dsh-tdai-memory 0.4.1 → 0.4.3

package/README.md

@@ -1,9 +1,17 @@
 # dsh-tdai-memory
+
+## DSH 0.2.1-rc.1
+
+Supports DSH 0.2.1-rc.1. Compatibility is declared through `engines.dsh` and matching peer ranges, including this exact prerelease. The existing 0.2.0-rc.2 support is retained.
 
+The plugin follows the DSH language setting (Chinese and English in DSH 0.2.0-rc.2), including configuration, status messages and plugin-list metadata. Language-pack locales use the host fallback chain. Switching languages preserves unsaved settings; there is no separate plugin language selector.
+
 ## Configuration page (DSH 0.2.0-rc.2 and later)
 
 Open **Plugins → Installed → dsh-tdai-memory** from the homepage sidebar to configure and save this plugin. The page uses the official `plugins.bundle.config` interface, without a duplicate entry in global Settings. Web and Desktop share the page. This version requires DSH 0.2.0-rc.2 or a later 0.2.x host; existing configuration is retained.
 
+Choose **Automatic memory** for capture, extraction, recall and search together; **Search existing memory only** disables capture, extraction and automatic recall while keeping search tools; **Pause memory features** disables all four. The main page contains the two model connections. Individual switches, storage paths and tuning remain under **Advanced**. Existing mixed switch values appear as **Custom combination** and are preserved. Select a mode, Save, then restart DSH for pipeline changes (the search-tool switch applies immediately). Only edited fields are written; blank keys and untouched advanced options remain unchanged.
+
 [![中文文档](https://img.shields.io/badge/%E4%B8%AD%E6%96%87%E6%96%87%E6%A1%A3-blue)](README.zh.md)
 
 **GitHub**: [Scorp1o117/dsh-tdai-memory](https://github.com/Scorp1o117/dsh-tdai-memory) · **npm**: [dsh-tdai-memory](https://www.npmjs.com/package/dsh-tdai-memory)
@@ -57,6 +65,12 @@ The data directory reuses the existing `~/.memory-tencentdb/memory-tdai`, so
 | dsh shell | `index.js`: config mapping, `session/event` + `session/flush` capture, `system-prompt/assemble` recall injection on `agent.ctx`, tool registration, lifecycle |
 | Fallback | `recall-inject.js`: preset-row recall injection (used when mounted inside an agent preset) |
 
+Recall caches expire after 30 seconds and retain at most 128 sessions per
+injection instance. Failed or timed-out recalls are retried on the next assembly.
+The preset row also accepts `timeoutMs` (default 4000); reaching this deadline
+continues prompt assembly without memory context. Deadline timers are cleared
+when recall settles early. Underlying recall work may continue after the deadline.
+
 Hard-won wiring details:
 
 - **Capture**: `session/flush` listener (await semantics; must complete before

package/README.zh.md

@@ -1,9 +1,17 @@
 # dsh-tdai-memory
+
+## DSH 0.2.1-rc.1
+
+支持 DSH 0.2.1-rc.1，通过 `engines.dsh` 和一致的 peer 范围声明兼容，明确包含该预发布版本，并保留已有 0.2.0-rc.2 支持。
 
+插件跟随 DSH 的语言设置（DSH 0.2.0-rc.2 内置中文和 English），配置页面、状态提示和插件列表名称/简介同步切换。扩展语言使用宿主的回退链。切换语言保留未保存的设置，无需单独选择插件语言。
+
 ## 配置入口（DSH 0.2.0-rc.2 起）
 
 在首页侧边栏打开 **插件 → 已安装 → dsh-tdai-memory**，直接在插件详情页配置并保存。配置页注册到官方的 `plugins.bundle.config` 接口；全局设置页不再重复显示配置入口。Web 与桌面版使用相同界面，本版要求 DSH 0.2.0-rc.2 或更新的 0.2.x 版本。现有配置无需迁移。
 
+选择 **自动记忆**，一次开启捕获、提取、召回和搜索；**仅搜索已有记忆** 关闭捕获、提取和自动召回，保留搜索工具；**暂停记忆功能** 关闭全部四项。主页面只展示两个模型的连接信息，独立开关、存储路径和调优参数保留在 **高级设置** 中。已有开关组合显示为 **自定义组合**，不会自动覆盖。选择模式后保存，并重启 DSH 使记忆流水线设置生效（搜索工具开关即时生效）。只写入编辑过的字段，密钥留空及未编辑的高级设置保持原值。
+
 **GitHub**: [Scorp1o117/dsh-tdai-memory](https://github.com/Scorp1o117/dsh-tdai-memory) · **npm**: [dsh-tdai-memory](https://www.npmjs.com/package/dsh-tdai-memory) · [English](README.md)
 
 [![Enhancement Suite](https://img.shields.io/badge/part%20of-Enhancement%20Suite-3964fe)](https://github.com/Scorp1o117/dsh-enhancement-suite) [![npm](https://img.shields.io/npm/v/dsh-enhancement-suite)](https://www.npmjs.com/package/dsh-enhancement-suite)
@@ -52,6 +60,11 @@ dsh plugin --profile desktop add dsh-tdai-memory@0.3.6
 | dsh 壳 | `index.js`：配置映射、`session/event`+`session/flush` 捕获、`agent.ctx` 上的 `system-prompt/assemble` waterfall 召回注入、工具注册、生命周期 |
 | 备用 | `recall-inject.js`：preset 行形态的召回注入（挂在 agent preset 内时用） |
 
