memory-eternal 0.9.7 → 0.9.9

package/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "$schema": "https://anthropic.com/claude-code/plugin.schema.json",
   "name": "memory-eternal",
-  "version": "0.9.7",
+  "version": "0.9.9",
   "description": "璁板繂鏍稿績锛圡emory Eternal锛夛細鐭ヨ瘑鍗¤嚜鍔ㄦ矇娣€/鍙洖锛岀粺涓€ ~/.dsh/memory-vault锛屼細璇濈粨鏉熻嚜鍔ㄥ叆搴撹繘寰呭鏍搞€?,
   "author": {
     "name": "EternalNight996",

package/.codex-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "memory-eternal",
-  "version": "0.9.7",
+  "version": "0.9.9",
   "description": "璁板繂鏍稿績锛圡emory Eternal锛夛細鐭ヨ瘑鍗¤嚜鍔ㄦ矇娣€/鍙洖锛岀粺涓€ ~/.dsh/memory-vault锛屼細璇濈粨鏉熻嚜鍔ㄥ叆搴撹繘寰呭鏍搞€?,
   "author": {
     "name": "EternalNight996",

package/.cursor-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "memory-eternal",
-  "version": "0.9.7",
+  "version": "0.9.9",
   "description": "璁板繂鏍稿績锛圡emory Eternal锛夛細鐭ヨ瘑鍗¤嚜鍔ㄦ矇娣€/鍙洖锛岀粺涓€ ~/.dsh/memory-vault锛屼細璇濈粨鏉熻嚜鍔ㄥ叆搴撹繘寰呭鏍搞€?,
   "author": {
     "name": "EternalNight996",

package/README.en.md

@@ -154,6 +154,23 @@ Three concepts, don't mix them:
 | Auto recall | on | AI auto-queries memory when it needs history |
 | Vault dir | `~/.dsh/memory-vault` | where memory lives, plain Markdown & git-able |
 
+#### Multiple vaults (per-project isolation)
+
+Two extra fields in `memory-eternal-config.json`:
+
+```json
+{
+  "vaultProfiles": [{ "name": "work", "path": "D:/vaults/work" }],
+  "activeVault": "work"
+}
+```
+
+Resolution order: `MEMORY_VAULT_DIR` env → the `path` of the `vaultProfiles` entry whose `name` matches `activeVault` → `vaultDir` → default `~/.dsh/memory-vault`.
+
+**Since v0.9.8 the CLI / MCP / hooks / standalone web / sweep share the host's exact order** — previously only the host honoured `activeVault`, so switching vaults left the terminal and hooks writing to the default vault and split your memory in two.
+
+> Note: the settings page still has no form for `vaultProfiles` / `activeVault` (edit the config file by hand), and per-workspace auto-selection is not implemented yet.
+
 ### 2. Save money (important)
 
 | Setting | Default | In plain words |
@@ -260,6 +277,8 @@ npm run build   # builds lib/client.js (DSH embed) + web/app.js (standalone web
 
