@shadowforge0/aquifer-memory 1.3.0 → 1.5.9

package/schema/006-insights.sql

@@ -0,0 +1,138 @@
+-- insights: higher-order reflection from session content (Q4).
+--
+-- Holds preferences, recurring patterns, frustrations, and successful
+-- workflows distilled from session_summaries over a window. Vector-indexed
+-- for natural-language recall via aquifer.recallInsights().
+--
+-- DROP-clean: no triggers/functions, no FK from anywhere else into this table.
+-- See scripts/drop-insights.sql.
+
+CREATE TABLE IF NOT EXISTS ${schema}.insights (
+  id                  BIGSERIAL    PRIMARY KEY,
+  tenant_id           TEXT         NOT NULL DEFAULT 'default',
+  agent_id            TEXT         NOT NULL,
+  insight_type        TEXT         NOT NULL
+    CHECK (insight_type IN ('preference', 'pattern', 'frustration', 'workflow')),
+  title               TEXT         NOT NULL CHECK (btrim(title) <> ''),
+  body                TEXT         NOT NULL CHECK (btrim(body) <> ''),
+  source_session_ids  TEXT[]       NOT NULL DEFAULT '{}',
+  evidence_window     TSTZRANGE    NOT NULL,
+  -- embedding: sized vector so HNSW can be built at migrate time. 1024 matches
+  -- the autodetect default (ollama bge-m3). Operators using a provider with
+  -- different dimensions (e.g. openai text-embedding-3-small = 1536) should
+  -- set `aquifer.embedding_dim` via GUC before running migrate(), or the
+  -- coerce block below will pick it up.
+  embedding           vector(1024),
+  importance          REAL         NOT NULL DEFAULT 0.5
+    CHECK (importance >= 0 AND importance <= 1),
+  status              TEXT         NOT NULL DEFAULT 'active'
+    CHECK (status IN ('active', 'stale', 'superseded')),
+  superseded_by       BIGINT       REFERENCES ${schema}.insights(id) ON DELETE SET NULL,
+  idempotency_key     TEXT,
+  metadata            JSONB        NOT NULL DEFAULT '{}'::jsonb,
+  created_at          TIMESTAMPTZ  NOT NULL DEFAULT now(),
+  updated_at          TIMESTAMPTZ  NOT NULL DEFAULT now()
+);
+
+-- Phase 2 C1: canonical_key_v2 identifies the CLAIM (type + canonicalClaim +
+-- entitySet). idempotency_key keeps its revision-level role. Old rows have
+-- canonical_key_v2 = NULL and are not retrofitted; new writes populate it.
+ALTER TABLE ${schema}.insights
+  ADD COLUMN IF NOT EXISTS canonical_key_v2 TEXT;
+
+-- Hot path: recall by agent + type, importance-ranked. Partial idx keeps
+-- the index small by skipping stale/superseded rows.
+CREATE INDEX IF NOT EXISTS idx_insights_active
+  ON ${schema}.insights (tenant_id, agent_id, insight_type, importance DESC, created_at DESC)
+  WHERE status = 'active';
+
+-- Idempotency: caller-supplied key writes once. Partial allows NULL keys.
+CREATE UNIQUE INDEX IF NOT EXISTS idx_insights_idempotency
+  ON ${schema}.insights (idempotency_key)
+  WHERE idempotency_key IS NOT NULL;
+
+-- Phase 2 C1: preflight lookup for canonical_key_v2 active row.
+-- NOT unique — canonical identity can have multiple revisions (legacy as
+-- 'superseded'); only the latest stays 'active'. Partial keeps index small.
+CREATE INDEX IF NOT EXISTS idx_insights_canonical_v2_active
+  ON ${schema}.insights (tenant_id, agent_id, insight_type, canonical_key_v2, created_at DESC)
+  WHERE status = 'active' AND canonical_key_v2 IS NOT NULL;
+
+-- Coerce pre-1.5.1 unsized `vector` column to a sized type so HNSW can be
+-- built. Pre-1.5.1 declared `embedding vector` (no dim) which makes HNSW
+-- creation permanently impossible — the "defer until first row" pattern
+-- was a broken diagnosis of the real problem (pgvector needs a dim on the
+-- COLUMN, not just the data). Idempotent: skipped if already sized.
