memory-eternal 0.10.16 → 0.10.17

package/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "$schema": "https://anthropic.com/claude-code/plugin.schema.json",
   "name": "memory-eternal",
-  "version": "0.10.16",
+  "version": "0.10.17",
   "description": "记忆核心（Memory Eternal）：对话结束后自动沉淀知识卡到本地记忆库（~/.dsh/memory-vault），支持检索 / 召回 / 审核中心。",
   "author": {
     "name": "EternalNight996",

package/.codex-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "memory-eternal",
-  "version": "0.10.16",
+  "version": "0.10.17",
   "description": "记忆核心（Memory Eternal）：对话结束后自动沉淀知识卡到本地记忆库（~/.dsh/memory-vault），支持检索 / 召回 / 审核中心。",
   "author": {
     "name": "EternalNight996",

package/.cursor-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "memory-eternal",
-  "version": "0.10.16",
+  "version": "0.10.17",
   "description": "记忆核心（Memory Eternal）：对话结束后自动沉淀知识卡到本地记忆库（~/.dsh/memory-vault），支持检索 / 召回 / 审核中心。",
   "author": {
     "name": "EternalNight996",

package/README.en.md

@@ -361,14 +361,15 @@ curl http://127.0.0.1:7999/memory-eternal/api/web-info    # is the standalone we
 
 | Version | Date | Highlights |
 |---|---|---|
+| **v0.10.17** | 2026-10-12 | **Root fix for the "resident is serving but nobody manages it" self-contradiction (issue #28)**. The scene: after upgrading to 0.10.16 the reporter had no resident instance at all, and the panel showed **neither a restart button nor any hint** (the banner only rendered when `vc.stale`, which is false when the host and the port report the same version), only an "auto-capture anomaly"; reproduced the other half locally: after `stop` the port was **still serving**, the lock was empty, and `status` said "no registered watchdog" — because that web had been spawned directly by the host's `ensureWeb()` (no `webPid` in the lock) and `stop` only killed the watchdog. Five fixes: ① **`/version-check` now reports `resident.managed`** (a live watchdog in the lock is what makes it managed) — "serving" is not "managed"; ② **panel**: both the **not-running** and the **unmanaged** resident show a restart button with dedicated wording ("not running" / "unmanaged") instead of only the version-drift case; ③ **`/restart-self`**: `alreadyCurrent` now requires a version match **and** a live manager, otherwise it replaces as usual (an orphan web does not get to stay just because its version is new); ④ **an explicit `dsh-memory stop` frees the port** (`forcePortOccupant: true`, same gate as ever: only processes `looksLikeOurWeb` recognises are terminated); ⑤ **the self-check now demands a fresh takeover**: `restartResident` no longer accepts "the port serves the expected version" (before and after an upgrade the version is often identical, so the **old** instance fooled the check — measuring "self-check passed (waited 0s)" immediately followed by `status` reporting no registered watchdog); the new `waitForTakeover()` requires a matching version **and** a live watchdog whose pid is not in the pre-replacement set, and failure reasons are split accordingly into "wrong version" vs "nobody took over". Measured on this machine in the same scenario: `41672/17432` before → `28136/31976` after, `restart` waited 13s and passed, and the **very next `status` showed the new watchdog registered** (no contradiction). `npm test` now runs **310 checks** |
 | **v0.10.16** | 2026-10-09 | **`--reap` no longer says "nothing was reaped" when it did reap something** (reported by alario-tang, measured on macOS with 0.10.13). The scene: after `kill -9` on the watchdog leaves a `PPID=1` orphan web, `dsh-memory watchdog --reap --port 7999` really freed the port and really killed the web, yet the screen read "**scanned 1 watchdog process, cleaned 0 orphans**" — because 0.10.13 made `scanned` include webs while `webKilled` was never printed. Fix: a pure `formatReapSummary()` unifies the wording (scan counts split into `watchdog A / web B`, cleaned entries listed per kind with pids), `reapStaleWatchdogs` gains `scannedWatchdogs` / `scannedWebs` (`scanned` stays their sum for older callers), and the CLI uses it. Two checks added (never say "cleaned 0 orphans" when a web was reaped / scan counts must be split), and `tests/watchdog-orphan-web.test.mjs` now records the reproduction note: **such an orphan can only be created with `kill -9` — SIGTERM runs the watchdog's graceful exit, which also stops the web it spawned** (the same coin as #19's "on Windows SIGTERM is unconditional termination"). `npm test` now runs **308 checks** |
-| **v0.10.15** | 2026-10-09 | **Three loose ends left by the upgrades.** ① **Recommended plan A no longer reinstalls the call we just saved**: v0.10.12 raised the default distiller output cap to 4000 (removing the wasted "hit 2000, then re-run the same candidate at 4000"), but the panel one-click preset "A (light)" still hard-coded `captureMaxTokens: 2000` — one click brought the retry back; it now matches the default, guarded by a source check (`tests/csrf-origin.test.mjs` asserts planA cap === default). ② **The README "stall watchdog" description now matches the new rule** (both languages): the old wording said "if no turn-stopping event arrives for 15 minutes", while it is actually "**zero stops since this process started** plus turns still starting inside the last 20 minutes → record a `warn`", and it now states that an in-flight **long turn is never mistaken for a stall**. ③ While at it, the other log wording `轮次收尾事件未触发（agent/turn-stopping 没到）` was verified to have last fired on **2026-09-24** — a historical remnant of the old heuristic with no such code path left today (nothing to fix). `npm test` now runs **306 checks** |
 
