dsh-prime-memory 0.11.0

package/dist/util/recall-budget.js

@@ -0,0 +1,84 @@
+/**
+ * 召回预算与超时(ADR-0001 / 规格 A 节)。
+ *
+ * 预算:单条记忆截断上限 + 整轮总量上限——超限截断并以后缀引导模型用记忆工具
+ * 查全文(截断是引流而不是损失:工具路径返回完整记录);总量超限按融合排名丢尾部。
+ * 超时:召回是增强能力,超时跳过本轮注入、绝不阻塞对话(CONTEXT.md「召回超时」语义
+ * 在本模块以 raceRecallTimeout 落地)。
+ */
+/** 截断后缀:显式告诉模型全文在工具侧(引导主动深挖)。注入字符串,逐字为契约。 */
+export const RECALL_TRUNCATION_SUFFIX = '…(已截断;可用 memory_search 或 conversation_search 查看详情)';
+/** 剩余预算小于该值时整条丢弃(截出比后缀还短的行没有意义)。 */
+const MIN_TRUNCATED_RECALL_LINE_CHARS = 40;
+function normalizeLimit(value) {
+    if (value == null || !Number.isFinite(value) || value <= 0)
+        return undefined;
+    return Math.floor(value);
+}
+/** 按 code point 计数截断(不劈开代理对),带引导后缀。 */
+export function truncateRecallLine(line, maxChars) {
+    const cps = Array.from(line);
+    if (cps.length <= maxChars)
+        return line;
+    if (maxChars <= RECALL_TRUNCATION_SUFFIX.length) {
+        return cps.slice(0, maxChars).join('');
+    }
+    return `${cps.slice(0, maxChars - RECALL_TRUNCATION_SUFFIX.length).join('').trimEnd()}${RECALL_TRUNCATION_SUFFIX}`;
+}
+/**
+ * 对召回行施加预算:先逐条截断,再按总量预算装填——装不下的尾部整条丢弃。
+ * 输入行应按相关性降序(低分先丢)。行间换行符计入总量预算。
+ */
+export function applyRecallBudget(lines, limits) {
+    const maxCharsPerMemory = normalizeLimit(limits.maxCharsPerMemory);
+    const maxTotalRecallChars = normalizeLimit(limits.maxTotalRecallChars);
+    if (!maxCharsPerMemory && !maxTotalRecallChars)
+        return lines;
+    const budgeted = [];
+    let usedChars = 0;
+    for (let i = 0; i < lines.length; i++) {
+        const line = lines[i];
+        const perBounded = maxCharsPerMemory ? truncateRecallLine(line, maxCharsPerMemory) : line;
+        if (!maxTotalRecallChars) {
+            budgeted.push(perBounded);
+            continue;
+        }
+        const separatorChars = budgeted.length > 0 ? 1 : 0;
+        const remaining = maxTotalRecallChars - usedChars - separatorChars;
+        if (remaining <= 0)
+            break;
+        if (perBounded.length > remaining) {
+            const canFit = remaining >= MIN_TRUNCATED_RECALL_LINE_CHARS;
+            if (canFit)
+                budgeted.push(truncateRecallLine(perBounded, remaining));
+            break;
+        }
+        budgeted.push(perBounded);
+        usedChars += separatorChars + perBounded.length;
+    }
+    return budgeted;
+}
+/**
+ * 召回总预算:超时返回 undefined(调用方跳过本轮注入),正常 resolve 返回原值。
+ * resolve 为空结果(空数组)与超时(undefined)语义不同,调用方据此区分日志。
+ */
+export async function raceRecallTimeout(promise, timeoutMs) {
+    if (!Number.isFinite(timeoutMs) || timeoutMs <= 0)
+        return promise;
+    let timer;
+    try {
+        return await Promise.race([
+            promise,
+            new Promise((resolve) => {
+                timer = setTimeout(() => resolve(undefined), timeoutMs);
+            }),
+        ]);
+    }
+    finally {
+        if (timer)
+            clearTimeout(timer);
+    }
+}
+/** 召回路径嵌入调用的内层钳制(固定值):给 FTS 降级留出总预算内的时间。
+ *  远程作用于 HTTP fetch;本地作用于 worker 代理的等待(race 放弃、迟到回复丢弃)。 */
+export const RECALL_EMBED_CAP_MS = 3_000;

