memory-eternal 0.10.9 → 0.10.11

package/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "$schema": "https://anthropic.com/claude-code/plugin.schema.json",
   "name": "memory-eternal",
-  "version": "0.10.9",
+  "version": "0.10.11",
   "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": "0.10.9",
+  "version": "0.10.11",
   "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": "0.10.9",
+  "version": "0.10.11",
   "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 +361,16 @@ curl http://127.0.0.1:7999/memory-eternal/api/web-info    # is the standalone we
 
 | Version | Date | Highlights |
 |---|---|---|
-| **v0.10.9** | 2026-10-09 | **Root fix for the upgrade-window port drift that turned the memory page into "connection refused" (issue #29)**. The scene: a detached instance left by the old package sat on 7999 without reporting a version → the host `ensureWebServer()` could not recognise it, drifted to 8000 and wrote that into `web-info`; the plugin's self-heal then reaped the instance on 8000 as a stray → `web-info` dangled on a dead port (the page loaded once, then "connection refused 127.0.0.1" after a refresh). ① **The library page now uses a host same-origin shell**: the host gains `/memory-eternal/ui/app` (same index.html as the config page `/ui/config`, where `<script src="app.js">` resolves to `/ui/app.js`), so the iframe address is **decoupled from whichever port the resident instance listens on** (an older host still falls back to `web-info`). ② **`/web-info` converges live**: before answering, it probes "configured port → recorded port", uses whichever is serving, and reports `alive:false` when neither is (no more advertising a dead address); it is also rewritten right after a successful self-heal restart and whenever the interval keeper sees the instance alive. ③ **No more blind port drift**: new `looksLikeOurWeb()` (lenient detection of "ours but not reporting a version": any `ok:true` from /overview, /web-info or /budget) gives self-heal a 6s takeover window before drifting, and logs it when it does drift. ④ **Backoff retry in the health check**: `inspectResident()` no longer concludes "old instance that does not report its version" from a single missed probe (that verdict triggers a needless replacement on every start), it retries with backoff first. Adds `tests/issue-fixes-5.test.mjs` (4 checks). `npm test` now runs **287 checks** |
-| **v0.10.8** | 2026-10-08 | **Saving on the standalone web page now takes effect on click: the pending apply moves from a plugin callback into a host HTTP request (design A, measurement-driven)**. The scene: a standalone-page save ended in a `dropped` patch after five retries, always with `HMR transactions cannot be nested` (0.10.7 only stopped it from masquerading as an auto-capture anomaly and filtered upgrade-window keys — the sync itself was not fixed). **Measured on the same host with the same `settings.update`**: called from a plugin **callback** (`setInterval` / `fs.watch`) → guard error and **nothing was committed at all** (a patch with only host-known keys and a genuinely different value failed the same way); called from a plugin **HTTP request handler** → **it really committed** (`recallSummaryLen` 130→132→130, `revision` 4→6). So the fix is not another API but another **calling context**: ① the host gains `POST /memory-eternal/api/drain-pending`, applying the pending patch inside a **request handler** (two gates: loopback origin plus a per-process token from the heartbeat file, so no other local page can write); ② the host publishes its web server's **listening port + token** in the heartbeat (`webServer.port` is a public getter), and a standalone save now **wakes the host** and reports the real result (new `savedApplied` toast); ③ a failed wake-up (old host without the route → 404, port mismatch, timeout) is **not** an error — it falls back to the 0.10.5 waiting path and the "change it under DSH Settings" hint. Adds `tests/host-drain.test.mjs` (6 checks: loopback, the three gates, the end-to-end wake-up, failure reporting, 404 fallback, source guards) plus `tests/e2e-host-drain.mjs` (2 checks: a **real spawned process** end-to-end "standalone save → wake the host → host applies → reports applied", and the fallback to "queued" when the heartbeat comes from an old host with no port/token). `npm test` now runs **283 checks** |