 | Version | Date | Highlights |
 |---|---|---|
+| **v0.9.9** | 2026-09-28 | **Fixed the sidebar "Memory" entry being pushed out of view (issue #2)**: the slot host comes in two shapes — the shell's own horizontal row (`.footerActions`) and a `div[data-slot="sidebar.footer.action"]` that other plugins (`dsh-diff-approval`, `dsh-footer-order`, …) turn into `flex-direction: column`. The old CSS applied `flex-wrap: wrap` plus `.me-footer { flex: 1 1 100% }` to both: inside a column container `wrap` means "new column" and `basis: 100%` means **100% height**, so the button was pushed into an overflow column and clipped to a sliver. `flex-wrap` is now scoped to the horizontal `.footerActions` only and `.me-footer` uses the direction-agnostic `flex: 0 0 auto`, guarded by a new SSR assertion. Also: bilingual READMEs gained a "Community & Support" section (group QR + WeChat pay/reward codes), plus `SPONSOR.md` and `.github/FUNDING.yml` |
+| **v0.9.8** | 2026-09-28 | **Backlog burn-down release.** ① **Capture reliability** (issues #3 / #4): distillation no longer hardcodes `providers[0]` — new `captureProvider` / `captureModel` config plus ordered fallback across providers; adapter errors carried by the stream's terminal `finish` block (missing credential, rate limit, …) are now **visible failures** (code/message in the capture log, red health state, raw-card fallback annotated with the cause) instead of a vague "distillation produced nothing"; the stall detector moved from cumulative-counter deltas to a **sliding window** (alerts only when turns keep starting with zero finishes inside 20 min), fixing both false alarms and misses. ② **Independent-process vault consistency** (dsh-memory-eternal#5): CLI / MCP / hooks / standalone web / sweep now share the host's resolution order (`MEMORY_VAULT_DIR` → `activeVault` → `vaultDir` → default), so switching vaults no longer splits your memory in two. ③ **Watchdog singleton lock** (dsh-memory-eternal#6): a new process exits when a live watchdog already owns the port, and stale pid records are cleaned, ending the orphan pile-up on every restart. ④ **Graph performance**: tag edges became "Top-K by shared tags" (measured **24,193 → 2,763** edges at 593 cards), similarity switched to a **rare-bigram inverted index** (cold **948 → ~520ms**, warm 2-4ms), payload **3,531KB → 579KB**, and the client dim pass now uses an **adjacency set** (**52ms → <1ms** per frame). ⑤ Five new test files (routing/finish, stall window, vault resolution, watchdog lock, graph perf baseline); `npm test` now covers **85 checks** |
 | **v0.9.7** | 2026-09-28 | **Cards open fully expanded (and can be collapsed) + new body typography**: the reader no longer folds to 300px by default — it **opens showing the whole card**, with a one-click collapse for long ones (and an "Expand ⌄" hint when folded). Frontmatter is lifted out of the body into **colored chips** (kind / audit status / tags / dates / source; kind reuses the graph palette) instead of leaking raw YAML. Body rendering is rewritten (`src/client/markdown.js`, shared by the host and the standalone web page): headings get an **emoji + color group by meaning** (conclusion ✅green / risk ⚠️amber / root cause 🔍blue / plan 🛠️purple / example 💡pink), consecutive `- ` lines become a real `<ul>`, `1.` becomes an `<ol>` with numbered circles, `- [x]` becomes a checkbox, `>` becomes an emoji callout, `---` becomes a divider, short `key: value` lines get an aligned style, and `**bold**` gets a highlighter underline; fenced code is protected so `#` / `-` inside it are no longer re-typeset. **Adds 7 typography unit tests** (XSS escaping, link-protocol allowlist, prose colons must not be misread as key/value). Also carries the PR #8 port: Electron-host subprocess node resolution (`lib/node-bin.js`, fixing "standalone web server never starts → sidebar popup white screen") and API routes registered after `webServer` is ready (fixing "settings page stuck loading") |
 | **v0.9.6** | 2026-09-28 | **Fixed the full-screen overlay covering the desktop shell's window controls (clicking Close quit the whole shell)**: the desktop shell draws minimize / maximize / close at the top of the page and, per the DSH convention, marks `<html>` with `data-windows-titlebar` and publishes `--dsh-windows-titlebar-height` (both confirmed present inside the official desktop build's `app.asar`); the memory overlay used `inset:0` plus an inline `100vh`, covering that whole strip, and our own top-right "×" landed exactly on the shell's window Close — one click quit the entire desktop shell. The full-screen state now uses the `.me-modal-full` class (no inline `100vh`, which CSS cannot override), and under `[data-windows-titlebar]` the `.me-overlay-top` / `.me-overlay` / `.me-modal-full` surfaces are pushed down by `--dsh-windows-titlebar-height` with height `calc(100vh - that)`. Browsers without the attribute get a 0px offset, so behaviour is unchanged there. The export-graph full-screen preview also moved from `100vh` to `100%` so it cannot overflow the shortened overlay. Adds 1 render regression asserting all three avoidance rules ship in the bundle |
 | **v0.9.5** | 2026-09-28 | **Fixed the plugin failing to mount under the official desktop build (schemastery 3.18.4)**: the desktop profile resolves **schemastery 3.18.4**, which turns every `volatile` Config field into a cosmokit **live reference** (`{ get(): snapshot }` branded with `Symbol.for('cosmokit.volatile.write')`), while 0.9.4 treated those fields as plain values — `cfg.vaultDir.trim()` threw `TypeError: cfg.vaultDir.trim is not a function` and the plugin died inside `apply` (the exact desktop error). The plugin now **deep-unwraps** references per the cosmokit protocol (snapshot each reference, recurse through arrays/objects), so `settings.get()` only ever exposes plain values; each read unwraps afresh, so in-place volatile hot updates stay immediately visible, and `JSON.stringify` no longer writes `{}` into the shared config file. Adds 2 regressions (the real 3.18.4 reference shape + nested arrays/objects) and verifies `apply` under the **desktop build's real module-resolution layout** (schemastery 3.18.4 + real cosmokit) |
@@ -278,6 +297,23 @@ npm run build   # builds lib/client.js (DSH embed) + web/app.js (standalone web
 