 <details>
-<summary>Older release notes (v0.10.14 and earlier · 41 releases, click to expand)</summary>
+<summary>Older release notes (v0.10.15 and earlier · 42 releases, click to expand)</summary>
 
 | Version | Date | Highlights |
 |---|---|---|
+| **v0.10.15** | 2026-10-09 | **Three loose ends left by the upgrades.** ① **Recommended plan A no longer reinstalls the call we just saved**: v0.10.12 raised the default distiller output cap to 4000 (removing the wasted "hit 2000, then re-run the same candidate at 4000"), but the panel one-click preset "A (light)" still hard-coded `captureMaxTokens: 2000` — one click brought the retry back; it now matches the default, guarded by a source check (`tests/csrf-origin.test.mjs` asserts planA cap === default). ② **The README "stall watchdog" description now matches the new rule** (both languages): the old wording said "if no turn-stopping event arrives for 15 minutes", while it is actually "**zero stops since this process started** plus turns still starting inside the last 20 minutes → record a `warn`", and it now states that an in-flight **long turn is never mistaken for a stall**. ③ While at it, the other log wording `轮次收尾事件未触发（agent/turn-stopping 没到）` was verified to have last fired on **2026-09-24** — a historical remnant of the old heuristic with no such code path left today (nothing to fix). `npm test` now runs **306 checks** |
 | **v0.10.14** | 2026-10-09 | **The "auto-capture anomaly: turn-stop event may be broken" false alarm is fixed at the root.** The scene: during a long turn lasting tens of minutes, `agent/inbox/claimed` fires **repeatedly inside one turn**, while `turn-stopping` only arrives when the turn really ends — so any 20-minute window necessarily shows "≥3 starts, 0 stops", and the check mistook an **in-flight** turn for a renamed event, pushing a red alert at the user (the "39/18" line measured in this session came from exactly that). Fix: one hard precondition — "the event was renamed/removed" is a **process-lifetime** property, not a window property: if this run has seen even one stop, the wiring works and 0 stops in the window merely means the current turn is still running (`evaluateStall` gains `everStopped` and reports `reason: wiring-proven | no-stop-ever`); only "zero stops since this process started **and** continuous starts inside the window" alerts. At the same time the alert is downgraded from `fail` to `warn` — an inference no longer paints capture health red nor masquerades as an auto-capture anomaly (real failures still turn it red on their own). Three checks in `tests/stall.test.mjs` lock the old and new behaviour (long turn stays quiet / a real rename alerts / `everStopped` injectable). `npm test` now runs **305 checks** |
 | **v0.10.13** | 2026-10-09 | **Orphan `web.js` processes are finally reachable (#11 item 1, the POSIX branch)**. The scene (alario-tang, macOS arm64): the watchdog died first and the lock file was gone, leaving a stale web on port 7999 with `PPID=1`; `stop`/`restart` read lock slots and found no target, `watchdog --reap` only matched `watchdog.js`, and the panel button rides the same path — so drift could be reported but never healed, and the user had to `kill` by hand. Two concrete gaps: ① **on POSIX the port occupant only came back as a process name** — `lsof -Fpcn` yields `node` for the `c` field rather than a command line, while the "should we reap it" test (`looksLikeOurWeb`) requires `memory-eternal` or this install web.js absolute path in the command line → always false → the fallback loops in `stopWatchdogs` / `clearResidentForTakeover` **never killed anything** (Windows uses `Get-CimInstance`, which returns the full command line, so this box never showed it); now a follow-up `ps -p <pid> -o args=` fills in the command line, falling back to the old behaviour when unavailable (better to miss than to misfire), with an injectable `platform` for tests; ② **`--reap` now covers our own webs**: a `listWebs` branch reaps "same port + our web + **its keeper is gone** (that slot watchdog is not alive) + outside the protected set (keepPid / current pid / parent / a `--reap` sibling)", while **a web whose keeper is alive is never touched**, and other ports or foreign `web.js` processes are never killed. Adds `tests/watchdog-orphan-web.test.mjs` (5 checks) and makes two watchdog-only cases state `listWebs: []` explicitly (otherwise reap enumerates the real process table). `npm test` now runs **302 checks** |
 | **v0.10.12** | 2026-10-09 | **Two quiet things that either burn money or leave a hole every day.** ① **Default distiller output cap: 2000 → 4000 (the hard ceiling)**: the logs kept showing "hit the maxTokens=2000 cap → retry the same candidate at 4000", and each occurrence **wasted one LLM call** (hit the cap, then re-run the same candidate at 4000). `maxTokens` is only a ceiling and billing follows what is actually generated, so the default now sits at the hard ceiling: those turns finish in one call; `maxTokenLadder` stays for users who **explicitly** lower `captureMaxTokens`. ② **A cross-site write gate on the plugin API**: `/memory-eternal/api/*` is unauthenticated, so any local web page could POST `/config`, `/audit/approve`, `/write`… New `crossOriginError()` — state-changing requests (non-GET/HEAD/OPTIONS) must be **same-origin with the Host of the request itself**: browsers always send `Origin` on cross-site requests (a mismatch returns 403 `CROSS_ORIGIN_BLOCKED`), while same-origin pages (DSH settings/memory page, the standalone page) and LAN access stay unaffected, curl / CLI / MCP calls without `Origin` keep working, and desktop-shell schemes like `app://` are allowed so nothing breaks. Both the intercepted host routes (`/config` POST, `/restart-self`, `/mcp/action`…) and the shared api entry carry the gate. Adds `tests/csrf-origin.test.mjs` (4 checks: decision matrix incl. LAN / no Origin / null / app://, a cross-site POST getting 403 without reaching business logic, both entry points present, and the default cap equal to the hard ceiling with no ladder retry). `npm test` now runs **297 checks** |

package/README.md

@@ -393,14 +393,15 @@ curl http://127.0.0.1:7999/memory-eternal/api/web-info    # 独立 web 是否活
 
