agents-can-communicate 0.10.1 → 0.10.3

package/docs/ARCHITECTURE.md

@@ -272,8 +272,14 @@ running. A binding that names no client process is judged by its session record
 exited, or, with no process recorded, when ACC's presence rule calls the session offline -
 closed, or 30 minutes without a heartbeat. A live recorded process, or a record that cannot be
 read, keeps it a hold. A client that wakes later runs its next hook through the launcher on
-the active generation. For a binding that names its client, presence TTL, `acc finish`, and
-delivery off do not prove native lifecycle end. Normal background updates never
+the active generation. A binding that names its client stops holding when that process exits,
+or when the process that now has its pid started after the binding was last written, since a
+client runs when it writes its binding (#276); a start time that cannot be read decides
+nothing. A zombie, a process that has exited but that its parent has not collected, counts as
+exited for a binding, an ACC lease and a pin (#280); a process state that cannot be read
+decides nothing. Otherwise presence TTL, `acc finish`, and delivery off do not prove native lifecycle
+end. A pending release that cannot activate is replaced by a newer one when `acc update` runs
+or an automatic check is due (#277). Normal background updates never
 restart a vendor daemon. An explicitly approved maintenance job can do so through the
 adapter lifecycle port: it waits for idle work, validates exact process identity, stops
 the service under the admission fence, refreshes integrations, and restores and verifies

package/docs/UPGRADING.md

@@ -57,6 +57,21 @@ closed. 0.10.1 judges such a binding by its session record and removes it while
 the update, so the update from 0.9.x finishes once no 0.9.x session runs. From 0.10.0, 0.10.1
 keeps the store contract, so it activates without waiting for open sessions.
 
+If `acc update` keeps retrying a pending 0.10.0 or 0.10.1 and names processes that are not your
+clients, or `unknown client pid`, clear the pending release and update again:
+
+<!-- test:illustration needs an installed ACC with a release pending; a fresh sandbox has neither -->
+```bash
+acc update --pin none
+acc update
+```
+
+`--pin none` keeps following stable releases and drops a pending release that is not pinned.
+From 0.10.2, the update does this by itself: a pending release that cannot activate is replaced
+by a newer one, and a binding whose pid another program now has no longer holds the update.
+From 0.10.3, a process that has exited but that its parent has not collected (a zombie, `Z` in
+`ps`) no longer holds it either.
+
 ## From 0.8.x
 
 Update as usual:
@@ -355,7 +370,10 @@ PID/start time, executable, socket ownership and versions before stopping the se
 It refreshes integrations, starts the service again and verifies the result. If refreshing
 integrations fails, it still restores the service and reports the incomplete update.
 A later update worker can resume an already approved recovery without another question.
+A daemon that its parent leaves as a zombie after the stop counts as stopped, and the worker
+starts the service again.
 `acc doctor --json` exposes `update.maintenance`; ordinary `acc doctor` describes its state.
+A failed or cancelled job is no longer reported once its release is active by another path.
 A current launcher can use the newer management reader while the old runtime is active.
 When admission is blocked, doctor returns an update-only report without inspecting the
 workspace or repairing it. Older launchers gain this reader after activation, or through

package/node_modules/@agents-can-communicate/adapter-antigravity/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@agents-can-communicate/adapter-antigravity",
-  "version": "0.10.1",
+  "version": "0.10.3",
   "private": true,
   "type": "module",
   "exports": {

package/node_modules/@agents-can-communicate/adapter-antigravity/plugin/plugin.json

@@ -1,5 +1,5 @@
 {
   "name": "acc",
-  "version": "0.10.1",
+  "version": "0.10.3",
   "description": "Coordinate this Antigravity CLI session with other AI agent sessions working in the same workspace."
 }

package/node_modules/@agents-can-communicate/adapter-claude-code/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@agents-can-communicate/adapter-claude-code",
-  "version": "0.10.1",
+  "version": "0.10.3",
   "private": true,
   "type": "module",
   "exports": {

package/node_modules/@agents-can-communicate/adapter-codex/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@agents-can-communicate/adapter-codex",
-  "version": "0.10.1",
+  "version": "0.10.3",
   "private": true,
   "type": "module",
   "exports": {

package/node_modules/@agents-can-communicate/adapter-codex/src/maintenance-host.mjs