+| **v0.10.11** | 2026-10-09 | **One more layer of tolerance in distiller-output parsing: unescaped quotes inside string values (`Expected ',' or '}' after property value`)**. The scene: `UNPARSEABLE_OUTPUT` dominates the `fail` log (103 of the last 600 entries); the freshest one carries the full picture — `JSON.parse 原始报错：Expected ',' or '}' after property value in JSON at position 1913`, with an excerpt whose body reads `运行 "npm test" 时…`: the model wrote a quote into the string as content, so the parser took it for the string terminator. That is a **different** class from the bare control characters fixed in #24 (a control character is illegal inside a string, a quote merely needs escaping). Fix: new `repairUnescapedQuotes()` — a `"` only ends the string when the next non-space character is `:`, `,`, `}`, `]` or end-of-text, otherwise it is escaped as content; it participates **only as the last candidate** and the result must still pass the card health check (usable title + body ≥20 chars) — better to fail than to write a bad card. Guarded by `tests/capture-json-repair.test.mjs` (4 checks: identity on valid JSON, the live unescaped-quote case, control chars + quotes combined, truncation still failing honestly). `npm test` now runs **293 checks** |
+| **v0.10.10** | 2026-10-09 | **The review centre can finally show card bodies, and you can approve/reject in place (issue #27)**. The scene (thinrflbtlm): the review centre listed only "title / kind / source / severity / reason" and **never the body**; and since v0.10.0 pending cards live physically quarantined, while the library lists approved cards only — so "no review without manual confirmation" made auditing pointless. ① Every row gains a ▸/▾ toggle (clicking the title works too) that **expands the frontmatter + body in place**, rendered with the **same Markdown pipeline as the card reader** (`renderMd` + `splitFrontmatter`), not raw text; ② the body is fetched **on demand** (`GET /card?path=…&status=pending|rejected`; the server still only allows paths that really sit in the queue, unknown paths keep returning 403 `CARD_NOT_APPROVED`), so the list endpoint never ships a whole store of bodies; ③ the expanded block carries **✓ approve / ✕ reject** directly (the recycle-bin tab uses ✓ restore / 🗑 delete), while bulk selection and the toolbar stay as they were; ④ approving/rejecting collapses and refreshes automatically, with both language dictionaries updated. Adds `tests/issue-fixes-6.test.mjs` (2 checks: a pending body is readable and leaves the queue once approved, plus source guards for on-demand fetching and the deny code). `npm test` now runs **289 checks** |
 
 <details>
-<summary>Older release notes (v0.10.7 and earlier · 34 releases, click to expand)</summary>
+<summary>Older release notes (v0.10.9 and earlier · 36 releases, click to expand)</summary>
 
 | Version | Date | Highlights |
 |---|---|---|
+| **v0.10.9** | 2026-10-09 | **Root fix for the upgrade-window port drift that turned the memory page into "connection refused" (issue #29)**. The scene: a detached instance left by the old package sat on 7999 without reporting a version → the host `ensureWebServer()` could not recognise it, drifted to 8000 and wrote that into `web-info`; the plugin's self-heal then reaped the instance on 8000 as a stray → `web-info` dangled on a dead port (the page loaded once, then "connection refused 127.0.0.1" after a refresh). ① **The library page now uses a host same-origin shell**: the host gains `/memory-eternal/ui/app` (same index.html as the config page `/ui/config`, where `<script src="app.js">` resolves to `/ui/app.js`), so the iframe address is **decoupled from whichever port the resident instance listens on** (an older host still falls back to `web-info`). ② **`/web-info` converges live**: before answering, it probes "configured port → recorded port", uses whichever is serving, and reports `alive:false` when neither is (no more advertising a dead address); it is also rewritten right after a successful self-heal restart and whenever the interval keeper sees the instance alive. ③ **No more blind port drift**: new `looksLikeOurWeb()` (lenient detection of "ours but not reporting a version": any `ok:true` from /overview, /web-info or /budget) gives self-heal a 6s takeover window before drifting, and logs it when it does drift. ④ **Backoff retry in the health check**: `inspectResident()` no longer concludes "old instance that does not report its version" from a single missed probe (that verdict triggers a needless replacement on every start), it retries with backoff first. Adds `tests/issue-fixes-5.test.mjs` (4 checks). `npm test` now runs **287 checks** |
