记忆系统：recall 结果接入 prompt + 可观测性（既有改动，代为提交）

## 说明 **这批改动不是本次会话写的**，它们在会话开始前就已在工作区里、一直未提交。我做的是**验证**它确实成立，然后按你的指示代为提交。出处：`docs/演示用/记忆系统排查报告-2026-09-14.md` 与同目录 `记忆系统修复文档-2026-09-14.md`（两份都在本次一并入库）。排查报告的结论是「记忆系统没有坏」——库里有真实数据、170 条抽取事件全部消费成功；真正的问题是「观测不到」+「召回结果没人消费」。 ## 改动内容（按两份文档的编号） - **F1 `RecalledMemory.content` 断头路**：`base.py` 新增 `memory_context_text()`， `risk_agent._agent_system_prompt` 接收并注入记忆段。无记忆时返回空串，因此 prompt 逐字不变 —— 这也是它能安全接线的理由。 - **F3 `governance.recall` 员工身份恒空**：补一条明确的语义日志。员工身份下召回的是"该用户自身作为客户"的记忆，恒为空属预期，但此前没有任何提示，运维看到 `count=0` 只会以为记忆坏了。 - **F4 可观测性**：`GET /api/v1/users/me/memories`（`stored` / `recalled` / `downstream` / `pending_events` 四段）+ 抽取与召回的 6 处日志 + 三个只读探针 `tools/probe_memory_state.py`、`probe_memory_detail.py`、 `probe_agent_types.py`。 **未实施**（文档明确留作待决，我也不代为决定）：F2 `known` 引用校验永不触发（需架构确认 memory 类 `source_references` 由业务填还是底座统一附加）、 F5 客服是否读写长期记忆（涉脱敏与复核，需产品+合规）。 ## 我做的验证（会话内实测，非照录文档） - 新接口 `GET /users/me/memories` 以 `cust_t` 调用 -> **HTTP 200**： stored: total=2, by_status={'active': 2} recalled: count=2, degraded=False 两条记忆：preference:horizon='约三年'(0.95)、preference:risk_level='稳健型'(0.98) 与排查报告 §〇列出的那两条**完全吻合**。 - `pytest tests/unit tests/contract` 全绿（这批改动没有破坏既有测试）。 ## 未验证的部分 `memory_context_text()` 接进 prompt 后的**端到端效果没有实测** —— 文档自己说明了原因：当前 `risk` Agent 的召回恒空（员工身份不是客户），所以接线后行为不变，要用测试替身才能验证注入。我没有为此编造证据。
2026-09-14 20:35:46 +08:00
parent c868f01e67
commit c0e5c80929
10 changed files with 537 additions and 10 deletions
@@ -0,0 +1,69 @@
+"""按 agent_type 统计运行分布 + 记忆写入时间点的对照（只读）。"""
+
+from __future__ import annotations
+
+import asyncio
+from typing import Any
+
+from sqlalchemy import text
+
+from app.infrastructure.db import SessionFactory
+
+
+async def q(session: Any, sql: str) -> list[Any]:
+    try:
+        return list((await session.execute(text(sql))).all())
+    except Exception as exc:
+        print(f"    ERROR {type(exc).__name__}: {exc}")
+        return []
+
+
+async def main() -> None:
+    async with SessionFactory() as session:
+        print("=" * 66)
+        print("A) agent_run 按 agent_type 分布")
+        print("=" * 66)
+        for r in await q(session, """
+            SELECT agent_type, status, COUNT(*) c, MIN(created_at), MAX(created_at)
+            FROM agent_run GROUP BY agent_type, status ORDER BY c DESC
+        """):
+            print(f"  {str(r[0]):<20} {str(r[1]):<12} {r[2]:<5}  {r[3]} ~ {r[4]}")
+
+        print()
+        print("=" * 66)
+        print("B) 记忆写入时刻前后，有没有非客服的 run")
+        print("=" * 66)
+        for r in await q(session, """
+            SELECT r.agent_type, r.status, r.created_at, r.run_id
+            FROM agent_run r
+            WHERE r.created_at < '2026-09-10 14:00:00'
+            ORDER BY r.created_at DESC LIMIT 12
+        """):
+            print(f"  {r[2]} | {str(r[0]):<20} | {str(r[1]):<10} | {str(r[3])[:8]}")
+
+        print()
+        print("=" * 66)
+        print("C) 记忆抽取事件与 run 的对应（最近 8 条 extraction 事件）")
+        print("=" * 66)
+        for r in await q(session, """
+            SELECT e.aggregate_id, e.status, e.retry_count, e.occurred_at, r.agent_type
+            FROM domain_event_outbox e
+            LEFT JOIN agent_run r ON r.run_id = e.aggregate_id
+            WHERE e.event_type = 'memory.extraction_requested'
+            ORDER BY e.id DESC LIMIT 8
+        """):
+            print(f"  {r[3]} | run={str(r[0])[:8]} | {r[1]:<9} | agent={r[4]}")
+
+        print()
+        print("=" * 66)
+        print("D) 客户 9001 的角色")
+        print("=" * 66)
+        for r in await q(session, """
+            SELECT r.role_code FROM sys_user_role ur
+            JOIN sys_role r ON r.id = ur.role_id WHERE ur.user_id = 9001
+        """):
+            print(f"  {r[0]}")
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
@@ -0,0 +1,101 @@
+"""记忆系统明细探针（只读）：回答「写入了什么 / 今天有没有活动 / 为什么 dead」。"""
+
+from __future__ import annotations
+
+import asyncio
+from typing import Any
+
+from sqlalchemy import text
+
+from app.infrastructure.db import SessionFactory