@@ -181,6 +181,12 @@ export async function observeMaintenanceProcess(pid, paths, run, { system = WIND
   if (response.status !== 0) return { state: "unknown" };
   const match = /^(.{24})\s+(.+)$/.exec(response.stdout.trim());
   if (!match || !startTimeValid(match[1])) return { state: "unknown" };
+  // A zombie has exited; only its parent has not collected it. `ps` names it
+  // `<defunct>` on macOS and `[name] <defunct>` on Linux. On 2026-10-06 the
+  // daemon `daemon stop` ended stayed one under Codex's own pid-update-loop,
+  // and reading it as alive left the stop unverified and the start refused
+  // (#280).
+  if (/<defunct>$/.test(match[2])) return { state: "dead" };
   return { state: "alive", processStartTime: match[1], command: match[2] };
 }
 

package/node_modules/@agents-can-communicate/adapter-gemini-cli/extension/gemini-extension.json

@@ -1,6 +1,6 @@
 {
   "name": "agents-can-communicate",
-  "version": "0.10.1",
+  "version": "0.10.3",
   "description": "Coordinate this Gemini CLI session with other AI agent sessions working in the same workspace.",
   "contextFileName": "skills/acc/SKILL.md"
 }

package/node_modules/@agents-can-communicate/adapter-gemini-cli/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@agents-can-communicate/adapter-gemini-cli",
-  "version": "0.10.1",
+  "version": "0.10.3",
   "private": true,
   "type": "module",
   "exports": {

package/node_modules/@agents-can-communicate/adapter-grok/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@agents-can-communicate/adapter-grok",
-  "version": "0.10.1",
+  "version": "0.10.3",
   "private": true,
   "type": "module",
   "exports": {

package/node_modules/@agents-can-communicate/adapter-kimi/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@agents-can-communicate/adapter-kimi",
-  "version": "0.10.1",
+  "version": "0.10.3",
   "private": true,
   "type": "module",
   "exports": {

package/node_modules/@agents-can-communicate/adapter-sdk/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@agents-can-communicate/adapter-sdk",
-  "version": "0.10.1",
+  "version": "0.10.3",
   "private": true,
   "type": "module",
   "exports": {

package/node_modules/@agents-can-communicate/adapter-sdk/src/index.mjs

@@ -29,7 +29,7 @@ export { clearSessionBinding, listSessionBindings, loadSessionBinding,
 
 export { clearNativeAttempt, loadNativeAttempt, storeNativeAttempt } from "./native-attempt.mjs";
 export { channelSocketDirectory } from "./channel-directory.mjs";
-export { readProcessArgs, readProcessTable, readWindowsProcess, splitWindowsCommandLine }
+export { processIsZombie, readProcessArgs, readProcessTable, readWindowsProcess, splitWindowsCommandLine }
   from "./process-table.mjs";
 export { npmShimTarget, resolveExecutable, runExecutable } from "./executables.mjs";
 export { pathOf } from "./executables.mjs";

package/node_modules/@agents-can-communicate/adapter-sdk/src/process-table.mjs

@@ -165,6 +165,26 @@ export async function readProcessTable({ platform = process.platform, from = pro
   return posixTable({ run, timeoutMs });
 }
 
+/**
+ * Whether a process has exited and only waits for its parent to collect it -
+ * the POSIX zombie, `ps` state `Z`. `kill(pid, 0)` succeeds for one and `ps`
+ * still lists its start time, so a liveness check that stops there calls it
+ * alive: on 2026-10-06 a stopped Codex daemon left as a zombie under Codex's
+ * own pid-update-loop kept an ACC update blocked and its restart refused
+ * (#280). True or false when `ps` answers; null when it does not. Windows has
+ * no such state.
+ */
+export async function processIsZombie(pid, { platform = process.platform, run = execFileAsync,
+  timeoutMs = 2_000 } = {}) {
+  if (platform === "win32") return false;
+  try {
+    const { stdout } = await run("/bin/ps", ["-p", String(pid), "-o", "stat="],
+      { timeout: timeoutMs, env: { ...process.env, LC_ALL: "C" } });
+    const state = String(stdout).trim();
+    return state === "" ? null : state.startsWith("Z");
+  } catch { return null; }
+}
+
 /**
  * The words one process was started with, or null when they cannot be read.
  *

package/node_modules/@agents-can-communicate/cli/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@agents-can-communicate/cli",
-  "version": "0.10.1",
+  "version": "0.10.3",
   "private": true,
   "type": "module",
   "exports": {

package/node_modules/@agents-can-communicate/cli/src/managed-runtime/activation.mjs

