memory-eternal 0.10.52 → 0.10.54

package/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "$schema": "https://anthropic.com/claude-code/plugin.schema.json",
   "name": "memory-eternal",
-  "version": "0.10.52",
+  "version": "0.10.54",
   "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": "0.10.52",
+  "version": "0.10.54",
   "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": "0.10.52",
+  "version": "0.10.54",
   "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

@@ -17,20 +17,36 @@
 </p>
 
 <p align="center">
-  <img src="https://raw.githubusercontent.com/EternalNight996/memory-eternal/main/assets/screen/memory-eternal.gif" width="880" alt="Auto-capture + visual library + knowledge graph (demo)" />
+  <img src="https://raw.githubusercontent.com/EternalNight996/memory-eternal/main/assets/screen/cards.png" width="900" alt="Knowledge card list: every conversation is distilled into a card you can search, audit and trace back" />
 </p>
 
-> **Auto-captures knowledge after every conversation and survives across sessions; recall fetches only the relevant chunks — saves tokens, less noise.**
-> Fully self-built, zero third-party memory framework, no DSH source changes, one vault shared by every Agent, SQLite persistent storage, zero external dependencies.
+> **Every conversation is distilled into knowledge cards; nothing is forgotten across sessions, and recall fetches only the relevant chunks — fewer tokens, less noise.**
+> Fully self-built, zero third-party memory framework, no DSH source changes; one vault shared by every Agent, SQLite persistent storage, zero external dependencies.
 
-<p align="center"><strong>⭐ Star it if you like it!</strong> <br/><sub>DSH one-liner: <code>dsh plugin --profile web add memory-eternal</code></sub></p>
+<p align="center"><strong>⭐ If it saved you from explaining the same context again, a star means a lot</strong><br/><sub>DSH one-liner: <code>dsh plugin --profile web add memory-eternal</code></sub></p>
 
 <p align="center">
-  <img src="https://raw.githubusercontent.com/EternalNight996/memory-eternal/main/assets/screen/memory-popup.png" width="32%" alt="Memory popup: knowledge cards / search / knowledge graph" />
-  <img src="https://raw.githubusercontent.com/EternalNight996/memory-eternal/main/assets/screen/memory-settings.png" width="32%" alt="DSH Settings → Memory: all options" />
-  <img src="https://raw.githubusercontent.com/EternalNight996/memory-eternal/main/assets/screen/memory-sidebar.png" width="32%" alt="Sidebar quick entry" />
+  <img src="https://raw.githubusercontent.com/EternalNight996/memory-eternal/main/assets/screen/graph.png" width="32%" alt="Knowledge graph: 897 nodes / 4257 edges, coloured by kind, cross-vault aggregation and export" />
+  <img src="https://raw.githubusercontent.com/EternalNight996/memory-eternal/main/assets/screen/health.png" width="32%" alt="Health overview: database / backups / auto-capture / versions — spot anything broken silently" />
+  <img src="https://raw.githubusercontent.com/EternalNight996/memory-eternal/main/assets/screen/audit.png" width="32%" alt="Audit centre: AI-written cards are reviewed before they land; batch approve / reject / recycle" />
 </p>
 
+<p align="center"><sub>▲ Card list · knowledge graph · health · audit centre — real screenshots from a running instance (1280×860)</sub></p>
+
+---
+
+## ✨ What problem does it actually solve
+
+| What you hit | Before | After |
+| --- | --- | --- |
+| A project you talked about for three days, forgotten in a new session | dig through history, explain it again | each turn is distilled into a card; one `memory_recall` brings it back |
+| Cramming memory into the prompt blows the token budget | paste everything, get more expensive | only the relevant chunk is recalled (capped), fewer tokens and less noise |
+| You cannot fully trust what the AI wrote | no trace, no rollback | cards carry source / time / status, audited before landing, deletions go to the recycle bin |
+
+- **Visible**: cards, graph, health overview and capture log — four layers of visibility;
+- **Controllable**: audit modes, dedup and contradiction scans, batch merge, daily backups and one-command restore;
+- **Portable**: the vault is local Markdown + SQLite — git-managed and exportable, never locked inside the plugin.
+
 ---
 
 ## 🚀 Get Started in 5 Minutes
@@ -170,7 +186,7 @@ Three concepts, don't mix them:
 > All settings live in the **vault left rail "Memory Config"** (or DSH Settings → Memory) — edit and hit "Save". Here's the full page (plugin info / Agent MCP mount status / auto-audit config):
 
 <p align="center">
-  <img src="https://raw.githubusercontent.com/EternalNight996/memory-eternal/main/assets/screen/memory-config.png" width="880" alt="Memory config page" />
+  <img src="https://raw.githubusercontent.com/EternalNight996/memory-eternal/main/assets/screen/config.png" width="880" alt="Memory config page" />
 </p>
 
