dsh-chat-tide 0.0.0-stage → 0.2.2

package/lib/attention.js

@@ -0,0 +1,191 @@
+// 关注点（attention）：只有这些主题值得主动搭话；其余只记录，不唤醒。
+import { createHash } from 'node:crypto';
+import { matchExpr, bigram } from './fts.js';
+import { scrubIdentifiers } from './scrub.js';
+
+// 指令腔：出现这些词的短句，多半是在吩咐助手，不是话题本身
+const INSTRUCTION_RE = /留意|关注|记得|提醒|看看|看一下|安排|准备|跟进|汇报|主动|可以|的时候/;
+
+/** 关注点标签规整：去掉主语和指令从句，保留描述性内容；不截断。 */
+export function normalizeAttentionLabel(text) {
+  let label = String(text ?? '').trim();
+  if (!label) return label;
+  // 开头的指令动词先剥掉
+  for (let i = 0; i < 3; i += 1) {
+    const before = label;
+    label = label.replace(/^(记得|提醒我?|帮我|留意|关注|看看|看一下)(一下|下)?/, '').trim();
+    if (label === before) break;
+  }
+  // 按标点切段，丢掉指令段（至少保留一段）
+  const parts = label.split(/[，,；;。]/).map((x) => x.trim()).filter(Boolean);
+  const kept = parts.filter((x) => !INSTRUCTION_RE.test(x));
+  label = (kept.length ? kept : parts).join('，');
+  // 去掉开头的主语与泛化限定词
+  label = label.replace(/^(主人|我|助手|用户)(对|的|固定|通常|一般|经常|偶尔)?/, '').trim();
+  // 收尾标点
+  label = label.replace(/[，,；;。!！?？\s]+$/g, '').trim();
+  return label.length >= 2 ? label : String(text ?? '').trim();
+}
+
+const now = () => Date.now();
+
+/** 默认屏蔽：开发过程、插件实现、模型自身运作不算主人的关注点 */
+const DEFAULT_BLOCK = ['插件', '记忆系统', 'prompt', '提示词', '模型运作', '系统设计', '时间窗', '候选集', '向量', '接口', '写入工具', '关注点机制', '关注点', '污染风险'];
+
+function isBlocked(label, blocked) {
+  const t = String(label ?? '');
+  const list = [...DEFAULT_BLOCK, ...(blocked ?? [])];
+  return list.some((p) => p && t.includes(p));
+}
+
+export function attentionIdFor(label) {
+  return 'att-' + createHash('sha1').update(String(label).trim()).digest('hex').slice(0, 16);
+}
+
+function indexAttention(db, attentionId, text) {
+  try {
+    db.prepare('DELETE FROM attention_fts WHERE attention_id = ?').run(attentionId);
+    db.prepare('INSERT INTO attention_fts(attention_id, text) VALUES (?, ?)').run(attentionId, bigram(text));
+  } catch {
+    /* 索引失败不阻断 */
+  }
+}
+
+export function addAttention(db, { kind = 'topic', refId = null, label, note = null, weight = 0.5, expiresAt = null, ttlDays = null, blockPatterns = [] }) {
+  const clean = scrubIdentifiers(normalizeAttentionLabel(label));
+  note = note === null || note === undefined ? note : scrubIdentifiers(note);
+  if (!clean) return null;
+  if (kind === 'topic' && isBlocked(clean, blockPatterns)) return null;
+  const id = attentionIdFor(clean);
+  const existing = db.prepare('SELECT attention_id, weight FROM attention WHERE attention_id = ?').get(id);
+  const exp = expiresAt ?? (ttlDays ? now() + ttlDays * 86_400_000 : null);
+  if (existing) {
+    db.prepare('UPDATE attention SET weight = MAX(weight, ?), note = COALESCE(?, note), status = ? WHERE attention_id = ?').run(
+      weight,
+      note,
+      'active',
+      id,
+    );
+  } else {
+    db.prepare(
+      `INSERT INTO attention(attention_id, kind, ref_id, label, note, weight, status, hits, created_at, expires_at)
+       VALUES (?,?,?,?,?,?,?,0,?,?)`,
+    ).run(id, kind, refId, clean, note, weight, 'active', now(), exp);
+  }
+  indexAttention(db, id, [clean, note].filter(Boolean).join(' '));
+  return id;
+}
+
+export function listAttention(db, { status = 'active', limit = 50 } = {}) {
+  return db
+    .prepare('SELECT attention_id AS id, kind, label, note, weight, status, hits, last_hit_at AS lastHitAt FROM attention WHERE status = ? ORDER BY weight DESC, created_at DESC LIMIT ?')
+    .all(status, limit);
+}
+
+export function retireAttention(db, id, reason = null) {
+  db.prepare("UPDATE attention SET status = 'retired', note = COALESCE(?, note) WHERE attention_id = ?").run(reason, id);
