feat(chat-intent): observe routing, keep discuss-first on chat, and escalate short follow-ups

Close the chat-session dead zone without flipping production flags: record 0.72 fallbacks, stop C-2 from treating「先聊聊」as execute, and promote「不对你去执行」after a completed direct turn on both h5direct and date sessions.

Co-authored-by: Cursor <cursoragent@cursor.com>

scripts/report-chat-intent-baseline.mjs

@@ -0,0 +1,93 @@
+#!/usr/bin/env node
+/**
+ * Stage A baseline for chat-task intent routing.
+ * Reads existing agent-run events; does not change routing.
+ *
+ * Usage:
+ *   node scripts/report-chat-intent-baseline.mjs
+ *   node scripts/report-chat-intent-baseline.mjs --hours 48
+ */
+import { loadH5Environment } from './load-env.mjs';
+import { createDbPool } from '../db.mjs';
+import {
+  aggregateChatIntentObservations,
+  observationFromIntentRoutedData,
+} from '../chat-intent-observation.mjs';
+
+loadH5Environment(import.meta.dirname);
+
+function parseHours(argv) {
+  const index = argv.indexOf('--hours');
+  if (index < 0) return 24;
+  const parsed = Number(argv[index + 1]);
+  return Number.isFinite(parsed) && parsed > 0 ? parsed : 24;
+}
+
+function parseJson(value) {
+  if (!value) return {};
+  if (typeof value === 'object') return value;
+  try {
+    return JSON.parse(String(value));
+  } catch {
+    return {};
+  }
+}
+
+async function main() {
+  const hours = parseHours(process.argv.slice(2));
+  const sinceMs = Date.now() - hours * 60 * 60 * 1000;
+  const pool = await createDbPool();
+  try {
+    const [rows] = await pool.query(
+      `SELECT event_type, data_json
+         FROM h5_agent_run_events
+        WHERE created_at >= ?
+          AND event_type IN (
+            'intent_routed',
+            'direct_session_escalated_to_deep_reasoning',
+            'direct_escalation_context_injected'
+          )
+        ORDER BY created_at ASC`,
+      [sinceMs],
+    );
+    const routed = [];
+    let escalatedCount = 0;
+    let contextInjectedCount = 0;
+    for (const row of rows ?? []) {
+      if (row.event_type === 'direct_session_escalated_to_deep_reasoning') {
+        escalatedCount += 1;
+        continue;
+      }
+      if (row.event_type === 'direct_escalation_context_injected') {
+        contextInjectedCount += 1;
+        continue;
+      }
+      routed.push(observationFromIntentRoutedData(parseJson(row.data_json)));
+    }
+    const summary = aggregateChatIntentObservations({
+      routed,
+      escalatedCount,
+      contextInjectedCount,
+    });
+    console.log(`==> chat intent baseline last ${hours}h`);
+    console.log(JSON.stringify(summary, null, 2));
+    if (summary.total === 0) {
+      console.log('NOTE  窗口内没有 intent_routed。阶段 A 需要先有聊天流量。');
+    } else {
+      console.log(
+        `NOTE  0.72 兜底 ${summary.ruleFallbackCount}/${summary.total}`
+        + ` (${summary.ruleFallbackRatio});`
+        + ` 升级 ${summary.escalatedCount};`
+        + ` 记忆召回仍走 direct ${summary.memoryRecallDirectRatio};`
+        + ` shadow wouldChangeRoute ${summary.wouldChangeRoute}/${summary.shadowObserved}`,
+      );
+    }
+  } finally {
+    await pool.end();
+  }
+}
+
+main().catch((err) => {
+  console.error(err instanceof Error ? err.message : err);
+  process.exitCode = 1;
+});

scripts/run-memind-tests.mjs

@@ -120,6 +120,8 @@ const SCOPE_RULES = [
       /session-snapshot/i,
       /direct-chat-service/i,
       /chat-intent-router/i,
+      /chat-intent-observation/i,
+      /chat-task-intent/i,
       /scripts\/check-memory-v2/i,
       /scripts\/smoke-memory-v2/i,
     ],
@@ -131,6 +133,8 @@ const SCOPE_RULES = [
       'episodic-memory.test.mjs',
       'direct-chat-service.test.mjs',
       'chat-intent-router.test.mjs',
+      'chat-intent-observation.test.mjs',
+      'chat-task-intent-config.test.mjs',
     ],
   },
   {