 ### 1. Most used
@@ -361,14 +377,16 @@ curl http://127.0.0.1:7999/memory-eternal/api/web-info    # is the standalone we
 
 | Version | Date | Highlights |
 |---|---|---|
-| **v0.10.52** | 2026-10-19 | **Redaction hole in the diagnostic bundle (privacy fix; upgrade recommended)**. The 0.10.51 one-command bundle had a gap: the home directory appears in **JSON-escaped form** (C:\\Users\\x) inside doctor/align --json output and meta.json, while only the plain and forward-slash variants were replaced, so the bundle could still be grepped for the home path. All three variants (plain, forward slash, JSON escaped) now become <home>, with a test assert added for the escaped form; a freshly generated bundle on this machine verifies as **no home path and no sk- keys**. If you already generated and shared a bundle with 0.10.51, please regenerate it before sharing. npm test now runs **433 checks** |
-| **v0.10.51** | 2026-10-19 | **One-command diagnostic bundle: dsh-memory support** (reliability line; in #47 the user-pasted context was truncated by GitHub so I could not see the full picture — this exists for exactly that kind of exchange). A single command produces a **zip** with meta (version matrix / platform / node / vault), doctor, align, status, the watchdog lock, the host heartbeat, the shared config, the tail of the capture log (200 lines by default) and a README; everything is **redacted before writing** (home directory to <home>, sk-/ghp_/github_pat/Bearer and key-like config fields to ***). Implemented with **no dependencies**: a hand-written store/deflate ZIP writer (lib/zip.js) whose output opens in Windows Explorer (verified by extracting with Expand-Archive). In restricted environments --no-spawn skips the sub-processes, while --out and --tail N control destination and log length. Adds tests/support-bundle.test.mjs (redaction, zip structure, no plaintext secrets inside the bundle, CLI wiring). npm test now runs **433 checks** |
+| **v0.10.54** | 2026-10-20 | **More reliable distillation with local models (#48), plus a refreshed README with new screenshots**. ① **#48**: a local 30B model takes 40-60s to cold start while the distill timeout was **hardcoded to 45s**, so every cold start aborted; the timeout is now configurable via **captureTimeoutMs** (still 45s by default, 120000+ recommended for local inference) and **a timeout retries the same candidate once with a longer budget** instead of switching models. ② **The second shape of #48**: when the provider does not report max-tokens, a truncated answer only shows up as unterminated JSON, so a dedicated **TRUNCATED_OUTPUT** code now joins MAX_TOKENS on the "double the output budget and retry the same candidate" ladder. ③ The settings page gains a distill timeout (ms) field. ④ **README refresh**: all real screenshots (cards, a graph with 897 nodes / 4257 edges, health overview, audit centre, settings) plus a "what problem does it solve" comparison table and a clearer value proposition and star call-to-action (both languages). npm test now runs **440 checks** |
+| **v0.10.53** | 2026-10-20 | **Health panel (reliability line): all previously invisible state in one place**. A new "Health" card at the top of the usage page shows: ① **database** (journal mode, writability, last health check, with doctor guidance when bad); ② **backups** (count and age of the newest, amber past 7 days, red at zero); ③ **auto capture** (last success/failure with reason); ④ **versions** (disk / resident / host, plus a Start check button because align needs the npm registry and cannot run automatically). Data comes from existing endpoints: /overview now includes a **backups summary** (count/newestAt/last five filenames) alongside /capture-log and /align. Also **fixes a white-screen bug**: the card initially used tf(), which only exists in the ConfigPanel scope, so opening the usage page blanked the whole app; the panel checker gained two gates — **page alive (textLen<20 counts as blank)** and **Runtime.exceptionThrown capture** — because a blank page previously passed the overflow assertions. Adds tests/health-panel.test.mjs (4 checks). npm test now runs **437 checks** |
 
 <details>
-<summary>Older release notes (v0.10.50 and earlier · 77 releases, click to expand)</summary>
+<summary>Older release notes (v0.10.52 and earlier · 79 releases, click to expand)</summary>
 
