@ardennguyen/zalo-agent-cli 2.5.9 → 3.0.0

package/src/core/self-heal.js

@@ -3,9 +3,8 @@
  * missed from Zalo's offline queue -- on its own socket, with no phone -- and
  * resolves a coverage gap only for the window that pull actually covered.
  *
- * This is what Zalo Web does on every authenticated connect (./offline-queue.js
- * has the bundle offsets and the capture): each message queue is asked for
- * everything after its saved cursor, page by page until the server says there
+ * This is what a client does on every authenticated connect (../zalo/queues/offline-queue.js): each message queue is
+ * asked for everything after its saved cursor, page by page until the server says there
  * is no more. It never involves the phone. The phone-backed restore
  * (transfer-sync-v2, cmd 590) is a different thing -- the full history, and the
  * one step that needs the owner's tap -- and nothing here ever starts it; only a
@@ -14,10 +13,18 @@
  * The rules, and why:
  *
  * - **Runs on every handshake.** The trigger is the cmd 1 handshake frame, read
- *   by ./socket-tap.js, not zca-js's `connected`: the handshake is what makes a
- *   socket usable, it carries the queue list (`qCmds`), and zca-js sets the
+ *   by ../zalo/listener/socket-tap.js, not the listener's `connected`: the handshake is what makes a
+ *   socket usable, it carries the queue list (`qCmds`), and the listener sets the
  *   cipher key from it synchronously just before. A daemon start is simply the
  *   first handshake.
+ * - **One queue at a time, one deadline.** The message queues in the handshake's `qCmds` order, then the control
+ *   queues, each drained from its own cursor once the one before it is done. Asking every queue in the same moment, as
+ *   this once did (5c99017), fails: live, Zalo then answered only the first queue
+ *   and never 511_1 or 603_1, at start or on a wake, reproduced both ways on the same cursors (2026-11-12). The whole
+ *   catch-up has one deadline (`offline_timeout`, 480 s), and one page is waited for at most 30 s inside it. A
+ *   queue unanswered then, or whose socket closed, is not finished: its gaps stay pending and the next connect or
+ *   wake asks again. The socket supervisor owns every reconnect, so at its deadline the run only says it did not
+ *   finish.
  * - **One of our stages at a time.** It takes the daemon's stage lock
  *   (daemon-channel.js `createStageLock`), so it waits behind a running `sync`
  *   stage or `msg history` fetch and never shares the socket with one. Live
@@ -30,8 +37,8 @@
  *   msgId). Moving it before the write would let a crash skip a message
  *   forever; a cursor that lags only re-fetches what insert-if-absent then
  *   ignores. With no saved cursor, a queue starts from its newest stored
- *   message, then from the other queue's cursor (the web keeps one shared
- *   `lastActionId` for queues it has no id of its own for).
+ *   message, then from the other queue's cursor (one shared `lastActionId`
+ *   serves queues that have no id of their own).
  * - **Written insert-if-absent, removals applied.** New messages go through
  *   the history writer (`storeHistoryMessage`, `raw_data.src` "offline"), so a
  *   row the cache already holds is never modified. A recall or a
@@ -46,13 +53,40 @@
  * - **A gap resolves only where it was covered.** A pending gap overlapping the
  *   covered window is resolved, and its uncovered remainder is recorded again
  *   as a pending gap, named with the `zalo-agent sync` command that closes it.
+ * - **Coverage never spans a device change.** Zalo keeps the offline queue per
+ *   device (imei): a new device's queue starts at its login, and a cursor from an
+ *   earlier device is answered with what the new queue holds -- often nothing,
+ *   `more: 0`, no `evict`. So coverage never starts before the time this device id
+ *   came into use (`deviceSince`, ./device-since.js), unless a phone-backed restore
+ *   finished after it and reached back past it. What lies before stays pending as
+ *   `new-device`, with the `zalo-agent sync --from` advice.
+ * - **Overflow repairs groups, and reports the rest.** A queue that reports
+ *   `evict` (or a reset) dropped messages. For the group queue the groups' cloud
+ *   store is read back to the loss (./overflow-repair.js, bounded) and, when that
+ *   reached every group, the group loss is closed without a tap. A 1-1 loss stays
+ *   pending. Either way a `queue_overflow` notice goes out (`onNotice`). A drain of
+ *   0 rows with no `evict` is a quiet queue, never an overflow.
  *
- * Out of scope here, on purpose: the 532 VERIFY_MSG acks and the 534/533/570
- * resend path. They repair a push lost in transit on a live socket; the
- * catch-up above is a cursor pull and does not depend on them -- Zalo Web does
- * not even verify offline batches (`skip_offline_verify`, bundle @~13896600).
+ * - **Items are acknowledged after the write, never before** (Y10, D7). Zalo wants to be told which queue
+ *   items a client holds (a `532`, see ../zalo/queues/queue-verify.js), and repairs what it was not told about. A live push
+ *   is acknowledged once every message in it is committed; a message the server re-sends as a `534` is stored first
+ *   and acknowledged after. An item whose row did not commit is not acknowledged, so the server delivers it again.
+ * - **Offline pages are never acknowledged, and nothing is while they drain.** Offline batches are not verified
+ *   (`skip_offline_verify`), and the verifier starts only once the offline catch-up is done
+ *   (`verify_on_done_offline`). So a connect holds the verifier until its catch-up
+ *   run is over, finished or not; live and `534` acks confirmed meanwhile go out then. Offline rows are still
+ *   stored and still move the cursor.
+ * - **A re-sent message is announced once** (design 6.1, FX 11). Our listener announces a `534` as `message` too,
+ *   after the tap stored it here and handed it on through `onRecovered`. `isRepeatedRepush` tells `listen` and
+ *   `mcp start` that such a message is already stored, so neither prints, buffers nor posts it a second time. The
+ *   `534` ids are noted even with self-heal off (read only: nothing sent, nothing written), so a re-push of a
+ *   message the daemon already got live stays one event there too.
+ *
+ * Out of scope here, on purpose: the 533/570 resend path (the server asking us for ids it says we hold). It
+ * repairs a push lost in transit on a live socket; the catch-up above is a cursor pull and does not depend on
+ * it.
  */
-import { GroupMessage, UserMessage, Undo } from "zca-js";
+import { GroupMessage, UserMessage, Undo } from "../zalo/listener/models.js";
 import {
     getMessageById,
     getMessages,
@@ -65,11 +99,22 @@ import {
     setSyncState,
 } from "./db.js";
 import { isRemovalMessage, storeHistoryMessage, storeLiveDelete, storeLiveUndo } from "./live-store.js";
-import { DEFAULT_MAX_PAGES, DEFAULT_PAGE_TIMEOUT_MS, drainOfflineQueue, messageQueuesFrom } from "./offline-queue.js";
-import { CMD_AUTHEN } from "./socket-tap.js";
+import { IDLE_EVICTED_REASON, IDLE_GAP_REASON } from "./listener-lifecycle.js";
+import {
+    DEFAULT_DRAIN_TIMEOUT_MS,
+    DEFAULT_MAX_PAGES,
+    controlQueuesFrom,
+    drainOfflineQueue,
+    messageQueuesFrom,
+} from "../zalo/queues/offline-queue.js";
+import { replayControlQueues } from "./control-replay.js";
+import { CMD_AUTHEN } from "../zalo/listener/socket-tap.js";
+import { CMD_VERIFY, createQueueVerifier, idsOf } from "../zalo/queues/queue-verify.js";
+import { controlCursorsOf, readControls } from "../zalo/queues/control-queue.js";
 import { describeGap } from "./sync-v2/gap-advice.js";
+import { queueOverflowEvent, repairGroupsFromCloud } from "./overflow-repair.js";
 
