fix: recover stuck background updates

install.sh

@@ -1395,7 +1395,7 @@ main() {
   unit_tmp=$(mktemp -d)
   sed "s#/opt/tallynote#$PREFIX#g; s#/var/lib/tallynote#$DATA_DIR#g" "$release_dir/systemd/tallynote.service" > "$unit_tmp/tallynote.service"
   sed "s#/opt/tallynote#$PREFIX#g; s#/var/lib/tallynote#$DATA_DIR#g; s#/var/lib/tallynote-backups#$(dirname -- "$DATA_DIR")/tallynote-backups#g" "$release_dir/systemd/tallynote-update.service" > "$unit_tmp/tallynote-update.service"
-  sed "s#/var/lib/tallynote#$DATA_DIR#g" "$release_dir/systemd/tallynote-update.path" > "$unit_tmp/tallynote-update.path"
+  sed "s#/opt/tallynote#$PREFIX#g; s#/var/lib/tallynote#$DATA_DIR#g" "$release_dir/systemd/tallynote-update.path" > "$unit_tmp/tallynote-update.path"
   sed "s#/opt/tallynote#$PREFIX#g; s#/etc/tallynote#$CONFIG_DIR#g" "$release_dir/bin/tallynote-admin-init" > "$unit_tmp/tallynote-admin-init"
   install -o root -g root -m 644 "$unit_tmp/tallynote.service" /etc/systemd/system/tallynote.service
   install -o root -g root -m 644 "$unit_tmp/tallynote-update.service" /etc/systemd/system/tallynote-update.service

package.json

@@ -1,6 +1,6 @@
 {
   "name": "tallynote",
-  "version": "1.1.13",
+  "version": "1.1.14",
   "private": true,
   "type": "module",
   "packageManager": "pnpm@9.0.6",

scripts/tallynote-update-runner.sh

@@ -41,7 +41,25 @@ if [[ "$request_operation" == download ]]; then
   [[ -n "$node_bin" ]] || die 'node runtime not found'
   cli="$CURRENT_LINK/dist/server/cli/update.js"
   [[ -f "$cli" ]] || die 'update CLI not found in current release'
-  "$node_bin" "$cli" --request-file "$REQUEST_FILE" || exit $?
+  set +e
+  "$node_bin" "$cli" --request-file "$REQUEST_FILE"
+  download_result=$?
+  set -e
+  if (( download_result != 0 )); then
+    # The CLI normally records failed itself. Retry the explicit finalization
+    # for failures that happen before its catch handler can persist the row,
+    # then remove the one-shot request so a failed download cannot keep the
+    # path unit in a permanently triggered state.
+    download_job_id=$(sed -n 's/.*"jobId"[[:space:]]*:[[:space:]]*"\([0-9a-f-]*\)".*/\1/p' "$REQUEST_FILE" | head -n 1)
+    if [[ "$download_job_id" =~ ^[0-9a-f-]{36}$ ]]; then
+      for _ in 1 2 3; do
+        if "$node_bin" "$cli" --finalize-job "$download_job_id" --finalize-status failed --message '更新下载失败' >/dev/null 2>&1; then break; fi
+        sleep 1
+      done
+    fi
+    rm -f -- "$REQUEST_FILE"
+    exit "$download_result"
+  fi
   rm -f -- "$REQUEST_FILE"
   exit 0
 fi
@@ -112,6 +130,27 @@ recover_stale_state() {
     done
     return 1
   fi
+  if [[ "$current_target" == "$state_old" ]]; then
+    # The process may have restored the old release before it was killed. In
+    # that case the old link is already safe to serve, but the database row
+    # can still be `applying`; finish it as failed before clearing recovery
+    # markers so the UI does not poll forever.
+    recovery_node="$CURRENT_LINK/runtime/bin/node"
+    [[ -x "$recovery_node" ]] || recovery_node=$(command -v node || true)
+    if finalize_state_job "$recovery_node" failed "$state_job"; then
+      rm -f -- "$REQUEST_FILE" 2>/dev/null || true
+      clear_update_state || true
+      return 11
+    fi
+    # A crash before the CLI created its job row is safe to retry. Preserve
+    # the request while dropping only the stale state marker.
+    if [[ -f "$REQUEST_FILE" && ! -L "$REQUEST_FILE" ]]; then
+      clear_update_state || true
+      return 0
+    fi
+    clear_update_state || true
+    return 0
+  fi
   if [[ "$current_target" != "$state_old" ]]; then
     rollback_link="$PREFIX/.current-recovery-$$-${RANDOM}.tmp"
     [[ ! -e "$rollback_link" && ! -L "$rollback_link" ]] || return 1
@@ -159,7 +198,12 @@ fi
 rollback_current() {
   local current_target rollback_link
   current_target=$(readlink -f -- "$CURRENT_LINK" 2>/dev/null || true)
-  [[ "$current_target" == "$old_target" ]] && return 0
+  if [[ "$current_target" == "$old_target" ]]; then
+    # An earlier failure branch may already have restored the link. Keep the
+    # marker truthful so the EXIT trap can still finalize the job.
+    switched=0
+    return 0
+  fi
   rollback_link="$PREFIX/.current-rollback-$$-${RANDOM}.tmp"
   [[ ! -e "$rollback_link" && ! -L "$rollback_link" ]] || return 1
   ln -s -- "$old_target" "$rollback_link" || return 1
@@ -172,8 +216,16 @@ rollback_current() {
 
