@theseus-whale/dsh-continuity 0.0.0-stage → 4.12.2

package/bin/eval-label-rules.mjs

@@ -0,0 +1,342 @@
+#!/usr/bin/env node
+/**
+ * eval-label-rules.mjs — 评估 A 表打标判据（top-1 / top-2 / 空值 / 精度 / 覆盖 / 误标率）。
+ *
+ * 输入是一个 JSONL，每行：
+ *   {"turn": 2056, "text": "……这一轮的正文……", "label": "跑团叙事"}
+ * 其中 label 是人工定的主话题（可以留空字符串表示"人工认为该判空"）。
+ *
+ * 用法：
+ *   node bin/eval-label-rules.mjs --input rows.jsonl
+ *   node bin/eval-label-rules.mjs --input rows.jsonl --baseline-file topic-baseline.json
+ *   node bin/eval-label-rules.mjs --input rows.jsonl --center-min 0.02 --lead 0.02 --json
+ *
+ * 选项：
+ *   --input <file>          必填，JSONL
+ *   --concepts <file>       A 表（默认 包内 templates/concepts.md；找不到则报错）
+ *   --baseline-file <file>  生产 baseline（支持 {"topic":0.51} 或 {version,window,samples:{topic:[...]}}）。
+ *                           不给则用输入集的 leave-one-out 均值——**仅用于评估**，不是生产语义。
+ *   --center-min <n>        centered 下限（默认 0.02）
+ *   --lead <n>              领先第二名的下限（默认 0.02）
+ *   --topic-min-score <n>   绝对下限保险（默认 0.33）
+ *   --min-score <n>         workflow 入选下限（默认 0.25）
+ *   --topk <n>              topic 最多返回几条（默认 2）
+ *   --min-precision <n>     低于该精度时退出码 1（默认不设）
+ *   --json                  只输出 JSON 摘要
+ *
+ * 注意：本脚本不读会话原文，也不写任何会话文件；输入里有什么就只算什么。
+ */
+import fs from 'node:fs';
+import path from 'node:path';
+import { fileURLToPath } from 'node:url';
+
+const here = path.dirname(fileURLToPath(import.meta.url));
+const pkgRoot = path.resolve(here, '..');
+
+const argv = process.argv.slice(2);
+const opt = (name, fallback = null) => {
+  const i = argv.indexOf(`--${name}`);
+  if (i < 0) return fallback;
+  const next = argv[i + 1];
+  return next === undefined || next.startsWith('--') ? true : next;
+};
+const num = (name, fallback) => {
+  const v = opt(name);
+  if (v === null || v === true) return fallback;
+  const n = Number(v);
+  return Number.isFinite(n) ? n : fallback;
+};
+
+const inputPath = opt('input');
+if (!inputPath || inputPath === true) {
+  console.error('缺 --input <jsonl>');
+  process.exit(2);
+}
+const conceptsPath = (() => {
+  const explicit = opt('concepts');
+  if (explicit && explicit !== true) return explicit;
+  const candidate = path.join(pkgRoot, 'templates', 'concepts.md');
+  return fs.existsSync(candidate) ? candidate : null;
+})();
+if (!conceptsPath || !fs.existsSync(conceptsPath)) {
+  console.error('找不到 A 表：请用 --concepts <concepts.md> 指定');
+  process.exit(2);
+}
+
+const centerMin = num('center-min', 0.02);
+const lead = num('lead', 0.02);
+const topicMinScore = num('topic-min-score', 0.33);
+const minScore = num('min-score', 0.25);
+const topk = num('topk', 2);
+const minPrecision = num('min-precision', null);
+const asJson = Boolean(opt('json', false));
+const baselineFile = opt('baseline-file');
+
+const { createBLabeler } = await import('file://' + path.join(pkgRoot, 'src/memory/bc-label.mjs'));
+const { loadConcepts } = await import('file://' + path.join(pkgRoot, 'src/memory/concepts-file.mjs'));
+const { embedTexts } = await import('file://' + path.join(pkgRoot, 'src/facets/embedding.mjs'));
+
+const rows = loadConcepts(conceptsPath);
+if (rows.length === 0) {
+  console.error('A 表为空：' + conceptsPath);
+  process.exit(2);
+}
+const labels = Object.fromEntries(rows.map((r) => [r.label, r.description]));
+const kinds = Object.fromEntries(rows.map((r) => [r.label, r.kind]));
