比价记录改后端 harvest: 透传壳落库替代客户端主动 POST

compare.py 透传壳新增后端 harvest(不再依赖客户端 POST /compare/record):
- 帧0(客户端首帧不带 trace_id)由 app-server 用 uuid 签发 trace_id、注入转发 body、
  回填响应顶层,并按 trace_id 建 running 行(harvest_running)
- done 帧(command=done 且 continue=false)更新为 success/failed 终态,并对本次新落成
  success 幂等发一次邀请奖(harvest_done + try_reward_on_compare)
- /trace/finalize(用户终止/Phase1 未识别,无 done)更新为 cancelled/failed,
  不降级已 success(harvest_abort)
- 软鉴权 OptionalUser(deps.py 新增 get_current_user_optional):新客户端带 JWT 绑
  user_id,老客户端匿名建行、user_id 暂空,由其后续 POST /compare/record 按 trace_id
  reconcile;新版覆盖够高后可收紧成硬鉴权
- 结构化日志(logging.py 加 trace_id_ctx ContextVar):请求级 trace_id 贯穿 harvest 各阶段

comparison_record 表结构调整(迁移 comparison_record_trace_unique):
- 唯一键 (user_id,trace_id) → trace_id 单列(harvest 帧0 建行时 user_id 可能暂缺)
- user_id 改可空

repositories/comparison.py 加 harvest_running/harvest_done/harvest_abort;
compare_record.py 的 POST /record 降级为灰度期老客户端兼容(按 trace_id 幂等、不降级 success);
补 tests/test_compare_harvest.py,test_compare_proxy/milestone 适配。

文档同步: docs/database/comparison_record.md 更新表结构(唯一键/user_id 可空/status 取值)
与写入模型; docs/api/compare/compare-record-report.md 标注 POST /record 灰度定位。

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

app/api/deps.py

@@ -54,5 +54,29 @@ def get_current_user(
     return user
 
 
+def get_current_user_optional(
+    credentials: Annotated[HTTPAuthorizationCredentials | None, Depends(_bearer)],
+    db: Annotated[Session, Depends(get_db)],
+) -> User | None:
+    """软鉴权:有合法 Bearer 就返回 user,否则(无 header / 无效 token / 用户禁用)一律返回
+    None,**不 raise**。
+
+    比价 step / finalize 灰度期用:新客户端带 JWT → 拿到 user_id 绑定 harvest 行;
+    老客户端(不带 JWT)→ None,harvest 行 user_id 暂空,由其后续 /compare/record 上报补齐。
+    等新版覆盖率够高,再把这几条端点从软鉴权收紧成硬 get_current_user。
+    """
+    if credentials is None or credentials.scheme.lower() != "bearer":
+        return None
+    try:
+        payload = decode_token(credentials.credentials, expected_type="access")
+        user = db.get(User, int(payload["sub"]))
+    except (TokenError, KeyError, ValueError, TypeError):
+        return None
+    if user is None or user.status != "active":
+        return None
+    return user
+
+
 CurrentUser = Annotated[User, Depends(get_current_user)]
+OptionalUser = Annotated[User | None, Depends(get_current_user_optional)]
 DbSession = Annotated[Session, Depends(get_db)]

app/api/v1/compare.py

@@ -1,49 +1,121 @@
-"""外卖比价业务透传端点。
+"""外卖比价业务透传端点 + 后端 harvest 落库。
 
-把客户端的 POST /api/v1/intent/recognize 和 /api/v1/price/step 请求原样转发给
-pricebot-backend 的 /api/intent/recognize 和 /api/price/step。MVP 阶段**不鉴权**
-(同 coupon/step,见 docs/待办与技术债.md P1:device_id 透传区分设备,待补 JWT +
-device_id↔user_id 绑定后才能做用户级画像)。
+把客户端 POST /api/v1/intent/* 和 /api/v1/price/step 转发给 pricebot,同时**由 app-server
+在透传里直接落库比价记录**(不再靠客户端 POST /compare/record):
+  - 帧0(客户端首帧不带 trace_id)→ app-server 用 uuid 签发 trace_id、注入转发 body、回给
+    客户端;并按 trace_id 建 running 行(harvest_running)。
+  - 最终 done 帧 → 更新成 success/failed + 结果(harvest_done)+ **幂等发一次邀请奖**。
+  - /trace/finalize(用户终止/Phase1 未识别,无 done)→ 更新成 cancelled/failed
+    (harvest_abort,**不降级 success**)。
+trace_url 从 pricebot 响应**顶层 trace_url** 取(pricebot 每帧都带,见其 goal_engine.process_step)。
 