-/** zca-js ThreadType.Group; anything else is a 1-1. */
+/** THREAD_TYPE.Group; anything else is a 1-1. */
 const THREAD_GROUP = 1;
 
 /** Live pushes, and the queue whose cursor each one's envelope carries. */
@@ -78,9 +123,28 @@ const PUSH_QUEUE = new Map([
     [521, { queueName: "511_1", rows: "groupMsgs" }],
 ]);
 
+/** The control frame: transfer-sync statuses and batches, unread marks, group and friend events. */
+const CMD_CONTROL = 601;
+
+/** A server re-push of a message it noticed we never confirmed (534), in the envelope of 501/521. */
+const CMD_PUSH_MISS = 534;
+
+/** At most this many re-pushed (534) message ids are remembered until their `message` event; the oldest go first. */
+const REPUSH_MAX = 2000;
+
+/** How long a live push waits for its rows to be stored before it is forgotten (never acknowledged). */
+const AWAIT_MS = 10 * 60 * 1000;
+const AWAIT_MAX = 500;
+
 /** recordGap's own floor: shorter than this is not a gap worth naming. */
 const MIN_GAP_MS = 1000;
 
+/** The reason of a window before this device's login, which the device's queue cannot reach (./device-since.js). */
+export const NEW_DEVICE_REASON = "new-device";
+
+/** The message queue of 1-1 messages and of group messages, by the cmd their name starts with. */
+const isGroupQueue = (queueName) => String(queueName).startsWith("511");
+
 /**
  * The sync_state key holding one queue's cursor, `{id, ts}` as JSON.
  *
@@ -105,25 +169,48 @@ const iso = (ts) => new Date(ts).toISOString();
 /** "uncovered:uncovered:startup-gap" helps nobody; keep the original cause. */
 const baseReason = (reason) => String(reason || "gap").replace(/^(uncovered:)+/, "");
 