+const concepts = rows.map((r) => r.label);
+const topics = rows.filter((r) => r.kind === 'topic').map((r) => r.label);
+const workflowNames = rows.filter((r) => r.kind === 'workflow').map((r) => r.label);
+
+const samples = fs.readFileSync(inputPath, 'utf8').replace(/^\uFEFF/, '')
+  .split(/\r?\n/)
+  .filter((line) => line.trim() !== '')
+  .map((line, index) => {
+    try {
+      const row = JSON.parse(line);
+      return {
+        turn: row.turn ?? index + 1,
+        text: String(row.text ?? ''),
+        label: String(row.label ?? ''),
+        // 给了 workflowLabel 才评 workflow：数组=人工标签（[] 表示该判空）；不写=不参与
+        workflowLabel: Array.isArray(row.workflowLabel)
+          ? row.workflowLabel.map((name) => String(name))
+          : (typeof row.workflowLabel === 'string'
+            ? row.workflowLabel.split(/[\/,，、]/).map((name) => name.trim()).filter(Boolean)
+            : undefined),
+      };
+    } catch (error) {
+      console.error(`第 ${index + 1} 行 JSON 解析失败：${error.message}`);
+      process.exit(2);
+    }
+  });
+if (samples.length === 0) {
+  console.error('输入为空');
+  process.exit(2);
+}
+
+const labeler = createBLabeler({ embed: embedTexts, labels, concepts, kinds });
+
+// baseline：生产文件优先；否则用输入集 leave-one-out 均值（仅评估用）
+// v2 文件形如 {version:2, topics:{...}, workflows:{...}}；v1 只有 samples（当作 topics）
+let baseline = null;
+let workflowBaseline = null;
+let baselineSource = 'leave-one-out';
+if (baselineFile && baselineFile !== true) {
+  const raw = JSON.parse(fs.readFileSync(baselineFile, 'utf8').replace(/^\uFEFF/, ''));
+  const topicSource = raw?.topics && typeof raw.topics === 'object'
+    ? raw.topics
+    : (raw?.samples && typeof raw.samples === 'object' ? raw.samples : raw);
+  const workflowSource = raw?.workflows && typeof raw.workflows === 'object' ? raw.workflows : null;
+
+  // 缺失必须进 NaN，不能进 0：0 会被当成有效 baseline（等于该项不居中），缺失项就会通吃
+  const buildMap = (source, names) => Object.fromEntries(names.map((name) => {
+    const value = source?.[name];
+    if (Array.isArray(value)) {
+      const nums = value.map((item) => Number(item)).filter((item) => Number.isFinite(item));
+      return [name, nums.length > 0 ? nums.reduce((sum, item) => sum + item, 0) / nums.length : Number.NaN];
+    }
+    if (value === undefined || value === null || value === '') return [name, Number.NaN];
+    const numeric = Number(value);
+    return [name, Number.isFinite(numeric) ? numeric : Number.NaN];
+  }));
+
+  const fill = (map, names) => {
+    const known = Object.values(map).filter((value) => Number.isFinite(value));
+    if (known.length === 0) return null;
+    const fallback = known.reduce((sum, value) => sum + value, 0) / known.length;
+    for (const name of names) {
+      if (!Number.isFinite(map[name])) map[name] = fallback;
+    }
+    return map;
+  };
+
+  baseline = fill(buildMap(topicSource, topics), topics);
+  if (workflowSource) {
+    workflowBaseline = fill(buildMap(workflowSource, workflowNames), workflowNames);
+  }
+  baselineSource = baselineFile;
+  if (!baseline && !workflowBaseline) {
+    // 一个样本都没有：按 raw bootstrap，不居中
+    baselineSource = 'raw-bootstrap';
+  }
+}
+
+// 先算一遍 raw，够用来做 leave-one-out 均值
+const results = [];
+for (const sample of samples) {
+  const res = await labeler.labelChunked(sample.text, {