@@ -1,10 +1,13 @@
+import { execFile } from "node:child_process";
 import { createHash } from "node:crypto";
 // Kept above 300 lines because activation and generation retirement must use
 // the same validated runtime/native hold definitions under the admission mutex.
 // Splitting their safety rules risks disagreeing about which users are live.
-import { readdir, rm } from "node:fs/promises";
+import { readdir, rm, stat } from "node:fs/promises";
 import path from "node:path";
 import { pathToFileURL } from "node:url";
+import { promisify } from "node:util";
+import { processIsZombie, readWindowsProcess } from "@agents-can-communicate/adapter-sdk";
 import { classifySessionPresence } from "@agents-can-communicate/core";
 import { assertPortableId } from "@agents-can-communicate/protocol";
 import { readSessionRecord, withWriterMutex } from "@agents-can-communicate/storage-filesystem";
@@ -49,6 +52,53 @@ async function isMcpContinuity(record, name, workspaceRoot, workspaceId) {
 const clientPidOf = record => record?.schemaVersion === 1 && Number.isSafeInteger(record.clientPid)
   && record.clientPid > 0 ? record.clientPid : null;
 
+const execFileAsync = promisify(execFile);
+
+/** When a process started, in milliseconds since the epoch; null when it cannot be read. */
+export async function processStartedAt(pid, { platform = process.platform, run = execFileAsync } = {}) {
+  try {
+    if (platform === "win32") {
+      const found = await readWindowsProcess(pid, { timeoutMs: 5_000 });
+      if (found?.pid !== pid || typeof found.start !== "string" || !/^\d+$/.test(found.start)) return null;
+      // A FILETIME: 100-nanosecond intervals since 1601-01-01 UTC.
+      return Number(BigInt(found.start) / 10_000n) - 11_644_473_600_000;
+    }
+    const { stdout } = await run("/bin/ps", ["-p", String(pid), "-o", "lstart="],
+      { timeout: 2_000, env: { ...process.env, LC_ALL: "C" } });
+    const started = Date.parse(stdout.trim());
+    return Number.isFinite(started) ? started : null;
+  } catch { return null; }
+}
+
+/** A client runs when it writes its binding, so a process that started after
+ * the binding was last written is not that client: the operating system gave
+ * its pid to another program. On 2026-10-06 chrome-devtools-mcp and two IntelliJ
+ * helpers held three such pids and kept 0.10.1 pending (#276). `ps` counts whole
+ * seconds and a filesystem may too; a second of slack keeps a client's own
+ * process from reading as another. A start time that cannot be read decides
+ * nothing. */
+async function pidReused(pid, file, startedAt) {
+  const [started, written] = await Promise.all([startedAt(pid),
+    stat(file).then(info => info.mtimeMs, () => null)]);
+  return started !== null && written !== null && started > written + 1_000;
+}
+
+// A zombie has exited; only its parent has not collected it, and kill(pid, 0)
+// still succeeds for it. On 2026-10-06 the Codex daemon `daemon stop` ended
+// stayed one under Codex's own pid-update-loop and kept 0.10.2 pending (#280).
+// A state that cannot be read decides nothing.
+const exitedProcess = async (pid, pidIsAlive, zombie) =>
+  await confirmedDead(pid, pidIsAlive) || await zombie(pid) === true;
+
+// One reading per pid for one pass: a vendor service can hold several bindings.
+const perPass = read => {
+  const seen = new Map();
+  return pid => {
+    if (!seen.has(pid)) seen.set(pid, Promise.resolve().then(() => read(pid)).catch(() => null));
+    return seen.get(pid);
+  };
+};
+
 /** A binding that names no client process cannot be judged by its own facts;
  * its session record can. It is stale when the record is gone, when the record
  * names a process that has exited, or, with no process recorded, when ACC's own
@@ -57,7 +107,7 @@ const clientPidOf = record => record?.schemaVersion === 1 && Number.isSafeIntege
  * process, or a record or store that cannot be read, keeps the binding a hold.
  * On 2026-10-06 the maintainer's 0.10.0 waited behind 32 such bindings from
  * 0.5.10 to 0.9.0 that no client could clear (#273). */
-async function staleWithoutPid(workspaceRoot, workspaceId, record, pidIsAlive) {
+async function staleWithoutPid(workspaceRoot, workspaceId, record, pidIsAlive, file, startedAt, zombie) {
   if (record?.schemaVersion !== 1) return false;
   let session;
   try {
@@ -65,15 +115,19 @@ async function staleWithoutPid(workspaceRoot, workspaceId, record, pidIsAlive) {
       sessionId: record.accSessionId, allowLegacyIdentity: true });
   } catch { return false; }
   if (session === null) return true;
-  if (Number.isSafeInteger(session.pid) && session.pid > 0) return confirmedDead(session.pid, pidIsAlive);
+  if (Number.isSafeInteger(session.pid) && session.pid > 0) {
+    return await exitedProcess(session.pid, pidIsAlive, zombie) || pidReused(session.pid, file, startedAt);
+  }
   // No process to ask, so the probe is never called.
   return classifySessionPresence(session, new Date().toISOString(), () => true) === "offline";
 }
 
