fix(mcp): Retry rate-limited GETs using Sentry rate-limit reset headers

README.md

@@ -177,6 +177,15 @@ self-hosted deployment that does run Seer, opt in with `MCP_SKILLS`:
 }
 ```
 
+### Rate Limits
+
+Tool calls are subject to the Sentry API's rate limits. The MCP server retries
+rate-limited requests, and transient Sentry server errors on read requests,
+after a short wait. When a limit lasts longer, the tool returns a **Rate
+Limited** message that says when to try again. Agents hit the per-second
+limits most often by running many Sentry tool calls in parallel, so run them
+one at a time if you see this message.
+
 ### MCP Inspector
 
 MCP includes an [Inspector](https://modelcontextprotocol.io/docs/tools/inspector), to easily test the service:

docs/contributing/api-patterns.md

@@ -73,7 +73,7 @@ fail to resolve for legacy mixed-case project slugs; keep them for display.
 Organization discovery uses `/api/0/organizations/`; a limit over 100 follows
 bounded pages through Sentry's Link cursors. Organization, team, and project
 lists use `@sentry/api` operations to build GET paths and queries. The MCP
-client still owns the request transport, token handling, GET retries, error
+client still owns the request transport, token handling, retries, error
 types, and Zod response validation. Public SaaS hosts use `sentry.io` to list
 organizations across regions. Single-tenant hosts under `*.my.sentry.io` and
 self-hosted instances keep their configured host for this request.
@@ -252,18 +252,35 @@ See error patterns in [common-patterns.md](common-patterns.md#error-handling).
 
 ### Automatic Retries
 
-`SentryApiService.request()` transparently retries transient upstream gateway
-failures (`502`/`503`/`504`) using exponential backoff. Retries are:
-
-- **GET-only** — non-idempotent methods (`POST`/`PUT`/`DELETE`) are never
-  retried, so an automatic retry can never replay a mutation.
-- **Gateway-only** — gated on `ApiServerError.isGatewayError()`. Persistent
-  `500`s, `4xx` client errors, and network/configuration errors fail fast.
+`SentryApiService.request()` transparently retries failed requests using the
+shared policy in `@sentry/toolkit-core/api-retry`. This page describes the
+behavior; the values live only in code, so read them there instead of copying
+them here:
+
+- Which statuses retry, and for which methods: `RETRYABLE_STATUSES` in
+  [`api-retry.ts`](../../packages/toolkit-core/src/api-retry.ts). In short, a
+  rate-limited request retries for any method, a transient server error only
+  for an idempotent one, and other errors never.
+- How many retries and how long they may wait: `REQUEST_RETRY_POLICY` in
+  [`client.ts`](../../packages/mcp-core/src/api-client/client.ts).
+
+Retries are:
+
+- **Rate-limit aware** — a `429` waits for its `Retry-After` header, or for
+  Sentry's `X-Sentry-Rate-Limit-Reset`. Concurrency-limit `429`s and `5xx`
+  errors use jittered exponential backoff.
+- **Capped** — by retry count, single wait, and total wait. A reset further
+  out than the single-wait cap fails at once with an `ApiRateLimitError` whose
+  `retryAfterMs` and user message say when to retry.
+- **Abortable** — aborting the request's `signal` ends the wait.
+- **Fail-fast otherwise** — network and configuration errors are not retried.
+  A rate-limited tool call returns a **Rate Limited** message with the retry
+  time, not an input error.
 - **Bypassed by `allowStatuses`** — a status the caller opts into is returned
   as-is from `requestOnce`, never retried.
 
 Callers and tools need no changes: a retried request either resolves normally or
-throws the same `ApiServerError` it would have thrown without retries.
+throws the same error it would have thrown without retries.