package/dist/util/sanitize.d.ts

@@ -0,0 +1,11 @@
+/**
+ * L0 捕获文本清洗。
+ * 保证:清除我们注入的召回标签(防反馈循环)、框架元数据块、媒体标记等。
+ * 正则清单即行为契约——漏一条,注入物就会经 L0→L1 反复自我强化。
+ */
+/** 剥离注入的记忆标签 + 框架元数据块 + 媒体标记。 */
+export declare function sanitizeText(text: string): string;
+/** 剥离助手回复中的围栏代码块(保留解释性文本)。 */
+export declare function stripCodeBlocks(text: string): string;
+/** L0 捕获过滤——宽松:只丢弃结构性无用消息。 */
+export declare function shouldCaptureL0(text: string): boolean;

package/dist/util/sanitize.js

@@ -0,0 +1,67 @@
+/**
+ * L0 捕获文本清洗。
+ * 保证:清除我们注入的召回标签(防反馈循环)、框架元数据块、媒体标记等。
+ * 正则清单即行为契约——漏一条,注入物就会经 L0→L1 反复自我强化。
+ */
+/** 剥离注入的记忆标签 + 框架元数据块 + 媒体标记。 */
+export function sanitizeText(text) {
+    let cleaned = text;
+    // 注入的记忆上下文标签(防止再捕获时反馈循环)
+    cleaned = cleaned.replace(/<relevant-memories>[\s\S]*?<\/relevant-memories>/g, '');
+    cleaned = cleaned.replace(/<user-persona>[\s\S]*?<\/user-persona>/g, '');
+    cleaned = cleaned.replace(/<relevant-scenes>[\s\S]*?<\/relevant-scenes>/g, '');
+    cleaned = cleaned.replace(/<scene-navigation>[\s\S]*?<\/scene-navigation>/g, '');
+    cleaned = cleaned.replace(/<memory-tools-guide>[\s\S]*?<\/memory-tools-guide>/g, '');
+    // 任务上下文注入块
+    cleaned = cleaned.replace(/<current_task_context>[\s\S]*?<\/current_task_context>/g, '');
+    cleaned = cleaned.replace(/<history_task_context[\s\S]*?<\/history_task_context>/g, '');
+    // 框架注入的入站元数据块(label + ```json ... ```)
+    cleaned = cleaned.replace(/(?:Conversation info|Sender|Thread starter|Replied message|Forwarded message context|Chat history since last reply)\s*\(untrusted[\s\S]*?\):\s*```json\s*[\s\S]*?```/g, '');
+    // 旧版会话元数据 JSON 块
+    cleaned = cleaned.replace(/```json\s*\{[\s\S]*?"session[\s\S]*?\}\s*```/g, '');
+    // 回复指令标签
+    cleaned = cleaned.replace(/\[\[reply_to[^\]]*\]\]\s*/g, '');
+    // Skill 选择包裹符
+    cleaned = cleaned.replace(/¥¥\[[\s\S]*?\]¥¥/g, '');
+    // 行首时间戳
+    cleaned = cleaned.replace(/^\[[\w\d\-:+ ]+\]\s*/gm, '');
+    // 媒体附件标记
+    cleaned = cleaned.replace(/\[media attached:[^\]]*\]\s*/g, '');
+    // 图片回复指令
+    cleaned = cleaned.replace(/To send an image back,[\s\S]*?(?:Keep caption in the text body\.)\s*/g, '');
+    // 系统执行块
+    cleaned = cleaned.replace(/^System:\s*\[[\s\S]*?$/gm, '');
+    // 内联 base64 图片
+    cleaned = cleaned.replace(/data:image\/[a-z+]+;base64,[A-Za-z0-9+/=]+/gi, '');
+    // 空字符 + 折叠空白
+    cleaned = cleaned.replace(/\0/g, '').replace(/\n{3,}/g, '\n\n').trim();
+    return cleaned;
+}
+/** 剥离助手回复中的围栏代码块(保留解释性文本)。 */
+export function stripCodeBlocks(text) {
+    return text.replace(/```[^\n]*\n[\s\S]*?```/g, '').replace(/\n{3,}/g, '\n\n').trim();
+}
+/** L0 捕获过滤——宽松:只丢弃结构性无用消息。 */
+export function shouldCaptureL0(text) {
+    if (!text || !text.trim())
+        return false;
+    if (isFrameworkNoise(text))
+        return false;
+    if (text.startsWith('/'))
+        return false;
+    return true;
+}
+function isFrameworkNoise(text) {
+    const t = text.trim();
+    if (t === '(session bootstrap)')
+        return true;
+    if (t.startsWith('A new session was started via'))
+        return true;
+    if (/^✅\s*New session started/.test(t))
+        return true;
+    if (t.startsWith('Pre-compaction memory flush'))
+        return true;
+    if (/^NO_REPLY\s*$/.test(t))
+        return true;
+    return false;
+}