+-- Dim priority: existing row dim > `aquifer.embedding_dim` GUC > 1024 default.
+-- Note: ${schema} is substituted to a quoted identifier by the loader, so
+-- we string-concat rather than format(%I, ...) to avoid double-quoting.
+DO $$
+DECLARE
+  is_unsized BOOLEAN;
+  existing_dim INT;
+  target_dim INT;
+BEGIN
+  SELECT format_type(atttypid, atttypmod) = 'vector'
+    INTO is_unsized
+    FROM pg_attribute
+    WHERE attrelid = '${schema}.insights'::regclass
+      AND attname = 'embedding';
+
+  IF is_unsized THEN
+    EXECUTE 'SELECT vector_dims(embedding) FROM ${schema}.insights WHERE embedding IS NOT NULL LIMIT 1'
+      INTO existing_dim;
+    target_dim := COALESCE(
+      existing_dim,
+      NULLIF(current_setting('aquifer.embedding_dim', true), '')::int,
+      1024
+    );
+    EXECUTE 'ALTER TABLE ${schema}.insights ALTER COLUMN embedding TYPE vector('
+         || target_dim::text
+         || ') USING embedding::vector('
+         || target_dim::text
+         || ')';
+    RAISE NOTICE '[aquifer] insights.embedding coerced from unsized vector to vector(%)', target_dim;
+  END IF;
+END$$;
+
+-- Vector index: HNSW for cosine distance, only over active insights with
+-- embeddings. Column is now sized so this builds on fresh installs too.
+-- Defer / out-of-memory / unavailable handlers kept as safety nets.
+DO $$
+BEGIN
+  EXECUTE 'CREATE INDEX IF NOT EXISTS idx_insights_embedding
+    ON ${schema}.insights USING hnsw (embedding vector_cosine_ops)
+    WHERE status = ''active'' AND embedding IS NOT NULL';
+EXCEPTION
+  WHEN undefined_object THEN
+    RAISE NOTICE '[aquifer] pgvector hnsw operator not available; skipping HNSW index on insights';
+  WHEN feature_not_supported THEN
+    RAISE NOTICE '[aquifer] HNSW not available on this pgvector; upgrade to >= 0.5.0 for index-accelerated insights recall';
+  WHEN out_of_memory THEN
+    RAISE WARNING '[aquifer] HNSW build on insights.embedding ran out of memory; raise maintenance_work_mem and re-run migrate()';
+  WHEN program_limit_exceeded THEN
+    RAISE WARNING '[aquifer] HNSW build on insights.embedding exceeded an internal limit; inspect pgvector logs';
+END$$;
+
+-- Diagnostic: who-references-which-session, for audit / re-extraction.
+CREATE INDEX IF NOT EXISTS idx_insights_source_sessions
+  ON ${schema}.insights USING GIN (source_session_ids)
+  WHERE status = 'active';
+
+COMMENT ON TABLE ${schema}.insights IS
+  'Higher-order observations distilled from sessions. NOT facts (use entity_state_history). NOT raw recap (use session_summaries). Reflection / skill memory.';
+
+COMMENT ON COLUMN ${schema}.insights.insight_type IS
+  'preference = stable user preference; pattern = recurring behaviour/decision; frustration = repeated pain point; workflow = reusable procedure that worked.';
+
+COMMENT ON COLUMN ${schema}.insights.evidence_window IS
+  'Time range of source sessions used to derive this insight. Half-open by convention.';
+
+COMMENT ON COLUMN ${schema}.insights.importance IS
+  'Caller-supplied [0,1]; recall ranking blends with semantic score and recency.';
+
+COMMENT ON COLUMN ${schema}.insights.canonical_key_v2 IS
+  'Phase 2 C1: stable claim identity = sha256(tenant|agent|type|normalizeCanonicalClaim(claim)|normalizeEntitySet(entities)). Survives LLM title drift. idempotency_key tracks revisions within a claim.';
+
+COMMENT ON COLUMN ${schema}.insights.idempotency_key IS
+  'Revision-level dedupe key. Default in writer: sha256(canonical_key_v2, normalized_body, sorted_session_ids, window). Same claim in same window with same body = duplicate; body change or window extend = new revision (old superseded).';