+    topK: 3,
+    minScore,
+    topicMinScore,
+    topicCenterMin: centerMin,
+    topicLead: lead,
+    topicTopK: topk,
+    topicBaseline: baseline ?? undefined,
+    workflowBaseline: workflowBaseline ?? undefined,
+  });
+  results.push({ sample, res });
+}
+
+let effectiveBaseline = baseline;
+if (!baseline) {
+  const sums = Object.fromEntries(topics.map((t) => [t, 0]));
+  let count = 0;
+  for (const { res } of results) {
+    const scores = res.topicScores ?? {};
+    const keys = topics.filter((t) => Number.isFinite(Number(scores[t])));
+    if (keys.length === 0) continue;
+    for (const t of keys) sums[t] += Number(scores[t]);
+    count += 1;
+  }
+  if (count > 0) {
+    effectiveBaseline = Object.fromEntries(topics.map((t) => [t, sums[t] / count]));
+    // leave-one-out：重新打一遍
+    for (const item of results) {
+      const scores = item.res.topicScores ?? {};
+      const loo = Object.fromEntries(topics.map((t) => {
+        const own = Number(scores[t]);
+        if (!Number.isFinite(own) || count <= 1) return [t, effectiveBaseline[t] ?? 0];
+        return [t, (sums[t] - own) / (count - 1)];
+      }));
+      item.res = await labeler.labelChunked(item.sample.text, {
+        topK: 3,
+        minScore,
+        topicMinScore,
+        topicCenterMin: centerMin,
+        topicLead: lead,
+        topicTopK: topk,
+        topicBaseline: loo,
+        workflowBaseline: workflowBaseline ?? undefined,
+      });
+    }
+  }
+}
+
+let nonEmpty = 0;
+let empty = 0;
+let correct = 0;
+let wrong = 0;
+let top2Hit = 0;
+let labelled = 0;
+const mistakes = [];
+const perRow = [];
+
+// workflow 侧统计（只统计给了 workflowLabel 的行；label=[] 表示"人工认为该判空"）
+let wfLabelled = 0;
+let wfNonEmpty = 0;
+let wfExact = 0;
+let wfAnyHit = 0;
+let wfMissing = 0;
+let wfSpurious = 0;
+let wfPrecisionSum = 0;
+let wfRecallSum = 0;
+let wfTop2Hit = 0;
+const wfMistakes = [];
+
+for (const { sample, res } of results) {
+  const picked = (res.topic ?? []).map((item) => item.concept);
+  const got = picked[0] ?? null;
+  const want = sample.label.trim();
+  const hasHuman = want !== '';
+  if (hasHuman) labelled += 1;
+  if (got === null) empty += 1;
+  else nonEmpty += 1;
+  const ok = got !== null && hasHuman && got === want;
+  const inTop2 = hasHuman && picked.slice(0, 2).includes(want);
+  if (ok) correct += 1;
+  if (got !== null && hasHuman && !ok) { wrong += 1; mistakes.push({ turn: sample.turn, got, want }); }
+  if (inTop2) top2Hit += 1;
+
+  const wfGot = (res.workflow ?? []).map((item) => item.concept);
+  const wfWant = sample.workflowLabel; // undefined = 未标注；[] = 人工认为该空
+  const wfHasHuman = Array.isArray(wfWant);
+  if (wfHasHuman) {
+    wfLabelled += 1;
+    if (wfGot.length > 0) wfNonEmpty += 1;
+    const inter = wfGot.filter((name) => wfWant.includes(name));
+    const sameSet = wfGot.length === wfWant.length && inter.length === wfGot.length;
+    if (sameSet) wfExact += 1;
+    if (inter.length > 0) wfAnyHit += 1;
+    if (wfWant.length > 0 && wfGot.length === 0) wfMissing += 1;
+    if (wfWant.length === 0 && wfGot.length > 0) wfSpurious += 1;
+    wfPrecisionSum += wfGot.length > 0 ? inter.length / wfGot.length : (wfWant.length === 0 ? 1 : 0);
+    wfRecallSum += wfWant.length > 0 ? inter.length / wfWant.length : (wfGot.length === 0 ? 1 : 0);
