fix: harden Service Worker activation and install lifecycle

SW-06: correlate activation, reset and drain replies by source object identity
against the captured waiting worker or controller, so an arbitrary same-origin
source cannot close this page's admission, and end a request immediately as
PROTOCOL_MISMATCH when the source is swapped instead of waiting for the drain
timeout. requestActivation() and resetOwnedCaches() are single-flight, so ten
concurrent callers share one nonce, listener and postMessage.

SW-07: an empty in-scope client set is vacuously drained rather than rejecting
a waiting worker when the requester already closed.

SW-08: isolate per-client postMessage failures. A client that cannot receive the
drain request fails immediately instead of holding pending state to the timeout,
skipWaiting() is the activation commit and its failure is a rejection, and the
accepted and reload notifications are sent afterwards as best effort.

SW-09: fence late install work. A fenced worker starts no new candidate work, a
late response body from a non-cooperative fetch is cancelled, a throwing digest
maps to a closed outcome, and a second exact delete of the owned candidate cache
is registered once the abandoned install settles - without extending the public
60s bound.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

docs/operations/adapter-remediation-ledger.md

@@ -144,10 +144,10 @@ Rollout state starts at `NOT_STARTED`; documented-unimplemented items start at
 | SW-03 | Composed when capability selected | `corepack pnpm exec vitest run tests/unit/service-worker-runtime.test.ts` | `fix: make Service Worker cache and removal outcomes truthful` | `FIXED_NOT_RELEASED` | false removal success | Red `unregister() === false` reported as UNREGISTERED → green FAILED |
 | SW-04 | Composed when capability selected | `corepack pnpm exec vitest run tests/unit/service-worker-runtime.test.ts` | `fix: make Service Worker cache and removal outcomes truthful` | `FIXED_NOT_RELEASED` | removal outcome misreport | Red removal modes always DISABLED → green outcome matrix (ABSENT/UNREGISTERED/PURGED→DISABLED, OWNERSHIP_MISMATCH→INCOMPATIBLE, FAILED→FAILED) |
 | SW-05 | Build gate | `corepack pnpm exec vitest run tests/unit/service-worker-build-input.test.ts` | `fix: make Service Worker cache and removal outcomes truthful` | `FIXED_NOT_RELEASED` | build admission rejection | Red tamper table (stale digest, byte length, cross-origin URL, dot segment, extension mismatch, unknown field, duplicate URL) → green; build gate decodes through the shared codec and recomputes the canonical digest |
-| SW-06 | Composed when capability selected | `corepack pnpm exec vitest run tests/unit/service-worker-runtime.test.ts` | — | `NOT_STARTED` | activation handshake failure | — |
-| SW-07 | Composed when capability selected | `corepack pnpm exec vitest run tests/unit/service-worker-runtime.test.ts` | — | `NOT_STARTED` | activation blocked with zero clients | — |
-| SW-08 | Composed when capability selected | `corepack pnpm exec vitest run tests/unit/service-worker-runtime.test.ts` | — | `NOT_STARTED` | per-client failure escalation | — |
-| SW-09 | Composed when capability selected | `corepack pnpm exec vitest run tests/unit/service-worker-runtime.test.ts` | — | `NOT_STARTED` | late install work observed | — |
+| SW-06 | Composed when capability selected | `corepack pnpm exec vitest run tests/unit/service-worker-runtime.test.ts` | `fix: harden Service Worker activation and install lifecycle` | `FIXED_NOT_RELEASED` | activation handshake failure | Red foreign-source drain, source swap and 10 concurrent activations → green 24/24; replies correlate by source object identity against the captured waiting worker or controller, and activation/reset are single-flight |
+| SW-07 | Composed when capability selected | `corepack pnpm exec vitest run tests/unit/service-worker-runtime.test.ts` | `fix: harden Service Worker activation and install lifecycle` | `FIXED_NOT_RELEASED` | activation blocked with zero clients | An empty in-scope client set is vacuously drained; `clients.matchAll()` failure still rejects |
+| SW-08 | Composed when capability selected | `corepack pnpm exec vitest run tests/unit/service-worker-runtime.test.ts` | `fix: harden Service Worker activation and install lifecycle` | `FIXED_NOT_RELEASED` | per-client failure escalation | Per-client `postMessage` isolation; `skipWaiting()` is the commit point and its failure is REJECTED, with accepted/reload notifications sent only afterwards as best effort |
+| SW-09 | Composed when capability selected | `corepack pnpm exec vitest run tests/unit/service-worker-runtime.test.ts` | `fix: harden Service Worker activation and install lifecycle` | `FIXED_NOT_RELEASED` | late install work observed | Red late-fetch case → green; a fenced worker starts no new candidate work, late response bodies are cancelled, digest throws map to a closed outcome, and a second exact-delete runs once the abandoned install settles without extending the public bound |
 | SW-10 | Protocol V2 migration | `corepack pnpm exec vitest run tests/unit/service-worker-runtime.test.ts` | — | `DEFERRED_TO_MIGRATION` | V1/V2 mismatch fail-close | Not closed here. Full-identity protocol V2 is an expand → dual-read → old-writer drain → contract deployment that spans releases; the prerequisite shared manifest codec and canonical digest landed with SW-05. |
 | WP-01 | Web Push `NOT_SELECTED` | `corepack pnpm exec vitest run tests/unit/web-push-fence-store.test.ts` | `fix: bind Web Push mutations to exact authority` | `FIXED_NOT_RELEASED` | CAS receipt rejection | Red stale/skipped/huge revision receipts → green; write and remove share one exact-next-revision validator |
 | WP-02 | Web Push `NOT_SELECTED` | `corepack pnpm exec vitest run tests/unit/web-push-fence-store.test.ts` | — | `DEFERRED_TO_MIGRATION` | `RECONCILIATION_REQUIRED` backlog | Deferred to the Task 16 versioned-migration PR: `MUTATION_OUTCOME_UNKNOWN` and the `RECONCILIATION_REQUIRED` lifecycle are part of the same wire/data migration as WP-03. |