+召回缓存默认 30 秒过期，每个注入实例最多保留 128 个会话。失败或超时的召回
+会在下次组装时重试。preset 行还支持 `timeoutMs`（默认 4000），到达期限后
+不注入记忆并继续组装提示词；提前完成时会清除超时定时器。期限到达后，底层
+召回任务可能仍会继续运行。
+
 关键接线（都是踩坑换来的）：
 
 - **捕获**：`session/flush` 监听器（await 语义，headless 退出前必完成）；`turn/start` 时间戳做 L0 cursor 下限；turn id 去重

package/client.js

@@ -36,6 +36,35 @@ window.__ModuleLoader__.load({
       ".__tm_status{font-size:12px;color:var(--dsw-alias-label-tertiary)}" +
       ".__tm_error{font-size:12px;color:var(--dsw-alias-state-error-primary)}" +
       ".__tm_unavailable{font-size:13px;color:var(--dsw-alias-label-tertiary)}";
+    // Scoped flat controls retain native keyboard and form behavior.
+    CSS += `
+.dsh-flat.__tm_root{width:100%;max-width:720px;gap:14px;font-size:13px;line-height:1.65;color:var(--dsw-alias-label-primary);--flat-accent:var(--dsw-alias-state-business-primary,#3964fe);--flat-border:var(--dsw-alias-border-l2,#dce2eb)}
+.dsh-flat.__tm_root *{box-sizing:border-box;min-width:0}
+.dsh-flat.__tm_root p{margin:0}
+.dsh-flat.__tm_root label[class$="_field"]{gap:7px}
+.dsh-flat.__tm_root [class$="_label"]{font-size:13px;font-weight:500}
+.dsh-flat.__tm_root [class$="_hint"]{font-size:12px;line-height:1.65}
+.dsh-flat.__tm_root input:not([type=checkbox]),.dsh-flat.__tm_root select,.dsh-flat.__tm_root textarea{width:100%;border:1px solid var(--flat-border);border-radius:6px;background:var(--dsw-alias-bg-layer-3);color:inherit;font:inherit;padding:9px 12px;min-height:40px;box-shadow:none;transition:border-color .15s}
+.dsh-flat.__tm_root input:hover:not(:disabled),.dsh-flat.__tm_root select:hover:not(:disabled),.dsh-flat.__tm_root textarea:hover:not(:disabled){border-color:var(--dsw-alias-label-tertiary)}
+.dsh-flat.__tm_root select{appearance:none;padding-right:34px;background-image:url("data:image/svg+xml,%3Csvg xmlns='http://www.w3.org/2000/svg' width='12' height='8' viewBox='0 0 12 8'%3E%3Cpath d='m2 2 4 4 4-4' fill='none' stroke='%23778091' stroke-width='1.5'/%3E%3C/svg%3E");background-repeat:no-repeat;background-position:right 12px center}
+.dsh-flat.__tm_root input[type=checkbox]{appearance:none;flex:none;width:30px;height:18px;margin:0;border:1px solid var(--flat-border);border-radius:12px;background:var(--dsw-alias-bg-layer-2);position:relative;cursor:pointer;transition:background .15s,border-color .15s}
+.dsh-flat.__tm_root input[type=checkbox]::before{content:"";position:absolute;left:2px;top:2px;width:12px;height:12px;border-radius:50%;background:var(--dsw-alias-label-secondary);transition:transform .15s}
+.dsh-flat.__tm_root input[type=checkbox]:checked{background:var(--flat-accent);border-color:var(--flat-accent)}
+.dsh-flat.__tm_root input[type=checkbox]:checked::before{transform:translateX(12px);background:#fff}
+.dsh-flat.__tm_root :is(input,select,textarea,button,summary,a):focus-visible{outline:2px solid var(--flat-accent);outline-offset:3px}
+.dsh-flat.__tm_root :is(input,select,textarea,button):disabled{opacity:.5;cursor:default}
+.dsh-flat.__tm_root [class$="_actions"]{flex-wrap:wrap;gap:10px;margin-top:4px;padding-top:16px;border-top:1px solid var(--flat-border)}
+.dsh-flat.__tm_root details{border-top:1px solid var(--flat-border);padding:0}
+.dsh-flat.__tm_root summary{list-style:none;display:flex;align-items:center;justify-content:space-between;gap:12px;padding:14px 0;font-size:13px;font-weight:500;cursor:pointer;color:var(--dsw-alias-label-secondary)}
+.dsh-flat.__tm_root summary::-webkit-details-marker{display:none}
+.dsh-flat.__tm_root summary::after{content:"+";font-size:18px;font-weight:400;flex:none}