 finalize_failed_job() {
   [[ "$job_id" =~ ^[0-9a-f-]{36}$ ]] || return 0
-  [[ -n "$old_node" && -f "$CURRENT_LINK/dist/server/cli/update.js" ]] || return 0
-  "$old_node" "$CURRENT_LINK/dist/server/cli/update.js" --finalize-job "$job_id" --finalize-status failed --message '新版本健康检查失败，已恢复上一版本' >/dev/null 2>&1
+  [[ -n "$old_node" && -f "$CURRENT_LINK/dist/server/cli/update.js" ]] || return 1
+  # Give SQLite a moment to release a transient lock before declaring the
+  # recovery itself failed.
+  for _ in 1 2 3; do
+    if "$old_node" "$CURRENT_LINK/dist/server/cli/update.js" --finalize-job "$job_id" --finalize-status failed --message '新版本健康检查失败，已恢复上一版本' >/dev/null 2>&1; then
+      return 0
+    fi
+    sleep 1
+  done
+  return 1
 }
 
 finalize_completed_job() {
@@ -187,7 +239,10 @@ cleanup_after_update() {
   local result=$? rollback_ok=1
   if (( result != 0 && handled == 0 )); then
     if ! rollback_current; then rollback_ok=0; fi
-    if (( rollback_ok == 1 && switched == 0 )); then
+    # Once the old release is active again, always try to close the job. The
+    # previous marker could remain set when an earlier branch had already
+    # rolled back before entering this EXIT trap, leaving `applying` forever.
+    if (( rollback_ok == 1 )); then
       if finalize_failed_job; then
         rm -f -- "$REQUEST_FILE"
         clear_update_state || true

scripts/test-installer.sh

@@ -4,6 +4,13 @@ root=$(cd "$(dirname "$0")/.." && pwd)
 bash -n "$root/install.sh" "$root/scripts/tallynote-update.sh"
 grep -Eq '^RestrictAddressFamilies=AF_UNIX AF_INET AF_INET6 AF_NETLINK$' "$root/systemd/tallynote.service"
 grep -Eq '^RestrictAddressFamilies=AF_UNIX AF_INET AF_INET6 AF_NETLINK$' "$root/systemd/tallynote-update.service"
+grep -Eq '^PathExists=/opt/tallynote/\.update-state$' "$root/systemd/tallynote-update.path"
+grep -Eq '^PathChanged=/opt/tallynote/\.update-state$' "$root/systemd/tallynote-update.path"
+grep -Eq '^PathChanged=/opt/tallynote$' "$root/systemd/tallynote-update.path"
+if grep -Eq '^ConditionPathExists=' "$root/systemd/tallynote-update.service"; then
+  echo 'update service must not require only the request file' >&2
+  exit 1
+fi
 output=$(bash "$root/install.sh" --dry-run --release-base-url https://releases.example.test/releases)
 grep -q 'dry-run' <<<"$output"
 grep -q '\[阶段\] 检查运行环境' <<<"$output"
@@ -192,6 +199,17 @@ bash -c '
   set_env_key test value
 ' _ "$installer_lib" "$release_archive" "$tmp/install-release"
 
