dsh-auto-memory 0.5.1 → 0.6.0

package/README.md

@@ -25,6 +25,26 @@ open a brand-new session tomorrow, ask *"what do you know about me?"*, and it
 
 ---
 
+## What's new in 0.6.0 (P2)
+
+- **Harness-sessions evaluation datasets** (`evals/datasets/`): synthetic
+  multi-session developer transcripts with gold-label facts — durable /
+  ephemeral-trap / updated labels, evidence sessions, abstention probes.
+  The coding-agent memory scenario no public benchmark covers.
+- **L1 injection recall** (zero-API runner): 100% on both datasets at the
+  perfect-consolidation ceiling — the injection side loses nothing.
+- **L3 consolidation quality** (one LLM call per session, same code path as
+  the plugin): **mini recall 100% / update-correctness 100% / pollution 0%;
+  standard recall 83.3% / precision 84.6% / update-correctness 100% /
+  pollution 0%** (deepseek-chat, temperature 0). Ephemeral traps were
+  correctly refused in every run.
+- **Product fix surfaced by the eval**: consolidation now writes
+  title/description in the transcript's language (Chinese transcripts used
+  to produce English memories — measured, then fixed).
+- Community: soft-mute (third state between keep and delete, #10), panel
+  current-workspace fix (#9), CJK bigram echo-dedup (#5), editor busy/error
+  UX (#8), GB11643 checksum before id-card redaction (#7).
+
 ## What's new in 0.5.0 (P2)
 
 - **Three-factor ranking** (Generative Agents formula, embedding-free):
@@ -220,7 +240,6 @@ Override in your profile's `cordis.patch.yml` (config replaces wholesale):
   resume); everything flows through standard `tool/call` / `tool/result`.
 
 Deep dives: [design decisions](docs/design.md) ·
-[dsh source-level research](docs/api-reports.md) ·
 [postmortem: shipping a PR to awesome-dsh-plugin](docs/postmortem-pr-5696.md)
 
 ## Roadmap

package/README.zh.md

@@ -24,6 +24,23 @@ dsh plugin --profile demo add dsh-auto-memory
 
 ---
 
+## 0.6.0 新增(P2)
+
+- **harness-sessions 评测数据集**(`evals/datasets/`):合成多会话开发者剧本
+  + 金标事实(durable / ephemeral 陷阱 / updated 推翻标签、evidence 会话、
+  拒答题)——公开基准都不覆盖的 coding-agent 记忆场景。
+- **L1 注入召回**(零 API runner):两个数据集在完美固化上界下均 100%——
+  注入侧零损耗。
+- **L3 固化质量**(每会话一次 LLM 调用,与插件同一条固化代码路径):
+  **mini 召回 100% / 更新正确率 100% / 污染 0%;standard 召回 83.3% /
+  精确率 84.6% / 更新正确率 100% / 污染 0%**(deepseek-chat,temperature 0)。
+  ephemeral 陷阱在每轮运行中都被正确拒绝。
+- **评测反哺的产品修复**:固化描述现在跟随会话语言(中文会话曾固化出
+  英文记忆——先测出,后修复)。
+- 社区贡献:软停用(保留与删除之间的第三档,#10)、面板当前工作区修复
+  (#9)、中文 bigram 回声去重(#5)、编辑器 busy/错误 UX(#8)、身份证
+  校验位前置验证(#7)。
+
 ## 0.5.0 新增(P2)
 
 - **三因子排序**(Generative Agents 公式的免 embedding 版):注入索引与
@@ -149,7 +166,7 @@ node scripts/demo.mjs   # 不要 API key、不开浏览器:看 写入 → 索引
 评测层已抓到过真实 bug:字节预算曾遗漏指导文本、整段超支 ~800 字节
 (已修复并带回归)。
 
-**78 项测试(含确定性评测层)。运行时依赖仅 `yaml`。安装体积 15 kB。**
+**93 项测试 + 三层评测(L0 确定性/L1 召回/L3 固化质量)。运行时依赖仅 `yaml`。**
 
 ## 安装
 
@@ -187,7 +204,6 @@ dsh --profile demo                               # 重启 profile 生效
   一切走标准 `tool/call` / `tool/result`。
 
 深度内容:[设计决策](docs/design.md) ·
-[dsh 源码级调研](docs/api-reports.md) ·
 [复盘:向 awesome-dsh-plugin 提 PR](docs/postmortem-pr-5696.md)
 
