Compare commits

..

3 Commits

Author SHA1 Message Date
guke e1a6095670 feat(compare-alert): 报警末帧列补「停留Xs」(dwell-only,不显总帧数) (#226)
描述

## Summary
- 飞书报警卡片「末帧」列:给 failed(T1/T2/T6) 与 cancelled 兜底两条末帧路径补「末帧所在屏停留 Xs」,
  区分「一到结算页就崩(停留<1s)」vs「在结算页干转 40s 才放弃」;T5 卡死路径不变。
- trace_stuck.py:抽 _last_segment(末段扫描+时长,卡死判据与末帧停留共用);StuckPoint 加 dwell_ms;
  last_step / read_stuck_points.last 附 dwell;删重构后无人调用的 _platform_stuck。
- compare_alert_worker.py:新增 dwell-only 的 _fmt_last(只显环节·页面+停留,**不显总帧数**,
  规避「总帧数配末段时长」误导);failed / cancelled 兜底两处接上。
- 无新增列、无落库、无迁移。

---------

Co-authored-by: guke <guke@autohome.com.cn>
Reviewed-on: #226
2026-08-07 19:13:21 +08:00
guke e8c2ebda13 feat(wallet): 金币记录按 trace_id 聚合比价/领券看广告金币 (#225)
## 背景

「金币变动记录」里看广告金币按**每条广告**统计,一次比价/领券的等候期会连续看多条信息流广告,于是列表刷出一长串「比价奖励 +x」「领券奖励 +x」,淹没其它记录、观感差。

## 改动

把**一次比价 / 一次领券**连续看广告获得的多条金币,按 `trace_id` **聚合成一条**展示(金额取该次会话合计)。

**纯后端**:比价、领券的结算上报都已把本场 `trace_id` 传到 `/feed-reward`(领券自 2026-07-15、客户端 `a98cab8` 起),无需 Android 改动即可生效。

### 具体
- `coin_transaction` 新增 `trace_id` 列 + 索引;`grant_coins` 增 `trace_id` 参数;`grant_feed_reward` 发奖时写入本场 trace_id。
- `GET /api/v1/wallet/coin-transactions` 改为**按 trace_id 分组的游标分页**(CTE 聚合),响应新增 `merged_count`(合并条数,未合并=1)。
- Alembic 迁移:加列 + 索引 + **回填历史**(从 `ad_feed_reward_record` 按 `ref_id == client_event_id` 补 trace_id)。
- 更正 `ad_feed_reward_record.trace_id` 过时注释(领券自 2026-07-15 也带)。
- **只改 App 用户接口**;admin 审计接口(`list_all_coin_transactions`)保持每条一行不动。

## 关键设计

- **分组键**:`biz_type ∈ (feed_ad_reward_comparison, feed_ad_reward_coupon)` 且 `trace_id` 非空 → 按 `trace_id` 合并;其余(激励视频/签到/通用信息流/空 trace 等)每条一行。
- **代表行**取组内 `MAX(id)`(供余额/时间/标题);`amount = SUM`;游标 `rep_id < cursor` 施于**外层**查询——CTE 在用户全量行上分组,**不**按 `id<cursor` 裁剪,避免会话行跨游标时产生「残组」重复(有专门回归测试锁死)。
- 返回 `CoinLedgerRow` dataclass(非 ORM,杜绝把聚合后的 amount 误写回底层流水)。
- SQL 全程 **SQLite/PG 双方言通用**(dev/测试 SQLite,prod PG)。

---------

Co-authored-by: guke <guke@autohome.com.cn>
Reviewed-on: #225
2026-08-07 19:12:58 +08:00
guke 3663b5b11c fix(compare): 「已下单」从店级改按 trace_id 精确对齐 (#224)
背景 / 问题
比价记录页「已下单」tab 混进大量本不该出现的记录:同一家店比价多次、只下过一单,却把这家店的每一条比价(含失败灰卡、没真正下单的成功比价)都塞进「已下单」。

根因
「已下单」原先是店级判定——下单上报(POST /order/report)不带 trace_id、savings_record 也没有指向具体某次比价的键,服务端只能用 savings_record.shop_name == comparison_record.store_name 做店名匹配 → 同店所有比价一并命中。属"先有省钱记账、后蹭已下单"的历史债(见 docs/database/savings_record.md 的 Join Key 注释)。

方案
把 trace_id 从比价会话打通到下单上报,「已下单」改为按 trace_id 精确对齐到那一条比价:

下单上报带上本次比价 trace_id → 落 savings_record.trace_id;
「已下单」过滤/打标(C 端 + admin)统一改成 savings_record.trace_id == comparison_record.trace_id(comparison 侧 trace_id 本就是 NOT NULL UNIQUE);
不做店名回退、不回填历史:没有 trace_id 的订单(历史/老客户端)对齐不上任何记录 → 不进「已下单」。

行为变化(评审必看)
上线后「已下单」只反映"更新版客户端下的新单":历史订单 + 老客户端过渡期订单没 trace_id,不再进「已下单」,随客户端铺量回血。
失败/终止比价不再进「已下单」(本就没真正下单;trace_id 只会指向有价可点的成功比价)。
后台看板「下单数」会掉:历史 savings_record 全是 trace_id=NULL,period_ordered_count 上线当天明显下跌——是口径变化、非回归,请知会看数据的同学。
必须与客户端一起上:只上服务端 → 订单永远不带 trace_id → 任何订单都对不上 → 「已下单」永远空。见关联 android PR。

