@nightpainters/dsh-ling 1.5.1

package/lib/host/summarizer.js

@@ -0,0 +1,103 @@
+// dsh-ling host — DSH 会话概述器 v1(轻量)。
+// 把 dsh_turns_raw 中每个顶层会话聚合成 source='dsh' 的 conv_overview:
+//   title = 首条真人消息单行截断;summary = 首问摘要 + 轮次;类别/关键词启发式。
+// 增量:每会话水位 kv('sumdsh:<sid>') = 已概述的最大 seq;有新增才重建。
+// 无 LLM 深摘要(v1);原文保留在 dsh_turns_raw,供 M2 概述升级。
+import { categorizeTitle } from './classify.js';
+import { heuristicTitle, TITLE_MAX_CHARS } from './retitle.js';
+import { keywordsFrom } from './inject-common.js';
+import { toIso } from './util.js';
+
+const MIN_MESSAGES = 2;
+const MIN_CHARS = 40;
+const TITLE_MAX = TITLE_MAX_CHARS; // D2(2026-09-17):46 → 58。旧值把一句话切成半句;此值仅作启发式失败时的兜底
+const SUMMARY_FIRST_MAX = 140;
+
+export function summarizeDsh(memory, { force = false } = {}) {
+  const groups = memory.db.prepare(
+    `SELECT session_id, COUNT(*) AS n, COALESCE(SUM(LENGTH(text)),0) AS chars,
+            MIN(ts) AS t0, MAX(ts) AS t1, MAX(seq) AS maxSeq
+     FROM dsh_turns_raw WHERE session_id NOT LIKE 'import:%' GROUP BY session_id`,
+  ).all();
+  const stats = { sessions: groups.length, created: 0, updated: 0, skipped: 0, watermarkSet: 0, archivedSkipped: 0 };
+  // 归档即清理(2026-09-16 用户定:清掉归档会话,未归档的测试会话不动):
+  // 已归档会话不再生成/重建概述 —— 否则清掉的会被下一轮概述器复活。force 仍可强制重建(CLI 显式动作)。
+  const archived = new Set(
+    memory.db.prepare('SELECT session_id FROM session_meta WHERE archived=1').all().map((r) => String(r.session_id)),
+  );
+
+  const firstUserStmt = memory.db.prepare(
+    `SELECT text FROM dsh_turns_raw WHERE session_id=? AND role='user' ORDER BY seq ASC LIMIT 1`,
+  );
+  const laterUserStmt = memory.db.prepare( // D2:首句是「你好」「下午好」这类时,往后取有实义的首句
+    `SELECT text FROM dsh_turns_raw WHERE session_id=? AND role='user' ORDER BY seq ASC LIMIT 6`,
+  );
+  const roleCountStmt = memory.db.prepare(
+    `SELECT role, COUNT(*) AS n FROM dsh_turns_raw WHERE session_id=? GROUP BY role`,
+  );
+
+  for (const g of groups) {
+    const sid = String(g.session_id);
+    if (!force && archived.has(sid)) {
+      stats.skipped += 1;
+      stats.archivedSkipped += 1;
+      continue;
+    }
+    const wmKey = 'sumdsh:' + sid;
+    const wm = Number(memory.kvGet(wmKey) || 0);
+    if (!force && wm >= Number(g.maxSeq)) {
+      stats.skipped += 1;
+      continue;
+    }
+    const roles = roleCountStmt.all(sid);
+    const users = Number((roles.find((r) => r.role === 'user') || {}).n || 0);
+    const asst = Number((roles.find((r) => r.role === 'assistant') || {}).n || 0);
+    const eligible = users >= 1 && users + asst >= MIN_MESSAGES && Number(g.chars) >= MIN_CHARS;
+    if (eligible) {
+      const first = String(firstUserStmt.get(sid)?.text || '').trim();
+      const oneLine = first.replace(/\s+/g, ' ');
+      // D2(2026-09-17):标题不再硬截首句。旧行为 60 条 dsh 里 26 条被切成半句话
+      // (「…继续走你的新生之路。 我刚刚对1.2.1的」),且「你好」「下午好」直接当标题。
+      // 三层:剥称呼/问候前缀 → 不好则往后取有实义的首句 → 在句读处收尾。
+      let title = heuristicTitle(oneLine);
+      if (!title) {
+        for (const t of laterUserStmt.all(sid).slice(1)) {
+          title = heuristicTitle(String(t.text || ''));
+          if (title) break;