+
+
+async def q(session: Any, sql: str, **params: Any) -> list[Any]:
+    try:
+        return list((await session.execute(text(sql), params)).all())
+    except Exception as exc:
+        print(f"    ERROR {type(exc).__name__}: {exc}")
+        return []
+
+
+async def main() -> None:
+    async with s_factory() as session:
+        print("=" * 66)
+        print("A) memory_unit 全部内容（写入是否成功的直接证据）")
+        print("=" * 66)
+        rows = await q(session, """
+            SELECT id, customer_id, memory_key, content, memory_type, status,
+                   confidence, source_type, evidence_count, created_at, updated_at
+            FROM memory_unit ORDER BY updated_at DESC LIMIT 20
+        """)
+        if not rows:
+            print("    (memory_unit 为空 —— 一条记忆都没写入)")
+        for r in rows:
+            print(f"  id={r[0]} customer={r[1]} [{r[5]}]")
+            print(f"    key   = {r[2]}")
+            print(f"    value = {r[3]!r}")
+            print(f"    type={r[4]} conf={r[6]} src={r[7]} evidence={r[8]}")
+            print(f"    created={r[9]}  updated={r[10]}")
+
+        print()
+        print("=" * 66)
+        print("B) 今天(2026-09-14)有没有任何新活动")
+        print("=" * 66)
+        for label, sql in (
+            ("agent_run 今天创建", "SELECT COUNT(*) FROM agent_run WHERE created_at >= '2026-09-14'"),
+            ("agent_run 最新一条", "SELECT MAX(created_at) FROM agent_run"),
+            ("conversation_message 最新", "SELECT MAX(created_at) FROM conversation_message"),
+            ("domain_event_outbox 最新", "SELECT MAX(created_at) FROM domain_event_outbox"),
+            ("outbox pending 总数", "SELECT COUNT(*) FROM domain_event_outbox WHERE status='pending'"),
+            ("outbox failed 总数", "SELECT COUNT(*) FROM domain_event_outbox WHERE status='failed'"),
+        ):
+            rows = await q(session, sql)
+            print(f"  {label:<26} {rows[0][0] if rows else '?'}")
+
+        print()
+        print("=" * 66)
+        print("C) 最近 10 条 outbox 事件（看最新动向）")
+        print("=" * 66)
+        rows = await q(session, """
+            SELECT event_type, status, retry_count, last_error, occurred_at
+            FROM domain_event_outbox ORDER BY id DESC LIMIT 10
+        """)
+        for r in rows:
+            err = (str(r[3])[:70] + "...") if r[3] and len(str(r[3])) > 70 else r[3]
+            print(f"  {r[4]} | {r[0]:<34} | {r[1]:<9} | retry={r[2]}")
+            if err:
+                print(f"      last_error: {err}")
+
+        print()
+        print("=" * 66)
+        print("D) agent.run_requested 死信的原因分布（取 5 条样本）")
+        print("=" * 66)
+        rows = await q(session, """
+            SELECT retry_count, last_error, occurred_at
+            FROM domain_event_outbox
+            WHERE event_type='agent.run_requested' AND status='dead'
+            ORDER BY id DESC LIMIT 5
+        """)
+        for r in rows:
+            err = (str(r[1])[:150] + "...") if r[1] and len(str(r[1])) > 150 else r[1]
+            print(f"  {r[2]} retry={r[0]}")
+            print(f"      {err}")
+
+        print()
+        print("=" * 66)
+        print("E) 最近 5 次 agent_run 的状态")
+        print("=" * 66)
+        rows = await q(session, """
+            SELECT run_id, agent_type, status, error_code, created_at
+            FROM agent_run ORDER BY id DESC LIMIT 5
+        """)
+        for r in rows:
+            print(f"  {r[4]} | {r[1]:<18} | {r[2]:<10} | err={r[3]} | {r[0][:8]}")
+
+
+def s_factory():
+    return SessionFactory()
+
+
+if __name__ == "__main__":
+    asyncio.run(main())
@@ -0,0 +1,121 @@
+"""记忆系统状态探针（只读）：一次看清「库里有没有」「事件消费到哪了」。
+
+用途：排查"记忆系统是否真的在工作"时，先跑这个脚本拿到事实基线，
+再去看日志定位是哪一段断了。全部是 SELECT，不写任何数据。
+
+用法：
+    D:/conda/envs/jr_py313/python.exe tools/probe_memory_state.py [customer_id]
+"""
+
+from __future__ import annotations
+
+import asyncio
+import sys
+from typing import Any
+
+from sqlalchemy import text
+
+from app.infrastructure.db import SessionFactory
+
+TABLES = (
+    "memory_unit",