 | Version | Date | Highlights |
 |---|---|---|
+| **v0.10.52** | 2026-10-19 | **Redaction hole in the diagnostic bundle (privacy fix; upgrade recommended)**. The 0.10.51 one-command bundle had a gap: the home directory appears in **JSON-escaped form** (C:\\Users\\x) inside doctor/align --json output and meta.json, while only the plain and forward-slash variants were replaced, so the bundle could still be grepped for the home path. All three variants (plain, forward slash, JSON escaped) now become <home>, with a test assert added for the escaped form; a freshly generated bundle on this machine verifies as **no home path and no sk- keys**. If you already generated and shared a bundle with 0.10.51, please regenerate it before sharing. npm test now runs **433 checks** |
+| **v0.10.51** | 2026-10-19 | **One-command diagnostic bundle: dsh-memory support** (reliability line; in #47 the user-pasted context was truncated by GitHub so I could not see the full picture — this exists for exactly that kind of exchange). A single command produces a **zip** with meta (version matrix / platform / node / vault), doctor, align, status, the watchdog lock, the host heartbeat, the shared config, the tail of the capture log (200 lines by default) and a README; everything is **redacted before writing** (home directory to <home>, sk-/ghp_/github_pat/Bearer and key-like config fields to ***). Implemented with **no dependencies**: a hand-written store/deflate ZIP writer (lib/zip.js) whose output opens in Windows Explorer (verified by extracting with Expand-Archive). In restricted environments --no-spawn skips the sub-processes, while --out and --tail N control destination and log length. Adds tests/support-bundle.test.mjs (redaction, zip structure, no plaintext secrets inside the bundle, CLI wiring). npm test now runs **433 checks** |
 | **v0.10.50** | 2026-10-19 | **Models that ignore JSON no longer burn a call every turn (a follow-up to #47)**. ① "The model replied with prose and produced no JSON object at all" is now separated from the generic UNPARSEABLE_OUTPUT failure and gets its own code **MODEL_NOT_JSON**, with wording that says the model likely does not follow structured output. ② New `lib/model-strikes.js` records which provider/model recently ignored the format (24h window, at most 50 entries) and the host moves those to the end of the candidate list when routing (never removing them; if every candidate was struck the original order is kept, so we never run out), logging a "candidates reordered" line. ③ Effect: models such as stepfun/step-api, which answer distillation prompts conversationally, stop being tried first from the second turn on, saving a pointless call each time. Adds `tests/model-strikes.test.mjs` (prose detection, stable ranking, record window and trimming, wiring guards). npm test now runs **429 checks** |
 | **v0.10.49** | 2026-10-19 | **Fixes "cannot choose the distillation model" in the desktop app (#47): the standalone page now proxies /llm-routes and /llm-test to the host, plus one more state in the matrix**. Root cause: the desktop memory popup is a **standalone page inside an iframe** with no DSH llm service, so /llm-routes always failed and provider/model had to be typed by hand. The standalone page now reads the host heartbeat (the port in <config>.host.json, 127.0.0.1 only, valid for 120s) and forwards those two routes to the host, marking the reply as proxied; when the host is offline the panel shows the concrete reason instead of silently falling back to text inputs. ② The version-card matrix gains restartRepair: when versions already match but nothing keeps the instance alive it says "re-register and keep it alive", no longer the self-contradictory "switch to v0.10.48 (currently v0.10.48)". npm test now runs **425 checks** |
 | **v0.10.48** | 2026-10-19 | **A visual gate plus exhaustive state coverage (the option we agreed on)**. ① **New `npm run panel-check`** (scripts/panel-check.mjs, zero dependencies): drives local headless Chrome over CDP, asserts **no horizontal overflow** on each screen and saves screenshots (list / reader / config / feedback modal). On its very first run it caught two real bugs: an unbreakable long run in a card title (`(begin/chunk/probe/complete/abort)`) overflowing by 45px, and the **feedback textarea overflowing by 22px** (width:100% plus padding under content-box) — exactly the layout/scrollbar anomaly reported. ② Fixes: word-breaking on `.mc-cardrow h4` and a shared border-box rule for form controls. ③ **The version-card decision is now a pure function** `src/client/version-state.js`: versionCardState returns {showButton, emphasis, warn, hint} and the component only maps keys to text; a table-driven test **enumerates 324 combinations** and enforces the invariant that the button is shown whenever version info exists — structurally preventing the 0.10.45 regression where a third state made it disappear. npm test now runs **420 checks** |

package/README.md

@@ -17,20 +17,36 @@
 </p>
 
 <p align="center">
-  <img src="https://raw.githubusercontent.com/EternalNight996/memory-eternal/main/assets/screen/memory-eternal.gif" width="880" alt="对话自动沉淀 + 图形化知识库 + 知识图谱（演示）" />
+  <img src="https://raw.githubusercontent.com/EternalNight996/memory-eternal/main/assets/screen/cards.png" width="900" alt="知识卡列表：对话结束自动沉淀成卡，可检索 / 可审核 / 可追溯来源" />
 </p>
 
-> **对话结束自动沉淀，跨会话不失忆；召回只取相关小块，省 token 少噪音。**
-> 全自研、零第三方记忆框架、不改 DSH 源码、一个记忆库所有 Agent 共享，SQLite 持久存储，零外部依赖。
+> **对话结束自动沉淀成知识卡，跨会话不失忆；召回只取相关小块，省 token 少噪音。**
+> 全自研、零第三方记忆框架、不改 DSH 源码；一个记忆库所有 Agent 共享，SQLite 持久存储，零外部依赖。
 