+        }
+      }
+      if (!title) title = oneLine.slice(0, TITLE_MAX);
+      const summary = first.length > SUMMARY_FIRST_MAX ? oneLine.slice(0, SUMMARY_FIRST_MAX) + '…' : oneLine;
+      const category = categorizeTitle(title);
+      const keywords = keywordsFrom((title + ' ' + oneLine).slice(0, 400));
+      const existing = memory.overviewById('dsh', sid);
+      const deepDone = !!memory.kvGet('deep:' + sid);
+      memory.upsertOverview({
+        conv_id: sid,
+        source: 'dsh',
+        title: title || '(未命名会话)',
+        // 修 A:MIN/MAX(ts) 在混合形态下会被遗留数字串带偏 —— 落库前一并归一为 ISO
+        started_at: toIso(g.t0),
+        updated_at: toIso(g.t1),
+        domain_tags: [category === 'knowledge' ? '知识' : category === 'feeling' ? '生活' : '日常'],
+        category,
+        keywords,
+        // 已深摘的会话保留深摘要,不因增量刷新被浅版覆盖
+        summary: deepDone && existing ? existing.summary : `${summary} — ${users + asst} 条消息`,
+        // G2(2026-09-17):概述器**不再往库里写** heat/importance/last_hit_at/hit_count ——
+        // 这四个是本机用户态(置顶、命中、热度):旧行为每轮重建都写 0 且被 UPSERT 无条件覆盖,
+        // 使置顶的删除保护(importance>=1)与热度排序同时失效。新建行由 upsertOverview 填默认 0。
+        overview_ok: 1,
+        origin: 'dsh-summarizer-v1',
+      });
+      if (existing) stats.updated += 1;
+      else stats.created += 1;
+    } else {
+      stats.skipped += 1;
+    }
+    memory.kvSet(wmKey, String(g.maxSeq));
+    stats.watermarkSet += 1;
+  }
+  if (groups.length) memory.kvSet('memory_version', String(Date.now()));
+  return stats;
+}

package/lib/host/tone-advice.js

@@ -0,0 +1,112 @@
+// dsh-ling host — 语气 P1:按记忆分类(知识↔生活)归纳语气候选(人审采纳,自由生长)。
+// 语义(2026-09-11 作者拍板):
+//   - 语气/风格属"自性类"→ 成长通道,锁定下免手术;契约类(名字/称呼/底线/世界观)仍属手术区;
+//   - 只做可逆动作:语气仅在四档枚举间替换;风格文本只追加、不删除用户写过的字;
+//   - 采纳=用户按下的那一下(候选不落盘)。
+import { extractJson } from './genesis.js';
+import { isPlaceholderSummary, snippetFromRaw } from './genesis.js';
+
+export const TONE_SET = ['natural', 'literary', 'concise', 'playful'];
+export const TONE_LABEL = { natural: '自然亲切', literary: '文雅', concise: '简洁直接', playful: '活泼俏皮' };
+
+/** 模型可能回中文/大写/近义词:统一归一到四档枚举,失败返回 ''(不猜)。 */
+export const TONE_ALIAS = {
+  natural: 'natural', 自然: 'natural', 自然亲切: 'natural', 亲切: 'natural', 温和: 'natural',
+  literary: 'literary', 文雅: 'literary', 文学: 'literary', 典雅: 'literary', 书面: 'literary',
+  concise: 'concise', 简洁: 'concise', 简洁直接: 'concise', 直接: 'concise', 精炼: 'concise', 干练: 'concise',
+  playful: 'playful', 活泼: 'playful', 俏皮: 'playful', 活泼俏皮: 'playful', 轻松: 'playful', 幽默: 'playful',
+};
+export function normalizeTone(raw) {
+  const s = String(raw ?? '').trim();
+  if (!s) return '';
+  const lower = s.toLowerCase();
+  if (TONE_SET.includes(lower)) return lower;
+  if (TONE_ALIAS[s]) return TONE_ALIAS[s];
+  if (TONE_ALIAS[lower]) return TONE_ALIAS[lower];
+  for (const k of TONE_SET) if (lower.includes(k)) return k; // 如 "concise(简洁直接)"
+  for (const [k, v] of Object.entries(TONE_ALIAS)) if (s.includes(k)) return v;
+  return '';
+}
+
+/** 两侧取样权重:工作侧看知识,生活侧看情感与日常 */
+export const TONE_SIDE_WEIGHTS = {