+/**
+ * What an uncovered part of a gap is called. A planned idle window (on-demand presence) that the queue no longer
+ * covers becomes `idle-evicted`: a real gap, alarmed on, closed by `sync --from`.
+ */
+const uncoveredReason = (reason, cause = null) =>
+    baseReason(reason) === IDLE_GAP_REASON || baseReason(reason) === IDLE_EVICTED_REASON
+        ? IDLE_EVICTED_REASON
+        : cause || `uncovered:${baseReason(reason)}`;
+
 /**
  * Build a daemon's self-heal. `listen` and `mcp start` each build exactly one.
  *
  * @param {object} args
  * @param {() => object} args.getApi - the daemon's CURRENT api; re-login replaces it
- * @param {ReturnType<import("./socket-tap.js").createSocketTap>} args.tap - the daemon's socket tap
+ * @param {ReturnType<import("../zalo/listener/socket-tap.js").createSocketTap>} args.tap - the daemon's socket tap
  * @param {ReturnType<import("./daemon-channel.js").createStageLock>} [args.lock] - the
  *   stage lock the daemon channel holds; without one the catch-up runs unguarded
- * @param {boolean} [args.enabled=true] - false (`--no-self-heal`): subscribe to
- *   nothing, send nothing, write nothing
+ * @param {boolean} [args.enabled=true] - false (`--no-self-heal`): send nothing,
+ *   write nothing, and subscribe to nothing but the read-only note of `534` ids
+ *   that `isRepeatedRepush` answers from
  * @param {(line: string) => void} [args.log] - one line per run; stderr only
  * @param {(items: Array<{msg: object, info: object}>) => void} [args.onRecovered] -
  *   the messages a run stored that the cache never had, oldest first, for the
  *   daemon's bot-facing output (the MCP buffer, `listen`'s events and webhook)
  * @param {() => number} [args.now]
- * @param {number} [args.timeoutMs] - wait per page
+ * @param {number} [args.timeoutMs] - the whole catch-up's deadline, every queue and page included
+ *   (`offline_timeout`); there is no per-page wait
  * @param {number} [args.maxPages] - pages per queue
+ * @param {boolean} [args.acks=true] - acknowledge stored queue items to Zalo (`532`); false sends nothing of the kind
+ * @param {(undo: Undo) => {stored: boolean}} [args.storeUndo] - writes a recall found on a queue page; the live
+ *   store's by default. A seam for tests: the recall's thread type does not reach zalo.db.
+ * @param {() => (number|{ts: number, source?: string}|null)} [args.deviceSince] - since when this account's device id
+ *   is in use (./device-since.js `deviceSinceOf`); coverage never starts before it. Null or absent: no floor.
+ * @param {(event: object) => void} [args.onNotice] - a `queue_overflow` event (./overflow-repair.js) for the daemon's
+ *   log, webhook and MCP buffer
+ * @param {object} [args.groupRepair] - options for ./overflow-repair.js `repairGroupsFromCloud` (tests: its seams);
+ *   `false` turns the group repair off
+ * @param {(done: {reason: string, ms: number, result: object|null}) => void} [args.onRun] - a catch-up (one run, with
+ *   the runs folded into it) finished: on-demand presence ends its start-up and pulse holds here
  * @returns {{attach: (listener: object) => void, run: (reason?: string) => Promise<object>,
  *   settled: () => Promise<void>, noteDelivered: (msg: object) => void,
+ *   isRepeatedRepush: (msg: object) => boolean,
  *   cursor: (queueName: string) => {id: string, ts: number}|null}}
  */
 export function createSelfHeal({
@@ -134,14 +221,35 @@ export function createSelfHeal({
     log = (line) => console.error(line),
     onRecovered = () => {},
     now = () => Date.now(),
-    timeoutMs = DEFAULT_PAGE_TIMEOUT_MS,
+    timeoutMs = DEFAULT_DRAIN_TIMEOUT_MS,
     maxPages = DEFAULT_MAX_PAGES,
+    acks = true,
+    storeUndo = storeLiveUndo,
+    deviceSince = null,
+    onNotice = () => {},
+    groupRepair = {},
+    onRun = () => {},
 } = {}) {
     const cursors = new Map();
+    // Tells Zalo which queue items are safely stored (Y10). Nothing is handed to it before the commit.
+    const verifier =
+        enabled && acks
+            ? createQueueVerifier({
+                  getListener: () => getApi?.()?.listener ?? null,
+                  settings: () => getApi?.()?.getContext?.()?.settings?.features?.socket ?? {},
+                  log: (line) => say(line),
+              })
+            : null;
+    /** The message ids of `534` frames whose `message` events have not come yet, oldest first. */
+    const repushed = new Set();
+    /** Live pushes whose rows are not all stored yet: {queueName, ids, msgIds: Set, at}. */
+    const awaiting = [];
     const attached = new WeakSet();
     let running = null;
     let again = false;
     let lastResult = null;
+    /** The repair a live overflow started, so they run one at a time. */
+    let overflowRun = Promise.resolve();
 
     const say = (line) => {
         try {
@@ -151,6 +259,43 @@ export function createSelfHeal({
         }
     };
 