 /** Native bindings outlive finish/TTL; MCP process lifetime is covered by its runtime lease. */
-export async function listNativeHolds(root, { pidIsAlive = defaultPidIsAlive, strict = false } = {}) {
+export async function listNativeHolds(root, { pidIsAlive = defaultPidIsAlive, strict = false,
+  startedAt = processStartedAt, zombie = processIsZombie } = {}) {
   const workspaces = path.join(path.dirname(await canonicalManagerRoot(root)), "workspaces");
   if (!await managedDirectory(workspaces)) return [];
+  const started = perPass(startedAt), zombieNow = perPass(zombie);
   const holds = [];
   for (const entry of await readdir(workspaces, { withFileTypes: true })) {
     if (!entry.isDirectory()) {
@@ -87,15 +141,18 @@ export async function listNativeHolds(root, { pidIsAlive = defaultPidIsAlive, st
     if (!await managedDirectory(bindings)) continue;
     for (const name of await readdir(bindings)) {
       if (!name.endsWith(".json")) continue;
-      const record = await readManagedJson(path.join(bindings, name)).catch(error => {
+      const file = path.join(bindings, name);
+      const record = await readManagedJson(file).catch(error => {
         if (strict) throw error;
         return null;
       });
       if (record === undefined) continue; // Concurrent lifecycle removal, before fencing.
       if (await isMcpContinuity(record, name, path.dirname(bindings), entry.name)) continue;
       const pid = clientPidOf(record);
-      if (pid !== null && await confirmedDead(pid, pidIsAlive)) continue;
-      if (pid === null && await staleWithoutPid(path.dirname(bindings), entry.name, record, pidIsAlive)) continue;
+      if (pid !== null && (await exitedProcess(pid, pidIsAlive, zombieNow)
+        || await pidReused(pid, file, started))) continue;
+      if (pid === null && await staleWithoutPid(path.dirname(bindings), entry.name, record, pidIsAlive,
+        file, started, zombieNow)) continue;
       holds.push({ kind: "native", pid, reason: pid === null ? "unknown_client_pid" : "client_running",
         storeVersion: Number.isSafeInteger(record?.storeVersion) ? record.storeVersion : null });
     }
@@ -109,17 +166,22 @@ const LIFECYCLE_LOCK_MS = 2_000;
 const lifecycleLocks = (workspaceRoot, harnessSessionId) => ({ locks: path.join(workspaceRoot,
   "lifecycle-locks", createHash("sha256").update(String(harnessSessionId)).digest("hex")) });
 