---------

Co-authored-by: guke <guke@autohome.com.cn>
Reviewed-on: #224
2026-08-07 18:16:55 +08:00
16 changed files with 2167 additions and 47 deletions
@@ -0,0 +1,60 @@
"""coin_transaction.trace_id (金币记录按会话聚合比价/领券看广告金币)
Revision ID: coin_transaction_trace_id
Revises: savings_record_trace_id
Create Date: 2026-08-07
比价/领券信息流发奖时把本场 trace_id 一并写入 coin_transaction;金币变动记录接口按
trace_id 把一次比价/领券的多条广告金币聚合成一条。历史行从 ad_feed_reward_record
回填(ref_id == client_event_id)。
"""
from typing import Sequence, Union
from alembic import op
import sqlalchemy as sa
# revision identifiers, used by Alembic.
revision: str = 'coin_transaction_trace_id'
down_revision: Union[str, Sequence[str], None] = 'savings_record_trace_id'
branch_labels: Union[str, Sequence[str], None] = None
depends_on: Union[str, Sequence[str], None] = None
def upgrade() -> None:
# SQLite 下 ADD COLUMN(可空) 与 CREATE INDEX 均原生支持,无需 batch_alter_table。
op.add_column(
'coin_transaction',
sa.Column('trace_id', sa.String(length=64), nullable=True),
)
op.create_index(
op.f('ix_coin_transaction_trace_id'),
'coin_transaction',
['trace_id'],
unique=False,
)
# 历史回填:从 ad_feed_reward_record 按 ref_id==client_event_id 补 trace_id。仅比价/领券两类、
# 仅当前为空、且广告行确有 trace_id 时补(EXISTS 守护);`trace_id IS NULL` 保证重跑幂等。
# 相关子查询 SQLite/PG 通用。两类 biz_type 在应用侧为 rewards.FEED_AD_SESSION_BIZ_TYPES,
# 此处按「迁移不可变」原则硬编码历史快照(勿改为 import 应用常量)。
# 大表(prod PG)如需可改按 id 区间分批;此处一次性 UPDATE。
op.execute(
"""
UPDATE coin_transaction SET trace_id = (
SELECT r.trace_id FROM ad_feed_reward_record r
WHERE r.client_event_id = coin_transaction.ref_id)
WHERE biz_type IN ('feed_ad_reward_comparison', 'feed_ad_reward_coupon')
AND trace_id IS NULL
AND EXISTS (SELECT 1 FROM ad_feed_reward_record r2
WHERE r2.client_event_id = coin_transaction.ref_id
AND r2.trace_id IS NOT NULL)
"""
)
def downgrade() -> None:
op.drop_index(
op.f('ix_coin_transaction_trace_id'),
table_name='coin_transaction',
)
op.drop_column('coin_transaction', 'trace_id')
+16 -6
View File
@@ -89,6 +89,16 @@ def _fmt_stuck(sp: trace_stuck.StuckPoint) -> str:
return s
def _fmt_last(sp: trace_stuck.StuckPoint) -> str:
"""末帧路径(failed/兜底):环节·页面 + 末段停留时长,不显总帧数(总帧数配末段时长会误导)。
dwell_ms 为 None(缺 ts/时钟回退) → 只显环节。"""
s = sp.label()
if sp.dwell_ms is not None:
sec = round(sp.dwell_ms / 1000)
s += f" 停留{sec}s" if sec >= 1 else " 停留<1s"
return s
def build_hits(
records: list,
*,
@@ -134,9 +144,9 @@ def build_hits(
cancelled_ms_threshold=cancelled_ms_threshold,
cancelled_step_threshold=cancelled_step_threshold,
)
# 没卡点但命中(超阈值放弃)→ 末帧标「退出前在哪屏」(平台·环节·页面)
# 没卡点但命中(超阈值放弃)→ 末帧标「退出前在哪屏 + 在那屏停多久」(dwell-only,不显总帧数)
if hit is not None and res is not None and res.last is not None:
hit = _dc_replace(hit, stuck_point=res.last.label())
hit = _dc_replace(hit, stuck_point=_fmt_last(res.last))
if hit is not None:
hits.append(hit)
else:
@@ -156,12 +166,12 @@ def build_hits(
):
td = _trace_dir(base, rec.trace_url)
if td is not None:
sp = trace_stuck.last_step(td)
sp = trace_stuck.last_step(td, max_tail=max_tail)
reads += 1
if sp is not None:
# failed 是「末帧停在哪」:last_step 返回的 frames 是该平台总帧数(非"卡住"帧数)、
# stuck_ms=None,带上帧数/时长会误导,故只用 label() 显示环节
hit = _dc_replace(hit, stuck_point=sp.label())
# failed 是「末帧停在哪 + 在那屏停多久」:只显环节·页面 + 末段停留 dwell,
# 不显总帧数(总帧数配末段时长会误导);缺 ts 时 _fmt_last 自动退化为只显环节
hit = _dc_replace(hit, stuck_point=_fmt_last(sp))
if hit is not None:
hits.append(hit)
return hits
+10
View File
@@ -26,6 +26,16 @@ SIGNIN_REWARDS: tuple[int, ...] = (
SIGNIN_CYCLE_LEN: int = len(SIGNIN_REWARDS)
# ===== 信息流广告「按会话聚合」的 biz_type =====
# 比价 / 领券等候期看的信息流广告,每条各写一条 coin_transaction;这两类在「金币变动记录」
# 里按 trace_id 聚合成一条展示(见 repositories/wallet.list_coin_transactions)。其余类型
# (reward_video / guide_video / 通用 feed_ad_reward / signin / task_* 等)不聚合。
FEED_AD_SESSION_BIZ_TYPES: tuple[str, str] = (
"feed_ad_reward_comparison",
"feed_ad_reward_coupon",
)
def signin_reward(cycle_day: int) -> int:
"""cycle_day 取值 1..SIGNIN_CYCLE_LEN。"""
return SIGNIN_REWARDS[cycle_day - 1]
+2 -2
View File
@@ -36,8 +36,8 @@ class AdFeedRewardRecord(Base):
# 点位场景:comparison(比价等待) / coupon(领券) / welfare(福利页)。比价与领券共用同一信息流
# 代码位,slot_id/our_code_id 分不出,只能客户端各调用点显式打标;NULL=历史/未升级客户端=未分类。
feed_scene: Mapped[str | None] = mapped_column(String(16), nullable=True)
# 本次比价 trace_id(仅 comparison 场景由客户端带上):把这场广告金币归属到对应比价记录,
# 比价记录页按 trace_id 聚合本场实发金币显示「比价赚 N 金币」。领券/福利/旧客户端 = NULL
# 本次会话 trace_id:比价一直带;领券自 2026-07-15(客户端 a98cab8)起也带;福利/旧客户端 = NULL。
# 用途:比价记录页按 trace_id 聚合「比价赚 N 金币」;金币记录列表把一次比价/领券的多条广告金币聚合成一条
trace_id: Mapped[str | None] = mapped_column(String(64), index=True, nullable=True)
# 来源(广告收益报表用):我们的应用环境 prod/test + 我们配置的代码位 104xxx,由客户端 feed-reward 上报带上。旧数据为 NULL。
app_env: Mapped[str | None] = mapped_column(String(16), nullable=True)
+3
View File
@@ -84,6 +84,9 @@ class CoinTransaction(Base):
# 关联业务 id(签到日期、任务 key 等),可空
ref_id: Mapped[str | None] = mapped_column(String(64), nullable=True)
remark: Mapped[str | None] = mapped_column(String(128), nullable=True)
# 会话键:仅比价/领券信息流发奖(feed_ad_reward_comparison/coupon)时写入本场 trace_id,
# 金币记录列表据此把一次比价/领券的多条广告金币聚合成一条。其余类型 = NULL。
trace_id: Mapped[str | None] = mapped_column(String(64), index=True, nullable=True)
created_at: Mapped[datetime] = mapped_column(
DateTime(timezone=True), server_default=func.now(), index=True, nullable=False
+1 -1
View File
@@ -219,7 +219,7 @@ def grant_feed_reward(
crud_wallet.grant_coins(
db, user_id, coin,
biz_type=reward_biz, ref_id=client_event_id,
remark=reward_remark,
remark=reward_remark, trace_id=trace_id,
)
rec = AdFeedRewardRecord(
client_event_id=client_event_id,
+83 -10
View File
@@ -10,9 +10,10 @@ import logging
import re
import unicodedata
import uuid
from dataclasses import dataclass
from datetime import datetime, timedelta, timezone
from sqlalchemy import func, select, update
from sqlalchemy import String, and_, case, cast, func, literal, select, update
from sqlalchemy.exc import IntegrityError
from sqlalchemy.orm import Session
@@ -170,6 +171,7 @@ def grant_coins(
biz_type: str,
ref_id: str | None = None,
remark: str | None = None,
trace_id: str | None = None,
) -> tuple[CoinAccount, CoinTransaction]:
"""金币变动入口(正数入账 / 负数出账)。更新余额 + 写流水,不 commit。
@@ -189,6 +191,7 @@ def grant_coins(
biz_type=biz_type,
ref_id=ref_id,
remark=remark,
trace_id=trace_id,
created_at=datetime.now(rewards.CN_TZ).replace(tzinfo=None), # 存北京 wall-clock(客户端原样切片显示)
)
db.add(txn)
@@ -257,24 +260,94 @@ def grant_invite_cash(
return acc, txn
@dataclass(frozen=True)
class CoinLedgerRow:
"""`list_coin_transactions` 的返回行。广告类按 trace_id 聚合后的展示行,
** ORM 对象**(避免把聚合后的 amount 误写回底层流水)"""
id: int
amount: int
balance_after: int
biz_type: str
ref_id: str | None
remark: str | None
created_at: datetime
merged_count: int
def list_coin_transactions(
db: Session,
user_id: int,
*,
limit: int = 20,
cursor: int | None = None,
) -> tuple[list[CoinTransaction], int | None]:
"""金币流水分页(按 id 倒序,游标式)。
) -> tuple[list[CoinLedgerRow], int | None]:
"""金币流水分页(游标式,按 id 倒序)。
cursor 为上一页最后一条的 id;返回 (本页列表, next_cursor)
next_cursor None 表示没有下一页
比价/领券两类信息流广告(rewards.FEED_AD_SESSION_BIZ_TYPES)且带 trace_id 的行,
trace_id 聚合成一条(一次比价/领券 = 一行):金额合计代表行取组内最新一条
(MAX(id))的余额/时间merged_count=组内条数其余每条一行
分组必须在该用户**全量**行上算真实 rep_id 后再按 rep_id 过滤**不可** CTE 输入
裁成 id<cursor,否则会话行交错跨游标时会算出与上页重复的残组
cursor 为上一页最后一条的 id(即其组 rep_id);返回 (本页列表, next_cursor)
"""
stmt = select(CoinTransaction).where(CoinTransaction.user_id == user_id)
if cursor is not None:
stmt = stmt.where(CoinTransaction.id < cursor)
stmt = stmt.order_by(CoinTransaction.id.desc()).limit(limit)
ct = CoinTransaction
group_key = case(
(
and_(
ct.biz_type.in_(rewards.FEED_AD_SESSION_BIZ_TYPES),
ct.trace_id.is_not(None),
),
literal("T:") + ct.trace_id,
),
else_=literal("I:") + cast(ct.id, String),
).label("group_key")
items = list(db.execute(stmt).scalars().all())
grp = (
select(
group_key,
func.max(ct.id).label("rep_id"),
func.sum(ct.amount).label("total_amount"),
func.count().label("merged_count"),
)
.where(ct.user_id == user_id)
.group_by(group_key)
.cte("grp")
)
stmt = (
select(
ct.id,
grp.c.total_amount.label("amount"),
ct.balance_after,
ct.biz_type,
ct.ref_id,
ct.remark,
ct.created_at,
grp.c.merged_count,
)
.select_from(grp)
.join(ct, ct.id == grp.c.rep_id)
)
if cursor is not None:
stmt = stmt.where(grp.c.rep_id < cursor)
stmt = stmt.order_by(grp.c.rep_id.desc()).limit(limit)
rows = db.execute(stmt).all()
items = [
CoinLedgerRow(
id=r.id,
amount=int(r.amount),
balance_after=r.balance_after,
biz_type=r.biz_type,
ref_id=r.ref_id,
remark=r.remark,
created_at=r.created_at,
merged_count=int(r.merged_count),
)
for r in rows
]
next_cursor = items[-1].id if len(items) == limit else None
return items, next_cursor
+1
View File
@@ -30,6 +30,7 @@ class CoinTransactionOut(BaseModel):
ref_id: str | None = None
remark: str | None = None
created_at: datetime
merged_count: int = Field(1, description="本行合并的底层流水条数(比价/领券按会话聚合;未合并=1)")
class CoinTransactionPage(BaseModel):
+35 -25
View File
@@ -47,6 +47,7 @@ class StuckPoint:
frames: int # 末段连续困住的帧数(上限 max_tail)
stuck_ms: int | None = None # 末段连续卡住的时长(ms);无 timestamp 时 None
detected_page: str | None = None # 末帧所在页面(pricebot detected_page 原值);无则 None
dwell_ms: int | None = None # 末帧所在屏停留时长(ms);末帧路径(failed/兜底)用,无 ts → None
def label(self) -> str:
p = PLATFORM_LABELS.get(self.platform, self.platform)
@@ -90,10 +91,15 @@ def _read_head(path: Path, nbytes: int = 4096) -> tuple[str | None, str | None,
return (ps.group(1) if ps else None, pg.group(1) if pg else None, ts.group(1) if ts else None)
def _platform_stuck(
platform: str, step_files: list[Path], threshold: int, max_tail: int
) -> StuckPoint | None:
"""末帧往前数连续同 (pipeline_step, detected_page) 的帧数 ≥threshold → 卡死。"""
def _last_segment(
step_files: list[Path], max_tail: int
) -> tuple[str, str | None, int, int | None] | None:
"""末帧往前数连续同 (pipeline_step, detected_page) 的一段。
返回 (pipeline_step, detected_page, count, dwell_ms);末帧 pipeline_step 抠不出 None
dwell_ms = 段末帧ts 段首帧ts(ms);两端 ts 不全可解析或负(帧钟回退) None
这是末段停留的唯一算法,卡死判据(threshold)与末帧停留(无门槛)都复用它
"""
tail = step_files[-max_tail:]
heads = [_read_head(p) for p in tail] # [(ps, pg, ts), ...]
last_ps, last_pg, _ = heads[-1]
@@ -105,15 +111,11 @@ def _platform_stuck(
seg_ts.append(ts)
else:
break
count = len(seg_ts)
if count < threshold:
return None
# seg_ts[0]=末帧, seg_ts[-1]=段首帧;两端都能解析才算时长
t_last, t_first = _parse_ts(seg_ts[0]), _parse_ts(seg_ts[-1])
stuck_ms = round((t_last - t_first).total_seconds() * 1000) if t_last and t_first else None
if stuck_ms is not None and stuck_ms < 0:
stuck_ms = None # 时钟不单调(帧 timestamp 回退)→ 降级为不显示时长
return StuckPoint(platform, last_ps, count, stuck_ms, detected_page=last_pg)
dwell_ms = round((t_last - t_first).total_seconds() * 1000) if t_last and t_first else None