+.dsh-flat.__tm_root details[open]>summary::after{content:"−"}
+.dsh-flat.__tm_root details>div{padding-bottom:18px}
+.dsh-flat.__tm_root .__tm_group{border:0;font-weight:600;font-size:14px;margin:8px 0 0;padding:0}
+@media(prefers-reduced-motion:reduce){.dsh-flat.__tm_root *,.dsh-flat.__tm_root input[type=checkbox]::before{transition:none}}
+.dsh-flat.__tm_root button{border-radius:6px;min-height:34px;padding:7px 14px;font:inherit;font-size:12px;box-shadow:none}
+.dsh-flat.__tm_root :is(h2,h3){margin:0;font-size:14px;font-weight:600}
+`;
     var tagId = "dsh-tdai-memory/main.css";
     if (typeof document !== "undefined" && document.querySelector("style[data-plugin-css=" + JSON.stringify(tagId) + "]") === null) {
       var tag = document.createElement("style");
@@ -50,10 +79,18 @@ window.__ModuleLoader__.load({
     var inject = ["slots", "locale", "configForms"];
     var zh = {
       nav: "记忆",
-      intro: "TDAI 记忆配置：L0 捕获 → L1 结构化提取 → 召回注入。密钥只写不读。TdaiCore 在启动时构建，修改后需重启生效。",
+      mode: "记忆模式",
+      modeAuto: "自动记忆（推荐）",
+      modeSearch: "仅搜索已有记忆",
+      modePaused: "暂停记忆功能",
+      modeCustom: "自定义组合",
+      modeHint: "自动：捕获、提取、召回和搜索全部开启；仅搜索：不记录新对话、不自动注入记忆；暂停：四项全部关闭。选择后点击保存。",
+      advanced: "高级设置：独立开关、存储与请求参数",
+      setupHint: "填写提取模型和向量模型的连接信息，再保存。已有自定义设置会保留；密钥留空不修改。",
+      intro: "记录对话、整理长期记忆，并在相关对话中自动召回。保存后请重启 DSH 生效；密钥只写不读。",
       groupData: "数据",
-      groupLlm: "提取 LLM（L1/L2/L3）",
-      groupEmbedding: "向量 Embedding",
+      groupLlm: "记忆整理模型",
+      groupEmbedding: "向量检索模型",
       groupCapture: "捕获",
       groupExtraction: "提取",
       groupRecall: "召回",
@@ -95,10 +132,18 @@ window.__ModuleLoader__.load({
     };
     var en = {
       nav: "Memory",
-      intro: "TDAI memory config: L0 capture → L1 extraction → recall injection. Keys are write-only. TdaiCore is built at startup; changes apply after a restart.",
+      mode: "Memory mode",
+      modeAuto: "Automatic memory (recommended)",
+      modeSearch: "Search existing memory only",
+      modePaused: "Pause memory features",
+      modeCustom: "Custom combination",
+      modeHint: "Automatic enables capture, extraction, recall and search. Search only records no new conversations and injects no memory. Pause disables all four. Click Save to apply your selection.",
+      advanced: "Advanced: individual switches, storage and request options",
+      setupHint: "Enter the extraction and embedding connections, then save. Existing custom settings are preserved; blank keys stay unchanged.",
+      intro: "Record conversations, organize long-term memory and recall it in relevant chats. Restart DSH after saving to apply changes; keys are write-only.",
       groupData: "Data",
-      groupLlm: "Extraction LLM (L1/L2/L3)",
-      groupEmbedding: "Embeddings",
+      groupLlm: "Memory extraction model",
+      groupEmbedding: "Memory search embeddings",
       groupCapture: "Capture",
       groupExtraction: "Extraction",
       groupRecall: "Recall",
@@ -165,6 +210,9 @@ window.__ModuleLoader__.load({
       { path: ["toolsEnabled"], label: "fieldToolsEnabled", type: "checkbox", group: "groupTools" }
     ];
     FIELDS.forEach(function (f) { f.key = f.path.join("."); });
+    var BASIC_KEYS = ["llm.baseUrl", "llm.apiKey", "llm.model", "embedding.baseUrl", "embedding.apiKey", "embedding.model"];
+    var MODE_KEYS = ["captureEnabled", "extraction.enabled", "recall.enabled", "toolsEnabled"];
+    var MODES = { auto: [true, true, true, true], search: [false, false, false, true], paused: [false, false, false, false] };
 
