claude-mem-lite 2.34.3 → 2.34.5

package/.claude-plugin/marketplace.json

@@ -10,7 +10,7 @@
   "plugins": [
     {
       "name": "claude-mem-lite",
-      "version": "2.34.3",
+      "version": "2.34.5",
       "source": "./",
       "description": "Lightweight persistent memory system for Claude Code — FTS5 search, episode batching, error-triggered recall"
     }

package/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-mem-lite",
-  "version": "2.34.3",
+  "version": "2.34.5",
   "description": "Lightweight persistent memory system for Claude Code — FTS5 search, episode batching, error-triggered recall",
   "author": {
     "name": "sdsrss"

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-mem-lite",
-  "version": "2.34.3",
+  "version": "2.34.5",
   "description": "Lightweight persistent memory system for Claude Code",
   "type": "module",
   "engines": {

package/scripts/prompt-search-utils.mjs

@@ -9,12 +9,23 @@ const CONFIRM_RE = /^(y(es)?|no?|ok|done|go|sure|lgtm|thanks?|ty|继续|确认|
 const SLASH_CMD_RE = /^\//;
 const PURE_OP_RE = /^(git\s+(commit|push|merge)|npm\s+(publish|deploy))\b/i;
 
+/**
+ * CJK-weighted effective length. CJK characters (CJK Unified Ideographs
+ * main + extension A) carry ~3x the semantic token density of Latin
+ * characters — a 5-char Chinese phrase like "优化数据库" encodes roughly
+ * the same information as a 15-char English equivalent. Used by every
+ * length gate downstream of the prompt hook so Latin-calibrated
+ * thresholds (8 / 15) don't falsely reject substantive CJK prompts.
+ */
+export function computeEffectiveLen(text) {
+  if (!text) return 0;
+  const cjkCount = (text.match(/[\u4e00-\u9fff\u3400-\u4dbf]/g) || []).length;
+  return (text.length - cjkCount) + cjkCount * 3;
+}
+
 export function shouldSkip(text) {
   if (!text) return true;
-  // CJK characters carry ~3x semantic weight per char vs Latin
-  const cjkCount = (text.match(/[\u4e00-\u9fff\u3400-\u4dbf]/g) || []).length;
-  const effectiveLen = (text.length - cjkCount) + cjkCount * 3;
-  if (effectiveLen < 8) return true;
+  if (computeEffectiveLen(text) < 8) return true;
   const trimmed = text.trim();
   if (CONFIRM_RE.test(trimmed)) return true;
   if (SLASH_CMD_RE.test(trimmed)) return true;

package/scripts/user-prompt-search.js

@@ -8,7 +8,7 @@ import { sanitizeFtsQuery, relaxFtsQueryToOr, truncate, typeIcon, inferProject,
 import { writeFileSync, readFileSync, existsSync, renameSync } from 'fs';
 import { join } from 'path';
 import Database from 'better-sqlite3';
-import { shouldSkip, detectIntent, shouldSkipByDedup, extractFiles, extractErrorSignature, DEDUP_STALE_MS, matchRegistrySkillName } from './prompt-search-utils.mjs';
+import { shouldSkip, computeEffectiveLen, detectIntent, shouldSkipByDedup, extractFiles, extractErrorSignature, DEDUP_STALE_MS, matchRegistrySkillName } from './prompt-search-utils.mjs';
 
 // ─── Constants ──────────────────────────────────────────────────────────────
 
@@ -29,9 +29,13 @@ const LOOKBACK_MS = 60 * 86400000; // 60 days
 // is TOP_REL_FLOOR below, which drops the whole FTS set when the best match
 // is weak.
 const BM25_MIN_SCORE = Number(process.env.CLAUDE_MEM_UPS_BM25_MIN || 1e-5);
-// Raw-character minimum length for the prompt. Additional to the CJK-weighted
-// `shouldSkip()` effective-length gate; catches medium-short Latin prompts that
-// survive `shouldSkip` but carry too few tokens to justify an FTS lookup.
+// CJK-weighted minimum length for the prompt. Catches medium-short Latin
+// prompts ("run tests", "fix bug now") that survive `shouldSkip`'s weaker 8-unit
+// floor but carry too few tokens to justify an FTS lookup.
+// v2.34.4: applied to `computeEffectiveLen(prompt)`, not raw char count — a
+// 14-char CJK prompt ("优化 hook 性能降低延迟") scores 30 effective units and
+// now reaches FTS, matching shouldSkip's CJK-weighted gate rather than silently
+// failing the raw-char one.
 const PROMPT_MIN_LENGTH = 15;
 
 // v2.33.1: follow-up prompts ("前面那个", "继续 X", "再看看 Y") are short by
@@ -159,6 +163,43 @@ function searchByFile(db, files, project, limit) {
   });
 }
 