if dwell_ms is not None and dwell_ms < 0:
dwell_ms = None # 时钟不单调(帧 timestamp 回退)→ 降级为不显示时长
return last_ps, last_pg, len(seg_ts), dwell_ms
def read_stuck_points(trace_dir: Path, *, threshold: int, max_tail: int) -> StuckResult:
@@ -132,15 +134,20 @@ def read_stuck_points(trace_dir: Path, *, threshold: int, max_tail: int) -> Stuc
if not step_files:
continue
any_frames = True
sp = _platform_stuck(pdir.name, step_files, threshold, max_tail)
if sp is not None:
points.append(sp)
# 末帧:取帧数最多平台的末帧(与 last_step 同口径),带 detected_page
seg = _last_segment(step_files, max_tail) # 每平台只扫一次末段
if seg is None:
# 末帧抠不出:整段跳过(不判卡死、也不当末帧候选)——与旧版 _platform_stuck→None
# + 独立 _read_head(末帧)→ps None 两处一并跳过等价(旧版两者都 key off 末帧)
continue
ps, pg, count, dwell_ms = seg
if count >= threshold:
points.append(StuckPoint(pdir.name, ps, count, dwell_ms, detected_page=pg))
# 末帧:取帧数最多平台的末帧(与 last_step 同口径),带 detected_page + 末段停留
if len(step_files) > best_n:
ps, pg, _ = _read_head(step_files[-1])
if ps is not None:
best_n = len(step_files)
last = StuckPoint(pdir.name, ps, len(step_files), detected_page=pg)
best_n = len(step_files)
last = StuckPoint(
pdir.name, ps, len(step_files), detected_page=pg, dwell_ms=dwell_ms
)
if not any_frames:
return StuckResult(readable=False, points=[])
return StuckResult(readable=True, points=points, last=last)
@@ -148,8 +155,9 @@ def read_stuck_points(trace_dir: Path, *, threshold: int, max_tail: int) -> Stuc
return StuckResult(readable=False, points=[])
def last_step(trace_dir: Path) -> StuckPoint | None:
"""failed 用:取帧数最多平台的末帧环节(不要求原地打转)。读不到 → None。"""
def last_step(trace_dir: Path, *, max_tail: int) -> StuckPoint | None:
"""failed 用:取帧数最多平台的末帧环节(不要求原地打转)。读不到 → None。
附末段停留 dwell_ms(末帧所在屏停留时长); ts/时钟回退 None"""
try:
if not trace_dir.is_dir():
return None
@@ -164,9 +172,11 @@ def last_step(trace_dir: Path) -> StuckPoint | None:
if best is None:
return None
_, platform, step_files = best
ps, pg, _ts = _read_head(step_files[-1])
if ps is None:
seg = _last_segment(step_files, max_tail)
if seg is None:
return None
return StuckPoint(platform, ps, len(step_files), detected_page=pg) # stuck_ms=None(failed 不算时长)
ps, pg, _count, dwell_ms = seg
# frames 仍=总帧数(选平台口径不变);dwell_ms=末帧所在屏停留
return StuckPoint(platform, ps, len(step_files), detected_page=pg, dwell_ms=dwell_ms)
except OSError:
return None
@@ -0,0 +1,761 @@
# 金币记录按会话汇总比价/领券看广告金币 Implementation Plan
> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
**Goal:** App「金币变动记录」里,一次比价 / 一次领券连续看广告获得的多条金币,按 `trace_id` 聚合成一条展示。
**Architecture:** 纯后端。给 `coin_transaction``trace_id` 列并在发奖时写入;`GET /api/v1/wallet/coin-transactions` 改为按 `trace_id` 分组的游标分页(比价/领券两类 feed 广告按会话合并,其余每条一行);历史行从 `ad_feed_reward_record` 回填。领券自 2026-07-15 起结算已带 trace_id,无需改 Android。仅动 App 用户接口,admin 审计接口不变。
**Tech Stack:** FastAPI · SQLAlchemy 2.0 · Alembic · Pydantic v2 · pytestSQLite)。
**Spec:** `docs/superpowers/specs/2026-08-07-coin-ledger-aggregate-ad-rewards-design.md`
**约定:** dev / pytest 均 **SQLite**prod PostgreSQL——所有 SQL 保持 SQLite/PG 通用(不用 PG 专有语法)。金额单位为金币(非分)。
---
## File Structure
| 文件 | 改动 | 职责 |
|---|---|---|
| `app/core/rewards.py` | 加常量 `FEED_AD_SESSION_BIZ_TYPES` | 「按会话聚合」的两类 biz_type 单一来源(查询 + 写入 + 测试共用) |
| `app/models/wallet.py` | `CoinTransaction``trace_id` 列 | 金币流水会话键 |
| `app/repositories/wallet.py` | `grant_coins``trace_id` 参;`list_coin_transactions` 改分组分页;加 `CoinLedgerRow` | 写入透传 + 聚合读取 |
| `app/repositories/ad_feed_reward.py` | `grant_feed_reward``grant_coins` 时传 `trace_id` | 把本场 trace_id 写进金币流水 |
| `app/schemas/welfare.py` | `CoinTransactionOut``merged_count` | 下发合并条数 |
| `app/models/ad_feed_reward.py` | 更正 `trace_id` 注释 | 文档(领券自 2026-07-15 也带) |
| `alembic/versions/coin_transaction_trace_id.py` | 新建迁移 | 加列 + 索引 + 回填历史 |
| `tests/test_welfare.py` | 追加测试 | 覆盖写入 / 回填 / 发奖透传 / 聚合 / 分页防残组 |
---
## Task 1: 常量 + `CoinTransaction.trace_id` 列 + `grant_coins` 透传
**Files:**
- Modify: `app/core/rewards.py`(加常量)
- Modify: `app/models/wallet.py:86``remark` 后加 `trace_id` 列)
- Modify: `app/repositories/wallet.py:165-196``grant_coins` 加参数 + 写入)
- Test: `tests/test_welfare.py`(追加)
- [ ] **Step 1: 写失败测试**
追加到 `tests/test_welfare.py` 末尾:
```python
def test_grant_coins_persists_trace_id(client) -> None:
"""grant_coins 传 trace_id 落库;不传则为 None。"""
phone = "13800002001"
_login(client, phone)
with SessionLocal() as db:
user = get_user_by_phone(db, phone)
assert user is not None
_, txn1 = crud_wallet.grant_coins(
db, user.id, 5, biz_type="feed_ad_reward_comparison",
ref_id="evt1", remark="比价奖励", trace_id="trace-A",
)
_, txn2 = crud_wallet.grant_coins(
db, user.id, 30, biz_type="signin", remark="每日签到奖励",
)
db.commit()
assert txn1.trace_id == "trace-A"
assert txn2.trace_id is None
```
- [ ] **Step 2: 跑测试确认失败**
Run: `pytest tests/test_welfare.py::test_grant_coins_persists_trace_id -q`
Expected: FAIL —— `TypeError: grant_coins() got an unexpected keyword argument 'trace_id'`
- [ ] **Step 3: 加常量**
`app/core/rewards.py` 的签到常量段之后(约第 32 行 `SIGNIN_CYCLE_LEN` 之后)插入:
```python
# ===== 信息流广告「按会话聚合」的 biz_type =====
# 比价 / 领券等候期看的信息流广告,每条各写一条 coin_transaction;这两类在「金币变动记录」
# 里按 trace_id 聚合成一条展示(见 repositories/wallet.list_coin_transactions)。其余类型
# (reward_video / guide_video / 通用 feed_ad_reward / signin / task_* 等)不聚合。
FEED_AD_SESSION_BIZ_TYPES: tuple[str, str] = (
"feed_ad_reward_comparison",
"feed_ad_reward_coupon",
)
```
- [ ] **Step 4: 加模型列**
`app/models/wallet.py`,把 `remark` 行(第 86 行)后面补一列:
```python
remark: Mapped[str | None] = mapped_column(String(128), nullable=True)
# 会话键:仅比价/领券信息流发奖(feed_ad_reward_comparison/coupon)时写入本场 trace_id,
# 金币记录列表据此把一次比价/领券的多条广告金币聚合成一条。其余类型 = NULL。
trace_id: Mapped[str | None] = mapped_column(String(64), index=True, nullable=True)
```
`String` 在该文件已导入。`index=True` 会在测试 `create_all` 时自动建 `ix_coin_transaction_trace_id`。)
- [ ] **Step 5: `grant_coins` 加参数 + 写入**
`app/repositories/wallet.py``grant_coins`:签名加 `trace_id`,构造 `CoinTransaction` 时带上。
签名(第 165-173 行)改为:
```python
def grant_coins(
db: Session,
user_id: int,
amount: int,
*,
biz_type: str,
ref_id: str | None = None,
remark: str | None = None,
trace_id: str | None = None,
) -> tuple[CoinAccount, CoinTransaction]:
```
`CoinTransaction(...)` 构造(第 185-193 行)改为:
```python
txn = CoinTransaction(
user_id=user_id,
amount=amount,
balance_after=acc.coin_balance,
biz_type=biz_type,
ref_id=ref_id,
remark=remark,
trace_id=trace_id,
created_at=datetime.now(rewards.CN_TZ).replace(tzinfo=None), # 存北京 wall-clock(客户端原样切片显示)
)
```
- [ ] **Step 6: 跑测试确认通过**
Run: `pytest tests/test_welfare.py::test_grant_coins_persists_trace_id -q`
Expected: PASS
- [ ] **Step 7: 回归 + lint**
Run: `pytest tests/test_welfare.py -q && ruff check app/core/rewards.py app/models/wallet.py app/repositories/wallet.py tests/test_welfare.py`
Expected: 原有用例仍 PASS,无新增 lint。
- [ ] **Step 8: 提交**
```bash
git add app/core/rewards.py app/models/wallet.py app/repositories/wallet.py tests/test_welfare.py
git commit -m "feat(wallet): coin_transaction 增 trace_id 列 + grant_coins 透传
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>"
```
---
## Task 2: 迁移(加列 + 索引 + 回填历史)
**Files:**
- Create: `alembic/versions/coin_transaction_trace_id.py`
- Test: `tests/test_welfare.py`(追加回填语义测试)
- [ ] **Step 1: 写失败测试(回填语义)**
`tests/test_welfare.py` 顶部 import 区补两行(若尚无):
```python
from sqlalchemy import text
from app.models.ad_feed_reward import AdFeedRewardRecord
```
追加测试:
```python
def test_backfill_coin_trace_id_from_ad_record(client) -> None:
"""回填:coin_transaction(trace_id 空)按 ref_id==client_event_id 从 ad_feed_reward_record 补 trace_id;
只补比价/领券两类,无关类型与无匹配的不动。SQL 与迁移 coin_transaction_trace_id 保持同步。"""
phone = "13800002002"
_login(client, phone)
with SessionLocal() as db:
user = get_user_by_phone(db, phone)
assert user is not None
# 广告表:两条带 trace_id 的记录(模拟客户端已上报)
db.add(AdFeedRewardRecord(
client_event_id="evt-cmp", user_id=user.id, reward_date="2026-08-07",
duration_seconds=10, unit_count=1, ecpm_raw="1000",
feed_scene="comparison", trace_id="trace-CMP", coin=5, status="granted",
))
db.add(AdFeedRewardRecord(
client_event_id="evt-cpn", user_id=user.id, reward_date="2026-08-07",
duration_seconds=10, unit_count=1, ecpm_raw="1000",
feed_scene="coupon", trace_id="trace-CPN", coin=7, status="granted",
))
db.commit()
# 老金币流水:trace_id 全空(模拟改动前),ref_id 指向上面的广告事件
_, c1 = crud_wallet.grant_coins(db, user.id, 5, biz_type="feed_ad_reward_comparison", ref_id="evt-cmp", remark="比价奖励")
_, c2 = crud_wallet.grant_coins(db, user.id, 7, biz_type="feed_ad_reward_coupon", ref_id="evt-cpn", remark="领券奖励")
_, c3 = crud_wallet.grant_coins(db, user.id, 30, biz_type="signin", remark="每日签到奖励")
db.commit()
assert c1.trace_id is None and c2.trace_id is None
# 执行与迁移 upgrade() 等价的回填 SQL(务必与迁移保持一致)
db.execute(text(
"""
UPDATE coin_transaction SET trace_id = (
SELECT r.trace_id FROM ad_feed_reward_record r
WHERE r.client_event_id = coin_transaction.ref_id)
WHERE biz_type IN ('feed_ad_reward_comparison', 'feed_ad_reward_coupon')
AND trace_id IS NULL
AND EXISTS (SELECT 1 FROM ad_feed_reward_record r2
WHERE r2.client_event_id = coin_transaction.ref_id
AND r2.trace_id IS NOT NULL)
"""
))
db.commit()
db.refresh(c1); db.refresh(c2); db.refresh(c3)
assert c1.trace_id == "trace-CMP" # 比价补上
assert c2.trace_id == "trace-CPN" # 领券补上
assert c3.trace_id is None # 签到不动
```
- [ ] **Step 2: 跑测试确认失败**
Run: `pytest tests/test_welfare.py::test_backfill_coin_trace_id_from_ad_record -q`
Expected: FAIL —— `sqlite3.OperationalError: no such column: trace_id`(Task 1 已加列则此步应改为直接 PASS;若已 PASS 说明 create_all 已含列,跳到 Step 4 建迁移)。
> 说明:本测试验证的是回填 SQL 的 JOIN 语义,不经 Alembic。Task 1 完成后列已在测试库存在,测试可能直接 PASS——这是预期的(回填 SQL 本身是正确逻辑)。真正要新建的产物是迁移文件(Step 4),供 dev/prod 使用。
- [ ] **Step 3: 跑测试确认通过**
Run: `pytest tests/test_welfare.py::test_backfill_coin_trace_id_from_ad_record -q`
Expected: PASS
- [ ] **Step 4: 建迁移文件**
Create `alembic/versions/coin_transaction_trace_id.py`
```python
"""coin_transaction.trace_id (金币记录按会话聚合比价/领券看广告金币)
Revision ID: coin_transaction_trace_id
Revises: comparison_updated_at
Create Date: 2026-08-07
比价/领券信息流发奖时把本场 trace_id 一并写入 coin_transaction;金币变动记录接口按
trace_id 把一次比价/领券的多条广告金币聚合成一条。历史行从 ad_feed_reward_record
回填(ref_id == client_event_id)。
"""
from typing import Sequence, Union
from alembic import op
import sqlalchemy as sa
# revision identifiers, used by Alembic.
revision: str = 'coin_transaction_trace_id'
down_revision: Union[str, Sequence[str], None] = 'comparison_updated_at'
branch_labels: Union[str, Sequence[str], None] = None
depends_on: Union[str, Sequence[str], None] = None
def upgrade() -> None:
# SQLite 下 ADD COLUMN(可空) 与 CREATE INDEX 均原生支持,无需 batch_alter_table。
op.add_column(