+    if (wfGot.slice(0, 2).some((name) => wfWant.includes(name))) wfTop2Hit += 1;
+    if (!sameSet) wfMistakes.push({ turn: sample.turn, got: wfGot, want: wfWant });
+  }
+
+  perRow.push({ turn: sample.turn, got, want, ok, top: picked, wfGot, wfWant: wfWant ?? null });
+}
+
+const total = results.length;
+const precision = nonEmpty > 0 ? correct / nonEmpty : 0;
+const coverage = total > 0 ? nonEmpty / total : 0;
+const mislabelRate = total > 0 ? wrong / total : 0;
+const top2Rate = labelled > 0 ? top2Hit / labelled : 0;
+
+const wfPrecision = wfLabelled > 0 ? wfPrecisionSum / wfLabelled : 0;
+const wfRecall = wfLabelled > 0 ? wfRecallSum / wfLabelled : 0;
+const wfExactRate = wfLabelled > 0 ? wfExact / wfLabelled : 0;
+const wfAnyRate = wfLabelled > 0 ? wfAnyHit / wfLabelled : 0;
+const wfTop2Rate = wfLabelled > 0 ? wfTop2Hit / wfLabelled : 0;
+
+const summary = {
+  rows: total,
+  labelled,
+  nonEmpty,
+  empty,
+  correct,
+  wrong,
+  precision: Number(precision.toFixed(3)),
+  coverage: Number(coverage.toFixed(3)),
+  mislabelRate: Number(mislabelRate.toFixed(3)),
+  top2Rate: Number(top2Rate.toFixed(3)),
+  centerMin,
+  lead,
+  topicMinScore,
+  baselineSource,
+  mistakes,
+  workflow: {
+    labelled: wfLabelled,
+    nonEmpty: wfNonEmpty,
+    exact: wfExact,
+    anyHit: wfAnyHit,
+    missing: wfMissing,
+    spurious: wfSpurious,
+    precision: Number(wfPrecision.toFixed(3)),
+    recall: Number(wfRecall.toFixed(3)),
+    exactRate: Number(wfExactRate.toFixed(3)),
+    anyRate: Number(wfAnyRate.toFixed(3)),
+    top2Rate: Number(wfTop2Rate.toFixed(3)),
+    mistakes: wfMistakes,
+  },
+};
+
+if (asJson) {
+  console.log(JSON.stringify({ summary, rows: perRow }, null, 2));
+} else {
+  console.log(`  A 表 ${rows.length} 条（topic ${topics.length} / workflow ${rows.length - topics.length}）· 输入 ${total} 行 · 有标签 ${labelled} 行`);
+  console.log(`  判据 centered≥${centerMin} 且 lead≥${lead} · 绝对下限 ${topicMinScore} · baseline ${baselineSource}`);
+  console.log('');
+  console.log(`  非空 ${nonEmpty}/${total} · 命中 ${correct} · 误标 ${wrong} · 空 ${empty}`);
+  console.log(`  精度 ${(precision * 100).toFixed(0)}% · 覆盖率 ${(coverage * 100).toFixed(0)}% · 误标率 ${(mislabelRate * 100).toFixed(0)}% · top-2 ${(top2Rate * 100).toFixed(0)}%`);
+  if (mistakes.length > 0) {
+    console.log('');
+    console.log('  误标：');
+    for (const m of mistakes) console.log(`    turn ${m.turn}: 机器 ${m.got} / 人工 ${m.want}`);
+  }
+  if (wfLabelled > 0) {
+    console.log('');
+    console.log(`  —— workflow（人工标注 ${wfLabelled} 行）——`);
+    console.log(`  机器非空 ${wfNonEmpty} · 集合完全一致 ${wfExact} · 至少一条命中 ${wfAnyHit} · 漏（人工有机器空）${wfMissing} · 多（人工空机器有）${wfSpurious}`);
+    console.log(`  平均精度 ${(wfPrecision * 100).toFixed(0)}% · 平均召回 ${(wfRecall * 100).toFixed(0)}% · 完全一致率 ${(wfExactRate * 100).toFixed(0)}% · 至少一条命中率 ${(wfAnyRate * 100).toFixed(0)}% · top-2 ${(wfTop2Rate * 100).toFixed(0)}%`);
+    if (wfMistakes.length > 0) {
+      console.log('  不一致：');
+      for (const m of wfMistakes) console.log(`    turn ${m.turn}: 机器 [${m.got.join('/')}] / 人工 [${m.want.join('/')}]`);
+    }
+  }
+}
+
+if (minPrecision !== null && precision < minPrecision) {
+  process.exitCode = 1;
+}

package/bin/export-turns.py

@@ -0,0 +1,68 @@
+#!/usr/bin/env python3
+"""Export one DSH session's human prompts as turns.jsonl for rebuild/relabel."""