+# The installed path unit must watch both the data-directory request and the
+# release-prefix recovery marker after custom paths are substituted.
+rendered_path="$tmp/rendered-update.path"
+sed "s#/opt/tallynote#$tmp/custom-prefix#g; s#/var/lib/tallynote#$tmp/custom-data#g" \
+  "$root/systemd/tallynote-update.path" > "$rendered_path"
+grep -Fxq "PathExists=$tmp/custom-data/update-request.json" "$rendered_path"
+grep -Fxq "PathChanged=$tmp/custom-data/update-request.json" "$rendered_path"
+grep -Fxq "PathExists=$tmp/custom-prefix/.update-state" "$rendered_path"
+grep -Fxq "PathChanged=$tmp/custom-prefix/.update-state" "$rendered_path"
+grep -Fxq "PathChanged=$tmp/custom-prefix" "$rendered_path"
+
 # The production admin wrapper must load a release-relative runtime, change to
 # the release root, and forward CLI arguments without requiring pnpm.
 wrapper_prefix="$tmp/wrapper-prefix"

server/app.ts

@@ -59,6 +59,7 @@ import {
   checkForUpdate,
   publicCheckFromCache,
   publicUpdateJob,
+  reconcileOrphanedUpdateJobs,
   readCachedRelease,
   writeUpdateRequest,
   type UpdateRequest,
@@ -939,6 +940,7 @@ export async function buildApp(database: DatabaseContext, config: AppConfig) {
     // Release metadata and task state should never be stored by an upstream
     // proxy or a shared browser cache.
     reply.header("Cache-Control", "no-store");
+    reconcileOrphanedUpdateJobs(database.sqlite, config);
     const cached = publicCheckFromCache(database.sqlite, config);
     const row = database.sqlite.prepare(`
       SELECT id, operation, status, version, platform, asset_name AS assetName,
@@ -956,6 +958,7 @@ export async function buildApp(database: DatabaseContext, config: AppConfig) {
 
   app.post("/api/update/check", { preHandler: guard(database, config), bodyLimit: 32 * 1024 }, async (request, reply) => {
     try {
+      reconcileOrphanedUpdateJobs(database.sqlite, config);
       // Disabled/dev installs do not contact a release endpoint, so repeated
       // checks are local status reads and should remain immediately usable.
       if (config.updateStrategy !== "disabled") {
@@ -995,6 +998,7 @@ export async function buildApp(database: DatabaseContext, config: AppConfig) {
     let applyAuditRecorded = false;
     let applyAuditTarget: string | undefined;
     try {
+    reconcileOrphanedUpdateJobs(database.sqlite, config);
     if (config.updateStrategy !== "systemd") {
       throw new AppError(503, "UPDATE_NOT_AVAILABLE", "当前安装方式未启用一键更新，请使用命令行更新");
     }
@@ -1145,6 +1149,7 @@ export async function buildApp(database: DatabaseContext, config: AppConfig) {
 
   app.post("/api/update/download", { preHandler: guard(database, config), bodyLimit: 32 * 1024 }, async (request, reply) => {
     const input = updateDownloadSchema.parse(request.body);
+    reconcileOrphanedUpdateJobs(database.sqlite, config);
     if (config.updateStrategy !== "systemd") throw new AppError(503, "UPDATE_NOT_AVAILABLE", "当前安装方式未启用一键更新，请使用命令行更新");
     const active = database.sqlite.prepare(`SELECT id FROM update_jobs WHERE status IN (${ACTIVE_UPDATE_STATUSES.map(() => "?").join(",")}) LIMIT 1`).get(...ACTIVE_UPDATE_STATUSES) as { id: string } | undefined;
     if (active) throw new AppError(409, "UPDATE_IN_PROGRESS", "已有更新任务正在进行，请等待完成");
@@ -1175,6 +1180,7 @@ export async function buildApp(database: DatabaseContext, config: AppConfig) {
 
   app.get("/api/update/jobs/:id", { preHandler: guard(database, config) }, async (request, reply) => {
     const id = z.string().uuid().parse((request.params as { id: string }).id);
+    reconcileOrphanedUpdateJobs(database.sqlite, config);
     const row = database.sqlite.prepare(`
       SELECT id, operation, status, version, platform, asset_name AS assetName,
         size_bytes AS sizeBytes, error_message AS errorMessage,