'coin_transaction',
sa.Column('trace_id', sa.String(length=64), nullable=True),
)
op.create_index(
op.f('ix_coin_transaction_trace_id'),
'coin_transaction',
['trace_id'],
unique=False,
)
# 历史回填:从 ad_feed_reward_record 按 ref_id==client_event_id 补 trace_id。仅比价/领券两类、
# 仅当前为空、且广告行确有 trace_id 时补(EXISTS 守护);`trace_id IS NULL` 保证重跑幂等。
# 相关子查询 SQLite/PG 通用。两类 biz_type 在应用侧为 rewards.FEED_AD_SESSION_BIZ_TYPES,
# 此处按「迁移不可变」原则硬编码历史快照(勿改为 import 应用常量)。
# 大表(prod PG)如需可改按 id 区间分批;此处一次性 UPDATE。
op.execute(
"""
UPDATE coin_transaction SET trace_id = (
SELECT r.trace_id FROM ad_feed_reward_record r
WHERE r.client_event_id = coin_transaction.ref_id)
WHERE biz_type IN ('feed_ad_reward_comparison', 'feed_ad_reward_coupon')
AND trace_id IS NULL
AND EXISTS (SELECT 1 FROM ad_feed_reward_record r2
WHERE r2.client_event_id = coin_transaction.ref_id
AND r2.trace_id IS NOT NULL)
"""
)
def downgrade() -> None:
op.drop_index(
op.f('ix_coin_transaction_trace_id'),
table_name='coin_transaction',
)
op.drop_column('coin_transaction', 'trace_id')
```
- [ ] **Step 5: 验证迁移可正反向应用(用一次性 scratch SQLite,不碰 dev 库)**
Run:
```bash
DATABASE_URL="sqlite:///./data/_migcheck.db" python -m alembic upgrade head && \
DATABASE_URL="sqlite:///./data/_migcheck.db" python -m alembic downgrade -1 && \
DATABASE_URL="sqlite:///./data/_migcheck.db" python -m alembic upgrade head && \
rm -f ./data/_migcheck.db
```
Expected: 三条 alembic 命令均无报错,最终 head 落在 `coin_transaction_trace_id``rm` 清掉临时库。
- [ ] **Step 6: lint**
Run: `ruff check alembic/versions/coin_transaction_trace_id.py tests/test_welfare.py`
Expected: 无新增 lint。
- [ ] **Step 7: 提交**
```bash
git add alembic/versions/coin_transaction_trace_id.py tests/test_welfare.py
git commit -m "feat(wallet): 迁移加 coin_transaction.trace_id + 索引 + 回填历史
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>"
```
---
## Task 3: `grant_feed_reward` 把 trace_id 写进金币流水
**Files:**
- Modify: `app/repositories/ad_feed_reward.py:219-223``grant_coins` 调用加 `trace_id`
- Test: `tests/test_welfare.py`(追加)
- [ ] **Step 1: 写失败测试**
`tests/test_welfare.py` 顶部 import 区补(若尚无):
```python
from sqlalchemy import select
from app.models.wallet import CoinTransaction
from app.repositories import ad_feed_reward as crud_feed
```
追加测试:
```python
def test_grant_feed_reward_sets_coin_trace_id(client) -> None:
"""grant_feed_reward(comparison) 把 trace_id 透传给 grant_coins,coin_transaction 带上本场 trace_id。"""
phone = "13800002003"
_login(client, phone)
with SessionLocal() as db:
user = get_user_by_phone(db, phone)
assert user is not None
rec = crud_feed.grant_feed_reward(
db, user.id,
client_event_id="evt-fr-1", ecpm="1000", duration_seconds=10,
feed_scene="comparison", trace_id="trace-FR", display_coin=5,
)
assert rec.status == "granted", rec.status
txn = db.execute(
select(CoinTransaction).where(
CoinTransaction.user_id == user.id,
CoinTransaction.ref_id == "evt-fr-1",
)
).scalar_one()
assert txn.biz_type == "feed_ad_reward_comparison"
assert txn.trace_id == "trace-FR"
```
- [ ] **Step 2: 跑测试确认失败**
Run: `pytest tests/test_welfare.py::test_grant_feed_reward_sets_coin_trace_id -q`
Expected: FAIL —— `AssertionError: assert None == 'trace-FR'``grant_coins` 尚未收到 trace_id)。
- [ ] **Step 3: 传 trace_id**
`app/repositories/ad_feed_reward.py` 第 219-223 行 `grant_coins` 调用改为:
```python
crud_wallet.grant_coins(
db, user_id, coin,
biz_type=reward_biz, ref_id=client_event_id,
remark=reward_remark, trace_id=trace_id,
)
```
- [ ] **Step 4: 跑测试确认通过**
Run: `pytest tests/test_welfare.py::test_grant_feed_reward_sets_coin_trace_id -q`
Expected: PASS
- [ ] **Step 5: lint**
Run: `ruff check app/repositories/ad_feed_reward.py tests/test_welfare.py`
Expected: 无新增 lint。
- [ ] **Step 6: 提交**
```bash
git add app/repositories/ad_feed_reward.py tests/test_welfare.py
git commit -m "feat(ad-feed): 发奖把本场 trace_id 写入金币流水
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>"
```
---
## Task 4: 聚合查询 —— `list_coin_transactions` 分组 + `CoinLedgerRow` + `merged_count`
**Files:**
- Modify: `app/repositories/wallet.py`imports + 加 `CoinLedgerRow` + 重写 `list_coin_transactions`
- Modify: `app/schemas/welfare.py:23-32``CoinTransactionOut``merged_count`
- Test: `tests/test_welfare.py`(追加 3 个聚合测试 + 一个 `_seed_coin` 帮手)
- [ ] **Step 1: 写失败测试**
`tests/test_welfare.py` 追加帮手 + 测试:
```python
def _seed_coin(db, user_id, amount, biz_type, *, trace_id=None, ref_id=None, remark=None):
crud_wallet.grant_coins(
db, user_id, amount, biz_type=biz_type, ref_id=ref_id, remark=remark, trace_id=trace_id
)
def test_coin_transactions_aggregate_by_trace(client) -> None:
"""一次比价的多条广告金币聚合成一条:金额合计、merged_count=条数、balance_after 取最后一条。"""
phone = "13800002004"
token = _login(client, phone)
with SessionLocal() as db:
user = get_user_by_phone(db, phone)
assert user is not None
_seed_coin(db, user.id, 5, "feed_ad_reward_comparison", trace_id="t1", ref_id="e1", remark="比价奖励")
_seed_coin(db, user.id, 3, "feed_ad_reward_comparison", trace_id="t1", ref_id="e2", remark="比价奖励")
_seed_coin(db, user.id, 4, "feed_ad_reward_comparison", trace_id="t1", ref_id="e3", remark="比价奖励")
db.commit()
r = client.get("/api/v1/wallet/coin-transactions", headers=_auth(token))
assert r.status_code == 200, r.text
items = r.json()["items"]
assert len(items) == 1
row = items[0]
assert row["biz_type"] == "feed_ad_reward_comparison"
assert row["amount"] == 12 # 5+3+4
assert row["merged_count"] == 3
assert row["balance_after"] == 12 # 末条到账后余额(本用户从 0 起)
def test_coin_transactions_distinct_traces_stay_separate(client) -> None:
"""不同 trace(两次比价 / 一次领券)各成一条;不同会话不合并。"""
phone = "13800002005"
token = _login(client, phone)
with SessionLocal() as db:
user = get_user_by_phone(db, phone)
assert user is not None
_seed_coin(db, user.id, 5, "feed_ad_reward_comparison", trace_id="cmpA", ref_id="a1", remark="比价奖励")
_seed_coin(db, user.id, 6, "feed_ad_reward_comparison", trace_id="cmpB", ref_id="b1", remark="比价奖励")
_seed_coin(db, user.id, 6, "feed_ad_reward_comparison", trace_id="cmpB", ref_id="b2", remark="比价奖励")
_seed_coin(db, user.id, 7, "feed_ad_reward_coupon", trace_id="cpnC", ref_id="c1", remark="领券奖励")
db.commit()
items = client.get("/api/v1/wallet/coin-transactions", headers=_auth(token)).json()["items"]
assert len(items) == 3
assert sorted(i["amount"] for i in items) == [5, 7, 12]
cpn = next(i for i in items if i["biz_type"] == "feed_ad_reward_coupon")
assert cpn["amount"] == 7 and cpn["merged_count"] == 1
def test_coin_transactions_non_session_rows_stay_per_row(client) -> None:
"""签到 / 无 trace 的通用信息流各自一行,不被聚合;夹在比价广告中间的签到不影响比价聚合。"""
phone = "13800002006"
token = _login(client, phone)
with SessionLocal() as db:
user = get_user_by_phone(db, phone)
assert user is not None
_seed_coin(db, user.id, 5, "feed_ad_reward_comparison", trace_id="t1", ref_id="e1", remark="比价奖励")
_seed_coin(db, user.id, 30, "signin", remark="每日签到奖励") # 夹在两条比价广告中间
_seed_coin(db, user.id, 4, "feed_ad_reward_comparison", trace_id="t1", ref_id="e2", remark="比价奖励")
_seed_coin(db, user.id, 8, "feed_ad_reward", trace_id=None, ref_id="w1", remark="信息流广告奖励")
db.commit()
items = client.get("/api/v1/wallet/coin-transactions", headers=_auth(token)).json()["items"]
assert len(items) == 3
cmp_row = next(i for i in items if i["biz_type"] == "feed_ad_reward_comparison")
assert cmp_row["amount"] == 9 and cmp_row["merged_count"] == 2
signin_row = next(i for i in items if i["biz_type"] == "signin")
assert signin_row["amount"] == 30 and signin_row["merged_count"] == 1
feed_row = next(i for i in items if i["biz_type"] == "feed_ad_reward")
assert feed_row["amount"] == 8 and feed_row["merged_count"] == 1
```
- [ ] **Step 2: 跑测试确认失败**
Run: `pytest tests/test_welfare.py::test_coin_transactions_aggregate_by_trace -q`
Expected: FAIL —— 未聚合,返回 3 条 / `merged_count` 字段缺失(`KeyError``len(items)==3`)。
- [ ] **Step 3: schema 加 `merged_count`**
`app/schemas/welfare.py``CoinTransactionOut`(第 23-32 行)在 `created_at` 后加一行:
```python
class CoinTransactionOut(BaseModel):
model_config = ConfigDict(from_attributes=True)
id: int
amount: int = Field(..., description="正=入账,负=出账")
balance_after: int
biz_type: str
ref_id: str | None = None
remark: str | None = None
created_at: datetime
merged_count: int = Field(1, description="本行合并的底层流水条数(比价/领券按会话聚合;未合并=1)")
```
- [ ] **Step 4: 补 imports + 加 `CoinLedgerRow`**
`app/repositories/wallet.py` 顶部:把 `from sqlalchemy import func, select, update`(第 15 行)改为:
```python
from sqlalchemy import String, and_, case, cast, func, literal, select, update
```
在 stdlib import 区(约第 7-13 行)加一行:
```python
from dataclasses import dataclass
```
`list_coin_transactions` 之前加返回行类型:
```python
@dataclass(frozen=True)
class CoinLedgerRow:
"""`list_coin_transactions` 的返回行。广告类按 trace_id 聚合后的展示行,
**非 ORM 对象**(避免把聚合后的 amount 误写回底层流水)。"""
id: int
amount: int
balance_after: int
biz_type: str
ref_id: str | None
remark: str | None
created_at: datetime
merged_count: int
```
- [ ] **Step 5: 重写 `list_coin_transactions`**
`app/repositories/wallet.py` 第 260-279 行整体替换为:
```python
def list_coin_transactions(
db: Session,
user_id: int,
*,
limit: int = 20,
cursor: int | None = None,
) -> tuple[list[CoinLedgerRow], int | None]:
"""金币流水分页(游标式,按 id 倒序)。
比价/领券两类信息流广告(rewards.FEED_AD_SESSION_BIZ_TYPES)且带 trace_id 的行,
按 trace_id 聚合成一条(一次比价/领券 = 一行):金额合计、代表行取组内最新一条
(MAX(id))的余额/时间、merged_count=组内条数。其余每条一行。
分组必须在该用户**全量**行上算真实 rep_id 后再按 rep_id 过滤——**不可**把 CTE 输入
裁成 id<cursor,否则会话行交错跨游标时会算出与上页重复的「残组」。
cursor 为上一页最后一条的 id(即其组 rep_id);返回 (本页列表, next_cursor)。
"""
ct = CoinTransaction
group_key = case(
(
and_(
ct.biz_type.in_(rewards.FEED_AD_SESSION_BIZ_TYPES),
ct.trace_id.is_not(None),
),
literal("T:") + ct.trace_id,
),
else_=literal("I:") + cast(ct.id, String),
).label("group_key")
grp = (
select(
group_key,
func.max(ct.id).label("rep_id"),
func.sum(ct.amount).label("total_amount"),
func.count().label("merged_count"),
)
.where(ct.user_id == user_id)
.group_by(group_key)
.cte("grp")
)
stmt = (
select(
ct.id,
grp.c.total_amount.label("amount"),
ct.balance_after,
ct.biz_type,
ct.ref_id,
ct.remark,
ct.created_at,
grp.c.merged_count,
)
.select_from(grp)
.join(ct, ct.id == grp.c.rep_id)
)
if cursor is not None:
stmt = stmt.where(grp.c.rep_id < cursor)
stmt = stmt.order_by(grp.c.rep_id.desc()).limit(limit)
rows = db.execute(stmt).all()
items = [
CoinLedgerRow(
id=r.id,
amount=int(r.amount),
balance_after=r.balance_after,
biz_type=r.biz_type,
ref_id=r.ref_id,
remark=r.remark,
created_at=r.created_at,
merged_count=int(r.merged_count),
)
for r in rows
]
next_cursor = items[-1].id if len(items) == limit else None
return items, next_cursor
```
(端点 `app/api/v1/wallet.py` 无需改:`CoinTransactionOut.model_validate(it)``CoinLedgerRow` dataclass 按 `from_attributes` 读取即可。)