 | 版本 | 日期 | 关键改动 |
 |---|---|---|
+| **v0.10.17** | 2026-10-12 | **「常驻实例在服务、却没人托管」这类状态自相矛盾的根治（issue #28）**。现场：用户升级到 0.10.16 后常驻实例没起来，面板**既没有重启按钮也没有任何提示**（横幅只在 `vc.stale` 时渲染，而宿主与端口版本号一致时 `stale` 为假），只看到一条「自动沉淀异常」；本机复现出另一半：`stop` 之后端口上**还活着**、锁里却空了、`status` 说「没有登记的看门狗」—— 因为端口上那个 web 是宿主 `ensureWeb()` 直接拉起的（锁里没有 `webPid`），`stop` 只杀了看门狗。五处一起修：① **`/version-check` 报出 `resident.managed`**（锁里有活着的看门狗才算被托管）—— 「在服务」≠「被托管」；② **面板**：常驻实例**没在跑**或**没被托管**时都显示重启按钮与专门文案（「未运行」/「未托管」），不再只认「版本落后」这一种；③ **`/restart-self`**：`alreadyCurrent` 必须同时满足「版本一致 **且** 被托管」，否则照常替换（孤儿 web 版本再新也得重新托管）；④ **显式 `dsh-memory stop` 把端口腾出来**（`forcePortOccupant: true`，闸门不变：只终止 `looksLikeOurWeb` 认得出是自家的进程）；⑤ **自检改为「新实例接管」**：`restartResident` 不再只看「端口服着期望版本」（升级前后版本号常常相同、旧实例本就是这个版本 → 自检会被**旧实例**骗过，实测打印「自检通过（等待 0s）」后紧接着 `status` 说没有登记的看门狗），新增 `waitForTakeover()`：必须「版本一致 **且** 锁里出现一个不在替换前 pid 集合里的活看门狗」才算成功，失败原因也据此分离成「版本不对」与「没人接管」。本机同一现场实测：替换前 `41672/17432` → 替换后 `28136/31976`，`restart` 等待 13s 后自检通过，**紧接着的 `status` 立刻显示新看门狗已登记**（不再矛盾）。`npm test` 共 **310 项** |
 | **v0.10.16** | 2026-10-09 | **`--reap` 的输出不再「明明收了却说没收」**（alario-tang 反馈，macOS + 0.10.13 实测）。现场：`kill -9` 掉 watchdog 造出 `PPID=1` 的孤儿 web 后跑 `dsh-memory watchdog --reap --port 7999`，端口确实被释放、web 确实被收掉，屏幕上却写着「**扫描到 1 个 watchdog 进程，清理孤儿 0 个**」—— 因为 0.10.13 让 `scanned` 含了 web、而 `webKilled` 一个字都没打印。修法：新增纯函数 `formatReapSummary()` 统一口径（扫描数拆成 `watchdog A / web B`、清理项按类分别列出 pid），`reapStaleWatchdogs` 返回值补 `scannedWatchdogs` / `scannedWebs`（`scanned` 仍是两者之和，兼容旧调用方），CLI 改用它。测试补 2 项（收了 web 时不许再说「清理孤儿 0 个」/ 扫描计数必须拆开），并在 `tests/watchdog-orphan-web.test.mjs` 头部记下复现要点：**这类孤儿只能用 `kill -9` 造 —— SIGTERM 会走 watchdog 的优雅退出、把它拉起的 web 一起停掉**（与 #19 里「Windows 上 SIGTERM 即无条件终止」是同一枚硬币的两面）。`npm test` 共 **308 项** |
-| **v0.10.15** | 2026-10-09 | **三处「升级留下的尾巴」**。① **推荐方案 A 不再把刚省掉的调用装回来**：v0.10.12 把「蒸馏输出上限」默认提到 4000（省掉「先撞 2000、再以 4000 重跑同一候选」那次白烧），但面板里「一键推荐配置 A（轻量）」仍写死 `captureMaxTokens: 2000` —— 用户点一下就又把重试装回来；现已与默认对齐，并加源码守卫（`tests/csrf-origin.test.mjs` 断言 planA 的上限 === 默认值）。② **README 的「停滞兜底」描述与新判据对齐**（中英两份）：旧文案写「15 分钟收不到收尾事件就报」，现在是「**本次运行至今 0 收尾** + 最近 20 分钟内持续有轮次开始 → 记一条 `warn`」，并写明正在进行的**长回合不会被误判**。③ 顺带核实日志里另一种措辞 `轮次收尾事件未触发（agent/turn-stopping 没到）` 最后出现于 **2026-09-24**，属于旧判据的历史残留，现行代码里已无这条路径（无需修复）。`npm test` 共 **306 项** |
 