+| **v0.10.8** | 2026-10-08 | **Saving on the standalone web page now takes effect on click: the pending apply moves from a plugin callback into a host HTTP request (design A, measurement-driven)**. The scene: a standalone-page save ended in a `dropped` patch after five retries, always with `HMR transactions cannot be nested` (0.10.7 only stopped it from masquerading as an auto-capture anomaly and filtered upgrade-window keys — the sync itself was not fixed). **Measured on the same host with the same `settings.update`**: called from a plugin **callback** (`setInterval` / `fs.watch`) → guard error and **nothing was committed at all** (a patch with only host-known keys and a genuinely different value failed the same way); called from a plugin **HTTP request handler** → **it really committed** (`recallSummaryLen` 130→132→130, `revision` 4→6). So the fix is not another API but another **calling context**: ① the host gains `POST /memory-eternal/api/drain-pending`, applying the pending patch inside a **request handler** (two gates: loopback origin plus a per-process token from the heartbeat file, so no other local page can write); ② the host publishes its web server's **listening port + token** in the heartbeat (`webServer.port` is a public getter), and a standalone save now **wakes the host** and reports the real result (new `savedApplied` toast); ③ a failed wake-up (old host without the route → 404, port mismatch, timeout) is **not** an error — it falls back to the 0.10.5 waiting path and the "change it under DSH Settings" hint. Adds `tests/host-drain.test.mjs` (6 checks: loopback, the three gates, the end-to-end wake-up, failure reporting, 404 fallback, source guards) plus `tests/e2e-host-drain.mjs` (2 checks: a **real spawned process** end-to-end "standalone save → wake the host → host applies → reports applied", and the fallback to "queued" when the heartbeat comes from an old host with no port/token). `npm test` now runs **283 checks** |
 | **v0.10.7** | 2026-10-08 | **"Still the old version after starting dsh" fixed at the root: version-drift self-heal + one-click resident restart (the positive answer to #19 / #23).** The scene: an npm upgrade only swaps files on disk, and the **resident web (port 7999 by default) outlives the DSH host** — the process on the port keeps running the code it loaded at startup, while activation policy was "there is already a live watchdog on this port → `delegated` (no spawn, no replacement)", so **no number of DSH restarts ever replaces it** (measured here: the host started at 10:19 on the new build and the same-origin settings panel showed 0.10.6, while the web on 7999 had started at 08:15 on 0.10.0 — `/config` and `/version-check` each reported a different version). ① **Startup self-heal (`autoRestartOnDrift`, default on)**: activation first runs `inspectResident()` — the version is always asked from **the port** (the lock's `pkgVersion` is only the watchdog's self-report, and an old instance wrote an empty string, so drift was invisible) — then `decideResidentAction()` decides: match → `delegate` (keeping #19's honest wording), drift → `restart`, own version unreadable → touch nothing, just replaced → debounce (a config change re-runs the effect). **"Updating the running program" is therefore defined as "restart that resident process on the new code", never as hot-swapping modules in memory**: ESM module cache plus closure state (settings / vault routing / SSE hub / `fs.watch` / the already-bound port) leaves no safe way to hot-swap, and a cache-busting re-import only produces a second module