claude-mem-lite 5.0.0 → 5.1.0

package/.claude-plugin/marketplace.json

@@ -10,7 +10,7 @@
   "plugins": [
     {
       "name": "claude-mem-lite",
-      "version": "5.0.0",
+      "version": "5.1.0",
       "source": "./",
       "description": "Persistent long-term memory for Claude Code via MCP — captures coding decisions, bugfixes, and context across sessions. Hybrid FTS5 + TF-IDF search with episode batching. Single SQLite DB, no external services. A lighter, lower-cost alternative to claude-mem (episode batching + a smaller model; cost savings are an internal estimate, not a measured benchmark)."
     }

package/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-mem-lite",
-  "version": "5.0.0",
+  "version": "5.1.0",
   "description": "Persistent long-term memory for Claude Code via MCP — captures coding decisions, bugfixes, and context across sessions. Hybrid FTS5 + TF-IDF search with episode batching. Single SQLite DB, no external services. A lighter, lower-cost alternative to claude-mem (episode batching + a smaller model; cost savings are an internal estimate, not a measured benchmark).",
   "author": {
     "name": "sdsrss"

package/README.md

@@ -75,7 +75,7 @@ How claude-mem-lite differs from the major neighbors in the LLM-memory space (ve
 
 ## Features
 
-- **Automatic capture** -- Hooks into Claude Code lifecycle (PostToolUse, SessionStart, Stop, UserPromptSubmit) to record observations without manual effort
+- **Automatic capture** -- Hooks into the Claude Code lifecycle (SessionStart, PreCompact, PreToolUse, PostToolUse, PostToolUseFailure, Stop, UserPromptSubmit — the seven events in `hooks/hooks.json`) to record observations without manual effort
 - **Hybrid search** -- FTS5 BM25 + TF-IDF vector cosine similarity, merged via Reciprocal Rank Fusion (RRF). FTS5 handles keyword matching; 512-dim TF-IDF vectors capture semantic similarity for recall beyond exact terms
 - **Timeline browsing** -- Navigate observations chronologically with anchor-based context windows
 - **Episode batching** -- Groups related file operations into coherent episodes before LLM encoding
@@ -105,14 +105,12 @@ How claude-mem-lite differs from the major neighbors in the LLM-memory space (ve
 - **Atomic writes** -- All file writes (episodes, CLAUDE.md) use write-to-tmp + rename to prevent corruption on crash
 - **Robust locking** -- PID-aware lock files with automatic stale/orphan cleanup (>30s timeout or dead PID)
 - **Stale session cleanup** -- Sessions active for >24h are automatically marked as abandoned on next start
-- **Unified resource discovery** -- Shared filesystem traversal layer (`resource-discovery.mjs`) used by both runtime scanner and offline indexer, supporting flat directories, plugin nesting, and loose `.md` files
-- **Domain synonym expansion** -- Registry search queries expand to domain synonyms (e.g., "fix" → debug, bugfix, troubleshoot, diagnose, repair)
+- **Domain synonym expansion** -- Search queries expand to domain synonyms (e.g., "fix" → debug, bugfix, troubleshoot, diagnose, repair)
 - **Multi-provider LLM mode** -- Provider priority `ANTHROPIC_API_KEY` (direct Anthropic API) → `OPENROUTER_API_KEY` (OpenRouter, OpenAI-compatible — point it at any model via `OPENROUTER_MODEL`) → `claude -p` CLI fallback when no key is set
 - **Lesson-learned indexing** -- `lesson_learned` field indexed in FTS5 with weight 8, making past debugging insights directly searchable
 - **Cross-source normalization** -- `mem_search` normalizes scores across observations, sessions, and prompts before merging, preventing any source from dominating results
 - **Exponential recency decay** -- Type-differentiated half-lives (decisions: 90d, discoveries: 60d, bugfixes: 14d, changes: 7d) consistently applied in all ranking paths
 - **Prompt-time memory injection** -- UserPromptSubmit hook automatically searches and injects relevant past observations with recency and importance weighting
-- **Smart skill invocation** -- Auto-loaded and searched managed skills/agents include portable `~` paths with `Read()` guidance; native plugin skills recommend `Skill("full:name")`; prevents `Skill()` misuse for managed resources that aren't registered with Claude Code's native handler
 - **Dual injection dedup** -- `user-prompt-search.js` and `handleUserPrompt` coordinate via temp file to prevent duplicate memory injection
 - **Plugin cache hook self-heal** -- Claude Code runtime reads plugin hooks from `~/.claude/plugins/cache/<mp>/<plugin>/<ver>/hooks/hooks.json`, not from the marketplace source. When `install.mjs`-managed `settings.json` hooks coexist with a stale cache `hooks.json` (e.g. from a previous marketplace install or a plugin auto-update), the runtime registers hooks twice → every session start / user prompt fires twice. `install.mjs` and `hook-update.mjs` now clear cache `hooks.json` in every version dir, and `hook.mjs session-start` self-heals on every session (gated by `hasInstallManagedHooks` so plugin-only users are not affected). `install.mjs status` reports cache pollution state (since v2.31.1/2.31.2).
 - **Result-dedup cooldown** -- User-prompt memory injection uses result-overlap detection (>80% ID overlap → skip) instead of time-based cooldown, allowing topic switches within seconds while preventing redundant injections
@@ -188,8 +186,8 @@ Source files stay in the cloned repo. Update via `git pull && node install.mjs i
 ### What happens during installation
 