 <details>
-<summary>更早的更新日志（v0.10.14 及以前 · 共 41 版，点开查看）</summary>
+<summary>更早的更新日志（v0.10.15 及以前 · 共 42 版，点开查看）</summary>
 
 | 版本 | 日期 | 关键改动 |
 |---|---|---|
+| **v0.10.15** | 2026-10-09 | **三处「升级留下的尾巴」**。① **推荐方案 A 不再把刚省掉的调用装回来**：v0.10.12 把「蒸馏输出上限」默认提到 4000（省掉「先撞 2000、再以 4000 重跑同一候选」那次白烧），但面板里「一键推荐配置 A（轻量）」仍写死 `captureMaxTokens: 2000` —— 用户点一下就又把重试装回来；现已与默认对齐，并加源码守卫（`tests/csrf-origin.test.mjs` 断言 planA 的上限 === 默认值）。② **README 的「停滞兜底」描述与新判据对齐**（中英两份）：旧文案写「15 分钟收不到收尾事件就报」，现在是「**本次运行至今 0 收尾** + 最近 20 分钟内持续有轮次开始 → 记一条 `warn`」，并写明正在进行的**长回合不会被误判**。③ 顺带核实日志里另一种措辞 `轮次收尾事件未触发（agent/turn-stopping 没到）` 最后出现于 **2026-09-24**，属于旧判据的历史残留，现行代码里已无这条路径（无需修复）。`npm test` 共 **306 项** |
 | **v0.10.14** | 2026-10-09 | **「自动沉淀异常：轮次收尾事件疑似失效」误报根治**。现场：几十分钟的长回合里，`agent/inbox/claimed` 在**一轮对话中途**也会多次触发，而 `turn-stopping` 要等回合真的结束才来 —— 于是 20 分钟窗口必然「开始 ≥3、收尾 0」，判据把**还在进行中**的轮次当成「事件被改名」，直接把红警报推给用户（本会话实测那条「累计 39/18」就是这么来的）。修法：加一条硬前提 —— 「事件被改名/移除」是**进程生命周期**属性，不是窗口属性：本次运行里只要收到过哪怕一次收尾，通路就是好的，窗口内 0 收尾只说明当前轮次还没结束（`evaluateStall` 新增 `everStopped`，并在返回值里给出 `reason: wiring-proven | no-stop-ever`）；只有「本次运行至今 **0** 收尾 + 窗口内持续开始」才报警。同时把这条**推断**从 `fail` 降级为 `warn` —— 不再由它把沉淀健康状态翻红、更不冒充「自动沉淀异常」（真故障会由写卡失败自己报红）。`tests/stall.test.mjs` 新增 3 项锁住新旧行为（长回合不报 / 真改名要报 / `everStopped` 可注入）。`npm test` 共 **305 项** |
 | **v0.10.13** | 2026-10-09 | **孤儿 `web.js` 终于收得掉了（#11「待做 1」的 POSIX 分支）**。现场（alario-tang，macOS arm64）：watchdog 先死 + 锁文件丢失 → 端口 7999 上留着一个 `PPID=1` 的旧 web；`stop`/`restart` 读锁槽找不到目标、`watchdog --reap` 只认 `watchdog.js`、面板按钮又走同一条链路 → 漂移只能报不能自愈，用户被迫手动 `kill`。