 ---
 
+## 💬 Community & Support
+
+**Questions, feature requests, or just want to talk about DSH plugin development — join the group chat:**
+
+<img src="https://raw.githubusercontent.com/EternalNight996/memory-eternal/main/assets/support/group-qr.jpg" width="260" alt="DeepSeek Harness community group QR code" />
+
+**If this plugin saved you some time, you can buy the author a coffee (either code works):**
+
+| WeChat Pay | WeChat Reward |
+|---|---|
+| <img src="https://raw.githubusercontent.com/EternalNight996/memory-eternal/main/assets/support/wechat-pay.jpg" width="230" alt="WeChat Pay QR code" /> | <img src="https://raw.githubusercontent.com/EternalNight996/memory-eternal/main/assets/support/wechat-reward.jpg" width="230" alt="WeChat reward QR code" /> |
+
+> Donations are entirely optional and change nothing about the plugin; it stays open source and free (MIT).
+> Picking the first custom entry in the repo's **Sponsor** dropdown lands on [SPONSOR.md](SPONSOR.md), which shows the same codes.
+
+---
+
 ## 📄 License
 
 MIT

package/README.md

@@ -165,6 +165,23 @@ dsh-memory watchdog [--port 7799]    # 看门狗保活 web（独立进程）
 | 自动召回 | 开 | AI 需要历史时自动帮你查记忆 |
 | 记忆库目录 | `~/.dsh/memory-vault` | 记忆存哪，纯 Markdown 可 git 管理 |
 
+#### 多库（按项目隔离记忆）
+
+配置里还有两个字段（写在 `memory-eternal-config.json`，或设置面板同步过来的共享配置）：
+
+```json
+{
+  "vaultProfiles": [{ "name": "work", "path": "D:/vaults/work" }],
+  "activeVault": "work"
+}
+```
+
+解析顺序：`MEMORY_VAULT_DIR` 环境变量 → `activeVault` 命中 `vaultProfiles[].name` 的 `path` → `vaultDir` → 默认 `~/.dsh/memory-vault`。
+
+**v0.9.8 起，CLI / MCP / hooks / 独立 web / sweep 与 DSH 宿主共用同一套顺序**：以前只有宿主认 `activeVault`，切库后终端与 hooks 仍往默认库写，记忆会被劈成两份。
+
+> 说明：配置页目前还没有 `vaultProfiles` / `activeVault` 的可视化表单（要手工写配置文件），按 workspace 自动选库也还没做。
+
 ### 二、省钱包（重要）
 
 | 配置项 | 默认 | 大白话说明 |
@@ -273,6 +290,8 @@ npm run build   # 构建 lib/client.js（DSH 内嵌）+ web/app.js（独立 web
 