- [ ] **Step 6: 跑三个聚合测试确认通过**
Run: `pytest tests/test_welfare.py -k "coin_transactions_aggregate_by_trace or distinct_traces_stay_separate or non_session_rows_stay_per_row" -q`
Expected: 3 PASS
- [ ] **Step 7: 全量回归 + lint**
Run: `pytest tests/test_welfare.py -q && ruff check app/repositories/wallet.py app/schemas/welfare.py tests/test_welfare.py`
Expected: 原有用例(含 `test_signin_flow` / `test_exchange_flow` 等对 coin-transactions 的断言)仍 PASS;无新增 lint。
- [ ] **Step 8: 提交**
```bash
git add app/repositories/wallet.py app/schemas/welfare.py tests/test_welfare.py
git commit -m "feat(wallet): 金币记录按 trace_id 聚合比价/领券看广告金币
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>"
```
---
## Task 5: 分页防残组回归测试
**Files:**
- Test: `tests/test_welfare.py`(追加)
- [ ] **Step 1: 写回归测试**
追加:
```python
def test_coin_transactions_pagination_no_phantom_regroup(client) -> None:
"""交错跨游标不产生残组:会话广告成员被其它记录隔开、rep_id 在游标上、成员在游标下时,
翻到下一页该会话不得以「残组」重复出现(锁死 spec §11 反面优化警示)。"""
phone = "13800002007"
token = _login(client, phone)
with SessionLocal() as db:
user = get_user_by_phone(db, phone)
assert user is not None
_seed_coin(db, user.id, 5, "feed_ad_reward_comparison", trace_id="t1", ref_id="e1", remark="比价奖励") # id=n+1
_seed_coin(db, user.id, 30, "signin", remark="每日签到奖励") # id=n+2
_seed_coin(db, user.id, 40, "signin", remark="每日签到奖励") # id=n+3
_seed_coin(db, user.id, 4, "feed_ad_reward_comparison", trace_id="t1", ref_id="e2", remark="比价奖励") # id=n+4 = t1 的 rep
db.commit()
# 第 1 页(limit=2):按 rep 降序 = [t1(rep=n+4, 合计 9), signin(n+3, 40)]
p1 = client.get("/api/v1/wallet/coin-transactions?limit=2", headers=_auth(token)).json()
assert len(p1["items"]) == 2
assert p1["items"][0]["biz_type"] == "feed_ad_reward_comparison"
assert p1["items"][0]["amount"] == 9 and p1["items"][0]["merged_count"] == 2
assert p1["items"][1]["biz_type"] == "signin" and p1["items"][1]["amount"] == 40
assert p1["next_cursor"] is not None
# 第 2 页:只剩另一条 signin(30);t1 的 rep 在游标上,成员虽在游标下也不得成残组重复
p2 = client.get(
f"/api/v1/wallet/coin-transactions?limit=2&cursor={p1['next_cursor']}",
headers=_auth(token),
).json()
assert len(p2["items"]) == 1
assert p2["items"][0]["biz_type"] == "signin" and p2["items"][0]["amount"] == 30
assert all(i["biz_type"] != "feed_ad_reward_comparison" for i in p2["items"])
assert p2["next_cursor"] is None
```
- [ ] **Step 2: 跑测试确认通过(锁定不变量)**
Run: `pytest tests/test_welfare.py::test_coin_transactions_pagination_no_phantom_regroup -q`
Expected: PASSTask 4 的全量分组实现已正确;若 FAIL 说明 CTE 被错误地按 cursor 裁剪,回到 Task 4 Step 5 修正)。
- [ ] **Step 3: lint**
Run: `ruff check tests/test_welfare.py`
Expected: 无新增 lint。
- [ ] **Step 4: 提交**
```bash
git add tests/test_welfare.py
git commit -m "test(wallet): 锁死金币记录分组分页不产生残组
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>"
```
---
## Task 6: 更正 `ad_feed_reward_record.trace_id` 过时注释
**Files:**
- Modify: `app/models/ad_feed_reward.py:39-40`
- [ ] **Step 1: 改注释**
`app/models/ad_feed_reward.py` 第 39-40 行两行注释替换为:
```python
# 本次会话 trace_id:比价一直带;领券自 2026-07-15(客户端 a98cab8)起也带;福利/旧客户端 = NULL。
# 用途:比价记录页按 trace_id 聚合「比价赚 N 金币」;金币记录列表把一次比价/领券的多条广告金币聚合成一条。
trace_id: Mapped[str | None] = mapped_column(String(64), index=True, nullable=True)
```
- [ ] **Step 2: 冒烟 + lint**
Run: `python -c "import app.models.ad_feed_reward" && ruff check app/models/ad_feed_reward.py`
Expected: 无报错、无 lint。
- [ ] **Step 3: 提交**
```bash
git add app/models/ad_feed_reward.py
git commit -m "docs(ad-feed): 更正 trace_id 注释(领券自 2026-07-15 也带)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>"
```
---
## 收尾验证
- [ ] **全量测试**`pytest -q`(对齐 memory「preexisting-test-lint-debt」:先前已知失败数不变、无本改动引入的新失败)。
- [ ] **lint**`ruff check .`(无本改动引入的新问题)。
- [ ] **迁移落 dev**`python -m alembic upgrade head`(把 dev SQLite 迁到最新,确认 `./run.sh` 正常)。
---
## Self-Review(写完计划后自查)
**Spec 覆盖:**
- §4 加列 → Task 1 Step 4 + Task 2。✓
- §5 写入透传 → Task 1 Step 5 + Task 3。✓
- §6 聚合查询(CASE 分组键、MAX(id) 代表/游标、biz_type 门控、常量)→ Task 4 Step 4-5 + Task 1 常量。✓
- §7 `merged_count` 下发 → Task 4 Step 3。✓
- §8 回填(幂等 + EXISTS 守护 + 通用 SQL)→ Task 2 Step 4。✓
- §9 注释更正 → Task 6。✓
- §10 App 端无需改 → 端点未改(Task 4 Step 5 注)。✓
- §11 反面优化警示(残组)→ Task 4 docstring + Task 5 回归测试。✓
- §3 只改用户接口、admin 不动 → 全程只碰 `crud_wallet.list_coin_transactions`,未触 `app/admin`。✓
- §12 测试(分组/隔离/空 trace/分页/回填)→ Task 2/4/5。✓
**占位扫描:** 无 TBD / TODO;每个代码步骤均给出完整代码与确切命令。✓
**类型一致:** `grant_coins(..., trace_id=...)`Task 1)↔ `grant_feed_reward` 调用(Task 3)↔ 查询 `rewards.FEED_AD_SESSION_BIZ_TYPES`(Task 4)↔ 迁移硬编码同两值(Task 2,有意快照)一致;`CoinLedgerRow` 字段(Task 4 Step 4)↔ `CoinTransactionOut` 字段(Task 4 Step 3)逐一对应(id/amount/balance_after/biz_type/ref_id/remark/created_at/merged_count)。✓
@@ -0,0 +1,538 @@
# 比价报警「末帧停留时长」实现计划
> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
**Goal:** 飞书报警卡片「末帧」列给 failed(T1/T2/T6) 与 cancelled 兜底两种末帧路径补上「末帧所在屏停留 Xs」,让运营一眼区分「一到结算页就崩(停留<1s)」vs「在结算页干转 40s 才放弃」。
**Architecture:** 在 `trace_stuck.py` 抽一个末段扫描 helper `_last_segment`(复用现有 `_platform_stuck` 段扫描+时长逻辑,去掉 threshold 门槛),供 `_platform_stuck`/`last_step`/`read_stuck_points` 三处复用;`StuckPoint` 新增 `dwell_ms` 字段承载「末帧所在屏停留」,与 `stuck_ms`(T5 卡死段)语义分离;worker 新增 dwell-only 格式化 `_fmt_last`(只显环节·页面+停留,**不显总帧数**,规避原注释担心的误导)。
**Tech Stack:** Python 3.11 / FastAPI / pytest。纯 CPU+本地文件读,无 DB、无迁移、无外部调用。
关联 spec`docs/superpowers/specs/2026-08-07-compare-alert-last-frame-dwell-design.md`
---
## File Structure
| 文件 | 职责 | 本计划改动 |
|---|---|---|
| `app/services/trace_stuck.py` | trace 末段读取+卡死判定(薄 IO+纯逻辑) | `StuckPoint``dwell_ms`;抽 `_last_segment``_platform_stuck`/`last_step`/`read_stuck_points` 复用它 |
| `app/core/compare_alert_worker.py` | 报警编排 | 新增 `_fmt_last`failed/cancelled 兜底两处末帧格式化换成它;`last_step` 调用传 `max_tail` |
| `tests/test_trace_stuck.py` | trace_stuck 单测 | 新增 dwell 用例;`last_step` 调用加 `max_tail=40` |
| `tests/test_compare_alert_stuck_worker.py` | worker 集成测 | 新增 `_frame_ts` helper + failed/兜底带 dwell 用例 + `_fmt_last` 单测 |
`app/services/compare_alert_format.py` **不改**(末帧列是自由字符串)。
**基线命令**(每个 Task 前后跑,避免全量 pytest 的先前债干扰):
```bash
pytest tests/test_trace_stuck.py tests/test_compare_alert_stuck_worker.py -q
```
---
## Task 1: `StuckPoint``dwell_ms` + 抽 `_last_segment`(重构,行为不变)
**Files:**
- Modify: `app/services/trace_stuck.py`(`StuckPoint` 定义 :43-57`_platform_stuck` :93-116)
- Test: `tests/test_trace_stuck.py`(现有测试作回归网,本 Task 不新增)
> 纯重构 + 加一个默认 `None` 的新字段。无新外部行为,靠现有测试保绿。
- [ ] **Step 1: 跑基线,确认现有测试全绿**
Run: `pytest tests/test_trace_stuck.py -q`
Expected: PASS(全绿)
- [ ] **Step 2: `StuckPoint``dwell_ms` 字段**
`app/services/trace_stuck.py``StuckPoint`(:43-57) 的字段区改为(仅加最后一行,`label()` 不动)
```python
@dataclass(frozen=True)
class StuckPoint:
platform: str
pipeline_step: str
frames: int # 末段连续困住的帧数(上限 max_tail)
stuck_ms: int | None = None # 末段连续卡住的时长(ms);无 timestamp 时 None
detected_page: str | None = None # 末帧所在页面(pricebot detected_page 原值);无则 None
dwell_ms: int | None = None # 末帧所在屏停留时长(ms);末帧路径(failed/兜底)用,无 ts → None
```
- [ ] **Step 3: 新增 `_last_segment` helper**
`app/services/trace_stuck.py``_platform_stuck` **之前**插入(紧跟 `_read_head` 之后)
```python
def _last_segment(
step_files: list[Path], max_tail: int
) -> tuple[str, str | None, int, int | None] | None:
"""末帧往前数连续同 (pipeline_step, detected_page) 的一段。
返回 (pipeline_step, detected_page, count, dwell_ms);末帧 pipeline_step 抠不出 → None。
dwell_ms = 段末帧ts 段首帧ts(ms);两端 ts 不全可解析、或负(帧钟回退) → None。
这是「末段停留」的唯一算法,卡死判据(≥threshold)与末帧停留(无门槛)都复用它。
"""
tail = step_files[-max_tail:]
heads = [_read_head(p) for p in tail] # [(ps, pg, ts), ...]
last_ps, last_pg, _ = heads[-1]
if last_ps is None:
return None
seg_ts: list[str | None] = [] # 连续段的 timestamp(逆序:末帧在前)
for ps, pg, ts in reversed(heads):
if ps == last_ps and pg == last_pg:
seg_ts.append(ts)
else:
break
t_last, t_first = _parse_ts(seg_ts[0]), _parse_ts(seg_ts[-1])
dwell_ms = round((t_last - t_first).total_seconds() * 1000) if t_last and t_first else None
if dwell_ms is not None and dwell_ms < 0:
dwell_ms = None # 时钟不单调(帧 timestamp 回退)→ 降级为不显示时长
return last_ps, last_pg, len(seg_ts), dwell_ms
```
- [ ] **Step 4: `_platform_stuck` 改为复用 `_last_segment`**
`app/services/trace_stuck.py` 的整个 `_platform_stuck`(:93-116) 替换为:
```python
def _platform_stuck(
platform: str, step_files: list[Path], threshold: int, max_tail: int
) -> StuckPoint | None:
"""末帧往前数连续同 (pipeline_step, detected_page) 的帧数 ≥threshold → 卡死。"""
seg = _last_segment(step_files, max_tail)
if seg is None:
return None
ps, pg, count, dwell_ms = seg
if count < threshold:
return None
# 卡死:frames=末段帧数、stuck_ms=末段时长(同段自洽);dwell_ms 字段留默认 None(T5 用 stuck_ms)
return StuckPoint(platform, ps, count, dwell_ms, detected_page=pg)
```
> `stuck_ms` 收的就是 `_last_segment``dwell_ms` 值——T5 场景「末段=卡死段」,二者本是同一个量,故行为与改前完全一致。
- [ ] **Step 5: 跑测试,确认行为不变(仍全绿)**
Run: `pytest tests/test_trace_stuck.py -q`
Expected: PASS(全绿;`test_stuck_ms_computed_from_timestamps` 等对 `stuck_ms`/`frames` 的断言不变)
- [ ] **Step 6: Commit**
```bash
git add app/services/trace_stuck.py
git commit -m "refactor(compare-alert): 抽 _last_segment、StuckPoint 加 dwell_ms 字段
末段扫描+时长算法抽成 _last_segment 供三处复用;StuckPoint 新增
dwell_ms(默认 None、承载末帧所在屏停留),T5 行为不变。
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>"
```
---
## Task 2: `last_step``dwell_ms`(failed 路径数据)
**Files:**
- Modify: `app/services/trace_stuck.py`(`last_step` :151-172)
- Modify: `app/core/compare_alert_worker.py`(`last_step` 调用 :159,本 Task 只传 `max_tail`、仍用 `label()`)
- Test: `tests/test_trace_stuck.py`
- [ ] **Step 1: 写失败测试(带 timestamp 的 dwell)**
`tests/test_trace_stuck.py` 末尾追加:
```python
def test_last_step_computes_dwell_ms(tmp_path):
# 帧数最多平台末段 5 帧都在 checkout,ts :00→:08(每帧+2s) → 停留 8s
m = tmp_path / "meituan"
_frame_ts(m, 0, "enter_store", "store", "2026-08-07T12:00:00.000000")
_frame_ts(m, 1, "enter_store", "store", "2026-08-07T12:00:02.000000")
_frame_ts(m, 2, "enter_store", "store", "2026-08-07T12:00:04.000000")
for i in range(3, 8): # step3..7 checkout,末段 5 帧
_frame_ts(m, i, "checkout", "checkout_page",
f"2026-08-07T12:00:{(i - 3) * 2:02d}.000000")
sp = last_step(tmp_path, max_tail=40)
assert sp.platform == "meituan"
assert sp.pipeline_step == "checkout"
assert sp.frames == 8 # 总帧数(非末段)
assert sp.dwell_ms == 8000 # 末段 checkout :00→:08 = 8s
def test_last_step_dwell_none_without_ts(tmp_path):
m = tmp_path / "meituan"
for i in range(5):
_frame(m, i, "checkout", "checkout_page") # 无 timestamp
sp = last_step(tmp_path, max_tail=40)
assert sp.dwell_ms is None
def test_last_step_dwell_zero_single_frame_segment(tmp_path):
# 末帧与前一帧不同屏 → 末段只有末帧 1 帧 → 停留 0(一到就是末屏)
m = tmp_path / "meituan"
for i in range(4):
_frame_ts(m, i, "enter_store", "store", f"2026-08-07T12:00:{i:02d}.000000")
_frame_ts(m, 4, "checkout", "checkout_page", "2026-08-07T12:00:10.000000")
