claude-mem-lite 6.3.0 → 6.5.0

package/.claude-plugin/marketplace.json

@@ -4,14 +4,14 @@
     "name": "sdsrss"
   },
   "metadata": {
-    "description": "Plugins by sdsrss",
-    "homepage": "https://github.com/sdsrss/claude-mem-lite"
+    "description": "Plugins by sdsrss"
   },
   "plugins": [
     {
       "name": "claude-mem-lite",
-      "version": "6.3.0",
+      "version": "6.5.0",
       "source": "./",
+      "homepage": "https://github.com/sdsrss/claude-mem-lite",
       "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": "6.3.0",
+  "version": "6.5.0",
   "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

@@ -598,7 +598,12 @@ Your data directory (`~/.claude-mem-lite/`) is untouched by install/rollback; sc
 
 ### doctor
 
-Checks Node.js version, dependencies, server/hook files, database integrity, FTS5 indexes, and stale processes.
+Checks Node.js version, dependencies, server/hook files, database integrity, FTS5 indexes,
+stale processes, MCP registration, and whether the marketplace clone can still be updated.
+
+It runs `claude mcp list` to answer the registration question, and that **health-checks every
+MCP server you have configured** — i.e. briefly launches each one, including remote endpoints.
+`status` deliberately does not: on a plugin install it answers from the manifest instead.
 
 ### status
 
@@ -617,9 +622,11 @@ claude-mem-lite repair
 **If `repair` itself fails** (the bin is older than v2.84.0, or the bin is also broken), run this one-liner — it pulls a fresh tarball into a temp dir and runs *that* tarball's `install.mjs`, bypassing every file on your disk:
 
 ```bash
-T=$(mktemp -d) && curl -sL https://api.github.com/repos/sdsrss/claude-mem-lite/tarball | tar xz -C "$T" --strip-components=1 && node "$T/install.mjs" install
+T=$(mktemp -d) && U=$(curl -sL https://api.github.com/repos/sdsrss/claude-mem-lite/releases/latest | grep -o '"tarball_url"[^,]*' | cut -d'"' -f4) && curl -sL "$U" | tar xz -C "$T" --strip-components=1 && node "$T/install.mjs" install
 ```
 
+It resolves the latest **release** tag first. A shell one-liner cannot verify the release signature the way `repair` does, so running it is a trust decision you are making explicitly — that is why it is the last resort and not the first suggestion.
+
 After it finishes, `~/.claude-mem-lite/` is back in sync with the latest release and `claude-mem-lite repair` is available for next time.
 
 ## Uninstall
@@ -643,6 +650,19 @@ Data in `~/.claude-mem-lite/` is preserved by default. Delete manually if needed
 rm -rf ~/.claude-mem-lite/
 ```
 
+**`/plugin uninstall` does not delete the plugin cache.** Claude Code materializes each
+version under `~/.claude/plugins/cache/`, with its own `node_modules`. While the plugin is
+installed these get pruned to the newest three (SessionStart does it, and so does the update
+path), so the directory is bounded — measured at 241 MB — not unbounded. But `/plugin
+uninstall` removes the manifest and there is no uninstall hook a plugin can attach to, so the
+hooks stop firing and whatever is left is never reclaimed. `claude-mem-lite uninstall` does
+reclaim it, but after `/plugin uninstall` that command may no longer be on your PATH. Either
+run it **first**, or delete the directory yourself:
+
+```bash
+rm -rf ~/.claude/plugins/cache/sdsrss/claude-mem-lite
+```
+
 ### Mixed-install residue (read this if you've used multiple install methods)
 
 `/plugin uninstall` only removes the plugin manifest — it **does not touch `~/.claude/settings.json`**. If you've ever run `claude-mem-lite install` (npx or git-clone path), hook entries pointing at `~/.claude-mem-lite/hook.mjs` were written into your user-global settings, and they keep firing after `/plugin uninstall`. If `~/.claude-mem-lite/hook.mjs` still exists they double-fire alongside the plugin; if you also ran `rm -rf ~/.claude-mem-lite/` they error every session.

package/README.zh-CN.md

@@ -478,7 +478,11 @@ npx claude-mem-lite doctor            # 诊断问题
 
 ### doctor
 
-检查 Node.js 版本、依赖、服务器/钩子文件、数据库完整性、FTS5 索引和残留进程。
+检查 Node.js 版本、依赖、服务器/钩子文件、数据库完整性、FTS5 索引、残留进程、MCP 注册状态，
+以及 marketplace clone 是否还能被更新。
+
+它用 `claude mcp list` 回答注册状态那一问，而这会**对你配置的每个 MCP server 做健康检查**——
+也就是逐个短暂启动，包括远程 endpoint。`status` 刻意不这么做：插件形态下它从 manifest 回答。
 
 ### status
 
@@ -497,9 +501,11 @@ claude-mem-lite repair
 **如果 `repair` 自己也跑不起来**（bin 比 v2.84.0 旧，或 bin 也坏了），用这条单行命令——它把最新 tarball 拉到临时目录、跑 *那份* tarball 里的 `install.mjs`，完全不依赖你磁盘上的任何文件：
 