server/cli/update.ts

@@ -26,7 +26,7 @@ import {
   type ReleaseMetadata,
   type UrlPolicy,
 } from "../update.js";
-import { attachSidecarHash } from "../update-service.js";
+import { ACTIVE_UPDATE_STATUSES, attachSidecarHash } from "../update-service.js";
 import type { UpdateJobStatus } from "../../shared/contracts.js";
 
 const updateRequestFileSchema = z.object({
@@ -153,7 +153,7 @@ function writeJob(sqlite: Database.Database | undefined, jobId: string, values:
       operation, status, version, platform, release_url, asset_name, asset_url,
       expected_sha256, actual_sha256, download_path, backup_path, size_bytes, error_message,
       created_at, updated_at, completed_at)
-    VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+    VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
     ON CONFLICT(id) DO UPDATE SET
       admin_id=COALESCE(excluded.admin_id, update_jobs.admin_id),
       session_hash=COALESCE(excluded.session_hash, update_jobs.session_hash),
@@ -355,7 +355,9 @@ export function finalizeUpdateJob(
     FROM update_jobs WHERE id=?
   `).get(jobId) as { id: string; status: UpdateJobStatus; version: string; platform: string; adminId: string | null; requestId: string | null; sessionHash: string | null } | undefined;
   if (!row) throw new Error("更新任务不存在");
-  if (row.status !== "applying" && row.status !== "completed" && row.status !== "failed") throw new Error("更新任务状态不允许完成");
+  const canComplete = row.status === "applying" || row.status === "completed";
+  const canFail = ACTIVE_UPDATE_STATUSES.includes(row.status) || row.status === "completed" || row.status === "failed";
+  if (status === "completed" ? !canComplete : !canFail) throw new Error("更新任务状态不允许完成");
   const now = Date.now();
   const safeFailureMessage = status === "failed" ? "新版本健康检查失败，已恢复上一版本" : null;
   sqlite.transaction(() => {
@@ -477,27 +479,34 @@ export async function main(config: AppConfig = loadConfig()): Promise<void> {
   const database = openDatabase(config);
   try {
     if (request?.operation === "apply") {
-      const staged = database.sqlite.prepare("SELECT download_path AS downloadPath, version FROM update_jobs WHERE id=? AND status='staged' AND operation='apply'").get(request.jobId) as { downloadPath: string | null; version: string } | undefined;
-      if (!staged?.downloadPath || staged.version !== request.version) throw new Error("更新暂存任务无效");
-      const root = path.resolve(config.updateWorkspaceDir);
-      const candidate = await validateStagedWorkspacePath(staged.downloadPath, root);
-      await applyStagedUpdate({
-        sqlite: database.sqlite,
-        jobId: request.jobId,
-        version: request.version,
-        stagedPath: candidate,
-        currentDir,
-        currentLink: request.currentLink,
-        releasesDir: request.releasesDir,
-        workspaceRoot: root,
-        ...(backupArchive ? { backupArchivePath: backupArchive } : {}),
-        ...(dataBackupArchive ? { dataBackupArchivePath: dataBackupArchive } : {}),
-        dataBackupSource: config.dataDir,
-        maxBytes: config.updateMaxBytes,
-        dataBackupMaxBytes: config.maxTotalBytes,
-      });
-      console.log(`更新已切换：${request.version}`);
-      return;
+      const staged = database.sqlite.prepare("SELECT status, operation, download_path AS downloadPath, version FROM update_jobs WHERE id=?").get(request.jobId) as { status: UpdateJobStatus; operation: "download" | "apply"; downloadPath: string | null; version: string } | undefined;
+      if (staged?.status === "staged" && staged.operation === "apply") {
+        if (!staged.downloadPath || staged.version !== request.version) throw new Error("更新暂存任务无效");
+        const root = path.resolve(config.updateWorkspaceDir);
+        const candidate = await validateStagedWorkspacePath(staged.downloadPath, root);
+        await applyStagedUpdate({
+          sqlite: database.sqlite,
+          jobId: request.jobId,
+          version: request.version,
+          stagedPath: candidate,
+          currentDir,
+          currentLink: request.currentLink,
+          releasesDir: request.releasesDir,
+          workspaceRoot: root,
+          ...(backupArchive ? { backupArchivePath: backupArchive } : {}),
+          ...(dataBackupArchive ? { dataBackupArchivePath: dataBackupArchive } : {}),
+          dataBackupSource: config.dataDir,
+          maxBytes: config.updateMaxBytes,
+          dataBackupMaxBytes: config.maxTotalBytes,
+        });
+        console.log(`更新已切换：${request.version}`);
+        return;
+      }
+      if (staged && !(staged.status === "queued" && staged.operation === "apply")) throw new Error("更新任务状态无效");
+      // A direct one-click request starts in queued/apply. Older clients do
+      // not have a separate download step, so fall through to runUpdate,
+      // which downloads, verifies, backs up, and switches the release in one
+      // transaction. A staged request still takes the branch above.
     }
     const result = await runUpdate({
       ...(effectiveMetadataUrl ? { metadataUrl: effectiveMetadataUrl } : {}),