package/dist/util/text.d.ts

@@ -0,0 +1,16 @@
+/**
+ * 中英混排分词:jieba 词元 ∪ 拉丁词 ∪ CJK 二元组,按首次出现顺序去重。
+ *
+ * 分词结果是 FTS 索引内容的定义——token 序列必须与既有索引构建算法严格一致,
+ * 否则旧库检索错位(这是格式级契约,不是可自由重设计项):
+ * - 词元给 BM25 提供高精度整词命中("负载均衡"作为词,idf 远高于碎片二元组);
+ * - 二元组保住子词召回底线:查询"负载"仍能命中只含"负载均衡"词元的行,且旧库
+ *   纯二元组索引无需迁移即可被新查询命中(新查询仍含二元组 token);
+ * - 去重防 2 字词与其自身二元组重复计数(FTS tf / 词频被同一出现双计);
+ * - jieba 加载失败时 jiebaCut 返回 undefined,自动退化为纯二元组。
+ */
+import type { ContentBlock } from '@deepseek-ai/dsh-llm';
+/** 把消息的 ContentBlock[] 展平成纯文本(仅 text 与 reasoning 块,换行连接)。 */
+export declare function blocksToText(blocks: readonly ContentBlock[] | undefined): string;
+/** FTS / BM25 共用分词入口(输入内部统一小写)。 */
+export declare function tokenize(text: string): string[];