src/adapters/service-worker/service-worker-lifecycle.ts

@@ -277,49 +277,64 @@ export function createServiceWorkerRuntime(
       parsed.message.sourceBuildId,
     );
     if (!drained) {
-      for (const client of clients) {
-        client.postMessage(
-          createServiceWorkerMessage({
-            kind: "ACTIVATE_REJECTED",
-            sourceBuildId: config.identity.buildId,
-            targetBuildId: parsed.message.sourceBuildId,
-            nonce,
-          }),
-        );
-      }
+      notifyClients(clients, "ACTIVATE_REJECTED", parsed.message.sourceBuildId, nonce);
       return "REJECTED";
     }
 
-    for (const client of clients) {
-      client.postMessage(
-        createServiceWorkerMessage({
-          kind: "ACTIVATE_ACCEPTED",
-          sourceBuildId: config.identity.buildId,
-          targetBuildId: parsed.message.sourceBuildId,
-          nonce,
-        }),
-      );
-    }
-    await scope.skipWaiting();
-    for (const client of clients) {
-      client.postMessage(
-        createServiceWorkerMessage({
-          kind: "ACTIVATED_RELOAD_REQUIRED",
-          sourceBuildId: config.identity.buildId,
-          targetBuildId: parsed.message.sourceBuildId,
-          nonce,
-        }),
-      );
+    // SW-08. `skipWaiting()` is the activation commit. It must succeed before
+    // any client is told the activation was accepted, and its failure is a
+    // rejection rather than an accepted-then-failed activation.
+    try {
+      await scope.skipWaiting();
+    } catch {
+      notifyClients(clients, "ACTIVATE_REJECTED", parsed.message.sourceBuildId, nonce);
+      return "REJECTED";
     }
+    // Post-commit notifications are per-client best effort.
+    notifyClients(clients, "ACTIVATE_ACCEPTED", parsed.message.sourceBuildId, nonce);
+    notifyClients(
+      clients,
+      "ACTIVATED_RELOAD_REQUIRED",
+      parsed.message.sourceBuildId,
+      nonce,
+    );
     return "ACCEPTED";
   }
 