 ## 路线图

package/lib/client.js

@@ -51,6 +51,7 @@ window.__ModuleLoader__.load({
 			const [groups, setGroups] = (0, react.useState)(null);
 			const [maxBytes, setMaxBytes] = (0, react.useState)(4096);
 			const [error, setError] = (0, react.useState)(false);
+			const [actionError, setActionError] = (0, react.useState)(null);
 			const [editing, setEditing] = (0, react.useState)(null);
 			const [busy, setBusy] = (0, react.useState)(false);
 			const load = (0, react.useCallback)(async () => {
@@ -68,12 +69,15 @@ window.__ModuleLoader__.load({
 			}, [load]);
 			const act = async (fn) => {
 				setBusy(true);
+				setActionError(null);
 				try {
 					const response = await fn();
-					if (!response.ok) console.warn("[auto-memory] action failed", response.status);
-					await load();
+					if (!response.ok) setActionError(response.status === 404 ? t("deleteFailed") : t("saveFailed"));
+				} catch {
+					setActionError(t("saveFailed"));
 				} finally {
 					setBusy(false);
+					await load();
 				}
 			};
 			const cwd = (group) => group.scope === "project" && group.current === true ? currentCwd() : void 0;
@@ -120,13 +124,19 @@ window.__ModuleLoader__.load({
 						style: { opacity: .75 },
 						children: t("intro")
 					}),
-					/* @__PURE__ */ (0, react_jsx_runtime.jsx)("p", { children: /* @__PURE__ */ (0, react_jsx_runtime.jsx)("button", {
+					/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("p", { children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)("button", {
 						onClick: () => {
 							load();
 						},
 						disabled: busy,
 						children: t("refresh")
-					}) }),
+					}), actionError !== null && /* @__PURE__ */ (0, react_jsx_runtime.jsx)("span", {
+						style: {
+							marginLeft: "12px",
+							color: "#d97706"
+						},
+						children: actionError
+					})] }),
 					total === 0 && /* @__PURE__ */ (0, react_jsx_runtime.jsx)("p", { children: t("empty") }),
 					groups.map((group) => {
 						const usage = group.indexBytes / maxBytes;
@@ -209,6 +219,10 @@ window.__ModuleLoader__.load({
 												title: t("pinned"),
 												children: " 📌"
 											}),
+											memory.disabled && /* @__PURE__ */ (0, react_jsx_runtime.jsx)("span", {
+												title: t("muted"),
+												children: " 🔇"
+											}),
 											" ",
 											/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("code", {
 												style: {
@@ -282,6 +296,36 @@ window.__ModuleLoader__.load({
 												children: memory.pinned ? t("unpinAction") : t("pinAction")
 											}),
 											" ",
+											/* @__PURE__ */ (0, react_jsx_runtime.jsx)("button", {
+												disabled: busy,
+												onClick: () => {
+													(async () => {
+														try {
+															const detail = await read({
+																name: memory.name,
+																scope: group.scope,
+																cwd: cwd(group)
+															});
+															await act(() => write({
+																cwd: cwd(group),
+																scope: group.scope,
+																name: memory.name,
+																title: memory.title,
+																description: memory.description,
+																type: memory.type,
+																body: detail.body,
+																pinned: memory.pinned,
+																disabled: !memory.disabled
+															}));
+														} catch (error) {
+															console.warn("[auto-memory] read failed, mute aborted", error);
+															await load();
+														}
+													})();
+												},
+												children: memory.disabled ? t("unmuteAction") : t("muteAction")
+											}),
+											" ",
 											/* @__PURE__ */ (0, react_jsx_runtime.jsx)("button", {
 												disabled: busy,
 												onClick: () => {
@@ -303,6 +347,7 @@ window.__ModuleLoader__.load({
 						t,
 						initial: editing.memory,
 						initialBody: editing.body,
+						busy,
 						onCancel: () => {
 							setEditing(null);
 						},
@@ -319,7 +364,7 @@ window.__ModuleLoader__.load({
 			});
 		}
 		/** 单条记忆编辑器(标题/摘要/正文,保存即重建索引、下次注入生效)。 */
-		function MemoryEditor({ t, initial, initialBody, onSave, onCancel }) {
+		function MemoryEditor({ t, initial, initialBody, busy, onSave, onCancel }) {
 			const [title, setTitle] = (0, react.useState)(initial.title ?? "");
 			const [description, setDescription] = (0, react.useState)(initial.description);
 			const [body, setBody] = (0, react.useState)(initialBody);
@@ -383,6 +428,7 @@ window.__ModuleLoader__.load({
 						style: { margin: "8px 0 0" },
 						children: [
 							/* @__PURE__ */ (0, react_jsx_runtime.jsx)("button", {
+								disabled: busy,
 								onClick: () => {
 									onSave({
 										name: initial.name,
@@ -397,6 +443,7 @@ window.__ModuleLoader__.load({
 							}),
 							" ",
 							/* @__PURE__ */ (0, react_jsx_runtime.jsx)("button", {
+								disabled: busy,
 								onClick: onCancel,
 								children: t("cancel")
 							})
@@ -436,6 +483,9 @@ window.__ModuleLoader__.load({
 			userScope: "用户级(所有工作区)",
 			projectScope: "项目级",
 			pinned: "置顶",
+			muted: "已停用(不注入)",
+			muteAction: "停用",
+			unmuteAction: "恢复",
 			reads: "次读取",
 			updated: "更新于",
 			injectedPreview: "下一次会话将注入的索引(点开预览)",
@@ -470,6 +520,9 @@ window.__ModuleLoader__.load({
 			userScope: "User scope (all workspaces)",
 			projectScope: "Project scope",
 			pinned: "pinned",
+			muted: "muted (not injected)",
+			muteAction: "Mute",
+			unmuteAction: "Unmute",
 			reads: "reads",
 			updated: "updated",
 			injectedPreview: "Index injected into the next session (click to preview)",
@@ -537,11 +590,19 @@ window.__ModuleLoader__.load({
 				currentCwd: () => currentCwd(ctx)
 			};
 		}
-		/** 从 sessions 快照取一个可用的工作区 cwd(官方 ui-reference 同款读取面)。 */
+		/** 从 sessions 快照解析当前工作区 cwd(#9:多工作区并存时不再取首个)。 */
 		function currentCwd(ctx) {
 			const byId = ctx.get("sessions")?.list?.getSnapshot?.()?.byId;
 			if (byId === void 0) return void 0;
-			for (const session of Object.values(byId)) if (session.cwd !== void 0 && session.cwd.length > 0) return session.cwd;
+			const rows = Object.values(byId).filter((session) => session.cwd !== void 0 && session.cwd.length > 0);
+			if (rows.length === 0) return void 0;
+			const running = rows.filter((session) => session.running === true);
+			const pool = running.length > 0 ? running : rows;
+			pool.sort((a, b) => {
+				const delta = (b.updatedAt ?? 0) - (a.updatedAt ?? 0);
+				return delta !== 0 ? delta : a.cwd < b.cwd ? -1 : 1;
+			});
+			return pool[0]?.cwd;
 		}
 		/** 注册侧栏入口与记忆管理面板。 */
 		function apply(ctx) {