-- Phase 1 /intent/recognize:从源平台(淘宝闪购/美团/京东外卖)购物车页识别店名+菜品+
-  价格,一次性。
-- Phase 2 /price/step:多轮循环,在目标平台复现订单读到手价,直到 done。
+鉴权:软鉴权(OptionalUser)——新客户端带 JWT → 绑 user_id;老客户端不带 → user_id 暂空,
+由其后续 /compare/record 上报补(灰度期两条写路径按 trace_id reconcile,success 不被降级)。
 
-真正的目标驱动比价逻辑在 pricebot-backend(GoalEngine,另一个 repo),本接口只是透传壳。
-电商(ecom)那两个端点 food MVP 暂不需要,以后放开 scene 时再加 /ecom/intent/recognize
-和 /ecom/step 两行即可。
-
-pricebot 协议文档:
-  pricebot-backend/docs/main/02_api_protocol.md
+真正的目标驱动比价逻辑在 pricebot-backend(GoalEngine,另一个 repo),本接口是透传壳 + 落库。
+pricebot 协议文档: pricebot-backend/docs/main/02_api_protocol.md
 """
 from __future__ import annotations
 
 import json
 import logging
+import time
+import uuid
 from typing import Any
 
 import httpx
 from fastapi import APIRouter, HTTPException, Request, status
+from fastapi.concurrency import run_in_threadpool
 
+from app.api.deps import OptionalUser
 from app.core.config import settings
+from app.core.logging import trace_id_ctx
 from app.core.pricebot_client import get_pricebot_client
 from app.core.pricebot_router import pick_pricebot
+from app.db.session import SessionLocal
+from app.repositories import comparison as crud_compare
+from app.repositories import invite as crud_invite
 
 logger = logging.getLogger("shagua.compare")
 
 router = APIRouter(prefix="/api/v1", tags=["compare"])
 
 
-async def _passthrough(request: Request, upstream_path: str) -> dict[str, Any]:
-    """把请求体原样转发给 pricebot-backend 的 {upstream_path}。
+# ============================================================
+# harvest 落库(阻塞 SQLAlchemy → run_in_threadpool,独立 SessionLocal,不阻塞事件循环;
+# 任何写库异常都吞掉、绝不连累比价透传返回 —— 同 coupon.py 现有 best-effort 写法)。
+# ============================================================
 