graph (double instances, double listeners, double writes). ② **One implementation for replacement**: new `restartResident()` spawns a detached `watchdog.js --replace` helper (reap the old watchdog / old web → **wait for the port to really free up** → take the lock → pull up the new web), then verifies "the version served on the port equals the local version" (#23 suggestion 2); if the port never frees up it **aborts the takeover** instead of falling back to `port+1` (fallback instances are exactly where orphan webs are born). `dsh-memory restart`, host startup self-heal and the panel button now share it. The helper is detached with `stdio:ignore` (nobody reads its stderr), so the takeover outcome **must** go into the capture log — otherwise "restarted but not in effect" becomes yet another untraceable silent failure. ③ **One-click fix in the panel**: new `POST /memory-eternal/api/restart-self` under three hard rules — this process may well **be** the stale web being replaced, so the response is sent **before** the replacement is scheduled; if the resident instance is already current it answers `alreadyCurrent` (the common case when clicking from the DSH panel — a healthy instance is never needlessly taken down); and only one replacement may run at a time. ④ **Panel semantics disambiguated**: "running" became "**running (process serving this page)**" (the resident web on the standalone page, the host inside DSH Settings — the very reason the two panels' numbers disagreed), plus a new "**resident instance vX**" chip so it is obvious which one to restart; the drift banner offers the button plus the equivalent command, and when the resident instance is already new but this page is not, it says plainly "restart the desktop app / DSH" instead of offering a button that would do nothing. ⑤ **`updateAvailable` now compares versions**: `latest !== onDisk` used to **advise downgrading** when running an unpublished local build (disk 0.10.7 / npm 0.10.6); it now uses `compareVersions` (with prerelease semantics, `rc.2 < rc.10`). ⑥ Also removed a duplicate `saving` key from the client dictionary (esbuild warned on every build). New `tests/watchdog-restart.test.mjs` (17 checks: decision matrix / inspection asking the port / replacement verification / non-lethal takeover cleanup / route responding before replacing + concurrency gate / version comparison). ⑦ **The README now ships as two separate files**: `README.md` is the **default** Simplified Chinese one (what npm and the DSH marketplace render) and `README.en.md` is the English mirror; both gained a language switch plus a "change one, change the other" note at the top, and the English self-diagnostics subsection that was missing; new `tests/readme-i18n.test.mjs` locks **1:1 structure** (heading-level sequence + code-block count + cross links + default-language markers) so future edits cannot land on one side only. ⑧ **`secretHint` credential hint + credential-read guard**: one line of your own convention gets injected into every session's systemPrompt so the agent looks up the **secret directory card** in memory first when it needs an API key / token (memory stores the directory only — names / locations / how to use — while **values stay in local credential stores**); `dsh-memory setup` now also registers the Claude Code `PreToolUse` guard (denies reads of `.credentials.yaml` / `.env` / `*-token` and points at the loader) and refreshes hook paths that still point at an older install. ⑨ **Three more live-site findings fixed before release**: ① the panel's **restart resident instance** button also showed when the *process serving this page* was the old one — clicking it must 404 (the old host has no such route) and the user only saw "unknown interface"; it now appears only when the resident instance on the port is behind the files on disk, and a 404/405 gets explicit "restart the desktop build / DSH, or run dsh-memory restart" guidance. ② When the standalone page (new) submits the whole form to an old host, the host drops the keys it does not know (e.g. `secretHint`) by whitelist → the read-back check can never match → five retries then `dropped`, with the user only seeing "HMR transactions cannot be nested"; the pending patch is now **filtered against the host schema** before it is applied (unknown keys take effect once the host upgrades), and "config sync failed" is kept strictly apart from "auto-capture anomaly" (`config-warn` / `config-fail`, no longer painting the capture health red). ③ `autoRestartOnDrift` now renders as **on** when the host snapshot has no such key (measured: 0.10.6 host + 0.10.7 panel → checkbox empty while the host actually treats it as on). Adds `tests/issue-fixes-4.test.mjs`. `npm test` now runs **274 checks** |