 | 版本 | 日期 | 关键改动 |
 |---|---|---|
+| **v0.9.9** | 2026-09-28 | **修复侧边栏「记忆」入口被挤出可视区（issue #2）**：槽位宿主有两种形态 —— 壳自己的横向行容器 `.footerActions`，以及被其他插件（`dsh-diff-approval` / `dsh-footer-order` 等）改成 `flex-direction: column` 的 `div[data-slot="sidebar.footer.action"]`。旧 CSS 对两者都套了 `flex-wrap: wrap` + `.me-footer { flex: 1 1 100% }`：纵向容器里 wrap 的语义是「换列」、`basis: 100%` 指的是**高度 100%**，叠加后按钮被挤进右侧溢出列、裁切到只剩右边缘一条缝。现在 `flex-wrap` 只留给横向 `.footerActions`，`.me-footer` 改用方向无关的 `flex: 0 0 auto`，并新增 SSR 回归断言。另：README 双语新增「交流群 & 打赏」（群二维码 + 微信收款码 / 赞赏码），新增 `SPONSOR.md` 与 `.github/FUNDING.yml`（FUNDING 只支持文字链接，故 custom 第一项指向 SPONSOR.md 页面直接展示二维码） |
+| **v0.9.8** | 2026-09-28 | **集中解决存量问题**。① **采集可信性**（issue #3 / #4）：蒸馏路由不再写死 `providers[0]` —— 新增 `captureProvider` / `captureModel` 配置，并按注册顺序**依次重试**；**流末尾 `finish` 块里的适配器错误**（缺凭证 `MISSING_CREDENTIAL`、限流等）现在成为**可见失败**（沉淀日志写明 code/message + 健康态亮红 + 兜底原文卡注明原因），不再笼统记成「蒸馏无输出」；停滞探测从「累积计数差额」改为**滑动窗口**（20 分钟内持续开始且零收尾才告警），误报与漏报一并修掉。② **独立进程 vault 一致性**（dsh-memory-eternal#5）：CLI / MCP / hooks / 独立 web / sweep 与宿主共用同一套优先级（`MEMORY_VAULT_DIR` → `activeVault` → `vaultDir` → 默认库），切库后不再数据分裂。③ **看门狗单例锁**（dsh-memory-eternal#6）：同端口已有活着的 watchdog 时新进程直接退出并清理陈旧 pid 记录，止住「每次重启多一个孤儿进程」。④ **图谱性能**：标签边改「共享标签数 Top-K」（593 卡实测 **24,193 → 2,763** 条边）、相似度改**稀有 bigram 倒排候选**（冷启动 **948 → ~520ms**，warm 2-4ms）、payload **3,531KB → 579KB**；客户端 dim 判定改**邻接 Set**（每帧 **52ms → <1ms**）。⑤ 新增 5 个测试文件（路由/finish、停滞窗口、vault 解析、看门狗锁、图谱性能基线），`npm test` **85 项** |
 | **v0.9.7** | 2026-09-28 | **打开卡片默认全部展开（可收起）＋ 正文排版重做**：卡片阅读器不再默认折成 300px，而是**打开即读全文**，长卡可一键收起（折叠后底部给「展开 ⌄」提示）。frontmatter 从正文里抽出来，改成 kind / 审核状态 / 标签 / 时间 / 来源的**彩色 chips**（kind 沿用图谱配色），不再把一段 YAML 当正文露出来。正文排版换成新方案（`src/client/markdown.js`，宿主与独立 web 共用同一套）：小标题按语义自动配 **emoji + 颜色分组**（结论 ✅绿 / 风险 ⚠️琥珀 / 根因 🔍蓝 / 方案 🛠️紫 / 示例 💡粉），连续 `- ` 合成真正的 `<ul>`、`1.` 合成带圆形序号的 `<ol>`、`- [x]` 变复选框、`>` 引用升级成带 emoji 的 callout、`---` 变分隔线、行首「短键：值」渲染成对齐样式、`**加粗**` 加荧光笔下划线；围栏代码整段保护，内部 `#` / `-` 不再被误排版。**新增 7 项排版单测**（含 XSS 转义、链接协议白名单、散文冒号不得误判成 key/value）。另含 PR #8 移植：Electron 宿主下的子进程 node 解析（`lib/node-bin.js`，修「独立 web server 拉不起来 → 侧边栏弹窗白屏」）与 `webServer` 就绪后再注册 API 路由（修「设置页永远加载中」） |
 | **v0.9.6** | 2026-09-28 | **修复官方桌面版全屏浮层挡住窗口控制面板（点「关闭」会退出整个桌面壳）**：桌面壳把「最小化 / 最大化 / 关闭」画在页面顶部，并按 DSH 约定在 `<html>` 上打 `data-windows-titlebar` + 给出 `--dsh-windows-titlebar-height`（已在官方桌面版 `app.asar` 里确认这两个标记确实存在）；记忆库全屏浮层原本是 `inset:0` + 内联 `100vh`，整条盖住标题栏，而我们右上角的「×」恰好压在窗口「关闭」上 —— 用户一点就直接退出整个桌面壳。现全屏态改用 `.me-modal-full` 类（不再写死内联 `100vh`：内联样式无法被 CSS 覆盖），并在 `[data-windows-titlebar]` 下把 `.me-overlay-top` / `.me-overlay` / `.me-modal-full` 一律下移 `--dsh-windows-titlebar-height`、高度改为 `calc(100vh - 该高度)`；浏览器里没有该属性时偏移 0px，行为与之前完全一致。导出图谱的全屏预览同步从 `100vh` 改为 `100%`，避免在缩短后的浮层里溢出。新增 1 项渲染回归断言（三处让位规则必须出现在打包产物中） |
 | **v0.9.5** | 2026-09-28 | **修复官方桌面版（schemastery 3.18.4）下插件整体挂不上**：桌面版 profile 解析到的是 **schemastery 3.18.4**，它会把标了 `volatile` 的 Config 字段解析成 cosmokit 的**活引用**（`{ get(): snapshot }`，品牌 `Symbol.for('cosmokit.volatile.write')`），而 0.9.4 的业务代码把字段当普通值用 —— `cfg.vaultDir.trim()` 当场抛 `TypeError: cfg.vaultDir.trim is not a function`，插件在 `apply` 里就崩（桌面版实测报错原文）。现按 cosmokit 协议统一**深解引用**（引用取快照、数组/对象递归展开），`settings.get()` 对外只给普通值；每次读取都重新解，loader 原地提交的 volatile 热更新照样立刻可见，`JSON.stringify` 写共享配置也不再变成 `{}`。新增 2 项回归（3.18.4 真实引用形态 + 嵌套数组/对象），并用**桌面版真实模块解析布局**（schemastery 3.18.4 + 真 cosmokit）跑通 `apply` 验证 |
@@ -291,6 +310,23 @@ npm run build   # 构建 lib/client.js（DSH 内嵌）+ web/app.js（独立 web
 
