@yolk_vat-y/dsh-project-memory 0.5.13 → 0.5.14

package/src/index.js

@@ -106,8 +106,11 @@ export const Config = Schema.object({
     // Schema 里声明过：走 cordis.patch.yml 配它们会被宿主按「not a declared property」拒掉，
     // 等于文档里的旋钮是假的。默认值以 cfgEngine 的兜底值为准（那里是权威，这里只负责暴露）。
     gateCooldownSteps: Schema.number().default(2),
-    maxItemsPerSession: Schema.number().default(12),
-    maxItemCharsPerSession: Schema.number().default(4000),
+    // 这两个是**保险丝**不是节流阀：节流由 gateCooldownSteps 与单轮 maxTokens 负责。
+    // 旧默认 12 / 4000 ≈ 11 条，实测真实日志里 11/81 个会话打满、打满后永久静默（1234 步失明）；
+    // 实测最长会话用 35 条 / 8640 字符，故抬到 60 / 24000。显式配小值可复现旧行为。
+    maxItemsPerSession: Schema.number().default(60),
+    maxItemCharsPerSession: Schema.number().default(24000),
     hintMinCoverage: Schema.number().default(0.45),
     hintMinMatched: Schema.number().default(2),
     hintMinSupport: Schema.number().default(0.15),
@@ -117,8 +120,10 @@ export const Config = Schema.object({
     // 影子记录（admission-shadow.jsonl）：**每步**一行，含全部候选的判据特征与场景。
     // 主审计只在真的注入时写，静默步零痕迹 → 无法离线重放"换个阈值会怎样"，也攒不出样本。
     // 只写盘、不进 prompt、不花 token，所以默认开。
+    // 512KB 而不是 2MB：轮转只留 `.1` 一代，所以单项目日志的硬上限是
+    // `2 × shadowMaxBytes + 2 × auditMaxBytes`（原 4.6MB）；512KB 下约 1.5MB。
     shadowLog: Schema.boolean().default(true),
-    shadowMaxBytes: Schema.number().default(2097152),
+    shadowMaxBytes: Schema.number().default(524288),
   }).default({}),
 })
 

package/src/insight-store.js

@@ -6,10 +6,11 @@
 // - 归档 = 软删（archived:true，不进召回/注入）；物理删除只发生在容量溢出
 // - 写盘前密文过滤：命中密钥/token/私钥形态 → 拒绝写入
 import { randomUUID } from 'node:crypto'
-import { mkdirSync, readFileSync, renameSync, writeFileSync } from 'node:fs'
+import { readFileSync, renameSync } from 'node:fs'
 import path from 'node:path'
 import os from 'node:os'
 import { normalizedTokenOverlap, findBestOverlapMatch, insightMatchText } from './similarity.js'
+import { writeJsonAtomic } from './util/fs.js'
 import { backfillDerivedTriggers } from './readiness.js'
 
 export const INSIGHT_KINDS = ['lesson', 'decision', 'procedure', 'experience']
@@ -332,22 +333,6 @@ function readJson(file, fallback) {
   }
 }
 
-function writeJsonAtomic(file, data) {
-  mkdirSync(path.dirname(file), { recursive: true })
-  const tmp = `${file}.${process.pid}.tmp`
-  try {
-    writeFileSync(tmp, JSON.stringify(data))
-    renameSync(tmp, file)
-  } catch (err) {
-    try {
-      renameSync(tmp, file) // fsync 无关紧要：单文件覆盖 + 崩溃后 .tmp 清理即可
-    } catch {
-      /* ignore */
-    }
-    throw err
-  }
-}
-
 function normalizeDoc(doc) {
   if (doc && typeof doc === 'object' && Array.isArray(doc.items)) {
     if (typeof doc.version !== 'number') doc.version = 1