 1. **Install dependencies** -- `npm install --omit=dev` (compiles native `better-sqlite3`)
-2. **Register MCP server** -- `mem-lite` server with 20 tools (9 core exposed via `tools/list` + 11 hidden-but-callable; see the Usage section for the full table). The pre-v2.78 generic server name `mem` is renamed to `mem-lite` for namespace hygiene; the tool names themselves (`mem_search`, `mem_recall`, ...) are unchanged.
-3. **Configure hooks** -- `PostToolUse`, `SessionStart`, `Stop`, `UserPromptSubmit` lifecycle hooks
+2. **Register MCP server** -- `mem-lite` server with 18 tools (9 core exposed via `tools/list` + 9 hidden-but-callable; see the Usage section for the full table). The pre-v2.78 generic server name `mem` is renamed to `mem-lite` for namespace hygiene; the tool names themselves (`mem_search`, `mem_recall`, ...) are unchanged.
+3. **Configure hooks** -- all seven lifecycle events: `SessionStart`, `PreCompact`, `PreToolUse`, `PostToolUse`, `PostToolUseFailure`, `Stop`, `UserPromptSubmit`
 4. **Create data directory** -- `~/.claude-mem-lite/` (hidden) for database, runtime, and managed resource files
 5. **Auto-migrate** -- If `~/.claude-mem/` (original claude-mem) or `~/claude-mem-lite/` (pre-v0.5 unhidden) exists, migrates database and runtime files to `~/.claude-mem-lite/`, preserving the original untouched
 6. **Initialize database** -- SQLite with WAL mode, FTS5 indexes created on first server start
@@ -228,8 +226,6 @@ rm -rf ~/claude-mem-lite/   # pre-v0.5 unhidden (if not auto-moved)
     ep-flush-*.json      # Flushed episodes awaiting processing
     reads-<project>.txt  # Read file paths (collected on flush)
   managed/
-    skills/              # Standalone skills: {name}/SKILL.md
-    agents/              # Agent plugins: {group}/agents/{name}.md + skills/*/SKILL.md
     repos/               # Shallow-cloned source repos
 ```
 
@@ -237,11 +233,14 @@ rm -rf ~/claude-mem-lite/   # pre-v0.5 unhidden (if not auto-moved)
 
 ### MCP Tools (used automatically by Claude)
 
-As of v2.70.0, the server registers 20 tools in total but only the 9 **core**
-tools appear in `tools/list`. The 11 **hidden** tools remain callable at the
+As of v2.70.0, the server registers 18 tools in total but only the 9 **core**
+tools appear in `tools/list`. The 9 **hidden** tools remain callable at the
 protocol layer (`tools/call` by exact name still routes normally); they're
-omitted from the list response so Claude Code sessions don't load 11 extra
-tool schemas at startup. Hidden tools are the maintenance / admin / browser
+omitted from the list response so Claude Code sessions don't load 9 extra
+tool schemas at startup. (It read 20 / 11 until v5.0.0 removed the two skill-registry
+tools — `tool-schemas.mjs` is the source of truth, and
+`tests/tool-count-docs.test.mjs` now holds this paragraph, both README tool tables,
+`README.zh-CN.md`, `llms.txt` and `docs/ARCHITECTURE.md` to it.) Hidden tools are the maintenance / admin / browser
 surface — reach them through the CLI column in the second table.
 
 **Core (9, exposed to Claude Code)**
@@ -258,7 +257,7 @@ surface — reach them through the CLI column in the second table.
 | `mem_defer_list` | List open deferred items for the current project. |
 | `mem_defer_drop` | Drop a deferred item without fixing it; requires a `reason` for the audit trail. |
 
-**Hidden-but-callable (11, CLI-routed)**
+**Hidden-but-callable (9, CLI-routed)**
 
 | Tool | CLI equivalent | Notes |
 |------|----------------|-------|
@@ -452,9 +451,6 @@ UserPromptSubmit (two parallel paths)
   -> [user-prompt-search.js] Auto-search memory via FTS5 + active file context
   -> [user-prompt-search.js] Inject relevant past observations with recency/importance weighting
   -> [user-prompt-search.js] Write injected IDs to temp file for dedup
-  -> [user-prompt-search.js] L1 skill auto-load: match managed skill names in prompt
-     -> Load content with portable ~ path + Read() guidance
-     -> source="managed-skill|managed-agent", path="~/.claude-mem-lite/managed/..."
   -> [hook.mjs handleUserPrompt] Capture user prompt text to user_prompts table
   -> [hook.mjs handleUserPrompt] Increment session prompt counter