     function getPath(obj, path) {
       var cur = obj;
@@ -176,6 +224,7 @@ window.__ModuleLoader__.load({
     }
 
     function MemorySection(props) {
+      useLocale(props.locale);
       var t = props.t;
       var scope = props.scope;
       var [snapshot, setSnapshot] = react.useState(function () { return scope.getSnapshot(); });
@@ -204,7 +253,7 @@ window.__ModuleLoader__.load({
       // call, so depending on snapshot.value would reset user input on every
       // render (typing appears dead).
       react.useEffect(function () {
-        if (ready) setDraft(Object.assign({}, valueToDraft(snapshot.value)));
+        if (ready) setDraft({});
         // eslint-disable-next-line react-hooks/exhaustive-deps
       }, [ready]);
 
@@ -226,6 +275,23 @@ window.__ModuleLoader__.load({
         setNotice(null);
         setError(null);
       }
+      function memoryMode() {
+        var flags = MODE_KEYS.map(function (key) {
+          return Object.prototype.hasOwnProperty.call(draft, key) ? draft[key] : getPath(value, key.split(".")) !== false;
+        });
+        return Object.keys(MODES).find(function (mode) {
+          return MODES[mode].every(function (flag, i) { return flag === flags[i]; });
+        }) || "custom";
+      }
+      function setMode(mode) {
+        if (!MODES[mode] || busy || !snapshot.writable) return;
+        setDraft(function (prev) {
+          var next = Object.assign({}, prev);
+          MODE_KEYS.forEach(function (key, i) { next[key] = MODES[mode][i]; });
+          return next;
+        });
+        setNotice(null); setError(null);
+      }
 
       function onSave() {
         if (busy || !snapshot.writable) return;
@@ -233,6 +299,7 @@ window.__ModuleLoader__.load({
         var ops = [];
         for (var i = 0; i < FIELDS.length; i += 1) {
           var f = FIELDS[i];
+          if (!Object.prototype.hasOwnProperty.call(draft, f.key)) continue;
           var d = fieldDraft(f);
           var current = getPath(value, f.path);
           if (f.type === "password") {
@@ -250,7 +317,7 @@ window.__ModuleLoader__.load({
           if (String(d).trim() === "" && getPath(user, f.path) === void 0) continue;
           ops.push(String(d).trim() === "" ? { op: "unset", path: f.path } : { op: "set", path: f.path, value: f.type === "number" ? Number(d) : d });
         }
-        if (ops.length === 0) { setBusy(false); setNotice(t("saved")); return; }
+        if (ops.length === 0) { setBusy(false); setNotice({ key: "saved" }); return; }
         commit(ops);
       }
 