 ```bash
-T=$(mktemp -d) && curl -sL https://api.github.com/repos/sdsrss/claude-mem-lite/tarball | tar xz -C "$T" --strip-components=1 && node "$T/install.mjs" install
+T=$(mktemp -d) && U=$(curl -sL https://api.github.com/repos/sdsrss/claude-mem-lite/releases/latest | grep -o '"tarball_url"[^,]*' | cut -d'"' -f4) && curl -sL "$U" | tar xz -C "$T" --strip-components=1 && node "$T/install.mjs" install
 ```
 
+它会先解析出最新 **release** 的 tag。shell 单行命令无法像 `repair` 那样校验 release 签名，所以跑它等于你自己做了一次信任决定——这也是它排在最后、而不是被优先推荐的原因。
+
 跑完之后，`~/.claude-mem-lite/` 就和最新 release 对齐，`claude-mem-lite repair` 下次再遇到类似问题也能直接用了。
 
 ## 卸载
@@ -523,6 +529,17 @@ npx claude-mem-lite uninstall --purge
 rm -rf ~/.claude-mem-lite/
 ```
 
+**`/plugin uninstall` 不会删 plugin cache。** Claude Code 会把每个版本展开到
+`~/.claude/plugins/cache/` 下，各带一份 `node_modules`。插件还装着的时候，这些会被裁剪到最新
+三个版本（SessionStart 会做，更新路径也会做），所以这个目录是有上限的——实测 241 MB——而不是无限
+增长。但 `/plugin uninstall` 只移除 manifest，插件又没有可挂的卸载生命周期钩子，于是钩子停止触发，
+剩下的东西再也没人回收。`claude-mem-lite uninstall` 能回收它，但 `/plugin uninstall` 之后这个命令
+可能已经不在 PATH 上了。所以要么**先**跑它，要么自己删：
+
+```bash
+rm -rf ~/.claude/plugins/cache/sdsrss/claude-mem-lite
+```
+
 ### 混装残留（用过多种安装方式的话务必看一下）
 
 `/plugin uninstall` 只删 plugin manifest，**不会动 `~/.claude/settings.json`**。如果你曾经跑过 `claude-mem-lite install`（npx 或 git-clone 路径），指向 `~/.claude-mem-lite/hook.mjs` 的 hook 条目就被写进了你的 user-global settings；`/plugin uninstall` 之后它们还在每会话触发。如果 `~/.claude-mem-lite/hook.mjs` 还在 → 与 plugin 双触发；如果你又 `rm -rf ~/.claude-mem-lite/` → 每次会话报错。