 | **v0.10.6** | 2026-10-07 | **Orphan webs are now reclaimed + a failed shared-config sync is no longer silent.** ① **Orphan webs**: the web child is spawned detached, so when a watchdog is hard-killed (on Windows `process.kill(pid,'SIGTERM')` terminates unconditionally) its exit handler never runs; if that web had **fallen back to port+1…** because the target port was taken, the lock entry naming it disappears together with the watchdog — leaving an unowned web parked on 8001/8002 (two were found on this machine, one of them 20 hours old, and `status` only probes the configured port so it could never see them). Now `startWatchdog` on startup, and `stop`/`restart` (the `checkPort` path), sweep the **fallback window** `[port+1, port+9]` and reap them; three gates must all pass (the port is inside the window / the command line **really is this plugin's web** / the pid is not any watchdog's `pid` or `webPid` in the lock), so **someone else's `node web.js` is never killed**. In addition `spawnWeb` no longer spawns a duplicate when the target port already holds our own web (the main source of fallback instances) — liveness is left to the tick probe. ② **`syncConfigFile()` no longer swallows errors**: it writes the read source shared by the standalone page / hooks / MCP, so a failed write shows up as "config changed but they can't see it" — the old code was `catch { /* silent */ }` (observed once: the shared file's mtime stayed at pre-activation while the host's own value had changed). Failures now reach stderr and the capture log as a `warn` (same reason reported once, no spam); because `logCapture` is defined later inside `apply()`, the log line is deliberately deferred by one tick to avoid the `const` TDZ. New `tests/watchdog-orphans.test.mjs` (7 checks); `npm test` now runs **245 checks** |
 | **v0.10.5** | 2026-10-07 | **Config saves from the standalone web page: no more empty promise, and no DSH install required.** ① **A host-independent write path (#21)**: if nothing consumes the pending file within 800 ms **and** no live DSH host is present (the heartbeat file `memory-eternal-config.host.json` is stale, or its pid is dead), the standalone side now **atomically merges** the change straight into the shared config, clears the pending file and answers `pendingOutcome: applied-direct` — so Codex / Claude Code / Cursor users, and anyone running only `dsh-memory serve`, get their save applied immediately (previously that path just wrote a "pending" file for a host that **does not exist**, while the UI promised "takes effect on next start" forever). When a host *is* alive the standalone side **never** writes directly: the shared file is the host's derived mirror, and a direct write would be overwritten by the next `syncConfigFile()` (a dedicated reverse test guards that boundary). ② **Failures now give directions, not just an error (#21)**: host-guard errors are recognised — `HMR transactions cannot be nested` (Cordis HMR hosts) and `root.events.emit is unavailable from a plugin activation` (dsh-tui and friends) are two phrasings of one defect family — and "this host does not let a plugin write config from its own callback context" plus where to change it instead (**DSH Settings → Memory**, or the profile patch layer) now lands in the give-up error, the diagnostics and the standalone page's hint; `POST /config` also returns `hostGuard` / `hint`. ③ **The UI no longer always says "Saved" (#21)**: the client consumes `pendingOutcome` — `queued` / `failed` use the non-success style and say "not confirmed applied yet", `applied-direct` states "no DSH host here, written straight into the shared config", and only a real apply shows "Saved". ④ Also fixed a cleanup race that could keep a host or test process **alive forever**: the heartbeat timer and the pending-file watcher are created after an async `import()`, so a dispose that happened before it resolved left them uncleaned (the heartbeat interval is additionally `unref()`ed) — this is exactly what hung `tests/settings-compat.test.mjs` when it ran `apply()` against a fake ctx. New `tests/issue-fixes-3.test.mjs` (9 checks); `npm test` now runs **238 checks** |