scripts/verify-chat-intent-followup.mjs

@@ -0,0 +1,195 @@
+#!/usr/bin/env node
+/**
+ * Local Stage D check: casual chat then a short "go do it" follow-up.
+ * Does not require Goose to finish; waits for intent_routed only.
+ *
+ *   H5_PORT=8086 node scripts/verify-chat-intent-followup.mjs
+ */
+import crypto from 'node:crypto';
+import { loadH5Environment } from './load-env.mjs';
+import { createDbPool } from '../db.mjs';
+import { createUserAuth, USER_COOKIE } from '../user-auth.mjs';
+import { CHAT_INTENT_ROUTE } from '../chat-intent-router.mjs';
+import { isDirectFollowupEscalateEnabledForUser, resolveDirectFollowupEscalatePolicy } from '../chat-task-intent-config.mjs';
+
+loadH5Environment(import.meta.dirname);
+
+const PORTAL = `http://127.0.0.1:${process.env.H5_PORT ?? 8081}`;
+const USERNAME = process.env.MEMIND_E2E_USERNAME ?? 'john';
+const PASSWORD = process.env.JOHN_PASSWORD ?? process.env.H5_ACCESS_PASSWORD ?? '888888';
+const DIRECT_MESSAGE = '我想了解车载冰箱，MPV 用，宽度不超过 60cm，先聊聊选购要点';
+const FOLLOWUP_MESSAGE = '不对，你去执行';
+const EVENT_WAIT_MS = 90_000;
+
+const issues = [];
+
+function pass(label, detail = '') {
+  console.log(`✔ ${label}${detail ? `: ${detail}` : ''}`);
+}
+
+function fail(label, detail = '') {
+  issues.push({ label, detail });
+  console.error(`✘ ${label}${detail ? `: ${detail}` : ''}`);
+}
+
+function parseEventData(raw) {
+  if (!raw) return null;
+  if (typeof raw === 'object') return raw;
+  try {
+    return JSON.parse(String(raw));
+  } catch {
+    return null;
+  }
+}
+
+async function loginViaApi() {
+  const response = await fetch(`${PORTAL}/auth/login`, {
+    method: 'POST',
+    headers: { 'Content-Type': 'application/json' },
+    body: JSON.stringify({ username: USERNAME, password: PASSWORD }),
+  });
+  const body = await response.json().catch(() => ({}));
+  if (response.ok && body?.authenticated) {
+    const setCookie = response.headers.getSetCookie?.() ?? [];
+    const cookieLine = setCookie.find((line) => line.startsWith(`${USER_COOKIE}=`))
+      ?? response.headers.get('set-cookie');
+    const match = String(cookieLine ?? '').match(new RegExp(`${USER_COOKIE}=([^;]+)`));
+    if (match?.[1]) {
+      return { token: decodeURIComponent(match[1]), userId: body.user?.id ?? null };
+    }
+  }
+  const pool = await createDbPool();
+  const auth = createUserAuth(pool);
+  const result = await auth.login({ username: USERNAME, password: PASSWORD, ip: '127.0.0.1' });
+  await pool.end();
+  if (!result.ok || !result.token) {
+    throw new Error(`登录失败: ${result.message ?? 'unknown'}`);
+  }
+  return { token: result.token, userId: result.user?.id ?? null };
+}
+
+async function createRun(token, { message, sessionId = null }) {
+  const response = await fetch(`${PORTAL}/api/agent/runs`, {
+    method: 'POST',
+    headers: {
+      'Content-Type': 'application/json',
+      Cookie: `${USER_COOKIE}=${token}`,
+    },
+    body: JSON.stringify({
+      request_id: crypto.randomUUID(),
+      session_id: sessionId,
+      user_message: {
+        id: crypto.randomUUID(),
+        role: 'user',
+        content: [{ type: 'text', text: message }],
+        metadata: { displayText: message, userVisible: true },
+      },
+    }),
+  });
+  const payload = await response.json().catch(() => ({}));
+  if (!response.ok) {
+    throw new Error(`POST /api/agent/runs ${response.status}: ${JSON.stringify(payload)}`);
+  }
+  const run = payload.run ?? payload;
+  return { runId: run.id, sessionId: run.sessionId ?? run.agent_session_id ?? sessionId ?? null };
+}
+
+async function getRun(token, runId) {
+  const response = await fetch(`${PORTAL}/api/agent/runs/${runId}`, {
+    headers: { Cookie: `${USER_COOKIE}=${token}` },
+  });