package/hook-context.mjs

@@ -585,7 +585,12 @@ export function buildSessionContextLines(
   } else if (!latestSummary && !effectiveQuiet()) {
     // Fallback: no summary AND no key observations — show recent activity.
     // Skipped under QUIET_HOOKS since the Recent table already carries titles.
-    const recentObs = (observations.length >= 3 ? observations : fallbackObs).slice(0, 3);
+    // Slice FIRST, then sort: the slice is the selection (top 3 by value density) and must
+    // stay that way; only the order they are printed in is corrected, same as the Recent
+    // table below. Sorting before the slice would silently change WHICH three are injected.
+    const recentObs = (observations.length >= 3 ? observations : fallbackObs)
+      .slice(0, 3)
+      .sort((a, b) => Date.parse(b.created_at) - Date.parse(a.created_at) || b.id - a.id);
     if (recentObs.length > 0) {
       summaryLines.push('### Recent Activity');
       for (const o of recentObs) {
@@ -695,8 +700,18 @@ export function buildSessionContextLines(
   }
 
   // 6. Recent observations table
+  //
+  // SELECTION order (greedy knapsack, value density) is not DISPLAY order. This block used
+  // to render the picks in the order the knapsack happened to take them, under a heading
+  // that says "Recent" next to a Time column — so row 1 was not the newest row, and both a
+  // human and the model read it as if it were. Sorting here is display-only: `obsToShow` is
+  // already chosen, so the token budget and the row set are untouched (pinned by a case in
+  // tests/hook-context.test.mjs). Tiebroken on id for the same reason D#9 gives — an
+  // untiebroken tie flips direction, and two saves in one millisecond are common.
   const obsLines = [];
-  const obsToShow = observations.length >= 3 ? observations : fallbackObs;
+  const obsToShow = [...(observations.length >= 3 ? observations : fallbackObs)].sort(
+    (a, b) => Date.parse(b.created_at) - Date.parse(a.created_at) || b.id - a.id,
+  );
   if (obsToShow.length > 0) {
     const today = now.toISOString().slice(0, 10);
     obsLines.push(`### Recent (${today})`);

package/hook-shared.mjs

@@ -28,6 +28,8 @@ import {
   shouldRecordSkew,
   SKEW_MARKER_PREFIX,
 } from './lib/schema-skew.mjs';
+import { isDbUnusableError, DB_UNUSABLE_MARKER_PREFIX } from './lib/db-unusable.mjs';
+import { shouldRecordOnce } from './lib/record-once.mjs';
 // Audit 2026-09-05 P1-2 (carried from 2026-09-02 P2-9): `callLLM`, the quiet/adoption
 // predicates and the handoff constants moved into `lib/` because two lib modules
 // imported them from here and dragged this file's whole import graph — haiku-client,
@@ -244,6 +246,9 @@ export const GC_PROJECT_MARKER_PREFIXES = Object.freeze([
   // v5.0.0; a prefix for files nothing writes any more is dead weight in a hot-path loop.
   'last-mark-compressible-', // per-project auto-compress 24h gate
   SKEW_MARKER_PREFIX, // per-project schema-skew log dedup; regenerated on the next skewed open
+  // Same shape, same reason, and it was missed here first time round — the note above is
+  // about exactly this defect, two entries up.
+  DB_UNUSABLE_MARKER_PREFIX, // per-project unopenable-DB log dedup; regenerated on the next failing open
 ]);
 
 // Records of a completed side effect — never age out. `ep-`/`ep-flush-`/
@@ -405,12 +410,28 @@ export function lastSchemaSkew() {
   return lastSkew;
 }
 