server/update-service.ts

@@ -1,7 +1,9 @@
+import { lstatSync, realpathSync } from "node:fs";
 import { chmod, mkdir, rename, writeFile } from "node:fs/promises";
 import path from "node:path";
 import { createPublicKey, randomUUID, verify as verifySignature } from "node:crypto";
 import type Database from "better-sqlite3";
+import { writeAudit } from "./audit.js";
 import { AppError } from "./errors.js";
 import type { AppConfig } from "./config.js";
 import {
@@ -30,6 +32,12 @@ export const ACTIVE_UPDATE_STATUSES: readonly UpdateJobStatus[] = [
   "applying",
 ];
 
+// A queued job normally starts within seconds and an applying job completes
+// after the service health check. This grace period only applies when both
+// hand-off markers are gone, so an active runner is never reclaimed midway
+// through a download, backup, or switch.
+export const ORPHANED_UPDATE_TIMEOUT_MS = 5 * 60 * 1000;
+
 export type CachedRelease = {
   checkedAt: number;
   metadataUrl: string;
@@ -355,3 +363,72 @@ export function publicUpdateJob(row: Record<string, unknown> | undefined): Recor
     ...(row.status === "applying" ? { restartWindowSeconds: 30 } : {}),
   };
 }
+
+function markerExists(filePath: string): boolean {
+  try {
+    const info = lstatSync(filePath);
+    return info.isFile() || info.isSymbolicLink();
+  } catch {
+    return false;
+  }
+}
+
+function currentReleaseVersion(config: AppConfig): string | null {
+  try {
+    const target = realpathSync(config.currentLink);
+    const releases = realpathSync(config.releasesDir);
+    if (!target.startsWith(`${releases}${path.sep}`)) return null;
+    return path.basename(target);
+  } catch {
+    return null;
+  }
+}
+
+/**
+ * Release an update row left behind after both privileged hand-off markers
+ * disappeared. This is deliberately conservative: staged downloads remain
+ * available for an explicit apply, and any visible marker means the runner
+ * still owns recovery.
+ */
+export function reconcileOrphanedUpdateJobs(database: Database.Database, config: AppConfig, now = Date.now()): number {
+  const placeholders = ACTIVE_UPDATE_STATUSES.map(() => "?").join(",");
+  const rows = database.prepare(`
+    SELECT id, status, version, admin_id AS adminId, request_id AS requestId,
+      updated_at AS updatedAt
+    FROM update_jobs
+    WHERE status IN (${placeholders})
+    ORDER BY updated_at ASC
+  `).all(...ACTIVE_UPDATE_STATUSES) as Array<{ id: string; status: UpdateJobStatus; version: string; adminId: string | null; requestId: string | null; updatedAt: number | null }>;
+  if (rows.length === 0) return 0;
+  const requestPresent = markerExists(config.updateRequestPath);
+  const statePresent = markerExists(path.join(config.installPrefix, ".update-state"));
+  if (requestPresent || statePresent) return 0;
+  const releaseVersion = currentReleaseVersion(config);
+  let reconciled = 0;
+  for (const row of rows) {
+    if (row.status === "staged" || typeof row.updatedAt !== "number" || now - row.updatedAt < ORPHANED_UPDATE_TIMEOUT_MS) continue;
+    const status: "completed" | "failed" = row.status === "applying" && releaseVersion === row.version ? "completed" : "failed";
+    const errorMessage = status === "failed" ? "更新任务超时，已释放更新队列" : null;
+    const changed = database.transaction(() => {
+      const result = database.prepare(`
+        UPDATE update_jobs
+        SET status=?, error_message=?, completed_at=?, updated_at=?
+        WHERE id=? AND status=? AND updated_at=?
+      `).run(status, errorMessage, now, now, row.id, row.status, row.updatedAt);
+      if (result.changes !== 1) return false;
+      writeAudit(database, {
+        requestId: row.requestId || randomUUID(),
+        actorAdminId: row.adminId,
+        action: "update.reconciled",
+        targetType: "update",
+        targetId: row.id,
+        outcome: status === "completed" ? "success" : "failure",
+        before: { status: row.status, version: row.version },
+        after: { status, version: row.version, reason: "orphaned_timeout" },
+      });
+      return true;
+    })();
+    if (changed) reconciled += 1;
+  }
+  return reconciled;
+}