+    /**
+     * The coverage floor: this device's login, lowered to where a phone-backed restore reached back when that restore
+     * finished after it (the restore then covers the window the device's queue cannot).
+     */
+    function coverageFloor() {
+        let floor = null;
+        let source = null;
+        try {
+            const v = typeof deviceSince === "function" ? deviceSince() : deviceSince;
+            const ts = typeof v === "object" && v !== null ? Number(v.ts) : Number(v);
+            if (Number.isFinite(ts) && ts > 0) {
+                floor = ts;
+                source = typeof v === "object" && v !== null ? v.source || null : null;
+            }
+        } catch (e) {
+            say(`could not read when this device logged in: ${e?.message || e}`);
+            return { ts: now(), source: "unknown" };
+        }
+        if (floor === null) return null;
+        try {
+            const okAt = Number(getSyncState("lastSyncOkAt"));
+            const okFrom = Number(getSyncState("lastSyncOkFrom"));
+            if (
+                getSyncState("lastSyncOkKind") === "transfer" &&
+                Number.isFinite(okAt) &&
+                Number.isFinite(okFrom) &&
+                okAt >= floor &&
+                okFrom <= floor
+            ) {
+                return { ts: okFrom, source: "restore", deviceTs: floor };
+            }
+        } catch {
+            /* no restore on record: the device's login stays the floor */
+        }
+        return { ts: floor, source, deviceTs: floor };
+    }
+
     function readCursor(queueName) {
         if (cursors.has(queueName)) return cursors.get(queueName);
         let c = null;
@@ -194,11 +339,37 @@ export function createSelfHeal({
             if (!getMessageById(String(id))) return;
             const queueName = msg.type === THREAD_GROUP ? "511_1" : "510_1";
             advance(queueName, { id: String(id), ts: Number(msg.data.ts) || now() });
+            settle(String(id));
         } catch (e) {
             say(`could not record the queue position: ${e?.message || e}`);
         }
     }
 
+    /** A row is committed: acknowledge every live push whose rows are now all committed. */
+    function settle(msgId) {
+        if (!verifier) return;
+        for (let i = awaiting.length - 1; i >= 0; i--) {
+            const entry = awaiting[i];
+            if (!entry.msgIds.delete(msgId) || entry.msgIds.size > 0) continue;
+            awaiting.splice(i, 1);
+            verifier.confirm(entry.queueName, entry.ids);
+        }
+    }
+
+    /** A live push named queue items: acknowledge them once its rows are committed. */
+    function trackPush(queueName, d, rowsKey) {
+        if (!verifier) return;
+        const ids = idsOf(d.queueStatus, queueName);
+        if (!ids.length) return;
+        const cutoff = now() - AWAIT_MS;
+        while (awaiting.length && (awaiting[0].at < cutoff || awaiting.length >= AWAIT_MAX)) awaiting.shift();
+        const rows = Array.isArray(d[rowsKey]) ? d[rowsKey] : [];
+        // Rows the storing handler committed before this frame was read are already safe.
+        const msgIds = new Set(rows.map((r) => String(r?.msgId ?? "")).filter((id) => isId(id) && !getMessageById(id)));
+        if (msgIds.size === 0) verifier.confirm(queueName, ids);
+        else awaiting.push({ queueName, ids, msgIds, at: now() });
+    }
+
     /** True when some pending gap already spans part of [fromTs, toTs]. */
     function overlapsPending(fromTs, toTs) {
         return getPendingSyncGaps().some((g) => Number(g.toTs) > fromTs && Number(g.fromTs) < toTs);
@@ -229,27 +400,48 @@ export function createSelfHeal({
             const evicted = Object.entries(d.queueStatus || {}).some(
                 ([q, st]) => q.startsWith(push.queueName.slice(0, 4)) && Number(st?.evict) === 1,
             );
-            if (evicted && cur) recordLoss(cur.ts, oldest, `queue-evicted:${push.queueName}`);
+            const lost = [];
+            if (evicted && cur && recordLoss(cur.ts, oldest, `queue-evicted:${push.queueName}`)) {
+                lost.push({ queueName: push.queueName, fromTs: cur.ts });
+            }
             const reset = d.resetLastActionId;
             if (isId(reset) && String(reset) !== "0") {
-                if (cur) recordLoss(cur.ts, now(), `queue-reset:${push.queueName}`);
+                if (cur && recordLoss(cur.ts, now(), `queue-reset:${push.queueName}`)) {
+                    lost.push({ queueName: push.queueName, fromTs: cur.ts });
+                }
                 writeCursor(push.queueName, { id: String(reset), ts: now() });
             }
+            if (lost.length) scheduleOverflow(lost);
+            trackPush(push.queueName, d, push.rows);
         } catch (e) {
             say(`could not read a live queue status: ${e?.message || e}`);
         }
     }
 