   -> [hook.mjs handleUserPrompt] Handoff: detect continuation intent → inject previous session context
@@ -547,7 +543,7 @@ Shows MCP registration, hook configuration, plugin disabled state, and database
 
 ### Recovery (stuck install / hook errors)
 
-If you see `ERR_MODULE_NOT_FOUND` on PreToolUse:Read/Edit/Skill hooks, or `claude-mem-lite` commands crash with import errors, you're likely hit by a partial auto-update — the updater copied new scripts but missed a sibling `lib/*` file, breaking the hook chain (and the next auto-update that would have healed it).
+If you see `ERR_MODULE_NOT_FOUND` on PreToolUse:Read/Edit hooks, or `claude-mem-lite` commands crash with import errors, you're likely hit by a partial auto-update — the updater copied new scripts but missed a sibling `lib/*` file, breaking the hook chain (and the next auto-update that would have healed it).
 
 **v2.84.0+** ships a `repair` subcommand that re-syncs from the latest GitHub release:
 
@@ -636,12 +632,11 @@ claude-mem-lite/
   scripts/
     setup.sh           # Setup hook: npm install + migration (hidden dir + old dir)
     post-tool-use.sh   # Bash pre-filter: skips noise in ~5ms, tracks Read paths
-    user-prompt-search.js # UserPromptSubmit hook: auto-search memory + L1 skill auto-load
-    pre-skill-bridge.js  # PreToolUse hook: L2 skill bridge for managed resources
+    user-prompt-search.js # UserPromptSubmit hook: auto-search memory on user prompts
     pre-tool-recall.js   # PreToolUse hook: file lesson recall before Edit/Write
+    post-tool-recall.js  # PostToolUse hook: error recall after a failed tool call
+    pre-agent-inject.sh  # PreToolUse hook: context for spawned agents
     prompt-search-utils.mjs # Shared logic: skip patterns, intent detection, name matching
-    convert-commands.mjs # Converts command .md → SKILL.md in managed plugins
-    index-managed.mjs  # Offline indexer for managed resources
   # Test & benchmark (dev only)
   tests/               # Unit, property, integration, contract, E2E, pipeline tests
   benchmark/           # BM25 search quality benchmarks + CI gate
@@ -765,7 +760,7 @@ claude-mem-lite.
 
 | Variable | Description | Default |
 |----------|-------------|---------|
-| `CLAUDE_MEM_ALL_TOOLS` | `1` exposes all 20 MCP tools in `tools/list` instead of the 9 core ones (pre-v2.34.0 behavior). The 11 hidden tools stay callable by exact name either way. | _(9 core)_ |
+| `CLAUDE_MEM_ALL_TOOLS` | `1` exposes all 18 MCP tools in `tools/list` instead of the 9 core ones (pre-v2.34.0 behavior). The 9 hidden tools stay callable by exact name either way. | _(9 core)_ |
 | `CLAUDE_MEM_FILE_INTEL` | `0` disables the file-intel block injected before `Read` (past observations about the file you are about to open). | _(on)_ |
 | `CLAUDE_MEM_FILE_INTEL_MIN_TOKENS` | Files smaller than this stay silent — file-intel only pays for itself on large files. | `800` |
 | `CLAUDE_MEM_REREAD_GUARD` | `0` disables the warning when the same file is read twice in a session. Never fires on `offset`/`limit` paging. | _(on)_ |
@@ -835,7 +830,7 @@ what is already stored — only whether new work runs.
 | `CLAUDE_MEM_SKIP_MAINTAIN` | Skip the 24h auto-maintain pass (decay, purge, backup). | _(runs)_ |
 | `CLAUDE_MEM_SKIP_OPTIMIZE` | Skip the LLM optimization pass (re-enrich, normalize, cluster-merge). | _(runs)_ |
 | `CLAUDE_MEM_SKIP_AUTO_DEDUP_FUZZY` | Skip the MinHash near-duplicate pass, keeping exact dedup. | _(runs)_ |
-| `CLAUDE_MEM_SKIP_MARKER_GC` | Skip the runtime-marker sweep. | _(runs)_ |
+| `CLAUDE_MEM_SKIP_MARKER_GC` | Skip the runtime-marker sweep. **Must be exactly `1`** — unlike the other `CLAUDE_MEM_SKIP_*` flags, which accept any truthy value, this one compares against the string `1`. That is deliberate: a truthy check makes `=0` mean "skip", which is the opposite of what anyone typing it intends. | _(runs)_ |
 | `CLAUDE_MEM_SKIP_UPDATE` | Skip the 24h auto-update check against GitHub Releases. | _(runs)_ |
 | `CLAUDE_MEM_SKIP_SIG_VERIFY` | Skip Ed25519 signature verification of a downloaded update. **Escape hatch — leaves updates unauthenticated.** | _(verifies)_ |