+import argparse
+import io
+import json
+import sys
+
+import zstandard
+
+
+def text_of(message):
+    parts = []
+    for block in message.get('content', []):
+        if isinstance(block, dict) and block.get('type') == 'text' and isinstance(block.get('text'), str):
+            parts.append(block['text'])
+    return '\n'.join(parts).strip()
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--session-file', required=True)
+    parser.add_argument('--out', required=True)
+    parser.add_argument('--session-id', default=None)
+    args = parser.parse_args()
+    turns = {}
+    current = None
+    with open(args.session_file, 'rb') as fh:
+        with zstandard.ZstdDecompressor().stream_reader(fh) as reader:
+            text = io.TextIOWrapper(reader, encoding='utf-8')
+            for line in text:
+                line = line.strip()
+                if not line:
+                    continue
+                try:
+                    event = json.loads(line)
+                except json.JSONDecodeError:
+                    continue
+                typ = event.get('type')
+                if typ == 'session':
+                    if args.session_id and event.get('id') != args.session_id:
+                        raise SystemExit(f"session id mismatch: {event.get('id')} != {args.session_id}")
+                    continue
+                if typ == 'turn/start':
+                    turn = event.get('data', {}).get('turn')
+                    current = turn if isinstance(turn, int) else None
+                    continue
+                if typ != 'user/message' or current is None:
+                    continue
+                data = event.get('data', {})
+                if data.get('source', {}).get('kind') != 'user':
+                    continue
+                body = text_of(data)
+                if body and current not in turns:
+                    turns[current] = body
+    with open(args.out, 'w', encoding='utf-8', newline='\n') as out:
+        for turn in sorted(turns):
+            out.write(json.dumps({'turn': turn, 'text': turns[turn]}, ensure_ascii=False) + '\n')
+    print(json.dumps({
+        'out': args.out,
+        'turns': len(turns),
+        'firstTurn': min(turns) if turns else None,
+        'lastTurn': max(turns) if turns else None,
+    }, ensure_ascii=False, indent=2))
+    return 0
+
+
+if __name__ == '__main__':
+    sys.exit(main())

package/bin/extract-user-paragraph.mjs

@@ -0,0 +1,14 @@
+import fs from 'node:fs';
+import { createExtractionProviderFromEnv } from '../src/memory/extraction-provider.mjs';
+import { createExtractionPreStep } from '../src/memory/extraction-pre-step.mjs';
+
+const config = 'C:/Users/35794/Documents/DeepSeek\'s Sandbox/coopanion-continuity/backups/continuity-tags-20261031/durable-memory/extraction-live.json';
+const text = '我要的是全自动化。这是你之前自己说的：这一套可以让LLM什么都不额外做，就正常聊天，然后自动系统在用户说到“我们接下来去 *** 吧” （比如跑团），时，就会自动给LLM补充一个“【可回忆】灯塔 / 钥匙 ---> turn49”这类的提示词。这个提示词就是自动算出facet，自动算出高置信度区间，回填turn49，然后LLM读到这个直接去load 49那里对应的seq就能找回原始上下文，然后它就直接想起来了。';
+const provider = createExtractionProviderFromEnv({ DSH_EXTRACTION_MODE: 'model', DSH_EXTRACTION_CONFIG: config });
+const result = await provider.extract({ text });
+const handler = createExtractionPreStep({ provider, mode: 'model' });