-/** Remove the bindings without a client pid that listNativeHolds no longer
- * counts. An older generation activates its successor with its own rule, which
- * counts every one of them as a live client; it loads the pending generation's
- * refresh before it lists blockers, so that refresh calls this (#273). Each one
- * is judged again and removed under its session's lifecycle lock, so a hook that
- * renews it is either seen or comes after (AI review of #274); a lock that stays
- * busy leaves the binding as it is. */
-export async function sweepStaleBindings(root, { pidIsAlive = defaultPidIsAlive } = {}) {
+/** Remove the bindings that listNativeHolds no longer counts and an older
+ * generation still does: those without a client pid that are stale (#273),
+ * those whose pid was given to another program (#276), and those whose client
+ * is a zombie (#280). An older generation
+ * activates its successor with its own rule, and loads the pending generation's
+ * refresh before it lists blockers, so that refresh calls this. A binding whose
+ * pid is merely dead holds nothing for either and stays, as the record a resumed
+ * conversation finds. Each one is judged again and removed under its session's
+ * lifecycle lock, so a hook that renews it is either seen or comes after (AI
+ * review of #274); a lock that stays busy leaves the binding as it is. */
+export async function sweepStaleBindings(root, { pidIsAlive = defaultPidIsAlive,
+  startedAt = processStartedAt, zombie = processIsZombie } = {}) {
   const workspaces = path.join(path.dirname(await canonicalManagerRoot(root)), "workspaces");
   if (!await managedDirectory(workspaces)) return 0;
   const clock = { now: () => new Date().toISOString() };
+  const started = perPass(startedAt), zombieNow = perPass(zombie);
   let removed = 0;
   for (const entry of await readdir(workspaces, { withFileTypes: true })) {
     if (!entry.isDirectory()) continue;
@@ -130,10 +192,16 @@ export async function sweepStaleBindings(root, { pidIsAlive = defaultPidIsAlive
     for (const name of await readdir(bindings)) {
       if (!name.endsWith(".json")) continue;
       const file = path.join(bindings, name);
-      const stale = async record => record !== undefined && record !== null
-        && clientPidOf(record) === null && typeof record.harnessSessionId === "string"
-        && !await isMcpContinuity(record, name, workspaceRoot, entry.name)
-        && staleWithoutPid(workspaceRoot, entry.name, record, pidIsAlive);
+      const stale = async record => {
+        if (record === undefined || record === null || typeof record.harnessSessionId !== "string"
+          || await isMcpContinuity(record, name, workspaceRoot, entry.name)) return false;
+        const pid = clientPidOf(record);
+        if (pid !== null) {
+          return !await confirmedDead(pid, pidIsAlive)
+            && (await zombieNow(pid) === true || pidReused(pid, file, started));
+        }
+        return staleWithoutPid(workspaceRoot, entry.name, record, pidIsAlive, file, started, zombieNow);
+      };
       const judged = await readManagedJson(file).catch(() => undefined);
       if (!await stale(judged)) continue;
       const gone = await withWriterMutex(lifecycleLocks(workspaceRoot, judged.harnessSessionId),
@@ -165,9 +233,13 @@ export const blocksActivation = (hold, incomingStoreVersion) =>
  * generation's, or when either side declares none.
  */
 export async function listActivationBlockers(root, { pidIsAlive = defaultPidIsAlive,
-  ignorePid = null, incomingStoreVersion = null, strictNative = false } = {}) {
-  const leases = (await listRuntimeHolds(root, { pidIsAlive })).filter(lease => lease.pid !== ignorePid);
-  const native = await listNativeHolds(root, { pidIsAlive, strict: strictNative });
+  ignorePid = null, incomingStoreVersion = null, strictNative = false, startedAt,
+  zombie = processIsZombie } = {}) {
+  const zombieNow = perPass(zombie);
+  const leases = (await listRuntimeHolds(root, { pidIsAlive, zombie: zombieNow }))
+    .filter(lease => lease.pid !== ignorePid);
+  const native = await listNativeHolds(root, { pidIsAlive, strict: strictNative, zombie: zombieNow,
+    ...(startedAt === undefined ? {} : { startedAt }) });
   const holds = [
     ...leases.map(({ pid, kind, runtime }) => ({ pid, kind, nativeBindings: 0,
       storeVersion: runtime?.storeVersion ?? null })),
@@ -246,8 +318,11 @@ async function resolveCandidate(value) {
 // reclaimed need to tell that from a pass that found nothing to remove.
 const POSTPONED = Object.freeze({ removed: [], postponed: true });
 