+  /**
+   * SW-08. One client's `postMessage()` throwing must not break the whole
+   * activation event; delivery is isolated per client.
+   */
+  function notifyClients(
+    clients: readonly WorkerClientLike[],
+    kind: "ACTIVATE_REJECTED" | "ACTIVATE_ACCEPTED" | "ACTIVATED_RELOAD_REQUIRED",
+    targetBuildId: string,
+    nonce: string,
+  ): void {
+    for (const client of clients) {
+      try {
+        client.postMessage(
+          createServiceWorkerMessage({
+            kind,
+            sourceBuildId: config.identity.buildId,
+            targetBuildId,
+            nonce,
+          }),
+        );
+      } catch {
+        // A dead client cannot change the already committed activation.
+      }
+    }
+  }
+
   async function drainClients(
     clients: readonly WorkerClientLike[],
     nonce: string,
     requesterBuildId: string,
   ): Promise<boolean> {
-    if (clients.length === 0) return false;
+    // SW-07. No in-scope client means nothing dirty to drain, so the set is
+    // vacuously drained. A `clients.matchAll()` failure still rejects upstream.
+    if (clients.length === 0) return true;
     const drained = new Promise<boolean>((resolve) => {
       const timer = setTimeout(() => {
         pendingActivations.delete(nonce);
@@ -336,15 +351,24 @@ export function createServiceWorkerRuntime(
         }),
       );
     });
+    // SW-08. A client that cannot receive the drain request can never
+    // acknowledge it, so it fails immediately instead of holding the pending
+    // state until the timeout.
     for (const client of clients) {
-      client.postMessage(
-        createServiceWorkerMessage({
-          kind: "CLIENT_DRAIN_REQUEST",
-          sourceBuildId: config.identity.buildId,
-          targetBuildId: requesterBuildId,
-          nonce,
-        }),
-      );
+      try {
+        client.postMessage(
+          createServiceWorkerMessage({
+            kind: "CLIENT_DRAIN_REQUEST",
+            sourceBuildId: config.identity.buildId,
+            targetBuildId: requesterBuildId,
+            nonce,
+          }),
+        );
+      } catch {
+        const pending = pendingActivations.get(nonce);
+        if (pending) settlePendingActivation(nonce, pending, false);
+        return await drained;
+      }
     }
     return drained;
   }

src/adapters/service-worker/service-worker-page-controller.ts

@@ -58,6 +58,9 @@ export function createServiceWorkerPageController(
   let registration: ServiceWorkerRegistration | null = null;
   let messageListener: ((event: MessageEvent) => void) | null = null;
   let updateTimer: ReturnType<typeof setInterval> | null = null;
+  /** SW-06. Single-flight command state. */
+  let activationInFlight: Promise<ServiceWorkerActivationOutcome> | null = null;
+  let resetInFlight: Promise<ServiceWorkerResetOutcome> | null = null;
   let stopped = false;
   const pendingStops = new Set<() => void>();
 
@@ -217,6 +220,13 @@ export function createServiceWorkerPageController(
           observe("client_drain", "MALFORMED");
           return;
         }
+        // SW-06. An arbitrary same-origin source must not be able to close this
+        // page's admission. The request has to come from the worker we are
+        // actually waiting on or the one currently controlling us.
+        if (!isExpectedWorkerSource(source)) {
+          observe("client_drain", "SOURCE_MISMATCH");
+          return;
+        }
         const rejected = isBlocked();
         source.postMessage(
           createServiceWorkerMessage({
@@ -234,6 +244,23 @@ export function createServiceWorkerPageController(
     container.addEventListener("message", messageListener);
   }
 