 | `CLAUDE_MEM_NO_LESSON_RETRY` | `1` disables the one-shot retry that re-asks for a missing `lesson_learned`. | _(retries)_ |

package/README.zh-CN.md

@@ -54,7 +54,7 @@
 
 ## 功能特性
 
-- **自动捕获** -- 挂载到 Claude Code 生命周期（PostToolUse、PreToolUse、SessionStart、Stop、UserPromptSubmit），无需手动操作即可记录观察
+- **自动捕获** -- 挂载到 Claude Code 生命周期（`hooks/hooks.json` 里的七个事件：SessionStart、PreCompact、PreToolUse、PostToolUse、PostToolUseFailure、Stop、UserPromptSubmit），无需手动操作即可记录观察
 - **FTS5 搜索** -- 基于 BM25 排名的全文搜索，覆盖观察、会话摘要和用户提示，支持重要度加权
 - **时间线浏览** -- 基于锚点的时间上下文窗口，按时间顺序浏览观察
 - **Episode 批处理** -- 将相关文件操作分组为连贯的 episode，再进行 LLM 编码
@@ -81,16 +81,13 @@
 - **原子写入** -- 所有文件写入（episode、CLAUDE.md）使用 write-to-tmp + rename 防止崩溃时损坏
 - **健壮锁机制** -- PID 感知的锁文件，自动清理过期（>30s）或孤儿（PID 已死）锁
 - **过期会话清理** -- 活跃超过 24 小时的会话在下次启动时自动标记为 abandoned
-- **统一资源发现** -- 共享文件系统遍历层（`resource-discovery.mjs`），运行时扫描器和离线索引器共用，支持扁平目录、插件嵌套和松散 `.md` 文件
-- **领域同义词扩展** -- 注册表搜索查询自动扩展领域同义词（如 "修复" → fix, debug, bugfix, repair, error）
-- **持久化冷却机制** -- 5 分钟跨会话冷却 + 同会话去重，避免重复推荐 skill 自动加载
+- **领域同义词扩展** -- 搜索查询自动扩展领域同义词（如 "修复" → fix, debug, bugfix, repair, error）
 - **多 provider LLM 调用** -- provider 优先级 `ANTHROPIC_API_KEY`（直连 Anthropic API）→ `OPENROUTER_API_KEY`（OpenRouter，OpenAI 兼容，可用 `OPENROUTER_MODEL` 指向任意模型）→ 无 key 时回退 `claude -p` CLI
 - **Haiku 熔断器** -- 连续 3 次 LLM 失败后，禁用 Haiku 调度 5 分钟，防止级联延迟
 - **否定意图感知** -- 正确处理 "不要测试了，先修 bug" 等复杂提示，排除被否定的意图，支持中英文混合输入
 - **可配置 LLM 模型** -- 通过 `CLAUDE_MEM_MODEL` 环境变量在 Haiku（快速/低成本）和 Sonnet（深度分析）之间切换
 - **数据库自动恢复** -- 启动时检测并清理损坏的 WAL/SHM 文件；定期 WAL checkpoint 防止无限增长
 - **Schema 自动迁移** -- 每次启动运行幂等的 `ALTER TABLE` 迁移，安全地添加新列和索引，不丢失数据
-- **探索奖励** -- 注册表中的新资源在复合排名中获得公平机会；高推荐零采纳的"僵尸"资源被惩罚
 - **LLM 并发控制** -- 基于文件的信号量将后台 worker 限制为 2 个并发 LLM 调用，防止资源争用
 - **stdin 溢出保护** -- Hook 输入在 256KB 处截断，对超大工具输出使用正则挽救关键信息
 - **跨会话交接** -- 在 `/clear` 或 `/exit` 时捕获会话状态（请求、已完成工作、后续步骤、关键文件），下次会话检测到继续意图时自动注入上下文（支持显式关键词和 FTS5 术语重叠匹配）
@@ -109,7 +106,7 @@
 
 ## 环境要求
 
-- **Node.js** >= 18
+- **Node.js** >= 22（v4.0.0 起：better-sqlite3 13 要求 >=22，Node 20 已于 2026-04 EOL；`package.json` 的 `engines` 是唯一事实来源）
 - **Claude Code** CLI 已安装并配置（`claude` 命令可用）
 - **SQLite3** 支持（由 `better-sqlite3` 提供，安装时编译）
 - **平台**：Linux 或 macOS（参见[平台支持](#平台支持)）
@@ -148,13 +145,13 @@ node install.mjs install
 ### 安装过程
 
 1. **安装依赖** -- `npm install --omit=dev`（编译原生 `better-sqlite3`）
-2. **注册 MCP 服务器** -- `mem-lite` 服务器，包含 20 个工具（9 个核心通过 `tools/list` 暴露 + 11 个隐藏但可调；完整表见 Usage 段）。v2.78 前服务器名为通用的 `mem`，现已改名为 `mem-lite` 避免与用户其它 `.mcp.json` 冲突；工具名（`mem_search`/`mem_recall` 等）保持不变。
+2. **注册 MCP 服务器** -- `mem-lite` 服务器，包含 18 个工具（9 个核心通过 `tools/list` 暴露 + 9 个隐藏但可调；完整表见 Usage 段）。v2.78 前服务器名为通用的 `mem`，现已改名为 `mem-lite` 避免与用户其它 `.mcp.json` 冲突；工具名（`mem_search`/`mem_recall` 等）保持不变。
 
 > **自动 adopt 会写进你的项目，且每次 SessionStart 都跑（v3.13+）。** 插件向**项目自己的 `<cwd>/CLAUDE.md`**（通常是会进 git 的文件）写入一个 slug 限定的**托管块**，外加 `<cwd>/.claude/plugin_claude_mem_lite.md` 详情文件。该块是一条提升 Claude 主动调用 `mem_recall` / `mem_save` 的 system-authority 指针；块以外的内容逐字保留，也能与其它插件的块共存于同一文件。这是**每次** SessionStart 都做的幂等同步，不只是第一次——块被删掉会重新写回，出货模板变了会刷新。**任何安装路径都生效**（npm、npx、`/plugin`、手动），**无需再手动跑 `/adopt`**。
 >
 > 关闭方式：项目级 `claude-mem-lite adopt --disable`（重新启用用 `--enable`）；全局 `export MEM_NO_AUTO_ADOPT=1`；只冻结模板刷新用 `CLAUDE_MEM_NO_TEMPLATE_REFRESH=1`。`claude-mem-lite unadopt` 可移除托管块与详情文件。手动 `/adopt` 仍保留用于编辑后重写或 `--all` 批量场景。
-3. **配置钩子** -- `PostToolUse`、`PreToolUse`、`SessionStart`、`Stop`、`UserPromptSubmit` 生命周期钩子
-4. **创建数据目录** -- `~/.claude-mem-lite/`（隐藏目录），存放数据库、运行时和托管资源文件
+3. **配置钩子** -- 全部七个生命周期事件：`SessionStart`、`PreCompact`、`PreToolUse`、`PostToolUse`、`PostToolUseFailure`、`Stop`、`UserPromptSubmit`
+4. **创建数据目录** -- `~/.claude-mem-lite/`（隐藏目录），存放数据库与运行时文件
 5. **自动迁移** -- 自动检测 `~/.claude-mem/`（原版 claude-mem）或 `~/claude-mem-lite/`（v0.5 前的非隐藏目录），将数据库和运行时文件迁移到 `~/.claude-mem-lite/`，原目录保持不变
 6. **初始化数据库** -- SQLite WAL 模式，FTS5 索引在服务器首次启动时创建
 
@@ -192,8 +189,6 @@ rm -rf ~/claude-mem-lite/   # v0.5 前的非隐藏目录（如未自动迁移）
     ep-flush-*.json      # 已刷新的 episode，等待处理
     reads-<project>.txt  # Read 文件路径（刷新时收集）
   managed/
-    skills/              # 独立 skill：{name}/SKILL.md
-    agents/              # Agent 插件：{group}/agents/{name}.md + skills/*/SKILL.md
     repos/               # 浅克隆的源代码仓库
 ```
 