+// v2.34.5 Gap 1: prompts-table fallback. When observations-based paths
+// (FTS / file-recall / sigRows / recent) all return empty, scan the user's
+// own past prompts — meta/UX/"did we discuss this" questions often match
+// prior prompts even when no observation was saved. Uses a simpler BM25
+// ranking with no scoring multipliers and no top-|rel| gate (prompts are
+// sparser and more surface-form than observations; the gate would rarely
+// fire and mostly kill real hits).
+function searchByUserPrompts(db, queryText, project, limit) {
+  const ftsQuery = sanitizeFtsQuery(queryText);
+  if (!ftsQuery) return [];
+
+  const cutoff = Date.now() - LOOKBACK_MS;
+  const sql = `
+    SELECT up.id, up.prompt_text, up.created_at_epoch,
+           bm25(user_prompts_fts) as relevance
+    FROM user_prompts_fts
+    JOIN user_prompts up ON up.id = user_prompts_fts.rowid
+    JOIN sdk_sessions s ON s.content_session_id = up.content_session_id
+    WHERE user_prompts_fts MATCH ?
+      AND s.project = ?
+      AND up.created_at_epoch > ?
+    ORDER BY relevance
+    LIMIT ?
+  `;
+
+  let rows = db.prepare(sql).all(ftsQuery, project, cutoff, limit);
+
+  if (rows.length === 0) {
+    const orQuery = relaxFtsQueryToOr(ftsQuery);
+    if (orQuery) {
+      try { rows = db.prepare(sql).all(orQuery, project, cutoff, limit); } catch {}
+    }
+  }
+
+  return rows;
+}
+
 function searchRecent(db, project, limit) {
   const cutoff = Date.now() - LOOKBACK_MS;
   // R1: exclude LOW_SIGNAL degraded titles from "recent" recall intent
@@ -221,6 +262,20 @@ function formatResults(rows) {
   return lines.join('\n');
 }
 
+// v2.34.5 Gap 1: distinct header signals to Claude that these are prior
+// *user questions*, not codebase lessons — helps the reader interpret the
+// row correctly (surface-form match, not a saved insight). Truncate to 80
+// chars (slightly longer than obs titles because prompts carry more context).
+function formatPromptResults(rows) {
+  if (!rows || rows.length === 0) return null;
+  const lines = ['[mem] Past similar questions:'];
+  for (const r of rows) {
+    const text = truncate((r.prompt_text || '').replace(/\s+/g, ' '), 80);
+    lines.push(`P#${r.id} 💬 ${text}`);
+  }
+  return lines.join('\n');
+}
+
 // ─── Registry Skill Pointer (T4 v2.31) ─────────────────────────────────────
 // Formerly "auto-load": we used to read the full SKILL.md body (up to 16KB)
 // and inject it into stdout on keyword match. Now we only emit a short
@@ -296,7 +351,7 @@ async function main() {
   // short continuations ("前面那个?", "does it work?") depend on prior context.
   const followUp = isFollowUpSession();
   const promptMinLen = followUp ? FOLLOWUP_PROMPT_MIN_LENGTH : PROMPT_MIN_LENGTH;
-  if (promptText.trim().length < promptMinLen) return;
+  if (computeEffectiveLen(promptText.trim()) < promptMinLen) return;
   const bm25Floor = followUp ? FOLLOWUP_BM25_MIN_SCORE : BM25_MIN_SCORE;
 
   let db;
@@ -372,10 +427,24 @@ async function main() {
       rows = [...sigRows, ...rows.filter(r => !sigIds.has(r.id))].slice(0, MAX_RESULTS);
     }
 
-    const candidateIds = rows.map(r => r.id);
+    // v2.34.5 Gap 1: if observations-based search drew a blank, try the
+    // user_prompts corpus. Only fires when `rows` is empty (obs hits
+    // suppress the fallback to avoid noise). Namespace prompt IDs with
+    // a "P" prefix so shouldSkipByDedup's Set comparison doesn't collide
+    // with future observation IDs.
+    let promptRows = [];
+    if (rows.length === 0) {
+      promptRows = searchByUserPrompts(db, promptText, project, 3);
+    }
+
+    const candidateIds = rows.length > 0
+      ? rows.map(r => r.id)
+      : promptRows.map(r => `P${r.id}`);
     const dedupSkip = shouldSkipByDedup(candidateIds, INJECTED_IDS_FILE);
 
-    const output = !dedupSkip ? formatResults(rows) : null;
+    const output = !dedupSkip
+      ? (rows.length > 0 ? formatResults(rows) : formatPromptResults(promptRows))
+      : null;
     if (output) {
       process.stdout.write(output + '\n');
       // Write injected IDs for dedup with hook.mjs handleUserPrompt + self-dedup