+    /** A live push reported a loss: repair and report it under the stage lock, after whatever runs now. */
+    function scheduleOverflow(lost) {
+        overflowRun = overflowRun
+            .then(async () => {
+                const release = lock ? await lock.acquire("self-heal") : () => {};
+                try {
+                    await handleOverflow(getApi?.(), lost);
+                } finally {
+                    release();
+                }
+            })
+            .catch((e) => say(`overflow repair failed: ${e?.message || e}`));
+    }
+
     /** Write one page's rows as the listener would have, had it seen them live. */
     function storeRows(queue, rows, ownId, cursorId) {
         const isGroup = queue.threadType === THREAD_GROUP;
-        const out = { added: [], untouched: 0, removals: 0, newest: null, oldestTs: null, older: 0 };
+        const out = { added: [], untouched: 0, removals: 0, newest: null, oldestTs: null, older: 0, failed: 0 };
         for (const raw of rows) {
             if (!raw || typeof raw !== "object") continue;
-            const row = { ...raw }; // zca-js's models rewrite their data in place
+            const row = { ...raw }; // the message models rewrite their data in place
             const ts = Number(row.ts);
             if (Number.isFinite(ts) && ts > 0) out.oldestTs = out.oldestTs === null ? ts : Math.min(out.oldestTs, ts);
-            // lastId means "after this" in every use Zalo Web makes of it; a
+            // lastId means "after this" in every use Zalo makes of it; a
             // row from before the cursor says the server read it differently.
             if (isId(row.msgId) && idGreater(cursorId, row.msgId)) out.older++;
             try {
@@ -260,8 +452,8 @@ export function createSelfHeal({
                     !Array.isArray(content) &&
                     Object.hasOwn(content, "deleteMsg")
                 ) {
-                    // A recall: zca-js turns exactly this shape into its `undo` event.
-                    if (storeLiveUndo(new Undo(ownId, row, isGroup)).stored) out.removals++;
+                    // A recall: the listener turns exactly this shape into its `undo` event.
+                    if (storeUndo(new Undo(ownId, row, isGroup)).stored) out.removals++;
                     continue;
                 }
                 const msg = isGroup ? new GroupMessage(ownId, row) : new UserMessage(ownId, row);
@@ -271,7 +463,11 @@ export function createSelfHeal({
                 }
                 const r = storeHistoryMessage(msg, { src: "offline" });
                 if (r.stored) out.added.push({ msg, info: r.info });
-                else out.untouched++;
+                else {
+                    out.untouched++;
+                    // A row deliberately not written is handled; one whose write failed is not.
+                    if (r.failed) out.failed++;
+                }
                 // The cursor may pass this row only because it is in zalo.db now.
                 const id = String(row.msgId ?? "");
                 if (isId(id) && getMessageById(id) && (!out.newest || idGreater(id, out.newest.id))) {
@@ -279,6 +475,7 @@ export function createSelfHeal({
                 }
             } catch (e) {
                 out.untouched++;
+                out.failed++;
                 say(`a recovered row was not stored: ${e?.message || e}`);
             }
         }
@@ -298,11 +495,11 @@ export function createSelfHeal({
                 left.push({
                     fromTs: f,
                     toTs: coverage.from,
-                    reason: coverage.cause || `uncovered:${baseReason(g.reason)}`,
+                    reason: uncoveredReason(g.reason, coverage.cause),
                 });
             }
             if (t - coverage.to >= MIN_GAP_MS) {
-                left.push({ fromTs: coverage.to, toTs: t, reason: `uncovered:${baseReason(g.reason)}` });
+                left.push({ fromTs: coverage.to, toTs: t, reason: uncoveredReason(g.reason) });
             }
             runInTransaction(() => {
                 resolveSyncGap(g.id);
@@ -343,6 +540,13 @@ export function createSelfHeal({
             `asking Zalo's offline queue on this socket what it missed (${reason}; no phone involved): ${asked.join(", ")}`,
         );
 
+        // One queue at a time, the message queues in `qCmds` order and then the control queues, all under one deadline
+        // from this moment. Not all at once: asked in the same moment, Zalo
+        // answered only the first (510_1) for this client and never 511_1 or 603_1, so group messages and group events
+        // missed while parked never came back (2026-11-12, reproduced both ways on the same cursors).
+        const deadline = now() + timeoutMs;
+        const left = () => Math.max(0, deadline - now());
+
         const results = [];
         for (const p of plan) {
             if (!p.cursor) {
@@ -354,11 +558,13 @@ export function createSelfHeal({
                 tap,
                 queue: p.queue,
                 lastId: p.cursor.id,
-                timeoutMs,
+                timeoutMs: left(),
                 maxPages,
+                log: say,
             });
             const s = storeRows(p.queue, r.rows, ownId, p.cursor.id);
             if (s.newest) advance(p.queue.queueName, s.newest);
+            // Not acknowledged: an offline batch is never verified (`skip_offline_verify`, see the file header).
             if (r.resetTo) writeCursor(p.queue.queueName, { id: r.resetTo, ts: now() });
             if (s.older) {
                 say(
@@ -371,6 +577,7 @@ export function createSelfHeal({
             const reason = r.complete && s.older ? "out-of-order" : r.reason;
             results.push({ queue: p.queue, cursor: p.cursor, ...r, ...s, complete, reason });
         }
+        const controlsDone = replayControls(listener, left());
 
         const added = results.flatMap((r) => r.added || []);
         if (added.length) {
@@ -381,7 +588,7 @@ export function createSelfHeal({
             }
         }
 