package/scripts/diagnose-fts-zh.js

@@ -4,11 +4,15 @@
  * Aquifer FTS 中文診斷
  *
  * 測 aquifer 實際搜尋主路徑（trigram ILIKE on search_text + similarity ranking）
- * vs fallback 路徑（tsvector @@ plainto_tsquery('simple', q)）對中文 query 的表現。
+ * vs fallback 路徑（tsvector @@ plainto_tsquery(<cfg>, q)）對中文 query 的表現。
+ * tsconfig 自動偵測：public.zhcfg 已存在就用 'zhcfg'（1.5.0+ 底層是 pg_jieba
+ * jiebaqry，1.4.0 底層是 zhparser），否則退回 'simple'。腳本會印出 zhcfg 實際
+ * parser 名稱——看到 'zhparser' 代表繁體分詞會退化 char-level。
  *
  * env:
  *   DATABASE_URL       — required
  *   AQUIFER_SCHEMA     — default 'public'
+ *   AQUIFER_FTS_CONFIG — override auto-detect ('zhcfg' or 'simple')
  *   DIAGNOSE_QUERIES   — comma-separated, overrides built-in set
  */
 
@@ -41,8 +45,37 @@ function pct(n, d) {
   return `${Math.round((n / d) * 100)}%`;
 }
 
+async function detectFtsConfig() {
+  if (process.env.AQUIFER_FTS_CONFIG === 'zhcfg' || process.env.AQUIFER_FTS_CONFIG === 'simple') {
+    return { cfg: process.env.AQUIFER_FTS_CONFIG, parser: null };
+  }
+  try {
+    const r = await pool.query(`
+      SELECT p.prsname AS parser
+      FROM pg_ts_config c JOIN pg_ts_parser p ON c.cfgparser = p.oid
+      WHERE c.cfgname = 'zhcfg' AND c.cfgnamespace = 'public'::regnamespace
+      LIMIT 1`);
+    if (r.rowCount > 0) return { cfg: 'zhcfg', parser: r.rows[0].parser };
+    return { cfg: 'simple', parser: null };
+  } catch {
+    return { cfg: 'simple', parser: null };
+  }
+}
+
+let FTS_CFG = 'simple';
+
 async function main() {
-  console.log(`=== Aquifer FTS 中文診斷 (schema=${SCHEMA}) ===\n`);
+  const detected = await detectFtsConfig();
+  FTS_CFG = detected.cfg;
+  const parserLabel = detected.parser
+    ? ` parser=${detected.parser}`
+    : '';
+  console.log(`=== Aquifer FTS 中文診斷 (schema=${SCHEMA}, tsconfig=${FTS_CFG}${parserLabel}) ===\n`);
+  if (detected.parser === 'zhparser') {
+    console.log('[warn] zhcfg 目前是 zhparser-backed。scws 內建字典是簡體字為主，對');
+    console.log('       繁體字會全退 char-level 分詞（「記憶」→ 記/憶 單字，等於');
+    console.log('       simple tokenizer）。考慮換 pg_jieba，見 CHANGELOG 1.5.0。\n');
+  }
 
   // -------------------------------------------------------------------------
   // 0. 覆蓋率：search_text NULL 率 → 看 fallback 觸發比例
@@ -94,7 +127,7 @@ async function main() {
   //
   // Ground truth = search_text ILIKE '%q%'（所有源欄位拼出的純文字 superset）
   // 主路徑        = search_text ILIKE（GIN trgm 加速，語意等價 ILIKE）
-  // Fallback     = search_tsv @@ plainto_tsquery('simple', q)
+  // Fallback     = search_tsv @@ plainto_tsquery(<cfg>, q)
   // -------------------------------------------------------------------------
   console.log('--- 2. 主路徑（trigram）vs fallback（tsvector）binary match ---');
   console.log('  query               | truth | trgm  | tsv   | trgm% | tsv%  | tsv-extra');
@@ -114,7 +147,7 @@ async function main() {
         SELECT search_text,
                search_tsv,
                (search_text ILIKE '%' || $1 || '%')                                 AS trgm_hit,
-               (search_tsv  @@ plainto_tsquery('simple', $2))                       AS tsv_hit
+               (search_tsv  @@ plainto_tsquery('${FTS_CFG}', $2))                   AS tsv_hit
         FROM ${qi(SCHEMA)}.session_summaries
         WHERE search_text IS NOT NULL
       )