@@ -201,12 +196,16 @@ rm -rf ~/claude-mem-lite/   # v0.5 前的非隐藏目录（如未自动迁移）
 
 ### MCP 工具
 
-v2.34.0 起服务端注册 17 个工具，但 `tools/list` 只暴露 6 个 **核心** 工具；其余
-11 个 **隐藏** 工具仍然注册在 MCP 层（按名 `tools/call` 仍命中），只是不会出现
-在列表响应里，以避免 Claude Code 会话启动时加载 11 份额外的工具 schema。隐藏
-工具走下面表格的 CLI 入口。
+v2.34.0 起服务端只把一部分工具暴露给 `tools/list`。当前是 18 个工具，其中 9 个
+**核心** 工具出现在列表里，另外 9 个 **隐藏** 工具仍然注册在 MCP 层（按名
+`tools/call` 仍命中），只是不出现在列表响应里，以避免 Claude Code 会话启动时
+多加载 9 份工具 schema。隐藏工具走下面表格的 CLI 入口。
 
-**核心（6 个，暴露给 Claude Code）**
+（这里长期写着 17 / 6 / 11，而英文 README 写着 20 / 9 / 11 —— 两个都不对。
+`tool-schemas.mjs` 是唯一事实来源，`tests/tool-count-docs.test.mjs` 现在把两份
+README 和 `docs/ARCHITECTURE.md` 都钉在它上面。）
+
+**核心（9 个，暴露给 Claude Code）**
 
 | 工具 | 描述 |
 |------|------|
@@ -216,8 +215,11 @@ v2.34.0 起服务端注册 17 个工具，但 `tools/list` 只暴露 6 个 **核
 | `mem_timeline` | 围绕锚点按时间顺序浏览观察。 |
 | `mem_get` | 获取指定观察 ID 的完整详情（包含重要度和关联 ID）。 |
 | `mem_save` | 手动保存记忆/观察。 |
+| `mem_defer` | 记录一条跨会话待办（deferred work）。 |
+| `mem_defer_list` | 列出当前项目未关闭的待办。 |
+| `mem_defer_drop` | 带理由地关闭一条待办。 |
 