package/src/readiness.js

@@ -447,10 +447,35 @@ export function auditTriggers(items, opts = {}) {
  * @returns {{coverage: number, matched: number, supported: number, terms: number}}
  */
 export function idfCoverage(query, itemText, corpusTexts) {
+  return idfCoveragePrepared(query, itemText, corpusTermSets(corpusTexts))
+}
+
+/**
+ * 语料 → 词集合数组。**对一批候选循环评分时必须只建一次**。
+ *
+ * `idfCoverage(query, item, corpusTexts)` 每次都把整个语料重新 tokenize 一遍；调用方对 C 个
+ * 候选各调一次，就是 C×C 次 tokenize —— 实测真实 store（72 条）单步 147 ms，提出循环外
+ * 降到 1.3 ms（109×），且 coverage 结果逐字段一致。热路径上每个 pre-step 都要跑一次，所以
+ * 这个差别是"每条模型消息"级别的。
+ * @param {string[]} texts - 语料正文
+ * @returns {Set<string>[]} 与输入同序的词集合
+ */
+export function corpusTermSets(texts) {
+  return (texts || []).map((t) => new Set(tokenize(t)))
+}
+
+/**
+ * 与 {@link idfCoverage} 同一套判据，但语料词集合由调用方预先建好（见 {@link corpusTermSets}）。
+ * @param {string} query - 查询文本
+ * @param {string} itemText - 被测条目的评分文本
+ * @param {Set<string>[]} corpusSets - {@link corpusTermSets} 的产物
+ * @returns {{coverage: number, matched: number, supported: number, terms: number}}
+ */
+export function idfCoveragePrepared(query, itemText, corpusSets) {
   const q = [...new Set(tokenize(query))]
   if (!q.length) return { coverage: 0, matched: 0, supported: 0, terms: 0 }
   const item = new Set(tokenize(itemText))
-  const corpus = (corpusTexts || []).map((t) => new Set(tokenize(t)))
+  const corpus = corpusSets || []
   const n = Math.max(corpus.length, 1)
   let total = 0
   let hit = 0

package/src/store.js

@@ -2,6 +2,7 @@ import { createHash, randomUUID } from 'node:crypto'
 import { mkdirSync, readFileSync, readdirSync, renameSync, rmSync, statSync, unlinkSync, writeFileSync } from 'node:fs'
 import path from 'node:path'
 import { rankExperience, tokenize, tokenizeRaw, extractCjkPhrases, makeSearchText } from './util/search.js'
+import { writeJsonAtomic } from './util/fs.js'
 import { backfillDerivedTriggers } from './readiness.js'
 
 const FORMAT_FILE = 'format.json'
@@ -188,21 +189,6 @@ function loadJson(filePath, fallback) {
   }
 }
 