+const decision = { kind: 'accept', messages: [{ role: 'user', content: [{ type: 'text', text }] }] };
+const injected = await handler({ step: 1 }, async () => decision);
+fs.writeFileSync('C:/Users/35794/Documents/DeepSeek\'s Sandbox/coopanion-continuity/backups/continuity-tags-20261031/durable-memory/build-a/EXTRACT-PARAGRAPH.json', JSON.stringify({ result, injected }, null, 2));
+console.log(JSON.stringify({ ok: result.ok, latencyMs: result.latencyMs, data: result.data, injected: injected.messages.at(-1)?.content?.[0]?.text }, null, 2));
+provider.close?.();

package/bin/extraction-composite-smoke.mjs

@@ -0,0 +1,28 @@
+import path from 'node:path';
+import { createCommandProvider } from '../src/memory/extraction-provider.mjs';
+import { createCompositeProvider } from '../src/memory/extraction-composite.mjs';
+import { PreTurnController } from '../src/memory/pre-turn.mjs';
+
+const cwd = process.cwd();
+const python311 = path.join(cwd, '.venvs', 'extraction311', 'Scripts', 'python.exe');
+const python38 = path.join(cwd, '.venvs', 'srl38', 'Scripts', 'python.exe');
+const python313 = path.join(cwd, '.venvs', 'relik313', 'Scripts', 'python.exe');
+const globalHf = 'C:\\Users\\35794\\.cache\\huggingface';
+const workHf = path.join(cwd, '..', 'backups', 'continuity-tags-20261031', 'durable-memory', 'hf-cache');
+const nerEnv = { ...process.env, HF_HOME: globalHf, HF_HUB_OFFLINE: '1' };
+const elEnv = { ...process.env, HF_HOME: workHf, HF_MODULES_CACHE: path.join(workHf, 'modules'), HF_HUB_OFFLINE: '1' };
+const provider = createCompositeProvider({
+  ner: createCommandProvider({ command: python311, args: [path.join(cwd, 'bin', 'extraction_service.py')], timeoutMs: 120000, env: nerEnv }),
+  coref: createCommandProvider({ command: python313, args: [path.join(cwd, 'bin', 'coref_service.py')], timeoutMs: 120000 }),
+  srl: createCommandProvider({ command: python38, args: [path.join(cwd, 'bin', 'srl_service.py')], timeoutMs: 180000 }),
+  el: createCommandProvider({ command: python311, args: [path.join(cwd, 'bin', 'relik_service.py')], timeoutMs: 180000, env: elEnv }),
+});
+const controller = new PreTurnController({ records: [] });
+const result = await controller.stepAsync({
+  extractionProvider: provider,
+  text: 'Michael Jordan kicked the ball into the net.',
+  facets: { activity: '对话', topic: '体育' },
+  confidence: 0.8,
+  nowTurn: 1,
+});
+console.log(JSON.stringify({ ok: result.ok, slots: result.slots, predicate: result.predicate, links: result.entityLinks, partial: result.extraction?.data?.partial }, null, 2));

package/bin/extraction-latency.mjs

@@ -0,0 +1,58 @@
+import path from 'node:path';
+import { createPersistentCommandProvider } from '../src/memory/extraction-provider.mjs';
+
+const cwd = process.cwd();
+const python311 = path.join(cwd, '.venvs', 'extraction311', 'Scripts', 'python.exe');
+const python38 = path.join(cwd, '.venvs', 'srl38', 'Scripts', 'python.exe');
+const python313 = path.join(cwd, '.venvs', 'relik313', 'Scripts', 'python.exe');
+const globalHf = 'C:/Users/35794/.cache/huggingface';
+const workHf = path.join(cwd, '..', 'backups', 'continuity-tags-20261031', 'durable-memory', 'hf-cache');
+const services = {
+  ner: {