-<p align="center"><strong>⭐ 觉得好用就点个 Star</strong>！ <br/><sub>DSH 一条命令：<code>dsh plugin --profile web add memory-eternal</code></sub></p>
+<p align="center"><strong>⭐ 如果它帮你少解释了一遍上下文，点个 Star 就是最大的支持</strong><br/><sub>DSH 一条命令：<code>dsh plugin --profile web add memory-eternal</code></sub></p>
 
 <p align="center">
-  <img src="https://raw.githubusercontent.com/EternalNight996/memory-eternal/main/assets/screen/memory-popup.png" width="32%" alt="记忆库弹窗：知识卡 / 检索 / 知识图谱" />
-  <img src="https://raw.githubusercontent.com/EternalNight996/memory-eternal/main/assets/screen/memory-settings.png" width="32%" alt="DSH 设置 → 记忆：全部配置项" />
-  <img src="https://raw.githubusercontent.com/EternalNight996/memory-eternal/main/assets/screen/memory-sidebar.png" width="32%" alt="侧边栏一键入口" />
+  <img src="https://raw.githubusercontent.com/EternalNight996/memory-eternal/main/assets/screen/graph.png" width="32%" alt="知识图谱：897 节点 / 4257 连线，按类型着色、支持跨库聚合与导出" />
+  <img src="https://raw.githubusercontent.com/EternalNight996/memory-eternal/main/assets/screen/health.png" width="32%" alt="健康概览：知识库 / 备份 / 自动沉淀 / 版本，一眼看清有没有静默坏掉的东西" />
+  <img src="https://raw.githubusercontent.com/EternalNight996/memory-eternal/main/assets/screen/audit.png" width="32%" alt="审核中心：AI 写的卡先审后入，可批量批准 / 驳回 / 回收" />
 </p>
 
+<p align="center"><sub>▲ 知识卡列表 · 知识图谱 · 健康概览 · 审核中心 —— 全部是真实运行截图（1280×860）</sub></p>
+
+---
+
+## ✨ 它到底解决什么问题
+
+| 你遇到的 | 以前的做法 | 用了之后 |
+| --- | --- | --- |
+| 聊了 3 天的项目，新会话全忘 | 手动翻历史 / 重新解释一遍 | 每轮结束自动沉淀成卡，开工前 `memory_recall` 一句就回来 |
+| 记忆塞进 prompt，token 爆掉 | 全量粘贴、越聊越贵 | 只召回相关的一小块（可配上限），省 token 也少噪音 |
+| AI 写的东西不敢直接信 | 无法追溯、无法回滚 | 卡片带来源 / 时间 / 状态，先审后入，删除进回收站 |
+
+- **看得见**：知识卡 / 图谱 / 健康概览 / 沉淀日志，四层可视化；
+- **控得住**：审核模式（全部人工 / 仅高风险）、去重与矛盾扫描、批量合并、每日备份与一键恢复；
+- **带得走**：知识库是本地 Markdown + SQLite，可 git 管理、可导出，不锁在插件里。
+
 ---
 
 ## 🚀 五分钟上手
@@ -190,7 +206,7 @@ dsh-memory audit reject <卡片路径...> --reason "..." # 人工批量驳回
 > 所有配置都在 **记忆库左栏「记忆配置」**（或 DSH 设置 → 记忆）里改，点「保存配置」生效。下图就是配置页全貌（含插件信息 / Agent MCP 挂载状态 / 自动审核配置）：
 
 <p align="center">
-  <img src="https://raw.githubusercontent.com/EternalNight996/memory-eternal/main/assets/screen/memory-config.png" width="880" alt="记忆配置页" />
+  <img src="https://raw.githubusercontent.com/EternalNight996/memory-eternal/main/assets/screen/config.png" width="880" alt="记忆配置页：蒸馏 provider / model 下拉、跟随会话、仅试首选、超时与输出上限" />
 </p>
 