查出**两个具体缺口**：① **POSIX 侧端口占用者只拿到进程名** —— `lsof -Fpcn` 的 `c` 字段是 `node` 而不是命令行，而「要不要收」的判据（`looksLikeOurWeb`）要求命令行含 `memory-eternal` 或本机 `web.js` 绝对路径 → 判据恒假 → `stopWatchdogs` / `clearResidentForTakeover` 的兜底循环**一次都不杀**（Windows 走 `Get-CimInstance` 拿完整命令行，所以本机从未暴露）；现在补一次 `ps -p <pid> -o args=`，拿不到才退回旧行为（宁可不杀、也不误杀），并让 `platform` 可注入以便单测。② **`--reap` 扩到自家 web**：新增 `listWebs` 分支，把「同端口 + 是自家 web + **托管者已死**（该槽 watchdog 不存活）+ 不在保护集合（keepPid / 当前 pid / 父进程 / 带 `--reap` 的同类）」的 web 一并回收；**托管者还活着的 web 一律不动**，别的端口、别人的 `web.js` 一个不杀。新增 `tests/watchdog-orphan-web.test.mjs`（5 项），并把两条只谈 watchdog 的旧用例显式声明 `listWebs: []`（否则 reap 会去列本机真实进程表）。`npm test` 共 **302 项** |
 | **v0.10.12** | 2026-10-09 | **两件「不显眼但每天都在扣钱 / 留洞」的事**。① **蒸馏输出上限默认 2000 → 4000（= 硬顶）**：日志实测高频出现「输出撞到 maxTokens=2000 上限 → 以 4000 重试同一候选」，每一次都是**白烧一次 LLM 调用**（先撞顶、再拿 4000 重跑同一个候选）。maxTokens 只是上限、按实际生成量计费，所以默认直接放到硬顶：同样的轮次一次调用就够；`maxTokenLadder` 仍保留给**显式**把 `captureMaxTokens` 调小的用户。② **插件 API 加跨站写入闸门**：`/memory-eternal/api/*` 无鉴权，本机任意网页都能 POST `/config`、`/audit/approve`、`/write`… 新增 `crossOriginError()` —— 状态变更请求（非 GET/HEAD/OPTIONS）必须与请求自身的 `Host` **同源**：浏览器跨站请求一定带 `Origin`（与 Host 不同源即 403 `CROSS_ORIGIN_BLOCKED`）；同源页面（DSH 设置页/记忆页、独立页）与局域网访问都同源不受影响；没有 Origin 的 curl / CLI / MCP 照常放行；`app://` 这类桌面壳自定义 scheme 也放行以免误伤。宿主自己的拦截路由（`/config` POST、`/restart-self`、`/mcp/action`…）与共享 api 入口**都**装了这道闸门。新增 `tests/csrf-origin.test.mjs`（4 项：判定矩阵含局域网 / 无 Origin / null / app://、跨站 POST 被 403 且不进业务逻辑、两处入口都在、默认上限=硬顶且无 ladder 重试）。`npm test` 共 **297 项** |