+    provider: createPersistentCommandProvider({
+      command: python311,
+      args: [path.join(cwd, 'bin', 'extraction_service.py'), '--serve'],
+      timeoutMs: 120000,
+      env: { ...process.env, HF_HOME: globalHf, HF_HUB_OFFLINE: '1' },
+    }),
+    text: '彭小军认为，国内银行现在走的是台湾的发卡模式',
+  },
+  coref: {
+    provider: createPersistentCommandProvider({
+      command: python313,
+      args: [path.join(cwd, 'bin', 'coref_service.py'), '--serve'],
+      timeoutMs: 120000,
+    }),
+    text: 'Coo went to the lamp room. She looked at the coat.',
+  },
+  srl: {
+    provider: createPersistentCommandProvider({
+      command: python38,
+      args: [path.join(cwd, 'bin', 'srl_service.py'), '--serve'],
+      timeoutMs: 180000,
+    }),
+    text: 'The player kicked the ball into the net.',
+  },
+  el: {
+    provider: createPersistentCommandProvider({
+      command: python311,
+      args: [path.join(cwd, 'bin', 'relik_service.py'), '--serve'],
+      timeoutMs: 180000,
+      env: { ...process.env, HF_HOME: workHf, HF_MODULES_CACHE: path.join(workHf, 'modules'), HF_HUB_OFFLINE: '1' },
+    }),
+    text: 'Michael Jordan played in the NBA.',
+    candidates: ['Michael Jordan', 'National Basketball Association'],
+  },
+};
+
+const result = {};
+for (const [name, { provider, text, candidates }] of Object.entries(services)) {
+  const latencies = [];
+  for (let index = 0; index < 3; index += 1) {
+    const response = await provider.extract(candidates ? { text, candidates } : { text });
+    latencies.push({ ok: response.ok, latencyMs: response.latencyMs, reason: response.reason ?? null });
+  }
+  result[name] = { latencies, coldMs: latencies[0]?.latencyMs ?? null, warmMs: latencies.slice(1).map((item) => item.latencyMs) };
+  provider.close();
+}
+console.log(JSON.stringify(result, null, 2));

package/bin/extraction-user-smoke.mjs

@@ -0,0 +1,28 @@
+import fs from 'node:fs';
+import path from 'node:path';
+import { createCommandProvider } from '../src/memory/extraction-provider.mjs';
+import { PreTurnController } from '../src/memory/pre-turn.mjs';
+
+const cwd = process.cwd();
+const base = path.join(cwd, '..', 'backups', 'continuity-tags-20261031', 'durable-memory', 'build-a');
+const provider = createCommandProvider({
+  command: path.join(cwd, '.venvs', 'extraction311', 'Scripts', 'python.exe'),
+  args: [path.join(cwd, 'bin', 'extraction_service.py')],
+  timeoutMs: 60000,
+});
+const controller = new PreTurnController({ records: [] });
+const lines = fs.readFileSync(path.join(cwd, '..', 'backups', 'continuity-tags-20261031', 'online400', 'user-only-messages.jsonl'), 'utf8').split(/\r?\n/).filter(Boolean).slice(0, 3);
+const rows = [];
+for (const line of lines) {
+  const item = JSON.parse(line);
+  const result = await controller.stepAsync({
+    extractionProvider: provider,
+    text: item.text,
+    facets: { activity: '对话', topic: '记忆' },
+    confidence: 0.8,
+    nowTurn: 1,
+  });
+  rows.push({ text: item.text, extraction: result.extraction, gate: result.gate?.action, slots: result.slots });
+}
+fs.writeFileSync(path.join(base, 'USER-SERVICE-SMOKE.json'), JSON.stringify({ rows }, null, 2));
+console.log(JSON.stringify({ n: rows.length, ok: rows.filter((row) => row.extraction?.ok).length }, null, 2));