 ### 一、最常用
@@ -393,14 +409,16 @@ 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 |
 |---|---|---|
-| **v0.10.52** | 2026-10-19 | **诊断包脱敏补漏（隐私修复，建议尽快升级）**。0.10.51 的「一键诊断包」有个漏点：doctor/align --json 与 meta.json 里的家目录是 **JSON 转义形态**（C:\\Users\\x），而我只替换了原样/正斜杠两种写法 → 包里仍能 grep 到家目录明文。现在三种写法（原样、正斜杠、JSON 转义）都替换成 <home>，并在测试里加了一条「JSON 转义形态不能漏」的断言；本机重新生成诊断包实测：**无家目录明文、无 sk- 密钥**。**如果你已经用 0.10.51 生成过诊断包并分享出去，建议重新生成一次再分享。** npm test 共 **433 项** |
-| **v0.10.51** | 2026-10-19 | **一键诊断包 dsh-memory support**（可靠性线；#47 里用户贴的现场被 GitHub 截断，我拿不到完整信息，这功能就是为这类沟通做的）。一条命令生成 **zip**：meta（版本矩阵/平台/node/vault）/ doctor / align / status / 看门狗锁 / 宿主心跳 / 共享配置 / 沉淀日志尾巴（默认 200 行）+ README；写入前**统一脱敏**（家目录→<home>，sk-/ghp_/github_pat/Bearer/配置里 key 类字段→***）。实现为**零依赖**：自写 store/deflate ZIP 写入器（lib/zip.js，Windows 资源管理器可直接打开，实测 Expand-Archive 解压正常）。受限环境可加 --no-spawn 跳过子进程；--out 指定输出、--tail N 控制日志条数。新增 tests/support-bundle.test.mjs（4 项：脱敏、zip 结构、包内不含明文密钥、CLI 接线）。npm test 共 **433 项** |
+| **v0.10.54** | 2026-10-20 | **本地大模型蒸馏更稳（#48）+ README 焕新与最新截图**。① **#48**：本地 30B 模型冷启动要 40–60 秒，而蒸馏超时**硬编码 45 秒** → 每次冷启动必报 ABORTED；现在超时可配 **captureTimeoutMs**（默认仍 45s，本地推理建议 120000+），且**超时后自动用更长超时重试同一候选一次**（不直接换模型）。② **#48 的第二种形态**：模型不给 max-tokens finish 时，截断只表现为「JSON 未闭合」→ 新增独立错误码 **TRUNCATED_OUTPUT**，与 MAX_TOKENS 一起走「加倍输出预算重试同一候选」的阶梯。③ 配置页新增「蒸馏超时（毫秒）」输入项。④ **README 焕新**：全部换成真实运行截图（知识卡 / 知识图谱 897 节点 4257 连线 / 健康概览 / 审核中心 / 配置页）+ 新增「它到底解决什么问题」对照表 + 更清晰的卖点与 Star 引导（中英同步）。npm test 共 **440 项** |
+| **v0.10.53** | 2026-10-20 | **健康面板（可靠性线）：把看不见的状态集中一处**。在「用量/今日」顶部新增「🩺 健康概览」：① **知识库**（journal 模式 / 可写 / 最后一次体检时间，异常时给 doctor 指引）；② **备份**（数量 + 最近一次多久前，>7 天标黄、0 备份标红）；③ **自动沉淀**（最近一次成功/失败与原因）；④ **版本**（磁盘 / 常驻 / 宿主；未检查时给「开始检查」按钮，因为 align 要联网比 npm 不能自动跑）。数据来自既有接口：/overview 新增 **backups 摘要**（count/newestAt/最近 5 个文件名），加上 /capture-log 与 /align。**顺带修一个白屏 bug**：健康卡最初误用了 ConfigPanel 作用域的 tf() → 点开「用量/今日」整页白屏；同时给 panel-check 补了两道闸 —— **页面存活（textLen<20 判空白页）** 与 **Runtime.exceptionThrown 捕获**，因为「白屏没有横向溢出」原本会让检查通过。新增 tests/health-panel.test.mjs（4 项）。npm test 共 **437 项** |
 
 <details>
-<summary>更早的更新日志（v0.10.50 及以前 · 共 77 版，点开查看）</summary>
+<summary>更早的更新日志（v0.10.52 及以前 · 共 79 版，点开查看）</summary>
 