systemd/tallynote-update.path

@@ -4,6 +4,15 @@ Description=Watch for TallyNote release update requests
 [Path]
 PathExists=/var/lib/tallynote/update-request.json
 PathChanged=/var/lib/tallynote/update-request.json
+# The recovery marker lives beside the release link. Watching it as well
+# allows systemd to resume reconciliation when the runner is interrupted
+# after consuming the request but before clearing its state file.
+PathExists=/opt/tallynote/.update-state
+PathChanged=/opt/tallynote/.update-state
+# Keep a directory-level fallback because some systemd/inotify versions skip
+# dotfiles when watching an individual path. State writes are atomic renames,
+# so the containing directory changes even when the marker itself is hidden.
+PathChanged=/opt/tallynote
 Unit=tallynote-update.service
 
 [Install]

systemd/tallynote-update.service

@@ -2,7 +2,6 @@
 Description=TallyNote privileged release updater
 After=network-online.target
 Wants=network-online.target
-ConditionPathExists=/var/lib/tallynote/update-request.json
 
 [Service]
 Type=oneshot
@@ -12,6 +11,10 @@ WorkingDirectory=/opt/tallynote/current
 EnvironmentFile=-/etc/tallynote/tallynote.env
 ExecStart=/usr/local/libexec/tallynote-update-runner
 Environment=PATH=/usr/sbin:/usr/bin:/sbin:/bin
+# Downloads, archive validation and data backups can exceed systemd's 90s
+# default start timeout on a slower server. Keep one update job alive long
+# enough to finish or reach its own health-check/recovery path.
+TimeoutStartSec=30min
 NoNewPrivileges=true
 CapabilityBoundingSet=
 AmbientCapabilities=

systemd/tallynote.service

@@ -14,6 +14,8 @@ Environment=PATH=/opt/tallynote/current/runtime/bin:/usr/sbin:/usr/bin:/sbin:/bi
 ExecStart=/opt/tallynote/current/bin/tallynote
 Restart=on-failure
 RestartSec=5s
+# Do not let a wedged Node process hold an update stop forever.
+TimeoutStopSec=30s
 NoNewPrivileges=true
 PrivateTmp=true
 ProtectSystem=strict

tests/update.test.ts

@@ -3,7 +3,7 @@ import { mkdir, readlink, symlink, writeFile, readFile, stat, readdir } from "no
 import { mkdtemp, rm } from "node:fs/promises";
 import { tmpdir } from "node:os";
 import path from "node:path";
-import { createHash, generateKeyPairSync, sign } from "node:crypto";
+import { createHash, generateKeyPairSync, randomUUID, sign } from "node:crypto";
 import {
   atomicSwitchRelease,
   createSafeArchive,
@@ -18,13 +18,13 @@ import {
   selectReleaseAsset,
   validateHttpsUrl,
 } from "../server/update.js";
-import { runUpdate } from "../server/cli/update.js";
+import { finalizeUpdateJob, runUpdate } from "../server/cli/update.js";
 import { validateUpdateRequest } from "../server/cli/update.js";
-import { checkForUpdate, verifyReleaseSignature } from "../server/update-service.js";
+import { checkForUpdate, ORPHANED_UPDATE_TIMEOUT_MS, reconcileOrphanedUpdateJobs, verifyReleaseSignature } from "../server/update-service.js";
 import { loadConfig, prepareDataDirectories } from "../server/config.js";
 import { openDatabase } from "../server/db/index.js";
 