package/lib/index.js

@@ -14,7 +14,7 @@ const USER_SCOPE_DIR = "_user";
 /**
 * 记忆存储层:类型化记忆文件 CRUD + MEMORY.md 索引维护。
 *
-* 关键设计决策(依据源码调研与对抗性审查,见 docs/api-reports.md):
+* 关键设计决策(依据源码调研与对抗性审查,见 internal research notes):
 * 1. **必须用 node:fs 而非 ctx.fs**——默认部署的 fs-sandbox 可写根不含 $DSH_HOME,
 *    走 ctx.fs 会抛 FS_SANDBOX_DENIED;官方先例 skill-filesystem 访问 $DSH_HOME
 *    下的文件同样直接用 node:fs。
@@ -114,7 +114,7 @@ function parseMemory(raw, scope) {
 	try {
 		const fm = parseFrontmatter(raw);
 		if (!fm) return null;
-		const { name, description, type, title, created, updated, lastRead, reads, pinned, importance } = fm.data;
+		const { name, description, type, title, created, updated, lastRead, reads, pinned, importance, disabled } = fm.data;
 		if (typeof name !== "string" || typeof description !== "string" || description.trim().length === 0) return null;
 		const parsedType = type === void 0 ? "reference" : asMemoryType(String(type));
 		const asMs = (value) => typeof value === "number" && Number.isSafeInteger(value) && value > 0 ? value : void 0;
@@ -128,6 +128,7 @@ function parseMemory(raw, scope) {
 			body: fm.body.trim(),
 			scope,
 			...pinned === true ? { pinned: true } : {},
+			...disabled === true ? { disabled: true } : {},
 			...asImportance(importance) !== void 0 ? { importance: asImportance(importance) } : {},
 			createdMs: asMs(created),
 			updatedMs: asMs(updated),
@@ -146,6 +147,7 @@ function serializeMemory(record) {
 		description: record.description,
 		type: record.type,
 		...record.pinned === true ? { pinned: true } : {},
+		...record.disabled === true ? { disabled: true } : {},
 		...record.importance !== void 0 ? { importance: record.importance } : {},
 		...record.createdMs !== void 0 ? { created: record.createdMs } : {},
 		...record.updatedMs !== void 0 ? { updated: record.updatedMs } : {},
@@ -171,7 +173,7 @@ function mergeLifecycleMeta(existing, now) {
 */
 function renderIndexBody(records) {
 	const now = Date.now();
-	const lines = [...records].sort((a, b) => {
+	const lines = [...records.filter((record) => record.disabled !== true)].sort((a, b) => {
 		const pa = a.pinned === true ? 0 : 1;
 		const pb = b.pinned === true ? 0 : 1;
 		if (pa !== pb) return pa - pb;
@@ -239,12 +241,54 @@ const SECRET_PATTERNS = [
 	}
 ];
 /**
+* GB11643 身份证校验位验证:18 位号码仅当末位与加权模 11 校验相符才算真身份证。
+* 纯长度匹配会把订单号/雪花 ID 等任意 18 位数字一并误杀(#7)。
+*/
+function isValidIdCard(candidate) {
+	if (!/^\d{17}[\dXx]$/.test(candidate)) return false;
+	const WEIGHTS = [
+		7,
+		9,
+		10,
+		5,
+		8,
+		4,
+		2,
+		1,
+		6,
+		3,
+		7,
+		9,
+		10,
+		5,
+		8,
+		4,
+		2
+	];
+	const CHECKS = [
+		"1",
+		"0",
+		"X",
+		"9",
+		"8",
+		"7",
+		"6",
+		"5",
+		"4",
+		"3",
+		"2"
+	];
+	let sum = 0;
+	for (let i = 0; i < 17; i++) sum += Number(candidate[i]) * WEIGHTS[i];
+	return candidate[17].toUpperCase() === CHECKS[sum % 11];
+}
+/**
 * 写入前脱敏(纯函数,可测):把高置信密钥/手机号/身份证替换为 [REDACTED:kind]。
-* 应用于 description 与 body——所有写路径(工具/固化/面板)都经 store.write,单点拦截。
+* 应用于 description/body——所有写路径(工具/固化/面板)都经 store.write,单点拦截。
 */
 function redactSecrets(text) {
 	let out = text;
-	for (const { kind, re } of SECRET_PATTERNS) out = out.replace(re, `[REDACTED:${kind}]`);
+	for (const { kind, re } of SECRET_PATTERNS) out = out.replace(re, (match) => kind === "id-card" && !isValidIdCard(match) ? match : `[REDACTED:${kind}]`);
 	return out;
 }
 /**