package/scripts/drop-entity-state-history.sql

@@ -0,0 +1,17 @@
+-- DROP-clean script for entity_state_history (Q3 bitter-lesson escape hatch).
+--
+-- Run this if you decide native long-context / agentic memory has obviated the
+-- temporal state-change layer. Removes the table and all dependent indexes;
+-- nothing else in Aquifer references it directly (FK is one-way: this table
+-- references entities/sessions, not the reverse).
+--
+-- Usage:
+--   psql $DATABASE_URL -v schema=miranda -f scripts/drop-entity-state-history.sql
+
+DROP TABLE IF EXISTS :"schema".entity_state_history CASCADE;
+
+-- Verify nothing remains.
+SELECT to_regclass(:'schema' || '.entity_state_history') AS table_after_drop;
+SELECT to_regclass(:'schema' || '.idx_entity_state_history_current') AS idx_current_after_drop;
+SELECT to_regclass(:'schema' || '.idx_entity_state_history_idempotency') AS idx_idempotency_after_drop;
+-- All three should report NULL.

package/scripts/drop-insights.sql

@@ -0,0 +1,12 @@
+-- DROP-clean script for insights (Q4 bitter-lesson escape hatch).
+--
+-- Removes the table and all dependent indexes. Nothing else in Aquifer
+-- references it directly, so DROP CASCADE is safe and complete.
+
+DROP TABLE IF EXISTS :"schema".insights CASCADE;
+
+-- Verify nothing remains.
+SELECT to_regclass(:'schema' || '.insights') AS table_after_drop;
+SELECT to_regclass(:'schema' || '.idx_insights_active') AS idx_active_after_drop;
+SELECT to_regclass(:'schema' || '.idx_insights_embedding') AS idx_embedding_after_drop;
+-- All three should report NULL.