-const envKeys = ["TALLYNOTE_DATA_DIR", "TALLYNOTE_PUBLIC_ORIGIN", "TALLYNOTE_COOKIE_SECURE", "TALLYNOTE_UPDATE_STRATEGY", "TALLYNOTE_UPDATE_METADATA_URL", "TALLYNOTE_UPDATE_ALLOWED_HOSTS", "TALLYNOTE_UPDATE_REQUIRE_SIGNATURE", "TALLYNOTE_UPDATE_PUBLIC_KEY"];
+const envKeys = ["TALLYNOTE_DATA_DIR", "TALLYNOTE_INSTALL_PREFIX", "TALLYNOTE_PUBLIC_ORIGIN", "TALLYNOTE_COOKIE_SECURE", "TALLYNOTE_UPDATE_STRATEGY", "TALLYNOTE_UPDATE_METADATA_URL", "TALLYNOTE_UPDATE_ALLOWED_HOSTS", "TALLYNOTE_UPDATE_REQUIRE_SIGNATURE", "TALLYNOTE_UPDATE_PUBLIC_KEY"];
 const originalFetch = globalThis.fetch;
 
 afterEach(() => {
@@ -207,6 +207,122 @@ describe("更新安全工具", () => {
     }
   });
 
+  it("更新器支持旧客户端创建的 queued/apply 直接更新请求", async () => {
+    const root = await mkdtemp(path.join(tmpdir(), "tallynote-update-direct-"));
+    const previousFetch = globalThis.fetch;
+    let database: ReturnType<typeof openDatabase> | undefined;
+    try {
+      const dataDir = path.join(root, "data");
+      const installPrefix = path.join(root, "install");
+      process.env.TALLYNOTE_DATA_DIR = dataDir;
+      process.env.TALLYNOTE_INSTALL_PREFIX = installPrefix;
+      process.env.TALLYNOTE_PUBLIC_ORIGIN = "http://127.0.0.1:3998";
+      process.env.TALLYNOTE_COOKIE_SECURE = "false";
+      process.env.TALLYNOTE_UPDATE_STRATEGY = "systemd";
+      process.env.TALLYNOTE_UPDATE_METADATA_URL = "https://updates.example/latest";
+      process.env.TALLYNOTE_UPDATE_ALLOWED_HOSTS = "updates.example";
+      process.env.TALLYNOTE_UPDATE_REQUIRE_SIGNATURE = "false";
+      const config = loadConfig();
+      prepareDataDirectories(config);
+      await mkdir(config.releasesDir, { recursive: true, mode: 0o755 });
+      const oldRelease = path.join(config.releasesDir, config.appVersion);
+      await mkdir(path.join(oldRelease, "dist"), { recursive: true, mode: 0o755 });
+      await writeFile(path.join(oldRelease, "dist", "marker"), "old");
+      await symlink(oldRelease, config.currentLink);
+
+      const source = path.join(root, "source");
+      await mkdir(path.join(source, "dist"), { recursive: true, mode: 0o755 });
+      await writeFile(path.join(source, "dist", "marker"), "new");
+      const archive = path.join(root, "release.tar.gz");
+      await createSafeArchive(source, archive);
+      const bytes = await readFile(archive);
+      const digest = createHash("sha256").update(bytes).digest("hex");
+      const jobId = randomUUID();
+      database = openDatabase(config);
+      const now = Date.now();
+      const assetName = `tallynote-1.2.0-${detectPlatform().target}.tar.gz`;
+      database.sqlite.prepare(`
+        INSERT INTO update_jobs(id, operation, status, version, platform, asset_url,
+          expected_sha256, created_at, updated_at, requested_at)
+        VALUES (?, 'apply', 'queued', '1.2.0', ?, ?, ?, ?, ?, ?)
+      `).run(jobId, detectPlatform().target, "https://updates.example/" + assetName, digest, now, now, now);
+      globalThis.fetch = (async (input: string | URL) => {
+        const url = input.toString();
+        if (url.endsWith("/latest")) return new Response(JSON.stringify({ tag_name: "v1.2.0", assets: [{ name: assetName, browser_download_url: `https://updates.example/${assetName}` }, { name: "SHA256SUMS", browser_download_url: "https://updates.example/SHA256SUMS" }] }));
+        if (url.endsWith("SHA256SUMS")) return new Response(`${digest}  ${assetName}\n`);
+        return new Response(bytes, { headers: { "content-length": String(bytes.length) } });
+      }) as typeof fetch;
+
+      await runUpdate({
+        metadataUrl: config.updateMetadataUrl,
+        version: "1.2.0",
+        currentVersion: config.appVersion,
+        currentDir: config.currentLink,
+        stagingDir: path.join(root, "staging"),