-    跟 coupon_step 同款透传壳:不鉴权、不做 schema 校验,仅读 device_id/trace_id/step
-    打日志。比价单帧是大上下文 / 逐帧 LLM,超时用 PRICEBOT_COMPARE_TIMEOUT_SEC(60s,
-    比领券的 30s 长)。
+
+def _harvest_running_blocking(
+    trace_id: str, user_id: int | None, business_type: str,
+    device_id: str | None, device_info: dict | None, trace_url: str | None,
+) -> None:
+    with SessionLocal() as db:
+        crud_compare.harvest_running(
+            db, trace_id=trace_id, user_id=user_id, business_type=business_type,
+            device_id=device_id, device_info=device_info, trace_url=trace_url,
+        )
+    logger.info(
+        "harvest running row (user=%s)", user_id,
+        extra={"phase": "harvest_running", "status": "running", "user_id": user_id},
+    )
+
+
+def _harvest_done_blocking(
+    trace_id: str, user_id: int | None, done_params: dict, business_type: str,
+    device_id: str | None, device_info: dict | None, trace_url: str | None,
+) -> None:
+    with SessionLocal() as db:
+        rec, newly_success = crud_compare.harvest_done(
+            db, trace_id=trace_id, user_id=user_id, done_params=done_params,
+            business_type=business_type, device_id=device_id,
+            device_info=device_info, trace_url=trace_url,
+        )
+        logger.info(
+            "harvest done → %s saved=%s newly=%s", rec.status,
+            rec.saved_amount_cents, newly_success,
+            extra={"phase": "harvest_done", "status": rec.status,
+                   "saved_cents": rec.saved_amount_cents,
+                   "best_platform": rec.best_platform_id, "newly_success": newly_success,
+                   "user_id": user_id},
+        )
+        # 本次**新**落成 success + 有登录用户 → 发一次邀请奖(try_reward 自身幂等,双保险)。
+        if newly_success and user_id is not None:
+            try:
+                crud_invite.try_reward_on_compare(db, user_id)
+                logger.info(
+                    "invite reward fired (user=%s)", user_id,
+                    extra={"phase": "reward", "user_id": user_id},
+                )
+            except Exception as e:  # noqa: BLE001 发奖失败不连累比价
+                logger.warning("harvest invite reward failed user=%s: %s", user_id, e)
+
+
+def _harvest_abort_blocking(
+    trace_id: str, status_hint: str, reason: str | None, trace_url: str | None,
+) -> None:
+    with SessionLocal() as db:
+        rec = crud_compare.harvest_abort(
+            db, trace_id=trace_id, status=status_hint, reason=reason, trace_url=trace_url,
+        )
+    logger.info(
+        "harvest abort → %s", (rec.status if rec else "no-row"),
+        extra={"phase": "harvest_abort",
+               "status": (rec.status if rec else None), "reason": reason},
+    )
+
+
+async def _forward(
+    request: Request, upstream_path: str, user, *, harvest_first_frame: bool = True,
+) -> tuple[dict[str, Any], str, dict[str, Any]]:
+    """透传给 pricebot 的 {upstream_path},并 mint trace_id + 首帧建 running 行。
+    返回 (resp_json, trace_id, meta)。
+
+    - 客户端首帧不带 trace_id → app-server 用 uuid 签发、写进转发 body、回填进响应顶层
+      `trace_id`(客户端据此拿到后续帧都带上)。带了(老客户端 / 后续帧)→ 原样用、走原始 bytes 快路。
+    - 仅**本次 mint(=首帧)**建 running 行(idempotent);后续帧不再写库。
     """
-    # 读原始字节,避免"反序列化→再序列化"的双重 JSON(省 ~一半透传 CPU,让 app-server
-    # 单 worker 也扛得住高并发)。只 json.loads 一次拿 trace_id 做亲和 + 打日志,转发时
-    # 直接发原始 bytes(content=raw),不重新 dumps。
     raw = await request.body()
     try:
         meta = json.loads(raw)
@@ -52,26 +124,41 @@ async def _passthrough(request: Request, upstream_path: str) -> dict[str, Any]:
     if not isinstance(meta, dict):
         meta = {}
 
-    # 按 trace_id 一致性 hash 选 pricebot 实例(同一比价的所有帧落同一进程,内存维护 state)
-    base = pick_pricebot(meta.get("trace_id"))
+    trace_id = meta.get("trace_id")
+    minted = False
+    if not trace_id:
+        trace_id = str(uuid.uuid4())
+        meta["trace_id"] = trace_id
+        raw = json.dumps(meta).encode()  # 仅首帧重新序列化(注入 trace_id);后续帧走原始 bytes
+        minted = True
+
+    # 请求级 trace_id 贯穿:此后本请求(含 run_in_threadpool 里的 harvest)每行日志自动带 trace_id
+    trace_id_ctx.set(trace_id)
+
+    base = pick_pricebot(trace_id)
     url = f"{base.rstrip('/')}{upstream_path}"
     timeout = settings.PRICEBOT_COMPARE_TIMEOUT_SEC
+    step = meta.get("step")
 
     logger.info(
-        "compare %s device_id=%s trace_id=%s step=%s",
-        upstream_path,
-        meta.get("device_id"),
-        meta.get("trace_id"),
-        meta.get("step"),
+        "→ pricebot %s step=%s%s", upstream_path, step, " [mint]" if minted else "",
+        extra={"phase": "forward", "endpoint": upstream_path, "step": step,
+               "device_id": meta.get("device_id"), "minted": minted,
+               "query": meta.get("query"), "user_id": (user.id if user else None)},
     )
 
+    t0 = time.monotonic()
     try:
         client = get_pricebot_client()
         resp = await client.post(
             url, content=raw, headers={"Content-Type": "application/json"}, timeout=timeout
         )
     except httpx.RequestError as e:
-        logger.error("[pricebot] request failed: %s", e)
+        logger.error(
+            "← pricebot %s 不可达: %s", upstream_path, e,
+            extra={"phase": "resp", "endpoint": upstream_path, "step": step,
+                   "error": "upstream_unreachable"},
+        )
         raise HTTPException(
             status_code=status.HTTP_502_BAD_GATEWAY,
             detail=f"pricebot upstream unreachable: {e}",
@@ -79,49 +166,99 @@ async def _passthrough(request: Request, upstream_path: str) -> dict[str, Any]:
 
     if resp.status_code >= 500:
         logger.error(
-            "[pricebot] 5xx status=%d body=%s",
-            resp.status_code,
-            resp.text[:500],
+            "← pricebot %s 5xx=%d", upstream_path, resp.status_code,
+            extra={"phase": "resp", "endpoint": upstream_path, "step": step,
+                   "http_status": resp.status_code, "error": "upstream_5xx"},
         )
         raise HTTPException(
             status_code=status.HTTP_502_BAD_GATEWAY,
             detail=f"pricebot upstream returned {resp.status_code}",
         )
 