 | 版本 | 日期 | 关键改动 |
 |---|---|---|
+| **v0.10.52** | 2026-10-19 | **诊断包脱敏补漏（隐私修复，建议尽快升级）**。0.10.51 的「一键诊断包」有个漏点：doctor/align --json 与 meta.json 里的家目录是 **JSON 转义形态**（C:\\Users\\x），而我只替换了原样/正斜杠两种写法 → 包里仍能 grep 到家目录明文。现在三种写法（原样、正斜杠、JSON 转义）都替换成 <home>，并在测试里加了一条「JSON 转义形态不能漏」的断言；本机重新生成诊断包实测：**无家目录明文、无 sk- 密钥**。**如果你已经用 0.10.51 生成过诊断包并分享出去，建议重新生成一次再分享。** npm test 共 **433 项** |
+| **v0.10.51** | 2026-10-19 | **一键诊断包 dsh-memory support**（可靠性线；#47 里用户贴的现场被 GitHub 截断，我拿不到完整信息，这功能就是为这类沟通做的）。一条命令生成 **zip**：meta（版本矩阵/平台/node/vault）/ doctor / align / status / 看门狗锁 / 宿主心跳 / 共享配置 / 沉淀日志尾巴（默认 200 行）+ README；写入前**统一脱敏**（家目录→<home>，sk-/ghp_/github_pat/Bearer/配置里 key 类字段→***）。实现为**零依赖**：自写 store/deflate ZIP 写入器（lib/zip.js，Windows 资源管理器可直接打开，实测 Expand-Archive 解压正常）。受限环境可加 --no-spawn 跳过子进程；--out 指定输出、--tail N 控制日志条数。新增 tests/support-bundle.test.mjs（4 项：脱敏、zip 结构、包内不含明文密钥、CLI 接线）。npm test 共 **433 项** |
 | **v0.10.50** | 2026-10-19 | **「模型不守 JSON」不再每轮白烧调用（#47 衍生改进）**。① 把「模型回了散文、输出里根本没有 JSON 对象」从笼统的 UNPARSEABLE_OUTPUT 里拆出来，给独立错误码 **MODEL_NOT_JSON**，文案直接说清「该模型可能不遵守结构化输出」。② 新增 `lib/model-strikes.js`：记下「哪个 provider/model 最近不守格式」（24h 窗口、最多 50 条），宿主选路时把它们**排到候选末尾**（不删除；全部命中过则按原顺序兜底，绝不没候选），并留一条「蒸馏候选已调整」日志。③ 效果：像 stepfun/step-api 这种会把蒸馏当聊天回的模型，第二轮起就不再优先尝试，省掉每次的无效调用。新增 `tests/model-strikes.test.mjs`（4 项：散文识别 / 排序稳定性 / 记录窗口与裁剪 / 接线守卫）。npm test 共 **429 项** |
 | **v0.10.49** | 2026-10-19 | **桌面版里「无法选择蒸馏模型」修复（#47）：独立页把 /llm-routes 与 /llm-test 代理给宿主 + 文案补一个状态**。根因：桌面版的「记忆」弹窗是 iframe 里的**独立页**，它没有 DSH 的 llm 服务 → /llm-routes 恒失败 → provider/model 只能手填。现在独立页读宿主心跳（<config>.host.json 里的端口，仅 127.0.0.1、心跳 120s 内有效），把这两个路由**转发给宿主**执行并标记 proxied；宿主不在线时给明确原因（面板直接显示，不再静默回退成手填）。② 版本卡状态矩阵补 restartRepair：**版本已对但没人保活**时提示「重新托管并保活」，不再出现「换成 v0.10.48（当前 v0.10.48）」这种自相矛盾的文案（截图实测）。npm test 共 **425 项** |
 | **v0.10.48** | 2026-10-19 | **界面验收手段 + 状态穷举（方案一落地）**。① **新增 `npm run panel-check`**（scripts/panel-check.mjs，零依赖）：用本机 Chrome headless + CDP 打开面板，逐屏断言**没有横向溢出**并截图（列表 / 阅读器 / 配置 / 反馈弹窗）。它第一次跑就抓到两个真 bug：卡片标题里的无断点长串（`（begin/chunk/probe/complete/abort）`）溢出 45px；**反馈弹窗 textarea 溢出 22px**（width:100% + padding 在 content-box 下）—— 正是用户报的「排版不对 / 滚动条异常」。② 修复：`.mc-cardrow h4` 断词、`.me-overlay input/textarea/select` 统一 border-box。③ **版本卡判定抽成纯函数** `src/client/version-state.js`：`versionCardState(...)` 返回 `{showButton, emphasis, warn, hint}`，组件只做 key→文案；新增表驱动测试**穷举 324 种组合**，并守住不变式「只要拿到版本信息，按钮永远显示」—— 0.10.45「第三种状态导致按钮消失」从结构上不可能再发生。npm test 共 **420 项** |