-**隐藏但可按名调用（11 个，走 CLI）**
+**隐藏但可按名调用（9 个，走 CLI）**
 
 | 工具 | 对应 CLI | 说明 |
 |------|----------|------|
@@ -368,9 +370,6 @@ PostToolUse（每次工具执行）
 UserPromptSubmit（两个并行路径）
   -> [user-prompt-search.js] 通过 FTS5 + 活跃文件上下文自动搜索记忆
   -> [user-prompt-search.js] 注入相关历史观察（按时效和重要性加权）
-  -> [user-prompt-search.js] L1 Skill 自动加载：匹配 prompt 中的 managed skill 名
-     -> 加载内容 + 便携 ~ 路径 + Read() 调用指引
-     -> source="managed-skill|managed-agent", path="~/.claude-mem-lite/managed/..."
   -> [hook.mjs] 捕获用户提示文本到 user_prompts 表
   -> [hook.mjs] 递增会话提示计数器
   -> [hook.mjs] 交接：检测继续意图 → 注入上一次会话上下文
@@ -430,7 +429,7 @@ npx claude-mem-lite doctor            # 诊断问题
 
 ### 故障恢复（安装卡死 / hook 报错）
 
-如果你看到 PreToolUse:Read/Edit/Skill hook 报 `ERR_MODULE_NOT_FOUND`，或者 `claude-mem-lite` 命令本身因为 import 错误崩溃，多半是被部分自动更新坑了——更新器复制了新脚本但漏了配套的 `lib/*` 文件，hook 链就此断掉（连下一次本可自愈的自动更新也跑不了）。
+如果你看到 PreToolUse:Read/Edit hook 报 `ERR_MODULE_NOT_FOUND`，或者 `claude-mem-lite` 命令本身因为 import 错误崩溃，多半是被部分自动更新坑了——更新器复制了新脚本但漏了配套的 `lib/*` 文件，hook 链就此断掉（连下一次本可自愈的自动更新也跑不了）。
 
 **v2.84.0+** 提供 `repair` 子命令，从 GitHub 最新 release 重新同步：
 
@@ -516,12 +515,11 @@ claude-mem-lite/
   scripts/
     setup.sh           # Setup 钩子：npm install + 迁移（隐藏目录 + 旧目录）
     post-tool-use.sh   # Bash 预过滤器：~5ms 跳过噪声，追踪 Read 路径
-    user-prompt-search.js # UserPromptSubmit 钩子：自动搜索记忆 + L1 skill 自动加载
-    pre-skill-bridge.js  # PreToolUse 钩子：L2 managed skill 桥接
+    user-prompt-search.js # UserPromptSubmit 钩子：用户提问时自动搜索记忆
     pre-tool-recall.js   # PreToolUse 钩子：Edit/Write 前文件教训回忆
+    post-tool-recall.js  # PostToolUse 钩子：工具失败后的错误召回
+    pre-agent-inject.sh  # PreToolUse 钩子：为子代理注入上下文
     prompt-search-utils.mjs # 共享逻辑：跳过模式、意图检测、名称匹配
-    convert-commands.mjs # 将 command .md 转换为托管插件中的 SKILL.md
-    index-managed.mjs  # 托管资源离线索引器
   # 测试和基准（仅开发）
   tests/               # 单元、属性、集成、契约、E2E、管线测试
   benchmark/           # BM25 搜索质量基准 + CI 门控
@@ -564,7 +562,7 @@ npm run benchmark:gate    # CI 门控：指标回退超过 5% 容差时失败
 
 | 变量 | 说明 | 默认值 |
 |------|------|--------|
-| `CLAUDE_MEM_DIR` | 自定义数据目录。所有数据库、运行时文件和托管资源均存储在此。 | `~/.claude-mem-lite/` |
+| `CLAUDE_MEM_DIR` | 自定义数据目录。所有数据库与运行时文件均存储在此。 | `~/.claude-mem-lite/` |
 | `CLAUDE_MEM_MODEL` | 后台 LLM 调用模型（Episode 提取、会话总结、调度）。可选 `haiku` 或 `sonnet`。 | `haiku` |
 | `ANTHROPIC_API_KEY` | Anthropic API key。设置后所有后台 LLM 调用直连 Anthropic Messages API（带 prompt caching），优先级最高。 | _(未设 → CLI)_ |
 | `OPENROUTER_API_KEY` | OpenRouter API key（OpenAI 兼容）。当**未设** `ANTHROPIC_API_KEY` 时用于后台 LLM 调用；两者都未设则回退到 `claude -p` CLI。 | _(未设)_ |