+    "memory_evidence",
+    "memory_conflict",
+    "memory_sync_outbox",
+    "user_facts",
+    "profile_snapshots",
+)
+
+
+async def _scalar(session: Any, sql: str, **params: Any) -> Any:
+    try:
+        return await session.scalar(text(sql), params)
+    except Exception as exc:  # 表不存在/连不上库都要能看，而不是整个脚本崩掉
+        return f"ERROR {type(exc).__name__}: {exc}"
+
+
+async def main() -> None:
+    customer_id = int(sys.argv[1]) if len(sys.argv) > 1 else None
+    async with SessionFactory() as session:
+        print("=" * 60)
+        print("1) 记忆相关表的总行数")
+        print("=" * 60)
+        for table in TABLES:
+            print(f"  {table:<22} {await _scalar(session, f'SELECT COUNT(*) FROM {table}')}")
+
+        print()
+        print("=" * 60)
+        print("2) memory_unit 按状态分组")
+        print("=" * 60)
+        try:
+            rows = (await session.execute(
+                text("SELECT status, COUNT(*) c FROM memory_unit GROUP BY status"))).all()
+            for row in rows:
+                print(f"  {row[0]:<12} {row[1]}")
+        except Exception as exc:
+            print(f"  ERROR {type(exc).__name__}: {exc}")
+
+        print()
+        print("=" * 60)
+        print("3) domain_event_outbox 按 (event_type, status) 分组")
+        print("   —— 这里能看到事件是不是堆在 pending（= Worker 没跑 / 处理失败）")
+        print("=" * 60)
+        try:
+            rows = (await session.execute(text(
+                "SELECT event_type, status, COUNT(*) c, MAX(occurred_at) latest "
+                "FROM domain_event_outbox GROUP BY event_type, status ORDER BY c DESC"
+            ))).all()
+            if not rows:
+                print("  (空)")
+            for row in rows:
+                print(f"  {row[0]:<38} {row[1]:<10} {row[2]:<6} latest={row[3]}")
+        except Exception as exc:
+            print(f"  ERROR {type(exc).__name__}: {exc}")
+
+        if customer_id is not None:
+            print()
+            print("=" * 60)
+            print(f"4) 客户 {customer_id} 的记忆明细")
+            print("=" * 60)
+            try:
+                rows = (await session.execute(text(
+                    "SELECT memory_uuid, memory_key, content, memory_type, status, "
+                    "confidence, source_type, updated_at "
+                    "FROM memory_unit WHERE customer_id=:cid ORDER BY updated_at DESC LIMIT 20"
+                ), {"cid": customer_id})).all()
+                if not rows:
+                    print("  (该客户没有任何 memory_unit 行)")
+                for row in rows:
+                    print(f"  [{row[4]}] {row[1]} = {row[2]!r}")
+                    print(f"      type={row[3]} conf={row[5]} src={row[6]} at={row[7]}")
+            except Exception as exc:
+                print(f"  ERROR {type(exc).__name__}: {exc}")
+
+            print()
+            print(f"  user_facts:")
+            try:
+                rows = (await session.execute(text(
+                    "SELECT fact_key, confidence FROM user_facts "
+                    "WHERE customer_id=:cid LIMIT 20"), {"cid": customer_id})).all()
+                for row in rows:
+                    print(f"      {row[0]} (conf={row[1]})")
+                if not rows:
+                    print("      (无)")
+            except Exception as exc:
+                print(f"      ERROR {type(exc).__name__}: {exc}")
+
+            print(f"  profile_snapshots:")
+            try:
+                rows = (await session.execute(text(
+                    "SELECT version, is_current, generated_at FROM profile_snapshots "
+                    "WHERE customer_id=:cid ORDER BY id DESC LIMIT 5"), {"cid": customer_id})).all()
+                for row in rows:
+                    print(f"      v{row[0]} current={row[1]} at={row[2]}")
+                if not rows:
+                    print("      (无)")
+            except Exception as exc:
+                print(f"      ERROR {type(exc).__name__}: {exc}")
+
+
+if __name__ == "__main__":
+    asyncio.run(main())