-function writeJsonAtomic(filePath, data) {
-  const tmp = `${filePath}.${process.pid}.tmp`
-  try {
-    writeFileSync(tmp, JSON.stringify(data))
-    renameSync(tmp, filePath)
-  } catch (err) {
-    try {
-      unlinkSync(tmp)
-    } catch {
-      // tmp already gone (rename succeeded) or undeletable; nothing to do
-    }
-    throw err
-  }
-}
-
 function shardRelPath(dir, rel) {
   return path.join(dir, SHARDS_DIR, createHash('sha256').update(rel).digest('hex') + '.json')
 }
@@ -315,8 +301,8 @@ export class ProjectMemoryStore {
         `[dsh-project-memory] migration dropped ${orphans.length} entry group(s) with no index record: ${orphans.slice(0, 3).join(', ')}${orphans.length > 3 ? ' …' : ''}`,
       )
     }
-    mkdirSync(path.join(this.dir, SHARDS_DIR), { recursive: true })
     for (const rel of Object.keys(files)) {
+      // 父目录由 writeJsonAtomic 保证
       writeJsonAtomic(shardRelPath(this.dir, rel), { relPath: rel, record: files[rel], entries: (entries[rel] || []).map(withoutPersistedDerived) })
     }
     writeJsonAtomic(formatPath, { version: 2, layout: 'sharded' })
@@ -524,7 +510,7 @@ export class ProjectMemoryStore {
     }
     for (const rel of this._dirtyShards) {
       if (this.files[rel]) {
-        mkdirSync(path.join(this.dir, SHARDS_DIR), { recursive: true })
+        // 父目录由 writeJsonAtomic 保证
         writeJsonAtomic(shardRelPath(this.dir, rel), { relPath: rel, record: this.files[rel], entries: (this.entries[rel] || []).map(withoutPersistedDerived) })
       } else {
         this._removedShards.add(rel)
@@ -580,11 +566,11 @@ export class ProjectMemoryStore {
     if (N === 0) return {}
     const df = {}
     for (const entry of entries) {
-      const text = entry.title || ''
-      const keywords = (entry.keywords || []).join(' ')
-      const summary = entry.summary || ''
-      const combined = `${text} ${text} ${text} ${text} ${text} ${keywords} ${summary}`.toLowerCase()
-      const terms = tokenizeRaw(combined)
+      // 语料必须与**检索**同源：检索走 search.js 的 weightedFieldText = title×5 + keywords
+      // + summary + **terms** + sourcePath。此前这里手抄了一份少了 `terms`/`sourcePath` 的
+      // 副本，于是只出现在 `terms` 里的词 df=0 → 被当成极稀有词拿虚高 IDF（而 `terms` 正是
+      // 0.5.x 为修"只有前 300 字符可检索"加的主检索面）。直接调 makeSearchText，权重自动对齐。
+      const terms = tokenizeRaw(makeSearchText(entry))
       const seen = new Set(terms)
       for (const t of seen) {
         df[t] = (df[t] || 0) + 1

package/src/tools/forget.js

@@ -11,7 +11,7 @@ export function forgetTool(config) {
       id_or_query: {
         type: 'string',
         required: true,
-        description: 'Experience note id (from remember/search output), or keywords to match.',
+        description: 'Experience note id (from remember or query_memory output), or keywords to match.',
       },
       root: {
         type: 'string',

package/src/tools/lesson-tools.js

@@ -16,6 +16,7 @@ export function lessonTool(config) {
     name: 'save_lesson',
     description:
       'Save a lesson/decision/procedure/experience into the tiered memory (single insight entity). ' +
+      'Prefer this over remember for anything new: entries here carry scope, merge/reinforce, and auto-promote. ' +
       'Call it when the task hit a real pitfall, got corrected, or made a deliberate choice worth remembering — not for routine work. ' +
       'Similar entries auto-merge (bidirectional token overlap >= 0.7) or reinforce (0.65~0.7, accumulating task hits); ' +
       'when the same insight is hit by 2+ tasks it auto-promotes task -> project, 3+ tasks project -> global. ' + kindDesc + ' ' + scopeDesc,

package/src/tools/remember.js

@@ -6,9 +6,11 @@ export function rememberTool(config) {
   return defineTool({
     name: 'remember',
     description:
-      'Save an experience note (problem -> solution) into project memory, e.g. a bug you just fixed or a decision made. ' +
-      'Retrieved later only when search_experience / query_memory matches the problem, never auto-injected. ' +
-      'If a note with a similar problem exists, it is superseded instead of duplicated.',
+      'Save an experience note (problem -> solution) into project memory, e.g. a bug you just fixed. ' +
+      'Retrieved later only when query_memory matches the problem, never auto-injected. ' +
+      'A note for a similar problem is superseded instead of duplicated. ' +
+      'This writes the standalone experience file: for lessons / decisions / procedures prefer save_lesson ' +
+      '(the tiered insight store) — query_memory searches both.',
     parameters: {
       problem: {
         type: 'string',