sp = last_step(tmp_path, max_tail=40)
assert sp.pipeline_step == "checkout"
assert sp.dwell_ms == 0
```
同时把现有 `test_last_step_returns_busiest_platform_last_env`(约 :87-95) 里的调用改为传 `max_tail`
```python
sp = last_step(tmp_path, max_tail=40)
```
(断言不变——该 fixture 无 timestamp`dwell_ms=None`=字段默认,精确 `StuckPoint(...)` 相等仍成立。)
- [ ] **Step 2: 跑测试,确认新用例失败**
Run: `pytest tests/test_trace_stuck.py -q`
Expected: FAIL — `last_step() got an unexpected keyword argument 'max_tail'`(签名还没加 `max_tail`)
- [ ] **Step 3: 改 `last_step`**
`app/services/trace_stuck.py` 的整个 `last_step`(:151-172) 替换为:
```python
def last_step(trace_dir: Path, *, max_tail: int) -> StuckPoint | None:
"""failed 用:取帧数最多平台的末帧环节(不要求原地打转)。读不到 → None。
附末段停留 dwell_ms(末帧所在屏停留时长);无 ts/时钟回退 → None。"""
try:
if not trace_dir.is_dir():
return None
best: tuple[int, str, list[Path]] | None = None
for pdir in sorted(trace_dir.iterdir()):
if not pdir.is_dir():
continue
step_files = sorted(pdir.glob("step_*.json"), key=_step_num)
# 平局(同帧数)时取字典序第一个平台(sorted 保证稳定)
if step_files and (best is None or len(step_files) > best[0]):
best = (len(step_files), pdir.name, step_files)
if best is None:
return None
_, platform, step_files = best
seg = _last_segment(step_files, max_tail)
if seg is None:
return None
ps, pg, _count, dwell_ms = seg
# frames 仍=总帧数(选平台口径不变);dwell_ms=末帧所在屏停留
return StuckPoint(platform, ps, len(step_files), detected_page=pg, dwell_ms=dwell_ms)
except OSError:
return None
```
- [ ] **Step 4: 同步 worker 的 `last_step` 调用(保持 `label()` 不变,避免签名破坏 worker 测试)**
`app/core/compare_alert_worker.py` 把 :159 一行:
```python
sp = trace_stuck.last_step(td)
```
改为:
```python
sp = trace_stuck.last_step(td, max_tail=max_tail)
```
(本 Task 只改调用签名;末帧格式化换成 `_fmt_last` 留到 Task 4。)
- [ ] **Step 5: 跑测试,确认全绿**
Run: `pytest tests/test_trace_stuck.py tests/test_compare_alert_stuck_worker.py -q`
Expected: PASS(新 dwell 用例过;worker 集成测因 `last_step` 仍用 `label()`、行为不变,全绿)
- [ ] **Step 6: Commit**
```bash
git add app/services/trace_stuck.py app/core/compare_alert_worker.py tests/test_trace_stuck.py
git commit -m "feat(compare-alert): last_step 附末段停留 dwell_ms
failed 末帧路径拿到「末帧所在屏停留」;frames 仍为总帧数、口径不变。
worker 调用同步传 max_tail(格式化留待接线)。
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>"
```
---
## Task 3: `read_stuck_points``last``dwell_ms`(cancelled 兜底数据)
**Files:**
- Modify: `app/services/trace_stuck.py`(`read_stuck_points` :119-148)
- Test: `tests/test_trace_stuck.py`
- [ ] **Step 1: 写失败测试**
`tests/test_trace_stuck.py` 末尾追加:
```python
def test_read_stuck_points_last_has_dwell(tmp_path):
# 没卡死(末段<threshold),但末帧末段带 ts → last.dwell_ms 有值
m = tmp_path / "meituan"
_frame_ts(m, 0, "enter_store", "store", "2026-08-07T12:00:00.000000")
_frame_ts(m, 1, "enter_store", "store", "2026-08-07T12:00:02.000000")
for i in range(2, 5): # 末段 checkout 3 帧 :00→:04
_frame_ts(m, i, "checkout", "checkout_page",
f"2026-08-07T12:00:{(i - 2) * 2:02d}.000000")
res = read_stuck_points(tmp_path, threshold=15, max_tail=40)
assert res.points == [] # 末段 3<15,没卡死
assert res.last.pipeline_step == "checkout"
assert res.last.frames == 5 # 总帧数
assert res.last.dwell_ms == 4000 # 末段 :00→:04 = 4s
```
- [ ] **Step 2: 跑测试,确认失败**
Run: `pytest tests/test_trace_stuck.py::test_read_stuck_points_last_has_dwell -q`
Expected: FAIL — `assert None == 4000`(`last.dwell_ms` 还没填)
- [ ] **Step 3: 改 `read_stuck_points`(每平台一次 `_last_segment`,零增量 IO)**
`app/services/trace_stuck.py` 的整个 `read_stuck_points`(:119-148) 替换为:
```python
def read_stuck_points(trace_dir: Path, *, threshold: int, max_tail: int) -> StuckResult:
"""逐平台判卡死。目录不存在/异常/无平台帧 → StuckResult(readable=False, points=[])。"""
try:
if not trace_dir.is_dir():
return StuckResult(readable=False, points=[])
points: list[StuckPoint] = []
any_frames = False
last: StuckPoint | None = None
best_n = -1
for pdir in sorted(trace_dir.iterdir()):
if not pdir.is_dir():
continue
step_files = sorted(pdir.glob("step_*.json"), key=_step_num)
if not step_files:
continue
any_frames = True
seg = _last_segment(step_files, max_tail) # 每平台只扫一次末段
if seg is None:
continue # 末帧抠不出:不判卡死、也不当末帧候选
ps, pg, count, dwell_ms = seg
if count >= threshold:
points.append(StuckPoint(pdir.name, ps, count, dwell_ms, detected_page=pg))
# 末帧:取帧数最多平台的末帧(与 last_step 同口径),带 detected_page + 末段停留
if len(step_files) > best_n:
best_n = len(step_files)
last = StuckPoint(
pdir.name, ps, len(step_files), detected_page=pg, dwell_ms=dwell_ms
)
if not any_frames:
return StuckResult(readable=False, points=[])
return StuckResult(readable=True, points=points, last=last)
except OSError:
return StuckResult(readable=False, points=[])
```
> 行为等价校验:`points``StuckPoint` 仍是 `frames=末段count / stuck_ms=末段时长`(T5 卡死,同改前)`last` 仍是 `frames=总帧数`,只是多带 `dwell_ms`。末帧 `ps is None` 的平台整段跳过(不判卡死、不更新 `last`),与改前 `_platform_stuck→None` + `if ps is not None` 一致。
- [ ] **Step 4: 跑测试,确认全绿**
Run: `pytest tests/test_trace_stuck.py -q`
Expected: PASS(新用例过;`test_read_stuck_points_returns_last_frame``test_per_platform_one_stuck_one_normal` 等精确断言仍相等)
- [ ] **Step 5: Commit**
```bash
git add app/services/trace_stuck.py tests/test_trace_stuck.py
git commit -m "feat(compare-alert): read_stuck_points.last 附 dwell_ms
cancelled 兜底末帧拿到末段停留;循环改为每平台一次 _last_segment,
判卡死与末帧候选共用同一次末段扫描,零增量 IO。
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>"
```
---
## Task 4: worker `_fmt_last` + 末帧列接线(dwell-only 显示)
**Files:**
- Modify: `app/core/compare_alert_worker.py`(新增 `_fmt_last`cancelled 兜底 :138-139、failed :161-164)
- Test: `tests/test_compare_alert_stuck_worker.py`
- [ ] **Step 1: 写失败测试(`_fmt_last` 三态 + failed/兜底集成)**
`tests/test_compare_alert_stuck_worker.py` 顶部把 import 改为(加 `_fmt_last`)
```python
from app.core.compare_alert_worker import _fmt_last, _fmt_stuck, build_hits
```
`_frame` helper(约 :23-28) 之后新增带 timestamp 的 fixture helper
```python
def _frame_ts(pdir: Path, idx: int, step: str, page: str, ts: str) -> None:
pdir.mkdir(parents=True, exist_ok=True)
body = {"pipeline_step": step, "detected_page": page, "timestamp": ts,
"windows": [{"n": ["x" * 200]}]}
(pdir / f"step_{idx:03d}.json").write_text(
json.dumps(body, ensure_ascii=False), encoding="utf-8"
)
```
在文件末尾追加:
```python
# ---- _fmt_last 单测(末帧路径:环节·页面 + 停留,不显总帧数)----
def test_fmt_last_with_dwell():
sp = StuckPoint("meituan", "checkout", frames=480,
detected_page="checkout_page", dwell_ms=8000)
assert _fmt_last(sp) == "美团·结算·checkout_page 停留8s"
def test_fmt_last_sub_second():
sp = StuckPoint("meituan", "checkout", frames=480,
detected_page="checkout_page", dwell_ms=300)
assert _fmt_last(sp) == "美团·结算·checkout_page 停留<1s"
def test_fmt_last_without_dwell():
sp = StuckPoint("meituan", "checkout", frames=480,
detected_page="checkout_page", dwell_ms=None)
assert _fmt_last(sp) == "美团·结算·checkout_page"
# ---- 末帧路径带 dwell 集成 ----
def test_failed_stuck_point_has_dwell(tmp_path):
# failed 末帧 5 帧都在 checkout,ts :00→:08 → stuck_point 附「停留8s」
p = tmp_path / "20260807_f" / "meituan"
for i in range(5):
_frame_ts(p, i, "checkout", "checkout_page",
f"2026-08-07T12:00:{i * 2:02d}.000000")
rec = _Rec(status="failed", fail_reason="启动超时",
trace_url="https://x/traces/20260807_f/")
hits = build_hits([rec], work_log_dir=str(tmp_path), **_KW)
assert len(hits) == 1
assert hits[0].alert_type == "T2"
assert hits[0].stuck_point == "美团·结算·checkout_page 停留8s"
def test_cancelled_fallback_stuck_point_has_dwell(tmp_path):
# cancelled 超阈值(95s)但末段 5<15 不卡死 → 兜底,末帧带 ts → 附「停留8s」
p = tmp_path / "20260807_c" / "eleme"
_frame_ts(p, 0, "set_address", "home", "2026-08-07T12:00:00.000000")
for i in range(1, 6): # enter_store 5 帧 :02→:10
_frame_ts(p, i, "enter_store", "store",
f"2026-08-07T12:00:{i * 2:02d}.000000")
rec = _Rec(status="cancelled", trace_url="https://x/traces/20260807_c/",
total_ms=95000, step_count=40)
hits = build_hits([rec], work_log_dir=str(tmp_path), **_KW)
assert len(hits) == 1
assert hits[0].alert_type == "T5"
assert hits[0].stuck_point == "饿了么·进店·store 停留8s"
```
- [ ] **Step 2: 跑测试,确认失败**
Run: `pytest tests/test_compare_alert_stuck_worker.py -q`
Expected: FAIL — `ImportError: cannot import name '_fmt_last'`
- [ ] **Step 3: 新增 `_fmt_last`**
`app/core/compare_alert_worker.py``_fmt_stuck`(:84-89) **之后**插入:
```python
def _fmt_last(sp: trace_stuck.StuckPoint) -> str:
"""末帧路径(failed/兜底):环节·页面 + 末段停留时长,不显总帧数(总帧数配末段时长会误导)。
dwell_ms 为 None(缺 ts/时钟回退) → 只显环节。"""
s = sp.label()
if sp.dwell_ms is not None:
sec = round(sp.dwell_ms / 1000)
s += f" 停留{sec}s" if sec >= 1 else " 停留<1s"
return s
```
- [ ] **Step 4: cancelled 兜底接线**
`app/core/compare_alert_worker.py` 把 :137-139
```python
# 没卡点但命中(超阈值放弃)→ 用末帧标「退出前在哪屏」(平台·环节·页面)
if hit is not None and res is not None and res.last is not None:
hit = _dc_replace(hit, stuck_point=res.last.label())
```
替换为:
```python
# 没卡点但命中(超阈值放弃)→ 末帧标「退出前在哪屏 + 在那屏停多久」(dwell-only,不显总帧数)
if hit is not None and res is not None and res.last is not None:
hit = _dc_replace(hit, stuck_point=_fmt_last(res.last))
```
- [ ] **Step 5: failed 接线 + 改注释**
`app/core/compare_alert_worker.py` 把 :161-164
```python
if sp is not None:
# failed 是「末帧停在哪」:last_step 返回的 frames 是该平台总帧数(非"卡住"帧数)、
# stuck_ms=None,带上帧数/时长会误导,故只用 label() 显示环节
hit = _dc_replace(hit, stuck_point=sp.label())
```
替换为:
```python
if sp is not None:
# failed 是「末帧停在哪 + 在那屏停多久」:只显环节·页面 + 末段停留 dwell,
# 不显总帧数(总帧数配末段时长会误导);缺 ts 时 _fmt_last 自动退化为只显环节
hit = _dc_replace(hit, stuck_point=_fmt_last(sp))
```
- [ ] **Step 6: 跑测试,确认全绿**
Run: `pytest tests/test_compare_alert_stuck_worker.py -q`
Expected: PASS(新 dwell 用例过;`test_failed_gets_stuck_point_appended``test_cancelled_readable_not_stuck_long_duration_reports` 等无 ts 用例因 `dwell_ms=None`→只显环节,断言仍成立)
- [ ] **Step 7: Commit**
```bash
git add app/core/compare_alert_worker.py tests/test_compare_alert_stuck_worker.py
git commit -m "feat(compare-alert): 末帧列显「停留Xs」(dwell-only,不显总帧数)
failed 与 cancelled 兜底两条末帧路径用 _fmt_last 显示环节·页面+末段停留;
只带 dwell、不带总帧数,规避原注释担心的「总帧数配末段时长」误导;
缺 ts/时钟回退降级只显环节。
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>"
```
---
## Task 5: 全量回归 + lint
**Files:** 无(仅校验)
- [ ] **Step 1: 跑两测试文件全绿**
Run: `pytest tests/test_trace_stuck.py tests/test_compare_alert_stuck_worker.py -q`
Expected: PASS(全绿)
- [ ] **Step 2: 跑其余 compare_alert 相关测试(确认没连带破坏)**
Run: `pytest tests/test_compare_alert_format.py tests/test_compare_alert_rules.py tests/test_compare_alert_fallback.py -q`
Expected: PASS(本计划未碰这些路径,应全绿)
- [ ] **Step 3: ruff 检查改动文件**
Run: `ruff check app/services/trace_stuck.py app/core/compare_alert_worker.py tests/test_trace_stuck.py tests/test_compare_alert_stuck_worker.py`
Expected: `All checks passed!`(如有可自动修的用 `ruff check --fix` 同名文件;有则改后重跑 Step 1)
- [ ] **Step 4: 如 Step 3 有 `--fix` 改动则 commit**
```bash