-export async function reclaimGenerations({ root, active = null, pidIsAlive = defaultPidIsAlive } = {}) {
+export async function reclaimGenerations({ root, active = null, pidIsAlive = defaultPidIsAlive,
+  zombie = processIsZombie } = {}) {
   root = await canonicalManagerRoot(root);
+  // A lease or a pin of a client that exited uncollected keeps no generation (#280).
+  const zombieNow = perPass(zombie);
   return withManagerLock(root, async () => {
     const generations = path.join(root, "generations");
     if (!await managedDirectory(generations)) return { removed: [] };
@@ -286,7 +361,7 @@ export async function reclaimGenerations({ root, active = null, pidIsAlive = def
       if (resolved.root) referenced.add(resolved.root);
     }
     let holds;
-    try { holds = await listRuntimeHolds(root, { pidIsAlive }); }
+    try { holds = await listRuntimeHolds(root, { pidIsAlive, zombie: zombieNow }); }
     catch { return POSTPONED; } // An unreadable lease is an unknown holder.
     for (const lease of holds) {
       const resolved = await resolveCandidate(lease.runtime?.root);
@@ -296,7 +371,7 @@ export async function reclaimGenerations({ root, active = null, pidIsAlive = def
     // A client that exits without a clean session end leaves its pin behind;
     // reap confirmed-dead ones before treating every remaining pin as live,
     // the same way admission already reaps confirmed-dead leases.
-    await reapPins({ root, pidIsAlive });
+    await reapPins({ root, pidIsAlive, zombie: zombieNow });
     const pins = path.join(root, "pins");
     if (await managedDirectory(pins)) {
       for (const name of await readdir(pins)) {
@@ -364,7 +439,7 @@ const recipe = c => JSON.stringify([c.active, c.pending, c.home, c.targets, c.au
 
 /** Detection is outside the fence; every integration write and commit stays inside it. */
 export async function activatePending(root, { prepare = prepareCandidate, env = process.env,
-  pidIsAlive = defaultPidIsAlive, ignorePid = null, beforeActivation = null } = {}) {
+  pidIsAlive = defaultPidIsAlive, ignorePid = null, beforeActivation = null, startedAt, zombie } = {}) {
   root = await canonicalManagerRoot(root);
   const before = await readControl(root);
   if (!before?.pending) return { activated: false, reason: "no_pending_update" };
@@ -374,8 +449,8 @@ export async function activatePending(root, { prepare = prepareCandidate, env =
     const current = await readControl(root);
     if (!current || recipe(current) !== recipe(before)) return { activated: false, reason: "state_changed" };
     if (beforeActivation) await beforeActivation(current);
-    const blockers = await listActivationBlockers(root, { pidIsAlive, ignorePid,
-      incomingStoreVersion: current.pending.storeVersion });
+    const blockers = await listActivationBlockers(root, { pidIsAlive, ignorePid, startedAt,
+      ...(zombie === undefined ? {} : { zombie }), incomingStoreVersion: current.pending.storeVersion });
     if (blockers.length) {
       const notice = activationBlockerNotice(current.pending.version, blockers);
       await writeControl(root, { ...current, notice });
@@ -409,7 +484,8 @@ export async function activatePending(root, { prepare = prepareCandidate, env =
   // turn an already-completed activation into a reported failure - it can
   // always run again on the next activation.
   if (outcome.activated) {
-    try { await reclaimGenerations({ root, active: activatedRoot, pidIsAlive }); }
+    try { await reclaimGenerations({ root, active: activatedRoot, pidIsAlive,
+      ...(zombie === undefined ? {} : { zombie }) }); }
     catch { /* Best effort; never fails an activation that already succeeded. */ }
     // The reclaim above is this, older, code's rule (#208). The activated
     // generation's own rule runs next, from its own files as its refresh did,

package/node_modules/@agents-can-communicate/cli/src/managed-runtime/diagnostics.mjs