-        // Coverage: every queue drained, each from a known point.
+        // Coverage: every queue drained, each from a known point -- and never from before this device's login.
         let coverage = null;
         if (results.every((r) => r.complete)) {
             let from = -Infinity;
@@ -397,6 +604,16 @@ export function createSelfHeal({
                     cause = r.evicted ? `queue-evicted:${r.queue.queueName}` : null;
                 }
             }
+            const floor = from !== null && Number.isFinite(from) ? coverageFloor() : null;
+            if (floor && floor.ts > from) {
+                say(
+                    `this device logged in at ${iso(floor.deviceTs ?? floor.ts)}, after the queue position ` +
+                        `(${iso(from)}): Zalo keeps the offline queue per device, so nothing before that can come ` +
+                        "from it, however empty its answer was",
+                );
+                from = floor.ts;
+                cause = NEW_DEVICE_REASON;
+            }
             if (from !== null && Number.isFinite(from)) coverage = { from, to: now(), cause };
         }
 
@@ -404,17 +621,21 @@ export function createSelfHeal({
 
         // What the server says it lost is a gap even with nothing pending over it.
         const losses = [];
+        const overflowed = [];
         for (const r of results) {
             if (!r.cursor) continue;
             if (r.evicted) {
                 const l = recordLoss(r.cursor.ts, r.oldestTs ?? now(), `queue-evicted:${r.queue.queueName}`);
                 if (l) losses.push(l);
+                overflowed.push({ queueName: r.queue.queueName, fromTs: r.cursor.ts });
             }
             if (r.resetTo) {
                 const l = recordLoss(r.cursor.ts, now(), `queue-reset:${r.queue.queueName}`);
                 if (l) losses.push(l);
+                overflowed.push({ queueName: r.queue.queueName, fromTs: r.cursor.ts });
             }
         }
+        const overflow = overflowed.length ? await handleOverflow(api, overflowed) : null;
 
         const untouched = results.reduce((n, r) => n + (r.untouched || 0), 0);
         const removals = results.reduce((n, r) => n + (r.removals || 0), 0);
@@ -431,6 +652,16 @@ export function createSelfHeal({
             }
         } else {
             const failed = results.filter((r) => !r.complete).map((r) => `${r.queue.queueName} (${r.reason})`);
+            // A planned idle window this run could not cover (a timeout, the page cap) is no longer planned.
+            const idle = getPendingSyncGaps().filter((g) => g.reason === IDLE_GAP_REASON);
+            if (idle.length) {
+                runInTransaction(() => {
+                    for (const g of idle) {
+                        resolveSyncGap(g.id);
+                        recordSyncGap(Number(g.fromTs), Number(g.toTs), IDLE_EVICTED_REASON);
+                    }
+                });
+            }
             const pendingGaps = getPendingSyncGaps();
             let hint = "";
             if (pendingGaps.length) {
@@ -441,9 +672,12 @@ export function createSelfHeal({
             say(`${summary}; not finished: ${failed.join(", ")}, so no window is marked covered${hint}`);
         }
 
+        const controls = await controlsDone;
+
         return {
             ran: true,
             reason,
+            controls,
             recovered: added.length,
             untouched,
             removals,
@@ -451,6 +685,7 @@ export function createSelfHeal({
             resolved: done.resolved,
             remainders: done.remainders,
             losses,
+            overflow,
             queues: results.map((r) => ({
                 queueName: r.queue.queueName,
                 lastId: r.cursor?.id ?? null,
@@ -463,6 +698,118 @@ export function createSelfHeal({
         };
     }
 