@@ -280,9 +324,19 @@ function memoryScore(record, nowMs) {
 /**
 * 描述相似度(Jaccard 词集,纯函数,可测):固化防回声的查重依据——
 * "既有记忆的复述不是新信息",相似度过高的候选跳过而非堆积。
+* CJK 感知:连续中文段按字符二元组(bigram)切分(检索领域标准做法)——
+* 整段单 token 会让中文近重复仅得 ~0.4,防回声对中文完全失效(#5)。
 */
 function descriptionSimilarity(a, b) {
-	const words = (text) => new Set(text.toLowerCase().split(/[^a-z0-9一-鿿]+/).filter((w) => w.length > 0));
+	const words = (text) => {
+		const tokens = text.toLowerCase().split(/[^a-z0-9一-鿿]+/).filter((w) => w.length > 0);
+		const out = /* @__PURE__ */ new Set();
+		for (const token of tokens) if (/^[一-鿿]+$/.test(token)) {
+			if (token.length === 1) out.add(token);
+			else for (let i = 0; i + 1 < token.length; i++) out.add(token.slice(i, i + 2));
+		} else out.add(token);
+		return out;
+	};
 	const wa = words(a);
 	const wb = words(b);
 	if (wa.size === 0 || wb.size === 0) return 0;
@@ -364,11 +418,13 @@ var MemoryStore = class {
 		await this.withLockRecovery(join(dir, INDEX_FILENAME), async () => {
 			const existing = parseMemory(await promises.readFile(file, "utf8").catch(() => ""), scope);
 			const pinned = record.pinned ?? existing?.pinned;
+			const disabled = record.disabled ?? existing?.disabled;
 			const importance = record.importance ?? existing?.importance;
 			const title = record.title ?? existing?.title;
 			written = {
 				...record,
 				...pinned === true ? { pinned: true } : {},
+				...disabled === true ? { disabled: true } : {},
 				...importance !== void 0 ? { importance } : {},
 				...title !== void 0 ? { title: sanitizeContent(title) } : {},
 				description: sanitizeContent(record.description),
@@ -615,6 +671,10 @@ function registerMemoryTools(ctx, store, enableUserScope) {
 				type: "boolean",
 				description: "Pin this memory: it sorts first in the index, survives budget truncation, and is never hidden by staleness eviction. Use when the user explicitly says to keep something forever; unpin by passing false"
 			},
+			disabled: {
+				type: "boolean",
+				description: "Soft-mute: keep the file but exclude it from the injected index (the memory stops reaching the model without being deleted). Use when the user says \"stop bringing this up\" or similar; unmute by passing false"
+			},
 			scope: {
 				type: "string",
 				enum: ["project", "user"],
@@ -661,6 +721,7 @@ function registerMemoryTools(ctx, store, enableUserScope) {
 				name,
 				title: args.title !== void 0 && args.title.trim().length > 0 ? args.title.trim().slice(0, 80) : void 0,
 				...args.pinned !== void 0 ? { pinned: args.pinned } : {},
+				...args.disabled !== void 0 ? { disabled: args.disabled } : {},
 				description: args.description.trim().slice(0, 160),
 				type: args.type,
 				body: args.body.slice(0, 2e3)
@@ -1252,6 +1313,9 @@ sessions for this user, following these rules:
 Return ONLY a JSON array, each element exactly:
 {"name":"kebab-case-id","title":"short human heading","description":"one line <=160 chars","type":"user|feedback|project|reference","body":"the fact in markdown; for feedback include **Why:** and **How to apply:** lines","scope":"project","importance":5}
 Use scope "user" only for user-global preferences; default "project".
+Write "name" in kebab-case English; write title/description/body in the SAME
+language as the transcript (a Chinese transcript gets Chinese memories —
+the user will read them).
 