@@ -27,7 +27,12 @@ export async function managedUpdateDiagnostic(root, manager, running) {
       : activationBlockerNotice(manager.pending.version, blockers),
     ...(blockers === null ? {} : { holds: blockers.length, blockers }) };
   const maintenance = await readMaintenance(root);
-  if (maintenance && (maintenance.target.root === (manager.pending ?? manager.active).root
+  // A job that failed or was cancelled for the release that is now active is
+  // history: another path activated it, and "run acc update to retry" asked
+  // for an update already done (#281). A completed job still reports its restart.
+  const settled = maintenance && manager.pending === null && maintenance.target.root === manager.active.root
+    && ["failed", "cancelled"].includes(maintenance.status);
+  if (maintenance && !settled && (maintenance.target.root === (manager.pending ?? manager.active).root
     || MAINTENANCE_ACTIVE.includes(maintenance.status))) {
     update.maintenance = maintenanceReport(maintenance);
     update.notice = maintenanceNotice(maintenance);

package/node_modules/@agents-can-communicate/cli/src/managed-runtime/leases.mjs

@@ -31,8 +31,10 @@ export async function acquireRuntime(root, { pid = process.pid, kind = "cli" } =
 /** Management decisions call this under withManagerLock to exclude admissions.
  * Records are immutable and uniquely named; confirmed-dead cleanup also works
  * outside a transaction. No release-on-return or timestamp expiry exists.
+ * A decision passes `zombie` so a holder that exited but was never collected
+ * holds nothing (#280); admission leaves it out and spawns no process.
  */
-export async function listRuntimeHolds(root, { pidIsAlive = defaultPidIsAlive } = {}) {
+export async function listRuntimeHolds(root, { pidIsAlive = defaultPidIsAlive, zombie = null } = {}) {
   root = await canonicalManagerRoot(root);
   if (!await managedDirectory(root)) return [];
   const directory = path.join(root, "leases");
@@ -50,7 +52,7 @@ export async function listRuntimeHolds(root, { pidIsAlive = defaultPidIsAlive }
       || typeof lease.kind !== "string" || !lease.kind
       || !Number.isFinite(Date.parse(lease.createdAt))) throw new Error("invalid runtime lease");
     const runtime = await validateRuntime(root, lease.runtime);
-    if (await confirmedDead(lease.pid, pidIsAlive)) {
+    if (await confirmedDead(lease.pid, pidIsAlive) || zombie !== null && await zombie(lease.pid) === true) {
       await rm(file, { force: true });
       removed = true;
     } else holds.push({ ...lease, runtime });

package/node_modules/@agents-can-communicate/cli/src/managed-runtime/pins.mjs

@@ -115,7 +115,7 @@ export async function resolvePinnedGeneration({ root, harnessSessionId, active }
  * already reaps dead leases; pins follow the same confirmed-death rule. A pin
  * that names no client, which 0.8.0 and earlier wrote when the pid was not yet
  * known, is reaped outright: nothing could ever prove its session over. */
-export async function reapPins({ root, pidIsAlive = defaultPidIsAlive } = {}) {
+export async function reapPins({ root, pidIsAlive = defaultPidIsAlive, zombie = null } = {}) {
   root = await canonicalManagerRoot(root);
   const directory = path.join(root, "pins");
   if (!await managedDirectory(directory)) return;
@@ -125,7 +125,8 @@ export async function reapPins({ root, pidIsAlive = defaultPidIsAlive } = {}) {
     const file = path.join(directory, name);
     const record = await readManagedJson(file).catch(() => null);
     if (record?.schemaVersion !== SCHEMA_VERSION) continue; // Unknown shape: an unknown holder.
-    if (!isPid(record.clientPid) || await confirmedDead(record.clientPid, pidIsAlive)) {
+    if (!isPid(record.clientPid) || await confirmedDead(record.clientPid, pidIsAlive)
+      || zombie !== null && await zombie(record.clientPid) === true) {
       await rm(file, { force: true });
       removed = true;
     }

package/node_modules/@agents-can-communicate/cli/src/managed-runtime/refresh.mjs

@@ -2,7 +2,9 @@ import path from "node:path";
 import { applyPlan, detectInstallation, livePolicyOf, loadOwnership, planInstallation }
   from "@agents-can-communicate/installer";
 import { ALL_ADAPTERS, clientContext, probeTimeout } from "../install-command.mjs";
+import { processIsZombie } from "@agents-can-communicate/adapter-sdk";
 import { sweepStaleBindings } from "./activation.mjs";
+import { listRuntimeHolds } from "./leases.mjs";
 import { stablePaths, writeLaunchers } from "./launchers.mjs";
 import { bridgeLegacyMaintenance } from "./legacy-maintenance.mjs";
 