package/index.js

@@ -77,6 +77,9 @@ export const Config = z.object({
   // 默认值 900 → 2000（issue #18）：prompt 要求 300-800 字正文，900 会让超长会话常态化撞上限，
   // 输出被截断后旧代码还会把它误报成「解析失败」，最终落一张 raw 噪声卡。
   captureMaxTokens: z.number().min(100).max(4000).default(DEFAULT_CAPTURE_MAX_TOKENS),
+  // issue #48：蒸馏单次超时（毫秒）。本地 30B 模型冷启动要 40–60s，45s 必超时；
+  // 默认仍 45s（云端模型够用），本地推理用户可调到 120000+。超时后宿主会自动加时重试一次。
+  captureTimeoutMs: z.number().min(5000).max(600000).default(45000),
   // 召回相关性阈值（minScore），越高召回越少越精越省
   recallMinScore: z.number().min(0).max(50).default(2),
   // 注入体积可配置（召回）
@@ -736,16 +739,28 @@ export function apply(ctx, config) {
       const maxTokens = Number(cfg.captureMaxTokens) || DEFAULT_CAPTURE_MAX_TOKENS
       for (let i = 0; i < routes.length; i++) {
         route = routes[i]
-        let detailed = await summarizeTurnDetailed(llm, route, text, { signal: AbortSignal.timeout(45000), existing: neighbors, maxTokens })
+        // issue #48：超时不再硬编码 45s（本地 30B 冷启动 40–60s 必超时），且超时后**同候选**用更长超时重试一次。
+        const baseTimeout = Math.max(5000, Number(cfg.captureTimeoutMs) || 45000)
+        const isTimeoutFailure = (x) => /ABORTED|TIMEOUT|ETIMEDOUT|aborted/i.test(String((x && (x.code + ' ' + x.message)) || ''))
+        const attempt = (budget, ms) => summarizeTurnDetailed(llm, route, text, { signal: AbortSignal.timeout(ms), existing: neighbors, maxTokens: budget })
+        let detailed = await attempt(maxTokens, baseTimeout)
+        if (detailed.failure && isTimeoutFailure(detailed.failure)) {
+          const longer = Math.min(Math.max(Math.round(baseTimeout * 2.5), baseTimeout + 30000), 180000)
+          logCapture(sessionId, 'listen', `蒸馏超时（${route.provider}/${route.model}）：可能是模型冷启动 → 以 ${Math.round(longer / 1000)}s 超时重试同一候选`)
+          detailed = await attempt(maxTokens, longer)
+        }
         // #18 建议 3 + #24：撞输出上限不是「模型坏了」，把上限一路翻倍到 schema 上限。
         // 旧实现只翻倍**一次**：1200 → 2400 仍不够就照样退成原文卡，尽管 4000 本来装得下
         // （issue #24 里 9 次失败中有一半正是被这一限制吃掉的）。
+        // #48：JSON 被截断（TRUNCATED_OUTPUT）与 MAX_TOKENS 同类处理 —— 靠加倍预算而不是换模型
+        const truncLike = (x) => Boolean(x && (x.code === 'MAX_TOKENS' || x.code === 'TRUNCATED_OUTPUT'))
         let prevBudget = maxTokens
         for (const bigger of maxTokenLadder(maxTokens)) {
-          logCapture(sessionId, 'listen', `输出撞到 maxTokens=${prevBudget} 上限 → 以 ${bigger} 重试同一候选（${route.provider}/${route.model}）`)
-          detailed = await summarizeTurnDetailed(llm, route, text, { signal: AbortSignal.timeout(45000), existing: neighbors, maxTokens: bigger })
+          if (!truncLike(detailed.failure)) break
+          logCapture(sessionId, 'listen', `输出撞到上限（${detailed.failure.code}，maxTokens=${prevBudget}）→ 以 ${bigger} 重试同一候选（${route.provider}/${route.model}）`)
+          detailed = await attempt(bigger, baseTimeout)
           prevBudget = bigger
-          if (!(detailed.failure && detailed.failure.code === 'MAX_TOKENS')) break
+
         }
         if (detailed.card !== undefined) { result = detailed.card; failure = null; break }
         if (detailed.failure) {
@@ -1194,6 +1209,7 @@ export function apply(ctx, config) {
                 captureProvider: cfg.captureProvider || '',
                 captureFollowSession: cfg.captureFollowSession === true,
                 captureFirstOnly: cfg.captureFirstOnly === true,
+                captureTimeoutMs: cfg.captureTimeoutMs ?? 45000,
                 captureModel: cfg.captureModel || '',
                 recallEmbedding: cfg.recallEmbedding || '',
                 sessionBudgetChars: cfg.sessionBudgetChars ?? 80000,