+// Same idea, other unhealable family: the file exists and SQLite will not open it. Held as a
+// boolean rather than the error, because the only thing SessionStart needs is "which notice",
+// and keeping an Error alive here would tempt a caller into rendering a stack trace at a user.
+let lastUnusable = false;
+
+/**
+ * True when the most recent openDb() returned null because the database file is not a usable
+ * database. Cleared by any successful open, so a repair mid-session stops the notice.
+ *
+ * @returns {boolean}
+ */
+export function lastDbUnusable() {
+  return lastUnusable;
+}
+
 export function openDb() {
   try {
     // WAL-corruption self-heal (was server.mjs-only): without it, hooks stayed
     // silently dead (null DB) on a corrupt WAL until the next MCP server start.
     const db = ensureDbWithWalRecovery();
     lastSkew = null;
+    lastUnusable = false;
     return db;
   } catch (e) {
     // Forward-incompat is its own family: it cannot be healed by anything this process can
@@ -423,10 +444,11 @@ export function openDb() {
     //
     // shouldRecordSkew is TOTAL by contract. Nothing in this catch may throw: the first cut
     // called getSessionId() here, which MINTS and writes a session id, so an unwritable
-    // runtime dir turned openDb() itself into a thrower. All 13 call sites are written to
+    // runtime dir turned openDb() itself into a thrower. All 12 openDb() call sites in hook.mjs are written to
     // no-op on null and none of them expects an exception.
     if (isSchemaSkewError(e)) {
       lastSkew = schemaSkewFromError(e) || { dbVersion: null, binaryVersion: null };
+      lastUnusable = false;
       // Guarded even though inferProject() reads env and cwd: "the only statement in this
       // catch cannot throw" was true of the original one-line body and stopped being true
       // the moment anything was added. An unscoped marker is a worse dedup, not a crash.
@@ -441,8 +463,31 @@ export function openDb() {
       }
       return null;
     }
-    // Still null, still no throw — a hook must never crash the host session, and all
-    // eight call sites in hook.mjs are written to no-op on null. But "returned null"
+    // The OTHER unhealable family, and it was the silent one. A file that is not a database
+    // repeats on every fire exactly like a skew does, and until now took the generic branch
+    // below: one full stack trace per SessionStart fire (measured: 20 fires → 20 records, ~860 B
+    // each), no dedup, and no user-visible word anywhere in the session. Same treatment as skew
+    // — record once per project per hour, and hand SessionStart a flag to speak with.
+    //
+    // Nothing in this branch may throw: `isDbUnusableError` is a regex over a string and
+    // `shouldRecordOnce` is total by contract, which is exactly the property the first cut of
+    // the skew dedup lost by calling a function that WRITES.
+    if (isDbUnusableError(e)) {
+      lastUnusable = true;
+      lastSkew = null; // the two flags are a set: whichever family fired last is the true one
+      let project = '';
+      try {
+        project = inferProject();
+      } catch {
+        /* total: the marker degrades to one shared file */
+      }
+      if (shouldRecordOnce(RUNTIME_DIR, DB_UNUSABLE_MARKER_PREFIX, project, 'unusable')) {
+        recordHookError('hook-shared:db-open', e, RUNTIME_DIR);
+      }
+      return null;
+    }
+    // Still null, still no throw — a hook must never crash the host session, and all 12
+    // openDb() call sites in hook.mjs are written to no-op on null. But "returned null"
     // used to be the ONLY trace: nothing reached runtime/hook-errors/, so `stats`
     // reported 0 and doctor printed "no recent silent hook breakage" while every
     // capture path was dead (audit B1, 2026-08-14 — the same blindness that hid the

package/hook-update.mjs

@@ -20,7 +20,12 @@ import {
 import { join, dirname, resolve } from 'node:path';
 import { pathToFileURL } from 'node:url';
 import { tmpdir, homedir } from 'node:os';
-import { DB_DIR, CODE_DIR } from './schema.mjs';
+// lib/data-paths.mjs, NOT schema.mjs: this module is what install.mjs::repair() imports to
+// reach the signature-verified release path, and schema.mjs statically imports
+// better-sqlite3 — so importing two path constants from there made the verified repair
+// unreachable on exactly the broken-install state it exists to repair (2026-09-08).
+// tests/repair-path-no-native-dep.test.mjs fails on any package edge reachable from here.
+import { DB_DIR, CODE_DIR } from './lib/data-paths.mjs';
 import { debugCatch, debugLog } from './utils.mjs';
 import { NATIVE_BINDING_SOURCE_BUILD_CMD } from './lib/binding-probe.mjs';
 // Local manifest is fallback only — the active manifest is loaded from the