package/scripts/extract-insights-from-recent-sessions.js

@@ -0,0 +1,315 @@
+#!/usr/bin/env node
+'use strict';
+
+/**
+ * Extract insights from recent sessions and commit them via aquifer.insights.
+ *
+ * Designed for cron: pulls the last N days of session_summaries for one
+ * agent, sends a single LLM call to distil higher-order insights, writes
+ * them to the insights table.
+ *
+ * This is "Route B" from spec.md Q4 — bypasses the cron prompt JSON-parse
+ * fragility and lets us own the LLM call + write atomically.
+ *
+ * Usage:
+ *   node scripts/extract-insights-from-recent-sessions.js \
+ *     --agent main \
+ *     [--days 14] \
+ *     [--max-sessions 50] \
+ *     [--types preference,pattern,frustration,workflow] \
+ *     [--schema miranda] \
+ *     [--tenant-id default] \
+ *     [--dry-run]
+ *
+ * env:
+ *   DATABASE_URL          required
+ *   EMBED_PROVIDER        recommended (vector recall otherwise won't work)
+ *   AQUIFER_LLM_PROVIDER  required (extraction LLM)
+ */
+
+const { Pool } = require('pg');
+const { spawn } = require('node:child_process');
+const aquiferIndex = require('..');
+const { createEmbedder } = require('..');
+const { resolveLlmFn } = require('../consumers/shared/llm-autodetect');
+
+// Optional adapter: spawn the `claude` CLI (Claude Code) for extraction.
+// Toggled by AQUIFER_INSIGHTS_CLI=claude. Uses OAuth from the user's
+// keychain (do NOT set --bare, which disables OAuth). Returns a function
+// with the same contract as resolveLlmFn's output: (prompt) => text.
+function createClaudeCliFn(env) {
+  const model = env.AQUIFER_INSIGHTS_CLI_MODEL || 'opus';
+  const bin = env.AQUIFER_INSIGHTS_CLI_BIN || 'claude';
+  const timeoutMs = parseInt(env.AQUIFER_INSIGHTS_CLI_TIMEOUT_MS || '600000', 10);
+  return function llmFn(prompt) {
+    return new Promise((resolve, reject) => {
+      const child = spawn(bin, ['-p', '--model', model, '--output-format', 'text'], {
+        stdio: ['pipe', 'pipe', 'pipe'],
+        env: process.env,
+      });
+      let stdout = '', stderr = '';
+      const timer = setTimeout(() => {
+        child.kill('SIGKILL');
+        reject(new Error(`[extract-insights] claude cli timeout after ${timeoutMs}ms`));
+      }, timeoutMs);
+      child.stdout.on('data', d => { stdout += d.toString('utf8'); });
+      child.stderr.on('data', d => { stderr += d.toString('utf8'); });
+      child.on('error', e => { clearTimeout(timer); reject(e); });
+      child.on('exit', code => {
+        clearTimeout(timer);
+        if (code === 0) return resolve(stdout);
+        reject(new Error(`[extract-insights] claude cli exit ${code}: ${stderr.slice(0, 800)}`));
+      });
+      child.stdin.end(prompt);
+    });
+  };
+}
+
+function parseArgs(argv) {
+  const args = {
+    agent: null,
+    days: 14,
+    maxSessions: 50,
+    types: ['preference', 'pattern', 'frustration', 'workflow'],
+    schema: process.env.AQUIFER_SCHEMA || 'miranda',
+    tenantId: process.env.AQUIFER_TENANT_ID || 'default',
+    dryRun: false,
+  };
+  for (let i = 0; i < argv.length; i++) {
+    const a = argv[i], v = argv[i + 1];
+    if (a === '--agent') { args.agent = v; i++; }
+    else if (a === '--days') { args.days = parseInt(v, 10); i++; }
+    else if (a === '--max-sessions') { args.maxSessions = parseInt(v, 10); i++; }
+    else if (a === '--types') { args.types = v.split(',').map(s => s.trim()); i++; }
+    else if (a === '--schema') { args.schema = v; i++; }
+    else if (a === '--tenant-id') { args.tenantId = v; i++; }
+    else if (a === '--dry-run') { args.dryRun = true; }
+    else if (a === '-h' || a === '--help') { args.help = true; }
+  }
+  return args;
+}
+
+function buildExtractionPrompt(sessions, types) {
+  const sessionsBlock = sessions.map(s => {
+    const summary = typeof s.structured_summary === 'object' ? s.structured_summary : {};
+    const title = summary.title || s.summary_text?.slice(0, 80) || '(untitled)';
+    const overview = summary.overview || s.summary_text || '';
+    return `### Session ${s.session_id} (${s.started_at})\n${title}\n${overview}`;
+  }).join('\n\n');
+
+  const typesList = types.join(' | ');
+
+  return `You distill HIGHER-ORDER INSIGHTS from a window of past sessions.
+NOT individual facts (those go to entity_state_history). NOT raw recap.
+Insights are stable observations about how the user works, what they prefer,
+where they get stuck, and which workflows succeed.
+
+Aim for 6-12 insights when the window has >50 sessions and >=3 distinct
+themes. Returning only 2-3 on a rich window means you're under-extracting.