package/lib/api.js

@@ -13,7 +13,8 @@
 
 import { gzipSync, brotliCompressSync, constants as zlibConstants } from 'node:zlib'
 import os from 'node:os'
-import { readFileSync } from 'node:fs'
+import fs, { readFileSync } from 'node:fs'
+import path from 'node:path'
 import { listCards, countCards, readCard, search, searchAll, graph, graphAll, overview, exportCards, deleteCard, writeCard, parseCard, stats, optimizeCandidates, optimizeApply, addFeedback, dailyCounts, mergeCards, ensureVault, auditQueue, recycleList, purgeExpired, setCardStatus, restoreCard, mainStoreBodies } from './vault.js'
 import { compressExcerpt, DEFAULT_CAPTURE_MAX_TOKENS } from './capture.js'
 import { buildDiagnostics } from './feedback.js'
@@ -230,11 +231,24 @@ export function createApi(deps = {}) {
         try { await ensureVault(vaultRoot) } catch (e) { return degraded((e && e.message) || e) }
         let data
         try { data = await overview(vaultRoot) } catch (e) { return degraded((e && e.message) || e) }
+        // 健康面板要看「最近一次备份多久前」：这里给摘要（最多 5 个文件名，避免把响应撑大）
+        let backups = { count: 0, newestAt: 0, files: [] }
+        try {
+          const dir = path.join(vaultRoot, 'backups')
+          const names = fs.readdirSync(dir).filter((n) => /\.db$/i.test(n)).sort()
+          let newestAt = 0
+          if (names.length) {
+            const newest = names[names.length - 1]
+            try { newestAt = fs.statSync(path.join(dir, newest)).mtimeMs } catch { newestAt = 0 }
+          }
+          backups = { count: names.length, newestAt, files: names.slice(-5) }
+        } catch { /* 没有备份目录 */ }
         json(res, 200, {
           ok: true,
           vaultDir: vaultRoot,
           version: servingVersion,
           db: dbStatus(vaultRoot),
+          backups,
           ...data,
         })
         return
@@ -888,7 +902,7 @@ export function createApi(deps = {}) {
           autoWeb: cfg.autoWeb ?? env.MEMORY_AUTO_WEB !== '0', autoWebMode: cfg.autoWebMode || 'init', webPort: Number(cfg.webPort) || Number(env.MEMORY_WEB_PORT) || 7999, webCheckIntervalMs: Number(cfg.webCheckIntervalMs) || 5000, webMaxRestart: Number(cfg.webMaxRestart) || 10, watchdogAutoSpawn: cfg.watchdogAutoSpawn ?? env.MEMORY_WATCHDOG === '1', autoRestartOnDrift: cfg.autoRestartOnDrift ?? true, autoMcpSetup: cfg.autoMcpSetup ?? false,
           auditMode: cfg.auditMode ?? 'all', auditExemptAgents: cfg.auditExemptAgents ?? [], auditExemptKinds: cfg.auditExemptKinds ?? [], recycleRetentionDays: cfg.recycleRetentionDays ?? 30,
           vaultProfiles: Array.isArray(cfg.vaultProfiles) ? cfg.vaultProfiles : [], activeVault: cfg.activeVault || '',
-          enabled: cfg.enabled !== false, vaultDir: cfg.vaultDir || '', captureProvider: cfg.captureProvider || '', captureFollowSession: cfg.captureFollowSession === true, captureFirstOnly: cfg.captureFirstOnly === true,
+          enabled: cfg.enabled !== false, vaultDir: cfg.vaultDir || '', captureProvider: cfg.captureProvider || '', captureFollowSession: cfg.captureFollowSession === true, captureFirstOnly: cfg.captureFirstOnly === true, captureTimeoutMs: cfg.captureTimeoutMs ?? 45000,
           captureModel: cfg.captureModel || '', recallEmbedding: cfg.recallEmbedding || '', sessionBudgetChars: cfg.sessionBudgetChars ?? 80000,
           webPort: cfg.webPort ?? 7999, webCheckIntervalMs: cfg.webCheckIntervalMs ?? 5000, webMaxRestart: cfg.webMaxRestart ?? 10,
         }

package/lib/capture.js

@@ -376,7 +376,12 @@ export async function summarizeTurnDetailed(llm, route, conversation, opts = {})
     }
     const why = out ? '模型输出无法解析为 JSON' : '模型返回空文本'
     // 括号没闭合 = 截断的强特征：即便宿主没给 max-tokens finish，也把方向指对（#18 建议 1）
-    const hint = out && looksTruncatedJson(out) ? '（JSON 括号未闭合，疑似输出被截断）' : ''
+    // issue #48：本地模型（ollama）不会可靠地给 max-tokens finish，输出被截断时只表现为「JSON 未闭合」。
+    // 给独立错误码，让调用方走「加大预算重试同一候选」的阶梯，而不是直接换模型。
+    if (out && looksTruncatedJson(out)) {
+      return { failure: { code: 'TRUNCATED_OUTPUT', message: '输出被截断（JSON 未闭合）—— 模型侧输出上限/超时导致的半截 JSON：调大「蒸馏输出上限」或缩短会话输入；本地模型也可能撞到 num_predict 上限' } }
+    }
+    const hint = ''
     // #24：JSON.parse 的原始报错（如 Bad control character in string literal at position N）
     // 以前被两个裸 catch{} 完全吞掉，排查只能靠猜；现场摘要也从「只留前 160 字」改成首尾各留一段。
     const raw = parsed.error ? '（JSON.parse 原始报错：' + parsed.error + '）' : ''