+  const payload = await response.json().catch(() => ({}));
+  return payload.run ?? payload;
+}
+
+async function readRunEvents(runId) {
+  const pool = await createDbPool();
+  const [rows] = await pool.query(
+    `SELECT event_type, data_json FROM h5_agent_run_events WHERE run_id = ? ORDER BY created_at ASC`,
+    [runId],
+  );
+  await pool.end();
+  return rows;
+}
+
+async function waitForRunEvents(token, runId, { requireCompleted = false } = {}) {
+  const started = Date.now();
+  while (Date.now() - started < EVENT_WAIT_MS) {
+    const run = await getRun(token, runId);
+    const events = await readRunEvents(runId);
+    const routed = events.find((row) => row.event_type === 'intent_routed');
+    const completed = events.find((row) => row.event_type === 'direct_chat_completed');
+    const sessionId = parseEventData(completed?.data_json)?.sessionId
+      ?? run?.sessionId
+      ?? run?.agent_session_id
+      ?? null;
+    if (routed && sessionId && (!requireCompleted || completed || ['succeeded', 'failed'].includes(run?.status))) {
+      return {
+        run,
+        events,
+        routed: parseEventData(routed.data_json),
+        sessionId,
+      };
+    }
+    await new Promise((resolve) => setTimeout(resolve, 1000));
+  }
+  throw new Error(`run ${runId} 未在 ${EVENT_WAIT_MS}ms 内拿到路由/session`);
+}
+
+async function main() {
+  const health = await fetch(`${PORTAL}/auth/status`).catch(() => null);
+  if (!health?.ok) throw new Error(`Portal 不可用: ${PORTAL}/auth/status`);
+  pass('Portal', PORTAL);
+
+  const { token, userId } = await loginViaApi();
+  pass('登录', `${USERNAME} (${userId ?? 'unknown'})`);
+
+  const policy = resolveDirectFollowupEscalatePolicy(process.env);
+  if (!isDirectFollowupEscalateEnabledForUser(userId, policy)) {
+    fail('阶段 D flag', `用户 ${userId} 未命中 DIRECT_FOLLOWUP_ESCALATE canary`);
+  } else {
+    pass('阶段 D flag', 'enabled for user');
+  }
+
+  const direct = await createRun(token, { message: DIRECT_MESSAGE });
+  pass('direct 提交', direct.runId);
+  const directResult = await waitForRunEvents(token, direct.runId, { requireCompleted: true });
+  if (directResult.routed?.route !== CHAT_INTENT_ROUTE.DIRECT_CHAT) {
+    fail('direct 路由', String(directResult.routed?.route ?? 'missing'));
+  } else {
+    pass('direct 路由', directResult.routed.reason ?? 'direct_chat');
+  }
+  const sessionId = directResult.sessionId;
+  if (!sessionId) fail('direct session', 'missing');
+  else pass('direct session', sessionId);
+
+  const followup = await createRun(token, { message: FOLLOWUP_MESSAGE, sessionId });
+  pass('followup 提交', followup.runId);
+  const followupResult = await waitForRunEvents(token, followup.runId);
+  const routed = followupResult.routed;
+  if (routed?.route !== CHAT_INTENT_ROUTE.AGENT) {
+    fail('阶段 D 路由', `期望 agent，实际 ${routed?.route ?? 'missing'} / ${routed?.reason ?? ''}`);
+  } else if (!String(routed.reason ?? '').includes('短句纠偏')) {
+    fail('阶段 D 原因', routed.reason ?? 'missing');
+  } else {
+    pass('阶段 D 路由', routed.reason);
+  }
+  const escalated = followupResult.events.some((row) =>
+    row.event_type === 'direct_session_escalated_to_deep_reasoning');
+  const isH5Direct = String(sessionId ?? '').startsWith('h5direct_');
+  if (escalated) pass('升级事件', 'direct_session_escalated_to_deep_reasoning');
+  else if (!isH5Direct) pass('升级事件', '日期 session 同会话升级，无需换 Goose session');
+  else fail('升级事件', '未看到 direct session 升级');
+
+  console.log('\n=== 汇总 ===');
+  if (issues.length) {
+    for (const item of issues) console.log(`  - ${item.label}: ${item.detail}`);
+    process.exit(1);
+  }
+  console.log('chat intent followup 本地验收通过');
+}
+
+main().catch((err) => {
+  console.error(err instanceof Error ? err.message : err);
+  process.exit(1);
+});