-    return resp.json()
+    resp_json = resp.json()
+    cost_ms = int((time.monotonic() - t0) * 1000)
+    if not isinstance(resp_json, dict):
+        return {}, trace_id, meta
+    # 回传 app-server 签发的 trace_id(新客户端首帧据此拿到,后续帧都带上它)
+    resp_json.setdefault("trace_id", trace_id)
+    _action = resp_json.get("action") or {}
+    logger.info(
+        "← pricebot %s cmd=%s cont=%s %dms", upstream_path,
+        _action.get("command"), resp_json.get("continue"), cost_ms,
+        extra={"phase": "resp", "endpoint": upstream_path, "step": step,
+               "command": _action.get("command"), "continue": resp_json.get("continue"),
+               "cost_ms": cost_ms},
+    )
+
+    # 首帧建 running 行(仅本次 mint;老客户端自带 trace_id → 不 mint → 由 done / 其 POST 建行)
+    if minted and harvest_first_frame:
+        try:
+            await run_in_threadpool(
+                _harvest_running_blocking, trace_id,
+                (user.id if user else None), "food",
+                meta.get("device_id"), meta.get("device_info"),
+                resp_json.get("trace_url"),
+            )
+        except Exception as e:  # noqa: BLE001
+            logger.warning("harvest_running failed trace=%s: %s", trace_id, e)
+
+    return resp_json, trace_id, meta
 
 
-@router.post("/intent/recognize", summary="外卖比价 Phase 1 意图识别 (透传到 pricebot)")
-async def intent_recognize(request: Request) -> dict[str, Any]:
-    return await _passthrough(request, "/api/intent/recognize")
+@router.post("/intent/recognize", summary="外卖比价 Phase 1 意图识别 (透传 + 建 running 行)")
+async def intent_recognize(request: Request, user: OptionalUser) -> dict[str, Any]:
+    resp, _, _ = await _forward(request, "/api/intent/recognize", user)
+    return resp
 
 
-@router.post("/intent/step", summary="外卖比价 Phase 1 多帧意图识别 (透传到 pricebot, 仅淘宝源)")
-async def intent_step(request: Request) -> dict[str, Any]:
-    # 多帧版意图识别(展开+滚动采集→提取): 循环调用直到 done(done 帧顶层带
-    # result+calibration)。目前仅淘宝源走这条, 其它源走上面单次 /intent/recognize。
-    return await _passthrough(request, "/api/intent/step")
+@router.post("/intent/step", summary="外卖比价 Phase 1 多帧意图识别 (透传, 仅淘宝源)")
+async def intent_step(request: Request, user: OptionalUser) -> dict[str, Any]:
+    resp, _, _ = await _forward(request, "/api/intent/step", user)
+    return resp
 
 
-@router.post(
-    "/intent/precoupon/step",
-    summary="外卖比价 Phase 0 意图识别前先用券 (透传到 pricebot, 仅美团源)",
-)
-async def intent_precoupon_step(request: Request) -> dict[str, Any]:
-    # 美团源平台『意图识别前先用券』多帧循环: 客户端在调 /intent/recognize 之前先循环
-    # 调本端点到 done(continue=false)。订单页底部有『点击使用X红包』就自动选最大免费券
-    # 用上, 已用券/无券则首帧秒过。与 /intent/step 同属 intent 域, 复用同一透传壳。
-    return await _passthrough(request, "/api/intent/precoupon/step")
+@router.post("/intent/precoupon/step", summary="外卖比价 Phase 0 识别前先用券 (透传, 仅美团源)")
+async def intent_precoupon_step(request: Request, user: OptionalUser) -> dict[str, Any]:
+    resp, _, _ = await _forward(request, "/api/intent/precoupon/step", user)
+    return resp
 