+    /**
+     * Zalo dropped (evict) or reset message queues. Groups: read back from the groups' cloud store, and the group loss
+     * closed when every group was reached. 1-1: stays pending. Then one `queue_overflow` notice.
+     *
+     * @param {object} api
+     * @param {Array<{queueName: string, fromTs: number}>} overflowed
+     */
+    async function handleOverflow(api, overflowed) {
+        const fromTs = Math.min(...overflowed.map((o) => o.fromTs));
+        const toTs = now();
+        const queues = [...new Set(overflowed.map((o) => o.queueName))];
+        const direct = queues.some((q) => !isGroupQueue(q));
+        let groups = null;
+        if (queues.some(isGroupQueue) && groupRepair !== false) {
+            say(`the group queue overflowed: reading the groups' cloud store back to ${iso(fromTs)} (no phone)`);
+            let repair;
+            try {
+                repair = await repairGroupsFromCloud({ api, fromTs, ...(groupRepair || {}) });
+            } catch (e) {
+                repair = { checked: 0, active: 0, added: [], complete: false, incomplete: [String(e?.message || e)] };
+            }
+            groups = { checked: repair.checked, active: repair.active, added: repair.added.length };
+            groups.complete = repair.complete;
+            if (repair.added.length) {
+                try {
+                    onRecovered(repair.added);
+                } catch (e) {
+                    say(`handing recovered messages on failed: ${e?.message || e}`);
+                }
+            }
+            if (repair.complete) {
+                // Every group was read back to the loss: the group queue's loss is covered.
+                const closing = getPendingSyncGaps().filter(
+                    (g) => /^queue-(evicted|reset):511/.test(String(g.reason)) && Number(g.fromTs) >= fromTs,
+                );
+                runInTransaction(() => {
+                    for (const g of closing) resolveSyncGap(g.id);
+                });
+                say(
+                    `groups repaired from the cloud store: ${repair.added.length} message(s) added from ` +
+                        `${repair.active} active group(s) of ${repair.checked}; closed ${closing.length} group gap(s)`,
+                );
+            } else {
+                say(
+                    `groups partly repaired from the cloud store (${repair.added.length} message(s) added, ` +
+                        `${repair.checked} group(s) read): ${repair.incomplete.join("; ")}; the group gap stays pending`,
+                );
+            }
+        }
+        const pending = getPendingSyncGaps().some((g) => Number(g.toTs) > fromTs && Number(g.fromTs) < toTs);
+        const ownId = String(api?.getOwnId?.() ?? api?.getContext?.()?.uid ?? "") || null;
+        const event = queueOverflowEvent({ accountId: ownId, fromTs, toTs, queues, direct, groups, pending });
+        try {
+            onNotice(event);
+        } catch (e) {
+            say(`the overflow notice failed: ${e?.message || e}`);
+        }
+        return event;
+    }
+
+    /**
+     * After the message queues: the control queues, replayed from their saved cursors (./control-replay.js), one at a
+     * time, in what is left of the same deadline. Each page reaches the same handlers as a live control push, which
+     * move the cursors; a queue with no saved cursor is not asked. Never rejects.
+     *
+     * @param {object} listener
+     * @param {number} budgetMs - what is left of the catch-up's deadline
+     */
+    async function replayControls(listener, budgetMs) {
+        const queues = controlQueuesFrom(tap?.auth()?.ctrl_qCmds);
+        let results;
+        try {
+            results = await replayControlQueues({
+                listener,
+                tap,
+                queues,
+                cursorOf: readCursor,
+                timeoutMs: budgetMs,
+                maxPages,
+                log: say,
+            });
+        } catch (e) {
+            say(`control replay failed: ${e?.message || e}`);
+            return [];
+        }
+        for (const r of results) {
+            const name = r.queue.queueName;
+            if (r.reason === "no-cursor") {
+                say(`${name}: no saved control cursor yet, nothing to replay (the next live control sets one)`);
+                continue;
+            }
+            if (r.resetTo) writeCursor(name, { id: r.resetTo, ts: now() });
+            say(
+                `${name}: replayed ${r.replayed} control(s) after ${r.cursor.id}` +
+                    (r.complete ? "" : `; not finished (${r.reason})`),
+            );
+            if (r.evicted) {
+                say(
+                    `${name}: Zalo dropped control events while this client was away; group, friend and read-state ` +
+                        "changes among them may be missing, and a `sync` restores what the phone holds",
+                );
+            }
+        }
+        return results.map((r) => ({
+            queueName: r.queue.queueName,
+            replayed: r.replayed,
+            complete: r.complete,
+            reason: r.reason,
+            evicted: r.evicted,
+        }));
+    }
+
     async function healUnderLock(reason) {
         const release = lock ? await lock.acquire("self-heal") : () => {};
         try {
@@ -479,6 +826,7 @@ export function createSelfHeal({
             again = true;
             return;
         }
+        const startedAt = now();
         running = (async () => {
             try {
                 let why = reason;
@@ -491,13 +839,119 @@ export function createSelfHeal({
                 say(`failed: ${e?.message || e}`);
             } finally {
                 running = null;
+                // The catch-up is over, finished or not: what was confirmed meanwhile may be acknowledged now.
+                verifier?.release();
+                try {
+                    onRun({ reason, ms: now() - startedAt, result: lastResult ?? null });
+                } catch {
+                    /* a hook failure is not the catch-up's */
+                }
             }
         })();
     }
 