@@ -269,36 +336,38 @@ window.__ModuleLoader__.load({
           var next = scope.getSnapshot();
           setSnapshot(next);
           setBusy(false);
-          if (!ok) { setError(t("error") + ": " + t("notApplied")); return; }
-          setNotice(t("saved"));
-          if (next.value) setDraft(valueToDraft(next.value));
+          if (!ok) { setError({ key: "error", detailKey: "notApplied" }); return; }
+          setNotice({ key: "saved" });
+          if (next.value) setDraft({});
         }).catch(function (e) {
-          setBusy(false); setError(t("error") + ": " + String(e && e.message || e));
+          setBusy(false); setError({ key: "error", detail: String(e && e.message || e) });
         });
       }
 
-      var nodes = [];
-      var lastGroup = null;
+      var nodes = [], advancedNodes = [];
+      var lastGroups = { basic: null, advanced: null };
       // forEach callback gives each handler its own `f` — a `for (var i)`
       // loop would share one `f` across every onChange closure, so typing
       // updated the LAST field's draft and the input appeared dead.
       FIELDS.forEach(function (f) {
-        if (f.group !== lastGroup) {
-          lastGroup = f.group;
-          nodes.push(h("div", { key: "g" + f.group, className: "__tm_group" }, t(f.group)));
+        var section = BASIC_KEYS.indexOf(f.key) !== -1 ? "basic" : "advanced";
+        var target = section === "basic" ? nodes : advancedNodes;
+        if (f.group !== lastGroups[section]) {
+          lastGroups[section] = f.group;
+          target.push(h("div", { key: "g" + f.group, className: "__tm_group" }, t(f.group)));
         }
         var overridden = getPath(user, f.path) !== void 0;
         if (f.type === "checkbox") {
-          nodes.push(h("label", { key: f.path.join("."), className: "__tm_field" },
+          target.push(h("label", { key: f.path.join("."), className: "__tm_field" },
             h("span", { className: "__tm_row" },
-              h("input", { className: "__tm_check", type: "checkbox", checked: Boolean(fieldDraft(f)), onChange: function (e) { setField(f, e.target.checked); } }),
+              h("input", { className: "__tm_check", type: "checkbox", disabled: busy || !snapshot.writable, checked: Boolean(fieldDraft(f)), onChange: function (e) { setField(f, e.target.checked); } }),
               h("span", { className: "__tm_label" }, t(f.label)),
               overridden ? h("span", { className: "__tm_override" }, t("overridden")) : null
             )
           ));
           return;
         }
-        nodes.push(h("label", { key: f.path.join("."), className: "__tm_field" },
+        target.push(h("label", { key: f.path.join("."), className: "__tm_field" },
           h("span", { className: "__tm_label" },
             t(f.label),
             overridden ? h("span", { className: "__tm_override" }, t("overridden")) : null
@@ -307,6 +376,7 @@ window.__ModuleLoader__.load({
             className: "__tm_input",
             type: f.type === "password" ? "password" : f.type === "number" ? "number" : "text",
             value: fieldDraft(f),
+            disabled: busy || !snapshot.writable,
             placeholder: f.type === "password" ? (overridden ? "••••••••" : t("secretHint")) : "",
             onChange: function (e) { setField(f, e.target.value); }
           }),
@@ -314,29 +384,45 @@ window.__ModuleLoader__.load({
         ));
       });
 
-      return h("div", { className: "__tm_root" },
+      return h("div", { className: "__tm_root dsh-flat" },
         h("p", { className: "__tm_hint", style: { margin: "0 0 4px" } }, t("intro")),
+        h("label", { className: "__tm_field" },
+          h("span", { className: "__tm_label" }, t("mode")),
+          h("select", { className: "__tm_input", "aria-label": t("mode"), value: memoryMode(), disabled: busy || !snapshot.writable,
+            onChange: function (e) { setMode(e.target.value); } },
+            h("option", { value: "auto" }, t("modeAuto")), h("option", { value: "search" }, t("modeSearch")),
+            h("option", { value: "paused" }, t("modePaused")), h("option", { value: "custom", disabled: true }, t("modeCustom"))),
+          h("span", { className: "__tm_hint" }, t("modeHint"))),
+        h("p", { className: "__tm_hint" }, t("setupHint")),
         nodes,
+        h("details", null, h("summary", { style: { cursor: "pointer" } }, t("advanced")),
+          h("div", { style: { display: "flex", flexDirection: "column", gap: 10, paddingTop: 12 } }, advancedNodes)),
         h("div", { className: "__tm_actions" },
           h("button", { type: "button", className: "__tm_btn __tm_btnPrimary", onClick: onSave, disabled: busy || !snapshot.writable }, t("save")),
           h("button", { type: "button", className: "__tm_btn", onClick: onReset, disabled: busy || !snapshot.writable }, t("reset")),
-          notice ? h("span", { className: "__tm_status" }, notice) : null,
+          notice ? h("span", { className: "__tm_status" }, messageText(t, notice)) : null,
           busy ? h("span", { className: "__tm_status" }, t("saving")) : null,
-          error ? h("span", { className: "__tm_error" }, error) : null
+          error ? h("span", { className: "__tm_error" }, messageText(t, error)) : null
         )
       );
     }
 