 
 ## Best Practices
 

packages/mcp-core/src/api-client/client.test.ts

@@ -11,6 +11,7 @@ import { SentryApiService } from "./client";
 import {
   ApiNotFoundError,
   ApiPermissionError,
+  ApiRateLimitError,
   ApiServerError,
   ApiValidationError,
 } from "./errors";
@@ -1442,8 +1443,32 @@ describe("transient 5xx retry", () => {
     expect(fetchMock).toHaveBeenCalledTimes(3);
   });
 
-  it("does not retry a non-gateway 500", async () => {
-    const fetchMock = vi.fn().mockResolvedValue(internalError());
+  it("retries a 500 on a GET", async () => {
+    const fetchMock = vi
+      .fn()
+      .mockResolvedValueOnce(internalError())
+      .mockResolvedValueOnce(authOk());
+    globalThis.fetch = fetchMock;
+
+    const apiService = new SentryApiService({
+      host: "sentry.io",
+      accessToken: "test-token",
+    });
+
+    const promise = apiService.getAuthenticatedUser();
+    await vi.runAllTimersAsync();
+
+    expect((await promise).id).toBe("1");
+    expect(fetchMock).toHaveBeenCalledTimes(2);
+  });
+
+  it("does not retry a deterministic 5xx such as 501", async () => {
+    const fetchMock = vi.fn().mockResolvedValue(
+      new Response(JSON.stringify({ detail: "Not Implemented" }), {
+        status: 501,
+        headers: { "Content-Type": "application/json" },
+      }),
+    );
     globalThis.fetch = fetchMock;
 
     const apiService = new SentryApiService({
@@ -1503,6 +1528,192 @@ describe("transient 5xx retry", () => {
   });
 });
 
+describe("rate limit retry", () => {
+  let originalFetch: typeof globalThis.fetch;
+  const now = Date.UTC(2026, 0, 1, 12, 0, 0, 300);
+  const nowSeconds = Math.floor(now / 1000);
+
+  // Shaped like Sentry's rate-limit middleware response for a fixed window.
+  const rateLimited = (resetSeconds: number, headers = {}) =>
+    new Response(
+      JSON.stringify({
+        detail:
+          "You are attempting to use this endpoint too frequently. Limit is 40 requests in 1 seconds",
+      }),
+      {
+        status: 429,
+        statusText: "Too Many Requests",
+        headers: {
+          "Content-Type": "application/json",
+          "X-Sentry-Rate-Limit-Remaining": "0",
+          "X-Sentry-Rate-Limit-Limit": "40",
+          "X-Sentry-Rate-Limit-Reset": String(resetSeconds),
+          ...headers,
+        },
+      },
+    );
+
+  const authOk = () =>
+    new Response(
+      JSON.stringify({ id: "1", name: "Test User", email: "[email protected]" }),
+      { status: 200, headers: { "Content-Type": "application/json" } },
+    );
+
+  const rawRequest = (apiService: SentryApiService) =>
+    (
+      apiService as unknown as {
+        request: (path: string, options?: RequestInit) => Promise<Response>;
+      }
+    ).request.bind(apiService);
+
+  const apiService = () =>
+    new SentryApiService({ host: "sentry.io", accessToken: "test-token" });
+
+  beforeEach(() => {
+    originalFetch = globalThis.fetch;
+    vi.useFakeTimers({ now });
+  });
+
+  afterEach(() => {
+    globalThis.fetch = originalFetch;
+    vi.useRealTimers();
+  });
+
+  it("waits for the rate-limit reset and retries a GET", async () => {
+    const fetchMock = vi
+      .fn()
+      .mockResolvedValueOnce(rateLimited(nowSeconds + 1))
+      .mockResolvedValueOnce(authOk());
+    globalThis.fetch = fetchMock;
+
+    const promise = apiService().getAuthenticatedUser();
+    // The reset is 700ms away; no retry before then.
+    await vi.advanceTimersByTimeAsync(699);
+    expect(fetchMock).toHaveBeenCalledTimes(1);
+    await vi.runAllTimersAsync();
+
+    expect((await promise).id).toBe("1");
+    expect(fetchMock).toHaveBeenCalledTimes(2);
+  });
+
+  it("retries a concurrency-limit 429 with backoff", async () => {
+    const concurrencyLimited = new Response(
+      JSON.stringify({
+        detail:
+          "You are attempting to go above the allowed concurrency for this endpoint. Concurrency limit is 25",
+      }),
+      {
+        status: 429,
+        headers: {
+          "Content-Type": "application/json",
+          "X-Sentry-Rate-Limit-Remaining": "30",
+          "X-Sentry-Rate-Limit-Reset": String(nowSeconds + 1),
+          "X-Sentry-Rate-Limit-ConcurrentRemaining": "0",
+          "X-Sentry-Rate-Limit-ConcurrentLimit": "25",
+        },
+      },
+    );
+    const fetchMock = vi
+      .fn()
+      .mockResolvedValueOnce(concurrencyLimited)
+      .mockResolvedValueOnce(authOk());
+    globalThis.fetch = fetchMock;
+
+    const promise = apiService().getAuthenticatedUser();
+    // The first backoff step is at most 250ms, well before the window reset.