+  try {
+    db.prepare('DELETE FROM attention_fts WHERE attention_id = ?').run(id);
+  } catch {
+    /* ignore */
+  }
+}
+
+export function bumpAttentionHit(db, id) {
+  db.prepare('UPDATE attention SET hits = hits + 1, last_hit_at = ? WHERE attention_id = ?').run(now(), id);
+}
+
+/** 查询是否落在某个关注点上；返回命中项或 null。 */
+export function matchAttention(db, text, { threshold = 2 } = {}) {
+  const clean = String(text ?? '').trim();
+  if (!clean) return null;
+  const grams = [...new Set(bigram(clean).split(' ').filter((g) => g.length >= 2))];
+  if (!grams.length) return null;
+  let rows = [];
+  try {
+    const expr = matchExpr(clean);
+    if (expr) {
+      rows = db
+        .prepare(
+          `SELECT a.attention_id AS id, a.label, a.note, a.weight, f.text
+             FROM attention_fts f JOIN attention a ON a.attention_id = f.attention_id
+            WHERE attention_fts MATCH ? AND a.status = 'active'
+            ORDER BY a.weight DESC LIMIT 10`,
+        )
+        .all(expr);
+    }
+  } catch {
+    rows = [];
+  }
+  if (!rows.length) {
+    rows = db
+      .prepare("SELECT attention_id AS id, label, note, weight, NULL AS text FROM attention WHERE status = 'active' ORDER BY weight DESC LIMIT 50")
+      .all();
+  }
+  let best = null;
+  for (const r of rows) {
+    const hay = String(r.text ?? '').replace(/ /g, '') + (r.label ?? '') + (r.note ?? '');
+    const hit = grams.filter((g) => hay.includes(g)).length;
+    if (hit >= threshold && (!best || hit > best.hit)) best = { id: r.id, label: r.label, note: r.note, weight: r.weight, hit };
+  }
+  return best;
+}
+
+/** 巡检：过期/低权重的关注点自动退休，保留最多 maxActive 条。 */
+export function pruneAttention(db, { now: t = now(), maxActive = 80, expireDays = 30 } = {}) {
+  const expired = db
+    .prepare("SELECT attention_id AS id FROM attention WHERE status = 'active' AND expires_at IS NOT NULL AND expires_at < ?")
+    .all(t);
+  for (const r of expired) retireAttention(db, r.id, 'expired');
+
+  const stale = db
+    .prepare(
+      `SELECT attention_id AS id FROM attention
+        WHERE status = 'active' AND hits = 0 AND created_at < ?
+        ORDER BY created_at ASC LIMIT 50`,
+    )
+    .all(t - expireDays * 86_400_000);
+  for (const r of stale) retireAttention(db, r.id, 'stale');
+
+  const over = db
+    .prepare(
+      `SELECT attention_id AS id FROM attention WHERE status = 'active'
+        ORDER BY weight ASC, last_hit_at ASC NULLS FIRST LIMIT 500`,
+    )
+    .all();
+  let retired = 0;
+  const activeCount = db.prepare("SELECT count(*) c FROM attention WHERE status = 'active'").get().c;
+  for (const r of over) {
+    if (activeCount - retired <= maxActive) break;
+    retireAttention(db, r.id, 'over-capacity');
+    retired++;
+  }
+  return { expired: expired.length, stale: stale.length, overCapacity: retired };
+}
+
+/** 关注点向量化的文本：主题 + 备注 + 关键词。 */
+export function attentionVectorText(row) {
+  return [row.label, row.note].filter(Boolean).join(' ');
+}
+
+/** attention 的隐式 rowid（vec_attention 用它对齐）。 */
+export function attentionRowid(db, attentionId) {
+  try {
+    return db.prepare('SELECT rowid FROM attention WHERE attention_id = ?').get(attentionId)?.rowid ?? null;
+  } catch {
+    return null;
+  }
+}
+
+/** 还没有向量的活跃关注点（启动补算用）。 */
+export function attentionNeedingVector(db, limit = 200) {
+  return db
+    .prepare(
+      `SELECT a.rowid AS rowid, a.attention_id AS id, a.label, a.note
+         FROM attention a
+        WHERE a.status = 'active'
+          AND NOT EXISTS (SELECT 1 FROM vec_attention v WHERE v.rowid = a.rowid)
+        LIMIT ?`,
+    )
+    .all(limit);
+}