package/dist/util/text.js

@@ -0,0 +1,61 @@
+import { jiebaCut } from './tokenizer.js';
+/** 把消息的 ContentBlock[] 展平成纯文本(仅 text 与 reasoning 块,换行连接)。 */
+export function blocksToText(blocks) {
+    if (!blocks)
+        return '';
+    const parts = [];
+    for (const b of blocks) {
+        if (b.type === 'text')
+            parts.push(b.text);
+        else if (b.type === 'reasoning')
+            parts.push(b.text);
+    }
+    return parts.join('\n');
+}
+const CJK_RE = /[\u3400-\u9fff\uf900-\ufaff]/;
+const WORD_RE = /[a-zA-Z0-9][a-zA-Z0-9_-]{1,}/g;
+/** token 里至少要有一个字母/数字/CJK 字(输入已小写;滤掉 jieba 切出的纯标点 token)。 */
+const TOKEN_KEEP_RE = /[a-z0-9\u3400-\u9fff\uf900-\ufaff]/;
+/** CJK 连续段二元组(首尾单字成 token;jieba 失败回退时的唯一分词,也是并集模式的子词召回底线)。 */
+function cjkBigrams(text) {
+    const tokens = [];
+    const cjk = text.replace(/[^\u3400-\u9fff\uf900-\ufaff]/g, ' ');
+    let i = 0;
+    while (i < cjk.length) {
+        const ch = cjk[i];
+        if (CJK_RE.test(ch)) {
+            const next = cjk[i + 1];
+            if (next && CJK_RE.test(next))
+                tokens.push(ch + next);
+            else
+                tokens.push(ch);
+        }
+        i += 1;
+    }
+    return tokens;
+}
+/** FTS / BM25 共用分词入口(输入内部统一小写)。 */
+export function tokenize(text) {
+    const lower = text.toLowerCase();
+    const seen = new Set();
+    const tokens = [];
+    const push = (t) => {
+        if (t.length >= 2 && TOKEN_KEEP_RE.test(t) && !seen.has(t)) {
+            seen.add(t);
+            tokens.push(t);
+        }
+        else if (t.length === 1 && CJK_RE.test(t) && !seen.has(t)) {
+            seen.add(t);
+            tokens.push(t);
+        }
+    };
+    const words = jiebaCut(lower);
+    if (words)
+        for (const w of words)
+            push(w.trim());
+    for (const m of lower.matchAll(WORD_RE))
+        push(m[0]);
+    for (const bg of cjkBigrams(lower))
+        push(bg);
+    return tokens;
+}

package/dist/util/tokenizer.d.ts

@@ -0,0 +1,9 @@
+export type TokenizerMode = 'jieba' | 'bigram';
+/** 启动期主动初始化并返回模式(MemoryDb 记日志用)。 */
+export declare function ensureTokenizer(): TokenizerMode;
+/** FTS 分词器版本戳(存 embedding_meta 表,键 fts_tokenizer)。 */
+export declare function tokenizerStamp(): string;
+/** 模式描述(日志用)。 */
+export declare function describeTokenizer(): string;
+/** jieba 切词(回退模式下返回 undefined,调用方走二元组路径)。 */
+export declare function jiebaCut(text: string): string[] | undefined;

package/dist/util/tokenizer.js

@@ -0,0 +1,50 @@
+/**
+ * 分词器装配:jieba 词级分词(@node-rs/jieba,Rust napi 预编译二进制)优先,
+ * 加载失败(平台无预编译二进制等)永久回退 CJK 二元组。
+ *
+ * - 惰性单例:首次调用即定死本进程的分词模式,运行中不漂移——FTS 读写两侧共用
+ *   同一实例,索引/查询天然对齐;
+ * - require 走 createRequire(与原生模块装配同款):失败只降级不抛出,由 MemoryDb
+ *   启动时主动 ensure 并记录模式日志;
+ * - 分词器版本戳 jieba-v1/bigram-v1 是 FTS 重建的触发键:戳 ≠ 索引构建时的戳,
+ *   FTS 表 drop 后从源表全量回灌——戳必须如实反映"构建当前 FTS 内容的分词器"。
+ */
+import { createRequire } from 'node:module';
+const require = createRequire(import.meta.url);
+let mode;
+let cutFn;
+/** 惰性初始化(永不抛出)。词典约 5MB,首次加载 ~100ms。 */
+function ensure() {
+    if (mode !== undefined)
+        return mode;
+    try {
+        const { Jieba } = require('@node-rs/jieba');
+        const { dict } = require('@node-rs/jieba/dict');
+        const jieba = Jieba.withDict(dict);
+        cutFn = (text) => jieba.cut(text, true);
+        mode = 'jieba';
+    }
+    catch {
+        cutFn = undefined;
+        mode = 'bigram';
+    }
+    return mode;
+}
+/** 启动期主动初始化并返回模式(MemoryDb 记日志用)。 */
+export function ensureTokenizer() {
+    return ensure();
+}
+/** FTS 分词器版本戳(存 embedding_meta 表,键 fts_tokenizer)。 */
+export function tokenizerStamp() {
+    return ensure() === 'jieba' ? 'jieba-v1' : 'bigram-v1';
+}
+/** 模式描述(日志用)。 */
+export function describeTokenizer() {
+    return ensure() === 'jieba'
+        ? 'jieba 词级分词(@node-rs/jieba)+ CJK 二元组并集'
+        : 'jieba 加载失败,回退 CJK 二元组分词(子词召回降级)';
+}
+/** jieba 切词(回退模式下返回 undefined,调用方走二元组路径)。 */
+export function jiebaCut(text) {
+    return ensure() === 'jieba' ? cutFn(text) : undefined;
+}