package/hook.mjs

@@ -62,7 +62,7 @@ import {
 } from './hook-episode.mjs';
 // CODE_DIR, not DB_DIR: the schema-skew notice asks which CODE homes exist, and those are
 // always homedir-rooted even when CLAUDE_MEM_DIR relocates the data.
-import { DB_DIR, CODE_DIR } from './schema.mjs';
+import { DB_DIR, DB_PATH, CODE_DIR } from './schema.mjs';
 import { cleanupClaudeMdLegacyBlock, buildSessionContextLines } from './hook-context.mjs';
 import { entry as preCompactEntry } from './hook-precompact.mjs';
 import {
@@ -85,6 +85,7 @@ import {
   sweepOrphanEpisodeFiles,
   sweepStaleProjectMarkers,
   lastSchemaSkew,
+  lastDbUnusable,
 } from './hook-shared.mjs';
 import { handleLLMEpisode, handleLLMSummary, saveEpisodeImmediate } from './hook-llm.mjs';
 import { readFastSummarySource, insertFastSummary, FAST_SUMMARY_LIMITS } from './lib/fast-summary.mjs';
@@ -2354,6 +2355,33 @@ async function emitSchemaSkewNotice() {
   }
 }
 
+/**
+ * The corruption twin of emitSchemaSkewNotice. Same channels, same reason they are BOTH used:
+ * queueHookContext reaches the model, queueHookSystemMessage reaches the human, and a notice
+ * whose whole job is handing the user a command must not depend on the assistant volunteering
+ * it (lib/hook-stdout.mjs names v3.70.0 for exactly that mistake).
+ *
+ * Dynamically imported like its twin: a cold path must not cost the healthy SessionStart a
+ * directory scan for backup snapshots.
+ */
+async function emitDbUnusableNotice() {
+  try {
+    if (!lastDbUnusable()) return;
+    const mod = await import('./lib/db-unusable.mjs');
+    // TWO DIFFERENT STRINGS, unlike the skew twin, and the asymmetry is the point: the
+    // human channel carries the repair command, the model channel carries only the fact.
+    // See formatDbUnusableModelNotice — this family's remedy overwrites the database.
+    queueHookSystemMessage(
+      mod.formatDbUnusableNotice({ dbPath: DB_PATH, remedy: mod.dbUnusableRemedy(DB_PATH) }),
+    );
+    queueHookContext('SessionStart', mod.formatDbUnusableModelNotice());
+  } catch (e) {
+    // A hook must never crash the host session, and a notice that cannot render is better
+    // handled by staying quiet than by taking SessionStart down with it.
+    debugCatch(e, 'session-start-db-unusable');
+  }
+}
+
 async function handleSessionStart() {
   // GC stale per-session cooldown files. Cheap (<5ms typical) and idempotent;
   // moved here from pre-tool-recall.js's hot path.
@@ -2542,7 +2570,13 @@ async function handleSessionStart() {
     // only other signal it produces is a `-32000 Connection closed` from the MCP host, which
     // names nothing. Measured 2026-09-08: a whole day of it, >=648 log lines, zero words to
     // the user. This is the surface the user actually reads.
+    //
+    // A database file that is not a database is the same shape and got the same treatment:
+    // unhealable without the user, disables every path, and every OTHER surface (CLI exit 1,
+    // `status`, `doctor` with an exact repair command) already reports it — leaving the
+    // in-session one as the only silent one.
     await emitSchemaSkewNotice();
+    await emitDbUnusableNotice();
     return;
   }