+  work: { knowledge: 1.0, daily: 0.3, feeling: 0.05 },
+  life: { feeling: 1.0, daily: 0.6, knowledge: 0.05 },
+};
+/** 置顶加成刻意小于最小类别差:只在同类内优先,不反转"工作侧看知识 / 生活侧看情感"的类别序。 */
+export const TONE_IMPORTANCE_BONUS = 0.35;
+
+export const TONE_ADVICE_SYS = `你是器灵的"语气观察员"。请阅读两组历史材料(工作侧 / 生活侧),分别判断用户更适应哪一种语气,并给出依据。
+四档语气(必须从中选择,不要创造新词):
+- natural 自然亲切:口语化但不轻浮,专业问题先严谨再谈温度
+- literary 文雅:文雅简洁,可适当用典,不堆砌辞藻
+- concise 简洁直接:直接、精炼、少铺垫、结论先行
+- playful 活泼俏皮:轻松活泼,适度俏皮,仍以有用为先
+要求:
+1. 依据必须来自材料里真实出现的痕迹(如反复要求结论先行、深夜更接受玩笑、讨厌冗长文言等),不要空话;
+2. 只输出一个 JSON 对象,键名严格为 work / life;每个值为 {"tone":"四档之一","note":"≤80 字语气注(她会怎么写自己)","evidence":"≤120 字判断依据"};
+3. 不要解释、不要 markdown 围栏以外的文字。`;
+
+/** 按侧取样:类别权重 + 置顶加成,降序取前 limit;并做三级取用(深摘→概述→片段/线索)。 */
+export function sampleToneRows(memory, { side = 'life', limit = 40, pronoun = '她' } = {}) {
+  const weights = TONE_SIDE_WEIGHTS[side] || TONE_SIDE_WEIGHTS.life;
+  const pool = memory.db.prepare(
+    'SELECT conv_id, title, summary, keywords, category, importance, updated_at FROM conv_overview',
+  ).all();
+  const rows = pool.map((r) => ({
+    conv_id: String(r.conv_id),
+    title: r.title,
+    summary: String(r.summary || ''),
+    keywords: (() => { try { const v = JSON.parse(String(r.keywords || '[]')); return Array.isArray(v) ? v : []; } catch { return []; } })(),
+    category: String(r.category),
+    importance: Number(r.importance) || 0,
+    updated_at: String(r.updated_at || ''),
+  }))
+    .map((r) => ({ ...r, score: (weights[r.category] ?? 0.05) + (r.importance > 0 ? TONE_IMPORTANCE_BONUS : 0) }))
+    .sort((a, b) => (b.score - a.score) || b.updated_at.localeCompare(a.updated_at))
+    .slice(0, Math.max(1, Math.min(limit, 200)));
+  const out = rows.map((r) => {
+    const deep = memory.kvGet('deep.sum:' + r.conv_id);
+    let summary = '';
+    if (deep) summary = String(deep);
+    else {
+      const s = r.summary.trim();
+      if (s && !isPlaceholderSummary(s)) summary = s;
+    }
+    return {
+      title: r.title,
+      summary,
+      snippet: summary ? '' : snippetFromRaw(memory, r.conv_id, { pronoun }),
+      keywords: summary ? [] : r.keywords,
+      category: r.category,
+    };
+  });
+  return { rows: out, sampled: out.length, pool: pool.length };
+}
+
+/** 清洗模型输出 → { work:{tone,note,evidence}, life:{...} }(tone 走别名归一,失败置空) */
+export function parseToneAdvice(text) {
+  const j = extractJson(text) || {};
+  const side = (a, b, c) => a || b || c || {};
+  const pick = (v) => {
+    const o = v && typeof v === 'object' ? v : {};
+    return {
+      tone: normalizeTone(o.tone),
+      note: String(o.note ?? '').trim().slice(0, 80),
+      evidence: String(o.evidence ?? '').trim().slice(0, 120),
+    };
+  };
+  return {
+    work: pick(side(j.work, j['工作'], j['工作侧'])),
+    life: pick(side(j.life, j['生活'], j['生活侧'])),
+  };
+}
+
+/** 风格文本追加:只追加、不删除;已含则原样返回(去重)。 */
+export function appendStyleNote(existing, note) {
+  const cur = String(existing || '');
+  const add = String(note || '').trim();
+  if (!add) return cur;
+  if (cur.includes(add)) return cur;
+  return (cur.trimEnd() ? cur.trimEnd() + '\n' : '') + add;
+}