package/dsh.plugin.json

@@ -0,0 +1,22 @@
+{
+  "id": "dsh-prime-memory",
+  "name": "dsh-prime-memory",
+  "version": "0.11.0",
+  "description": {
+    "en": "Layered distilled memory plugin (L0 capture → L1 atomic memory → L2 scenes → L3 persona) with automatic recall injection before each model step.",
+    "zh": "L0~L3 分层蒸馏记忆插件:自动捕获对话、抽取原子记忆、整合场景块、蒸馏核心画像,并在模型步骤前自动召回注入。"
+  },
+  "homepage": "https://github.com/drscrewdriver/dsh-prime-memory",
+  "repository": {
+    "type": "git",
+    "url": "https://github.com/drscrewdriver/dsh-prime-memory.git"
+  },
+  "license": "MIT",
+  "engines": {
+    "dsh": ">=0.1.1-rc.2 <0.2.0-0"
+  },
+  "components": {
+    "host": "dist/index.js",
+    "client": "dist/client.js"
+  }
+}

package/package.json

@@ -0,0 +1,118 @@
+{
+  "name": "dsh-prime-memory",
+  "version": "0.11.0",
+  "description": "L0~L3 分层蒸馏记忆插件 for DeepSeek Harness：自动捕获对话（L0）、抽取原子记忆（L1）、整合场景块（L2）、蒸馏核心画像/团队方法论（L3），并在模型步骤前自动召回注入。移植自 MemoryCore (TencentDB Agent Memory) 的管线设计。",
+  "type": "module",
+  "main": "dist/index.js",
+  "types": "dist/index.d.ts",
+  "exports": {
+    ".": {
+      "types": "./dist/index.d.ts",
+      "default": "./dist/index.js"
+    },
+    "./client": "./dist/client.js",
+    "./package.json": "./package.json"
+  },
+  "dsh": {
+    "bundle": {
+      "patch": "./cordis.patch.yml"
+    },
+    "client": {
+      "inject": [
+        "@deepseek-ai/dsh-client-runtime",
+        "@deepseek-ai/dsh-client-connection"
+      ],
+      "platform": "web"
+    }
+  },
+  "files": [
+    "dist",
+    "assets",
+    "cordis.patch.yml",
+    "dsh.plugin.json",
+    "README.md",
+    "README.en.md",
+    "README.ja.md",
+    "README.ko.md",
+    "INSTALL.md",
+    "INSTALL.en.md",
+    "INSTALL.ja.md",
+    "INSTALL.ko.md",
+    "CHANGELOG.md",
+    "CHANGELOG.en.md",
+    "CHANGELOG.ja.md",
+    "CHANGELOG.ko.md",
+    "ENGINEERING-NOTES.md"
+  ],
+  "scripts": {
+    "build": "tsc -p tsconfig.json && node scripts/build-client.mjs && node scripts/copy-client.mjs",
+    "smoke": "npm run build:smoke && node dist-smoke/smoke.js",
+    "verify-catalog": "npm run build && node scripts/verify-catalog.mjs",
+    "typecheck": "tsc -p tsconfig.json --noEmit && tsc -p tsconfig.client.json --noEmit && tsc -p tsconfig.test.json",
+    "build:smoke": "tsc src/smoke.ts --outDir dist-smoke --module nodenext --moduleResolution nodenext --target es2022 --strict --skipLibCheck --esModuleInterop",
+    "lint": "eslint src client/src scripts tests",