+        currentLink: config.currentLink,
+        releasesDir: config.releasesDir,
+        allowedHosts: config.updateAllowedHosts,
+        maxBytes: config.updateMaxBytes,
+        dataBackupMaxBytes: config.maxTotalBytes,
+        deferCompletion: true,
+        operation: "apply",
+        jobId,
+        sqlite: database.sqlite,
+        fetchImpl: globalThis.fetch,
+      });
+      expect(await readFile(path.join(config.currentLink, "dist", "marker"), "utf8")).toBe("new");
+      const row = database.sqlite.prepare("SELECT operation, status FROM update_jobs WHERE id=?").get(jobId);
+      expect(row).toEqual({ operation: "apply", status: "applying" });
+      finalizeUpdateJob(database.sqlite, jobId, "failed");
+      expect(database.sqlite.prepare("SELECT status FROM update_jobs WHERE id=?").get(jobId)).toEqual({ status: "failed" });
+    } finally {
+      globalThis.fetch = previousFetch;
+      if (database) database.sqlite.close();
+      await rm(root, { recursive: true, force: true });
+    }
+  });
+
+  it("在请求和恢复标记丢失后收敛孤儿任务，但保留 staged 下载", async () => {
+    const root = await mkdtemp(path.join(tmpdir(), "tallynote-update-reconcile-"));
+    let database: ReturnType<typeof openDatabase> | undefined;
+    try {
+      const dataDir = path.join(root, "data");
+      const installPrefix = path.join(root, "install");
+      process.env.TALLYNOTE_DATA_DIR = dataDir;
+      process.env.TALLYNOTE_INSTALL_PREFIX = installPrefix;
+      process.env.TALLYNOTE_PUBLIC_ORIGIN = "http://127.0.0.1:3998";
+      process.env.TALLYNOTE_COOKIE_SECURE = "false";
+      process.env.TALLYNOTE_UPDATE_STRATEGY = "systemd";
+      process.env.TALLYNOTE_UPDATE_METADATA_URL = "https://updates.example/latest";
+      process.env.TALLYNOTE_UPDATE_ALLOWED_HOSTS = "updates.example";
+      process.env.TALLYNOTE_UPDATE_REQUIRE_SIGNATURE = "false";
+      const config = loadConfig();
+      prepareDataDirectories(config);
+      await mkdir(path.join(config.releasesDir, config.appVersion, "dist"), { recursive: true });
+      await symlink(path.join(config.releasesDir, config.appVersion), config.currentLink);
+      database = openDatabase(config);
+      const staleAt = Date.now() - ORPHANED_UPDATE_TIMEOUT_MS - 1;
+      const insert = database.sqlite.prepare(`
+        INSERT INTO update_jobs(id, operation, status, version, platform, asset_url, created_at, updated_at)
+        VALUES (?, ?, ?, ?, ?, ?, ?, ?)
+      `);
+      const queuedId = randomUUID();
+      const applyingId = randomUUID();
+      const stagedId = randomUUID();
+      insert.run(queuedId, "apply", "queued", "1.2.0", "linux-x64", "https://updates.example/queued.tar.gz", staleAt, staleAt);
+      insert.run(applyingId, "apply", "applying", config.appVersion, "linux-x64", "https://updates.example/applying.tar.gz", staleAt, staleAt);
+      insert.run(stagedId, "download", "staged", "1.2.0", "linux-x64", "https://updates.example/staged.tar.gz", staleAt, staleAt);
+      const now = Date.now();
+      expect(reconcileOrphanedUpdateJobs(database.sqlite, config, now)).toBe(2);
+      expect(database.sqlite.prepare("SELECT status FROM update_jobs WHERE id=?").get(queuedId)).toEqual({ status: "failed" });
+      expect(database.sqlite.prepare("SELECT status FROM update_jobs WHERE id=?").get(applyingId)).toEqual({ status: "completed" });
+      expect(database.sqlite.prepare("SELECT status FROM update_jobs WHERE id=?").get(stagedId)).toEqual({ status: "staged" });
+    } finally {
+      if (database) database.sqlite.close();
+      await rm(root, { recursive: true, force: true });
+    }
+  });
+
   it("流式解包在展开大小上限前拒绝高压缩比归档，并修正发布树权限", async () => {
     const root = await mkdtemp(path.join(tmpdir(), "tallynote-update-stream-"));
     try {