package/README.md

@@ -393,14 +393,16 @@ 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 |
 |---|---|---|
-| **v0.10.9** | 2026-10-09 | **升级期端口漂移把记忆页打成「拒绝连接」的根治（issue #29）**。现场：旧包遗留的 detached 实例占着 7999 且不自报版本 → 宿主 `ensureWebServer()` 认不出它、向后漂到 8000 并写进 `web-info`；而插件的自愈随后把 8000 上的实例当野实例收掉 → `web-info` 永久悬空，客户端 iframe 指向死端口（首次能进、刷新即「拒绝连接 127.0.0.1」）。① **记忆库页改走 host 同源壳**：宿主新增 `/memory-eternal/ui/app`（与配置页 `/ui/config` 同一份 index.html，其中 `<script src="app.js">` 相对路径正好解析到 `/ui/app.js`），iframe 地址从此与「常驻实例监听哪个端口」**彻底解耦**（旧宿主 404 时仍回落 `web-info`）。② **`/web-info` 实时收敛**：每次读之前先探「配置端口 → 记录端口」，谁在服务就用谁，都不通就如实 `alive:false`（不再把死地址当活的）；自愈替换成功后与 interval 保活探到活时都就地回写。③ **不再盲目漂端口**：新增 `looksLikeOurWeb()`（宽松识别「自家但不自报版本」的占用者：/overview、/web-info、/budget 任一 ok:true 即认），疑似自家时先给自愈留 6s 接管窗口，等不到才漂并留日志。④ **体检退避重试**：`inspectResident()` 一次探针打空不再直接判「旧版不自报版本」（那个结论会触发误替换，每次启动都白起一次实例），退避重试后才下结论。新增 `tests/issue-fixes-5.test.mjs`（4 项：宽松识别真/假阳性、等自愈接管不漂端口、宿主与客户端源码守卫）。`npm test` 共 **287 项** |
-| **v0.10.8** | 2026-10-08 | **独立 Web 端保存「点一下即生效」：把应用 pending 的上下文从插件回调搬到宿主 HTTP 请求（方案 A，实测驱动）**。现场：独立页保存 → 宿主连续重试 5 次后 `dropped`，报错恒为 `HMR transactions cannot be nested`（0.10.7 只修掉了它冒充「自动沉淀异常」与升级期未知键，同步本身没修）。**实测定位**（同一宿主、同一个 `settings.update`）：从插件**回调**（`setInterval` / `fs.watch`）里调 → 抛守卫错且**一个字节都没提交**（只含宿主认识的键、值真的变了，同样失败）；从插件 **HTTP 请求处理器**里调 → **真的提交**（`recallSummaryLen` 130→132→130，`revision` 4→6）。所以修法不是换 API 而是换**调用上下文**：① 宿主新增 `POST /memory-eternal/api/drain-pending`，在**请求处理器**里应用 pending（两道闸门：只接受本机来源 + 心跳里的一次性令牌，防本机任意页面乱写）；② 宿主把自己 web server 的**监听端口 + 令牌**写进心跳（`webServer.port` 是服务公开的 getter），独立端保存后据此**主动叫醒宿主**，拿到真实结果再回给界面（新增 `savedApplied`「已保存并生效」）；③ 叫醒失败（旧宿主没这条路由 404 / 端口对不上 / 超时）**不算错误** —— 回落到 0.10.5 的等待路径与「去 DSH 设置里改」的指路。新增 `tests/host-drain.test.mjs`（6 项：回环判定 / 守卫三件套 / 端到端叫醒 / 失败如实回报 / 404 回落 / 源码守卫）+ `tests/e2e-host-drain.mjs`（2 项：**真起进程**的端到端「独立页保存 → 叫醒宿主 → 宿主应用 → 回报已生效」，以及旧宿主心跳（无端口/令牌）时回落成「排队等待」）。`npm test` 共 **283 项** |
+| **v0.10.11** | 2026-10-09 | **蒸馏输出解析再补一层容错：字符串里的裸引号（`Expected ',' or '}' after property value`）**。现场：`fail` 日志里 `UNPARSEABLE_OUTPUT` 占大头（最近 600 条里 103 条），最近一条带全了现场 —— `JSON.parse 原始报错：Expected ',' or '}' after property value in JSON at position 1913`，摘录里能看到正文写着 `运行 "npm test" 时…`：模型把引号当正文直接写进字符串，于是它被当成字符串结束符。这与 #24 修过的「裸控制字符」是**不同**的一类（控制字符在字符串内非法，引号只是需要转义）。修法：新增 `repairUnescapedQuotes()` —— 一个 `"` 后面（跳过空白）紧跟 `:` `,` `}` `]` 或文本结尾才算字符串结束，否则当正文补 `\"`；它**只作为最后一个候选**参与解析，且必须仍通过卡片体检（标题合格 + 正文 ≥20 字）才被接受 —— 宁可不修，也不写坏卡。回归守卫 `tests/capture-json-repair.test.mjs`（4 项：合法 JSON 恒等变换、裸引号现场、控制字符与裸引号叠加、截断仍如实失败）。`npm test` 共 **293 项** |
+| **v0.10.10** | 2026-10-09 | **审核中心能看正文了 + 就地批准/驳回（issue #27）**。现场（thinrflbtlm）：审核中心过去只列「标题 / 类型 / 来源 / 严重度 / 原因」，**看不到正文**；而 v0.10.0 之后待审卡物理隔离在 `quarantine`、知识卡界面只列已审核卡 —— 于是「不人工确认就没法审」，审核形同虚设。① 每行新增 ▸/▾（点标题也行）**就地展开 frontmatter + 正文**，用与卡片详情**同一套 Markdown 渲染**（`renderMd` + `splitFrontmatter`），不是裸文本；② 正文**按需拉取**（`GET /card?path=…&status=pending|rejected`，服务端只放行确实在审核队列里的 path，未知 path 仍然 403 `CARD_NOT_APPROVED`），列表接口不塞整库正文；③ 展开块里直接给 **✓ 批准 / ✕ 驳回**（回收站页签是 ✓ 恢复 / 🗑 删除），批量勾选与外层工具条保持原样；④ 批/驳后自动收起并刷新，中英双语词条齐全。新增 `tests/issue-fixes-6.test.mjs`（2 项：待审正文读得到且批准后离开队列、源码守卫含「按需拉取」与「越权拒绝码」）。`npm test` 共 **289 项** |
 
 <details>
-<summary>更早的更新日志（v0.10.7 及以前 · 共 34 版，点开查看）</summary>
+<summary>更早的更新日志（v0.10.9 及以前 · 共 36 版，点开查看）</summary>
 
 | 版本 | 日期 | 关键改动 |
 |---|---|---|