-    function valueToDraft(value) {
-      var out = {};
-      for (var i = 0; i < FIELDS.length; i += 1) {
-        var f = FIELDS[i];
-        out[f.key] = f.secret ? "" : f.type === "checkbox" ? Boolean(getPath(value, f.path)) : String(getPath(value, f.path) ?? "");
-      }
-      return out;
-    }
-
     // ── plugin ────────────────────────────────────────────────────────────
+
+    // Follow the host language without remounting the form or losing drafts.
+    function useLocale(locale) {
+      var refresh = react.useState(0)[1];
+      react.useEffect(function () {
+        if (!locale || typeof locale.subscribe !== "function") return;
+        return locale.subscribe(function () { refresh(function (revision) { return revision + 1; }); });
+      }, [locale]);
+    }
+    // Keep translation keys in state so feedback follows later language changes.
+    function messageText(t, message) {
+      if (!message) return "";
+      return t(message.key) + (message.detailKey ? ": " + t(message.detailKey) : message.detail ? ": " + message.detail : "");
+    }
+
     function apply(ctx) {
       var t = ctx.locale.bind(NS);
       ctx.effect(function () { return ctx.locale.register(NS, { zh: zh, en: en }); }, "dsh-tdai-memory: dictionaries");
@@ -347,7 +433,7 @@ window.__ModuleLoader__.load({
           key: "dsh-tdai-memory",
           locale: NS
         }, function (props) {
-          return h(MemorySection, Object.assign({}, props, { scope: scope, t: t }));
+          return h(MemorySection, Object.assign({}, props, { scope: scope, t: t, locale: ctx.locale }));
         });
       });
     }

package/index.js

@@ -24,6 +24,7 @@ import { defineTool } from "@deepseek-ai/dsh-tools";
 import { TdaiCore } from "./vendor/tdai/core/tdai-core.js";
 import { parseConfig } from "./vendor/tdai/config.js";
 import { StandaloneHostAdapter } from "./vendor/tdai/adapters/standalone/host-adapter.js";
+import { createRecallCache, recallWithTimeout } from "./recall-runtime.js";
 
 /** Cordis plugin name. */
 const name = "tdai-memory";
@@ -145,8 +146,7 @@ function apply(ctx, config) {
   /** session keys seen this process (for final L1 flush on teardown). */
   const seenSessions = new Set();
   /** recall cache per session: same user text within TTL reuses the result. */
-  const recallCache = new Map();
-  const RECALL_CACHE_TTL_MS = 30_000;
+  const recallCache = createRecallCache();
 
   // One-shot mode (dsh --profile headless): the process exits right after
   // flush, so L1 must finish inside the flush listener. In web mode the
@@ -231,6 +231,7 @@ function apply(ctx, config) {
     // new core; initialize() only sets up stores/pipeline.
     tdaiConfig = nextConfig;
     core = nextCore;
+    recallCache.clear();
     await nextCore.initialize();
     const key = nextConfig.llm.apiKey || "";
     log("info", `initialized (dataDir=${dataDir}, llm=${nextConfig.llm.model}, apiKey=${key ? `${key.slice(0, 6)}…` : "<empty>"}, embedding=${nextConfig.embedding.model})`);
@@ -321,16 +322,15 @@ function apply(ctx, config) {
             if (!lastUser) return assembly;
             const text = extractText(lastUser.content);
             if (!text) return assembly;
-            const now = Date.now();
-            const cached = recallCache.get(session.id);
-            if (cached && cached.text === text && now - cached.ts < RECALL_CACHE_TTL_MS) {
-              return injectRecall(assembly, cached.result);
+            const cached = recallCache.get(session.id, text);
+            if (cached !== undefined) {
+              return injectRecall(assembly, cached);
             }
-            const result = await Promise.race([
-              core.handleBeforeRecall(text, session.id),
-              new Promise((resolve) => setTimeout(() => resolve(null), tdaiConfig.recall.timeoutMs + 1000)),
-            ]).catch(() => null);
-            recallCache.set(session.id, { text, result: result ?? {}, ts: now });
+            const recallCore = core;
+            const result = await recallWithTimeout(
+              () => recallCore.handleBeforeRecall(text, session.id), tdaiConfig.recall.timeoutMs + 1000,
+            );
+            if (core === recallCore) recallCache.set(session.id, text, result);
             return injectRecall(assembly, result ?? {});
           } catch (error) {
             ctx.logger.warn(`[tdai-memory] recall failed: ${String(error)}`);

package/locale/en.json

@@ -0,0 +1,6 @@
+{
+  "meta": {
+    "title": "Long-term memory",
+    "description": "Organize long-term memories from conversations and recall relevant context, with search-only and paused modes."
+  }
+}

package/locale/zh.json

@@ -0,0 +1,6 @@
+{
+  "meta": {
+    "title": "长期记忆",
+    "description": "自动整理对话中的长期记忆，并在相关对话中召回；支持仅搜索和暂停模式。"
+  }
+}