package/bin/dsh-memory.mjs

@@ -167,7 +167,10 @@ async function main() {
       const { stopWatchdogs } = await import('../lib/watchdog.js')
       const port = argOf('--port')
       // checkPort：停完再问一次「端口上还有谁在服务」，把「停了却还占着端口」显式说出来（#23）
-      const out = await stopWatchdogs({ port: port || undefined, checkPort: true })
+      // forcePortOccupant：**显式 stop 就该把端口腾出来** —— 端口上那个自家 web 很可能不是看门狗拉起的
+      // （宿主 ensureWeb 直接起、或看门狗被 kill -9 后残留的孤儿），锁里没有 webPid 就谁也收不到它。
+      // 闸门没松：只有 looksLikeOurWeb(命令行) 为真才会终止，认不出的照旧只告警。
+      const out = await stopWatchdogs({ port: port || undefined, checkPort: true, forcePortOccupant: true })
       if (has('--json')) { console.log(JSON.stringify({ ok: out.failed.length === 0, ...out }, null, 2)); return }
       if (!out.stopped.length && !out.failed.length && !out.webStopped.length) console.log('没有需要停止的常驻看门狗（锁里没有活着的实例）。')
       if (out.stopped.length) console.log(`已停止看门狗：pid ${out.stopped.join(', ')}`)

package/index.js

@@ -1366,7 +1366,9 @@ export function apply(ctx, config) {
             // 自愈刚把实例挪到配置端口：立刻收敛 web-info，别再让面板指向被收掉的那个端口（issue #29）
             if (out.ok) refreshWebInfo({ url: `http://127.0.0.1:${wdPort}`, port: wdPort, spawned: false })
             console.error(`[memory-eternal] ${done}`)
-            logCapture('system', out.ok ? 'boot' : 'fail', done)
+            // 用 warn 不用 fail：这是**服务管理**的失败（常驻实例没起来），不是沉淀管线坏了。
+            // 过去这里记 fail 会把沉淀健康状态翻红，用户看到的是「自动沉淀异常」（实测 #28 的现场）。
+            logCapture('system', out.ok ? 'boot' : 'warn', done)
             return
           }
 

package/lib/api.js

@@ -20,7 +20,7 @@ import { buildDiagnostics } from './feedback.js'
 import { getSetupStatus, runSetup, mcpAgentAction } from './setup.js'
 import { backupDb } from './db.js'
 import { createHub } from './sse.js'
-import { probeServedVersion, restartResident, currentPkgVersion } from './watchdog.js'
+import { probeServedVersion, restartResident, currentPkgVersion, readWatchdogLock, isPidAlive } from './watchdog.js'
 