+| **v0.10.9** | 2026-10-09 | **升级期端口漂移把记忆页打成「拒绝连接」的根治（issue #29）**。现场：旧包遗留的 detached 实例占着 7999 且不自报版本 → 宿主 `ensureWebServer()` 认不出它、向后漂到 8000 并写进 `web-info`；而插件的自愈随后把 8000 上的实例当野实例收掉 → `web-info` 永久悬空，客户端 iframe 指向死端口（首次能进、刷新即「拒绝连接 127.0.0.1」）。① **记忆库页改走 host 同源壳**：宿主新增 `/memory-eternal/ui/app`（与配置页 `/ui/config` 同一份 index.html，其中 `<script src="app.js">` 相对路径正好解析到 `/ui/app.js`），iframe 地址从此与「常驻实例监听哪个端口」**彻底解耦**（旧宿主 404 时仍回落 `web-info`）。② **`/web-info` 实时收敛**：每次读之前先探「配置端口 → 记录端口」，谁在服务就用谁，都不通就如实 `alive:false`（不再把死地址当活的）；自愈替换成功后与 interval 保活探到活时都就地回写。③ **不再盲目漂端口**：新增 `looksLikeOurWeb()`（宽松识别「自家但不自报版本」的占用者：/overview、/web-info、/budget 任一 ok:true 即认），疑似自家时先给自愈留 6s 接管窗口，等不到才漂并留日志。④ **体检退避重试**：`inspectResident()` 一次探针打空不再直接判「旧版不自报版本」（那个结论会触发误替换，每次启动都白起一次实例），退避重试后才下结论。新增 `tests/issue-fixes-5.test.mjs`（4 项：宽松识别真/假阳性、等自愈接管不漂端口、宿主与客户端源码守卫）。`npm test` 共 **287 项** |
+| **v0.10.8** | 2026-10-08 | **独立 Web 端保存「点一下即生效」：把应用 pending 的上下文从插件回调搬到宿主 HTTP 请求（方案 A，实测驱动）**。现场：独立页保存 → 宿主连续重试 5 次后 `dropped`，报错恒为 `HMR transactions cannot be nested`（0.10.7 只修掉了它冒充「自动沉淀异常」与升级期未知键，同步本身没修）。**实测定位**（同一宿主、同一个 `settings.update`）：从插件**回调**（`setInterval` / `fs.watch`）里调 → 抛守卫错且**一个字节都没提交**（只含宿主认识的键、值真的变了，同样失败）；从插件 **HTTP 请求处理器**里调 → **真的提交**（`recallSummaryLen` 130→132→130，`revision` 4→6）。所以修法不是换 API 而是换**调用上下文**：① 宿主新增 `POST /memory-eternal/api/drain-pending`，在**请求处理器**里应用 pending（两道闸门：只接受本机来源 + 心跳里的一次性令牌，防本机任意页面乱写）；② 宿主把自己 web server 的**监听端口 + 令牌**写进心跳（`webServer.port` 是服务公开的 getter），独立端保存后据此**主动叫醒宿主**，拿到真实结果再回给界面（新增 `savedApplied`「已保存并生效」）；③ 叫醒失败（旧宿主没这条路由 404 / 端口对不上 / 超时）**不算错误** —— 回落到 0.10.5 的等待路径与「去 DSH 设置里改」的指路。新增 `tests/host-drain.test.mjs`（6 项：回环判定 / 守卫三件套 / 端到端叫醒 / 失败如实回报 / 404 回落 / 源码守卫）+ `tests/e2e-host-drain.mjs`（2 项：**真起进程**的端到端「独立页保存 → 叫醒宿主 → 宿主应用 → 回报已生效」，以及旧宿主心跳（无端口/令牌）时回落成「排队等待」）。`npm test` 共 **283 项** |
 | **v0.10.7** | 2026-10-08 | **「启动 dsh 后运行中还是旧版本」根治：版本漂移自愈 + 一键重启常驻实例（#19 / #23 的正面解法）**。现场：升级 npm 包只换了磁盘文件，**常驻 web（默认 7999）比 DSH 宿主活得久** —— 端口上那个进程仍在跑「启动时加载进内存」的旧代码；而宿主激活时的策略是「同端口已有活着的 watchdog 就 `delegated`（不再 spawn、更不替换）」，于是**重启多少次 DSH 都不会换掉它**（实测本机：宿主 10:19 启动已是新版、设置页同源显示 0.10.6，而 7999 上的 web 是 08:15 启动的 0.10.0，`/config` 与 `/version-check` 各报一个版本）。① **启动自愈（`autoRestartOnDrift`，默认开）**：激活时先 `inspectResident()` 体检 —— 版本一律**问端口**（锁里的 `pkgVersion` 只是 watchdog 自述，旧实例写的还是空串，根本看不出版本漂移），再 `decideResidentAction()` 决策：一致 → `delegate`（保持 #19 的如实口径）、漂移 → `restart`、读不到本机版本 → 绝不乱动、刚替换过 → 防抖不重来（配置改动会让 effect 重跑）。**「更新运行中的程序」的正确形态由此定为「用新代码重启那个常驻进程」，而不是热替换内存里的模块**：ESM 模块缓存 + 闭包状态（settings / vault 路由 / SSE hub / `fs.watch` / 已监听端口）没有安全热换的余地，缓存失效重 import 只会得到第二份模块图（双实例、双监听、双写）。② **替换机制统一到一处**：新增 `restartResident()` = spawn 一个 detached 的 `watchdog.js --replace` 助手（收旧 watchdog / 旧 web → **等端口真的空出来** → 抢锁 → 拉起新 web），再自检「端口上真正服务的版本 = 本机版本」（#23 建议 2）；端口没空出来就**中止接管**、不退让到 `port+1`（退让实例正是孤儿 web 的诞生地）。`dsh-memory restart`、宿主启动自愈、面板按钮现在共用这一份实现。助手是 detached + `stdio:ignore`（stderr 没人看），所以接管结论**必须**写进「自动沉淀日志」—— 否则「重启了却没生效」又会变成一条无迹可查的静默失败。③ **面板一键修复**：新增 `POST /memory-eternal/api/restart-self`，三条硬约束 —— 本进程很可能**就是**要被替换的旧 web，所以**先回响应再调度**替换；常驻实例已经是最新就回 `alreadyCurrent`（宿主面板点这个按钮时最常见，绝不为一个健康实例白折腾一次停机）；同一时刻只允许一个替换在跑。④ **面板语义消歧**：「运行中」改为「**运行中（服务本页的进程）**」（独立 Web 页里它是常驻 web，DSH 设置页里是宿主 —— 这正是两个面板数字「打架」的根源），并新增「**常驻实例 vX**」徽标：两者不同时一眼看出该重启谁；漂移横幅给按钮 + 等价命令，若常驻实例已新而本页仍旧，则明说「需重启桌面版 / DSH」，不再给一个点了也没用的按钮。⑤ **`updateAvailable` 改按版本比较**：`latest !== onDisk` 在本地跑未发布版本时会**反过来劝你装回旧版**（磁盘 0.10.7 / npm 0.10.6），现在用 `compareVersions`（含预发布语义 `rc.2 < rc.10`）。⑥ 顺带清掉客户端词典里重复的 `saving` 键（esbuild 每次构建都在报警告）。新增 `tests/watchdog-restart.test.mjs`（17 项：决策矩阵 / 体检问端口 / 替换自检 / 接管前清场不误杀 / 路由先响应后替换 + 并发闸门 / 版本比较）。⑦ **README 改为中英两份独立文件**：`README.md` 是**默认**的简体中文（npm 与 DSH 插件市场展示这一份），`README.en.md` 是英文镜像；顶部加语言切换与「改一处要同步另一处」的说明，并补齐英文版缺失的「环境自检」小节；新增 `tests/readme-i18n.test.mjs` 锁住**结构 1:1**（标题层级序列 + 代码块数量 + 互链 + 默认语言标记），防止以后只改一份。⑧ **`secretHint` 凭证提示 + 凭据读取守卫**：填一句自己的约定就会被注入每个会话的 systemPrompt，让 agent 需要 API key / token 时**先查记忆里的「密钥目录」卡**（记忆库只存目录：名字/位置/取用方式，**值留在本机凭据库**）；`dsh-memory setup` 同时注册 Claude Code 的 `PreToolUse` 守卫（拒读 `.credentials.yaml` / `.env` / `*-token` 并指路到加载器），并会刷新仍指向旧安装副本的 hook 路径。⑨ **发布前又抓到三个实测现场并修掉**：① 面板「重启常驻实例」按钮在**服务本页的进程**是旧版时也显示 —— 点了必然 404（旧宿主没有这条路由），用户只看到一句「未知接口」；现在按钮只在「端口上那个常驻实例落后于磁盘」时出现，404/405 还会给一条「重启桌面版 / DSH 或执行 dsh-memory restart」的指路。② 独立页（新版）提交整张表单、宿主还是旧版时：宿主按白名单丢掉它不认识的键（如 `secretHint`）→ 回读校验必然对不上 → 连试 5 次 `dropped`，用户只看到「HMR transactions cannot be nested」；现在应用前按宿主 schema **过滤未知键**（宿主升级后自然生效），并把「配置同步失败」与「自动沉淀异常」彻底分开（`config-warn` / `config-fail`，不再翻红沉淀健康状态）。③ `autoRestartOnDrift` 在宿主快照里没有这个键时按**默认开**渲染（实测现场：0.10.6 宿主 + 0.10.7 面板 → 复选框空着，而宿主实际按开处理，面板与真实行为不一致）。新增 `tests/issue-fixes-4.test.mjs`。`npm test` 共 **274 项** |