package/package.json

@@ -1,97 +1,102 @@
-{
-    "name": "dsh-tdai-memory",
-    "version": "0.4.1",
-    "description": "DeepSeek Harness 的 TencentDB Agent Memory 移植：L0 对话捕获 → L1 结构化记忆提取 → L2 场景/L3 画像，自动召回注入 + 记忆/对话搜索工具；复用现有 ~/.memory-tencentdb/memory-tdai 数据；附 Web UI 设置栏。",
-    "type": "module",
-    "main": "index.js",
-    "exports": {
-        ".": "./index.js",
-        "./client": "./client.js",
-        "./package.json": "./package.json"
-    },
-    "dsh": {
-        "bundle": {
-            "patch": "./cordis.patch.yml"
-        },
-        "client": {
-            "platform": "web"
-        },
-        "compatibility": {
-            "dsh": ">=0.2.0-rc.2 <0.3.0",
-            "dshReleases": {
-                "0.1.5-rc.2": "incompatible",
-                "0.1.5-rc.3": "incompatible",
-                "0.1.6-alpha.1": "unknown",
-                "0.1.6-alpha.2": "unknown",
-                "0.1.7-alpha.1": "unknown",
-                "0.1.7-alpha.2": "unknown",
-                "0.1.7-rc.1": "incompatible",
-                "0.1.7-rc.2": "incompatible",
-                "0.2.0-rc.2": "compatible"
-            },
-            "node": ">=22.18",
-            "profiles": [
-                "web",
-                "desktop"
-            ]
-        }
-    },
-    "homepage": "https://github.com/Scorp1o117/dsh-tdai-memory",
-    "files": [
-        "index.js",
-        "recall-inject.js",
-        "client.js",
-        "cordis.patch.yml",
-        "vendor",
-        "README.md",
-        "README.zh.md",
-        "LICENSE",
-        "THIRD-PARTY-NOTICES.md"
-    ],
-    "keywords": [
-        "dsh",
-        "dsh-plugin",
-        "deepseek-harness",
-        "memory",
-        "tencentdb",
-        "agent-memory",
-        "rag",
-        "vector-search",
-        "persona"
-    ],
-    "license": "MIT",
-    "engines": {
-        "node": ">=22.18"
-    },
-    "repository": {
-        "type": "git",
-        "url": "git+https://github.com/Scorp1o117/dsh-tdai-memory.git"
-    },
-    "bugs": {
-        "url": "https://github.com/Scorp1o117/dsh-tdai-memory/issues"
-    },
-    "dependencies": {
-        "@ai-sdk/openai": "^3.0.53",
-        "@deepseek-ai/schemastery": "^3.18.4",
-        "@node-rs/jieba": "^2.0.1",
-        "@tencentdb-agent-memory/tcvdb-text": "^0.1.1",
-        "ai": "^6.0.164",
-        "json5": "^2.2.3",
-        "sqlite-vec": "0.1.7-alpha.2",
-        "undici": "^8.1.0"
-    },
-    "peerDependencies": {
-        "react": "^18.2.0",
-        "@deepseek-ai/dsh-tools": ">=0.2.0-rc.2 <0.3.0",
-        "@deepseek-ai/dsh-settings": ">=0.2.0-rc.2 <0.3.0",
-        "node-llama-cpp": "^3.19.1"
-    },
-    "peerDependenciesMeta": {
-        "node-llama-cpp": {
-            "optional": true
-        }
-    },
-    "scripts": {
-        "test": "node tests/settings-scope-compat.test.mjs && node tests/install-contract.test.mjs && node --test tests/plugin-page.test.mjs tests/client-write-flow.test.mjs tests/tools-registration.test.mjs"
-    }
-}
+{
+    "name": "dsh-tdai-memory",
+    "version": "0.4.3",
+    "description": "DeepSeek Harness 的 TencentDB Agent Memory 移植：L0 对话捕获 → L1 结构化记忆提取 → L2 场景/L3 画像，自动召回注入 + 记忆/对话搜索工具；复用现有 ~/.memory-tencentdb/memory-tdai 数据；附 Web UI 设置栏。",
+    "type": "module",
+    "main": "index.js",
+    "exports": {
+        ".": "./index.js",
+        "./client": "./client.js",
+        "./package.json": "./package.json",
+        "./locale/*.json": "./locale/*.json"
+    },
+    "dsh": {
+        "bundle": {
+            "patch": "./cordis.patch.yml"
+        },
+        "client": {
+            "platform": "web"
+        },
+        "compatibility": {
+            "dsh": ">=0.2.0-rc.2 <0.3.0 || 0.2.1-rc.1",
+            "dshReleases": {
+                "0.1.5-rc.2": "incompatible",
+                "0.1.5-rc.3": "incompatible",
+                "0.1.6-alpha.1": "unknown",
+                "0.1.6-alpha.2": "unknown",
+                "0.1.7-alpha.1": "unknown",
+                "0.1.7-alpha.2": "unknown",
+                "0.1.7-rc.1": "incompatible",
+                "0.1.7-rc.2": "incompatible",
+                "0.2.0-rc.2": "compatible",
+                "0.2.1-rc.1": "compatible"
+            },
+            "node": ">=22.18",
+            "profiles": [
+                "web",
+                "desktop"
+            ]
+        }
+    },
+    "homepage": "https://github.com/Scorp1o117/dsh-tdai-memory",
+    "files": [
+        "locale",
+        "index.js",
+        "recall-inject.js",
+        "recall-runtime.js",
+        "client.js",
+        "cordis.patch.yml",
+        "vendor",
+        "README.md",
+        "README.zh.md",
+        "LICENSE",
+        "THIRD-PARTY-NOTICES.md"
+    ],
+    "keywords": [
+        "dsh",
+        "dsh-plugin",
+        "deepseek-harness",
+        "memory",
+        "tencentdb",
+        "agent-memory",
+        "rag",
+        "vector-search",
+        "persona"
+    ],
+    "license": "MIT",
+    "engines": {
+        "node": ">=22.18",
+        "dsh": ">=0.2.0-rc.2 <0.3.0 || 0.2.1-rc.1"
+    },
+    "repository": {
+        "type": "git",
+        "url": "git+https://github.com/Scorp1o117/dsh-tdai-memory.git"
+    },
+    "bugs": {
+        "url": "https://github.com/Scorp1o117/dsh-tdai-memory/issues"
+    },
+    "dependencies": {
+        "@ai-sdk/openai": "^3.0.53",
+        "@deepseek-ai/schemastery": "^3.18.4",
+        "@node-rs/jieba": "^2.0.1",
+        "@tencentdb-agent-memory/tcvdb-text": "^0.1.1",
+        "ai": "^6.0.164",
+        "json5": "^2.2.3",
+        "sqlite-vec": "0.1.7-alpha.2",
+        "undici": "^8.1.0"
+    },
+    "peerDependencies": {
+        "react": "^18.2.0",
+        "@deepseek-ai/dsh-tools": ">=0.2.0-rc.2 <0.3.0 || 0.2.1-rc.1",
+        "@deepseek-ai/dsh-settings": ">=0.2.0-rc.2 <0.3.0 || 0.2.1-rc.1",
+        "node-llama-cpp": "^3.19.1"
+    },
+    "peerDependenciesMeta": {
+        "node-llama-cpp": {
+            "optional": true
+        }
+    },
+    "scripts": {
+        "test": "node tests/settings-scope-compat.test.mjs && node tests/install-contract.test.mjs && node --test tests/plugin-page.test.mjs tests/client-write-flow.test.mjs tests/tools-registration.test.mjs tests/recall-runtime.test.mjs && node --test tests/locale.test.mjs"
+    }
+}