 ---
 
+## 💬 交流群 & 打赏
+
+**遇到问题、想提需求、或者想聊聊 DSH 插件开发 —— 欢迎进群：**
+
+<img src="https://raw.githubusercontent.com/EternalNight996/memory-eternal/main/assets/support/group-qr.jpg" width="260" alt="DeepSeek Harness 交流群二维码" />
+
+**如果这个插件帮你省了时间，可以请作者喝杯咖啡（两个码任选其一）：**
+
+| 微信收款码 | 微信赞赏码 |
+|---|---|
+| <img src="https://raw.githubusercontent.com/EternalNight996/memory-eternal/main/assets/support/wechat-pay.jpg" width="230" alt="微信收款码" /> | <img src="https://raw.githubusercontent.com/EternalNight996/memory-eternal/main/assets/support/wechat-reward.jpg" width="230" alt="微信赞赏码" /> |
+
+> 打赏完全自愿，不影响任何功能；插件本体始终开源免费（MIT）。
+> 仓库页的 **Sponsor** 按钮下拉里选 custom 第一项，会落到 [SPONSOR.md](SPONSOR.md)（同样是这一页的码）。
+
+---
+
 ## 📄 License
 
 MIT

package/index.js

@@ -27,7 +27,8 @@ import z from '@deepseek-ai/schemastery'
 import { defineTool } from '@deepseek-ai/dsh-tools'
 import { ensureVault, search } from './lib/vault.js'
 import { migrateFromMarkdown, setAuditConfig, backupDb } from './lib/db.js'
-import { summarizeTurn, extractLastTurn, sliceNewEvents, sessionEvents, sessionEventApi, createCaptureHealth, resolveRoute, captureCard, captureUpdate, pickNeighbors } from './lib/capture.js'
+import { evaluateStall, MAX_STAMPS } from './lib/stall.js'
+import { summarizeTurn, summarizeTurnDetailed, routeCandidates, extractLastTurn, sliceNewEvents, sessionEvents, sessionEventApi, createCaptureHealth, resolveRoute, captureCard, captureUpdate, pickNeighbors } from './lib/capture.js'
 import { createApi, json, encodeBody } from './lib/api.js'
 import { appendCaptureLog, readCaptureLog, rotateCaptureLog } from './lib/capture-log.js'
 import { nodeBinary, childEnv } from './lib/node-bin.js'
@@ -57,6 +58,10 @@ export const Config = z.object({
   distillEnabled: z.boolean().default(true),
   // 语义去重：true=把已有卡索引喂 LLM 决定「新建 vs 追加」；false=纯词法去重（省一次蒸馏前的 LLM 调用）
   dedupByLLM: z.boolean().default(true),
+  // 蒸馏用的 provider / model（留空 = 按注册顺序自动挑，见 routeCandidates）
+  // 多 provider 环境下务必显式指定，否则可能撞上没配 key 的官方 provider（issue #3）
+  captureProvider: z.string().default(''),
+  captureModel: z.string().default(''),
   // 蒸馏单次输出上限（token），越高越准越贵
   captureMaxTokens: z.number().min(100).max(4000).default(900),
   // 召回相关性阈值（minScore），越高召回越少越精越省
@@ -394,13 +399,30 @@ export function apply(ctx, config) {
         else logCapture(sessionId, 'fail', `写卡失败：${out.reason || '未知原因'}`)
         return
       }
-      const route = await resolveRoute(llm)
-      if (!route) { logCapture(sessionId, 'fail', '取不到模型路由（llm.listProviders 为空）'); return }
+      // 候选路由：显式配置优先，其余按注册顺序兜底（多 provider 下 providers[0] 可能是没配 key 的官方 provider → issue #3）。
+      const routes = await routeCandidates(llm, { provider: cfg.captureProvider, model: cfg.captureModel })
+      if (!routes.length) { logCapture(sessionId, 'fail', '取不到模型路由（llm.listProviders 为空）'); return }
       // 语义去重近邻：把已有卡片索引喂给模型，让模型决定新建 vs 追加。
       // 成本控制：dedupByLLM=false 时跳过喂 LLM 的近邻采样（纯词法去重兜底）。
       const draft = { title: '', body: text.slice(0, 400) }
       const neighbors = cfg.dedupByLLM === false ? [] : await pickNeighbors(vaultDir(), draft, 8)
-      const result = await summarizeTurn(llm, route, text, { signal: AbortSignal.timeout(45000), existing: neighbors, maxTokens: cfg.captureMaxTokens ?? 900 })
+      let route = routes[0]
+      let result = null
+      let failure = null
+      for (let i = 0; i < routes.length; i++) {
+        route = routes[i]
+        const detailed = await summarizeTurnDetailed(llm, route, text, { signal: AbortSignal.timeout(45000), existing: neighbors, maxTokens: cfg.captureMaxTokens ?? 900 })
