app/api/chats/route.ts

@@ -5,7 +5,9 @@ export const runtime = "nodejs";
 
 // GET /api/chats          → list the user's chats (summaries, newest first)
 // GET /api/chats?q=<text>  → fuzzy-search the user's chats by what was said
-// The active user is identified by the `x-user-id` header, so each user sees only their own chats.
+// ⚠️ DEMO ONLY: the user is identified by the client-supplied `x-user-id` header (allow-listed to two
+// fixed demo users by `normalizeUser`). In production, derive `userId` from a VERIFIED server-side
+// session (Clerk, Auth.js/NextAuth, Supabase Auth, Auth0, …), never from a client-supplied value.
 export async function GET(req: Request) {
   const history = getHistory();
   const userId = normalizeUser(req.headers.get(USER_HEADER));

packages/ai-sdk/README.md

@@ -28,9 +28,12 @@ const history = createChatHistory({
 });
 ```
 
-Every method takes a single object; `userId` is **required, non-empty, and must be unique per user**
-(your auth subject id). It's the tenant boundary — a chat can't be read or overwritten under a
-different `userId`. `saveChat` overwrites the **whole** message array — `useChat` sends the full
+Every method takes a single object; `userId` is **required, non-empty, and may not contain `:`**. It's
+the tenant boundary, so **derive it from a verified server-side auth source** — the subject/user id
+from your auth provider (Clerk, Auth.js/NextAuth, Supabase Auth, Auth0, …) — and **never from a
+client-supplied header, query param, or body** (e.g. read it from the session in your route, not from
+the request the browser controls). A chat can't be read or overwritten under a different `userId`.
+`saveChat` overwrites the **whole** message array — `useChat` sends the full
 conversation, so there's no transport trimming and no delta to merge. Persist from your route's
 `onFinish`:
 
@@ -85,9 +88,10 @@ const tools = createMemoryTools({
 await generateText({ model, tools, stopWhen: stepCountIs(5), prompt: "What do you know about me?" });
 ```
 
-> **`userId` is required and must be non-empty** — it's the only tenant boundary for memory.
-> Make it **unique per user** (pass the user id, or a `(input, options) => string` deriving it); an
-> empty value throws rather than collapsing every user into one shared bucket.
+> **`userId` is required, non-empty, and may not contain `:`** — it's the only tenant boundary for
+> memory. **Derive it from a verified server-side auth source** (the subject/user id from Clerk,
+> Auth.js/NextAuth, Supabase Auth, Auth0, …), passed as a string or a `(input, options) => string`;
+> **never trust a client-supplied value.** An empty/separator-bearing value throws.
 
 ## Search tools
 

packages/ai-sdk/src/memory.ts

@@ -29,7 +29,7 @@ export interface CreateMemoryToolsConfig {
 
 /**
  * Build `recall_memory` and `save_memory` AI SDK tools backed by long-term {@link AgentMemory}. Spread
- * the returned map into `generateText({ tools })`. Pass only a `scope`; `redis` defaults to
+ * the returned map into `generateText({ tools })`. Pass only a `userId`; `redis` defaults to
  * `Redis.fromEnv()`.
  *
  * ```ts

packages/eve/README.md

@@ -7,7 +7,7 @@ code-execution **sandbox backend** powered by [Upstash Box](https://github.com/u
 cached tools.
 
 ```bash
-pnpm add @upstash/agentkit-eve @upstash/agentkit-sdk @upstash/redis
+pnpm add @upstash/agentkit-eve @upstash/redis
 # in your app (Eve + the OpenAI provider, plus Box only if you use /sandbox):
 pnpm add eve @ai-sdk/openai @upstash/box
 ```
@@ -47,6 +47,12 @@ export default defineMemorySaveTool({
 });
 ```
 
+> **`userId` is the tenant boundary** (required, non-empty, no `:`). Derive it from eve's **verified
+> session auth** — `ctx.session.auth.current?.principalId` (the authenticated principal, gated by your
+> channel's `auth` walk) — as above, not from anything the client supplies. Configure a real
+> authenticator (`vercelOidc()`, an OIDC/JWT provider like Clerk, …) so `principalId` is trustworthy;
+> the `?? ctx.session.id` fallback only applies when a request is unauthenticated.
+
 ## Search tools (`agent/tools/*.ts`)
 
 `defineSearchTools` builds `search` / `aggregate` / `count` eve tools over an Upstash Redis Search
@@ -107,6 +113,12 @@ export default eveChannel({
 > one abusive caller can exhaust the window for everyone, so for per-user limiting derive it per
 > request (an authenticated user id, an API key, or `x-forwarded-for` for per-IP).
 
+> **Only `POST` requests are counted.** eve runs each turn as two authenticated requests — the message
+> `POST` (which invokes the model) and a follow-up `GET …/stream` that opens the reply stream — and the
+> `auth` walk runs on both. `createRateLimitAuth` throttles only the `POST`s, so **one turn costs one
+> token** of your limiter (a `Ratelimit.slidingWindow(20, "1 m")` allows 20 turns/min, not 10); the
+> session-read `GET`s fall through unthrottled.
+
 ## Code-execution sandbox (`agent/sandbox.ts`)
 
 `upstash()` is a drop-in replacement for Eve's `vercel()` backend, powered by Upstash Box. Swap the

packages/eve/src/auth.test.ts

@@ -4,7 +4,8 @@ import { ForbiddenError } from "eve/channels/auth";
 import { Ratelimit, createRateLimitAuth } from "./index.js";
 import { hasRedisCreds, testRedis } from "./test-support.js";
 