package/src/tools/task-tools.js

@@ -149,13 +149,14 @@ export function showTaskPanelTool(config) {
       'Show the task panel in the UI. Call this when the user asks to see the task list or when you want to display the task panel to the user.',
     parameters: {},
     output: { schema: { type: 'string' }, render: (_a, v) => [{ type: 'text', text: v }] },
-    async execute(_args, exec) {
-      const ctx = exec?.ctx
-      if (!ctx) return '无法获取上下文'
-      // 通过 BroadcastChannel 通知客户端打开面板
-      try {
-        ctx.events?.emit?.('dsh:task-panel:show', {})
-      } catch { /* ignore */ }
+    async execute() {
+      // 开面板的动作在**客户端**：这个工具名被 src/client/ShowTaskPanelNode.tsx 认领，
+      // 结果节点一渲染就由浏览器侧调 `taskUIStore.open()`。
+      //
+      // 宿主进程没有触碰面板状态的通道 —— 面板的 closed/minimized 是浏览器侧 UI store，
+      // 而 `ToolRunContext` 只有 `deferContext()` / `concludeTurn()` 两个自有成员
+      // （packages/core/tools/src/index.ts:418），**没有 `ctx`**。旧实现读 `exec.ctx` 再
+      // `emit('dsh:task-panel:show')`，于是每次都返回「无法获取上下文」，面板从未打开过。
       return '已请求打开任务面板。'
     },
   })

package/src/tools/watch-repo.js

@@ -9,7 +9,7 @@ export function watchRepoTool(watchManager, config) {
     description:
       'Start silent auto-refresh for a project: the plugin polls the root (configurable interval), ' +
       'detects changed/new docs and code files via mtime+content-hash, and re-indexes only those silently. ' +
-      'No GUI, no manual re-index needed after this. Stop by reloading the plugin or calling with watch=false.',
+      'No manual re-index needed after this. Call again with watch=false to stop.',
     parameters: {
       root: {
         type: 'string',

package/src/util/fs.js

@@ -1,9 +1,41 @@
-import { accessSync, constants, existsSync, readdirSync, readFileSync, statSync } from 'node:fs'
+import {
+  accessSync, constants, existsSync, mkdirSync, readdirSync, readFileSync, renameSync, statSync,
+  unlinkSync, writeFileSync,
+} from 'node:fs'
 import { readFile } from 'node:fs/promises'
 import { createHash } from 'node:crypto'
 import os from 'node:os'
 import path from 'node:path'
 
+/**
+ * 原子写 JSON：先写同目录的 `<file>.<pid>.tmp`，再 rename 覆盖目标。
+ *
+ * 两点语义，别再各写一份副本：
+ * - 先建父目录（global insight 文件首次写入时父目录可能不存在）。
+ * - 失败时删掉自己的 `.tmp`，不重试 rename：同样的参数重试没有意义，而 `.tmp` 留下只有
+ *   `cleanStaleTmp()` 会清，它扫不到 `~/.config/dsh-project-memory/`。
+ *
+ * 崩溃（kill -9 / 断电）走不到 catch，那种残留由 `cleanStaleTmp()` 兜底。
+ * @param {string} file - 目标文件绝对路径。
+ * @param {unknown} data - 要序列化的 JSON 值。
+ * @throws 序列化失败或写盘/rename 失败时原样抛出（调用方决定回退策略）。
+ */
+export function writeJsonAtomic(file, data) {
+  mkdirSync(path.dirname(file), { recursive: true })
+  const tmp = `${file}.${process.pid}.tmp`
+  try {
+    writeFileSync(tmp, JSON.stringify(data))
+    renameSync(tmp, file)
+  } catch (err) {
+    try {
+      unlinkSync(tmp)
+    } catch {
+      // tmp 已不在（rename 其实成功了）或不可删；两种都无需处理
+    }
+    throw err
+  }
+}
+
 export function assertReadableFile(filePath, maxFileSizeMb) {
   if (typeof filePath !== 'string' || !filePath) {
     throw new Error('file_path must be a non-empty string')