+    "test": "vitest run"
+  },
+  "engines": {
+    "node": ">=22.16.0"
+  },
+  "keywords": [
+    "dsh",
+    "deepseek-harness",
+    "memory",
+    "agent-memory",
+    "cordis-plugin"
+  ],
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/drscrewdriver/dsh-prime-memory.git"
+  },
+  "author": "drscrewdriver <597093681@qq.com>",
+  "license": "MIT",
+  "dependencies": {
+    "@deepseek-ai/schemastery": "3.18.1",
+    "@node-rs/jieba": "^2.0.2",
+    "sqlite-vec": "^0.1.7-alpha.2",
+    "undici": "^7.29.0"
+  },
+  "peerDependencies": {
+    "@deepseek-ai/cordis": "^4.0.1",
+    "@deepseek-ai/dsh-agent": "^0.1.1-rc.2 || ^0.1.2-rc.1 || ^0.1.3-rc.1 || ^0.1.5-rc.2",
+    "@deepseek-ai/dsh-home-paths": "^0.1.1-rc.2 || ^0.1.2-rc.1 || ^0.1.3-rc.1 || ^0.1.5-rc.2",
+    "@deepseek-ai/dsh-llm": "^0.1.1-rc.2 || ^0.1.2-rc.1 || ^0.1.3-rc.1 || ^0.1.5-rc.2",
+    "@deepseek-ai/dsh-session": "^0.1.1-rc.2 || ^0.1.2-rc.1 || ^0.1.3-rc.1 || ^0.1.5-rc.2",
+    "@deepseek-ai/dsh-settings": "^0.1.1-rc.2 || ^0.1.2-rc.1 || ^0.1.3-rc.1 || ^0.1.5-rc.2",
+    "@deepseek-ai/dsh-system-prompt": "^0.1.1-rc.2 || ^0.1.2-rc.1 || ^0.1.3-rc.1 || ^0.1.5-rc.2",
+    "@deepseek-ai/dsh-tools": "^0.1.1-rc.2 || ^0.1.2-rc.1 || ^0.1.3-rc.1 || ^0.1.5-rc.2"
+  },
+  "peerDependenciesMeta": {
+    "@deepseek-ai/dsh-agent": { "optional": true },
+    "@deepseek-ai/dsh-home-paths": { "optional": true },
+    "@deepseek-ai/dsh-llm": { "optional": true },
+    "@deepseek-ai/dsh-session": { "optional": true },
+    "@deepseek-ai/dsh-settings": { "optional": true },
+    "@deepseek-ai/dsh-system-prompt": { "optional": true },
+    "@deepseek-ai/dsh-tools": { "optional": true }
+  },
+  "devDependencies": {
+    "@deepseek-ai/cordis": "4.0.1",
+    "@deepseek-ai/dsh-agent": "0.1.1-rc.2",
+    "@deepseek-ai/dsh-agent-default-model": "0.1.1-rc.2",
+    "@deepseek-ai/dsh-client-connection": "0.1.1-rc.2",
+    "@deepseek-ai/dsh-home-paths": "0.1.1-rc.2",
+    "@deepseek-ai/dsh-llm": "0.1.1-rc.2",
+    "@deepseek-ai/dsh-session": "0.1.1-rc.2",
+    "@deepseek-ai/dsh-settings": "0.1.1-rc.2",
+    "@deepseek-ai/dsh-system-prompt": "0.1.1-rc.2",
+    "@deepseek-ai/dsh-tools": "0.1.1-rc.2",
+    "@eslint/js": "^10.0.1",
+    "@types/node": "^22.0.0",
+    "@types/react": "^19.2.18",
+    "esbuild": "^0.28.2",
+    "eslint": "^10.9.1",
+    "globals": "^17.12.0",
+    "typescript": "^5.6.0",
+    "typescript-eslint": "^8.69.0",
+    "vitest": "^4.1.11"
+  }
+}