+  /**
+   * SW-06. Source identity is checked by object identity against the
+   * registration's waiting/installing/active worker and the container's
+   * controller. An empty `event.origin` is never used as a trust signal.
+   */
+  function isExpectedWorkerSource(source: unknown): boolean {
+    const expected = [
+      registration?.waiting,
+      registration?.installing,
+      registration?.active,
+      dependencies.container?.controller,
+    ];
+    return expected.some(
+      (candidate) => candidate != null && candidate === source,
+    );
+  }
+
   function scheduleUpdateChecks(): void {
     // §17.14. At most one check per 6 hours, and none while the page is hidden.
     if (updateTimer) return;
@@ -251,7 +278,16 @@ export function createServiceWorkerPageController(
    * §17.11. Activation is a handshake: every controlled client must close new
    * admission and acknowledge within 30s. One missing client rejects it.
    */
-  async function requestActivation(): Promise<ServiceWorkerActivationOutcome> {
+  function requestActivation(): Promise<ServiceWorkerActivationOutcome> {
+    // SW-06. Concurrent callers share one command: a second call must not issue
+    // a second nonce, a second listener or a second postMessage.
+    activationInFlight ??= runActivation().finally(() => {
+      activationInFlight = null;
+    });
+    return activationInFlight;
+  }
+
+  async function runActivation(): Promise<ServiceWorkerActivationOutcome> {
     const waiting = registration?.waiting;
     if (!waiting) return Object.freeze({ kind: "NO_WAITING_WORKER" as const });
     if (isBlocked()) {
@@ -287,6 +323,17 @@ export function createServiceWorkerPageController(
           ) {
             return;
           }
+          // SW-06. The reply must come from the exact worker this request was
+          // sent to. A source swap ends the request immediately as a protocol
+          // mismatch rather than waiting for the drain timeout.
+          if (event.source !== null && event.source !== waiting) {
+            finish(Object.freeze({ kind: "PROTOCOL_MISMATCH" as const }));
+            return;
+          }
+          if (registration?.waiting !== waiting) {
+            finish(Object.freeze({ kind: "PROTOCOL_MISMATCH" as const }));
+            return;
+          }
           if (
             parsed.message.kind === "ACTIVATE_REJECTED" &&
             parsed.message.nonce === nonce
@@ -329,11 +376,21 @@ export function createServiceWorkerPageController(
   }
 
   /** §18.10. Static caches only; the registration itself is left in place. */
-  async function resetOwnedCaches(): Promise<ServiceWorkerResetOutcome> {
+  function resetOwnedCaches(): Promise<ServiceWorkerResetOutcome> {
+    // SW-06. Single-flight, like activation.
+    resetInFlight ??= runReset().finally(() => {
+      resetInFlight = null;
+    });
+    return resetInFlight;
+  }
+
+  async function runReset(): Promise<ServiceWorkerResetOutcome> {
     const container = dependencies.container;
     if (!container?.controller) {
       return Object.freeze({ kind: "NOT_CONTROLLED" as const });
     }
+    // SW-06. The reply must come from the controller this request was sent to.
+    const requestedController = container.controller;
     const nonce = nonces.issue();
     return new Promise<ServiceWorkerResetOutcome>((resolve) => {
       let settled = false;
@@ -357,6 +414,18 @@ export function createServiceWorkerPageController(
         ) {
           return;
         }
+        if (
+          (event.source !== null && event.source !== requestedController) ||
+          container.controller !== requestedController
+        ) {
+          finish(
+            Object.freeze({
+              kind: "FAILED" as const,
+              code: "PROTOCOL_MISMATCH",
+            }),
+          );
+          return;
+        }
         finish(
           Object.freeze({
             kind: "RESET" as const,