package/lib/budget.js

@@ -0,0 +1,37 @@
+// 路由预算：按窗口累计花费，超限记 budget_skipped（与熔断分开）。
+export const DEFAULT_WINDOW_MS = 5 * 60_000;
+
+/** 读当前窗口已花费，不写入。 */
+export function peekBudget(db, conversationKey, { windowMs = DEFAULT_WINDOW_MS, now = Date.now() } = {}) {
+  const windowStart = Math.floor(now / windowMs) * windowMs;
+  const row = db
+    .prepare('SELECT spent FROM route_budget WHERE conversation_key = ? AND window_start = ?')
+    .get(conversationKey, windowStart);
+  return { windowStart, spent: row?.spent ?? 0 };
+}
+
+/**
+ * 尝试花掉一次预算。
+ * @returns {{allowed:boolean, spent:number, remaining:number, windowStart:number}}
+ */
+export function consumeBudget(db, conversationKey, { windowMs = DEFAULT_WINDOW_MS, budget = 1000, cost = 1, now = Date.now() } = {}) {
+  const windowStart = Math.floor(now / windowMs) * windowMs;
+  const row = db
+    .prepare('SELECT spent FROM route_budget WHERE conversation_key = ? AND window_start = ?')
+    .get(conversationKey, windowStart);
+  const spent = row?.spent ?? 0;
+  if (spent + cost > budget) return { allowed: false, spent, remaining: Math.max(0, budget - spent), windowStart };
+  if (row) {
+    db.prepare('UPDATE route_budget SET spent = spent + ? WHERE conversation_key = ? AND window_start = ?').run(cost, conversationKey, windowStart);
+  } else {
+    db.prepare('INSERT INTO route_budget(conversation_key, window_start, spent) VALUES (?,?,?)').run(conversationKey, windowStart, cost);
+  }
+  return { allowed: true, spent: spent + cost, remaining: Math.max(0, budget - spent - cost), windowStart };
+}
+
+/** 清理过期窗口（低频调用）。 */
+export function pruneBudgets(db, { windowMs = DEFAULT_WINDOW_MS, now = Date.now(), keepWindows = 12 } = {}) {
+  const cutoff = Math.floor(now / windowMs) * windowMs - keepWindows * windowMs;
+  const res = db.prepare('DELETE FROM route_budget WHERE window_start < ?').run(cutoff);
+  return { pruned: res.changes };
+}

package/lib/circuit.js

@@ -0,0 +1,42 @@
+// 熔断：连续失败打开、冷却后半开、成功关闭。与预算跳过分开统计。
+export function createCircuit({ threshold = 5, cooldownMs = 60_000, now = () => Date.now() } = {}) {
+  const state = new Map(); // name -> { failures, openedAt }
+
+  function get(name) {
+    let s = state.get(name);
+    if (!s) {
+      s = { failures: 0, openedAt: 0 };
+      state.set(name, s);
+    }
+    return s;
+  }
+
+  return {
+    isOpen(name) {
+      const s = get(name);
+      if (!s.openedAt) return false;
+      if (now() - s.openedAt >= cooldownMs) {
+        // 半开：放一个请求试探
+        s.openedAt = 0;
+        s.failures = 0;
+        return false;
+      }
+      return true;
+    },
+    recordFailure(name) {
+      const s = get(name);
+      s.failures += 1;
+      if (s.failures >= threshold && !s.openedAt) s.openedAt = now();
+      return s.failures >= threshold;
+    },
+    recordSuccess(name) {
+      const s = get(name);
+      s.failures = 0;
+      s.openedAt = 0;
+    },
+    snapshot(name) {
+      const s = get(name);
+      return { ...s, open: this.isOpen(name) };
+    },
+  };
+}