+        if (detailed.card !== undefined) { result = detailed.card; failure = null; break }
+        if (detailed.failure) {
+          failure = { ...detailed.failure, provider: route.provider, model: route.model }
+          const more = i < routes.length - 1
+          logCapture(sessionId, 'fail', `蒸馏调用失败（${route.provider}/${route.model}）：${failure.code} ${failure.message}${more ? ' → 换下一个 provider 重试' : ''}`)
+          if (more) continue
+        }
+        break // skip（太短 / 不值得保存）或没有更多候选
+      }
+      // 真失败要亮红并进提示段：此前只有笼统的「蒸馏无输出」，真实原因（如缺凭证）完全不可见。
+      if (failure) { health.fail(`蒸馏失败（${failure.provider}）：${failure.code} ${failure.message}`); touchPrompt() }
       if (!result) {
         // 蒸馏失败不能让内容白丢：退成原文卡（与「关闭蒸馏」同一条降级路径）。
         const raw = await captureCard(vaultDir(), {
@@ -414,7 +436,8 @@ export function apply(ctx, config) {
           severity: 'info',
           reason: 'AI 自动沉淀（蒸馏无输出 → 原文卡兜底）',
         }, { threshold: cfg.dedupThreshold })
-        if (raw.ok) { countWrite(); logCapture(sessionId, 'created', '蒸馏无输出 → 原文卡兜底', { path: raw.path ?? raw.rel, kind: 'content', model: route.model }) }
+        const why = failure ? `${failure.code} ${failure.message}` : '无可用产出'
+        if (raw.ok) { countWrite(); logCapture(sessionId, 'created', `蒸馏失败（${why}）→ 原文卡兜底`, { path: raw.path ?? raw.rel, kind: 'content', model: route.model }) }
         else if (raw.duplicate) { countWrite(); logCapture(sessionId, 'appended', '蒸馏无输出 + 与已有卡重复 → 追加更新', { path: raw.duplicate.path, model: route.model }) }
         else logCapture(sessionId, 'fail', `蒸馏无输出，且兜底原文卡也失败：${raw.reason || '未知原因'}`, { model: route.model })
         return
@@ -469,10 +492,14 @@ export function apply(ctx, config) {
   // 就说明收尾事件没到——这是「监听器整个没被调用」这类静默死亡的兜底探测。
   let turnsStarted = 0
   let turnsStopped = 0
-  let lastClaimAt = 0
-  ctx.on('agent/inbox/claimed', () => { turnsStarted += 1; lastClaimAt = Date.now() })
+  // 滑动窗口用的时间戳（旧版只留累积计数，导致误报+漏报，见 lib/stall.js）
+  const claimedAt = []
+  const stoppedAt = []
+  const stamp = (arr) => { arr.push(Date.now()); if (arr.length > MAX_STAMPS) arr.shift() }
+  ctx.on('agent/inbox/claimed', () => { turnsStarted += 1; stamp(claimedAt) })
   ctx.on('agent/turn-stopping', ({ agent }) => {
     turnsStopped += 1
+    stamp(stoppedAt)
     const sessionId = agent?.session?.id ?? agent?.id ?? 'unknown'
     try {
       const api = sessionEventApi(agent?.session)
@@ -594,14 +621,12 @@ export function apply(ctx, config) {
   // 沉淀停滞探测：只有「有轮次在跑」且「管线 15 分钟一个活口都没有」才报（真死才报）。
   // 计数器对不上是常态（子代理轮次、被取消的轮次都不发 turn-stopping），单看计数会误报刷屏。
   const stallTimer = setInterval(() => {
-    const idleMs = Date.now() - lastActivityAt
-    const claimIdleMs = lastClaimAt ? Date.now() - lastClaimAt : 0
-    const suspicious = turnsStarted > turnsStopped && claimIdleMs > 15 * 60 * 1000 && idleMs > 15 * 60 * 1000
-    if (suspicious && !stallAlerted) {
+    const { alert, started, stopped, windowMs } = evaluateStall({ claimedAt, stoppedAt })
+    if (alert && !stallAlerted) {
       stallAlerted = true
-      logCapture('system', 'fail', `轮次收尾事件未触发（agent/turn-stopping 没到）：已开始 ${turnsStarted} / 已收尾 ${turnsStopped}，且 15 分钟无任何沉淀活动——DSH 可能改了事件名或作用域`)
+      logCapture('system', 'fail', `轮次收尾事件疑似失效：最近 ${Math.round(windowMs / 60000)} 分钟内开始 ${started} 个轮次、收尾 ${stopped} 个（累计 ${turnsStarted}/${turnsStopped}）——DSH 可能改了事件名或作用域`)
     }
-  }, 10 * 60 * 1000)
+  }, 5 * 60 * 1000)
   ctx.effect(() => () => clearInterval(stallTimer), 'memory-eternal: capture stall watch')
 