src/adapters/service-worker/service-worker-static-assets.ts

@@ -149,6 +149,19 @@ export async function installStaticAssets(
 
   if (outcome.kind === "REJECTED") {
     await dependencies.caches.delete(cacheName).catch(() => false);
+    // SW-09. A non-cooperative fetch, digest or `cache.put` started before the
+    // deadline cannot be cancelled, so it may recreate the candidate cache
+    // after that delete. The public result already closed at the deadline; a
+    // second exact delete is registered once the abandoned work settles. It is
+    // deliberately not awaited, so the public bound is not extended.
+    if (deadlineExceeded) {
+      void installation
+        .catch(() => undefined)
+        .then(async () => {
+          await dependencies.caches.delete(cacheName).catch(() => false);
+        })
+        .catch(() => undefined);
+    }
   }
   return outcome;
 }
@@ -173,6 +186,11 @@ async function installCandidate(
   const worker = async (): Promise<void> => {
     for (;;) {
       if (failure) return;
+      // SW-09. Once fenced, no new candidate work is started.
+      if (signal.aborted) {
+        failure ??= rejected("INSTALL_DEADLINE_EXCEEDED");
+        return;
+      }
       const asset = queue.shift();
       if (!asset) return;
       const outcome = await storeAsset(asset, cache, dependencies, signal);
@@ -205,15 +223,22 @@ async function storeAsset(
   if (signal.aborted) return rejected("FETCH_FAILED");
   let response: Response;
   try {
-    const fetched = await abortable(
-      dependencies.fetcher(asset.url, {
-        cache: "no-store",
-        credentials: "omit",
-        redirect: "error",
-        signal,
-      }),
+    // SW-09. A non-cooperative fetch that ignores the signal still settles
+    // later; its body is compensated so an abandoned response is not left open.
+    const pending = dependencies.fetcher(asset.url, {
+      cache: "no-store",
+      credentials: "omit",
+      redirect: "error",
       signal,
-    );
+    });
+    const fetched = await abortable(pending, signal);
+    if (fetched === ABORTED) {
+      void pending
+        .then(async (late) => {
+          await late.body?.cancel();
+        })
+        .catch(() => undefined);
+    }
     if (fetched === ABORTED) return rejected("FETCH_FAILED");
     response = fetched;
   } catch {
@@ -234,7 +259,17 @@ async function storeAsset(
   if (!body.ok) return rejected(body.code);
   const bytes = body.bytes;
 
-  const digest = await abortable(dependencies.digest(bytes), signal);
+  // SW-09. A digest dependency that throws becomes a closed typed outcome
+  // rather than an escaping rejection.
+  let digest: string | typeof ABORTED;
+  try {
+    digest = await abortable(
+      Promise.resolve(dependencies.digest(bytes)),
+      signal,
+    );
+  } catch {
+    return rejected("INTEGRITY_MISMATCH");
+  }
   if (digest === ABORTED) return rejected("FETCH_FAILED");
   if (digest !== asset.sha256) return rejected("INTEGRITY_MISMATCH");
 

tests/unit/service-worker-runtime.test.ts

@@ -49,6 +49,8 @@ function pageContainer(options: { waiting?: boolean; controlled?: boolean } = {}
   } as unknown as ServiceWorkerContainer;
   return {
     container,
+    waiting,
+    controlled,
     registration,
     waitingMessages,
     controllerMessages,
@@ -384,7 +386,12 @@ describe("service worker page protocol", () => {
     const browser = pageContainer({ waiting: true });
     const controller = pageController(browser.container);
     await controller.start();
-    const source = { postMessage: vi.fn() };
+    // SW-06. Replies are correlated by source identity, so the fake request
+    // comes from the registration's waiting worker.
+    const source = browser.waiting as unknown as {
+      postMessage(message: unknown): void;
+    };
+    const sourceMessages = vi.spyOn(source, "postMessage");
 
     browser.dispatch(
       createServiceWorkerMessage({
@@ -396,7 +403,7 @@ describe("service worker page protocol", () => {
       source,
     );
 
-    expect(source.postMessage).toHaveBeenCalledWith(
+    expect(sourceMessages).toHaveBeenCalledWith(
       expect.objectContaining({
         kind: "CLIENT_DRAINED",
         sourceBuildId: "page-build",
@@ -444,7 +451,7 @@ describe("service worker page protocol", () => {
         nonce: request.nonce,
       }),
       cachesDeleted: 2,
-    });
+    }, browser.controlled ?? undefined);
 
     await expect(result).resolves.toEqual({ kind: "RESET", cachesDeleted: 2 });
     await controller.stop();
@@ -700,7 +707,7 @@ describe("service worker static asset install", () => {
 