 | **v0.10.6** | 2026-10-07 | **孤儿 web 有了回收机制 + 共享配置同步失败不再静默**。① **孤儿 web**：web 子进程是 detached 启动的，watchdog 被硬终止时（Windows 上 `process.kill(pid,'SIGTERM')` 即无条件终止）它的退出处理器不执行；若这个 web 当初因为目标端口被占而**退让到 port+1…**，登记它的锁条目又随 watchdog 一起消失 —— 于是 8001/8002 这类端口上会常驻没人认领的 web（实测本机就躺着两个，其中一个已跑了 20 小时，而 `status` 只探配置端口，永远发现不了）。现在 `startWatchdog` 启动时、以及 `stop`/`restart`（`checkPort` 路径）都会扫**退让窗口** `[port+1, port+9]` 并收掉它们；三道门缺一不可（端口在窗口内 / 命令行**确实是本插件的 web** / pid 不在锁里任何 watchdog 的 `pid` 或 `webPid` 上），**别人的 `node web.js` 一个都不杀**。另外 `spawnWeb` 发现目标端口上已有自家 web 时**不再重复拉起**（退让实例的主要来源），只交给 tick 探活。② **`syncConfigFile()` 不再静默吞错**：它写的是独立页 / hooks / MCP 共用的读源，写不进去就表现为「配置改了但他们看不到」——旧实现是 `catch { /* 静默 */ }`（实测到过一次：共享文件 mtime 停在激活前，而宿主里的值已经变了）。现在失败会进 stderr 与「自动沉淀日志」的 warn（同一原因只报一次，不刷屏）；因为 `logCapture` 定义在 `apply()` 更靠后的位置，日志刻意**延迟一拍**再写，避开 `const` 的 TDZ。新增 `tests/watchdog-orphans.test.mjs`（7 项），`npm test` 共 **245 项** |
 | **v0.10.5** | 2026-10-07 | **独立 Web 端的配置保存：不再空头承诺，也不再要求「必须装了 DSH」**。① **宿主无关的直写路径（#21）**：独立端保存后若 800ms 内没人消费 pending 文件、**且**本机没有活着的 DSH 宿主（心跳文件 `memory-eternal-config.host.json` 不新鲜、或里面的 pid 已死），就直接把改动**原子合并**写进共享配置、清掉 pending，响应回 `pendingOutcome: applied-direct` —— 于是 Codex / Claude Code / Cursor、以及只跑 `dsh-memory serve` 的用户，保存当场生效（此前那条路只会写「待应用」等一个**根本不存在**的宿主，界面上那句「下次启动生效」永远不会兑现）。宿主活着时**绝不**直写：共享配置是宿主的派生镜像，绕过它会被下一次 `syncConfigFile()` 盖回去（有专门的反向测试守着这条边界）。② **失败要指路，不只是重复报错（#21）**：识别「宿主守卫类」报错 —— `HMR transactions cannot be nested`（Cordis HMR 宿主）与 `root.events.emit is unavailable from a plugin activation`（dsh-tui 一类）是同一族缺陷的两种措辞 —— 并把「此宿主不允许插件从自己的回调上下文写配置」与「去哪改」（**DSH 设置 → 记忆**，或 profile 的配置补丁层）写进放弃时的报错、诊断信息与独立页提示；`POST /config` 同时回 `hostGuard` / `hint` 字段。③ **界面不再一律弹「已保存」（#21）**：客户端开始消费 `pendingOutcome` —— `queued` / `failed` 用**非成功样式**并说明「尚未确认生效」，`applied-direct` 明确写「本机没有 DSH 宿主，已直接写入共享配置」，只有真生效才显示「已保存」。④ 顺带修掉一个会让宿主/测试**无法退出**的清理竞态：心跳定时器与 pending 文件监听都是在异步 `import()` 之后才建立，若在它 resolve 之前就 dispose，回调就再没人清理（心跳定时器还额外 `unref()`）—— `tests/settings-compat.test.mjs` 用假 ctx 跑 `apply()` 时正是被它挂住的。新增 `tests/issue-fixes-3.test.mjs`（9 项），`npm test` 共 **238 项** |

package/lib/capture.js

@@ -418,6 +418,44 @@ export function repairJsonControlChars(text) {
   return { text: out, repaired }
 }
 
+/**
+ * 修复「字符串值里未转义的裸引号」（10/8 实测现场）。
+ *
+ * 现场：`JSON.parse 原始报错：Expected ',' or '}' after property value in JSON at position 1913`，
+ * 摘录里能看到正文写着 `运行 "npm test" 时…` —— 模型把引号当正文内容直接写进字符串，于是这个
+ * 引号被当成字符串结束符，后面的中文就变成「属性值后面还有东西」。这与已经修过的「裸控制字符」
+ * （#24）是**不同**的一类：控制字符在字符串内部是非法的，而引号在字符串内部是「合法但需要转义」。
+ *
+ * 启发式（只作为**最后一个候选**，见 parseCaptureJsonDetailed）：一个 `"` 只有当它后面（跳过空白）
+ * 紧跟 `:` `,` `}` `]` 或文本结尾时才算「字符串结束」，否则补 `\"`。
+ * 因为可能误修，调用方必须再通过卡片体检（`cardFromParsedJson`）才接受 —— 宁可不修，也不写坏卡。
+ *
+ * @returns {{ text: string, repaired: boolean }} repaired=true 表示确实补过反斜杠
+ */
+export function repairUnescapedQuotes(text) {
+  const s = String(text || '')
+  let out = ''
+  let inStr = false
+  let esc = false
+  let repaired = false
+  for (let i = 0; i < s.length; i++) {
+    const ch = s[i]
+    if (esc) { esc = false; out += ch; continue }
+    if (ch === '\\') { esc = true; out += ch; continue }
+    if (ch !== '"') { out += ch; continue }
+    if (!inStr) { inStr = true; out += ch; continue }
+    // 已在字符串里：看这个引号后面是否紧跟结构字符
+    let j = i + 1
+    while (j < s.length && /\s/.test(s[j])) j++
+    const next = j < s.length ? s[j] : ''
+    const endsString = next === '' || next === ':' || next === ',' || next === '}' || next === ']'
+    if (endsString) { inStr = false; out += ch; continue }
+    repaired = true
+    out += '\\"'
+  }
+  return { text: out, repaired }
+}
+
 /**
  * 失败现场摘要：首尾各留一段，中间省略多少字写清楚。
  *
@@ -479,6 +517,12 @@ export function parseCaptureJsonDetailed(text) {
     const fixed = repairJsonControlChars(base)
     if (fixed.repaired) candidates.push({ text: fixed.text, repaired: true })
   }
+  // 最后一层：字符串内裸引号（10/8 现场）。它对合法 JSON 是恒等变换（repaired=false 时不入列），
+  // 且必须仍能通过卡片体检才会被接受。
+  for (const candidate of [...candidates]) {
+    const fixed = repairUnescapedQuotes(candidate.text)
+    if (fixed.repaired) candidates.push({ text: fixed.text, repaired: true })
+  }
   let lastError = ''
   for (const candidate of candidates) {
     let parsed