+    /** Control kinds already reported as unhandled, so each is named once. */
+    const unhandled = new Set();
+
+    /**
+     * A 601: keep the control queues' cursors (forward only) for the pull that will use them, and name any
+     * control this client has no handler for, once, instead of letting it pass unseen (X4).
+     */
+    function noteControls(frame) {
+        const d = frame.data;
+        if (!d) return;
+        try {
+            for (const { queueName, id } of controlCursorsOf(d)) advance(queueName, { id, ts: now() });
+            for (const c of readControls(d)) {
+                if (c.handled) continue;
+                const key = `${c.actType}/${c.act}`;
+                if (unhandled.has(key)) continue;
+                unhandled.add(key);
+                say(`a control this client has no handler for: act_type "${c.actType}", act "${c.act}"`);
+            }
+        } catch (e) {
+            say(`could not read a control frame: ${e?.message || e}`);
+        }
+    }
+
+    /** The server re-sends a message it noticed we never confirmed: store it, then acknowledge it. */
+    function noteMissedPush(frame) {
+        const d = frame.data;
+        if (!d) return;
+        try {
+            const api = getApi?.();
+            const ownId = String(api?.getOwnId?.() ?? api?.getContext?.()?.uid ?? "");
+            const kinds = [
+                { queueName: "510_1", rows: "msgs", threadType: 0 },
+                { queueName: "511_1", rows: "groupMsgs", threadType: THREAD_GROUP },
+            ];
+            const added = [];
+            for (const kind of kinds) {
+                const rows = Array.isArray(d[kind.rows]) ? d[kind.rows] : [];
+                if (!rows.length) continue;
+                const s = storeRows(kind, rows, ownId, "0");
+                added.push(...s.added);
+                if (s.newest) advance(kind.queueName, s.newest);
+                const ids = idsOf(d.queueStatus, kind.queueName);
+                if (verifier && s.failed === 0 && ids.length) verifier.confirm(kind.queueName, ids);
+            }
+            if (added.length) {
+                say(`Zalo re-sent ${added.length} message(s) it noticed this client never confirmed; stored`);
+                onRecovered(added);
+            }
+        } catch (e) {
+            say(`could not store a re-sent message: ${e?.message || e}`);
+        }
+    }
+
+    /** A 534 frame: note its message ids before the listener announces them (the tap sees a frame first). */
+    function noteRepushIds(frame) {
+        const d = frame.data;
+        if (!d) return;
+        for (const key of ["msgs", "groupMsgs", "pageMsgs"]) {
+            for (const row of Array.isArray(d[key]) ? d[key] : []) {
+                if (!isId(row?.msgId)) continue;
+                const id = String(row.msgId);
+                repushed.delete(id);
+                repushed.add(id);
+                if (repushed.size > REPUSH_MAX) repushed.delete(repushed.values().next().value);
+            }
+        }
+    }
+
+    /**
+     * True for the `message` event of a message the server re-sent (534) that zalo.db already holds: it was stored
+     * and announced already (live, or by the re-push handling above), so the daemon must not announce it again. A
+     * re-sent message that is not stored yet is false: the daemon announces and stores it as a live one. Each id
+     * answers true once.
+     */
+    function isRepeatedRepush(msg) {
+        const id = msg?.data?.msgId;
+        if (!isId(id) || !repushed.has(String(id))) return false;
+        repushed.delete(String(id));
+        try {
+            return !!getMessageById(String(id));
+        } catch {
+            return false;
+        }
+    }
+
+    if (tap) tap.on(CMD_PUSH_MISS, noteRepushIds);
     if (enabled && tap) {
-        tap.on(CMD_AUTHEN, () => trigger("connect"));
+        tap.on(CMD_AUTHEN, () => {
+            // A new connection: what was waiting to be acknowledged belonged to the old one. Nothing is acknowledged
+            // until this connection's catch-up is over (`verify_on_done_offline`).
+            verifier?.reset();
+            verifier?.hold();
+            awaiting.length = 0;
+            trigger("connect");
+        });
         for (const cmd of PUSH_QUEUE.keys()) tap.on(cmd, noteEnvelope);
+        tap.on(CMD_CONTROL, noteControls);
+        if (verifier) {
+            tap.on(CMD_VERIFY, (frame) => verifier.handleReply(frame));
+            tap.on(CMD_PUSH_MISS, noteMissedPush);
+        }
     }
 
     return {
@@ -514,9 +968,14 @@ export function createSelfHeal({
             return lastResult;
         },
         settled() {
-            return running ? running.then(() => {}) : Promise.resolve();
+            return Promise.all([running ?? Promise.resolve(), overflowRun]).then(() => {});
         },
         noteDelivered,
+        isRepeatedRepush,
         cursor: (queueName) => readCursor(queueName),
+        /** The queue acknowledgements so far, or null when they are off. */
+        queueAcks: () => verifier?.stats() ?? null,
+        /** Send the pending acknowledgements now (a test and shutdown seam). */
+        flushAcks: () => verifier?.flush() ?? 0,
     };
 }