-const req = () => new Request("http://localhost/agent");
+// Only POST requests (the model-invoking message submissions) are counted, so default to POST here.
+const req = () => new Request("http://localhost/agent", { method: "POST" });
 
 describe.skipIf(!hasRedisCreds)("createRateLimitAuth (live Redis)", () => {
   const redis = testRedis();
@@ -35,11 +36,27 @@ describe.skipIf(!hasRedisCreds)("createRateLimitAuth (live Redis)", () => {
     });
 
     const reqFor = (user: string) =>
-      new Request("http://localhost/agent", { headers: { "x-user": user } });
+      new Request("http://localhost/agent", { method: "POST", headers: { "x-user": user } });
 
     // Exhaust A's bucket; B's is untouched and still passes.
     expect(await auth(reqFor("a"))).toBeNull();
     await expect(auth(reqFor("a"))).rejects.toBeInstanceOf(ForbiddenError);
     expect(await auth(reqFor("b"))).toBeNull();
   });
+
+  // eve runs the auth walk on the follow-up `GET …/stream` too; only POSTs (which invoke the model)
+  // are counted, so a turn isn't charged twice. A GET always falls through without touching the limiter.
+  it("does not count non-POST requests (e.g. the stream GET)", async () => {
+    const auth = createRateLimitAuth({
+      redis,
+      limiter: Ratelimit.slidingWindow(1, "60 s"),
+      identifier: `test:${randomUUID().slice(0, 8)}`,
+    });
+    const get = () => new Request("http://localhost/agent", { method: "GET" });
+
+    // Many GETs in a row never exhaust the (size-1) window — they're not counted at all.
+    expect(await auth(get())).toBeNull();
+    expect(await auth(get())).toBeNull();
+    expect(await auth(get())).toBeNull();
+  });
 });

packages/eve/src/auth.ts

@@ -24,6 +24,12 @@ export interface RateLimitAuthConfig extends Omit<RateLimitConfig, "redis"> {
  * Over the limit it throws a `ForbiddenError` (HTTP 403). Backed by {@link createRateLimit}; keys are
  * `agentkit:rateLimit:<identifier>`.
  *
+ * **Only `POST` requests are counted** (the message-submitting routes that actually invoke the model:
+ * `POST /eve/v1/session` and `POST /eve/v1/session/:id`). eve drives each turn as two authenticated
+ * requests — the message `POST` **and** a follow-up `GET …/stream` that opens the reply stream — and
+ * the auth walk runs on both. Counting both would charge every turn twice; gating only the `POST`
+ * makes one turn cost exactly one token, while the session-read `GET`s fall through unthrottled.
+ *
  * ```ts
  * // agent/channels/eve.ts
  * import { createRateLimitAuth, Ratelimit } from "@upstash/agentkit-eve";
@@ -47,6 +53,10 @@ export function createRateLimitAuth(config: RateLimitAuthConfig): AuthFn<Request
   const ratelimit = createRateLimit({ ...rest, redis: redis ?? Redis.fromEnv() });
 
   return async (request) => {
+    // Only throttle the model-invoking message submissions (POST). The follow-up `GET …/stream` (and
+    // other session reads) share this auth walk but shouldn't each cost a token — let them through so
+    // a single turn = a single increment.
+    if (request.method !== "POST") return null;
     const id = typeof identifier === "function" ? await identifier(request) : identifier;
     const { success } = await ratelimit.limit(id);
     if (!success) {

packages/sdk/README.md

@@ -81,9 +81,12 @@ const hits = await history.searchChats({
 await history.deleteChat({ userId: "user-123", sessionId: "session-abc" }); // delete (also de-indexes it)
 ```
 
-> **`userId` and `sessionId` are required and must be non-empty** — they are the only tenant
-> boundary, so an empty value throws rather than silently mis-scoping a chat. Make `userId` unique
-> per user (e.g. your auth subject id); a chat can't be read or overwritten by a different `userId`.
+> **`userId` and `sessionId` are required, non-empty, and may not contain `:`** (the key separator) —
+> they are the only tenant boundary, so an empty or separator-bearing value throws rather than
+> silently mis-scoping a chat. **Derive `userId` from a verified server-side auth source** — the
+> subject/user id from your auth provider (Clerk, Auth.js/NextAuth, Supabase Auth, Auth0, …) — and
+> **never from a client-supplied header, query param, or request body**, or a caller can impersonate
+> any user. A chat can't be read or overwritten under a different `userId`.
 
 ### Agent memory
 
@@ -113,9 +116,11 @@ const hits = await memory.recall({
 await memory.forget("pref-lang", { userId: "user-123" }); // required, non-empty userId
 ```
 
-> **`userId` is required and must be non-empty** on every method — it's the only tenant boundary
-> for memory, so an empty value throws rather than collapsing all callers into one shared bucket.
-> Make it **unique per user** (e.g. the user id) to keep each user's memories isolated.
+> **`userId` is required, non-empty, and may not contain `:`** on every method — it's the only tenant
+> boundary for memory, so an empty or separator-bearing value throws rather than collapsing or
+> colliding callers. **Derive it from a verified server-side auth source** (the subject/user id from
+> Clerk, Auth.js/NextAuth, Supabase Auth, Auth0, …) — **never from a client-supplied value** — to keep
+> each user's memories isolated.
 
 ### Search tools
 
@@ -185,9 +190,10 @@ const getWeather = tools.wrap(
 );
 ```
 
-> **`userId` and `toolName` are both required and must be non-empty** (`get`/`set`/`invalidate`/`wrap`
-> all throw on an empty value). The entry is scoped to the user first, so one user's cached result is
-> never served to another.
+> **`userId` and `toolName` are both required, non-empty, and may not contain `:`** (`get`/`set`/
+> `invalidate`/`wrap` all throw otherwise). The entry is scoped to the user first, so one user's cached
+> result is never served to another — provided `userId` comes from a verified auth source, not a
+> client-supplied value.
 
 ## Testing