+    await vi.advanceTimersByTimeAsync(250);
+
+    expect((await promise).id).toBe("1");
+    expect(fetchMock).toHaveBeenCalledTimes(2);
+  });
+
+  it("gives up after the retry budget with the delay in the message", async () => {
+    const fetchMock = vi
+      .fn()
+      .mockImplementation(() => rateLimited(Math.floor(Date.now() / 1000) + 1));
+    globalThis.fetch = fetchMock;
+
+    const promise = apiService().getAuthenticatedUser();
+    promise.catch(() => {});
+    await vi.runAllTimersAsync();
+
+    const error = await promise.catch((e: unknown) => e);
+    expect(error).toBeInstanceOf(ApiRateLimitError);
+    expect((error as ApiRateLimitError).toUserMessage()).toMatch(
+      /Retry in 1 second\.$/,
+    );
+    // Initial attempt + 2 retries.
+    expect(fetchMock).toHaveBeenCalledTimes(3);
+  });
+
+  it("does not wait for a reset beyond the per-wait cap", async () => {
+    const fetchMock = vi.fn().mockResolvedValue(rateLimited(nowSeconds + 30));
+    globalThis.fetch = fetchMock;
+
+    const error = await apiService()
+      .getAuthenticatedUser()
+      .catch((e: unknown) => e);
+
+    expect(error).toBeInstanceOf(ApiRateLimitError);
+    expect((error as ApiRateLimitError).retryAfterMs).toBe(29_700);
+    expect((error as ApiRateLimitError).toUserMessage()).toBe(
+      "API error (429): You are attempting to use this endpoint too frequently. Limit is 40 requests in 1 seconds. Retry in 30 seconds.",
+    );
+    expect(fetchMock).toHaveBeenCalledTimes(1);
+  });
+
+  it("honors a proxy's Retry-After over the Sentry reset", async () => {
+    const fetchMock = vi
+      .fn()
+      .mockResolvedValue(rateLimited(nowSeconds + 1, { "Retry-After": "120" }));
+    globalThis.fetch = fetchMock;
+
+    const error = await apiService()
+      .getAuthenticatedUser()
+      .catch((e: unknown) => e);
+
+    expect((error as ApiRateLimitError).retryAfterMs).toBe(120_000);
+    expect(fetchMock).toHaveBeenCalledTimes(1);
+  });
+
+  it("retries a rate-limited POST, which Sentry rejected before running it", async () => {
+    const fetchMock = vi
+      .fn()
+      .mockImplementation(() => rateLimited(Math.floor(Date.now() / 1000) + 1));
+    globalThis.fetch = fetchMock;
+
+    const promise = apiService().createTeam({
+      organizationSlug: "my-org",
+      name: "My Team",
+    });
+    promise.catch(() => {});
+    await vi.runAllTimersAsync();
+
+    await expect(promise).rejects.toBeInstanceOf(ApiRateLimitError);
+    expect(fetchMock).toHaveBeenCalledTimes(3);
+    for (const [, init] of fetchMock.mock.calls) {
+      expect(init.method).toBe("POST");
+    }
+  });
+
+  it("stops waiting when the request is aborted", async () => {
+    const fetchMock = vi.fn().mockResolvedValue(rateLimited(nowSeconds + 1));
+    globalThis.fetch = fetchMock;
+    const controller = new AbortController();
+    const reason = new Error("tool call cancelled");
+
+    const promise = rawRequest(apiService())("/auth/", {
+      signal: controller.signal,
+    });
+    promise.catch(() => {});
+    await vi.advanceTimersByTimeAsync(100);
+    controller.abort(reason);
+
+    await expect(promise).rejects.toBe(reason);
+    expect(fetchMock).toHaveBeenCalledTimes(1);
+  });
+});
+
 describe("request headers", () => {
   let originalFetch: typeof globalThis.fetch;