@@ -12,6 +14,9 @@ export async function prepareRefresh({ control, root, env = process.env, callerP
   // An older generation lists its blockers right after this returns, by a rule
   // that counts a binding without a client pid as a live client forever (#273).
   await sweepStaleBindings(root).catch(() => {});
+  // It counts a zombie's lease as a live ACC process too (#280); a zombie's lease
+  // is removed like a dead process's.
+  await listRuntimeHolds(root, { zombie: processIsZombie }).catch(() => {});
   const dataHome = path.dirname(path.dirname(root));
   const wanted = new Set(control.targets);
   const adapters = ALL_ADAPTERS().filter(adapter => wanted.has(adapter.id));

package/node_modules/@agents-can-communicate/cli/src/managed-runtime/worker.mjs

@@ -32,8 +32,18 @@ export async function performUpdate(root, { force = false, check = false, env =
   }
   if (!force && !check && !initial.auto) return { ...data, checked: false, reason: "auto_off" };
   // Recovery is local; the hard no-network override must not strand a partial refresh.
+  let baseline = initial.active.version;
+  let attempt = null;
   if (initial.pending && !check && (force || !networkDisabled(env))) {
-    return { ...data, ...await activate(root, { env, ignorePid }), checked: false };
+    attempt = await activate(root, { env, ignorePid });
+    // A pending release that cannot activate kept every later one out: on
+    // 2026-10-06 0.10.0 stayed pending behind stale bindings, and the 0.10.1
+    // that fixed them reached the machine only after its pending release was
+    // cleared by hand (#277). Once a check is due, or when asked, a newer
+    // release is staged in its place.
+    if (attempt.activated || attempt.reason !== "processes_active" || networkDisabled(env)
+      || !force && !checkDue(initial)) return { ...data, ...attempt, checked: false };
+    baseline = initial.pending.version;
   }
   if (networkDisabled(env)) return { ...data, checked: false, reason: "network_disabled" };
   if (!force && !check && !checkDue(initial)) return { ...data, checked: false, reason: "not_due" };
@@ -44,7 +54,9 @@ export async function performUpdate(root, { force = false, check = false, env =
   const release = await discover({ pin: initial.pin, env });
   const newer = newerVersion(release.version, initial.active.version);
   const checked = { ...data, checked: true, latest: release.version, newer };
-  if (check || !newer) return checked;
+  // Nothing newer than a blocked pending release: report the activation that
+  // waits, as before - the caller asks to restart a client service from it.
+  if (check || !newerVersion(release.version, baseline)) return attempt === null ? checked : { ...checked, ...attempt };
   const generation = await download(root, release, { env });
   // generation.hold has protected this candidate since it was staged;
   // release it once this publish attempt resolves either way - published,

package/node_modules/@agents-can-communicate/core/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@agents-can-communicate/core",
-  "version": "0.10.1",
+  "version": "0.10.3",
   "private": true,
   "type": "module",
   "exports": {

package/node_modules/@agents-can-communicate/delivery-router/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@agents-can-communicate/delivery-router",
-  "version": "0.10.1",
+  "version": "0.10.3",
   "private": true,
   "type": "module",
   "exports": {

package/node_modules/@agents-can-communicate/hook-runner/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@agents-can-communicate/hook-runner",
-  "version": "0.10.1",
+  "version": "0.10.3",
   "private": true,
   "type": "module",
   "exports": {

package/node_modules/@agents-can-communicate/installer/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@agents-can-communicate/installer",
-  "version": "0.10.1",
+  "version": "0.10.3",
   "private": true,
   "type": "module",
   "exports": {

package/node_modules/@agents-can-communicate/mcp-server/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@agents-can-communicate/mcp-server",
-  "version": "0.10.1",
+  "version": "0.10.3",
   "private": true,
   "type": "module",
   "exports": {

package/node_modules/@agents-can-communicate/protocol/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@agents-can-communicate/protocol",
-  "version": "0.10.1",
+  "version": "0.10.3",
   "private": true,
   "type": "module",
   "exports": {

package/node_modules/@agents-can-communicate/storage-filesystem/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@agents-can-communicate/storage-filesystem",
-  "version": "0.10.1",
+  "version": "0.10.3",
   "private": true,
   "type": "module",
   "exports": {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agents-can-communicate",
-  "version": "0.10.1",
+  "version": "0.10.3",
   "accManagedUpdateProtocol": 2,
   "accStoreVersion": 7,
   "type": "module",