package/commands/mem.md

@@ -30,9 +30,9 @@ When the user invokes `/mem`, parse their intent:
 - `/mem save <text>` → call `mem_save` MCP tool with the text as content
 - `/mem stats` → run `node ${CLAUDE_PLUGIN_ROOT}/cli.mjs stats` via Bash
 - `/mem get <ids>` → run `node ${CLAUDE_PLUGIN_ROOT}/cli.mjs get <ids>` via Bash
-- `/mem cleanup` → run `mem_maintain(action="scan")`, report pending purge count and stale items to user, ask for confirmation, then run `mem_maintain(action="execute", operations=["purge_stale"])` if confirmed
-- `/mem cleanup Nd` (e.g. `60d`) → same as above but use `retain_days=N` to only purge items older than N days
-- `/mem cleanup keep Nd` (e.g. `keep 14d`) → same as above with `retain_days=N`
+- `/mem cleanup` → run `mem_maintain(action="scan")`, report pending purge count and stale items to user, ask for confirmation, then run `mem_maintain(action="execute", operations=["purge_stale"], confirm=true)` if confirmed. **`confirm=true` is required and is not optional politeness:** without it the call returns a dry-run PREVIEW and deletes nothing, while still succeeding — so you would report a cleanup that never happened.
+- `/mem cleanup Nd` (e.g. `60d`) → same as above but add `retain_days=N` to only purge items older than N days. **`retain_days` must be between 7 and 365**; anything outside that range is rejected by the schema, so `/mem cleanup 3d` cannot be honoured — say so rather than silently substituting the default.
+- `/mem cleanup keep Nd` (e.g. `keep 14d`) → same as above with `retain_days=N`, same 7–365 range.
 - `/mem <query>` (no subcommand) → treat as search, run `node ${CLAUDE_PLUGIN_ROOT}/cli.mjs search <query>` via Bash
 
 Use Bash commands first. For detailed data, use `node ${CLAUDE_PLUGIN_ROOT}/cli.mjs get <id>` via Bash.

package/commands/update.md

@@ -22,7 +22,7 @@ When the user invokes `/mem:update`, perform the following maintenance cycle:
 3. Call `mem_maintain(action="execute", operations=["cleanup","decay","boost"])` to apply safe automatic changes
 4. If duplicates were found in scan, review them and call `mem_maintain(action="execute", operations=["dedup"], merge_ids=[[keepId, removeId1, ...], ...])` — keep the more important/recent observation in each pair
 5. Run `mem_compress(preview=false)` for old low-value observations
-6. **If pending purge items > 0**: Report the count to the user and ask for confirmation. If confirmed, call `mem_maintain(action="execute", operations=["purge_stale"])`. User may optionally specify `retain_days` (default 30) to control how many days of data to keep. Do NOT purge without explicit user confirmation.
+6. **If pending purge items > 0**: Report the count to the user and ask for confirmation. If confirmed, call `mem_maintain(action="execute", operations=["purge_stale"], confirm=true)`. **`confirm=true` is required:** without it the call returns a dry-run preview and deletes nothing while still succeeding, so you would report a purge that never happened. User may optionally specify `retain_days` (default 30, allowed range 7–365) to control how many days of data to keep. Do NOT purge without explicit user confirmation.
 
 ### Phase 2: Summary
 

package/haiku-client.mjs

@@ -6,13 +6,40 @@
 // overridable via OPENROUTER_MODEL
 
 import { execFileSync, spawn } from 'child_process';
+import { mkdirSync } from 'fs';
 import { readFileSync } from 'fs';
 import { join } from 'path';
 import { randomUUID } from 'crypto';
 import { debugLog, debugCatch, parseJsonFromLLM } from './utils.mjs';
 import { DB_DIR } from './schema.mjs';
+import { resolveRuntimeDir } from './lib/resolve-data-dir.mjs';
 import { httpConnectProxyFor, postViaConnectProxy } from './lib/proxy-fetch.mjs';
 
+/**
+ * cwd for every `claude -p` spawn. R10 P2-13.
+ *
+ * This was `/tmp`, which is world-writable. Claude Code loads a project-level CLAUDE.md
+ * and .claude/settings.json from its cwd, so on a shared host ANY local account could
+ * create /tmp/CLAUDE.md and inject instructions into every episode summary, session
+ * summary and optimize call this process makes — and the CLI leg is the fallback every
+ * keyed provider failure lands on, so it is not an exotic path.
+ *
+ * The original reason for /tmp was ghost sessions in the user's /resume list, and that is
+ * already solved by --no-session-persistence on the same spawns. A private directory under
+ * the runtime dir (whose parent is 0700) keeps that property and removes the injection
+ * surface. Created lazily; if creation fails we still do not fall back to /tmp — an
+ * unwritable cwd fails the spawn loudly, which is the better failure.
+ */
+function cliSpawnCwd() {
+  const dir = join(resolveRuntimeDir(DB_DIR), 'cli-cwd');
+  try {
+    mkdirSync(dir, { recursive: true, mode: 0o700 });
+  } catch {
+    /* already there, or unwritable — the spawn reports it */
+  }
+  return dir;
+}
+
 // ─── Model Resolution ────────────────────────────────────────────────────────
 
 // CLI name → API model ID mapping
@@ -237,7 +264,10 @@ export async function callHaikuJSON(prompt, opts) {
 