+Returning 0 is only correct when the window is genuinely sparse.
+
+## Insight types
+- preference: stable user preference (e.g. "MK prefers terse responses with no trailing summaries")
+- pattern: recurring behaviour or decision (e.g. "MK runs /develop before any non-trivial schema change")
+- frustration: repeated pain point (e.g. "Cron jobs.json prompt parse keeps breaking on minor LLM output drift")
+- workflow: reusable procedure that worked (e.g. "Aquifer release: pack tarball -> bump gateway pkg -> migrate -> restart")
+
+## What to look for — don't just describe incidents
+
+Technical bug patterns (timeouts, drift, regressions) are easy to spot but
+shallow. The *high-value* insights are META-LEVEL signals about how the user
+operates that you'd only see by reading MULTIPLE sessions back-to-back:
+
+- **Behavioural preferences the user re-states or re-enforces.** If the user
+  corrects the agent's tone, format, or process more than once across
+  sessions (e.g. "stop using bullet lists", "查歷史再動手", "不要客套"),
+  that's a preference worth recording. These directly shape how the agent
+  should behave next time — importance 0.85-0.95.
+- **Discipline gaps the user flags repeatedly.** Things like "未驗證就回答",
+  "未查 context 就動手", "重複早上做過的事" are frustration insights about
+  the agent's own behaviour, not about external systems. These are the
+  highest-leverage insights because they prevent future trust erosion.
+- **Decision-style signatures.** How the user makes calls under ambiguity:
+  "prefer direct over indirect routing", "拔掉不再用的 infra 不留以後可能用",
+  "選穩定版不追最新". These are rarely stated once but emerge as a shape
+  across many sessions.
+- **Workflows that succeeded AND the scaffolding that made them succeed.**
+  Not just "user did X", but "user's X works because of Y precondition".
+
+If you only surface technical bug frustrations and miss the meta-level
+behavioural signal, you've failed at this task — a shallow extractor would
+do the same.
+
+## Strict rules
+1. Insights must be TRUE ACROSS MULTIPLE SESSIONS (>=2). One-off events don't count.
+2. title: <= 80 chars, declarative. The display surface — can be colourful.
+3. canonicalClaim: <= 80 chars, DECLARATIVE AND STABLE. The *identity* of this
+   insight. No rhetoric, no examples, no time words, no emphasis. If the same
+   underlying claim shows up under a different title next run, canonicalClaim
+   should be identical. Example: canonicalClaim="mk prefers prose over bullet
+   lists", while title could be "散文段落，禁 bullet" or "prose-only formatting".
+4. entities: array of proper-noun subjects the claim is ABOUT. Tool names,
+   project names, persona names, components. Empty array [] is valid when the
+   claim is generic. Example: ["Aquifer", "insights-cron"] or ["Claude Code"].
+5. body: 2-4 sentences. Cite the pattern AND the root cause or user motivation,
+   not just restate facts.
+6. importance: 0..1.
+   - 0.85-0.95: meta-level preferences + discipline gaps that directly change
+     how the agent should behave (highest leverage — these go here, not lower).
+   - 0.65-0.80: stable technical patterns / workflows.
+   - 0.45-0.60: useful but lower-leverage observations.
+   Don't bunch everything in 0.70-0.85 out of caution — spread the scale.
+7. sourceSessionIds: list every session_id that contributes evidence.
+   >=2 required; >=3 strongly preferred for meta-level insights.
+8. type must be one of: ${typesList}.
+9. Do NOT output {"insights":[]} just because you're uncertain on individual
+   items. Extract what has clear evidence; omit only what lacks it.
+
+## Output
+Single JSON object, no prose, no fence:
+{
+  "insights": [
+    {
+      "type": "preference|pattern|frustration|workflow",
+      "title": "...",
+      "canonicalClaim": "...",
+      "entities": ["..."],
+      "body": "...",
+      "importance": 0.7,
+      "sourceSessionIds": ["sess_a", "sess_b"]
+    }
+  ]
+}
+
+## Sessions in window
+${sessionsBlock}
+`;
+}
+
+function extractJsonBlock(text) {
+  if (!text || typeof text !== 'string') return null;
+  let s = text.trim();
+  const fence = s.match(/```(?:json)?\s*([\s\S]*?)```/);
+  if (fence) s = fence[1].trim();
+  const first = s.indexOf('{'), last = s.lastIndexOf('}');
+  if (first < 0 || last < first) return null;
+  try { return JSON.parse(s.slice(first, last + 1)); } catch { return null; }
+}
+
+async function main() {
+  const args = parseArgs(process.argv.slice(2));
+  if (args.help || !args.agent) {
+    console.error('Usage: --agent <id> [--days 14] [--max-sessions 50] [--types ...] [--dry-run]');