   // 回收站清理：每 30 分钟永久删除超过保留期（默认 30 天）的软删卡。

package/lib/capture-run.js

@@ -8,12 +8,31 @@
 import path from 'node:path'
 import os from 'node:os'
 import fs from 'node:fs'
-import { captureCard, captureUpdate, pickNeighbors, summarizeTurn } from './capture.js'
+import { captureCard, captureUpdate, pickNeighbors, summarizeTurnDetailed, routeCandidates } from './capture.js'
 import { createLlmFromEnv } from './llm-openai.js'
 
+/**
+ * 解析当前 vault 目录。
+ *
+ * 优先级必须与宿主 index.js 的 vaultDir() **完全一致**：
+ *   MEMORY_VAULT_DIR → activeVault 命中 vaultProfiles → vaultDir → 默认 ~/.dsh/memory-vault。
+ * 旧实现只认 MEMORY_VAULT_DIR / 默认路径，于是用户在宿主里切库后，CLI / MCP / hooks /
+ * 独立 web / sweep 仍然写默认库 —— 同一台机器上记忆被劈成两份（dsh-memory-eternal#5）。
+ *
+ * @param {NodeJS.ProcessEnv} [env]
+ * @returns {string} 绝对路径
+ */
 export function defaultVaultDir(env = process.env) {
   if (env.MEMORY_VAULT_DIR && env.MEMORY_VAULT_DIR.trim()) return path.resolve(env.MEMORY_VAULT_DIR.trim())
   const home = env.DSH_HOME || path.join(os.homedir(), '.dsh')
+  try {
+    const cfg = readConfig(env) || {}
+    const profiles = Array.isArray(cfg.vaultProfiles) ? cfg.vaultProfiles : []
+    const active = String(cfg.activeVault || '').trim()
+    const hit = active && profiles.find((p) => p && p.name === active)
+    if (hit && typeof hit.path === 'string' && hit.path.trim()) return path.resolve(hit.path.trim())
+    if (typeof cfg.vaultDir === 'string' && cfg.vaultDir.trim()) return path.resolve(cfg.vaultDir.trim())
+  } catch { /* 读不到共享配置就回落默认库 */ }
   return path.join(home, 'memory-vault')
 }
 
@@ -50,13 +69,20 @@ export async function runStandaloneCapture(vaultRoot, text, { source = '', llm,
   }
 
   const llmClient = llm ?? createLlmFromEnv(env)
-  const route = llmClient ? { provider: llmClient.listProviders()[0].id, model: (await llmClient.listModels())[0].id } : null
+  // 与宿主同样的候选路由：显式配置优先，其余按注册顺序兜底（别写死 providers[0]，见 issue #3）
+  const routes = llmClient ? await routeCandidates(llmClient, { provider: shared?.captureProvider, model: shared?.captureModel }) : []
+  const route = routes[0] ?? null
 
   // 有 LLM：蒸馏判定链（与 DSH 侧一致）
   if (llmClient && route) {
     try {
       const neighbors = await pickNeighbors(vaultRoot, { title: '', body: clean.slice(0, 400) }, 8)
-      const result = await summarizeTurn(llmClient, route, clean, { existing: neighbors, signal: AbortSignal.timeout(45000) })
+      const detailed = await summarizeTurnDetailed(llmClient, route, clean, { existing: neighbors, signal: AbortSignal.timeout(45000) })
+      const result = detailed.card
+      // 独立进程也必须把真实失败暴露出来（走 stderr），别再静默降级成「无输出」
+      if (detailed.failure) {
+        process.stderr.write(`[memory-eternal] 蒸馏失败（${route.provider}/${route.model}）：${detailed.failure.code} ${detailed.failure.message}\n`)
+      }
       if (result && result.save === true) {
         if (result.append_to) {
           await captureUpdate(vaultRoot, result.append_to, result.update, { threshold: 0.62 })

package/lib/capture.js

@@ -41,9 +41,23 @@ export function looksDurable(text) {
  * @returns Promise<null | { save:boolean } | { save:true, title, kind, tags, body } | { append_to, update }>
  */
 export async function summarizeTurn(llm, route, conversation, opts = {}) {
+  const r = await summarizeTurnDetailed(llm, route, conversation, opts)
+  return r.card ?? null
+}
+
+/**
+ * summarizeTurn 的「带诊断」版本：把「跳过」与「真失败」分开。
+ *
+ * 关键：适配器错误（缺凭证 / 网络 / 限流）不会抛给调用方，而是以
+ * \`{ kind:'error'|'aborted', failure:{ code, message } }\` 终结流。旧实现只收集 text 块，
+ * 于是空文本被笼统记成「蒸馏无输出」，真实原因（例如 MISSING_CREDENTIAL）完全不可见。
+ *
+ * @returns {Promise<{card?:any, skip?:string, failure?:{code:string,message:string}}>}
+ */
+export async function summarizeTurnDetailed(llm, route, conversation, opts = {}) {
   const text = String(conversation || '').trim()
-  if (text.length < 120) return null // 太短，没有沉淀价值
-  if (!looksDurable(text)) return null // 启发式预筛：无可复用信号则跳过 LLM 判定（省 token）
+  if (text.length < 120) return { skip: 'too-short' } // 太短，没有沉淀价值
+  if (!looksDurable(text)) return { skip: 'not-durable' } // 启发式预筛：无可复用信号则跳过 LLM 判定（省 token）
 