package/recall-inject.js

@@ -1,86 +1,88 @@
-/**
- * tdai-recall-inject — agent-plane recall injection row for dsh-tdai-memory.
- *
- * Mounted INSIDE an agent preset (agent.cordis.yml), this row listens on the
- * `system-prompt/assemble` waterfall in the preset's scope, so it runs for
- * every assembly of every agent joined to that preset (agent -> preset scope
- * chain). It asks the `tdaiMemory` service (provided by the root-level
- * dsh-tdai-memory plugin) for relevant long-term memories and appends them
- * as dynamic context sections (user-role snapshots):
- *
- *   - tdai:recall  — relevant L1 memories for the current user message
- *   - tdai:profile — stable persona / scene navigation context
- *
- * The recall call is bounded by `recall.timeoutMs`; failures degrade to no
- * injection and never break the assembly.
- */
-import z from "@deepseek-ai/schemastery";
-
-/** Cordis plugin name. */
-const name = "tdai-recall-inject";
-/** Services consumed: the prompt registry and the memory core service. */
-const inject = ["systemPrompt", "tdaiMemory"];
-
-/** Runtime schema for the recall-inject row. */
-const Config = z.object({
-  /** Skip recall when the user message is shorter than this. */
-  minUserTextChars: z.number().default(1),
-  /** Per-session cache TTL for identical user text (ms). */
-  cacheTtlMs: z.number().default(30_000),
-});
-
-function apply(ctx, config) {
-  const cache = new Map();
-
-  ctx.on("system-prompt/assemble", async (assembly, context) => {
-    try {
-      const agent = context?.agent;
-      if (!agent) return assembly;
-      const session = agent.session;
-      if (!session) return assembly;
-      // Only real human input triggers recall. DSH injects synthetic
-      // user-role context (system reminders, runtime context, skill
-      // catalogs, …) whose source.kind is not "user"; recalling on those
-      // would pollute the prompt with irrelevant memories (issue #1).
-      const lastUser = [...session.deriveMessages()]
-        .reverse()
-        .find((m) => m.role === "user" && (m.source == null || m.source.kind === "user"));
-      if (!lastUser) return assembly;
-      const text = extractText(lastUser.content);
-      if (!text || text.length < config.minUserTextChars) return assembly;
-
-      const now = Date.now();
-      const cached = cache.get(session.id);
-      if (cached && cached.text === text && now - cached.ts < config.cacheTtlMs) {
-        return injectRecall(assembly, cached.result);
-      }
-      const result = await ctx.tdaiMemory
-        .handleBeforeRecall(text, session.id)
-        .catch(() => null);
-      cache.set(session.id, { text, result: result ?? {}, ts: now });
-      return injectRecall(assembly, result ?? {});
-    } catch {
-      return assembly;
-    }
-  });
-}
-
-function extractText(content) {
-  if (typeof content === "string") return content.trim();
-  if (!Array.isArray(content)) return "";
-  return content
-    .filter((b) => b && b.type === "text" && typeof b.text === "string")
-    .map((b) => b.text)
-    .join("\n")
-    .trim();
-}
-
-function injectRecall(assembly, result) {
-  const extras = [];
-  if (result.prependContext) extras.push({ name: "tdai:recall", order: 1000, text: result.prependContext });
-  if (result.appendSystemContext) extras.push({ name: "tdai:profile", order: 1001, text: result.appendSystemContext });
-  if (extras.length === 0) return assembly;
-  return { ...assembly, contexts: [...assembly.contexts, ...extras] };
-}
-
-export { Config, apply, inject, name };
+/**
+ * tdai-recall-inject — agent-plane recall injection row for dsh-tdai-memory.
+ *
+ * Mounted INSIDE an agent preset (agent.cordis.yml), this row listens on the
+ * `system-prompt/assemble` waterfall in the preset's scope, so it runs for
+ * every assembly of every agent joined to that preset (agent -> preset scope
+ * chain). It asks the `tdaiMemory` service (provided by the root-level
+ * dsh-tdai-memory plugin) for relevant long-term memories and appends them
+ * as dynamic context sections (user-role snapshots):
+ *
+ *   - tdai:recall  — relevant L1 memories for the current user message
+ *   - tdai:profile — stable persona / scene navigation context
+ *
+ * The recall call is bounded by this row's `timeoutMs`; failures degrade to no
+ * injection and never break the assembly.
+ */
+import z from "@deepseek-ai/schemastery";
+import { createRecallCache, recallWithTimeout } from "./recall-runtime.js";
+
+/** Cordis plugin name. */
+const name = "tdai-recall-inject";
+/** Services consumed: the prompt registry and the memory core service. */
+const inject = ["systemPrompt", "tdaiMemory"];
+
+/** Runtime schema for the recall-inject row. */
+const Config = z.object({
+  /** Skip recall when the user message is shorter than this. */
+  minUserTextChars: z.number().default(1),
+  /** Per-session cache TTL for identical user text (ms). */
+  cacheTtlMs: z.number().default(30_000),
+  /** Assembly deadline even when the memory service does not settle. */
+  timeoutMs: z.number().default(4000),
+});
+
+function apply(ctx, config) {
+  const cache = createRecallCache(config.cacheTtlMs ?? 30_000);
+
+  ctx.on("system-prompt/assemble", async (assembly, context) => {
+    try {
+      const agent = context?.agent;
+      if (!agent) return assembly;
+      const session = agent.session;
+      if (!session) return assembly;
+      // Only real human input triggers recall. DSH injects synthetic
+      // user-role context (system reminders, runtime context, skill
+      // catalogs, …) whose source.kind is not "user"; recalling on those
+      // would pollute the prompt with irrelevant memories (issue #1).
+      const lastUser = [...session.deriveMessages()]
+        .reverse()
+        .find((m) => m.role === "user" && (m.source == null || m.source.kind === "user"));
+      if (!lastUser) return assembly;
+      const text = extractText(lastUser.content);
+      if (!text || text.length < config.minUserTextChars) return assembly;
+
+      const cached = cache.get(session.id, text);
+      if (cached !== undefined) {
+        return injectRecall(assembly, cached);
+      }
+      const result = await recallWithTimeout(
+        () => ctx.tdaiMemory.handleBeforeRecall(text, session.id), config.timeoutMs,
+      );
+      cache.set(session.id, text, result);
+      return injectRecall(assembly, result ?? {});
+    } catch {
+      return assembly;
+    }
+  });
+}
+
+function extractText(content) {
+  if (typeof content === "string") return content.trim();
+  if (!Array.isArray(content)) return "";
+  return content
+    .filter((b) => b && b.type === "text" && typeof b.text === "string")
+    .map((b) => b.text)
+    .join("\n")
+    .trim();
+}
+
+function injectRecall(assembly, result) {
+  const extras = [];
+  if (result.prependContext) extras.push({ name: "tdai:recall", order: 1000, text: result.prependContext });
+  if (result.appendSystemContext) extras.push({ name: "tdai:profile", order: 1001, text: result.appendSystemContext });
+  if (extras.length === 0) return assembly;
+  return { ...assembly, contexts: [...assembly.contexts, ...extras] };
+}
+
+export { Config, apply, inject, name };

package/recall-runtime.js

@@ -0,0 +1,42 @@
+/** Bound recall latency without leaving a timer behind after a fast result.
+ * The underlying operation may continue; late failures remain handled. */
+export async function recallWithTimeout(operation, timeoutMs = 4000) {
+  const delay = Number.isFinite(timeoutMs) && timeoutMs > 0 ? timeoutMs : 4000;
+  let timer;
+  try {
+    return await Promise.race([
+      Promise.resolve().then(operation),
+      new Promise(resolve => { timer = setTimeout(() => resolve(null), delay); }),
+    ]);
+  } catch {
+    return null;
+  } finally {
+    clearTimeout(timer);
+  }
+}
+
+/** Expiring, capacity-limited per-session cache for long-running hosts. */
+export function createRecallCache(ttlMs = 30_000, capacity = 128, now = Date.now) {
+  const entries = new Map();
+  function prune() {
+    const time = now();
+    for (const [key, entry] of entries) {
+      if (time - entry.ts >= ttlMs) entries.delete(key);
+    }
+  }
+  return {
+    get(sessionId, text) {
+      prune();
+      const entry = entries.get(sessionId);
+      return entry?.text === text ? entry.result : undefined;
+    },
+    set(sessionId, text, result) {
+      prune();
+      if (ttlMs <= 0 || capacity <= 0 || result == null) return;
+      entries.delete(sessionId);
+      entries.set(sessionId, { text, result, ts: now() });
+      while (entries.size > capacity) entries.delete(entries.keys().next().value);
+    },
+    clear() { entries.clear(); },
+  };
+}