 
-@router.post("/price/step", summary="外卖比价 Phase 2 步进 (透传到 pricebot)")
-async def price_step(request: Request) -> dict[str, Any]:
-    return await _passthrough(request, "/api/price/step")
+@router.post("/price/step", summary="外卖比价 Phase 2 步进 (透传 + done 落库 + 发奖)")
+async def price_step(request: Request, user: OptionalUser) -> dict[str, Any]:
+    resp, trace_id, meta = await _forward(request, "/api/price/step", user)
+    # 最终 done 帧(command=done 且 continue=false)→ harvest 更新成终态 + 幂等发奖。
+    # (单平台中途 done 被 pricebot 改写成 wait+continue=true,不会命中这里,同 coupon 语义。)
+    action = resp.get("action") or {}
+    if action.get("command") == "done" and not resp.get("continue", True):
+        done_params = action.get("params") or {}
+        try:
+            await run_in_threadpool(
+                _harvest_done_blocking, trace_id, (user.id if user else None),
+                done_params, "food",
+                meta.get("device_id"), meta.get("device_info"),
+                resp.get("trace_url") or done_params.get("trace_url"),
+            )
+        except Exception as e:  # noqa: BLE001
+            logger.warning("harvest_done failed trace=%s: %s", trace_id, e)
+    return resp
 
 
-@router.post("/trace/finalize", summary="比价 trace 收尾上云 (透传到 pricebot, 终止/未识别拿 trace_url)")
-async def trace_finalize(request: Request) -> dict[str, Any]:
-    # 用户终止 / Phase1 未识别没走到 done 帧, pricebot 没上云也没回传 trace_url。客户端收尾时
-    # 打这个, _passthrough 按 trace_id 一致性 hash 落到处理这条 trace 的同一 pricebot 进程
-    # (dir_cache 在那, 才能算对 trace 目录), 由后者打包上云返回 {trace_url}。
-    return await _passthrough(request, "/api/trace/finalize")
+@router.post("/trace/finalize", summary="比价 trace 收尾上云 (透传 + 夭折落库)")
+async def trace_finalize(request: Request, user: OptionalUser) -> dict[str, Any]:
+    # 用户终止 / Phase1 未识别没到 done 帧: pricebot 打包半截上云返回 {trace_url};
+    # app-server 顺手把该 trace 的 running 行更新成夭折终态(**不降级 success**)。
+    # 客户端 finalize body 带 status(cancelled/failed)+ reason;老客户端只带 trace_id →
+    # 默认 cancelled,其后续 /compare/record 上报再补精确态。
+    resp, trace_id, meta = await _forward(
+        request, "/api/trace/finalize", user, harvest_first_frame=False,
+    )
+    try:
+        await run_in_threadpool(
+            _harvest_abort_blocking, trace_id,
+            (meta.get("status") or "cancelled"),
+            (meta.get("reason") or meta.get("information")),
+            (resp.get("trace_url") if isinstance(resp, dict) else None),
+        )
+    except Exception as e:  # noqa: BLE001
+        logger.warning("harvest_abort failed trace=%s: %s", trace_id, e)
+    return resp

app/api/v1/compare_record.py

@@ -1,15 +1,13 @@
-"""比价记录 endpoint(「我的比价记录」数据源)。
+"""比价记录 endpoint(「我的比价记录」+ admin 数据源)。
 
 路由前缀 `/api/v1/compare`:
-  POST /record         上报一次比价结果(幂等:同 user+trace_id 覆盖)
+  POST /record         (灰度期兼容)老客户端上报,按 **trace_id** 幂等 + 不降级 success
   GET  /records        比价记录列表(游标分页)
   GET  /records/{id}   单条详情(含 raw_payload 全量)
 
-**均需鉴权**(CurrentUser)——与同文件无关的不鉴权透传 `compare.py` 分开:那个是
-转发壳(MVP 不鉴权),这里是按用户维度落库的业务接口,必须有 user_id。
-
-注:本轮只做 server 端,客户端(android 仓)在 done 帧后调 POST /record 上报的改动
-另起一轮(见 app-server docs/待办与技术债.md P1)。
+**均需鉴权**(CurrentUser)。⚠️ 写路径现以 `compare.py` 透传壳的**后端 harvest** 为主
+(帧0 建 running 行 → done/finalize 落终态,新客户端不再 POST);本 POST /record 仅灰度期
+给老客户端用,与 harvest 按 trace_id reconcile。新版覆盖够高后可下线本 POST(阶段3)。
 """
 from __future__ import annotations