 // 独立 web 的 SSE 中心：配置/数据变更即时推给已打开的页面（无需手动刷新）
 const hub = createHub()
@@ -251,13 +251,21 @@ export function createApi(deps = {}) {
         const residentPort = selfPort() || Number(cfgVc.webPort) || Number(process.env.MEMORY_WEB_PORT) || 7999
         let residentServed = ''
         try { residentServed = String((await probeResidentVersion(residentPort)) || '') } catch { residentServed = '' }
+        // 「在服务」≠「被托管」：web 可能是宿主 ensureWeb() 直接拉起的、或看门狗被 stop/kill 后残留的孤儿
+        // —— 此时锁里没有活着的看门狗，端口上的实例没人保活、也不会随版本升级被替换。
+        // #28 的现场就是这一种：host 与端口版本号一致 → 面板认为「无需重启」，而其实该重启（重新托管）。
+        let residentManaged = false
+        try {
+          residentManaged = readWatchdogLock(process.env).watchdogs
+            .some((w) => Number(w.port) === Number(residentPort) && isPidAlive(w.pid))
+        } catch { residentManaged = false }
         json(res, 200, {
           ok: true,
           loaded,
           onDisk,
           latest,
           checkError,
-          resident: { port: residentPort, served: residentServed },
+          resident: { port: residentPort, served: residentServed, managed: residentManaged },
           stale: !!(loaded && onDisk && loaded !== onDisk),
           // 只在 **npm 上确实更新** 时提示升级：本地跑着未发布版本（磁盘 0.10.7 / npm 0.10.6）时
           // 旧的 `latest !== onDisk` 会反过来劝用户装回旧版。
@@ -280,10 +288,15 @@ export function createApi(deps = {}) {
         const port = Number(query.get('port')) || selfPort() || Number(cfgR.webPort) || Number(process.env.MEMORY_WEB_PORT) || 7999
         const expect = pkgVersion()
         const served = await probeResidentVersion(port).catch(() => '')
-        if (served && expect && String(served) === String(expect)) {
+        // 「版本一致」还不够：还要**有活着的看门狗托管它**。孤儿 web（看门狗被 stop / 被 kill -9 后残留）
+        // 版本号可能正好也是新的，但没人保活、也不会被后续升级替换 —— 那时一键重启才是唯一出路（#28）。
+        const managed = (() => {
+          try { return readWatchdogLock(process.env).watchdogs.some((w) => Number(w.port) === Number(port) && isPidAlive(w.pid)) } catch { return false }
+        })()
+        if (served && expect && String(served) === String(expect) && managed) {
           return json(res, 200, {
-            ok: true, alreadyCurrent: true, port, served: String(served), expect,
-            note: `常驻实例（端口 ${port}）已经在服务 v${served}，无需重启。若「运行中」仍显示旧版本号，说明那一行来自另一个进程（例如 DSH 宿主），需重启桌面版 / DSH。`,
+            ok: true, alreadyCurrent: true, port, served: String(served), expect, managed,
+            note: `常驻实例（端口 ${port}）已经在服务 v${served} 且有看门狗保活，无需重启。若「运行中」仍显示旧版本号，说明那一行来自另一个进程（例如 DSH 宿主），需重启桌面版 / DSH。`,
           })
         }
         if (residentRestarting) {
@@ -304,7 +317,8 @@ export function createApi(deps = {}) {
               try {
                 const { appendCaptureLog } = await import('./capture-log.js')
                 await appendCaptureLog({
-                  time: Date.now(), sessionId: 'system', action: out && out.ok ? 'boot' : 'fail',
+                  // 同上：服务管理失败用 warn（别翻红沉淀健康状态、别冒充「自动沉淀异常」）
+                  time: Date.now(), sessionId: 'system', action: out && out.ok ? 'boot' : 'warn',
                   reason: out && out.ok
                     ? `面板一键重启常驻实例成功：端口 ${port} 正在服务 v${out.served}`
                     : `面板一键重启常驻实例未确认生效：${(out && out.error) || '未知原因'}`,