   const existing = Array.isArray(opts.existing) ? opts.existing.slice(0, DEDUP_NEIGHBORS) : []
   const existingBlock = existing.length
@@ -86,8 +100,23 @@ export async function summarizeTurn(llm, route, conversation, opts = {}) {
     signal: opts.signal ?? AbortSignal.timeout(45000),
   }
   for await (const chunk of llm.stream(options)) assembler.push(chunk)
+  // finish 块才是失败真相（见上）；没有 finish 的兼容适配器按 stop 处理。
+  const finish = assembler.finish && typeof assembler.finish === 'object' ? assembler.finish : { kind: 'stop' }
+  if (finish.kind === 'error' || finish.kind === 'aborted') {
+    const f = finish.failure || {}
+    return {
+      failure: {
+        code: String(f.code || f.kind || finish.kind),
+        message: String(f.message || f.detail || '模型调用失败'),
+      },
+    }
+  }
   const out = assembler.blocks().filter((b) => b.type === 'text').map((b) => b.text).join('').trim()
-  return parseCaptureJson(out)
+  const parsed = parseCaptureJson(out)
+  if (!parsed) {
+    return { failure: { code: 'UNPARSEABLE_OUTPUT', message: out ? '模型输出无法解析为 JSON：' + out.slice(0, 160) : '模型返回空文本' } }
+  }
+  return { card: parsed }
 }
 
 /**
@@ -321,15 +350,48 @@ export function compressExcerpt(text, maxChars = 2400) {
   return out
 }
 
-/** 解析当前模型路由：取第一个 provider 的旗舰模型（与驯兽师一致）。 */
-export async function resolveRoute(llm) {
+/**
+ * 按优先级给出候选路由列表。
+ *
+ * 为什么不直接取 \`providers[0]\`：DSH 的 listProviders() 按**注册顺序**返回，内置的
+ * deepseek-official 通常排第一；多 provider 用户（例如自建中转）本机没配官方 key 时，
+ * 请求会在**发出去之前**就以「缺凭证」失败，而这条失败只体现在流末尾的 finish 块里
+ * —— 旧代码只看 text，于是被误报成「蒸馏无输出」，管线 100% 静默失败（issue #3）。
+ * 配置了 captureProvider / captureModel 时以配置为准，其余按注册顺序依次兜底尝试。
+ *
+ * @param {object} llm DSH llm 服务（或 lib/llm-openai.js 的兼容适配器）
+ * @param {{provider?:string, model?:string, limit?:number}} [opts]
+ * @returns {Promise<Array<{provider:string, model:string}>>}
+ */
+export async function routeCandidates(llm, opts = {}) {
+  const out = []
+  const want = String(opts.provider || '').trim()
+  const wantModel = String(opts.model || '').trim()
+  const limit = Number(opts.limit) > 0 ? Number(opts.limit) : 4
   try {
-    const providers = llm.listProviders()
-    if (!providers || providers.length === 0) return null
-    const models = await llm.listModels(providers[0].id)
-    if (!models || models.length === 0) return null
-    return { provider: providers[0].id, model: models[0].id }
-  } catch {
-    return null
-  }
+    const providers = llm.listProviders() || []
+    if (!providers.length) return out
+    if (want) {
+      const hit = providers.find((p) => p && p.id === want)
+      if (hit) {
+        const models = (await llm.listModels(hit.id)) || []
+        const model = wantModel || (models[0] && models[0].id)
+        if (model) out.push({ provider: hit.id, model })
+      }
+    }
+    for (const p of providers) {
+      if (!p || !p.id || out.some((r) => r.provider === p.id)) continue
+      const models = (await llm.listModels(p.id)) || []
+      const model = (p.id === want && wantModel) || (models[0] && models[0].id)
+      if (model) out.push({ provider: p.id, model })
+      if (out.length >= limit) break
+    }
+  } catch { /* 路由解析失败按空列表返回，由调用方兜底 */ }
+  return out
+}
+
+/** 解析当前模型路由：显式配置优先，其次第一个 provider 的首个模型（保持旧行为作兜底）。 */
+export async function resolveRoute(llm, opts = {}) {
+  const list = await routeCandidates(llm, opts)
+  return list[0] ?? null
 }