 <transcript>
 ${transcript}
@@ -1399,6 +1463,7 @@ async function readGroup(dir, scope, slug) {
 			description: record.description,
 			type: record.type,
 			pinned: record.pinned === true,
+			disabled: record.disabled === true,
 			reads: record.reads ?? 0,
 			...record.updatedMs !== void 0 ? { updatedMs: record.updatedMs } : {},
 			bytes: Buffer.byteLength(raw, "utf8")
@@ -1470,7 +1535,8 @@ function registerMemoryApi(ctx, store, rootDir, config) {
 					description: record.description,
 					type: record.type,
 					body: record.body,
-					pinned: record.pinned === true
+					pinned: record.pinned === true,
+					disabled: record.disabled === true
 				});
 			} catch (error) {
 				return json({ error: String(error) }, 400);
@@ -1493,6 +1559,7 @@ function registerMemoryApi(ctx, store, rootDir, config) {
 					name,
 					title: typeof body.title === "string" && body.title.trim().length > 0 ? body.title.trim().slice(0, 80) : void 0,
 					...body.pinned !== void 0 ? { pinned: body.pinned } : {},
+					...body.disabled === true ? { disabled: true } : {},
 					description: body.description.trim().slice(0, 160),
 					type: body.type,
 					body: typeof body.body === "string" ? body.body.slice(0, 2e3) : ""

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-auto-memory",
-  "version": "0.5.1",
+  "version": "0.6.0",
   "description": "Claude Code style auto-memory plugin for DeepSeek Harness: typed memory files + MEMORY.md index auto-injected into the system prompt",
   "type": "module",
   "license": "MIT",