   it("aborts and rolls back a candidate cache at the overall install deadline", async () => {
     vi.useFakeTimers();
-    const deleteCache = vi.fn(async () => true);
+    const deleteCache = vi.fn(async (_name: string) => true);
     const fetcher = vi.fn(
       (_input: RequestInfo | URL, init?: RequestInit) =>
         new Promise<Response>((_resolve, reject) => {
@@ -725,7 +732,47 @@ describe("service worker static asset install", () => {
       code: "INSTALL_DEADLINE_EXCEEDED",
     });
     expect(fetcher.mock.calls[0]?.[1]?.signal?.aborted).toBe(true);
-    expect(deleteCache).toHaveBeenCalledTimes(1);
+    // SW-09. The public result closes at the deadline with one exact delete,
+    // and a second exact delete is registered once the abandoned install work
+    // actually settles. Both target the same owned candidate cache.
+    await vi.advanceTimersByTimeAsync(0);
+    expect(deleteCache).toHaveBeenCalledTimes(2);
+    expect(new Set(deleteCache.mock.calls.map((call) => call[0])).size).toBe(1);
+    vi.useRealTimers();
+  });
+
+  it("observes and cleans non-cooperative late install work", async () => {
+    vi.useFakeTimers();
+    const deleteCache = vi.fn(async (_name: string) => true);
+    const cancel = vi.fn(async () => {});
+    let releaseFetch: ((response: Response) => void) | undefined;
+    // A fetch that ignores the abort signal entirely.
+    const fetcher = vi.fn(
+      () =>
+        new Promise<Response>((resolve) => {
+          releaseFetch = resolve;
+        }),
+    );
+
+    const result = installStaticAssets(manifest, {
+      caches: {
+        open: vi.fn(async () => ({ put: vi.fn() }) as unknown as Cache),
+        delete: deleteCache,
+      },
+      fetcher: fetcher as typeof fetch,
+      digest: vi.fn(),
+    });
+    await vi.advanceTimersByTimeAsync(SERVICE_WORKER_BOUNDS.installDeadlineMs);
+    await expect(result).resolves.toEqual({
+      kind: "REJECTED",
+      code: "INSTALL_DEADLINE_EXCEEDED",
+    });
+
+    // The late response arrives after the public bound; its body is cancelled
+    // and no unhandled rejection escapes.
+    releaseFetch?.({ body: { cancel } } as unknown as Response);
+    await vi.advanceTimersByTimeAsync(0);
+    expect(cancel).toHaveBeenCalledOnce();
     vi.useRealTimers();
   });
 
@@ -807,4 +854,65 @@ describe("service worker static asset install", () => {
       vi.useRealTimers();
     }
   });
+
+  it("accepts replies only from the captured waiting or controller source", async () => {
+    const browser = pageContainer({ waiting: true });
+    const controller = pageController(browser.container);
+    await controller.start();
+    const foreign = { postMessage: vi.fn() };
+
+    // SW-06. A same-origin but unrecognised source must not close admission.
+    browser.dispatch(
+      createServiceWorkerMessage({
+        kind: "CLIENT_DRAIN_REQUEST",
+        sourceBuildId: "worker-build",
+        targetBuildId: "page-build",
+        nonce: "drain-foreign",
+      }),
+      foreign,
+    );
+    expect(foreign.postMessage).not.toHaveBeenCalled();
+    await controller.stop();
+  });
+
+  it("coalesces concurrent activation and reset commands", async () => {
+    const browser = pageContainer({ waiting: true });
+    const controller = pageController(browser.container);
+    await controller.start();
+
+    // SW-06. Ten concurrent callers issue exactly one request.
+    const activations = Array.from({ length: 10 }, () =>
+      controller.requestActivation(),
+    );
+    await Promise.resolve();
+    expect(browser.waitingMessages).toHaveLength(1);
+    expect(new Set(activations).size).toBe(1);
+
+    await controller.stop();
+    await Promise.allSettled(activations);
+  });
+
+  it("ends an activation whose reply source was swapped", async () => {
+    const browser = pageContainer({ waiting: true });
+    const controller = pageController(browser.container);
+    await controller.start();
+    const activation = controller.requestActivation();
+    await Promise.resolve();
+
+    const request = browser.waitingMessages.at(-1) as { nonce?: string };
+    // A different worker answers: terminate immediately rather than waiting for
+    // the drain timeout.
+    browser.dispatch(
+      createServiceWorkerMessage({
+        kind: "ACTIVATED_RELOAD_REQUIRED",
+        sourceBuildId: "worker-build",
+        targetBuildId: "page-build",
+        nonce: request.nonce,
+      }),
+      { postMessage: vi.fn() },
+    );
+
+    await expect(activation).resolves.toEqual({ kind: "PROTOCOL_MISMATCH" });
+    await controller.stop();
+  });
 });