 /**
  * Non-blocking sibling of callHaikuJSON for callers reachable from an MCP request
- * handler (registry enrichment: mem_registry `enrich` / `import_url`). Same
+ * handler. R10 P3-28: this used to name `mem_registry enrich / import_url` as the caller,
+ * a tool removed in v5.0.0 — read as a live example, it sent readers looking for a handler
+ * that does not exist. The REASON is what still applies to whatever calls it next: an MCP
+ * request handler must not block the server event loop. Same
  * provider priority; the CLI leg — primary AND post-provider-failure fallback —
  * is the async spawn, so a keyed-provider outage cannot freeze the server event
  * loop for BG_LLM_TIMEOUT_MS (D#138 MEDIUM-3).
@@ -245,8 +275,9 @@ export async function callHaikuJSON(prompt, opts) {
  * `resolveModel().cli`, NOT the literal 'haiku': despite the name, callHaikuJSON
  * reaches the model through resolveModel() on ALL three legs (callHaikuAPI,
  * callOpenRouterAPI, callHaikuCLI), so it honours the documented CLAUDE_MEM_MODEL
- * knob. Pinning 'haiku' here would silently downgrade registry enrichment for
- * every user who set CLAUDE_MEM_MODEL=sonnet — pre-tag review finding, v3.68.0.
+ * knob. Pinning 'haiku' here would silently downgrade any caller's model for every user
+ * who set CLAUDE_MEM_MODEL=sonnet — pre-tag review finding, v3.68.0, when the caller in
+ * question was registry enrichment.
  *
  * Defaults also mirror callHaiku (10s / 500 tokens), not callModelJSONAsync's
  * 15s / 1000: a caller that omits opts must get the sync twin's budget.
@@ -565,7 +596,7 @@ export function execClaudeCliSync(modelName, { input, timeout }) {
     encoding: 'utf8',
     env: { ...process.env, CLAUDE_MEM_HOOK_RUNNING: '1', DISABLE_CLAUDEMD_HOOKS: '1' },
     stdio: ['pipe', 'pipe', 'pipe'],
-    cwd: '/tmp', // Prevent ghost sessions in the user's /resume list
+    cwd: cliSpawnCwd(), // private dir, not /tmp — see cliSpawnCwd (R10 P2-13)
   };
   const args = claudeArgs(modelName);
   const started = Date.now();
@@ -650,7 +681,7 @@ export async function callModelCLIAsync(prompt, model, { timeout }) {
         // Same headless-tax flags + flag-compat retry as callModelCLI (rationale there).
         child = spawn(getClaudePath(), args, {
           env: { ...process.env, CLAUDE_MEM_HOOK_RUNNING: '1', DISABLE_CLAUDEMD_HOOKS: '1' },
-          cwd: '/tmp',
+          cwd: cliSpawnCwd(), // private dir, not /tmp — see cliSpawnCwd (R10 P2-13)
           stdio: ['pipe', 'pipe', 'pipe'],
         });
       } catch (e) {

package/hash-utils.mjs

@@ -53,6 +53,27 @@ export function computeMinHash(text, numHashes = 64) {
     .replace(/[^a-z0-9\s]/g, ' ')
     .split(/\s+/)
     .filter((t) => t.length > 2);
+  // R10 P3-10: CJK fallback, and ONLY as a fallback. The tokenizer above deletes every
+  // non-ASCII character, so a title or narrative written entirely in Chinese or Japanese
+  // produced zero tokens and this returned null — and a null signature makes the row
+  // invisible to the MinHash prefilter that findDuplicates and selectFuzzyDedupeIds run,
+  // so CJK-only rows could never be deduplicated against anything.
+  //
+  // Confined to rows that get NOTHING from the ASCII path, which is the whole point.
+  // Signatures are STORED, and estimateJaccardFromMinHash compares a stored signature
+  // against a freshly computed one; widening the tokenization for text that already
+  // signs would make the entire existing corpus incomparable with everything written
+  // afterwards — dedup degrading everywhere, silently, until a full rebuild. Every text
+  // that signs today still produces the same bytes.
+  //
+  // Character bigrams rather than nlp.mjs's dictionary segmentation: this module is
+  // dependency-free on purpose, and bigrams are the standard CJK shingle for exactly this
+  // job — no vocabulary to maintain and no dependence on a dictionary that lags usage.
+  if (tokens.length === 0) {
+    for (const run of text.match(/[\u4e00-\u9fff\u3400-\u4dbf\u3040-\u30ff]{2,}/g) || []) {
+      for (let i = 0; i + 2 <= run.length; i++) tokens.push(run.slice(i, i + 2));
+    }
+  }
   // Require at least 3 tokens for meaningful signature (avoids high collision on short texts)
   if (tokens.length < 3) return null;
 

package/hook-episode.mjs

@@ -248,7 +248,13 @@ export function mergePendingEntries(episode) {
   let files;
   try {
     files = readdirSync(RUNTIME_DIR)
-      .filter((f) => f.startsWith('pending-'))
+      // R10 P3-4: `.json.tmp` is the WRITER's in-flight temp, not a pending entry.
+      // writePendingEntry writes `pending-<ts>-<rand>.json.tmp` then renames it into place;
+      // the bare `pending-` prefix matched the temp too, this loop's JSON.parse failed on a
+      // partial file, the catch deleted it as corrupt, and the writer's rename then failed
+      // ENOENT inside its own swallowed catch. One tool entry lost, silently, every time
+      // the two raced.
+      .filter((f) => f.startsWith('pending-') && f.endsWith('.json'))
       .sort();
   } catch {
     return;