+    process.exit(args.help ? 0 : 2);
+  }
+
+  const dbUrl = process.env.DATABASE_URL || process.env.AQUIFER_DB_URL;
+  if (!dbUrl) { console.error('DATABASE_URL is required'); process.exit(2); }
+
+  const pool = new Pool({ connectionString: dbUrl });
+
+  const useCli = (process.env.AQUIFER_INSIGHTS_CLI || '').toLowerCase() === 'claude';
+  const llmFn = useCli
+    ? createClaudeCliFn(process.env)
+    : resolveLlmFn(null, process.env);
+  if (!llmFn) { console.error('AQUIFER_LLM_PROVIDER + key required (or set AQUIFER_INSIGHTS_CLI=claude)'); process.exit(2); }
+  console.log('[extract-insights] llm backend:', useCli ? `claude cli (${process.env.AQUIFER_INSIGHTS_CLI_MODEL || 'opus'})` : 'api provider');
+
+  const qi = (s) => `"${String(s).replace(/"/g, '""')}"`;
+  const sessionsRes = await pool.query(
+    `SELECT s.session_id, s.started_at, ss.summary_text, ss.structured_summary
+     FROM ${qi(args.schema)}.sessions s
+     JOIN ${qi(args.schema)}.session_summaries ss ON ss.session_row_id = s.id
+     WHERE s.tenant_id = $1
+       AND s.agent_id = $2
+       AND s.started_at >= now() - ($3 || ' days')::interval
+       AND ss.summary_text IS NOT NULL
+     ORDER BY s.started_at DESC
+     LIMIT $4`,
+    [args.tenantId, args.agent, String(args.days), args.maxSessions]
+  );
+
+  const sessions = sessionsRes.rows;
+  console.log(`[extract-insights] ${sessions.length} sessions in last ${args.days}d for agent=${args.agent}`);
+  if (sessions.length === 0) {
+    console.log('[extract-insights] nothing to do, exiting clean');
+    await pool.end();
+    return;
+  }
+
+  const prompt = buildExtractionPrompt(sessions, args.types);
+  console.log('[extract-insights] sending to LLM...');
+  let raw;
+  try {
+    raw = await llmFn(prompt);
+  } catch (e) {
+    console.error('[extract-insights] llm call failed:', e.message);
+    await pool.end();
+    process.exit(1);
+  }
+
+  const parsed = extractJsonBlock(raw);
+  if (!parsed || !Array.isArray(parsed.insights)) {
+    console.error('[extract-insights] malformed LLM output, dumping raw:\n', raw);
+    await pool.end();
+    process.exit(1);
+  }
+  console.log(`[extract-insights] ${parsed.insights.length} insights returned`);
+
+  if (args.dryRun) {
+    console.log(JSON.stringify(parsed.insights, null, 2));
+    await pool.end();
+    return;
+  }
+
+  // Build embedFn (optional — without it insights still write but recall via
+  // semantic query won't work).
+  let embedFn = null;
+  try {
+    const e = createEmbedder({});
+    embedFn = (texts) => e.embedBatch(texts);
+  } catch {
+    console.warn('[extract-insights] embed unavailable, insights will save without vector index entries');
+  }
+
+  const aquifer = aquiferIndex.createAquifer({
+    db: pool,
+    schema: args.schema,
+    tenantId: args.tenantId,
+    embed: embedFn ? { fn: embedFn } : undefined,
+  });
+
+  // Window = oldest..newest source session timestamp (fallback to now).
+  const sortedTimes = sessions.map(s => new Date(s.started_at)).sort((a, b) => a - b);
+  const windowFrom = sortedTimes[0]?.toISOString() || new Date().toISOString();
+  const windowTo = sortedTimes[sortedTimes.length - 1]?.toISOString() || new Date().toISOString();
+
+  let written = 0, duplicates = 0, failed = 0;
+  for (const ins of parsed.insights) {
+    if (!ins || !ins.type || !args.types.includes(ins.type)) { failed++; continue; }
+    const r = await aquifer.insights.commitInsight({
+      agentId: args.agent,
+      type: ins.type,
+      title: ins.title,
+      canonicalClaim: typeof ins.canonicalClaim === 'string' ? ins.canonicalClaim : undefined,
+      entities: Array.isArray(ins.entities) ? ins.entities : [],
+      body: ins.body,
+      sourceSessionIds: Array.isArray(ins.sourceSessionIds) ? ins.sourceSessionIds : [],
+      evidenceWindow: { from: windowFrom, to: windowTo },
+      importance: ins.importance,
+      metadata: { extractor: 'extract-insights-from-recent-sessions', windowDays: args.days },
+    });
+    if (!r.ok) { failed++; console.warn(`  fail ${ins.type}: ${r.error.code} ${r.error.message}`); }
+    else if (r.data.duplicate) { duplicates++; console.log(`  dup  ${ins.type}: ${ins.title}`); }
+    else { written++; console.log(`  ok   ${ins.type} (id=${r.data.insight.id}): ${ins.title}`); }
+  }
+  console.log(`[extract-insights] written=${written} dup=${duplicates} failed=${failed}`);
+
+  await aquifer.close?.().catch(() => {});
+  await pool.end().catch(() => {});
+}
+
+main().catch(err => {
+  console.error('[extract-insights] fatal:', err.stack || err.message);
+  process.exit(1);
+});

package/scripts/find-dburl-hints.js

@@ -0,0 +1,29 @@
+const fs = require('fs');
+const path = require('path');
+
+const root = process.argv[2] || process.cwd();
+const exts = new Set(['.md', '.js', '.json', '.yaml', '.yml', '.sh', '.sql', '.txt']);
+const needles = [/DATABASE_URL/g, /AQUIFER_DB_URL/g, /postgresql:\/\//g, /postgres:\/\//g];
+
+function walk(dir, out=[]) {
+  for (const entry of fs.readdirSync(dir, { withFileTypes: true })) {
+    if (entry.name === 'node_modules' || entry.name === '.git') continue;
+    const p = path.join(dir, entry.name);
+    if (entry.isDirectory()) walk(p, out);
+    else out.push(p);
+  }
+  return out;
+}
+
+for (const f of walk(root)) {
+  if (!exts.has(path.extname(f))) continue;
+  let txt;
+  try { txt = fs.readFileSync(f, 'utf8'); } catch { continue; }
+  if (!needles.some(re => re.test(txt))) continue;
+  const lines = txt.split('\n');
+  lines.forEach((line, i) => {
+    if (needles.some(re => re.test(line))) {
+      console.log(`${path.relative(root, f)}:${i+1}: ${line}`);
+    }
+  });
+}

package/scripts/queries.json

@@ -0,0 +1,45 @@
+{
+  "version": 1,
+  "queries": [
+    {
+      "id": "q-001",
+      "lang": "en",
+      "text": "How do I set up Aquifer memory storage with PostgreSQL?"
+    },
+    {
+      "id": "q-002",
+      "lang": "en",
+      "text": "What is the difference between memory_search and session_recall in Aquifer?"
+    },
+    {
+      "id": "q-003",
+      "lang": "zh",
+      "text": "Aquifer 的 session recall 是怎麼做 hybrid 檢索的？"
+    },
+    {
+      "id": "q-004",
+      "lang": "zh",
+      "text": "為什麼 zhcfg 會依賴 jieba 或 zhparser？"
+    },
+    {
+      "id": "q-005",
+      "lang": "mixed",
+      "text": "How to debug fts-zhcfg pipeline 在 jieba migration 後失敗的問題？"
+    },
+    {
+      "id": "q-006",
+      "lang": "mixed",
+      "text": "memory_search 找不到結果時，應該先看哪個 log 或 table？"
+    },
+    {
+      "id": "q-007",
+      "lang": "en",
+      "text": "How does hybrid-rerank differ from hybrid mode in retro recall bench?"
+    },
+    {
+      "id": "q-008",
+      "lang": "zh",
+      "text": "Aquifer 初始化後要如何驗證 embeddings pipeline 有正常工作？"
+    }
+  ]
+}