git add -A
git commit -m "style(compare-alert): ruff 清理
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>"
```
---
## 完成标准(Definition of Done)
- 飞书卡片「末帧」列:failed 与 cancelled 兜底两种路径显示 `平台·环节·页面 停留Xs`(有 ts 时)或 `平台·环节·页面`(缺 ts)。
- T5 卡死路径显示不变(`平台·环节 N帧/Xs`)。
- `pytest tests/test_trace_stuck.py tests/test_compare_alert_stuck_worker.py -q` 全绿。
- `ruff check` 改动文件通过。
- 无新增列、无落库、无迁移。
@@ -0,0 +1,190 @@
# 金币记录:比价/领券看广告金币按会话汇总成一条 设计
- 日期:2026-08-07
- 分支:feat/coin-ledger-aggregate-ad-rewards
- 相关:`app/repositories/wallet.py`(发币/流水)、`app/repositories/ad_feed_reward.py`(信息流发奖)、`app/api/v1/wallet.py`(金币流水接口)、Android `CoinHistoryViewModel`(列表渲染)
## 1. 背景与问题
App「金币变动记录」列表按**每看一次广告一条**展示。一次比价的等候期会连续看多条信息流广告,一次领券流程同理,于是列表里一次比价/领券会刷出一长串「比价奖励 +x」「领券奖励 +x」,淹没其它记录、观感差。
需求:**一次比价的所有看广告金币合并成一条、一次领券的合并成一条**,金额取该次会话合计。
## 2. 现状(关键事实)
- 发币唯一入口 `crud_wallet.grant_coins``app/repositories/wallet.py:165`):每条广告写一条 `CoinTransaction`。金币账本 `coin_transaction` 与现金账本物理分离。
- 信息流发奖 `grant_feed_reward``app/repositories/ad_feed_reward.py:71`)按点位场景拆 `biz_type`
- `feed_ad_reward_comparison`(比价,remark「比价奖励」)
- `feed_ad_reward_coupon`(领券,remark「领券奖励」)
- `feed_ad_reward`(通用/福利/旧端)
- **trace_id 已全链贯通到结算**
- 比价:客户端结算上报早已带 trace_id。
- 领券:`CouponForegroundService.reportFeedReward`**2026-07-15**commit `a98cab8`)起带 `feedScene="coupon" + traceId=sessionTraceId`
- 两者的 trace_id 都落到了 `ad_feed_reward_record.trace_id``grant_feed_reward` 写入)。
- **唯一缺口**`grant_feed_reward``grant_coins` 时**没把 trace_id 透传下去**`coin_transaction` 表本身**没有 trace_id 列** → 金币流水层无法按会话分组。
- 列表接口 `GET /api/v1/coin-transactions``app/api/v1/wallet.py:61``list_coin_transactions` `app/repositories/wallet.py:260`):按 `id` 倒序游标分页。
- 展示文案权在客户端:Android `CoinHistoryViewModel.coinTitle``bizType` 直显固定标题(比价奖励/领券奖励),后端 `remark` 只作兜底。
结论:这是**纯后端**改动,领券不需要额外 Android 改动即可生效。
## 3. 目标与范围
- 目标:金币记录列表里,同一次比价的多条「比价奖励」合并成一条、同一次领券的多条「领券奖励」合并成一条;金额为该次会话合计。
- 会话键:`trace_id`(一次比价/一次领券 = 一个 trace_id)。与「比价记录页」现有的 trace_id 聚合口径一致。
- **范围**:只圈两类 `biz_type` —— `feed_ad_reward_comparison``feed_ad_reward_coupon`。激励视频 `reward_video`、引导视频 `guide_video`、通用信息流 `feed_ad_reward`、签到、任务、兑换等**一律不动**,仍每条一行。
- 聚合位置:**后端**`/coin-transactions` 直接返回合并后的行)。客户端聚合被否决(见 §11)。
- **只改 App 用户接口**:聚合仅作用于 `crud_wallet.list_coin_transactions``GET /api/v1/wallet/coin-transactions`)。**admin 接口不动**——`app/admin/routers/wallet.py` 走独立的 `queries.list_all_coin_transactions`(跨用户、可按 `biz_type` 筛),审计/客服必须能看到**每一条**广告发币,保持每条一行。两者物理分离,聚合天然不波及 admin,此处显式声明防误改。
- 历史:**加列 + 一次性回填**,历史记录也合并。
## 4. 数据模型改动
`coin_transaction` 新增列:
| 列 | 类型 | 约束 | 说明 |
|---|---|---|---|
| `trace_id` | `String(64)` | nullable, index | 会话键;仅比价/领券信息流发奖时写入,其余为 NULL |
- 新增 Alembic 迁移(`render_as_batch` 兼容 SQLite):加列 + 索引 `ix_coin_transaction_trace_id`
- 可选复合索引 `(user_id, trace_id)` 辅助分组扫描(视线上量级决定,MVP 可先只加单列索引)。
## 5. 写入路径改动
- `grant_coins(...)` 增加可选参数 `trace_id: str | None = None`,写入 `CoinTransaction.trace_id`。默认 None → 其余调用方(签到/任务/兑换/激励视频)零改动、保持 NULL。
- `grant_feed_reward``grant_coins` 时透传 `trace_id=trace_id`(比价/领券自然有值;welfare/旧端为 None)。
- 幂等不变:`client_event_id` 仍是幂等键;`grant_coins` 仍不 commit,由 `_commit_record` 同事务提交。
## 6. 读取/聚合查询(核心)
`list_coin_transactions` 改为「分组游标分页」。用 CTE 先分组、再 JOIN 回代表行取展示字段。
分组规则:
```
group_key =
若 biz_type ∈ (feed_ad_reward_comparison, feed_ad_reward_coupon) 且 trace_id 非空
→ 'T:' || trace_id # 同一会话所有广告归一组
否则
→ 'I:' || id # 每条自成一组(行为等同现状)
```
每组取:`rep_id = MAX(id)``total = SUM(amount)``merged_count = COUNT(*)`;再 JOIN `coin_transaction``rep_id` 那条的 `balance_after / biz_type / ref_id / remark / created_at / trace_id`
等价 SQLSQLite / PostgreSQL 通用):
```sql
WITH grp AS (
SELECT
CASE WHEN biz_type IN ('feed_ad_reward_comparison','feed_ad_reward_coupon')
AND trace_id IS NOT NULL
THEN 'T:' || trace_id
ELSE 'I:' || CAST(id AS TEXT) END AS group_key,
MAX(id) AS rep_id,
SUM(amount) AS total_amount,
COUNT(*) AS merged_count
FROM coin_transaction
WHERE user_id = :uid
GROUP BY group_key
)
SELECT ct.id, grp.total_amount AS amount, ct.balance_after,
ct.biz_type, ct.ref_id, ct.remark, ct.trace_id, ct.created_at,
grp.merged_count
FROM grp
JOIN coin_transaction ct ON ct.id = grp.rep_id
WHERE (:cursor IS NULL OR grp.rep_id < :cursor)
ORDER BY grp.rep_id DESC
LIMIT :limit;
```
返回行字段口径:
- `id` = 组内 `MAX(id)`:做列表 key + 下一页游标。组间不重复(每个 id 只属一组),`rep_id DESC` 是全序,游标 `rep_id < cursor` 干净。
- `amount` = 组内合计(该次会话总金币;这些广告行金额恒正)。
- `balance_after` / `created_at` / `remark` / `ref_id` = 代表行(最后一条)的值。`balance_after` 即该次会话最后一条广告到账后的余额,正确。
- `merged_count` = 合并条数(未合并 = 1)。
- `next_cursor` = 本页最后一行的 `id`= 其组 rep_id),够 limit 才给,否则 None。
非分组类型、`trace_id` 为空的旧广告行 → 各自成组(`'I:'||id`),行为与现状完全一致。
> 说明(分组安全性):分组键按 `biz_type` **门控**——通用 `feed_ad_reward`(福利/旧端)即便偶带 trace_id 也走 `'I:'||id` 保持每条一行,只有比价/领券两类才按 trace_id 合并。trace_id 由后端按会话签发、全局唯一(一个 trace = 一次比价**或**一次领券),故仅按 trace_id 分组不会把两类混并,代表行的 `biz_type` 唯一确定展示标题。
>
> 实现约定:这两类 biz_type 抽成模块常量 `FEED_AD_SESSION_BIZ_TYPES`,**查询 / 回填 / 测试共用一处**,避免字符串散落三地漂移。
## 7. 接口契约
`GET /api/v1/coin-transactions` 响应新增 `merged_count``CoinTransactionOut` 加字段,Pydantic 默认 1):
```jsonc
{
"items": [
{ "id": 1520, "amount": 12, "balance_after": 3380,
"biz_type": "feed_ad_reward_comparison", "ref_id": "evt_...",
"remark": "比价奖励", "created_at": "2026-08-07T12:03:11",
"merged_count": 5 }, // 本次比价看了 5 条广告,合计 +12
{ "id": 1512, "amount": 30, "balance_after": 3368,
"biz_type": "signin", "remark": "每日签到奖励",
"merged_count": 1 }
],
"next_cursor": 1490
}
```
`trace_id` 是否在响应里下发:**可选**,MVP 不下发(客户端用不到,标题按 biz_type、金额已合计)。若后续要做「点开看明细」再补。
## 8. 历史回填
迁移里一次性把历史 `coin_transaction.trace_id``ad_feed_reward_record` 补齐(相关子查询 UPDATESQLite/PG 通用):
```sql
UPDATE coin_transaction SET trace_id = (
SELECT r.trace_id FROM ad_feed_reward_record r
WHERE r.client_event_id = coin_transaction.ref_id)
WHERE biz_type IN ('feed_ad_reward_comparison','feed_ad_reward_coupon')
AND trace_id IS NULL
AND EXISTS (SELECT 1 FROM ad_feed_reward_record r2
WHERE r2.client_event_id = coin_transaction.ref_id
AND r2.trace_id IS NOT NULL);
```
- 关联键:`coin_transaction.ref_id == ad_feed_reward_record.client_event_id``grant_feed_reward` 里两者同源,见 `app/repositories/ad_feed_reward.py:221`)。`client_event_id` 有唯一约束 → 相关子查询至多一行,无「子查询多行」风险。
- 只补两类 biz_type;只在广告行确有 trace_id 时补(`EXISTS` 守护)。
- **幂等**`AND trace_id IS NULL` 保证重跑不重复改。
- **大表守护**`coin_transaction` 线上若很大,单条全表 UPDATE 在 PG 上是一次性长事务/行锁;必要时按 `id` 区间**分批**(每批数万行)跑。SQLite(dev)无所谓。
- 2026-07-15 前的领券老行 `ad_feed_reward_record.trace_id` 本就为空 → 回填后仍为空 → 保持每条一行(可接受,会随时间自然淘汰)。
## 9. 顺手更正
更新 `app/models/ad_feed_reward.py:39``trace_id` 注释:现「仅 comparison 场景由客户端带上……领券/福利/旧客户端 = NULL」已过时,改为「比价一直带;领券自 2026-07-15 起也带;福利/旧客户端仍 NULL」。
## 10. App 端
- **核心无需改**:接口返回合并后的行,`CoinHistoryViewModel.coinTitle` 仍按 `bizType` 显示「比价奖励/领券奖励」,金额显示合计值。
- **可选增强(后续、非本次必须)**:客户端读到 `merged_count > 1` 时把标题渲染成「比价奖励 ×5」或副行「看了 5 条广告」。做不做都不影响本功能生效。
## 11. 边界与取舍
- **时间交错重排**:一次比价的广告中间夹了签到等其它记录时,合并行锚定在「该会话最后一条广告」的 id 上,夹在中间的签到会排到合并行之后。这是折叠的固有效果,可接受。
- **余额不连续**:相邻两展示行的 `balance_after` 之间存在被折叠的中间变动,属聚合固有现象;每行 `balance_after` 仍是该点真实运行余额。
- **迟到广告**:会话已过、同 trace 的迟到广告到账后并入该组、金额变大、合并行上移(因 rep_id 变大);用户未刷新则短暂 stale。会话结束后广告即时结算,迟到罕见,可接受。
- **单条会话**:一次会话只看了 1 条广告 → 组内 1 行、`merged_count=1`、展示与现状一致,无需特判。
- **查询成本**:分组 CTE 每页对该用户**全部** `coin_transaction` 行分组后再分页,O(N)/页,比原「`id < cursor LIMIT n`」的索引 seek 重。金币记录非热路径(偶尔打开、翻几页);粗估重度用户万级行/年、单页分组数毫秒级(估算,非实测),可接受。
- **索引口径**:分组按 CASE 表达式(`'T:'||trace_id` / `'I:'||id`)在内存做,`(user_id, trace_id)` 复合索引**并不加速**该 group-by;真正需要的只是「按 user_id 取该用户全部行」,现有 `ix_coin_transaction_user_id` 已够。故本设计**只加 `trace_id` 单列索引**(回填 / 潜在按 trace 查用),不加 `(user_id, trace_id)`;若要 index-only 可选覆盖索引 `(user_id, id, biz_type, trace_id, amount)`MVP 不加。
- ⚠️ **反面优化警示****不要**把 CTE 输入裁成 `WHERE id < :cursor` 求快。会话行若被其它记录交错、成员跨越游标(部分成员 `id < cursor`、但组的 `rep_id ≥ cursor`),裁剪后会算出一个 rep_id 更小的「残组」→ 与上一页已展示的整组**重复出现**。分组必须在该用户全量行上算出真实 `rep_id` 后,再 `rep_id < cursor` 过滤。
- **未来读优化(YAGNI,暂不做)**:若量级压力显现,加只读会话聚合投影表 `coin_ad_session_agg(user_id, trace_id, total_coin, cnt, last_txn_id, last_created_at)`,写广告流水时 upsert,列表用「非广告行 UNION 该投影」走索引分页。这是**读优化**(明细账本仍每条一行、审计不动),与 §13 否决的「写时账本聚合」不同。
## 12. 测试
- **分组**:同 trace 多条比价广告 → 一行(`amount` 合计 / `merged_count=N` / `balance_after` 取最后 / `id=MAX`);两次不同 trace 的比价 → 两行;领券同 trace → 一行。
- **隔离**:比价广告序列中间夹一条签到 → 比价折叠成一行、签到独立成行;`reward_video` / `signin` / `exchange_out` → 仍 1:1`merged_count=1`
- **空 trace_id**`feed_ad_reward`(通用)或 trace_id 为 NULL 的广告行 → 每条一行。
- **分页**:多组跨页时按 `rep_id` 游标翻页不重不漏;一个大组(跨越 limit 边界的多条底层行)不被拆成两页;`next_cursor` 到底返回 None。
- **交错跨游标不重复(防残组回归)**:造一个会话,其广告成员 id 被一条非广告行隔开(部分成员在游标下、`rep_id` 在游标上),翻到下一页时该组**不得**再以「残组」重复出现——锁死 §11 的反面优化警示。
- **回填迁移**:造 `coin_transaction`(比价/领券行,ref_id 指向带 trace_id 的 `ad_feed_reward_record`)+ 无关行,跑迁移后仅两类被正确补 trace_id、无关行不动。
- **既有测试**`/coin-transactions` 现有用例(非广告类型 1:1、分页)保持通过。
## 13. 不做(YAGNI
- 不做写时聚合(会破坏实时到账 / 幂等 / 余额连续性)。
- 不做客户端聚合(SWR 缓存 + 游标分页下,一个会话可能跨页,跨页分组脆弱)。
- 不动激励视频 / 引导视频 / 通用信息流 / 签到等其它类型。
- 不动「比价记录页」——它的「比价赚 N 金币」是独立查询(直接聚合 `ad_feed_reward_record`),不受本改动影响。
- MVP 不下发 trace_id、不做「点开看明细」。
@@ -0,0 +1,121 @@
# 比价报警「末帧停留时长」增强设计
- 日期:2026-08-07
- 分支:feat-compare-alert-last-frame-dwell
- 关联:
- `docs/superpowers/specs/2026-08-04-compare-fail-alert-design.md`(一期报警)
- `docs/superpowers/specs/2026-08-05-compare-stuck-detection-design.md`(卡死定位,本期母 spec
## 1. 背景与问题
飞书报警卡片已有独立的「末帧」列(`app/services/compare_alert_format.py` `_TABLE_COLUMNS`,值 = `AlertHit.stuck_point`)。当前三种末帧口径里,**只有 T5 卡死带时长**,另两种只显「停在哪屏」、不显「在那屏停了多久」:
| 场景 | 「末帧」列现状 | 有时长? |
|---|---|---|
| T5 cancelled·判出原地卡死 | `平台·环节 N帧/Xs``_fmt_stuck` | ✅ `stuck_ms` |
| cancelled·兜底(超阈值放弃、没判出卡点) | `平台·环节·页面``res.last.label()` | ❌ |
| failedT1/T2/T6 | `平台·环节·页面``last_step().label()` | ❌ |
问题:后两种看不出「碰一下结算页就崩」和「在结算页干转 40s 才放弃」的区别——而这个区别对定位 failed / 深度放弃很关键。卡片已有的「用时」列量的是**整场**耗时,不是**末屏**停留,二者互补不重复。
### 1.1 为什么现在没有
不是缺数据,是当初**故意**没算:`last_step()` / `read_stuck_points().last` 返回的 `frames` 是该平台**总帧数**(非末段停留帧),`stuck_ms=None``compare_alert_worker.py` 里有注释明确「带上帧数/时长会误导」,所以只用了 `label()`。数据其实现成——每帧 `timestamp` 已被 `_read_head` 抠出,末段时长算法已在 `_platform_stuck` 里(`stuck_ms`)。
## 2. 目标
- 给 **failed****cancelled 兜底** 两种末帧路径补上「末帧所在屏停留 Xs」。
- 口径与 T5 的 `stuck_ms` 一致(同一种「末段连续同屏时长」),一个卡片里不出现两种「时长含义」。
- **规避原注释担心的误导**:末帧路径只显停留时长、**不显总帧数**。
- 稳:缺时间戳 / 时钟回退 / 读不到 trace → 降级只显环节,**绝不阻断报警**(延续母 spec 铁律)。
## 3. 口径定义
**末段停留 `dwell_ms`** = 从末帧往前、连续 `(pipeline_step, detected_page)` 都与末帧相同的那一段的时长(= 段末帧 `timestamp` 段首帧 `timestamp`round 到 ms)。
- 与卡死判据 `stuck_ms` **同一算法**,唯一区别:**去掉 `count ≥ threshold` 门槛**(末帧路径不要求原地打转,只问「末屏停了多久」)。
- 两端 `timestamp` 都能解析才有值;负时长(帧钟非单调/回退)→ `None`(沿用 `_platform_stuck` 现有降级)。
- failed 与 cancelled 兜底都取「帧数最多平台」的末段停留(与既有 `last` / `last_step` 选平台口径一致)。
## 4. 数据模型
`app/services/trace_stuck.py``StuckPoint` 新增一个字段:
```python
@dataclass(frozen=True)
class StuckPoint:
platform: str
pipeline_step: str
frames: int # 末段连续困住的帧数(上限 max_tail)
stuck_ms: int | None = None # 判为卡死那段的时长;T5 用
detected_page: str | None = None
dwell_ms: int | None = None # 新增:末帧所在屏停留时长;末帧路径用
```
`stuck_ms``dwell_ms` **并存、语义分离**
| 字段 | 含义 | 谁填/谁用 | 与 `frames` 关系 |
|---|---|---|---|
| `stuck_ms` | 判为卡死那段的时长 | `_platform_stuck` 填、T5 显 | `frames`=末段卡住帧数,**同段自洽** |
| `dwell_ms` | 末帧所在屏停留 | 末帧路径填、末帧列显 | `frames`=平台总帧数,**不参与显示** |
> 为什么不复用 `stuck_ms`:末帧路径返回的 `StuckPoint``frames` 是**总帧数**,若把末段停留塞进 `stuck_ms`,对象内部「帧数(总)」与「时长(末段)」不同段、不自洽,且会诱使误用 `_fmt_stuck` 打印出「总帧数 / 末段时长」——正是第 1.1 节要规避的误导。新加独立字段 + 专用 dwell-only 格式化,语义干净。
## 5. 改动
### 5.1 `services/trace_stuck.py`
- 抽末段扫描逻辑(复用现有 `_platform_stuck` 的段扫描 + 时长计算,去掉 `count ≥ threshold` 门槛),产出末段 `(count, dwell_ms, detected_page)`
- `last_step()`:除末帧 head 外,读该平台末段几帧 head,算 `dwell_ms` 填入返回的 `StuckPoint``frames` 仍 = 总帧数,语义不变)。
- `read_stuck_points()``last`:补 `dwell_ms`。该平台 tail 在逐平台判卡死时已读过,几乎零增量 IO。
### 5.2 `core/compare_alert_worker.py`
新增 dwell-only 格式化(**不显总帧数**是规避误导的关键):
```python
def _fmt_last(sp: trace_stuck.StuckPoint) -> str:
"""末帧路径:环节·页面 + 停留时长,不显总帧数。dwell_ms 为 None → 只显环节。"""
s = sp.label()
if sp.dwell_ms is not None:
sec = round(sp.dwell_ms / 1000)
s += f" 停留{sec}s" if sec >= 1 else " 停留<1s"
return s
```
- failed 路径:`sp.label()``_fmt_last(sp)`
- cancelled 兜底:`res.last.label()``_fmt_last(res.last)`
- 改掉原「带上帧数/时长会误导」注释(现在只带 dwell、不带总帧数,不再误导)。
### 5.3 `services/compare_alert_format.py`
**不改**。「末帧」列本就是自由字符串。
## 6. 显示效果
| 场景 | 改前 | 改后 |
|---|---|---|
| failed 结算页秒崩 | 美团·结算·checkout | 美团·结算·checkout **停留<1s** |
| failed 结算页干转 | 美团·结算·checkout | 美团·结算·checkout **停留40s** |
| cancelled 兜底 | 饿了么·进店·store | 饿了么·进店·store **停留8s** |
| 缺 ts / 时钟回退 | 只环节 | 只环节(无停留) |
| T5 卡死 | 美团·加菜 110帧/32s | 不变 |
## 7. 降级与物理边界
- **降级(不阻断报警)**:缺 `timestamp` / 帧钟非单调 / 读不到 trace → `dwell_ms=None` → 只显环节。任何 trace 异常仍在 `trace_stuck` 内降级。
- **物理边界**:帧 `timestamp` 只到末帧。若比价在**写完末帧之后**才彻底冻死(不再落帧),这段测不到 → `dwell≈0`。要覆盖它得用 `abort时间 末帧ts`,但那是 **DB `updated_at`SQLite UTCvs pricebot 帧钟**、跨源跨时区——worker 对「别混钟」很谨慎(见冷启动水位注释),**不引入混钟**。所以报的是「帧级末段停留」:`dwell≈0` = 一到这屏就死,`dwell=30s` = 在这屏干转——低估本身也是信号。
## 8. 测试
- **`trace_stuck` 单测**
- 新增:带 `timestamp` 的末帧 `dwell_ms` 用例(仿 `test_stuck_ms_computed_from_timestamps`)——覆盖末段多帧算出停留、末段单帧 → 0、无 ts → None、时钟回退 → None。
- 更新:`test_last_step_returns_busiest_platform_last_env``test_read_stuck_points_returns_last_frame` 的精确 `StuckPoint` 断言(多 `dwell_ms` 字段)。
- **worker 集成**:新增「failed / cancelled 兜底带 dwell」用例(fixture 帧带 `timestamp`);现有不带 ts 的用例不受影响(`dwell=None` → 只显环节)。
- **`_fmt_last` 单测**:有 dwell(≥1s/ `<1s`(sec 四舍五入为 0)/ 无 dwell 三态。
## 9. 不做(YAGNI
- 不混 DB 钟补「纯末尾冻死」。
- 不动 T5 卡死路径 / `timing.json` / 逐帧 profile(母 spec 第 12 节「不做每帧耗时」指 `timing.json` 逐帧 profile;本期用帧 `timestamp` 算的段时长是两回事,数据现成、报警用得上)。
- 不加列、不落库、不迁移。
+68 -1
View File
@@ -2,7 +2,7 @@
import json
from pathlib import Path
from app.core.compare_alert_worker import _fmt_stuck, build_hits
from app.core.compare_alert_worker import _fmt_last, _fmt_stuck, build_hits
from app.services.trace_stuck import StuckPoint
@@ -28,6 +28,15 @@ def _frame(pdir: Path, idx: int, step: str, page: str) -> None:
)
def _frame_ts(pdir: Path, idx: int, step: str, page: str, ts: str) -> None:
pdir.mkdir(parents=True, exist_ok=True)
body = {"pipeline_step": step, "detected_page": page, "timestamp": ts,
"windows": [{"n": ["x" * 200]}]}
(pdir / f"step_{idx:03d}.json").write_text(
json.dumps(body, ensure_ascii=False), encoding="utf-8"
)
_KW = dict(
stuck_threshold=15, max_tail=40, max_trace_reads=30,
cancelled_ms_threshold=90000, cancelled_step_threshold=30,
@@ -144,3 +153,61 @@ def test_fmt_stuck_with_ms():
def test_fmt_stuck_without_ms():
sp = StuckPoint(platform="meituan", pipeline_step="add_one_dish", frames=110, stuck_ms=None)
assert _fmt_stuck(sp) == "美团·加菜 110帧"
# ---- _fmt_last 单测(末帧路径:环节·页面 + 停留,不显总帧数)----
def test_fmt_last_with_dwell():
sp = StuckPoint("meituan", "checkout", frames=480,
detected_page="checkout_page", dwell_ms=8000)
assert _fmt_last(sp) == "美团·结算·checkout_page 停留8s"
def test_fmt_last_sub_second():
sp = StuckPoint("meituan", "checkout", frames=480,
detected_page="checkout_page", dwell_ms=300)
assert _fmt_last(sp) == "美团·结算·checkout_page 停留<1s"
# <1s 阈值边界:round(500/1000)=0 → <1s;round(999/1000)=1 → 停留1s
sp500 = StuckPoint("meituan", "checkout", frames=480,
detected_page="checkout_page", dwell_ms=500)
assert _fmt_last(sp500) == "美团·结算·checkout_page 停留<1s"
sp999 = StuckPoint("meituan", "checkout", frames=480,
detected_page="checkout_page", dwell_ms=999)
assert _fmt_last(sp999) == "美团·结算·checkout_page 停留1s"
def test_fmt_last_without_dwell():
sp = StuckPoint("meituan", "checkout", frames=480,
detected_page="checkout_page", dwell_ms=None)
assert _fmt_last(sp) == "美团·结算·checkout_page"
# ---- 末帧路径带 dwell 集成 ----
def test_failed_stuck_point_has_dwell(tmp_path):
# failed 末帧 5 帧都在 checkout,ts :00→:08 → stuck_point 附「停留8s」
p = tmp_path / "20260807_f" / "meituan"
for i in range(5):
_frame_ts(p, i, "checkout", "checkout_page",
f"2026-08-07T12:00:{i * 2:02d}.000000")
rec = _Rec(status="failed", fail_reason="启动超时",
trace_url="https://x/traces/20260807_f/")
hits = build_hits([rec], work_log_dir=str(tmp_path), **_KW)
assert len(hits) == 1
assert hits[0].alert_type == "T2"
assert hits[0].stuck_point == "美团·结算·checkout_page 停留8s"
def test_cancelled_fallback_stuck_point_has_dwell(tmp_path):
# cancelled 超阈值(95s)但末段 5<15 不卡死 → 兜底,末帧带 ts → 附「停留8s」
p = tmp_path / "20260807_c" / "eleme"
_frame_ts(p, 0, "set_address", "home", "2026-08-07T12:00:00.000000")
for i in range(1, 6): # enter_store 5 帧 :02→:10
_frame_ts(p, i, "enter_store", "store",
f"2026-08-07T12:00:{i * 2:02d}.000000")
rec = _Rec(status="cancelled", trace_url="https://x/traces/20260807_c/",
total_ms=95000, step_count=40)
hits = build_hits([rec], work_log_dir=str(tmp_path), **_KW)
assert len(hits) == 1
assert hits[0].alert_type == "T5"
assert hits[0].stuck_point == "饿了么·进店·store 停留8s"
+69 -1
View File
@@ -91,7 +91,7 @@ def test_last_step_returns_busiest_platform_last_env(tmp_path):
e = tmp_path / "eleme"
for i in range(3):
_frame(e, i, "enter_store", "store")
sp = last_step(tmp_path)
sp = last_step(tmp_path, max_tail=40)
assert sp == StuckPoint("meituan", "add_one_dish", 20, detected_page="meal_detail_popup")
@@ -172,3 +172,71 @@ def test_stuck_ms_none_when_clock_goes_backwards(tmp_path):
res = read_stuck_points(tmp_path, threshold=15, max_tail=40)
assert len(res.points) == 1
assert res.points[0].stuck_ms is None # 负时长降级为 None
def test_last_step_computes_dwell_ms(tmp_path):
# 帧数最多平台末段 5 帧都在 checkout,ts :00→:08(每帧+2s) → 停留 8s
m = tmp_path / "meituan"
_frame_ts(m, 0, "enter_store", "store", "2026-08-07T12:00:00.000000")
_frame_ts(m, 1, "enter_store", "store", "2026-08-07T12:00:02.000000")
_frame_ts(m, 2, "enter_store", "store", "2026-08-07T12:00:04.000000")
for i in range(3, 8): # step3..7 checkout,末段 5 帧
_frame_ts(m, i, "checkout", "checkout_page",
f"2026-08-07T12:00:{(i - 3) * 2:02d}.000000")
sp = last_step(tmp_path, max_tail=40)
assert sp.platform == "meituan"
assert sp.pipeline_step == "checkout"
assert sp.frames == 8 # 总帧数(非末段)
assert sp.dwell_ms == 8000 # 末段 checkout :00→:08 = 8s
def test_last_step_dwell_none_without_ts(tmp_path):
m = tmp_path / "meituan"
for i in range(5):
_frame(m, i, "checkout", "checkout_page") # 无 timestamp
sp = last_step(tmp_path, max_tail=40)
assert sp.dwell_ms is None
def test_last_step_dwell_zero_single_frame_segment(tmp_path):
# 末帧与前一帧不同屏 → 末段只有末帧 1 帧 → 停留 0(一到就是末屏)
m = tmp_path / "meituan"
for i in range(4):
_frame_ts(m, i, "enter_store", "store", f"2026-08-07T12:00:{i:02d}.000000")
_frame_ts(m, 4, "checkout", "checkout_page", "2026-08-07T12:00:10.000000")
sp = last_step(tmp_path, max_tail=40)
assert sp.pipeline_step == "checkout"
assert sp.dwell_ms == 0
def test_read_stuck_points_last_has_dwell(tmp_path):
# 没卡死(末段<threshold),但末帧末段带 ts → last.dwell_ms 有值
m = tmp_path / "meituan"
_frame_ts(m, 0, "enter_store", "store", "2026-08-07T12:00:00.000000")
_frame_ts(m, 1, "enter_store", "store", "2026-08-07T12:00:02.000000")
for i in range(2, 5): # 末段 checkout 3 帧 :00→:04
_frame_ts(m, i, "checkout", "checkout_page",
f"2026-08-07T12:00:{(i - 2) * 2:02d}.000000")
res = read_stuck_points(tmp_path, threshold=15, max_tail=40)
assert res.points == [] # 末段 3<15,没卡死
assert res.last.pipeline_step == "checkout"
assert res.last.frames == 5 # 总帧数
assert res.last.dwell_ms == 4000 # 末段 :00→:04 = 4s
def test_read_stuck_points_last_falls_through_when_busiest_last_frame_corrupt(tmp_path):
# 最忙平台末帧损坏(抠不出环节)→ _last_segment=None → 整段跳过(重构 continue 分支)
# → last 落到次忙的干净平台。锁定 read_stuck_points 重构的最险等价分支。
m = tmp_path / "meituan"
m.mkdir()
for i in range(7):
_frame(m, i, "add_one_dish", "menu")
(m / "step_007.json").write_bytes(b'{"pipeline_step": "add\xff') # 末帧截断 UTF-8
e = tmp_path / "eleme"
for i in range(3):
_frame(e, i, "enter_store", "store")
res = read_stuck_points(tmp_path, threshold=15, max_tail=40)
assert res.points == [] # 谁都没卡死
assert res.last is not None
assert res.last.platform == "eleme" # 最忙的 meituan(8帧)末帧损坏被跳过,last 落到 eleme
assert res.last.pipeline_step == "enter_store"
+209 -1
View File
@@ -4,7 +4,7 @@
"""
from __future__ import annotations
import json
from sqlalchemy import select, text
from app.core.config import settings
from app.core.rewards import (
@@ -17,6 +17,9 @@ from app.core.rewards import (
)
from app.db.session import SessionLocal
from app.integrations import pangle
from app.models.ad_feed_reward import AdFeedRewardRecord
from app.models.wallet import CoinTransaction
from app.repositories import ad_feed_reward as crud_feed
from app.repositories import wallet as crud_wallet
from app.repositories.user import get_user_by_phone
@@ -384,3 +387,208 @@ def test_endpoints_require_auth(client) -> None:
assert client.get("/api/v1/savings/summary").status_code == 401
assert client.get("/api/v1/savings/battle").status_code == 401
assert client.get("/api/v1/savings/records").status_code == 401
def test_grant_coins_persists_trace_id(client) -> None:
"""grant_coins 传 trace_id 落库;不传则为 None。"""
phone = "13800002001"
_login(client, phone)
with SessionLocal() as db:
user = get_user_by_phone(db, phone)
assert user is not None
_, txn1 = crud_wallet.grant_coins(
db, user.id, 5, biz_type="feed_ad_reward_comparison",
ref_id="evt1", remark="比价奖励", trace_id="trace-A",
)
_, txn2 = crud_wallet.grant_coins(
db, user.id, 30, biz_type="signin", remark="每日签到奖励",
)
db.commit()
assert txn1.trace_id == "trace-A"
assert txn2.trace_id is None
def test_backfill_coin_trace_id_from_ad_record(client) -> None:
"""回填:coin_transaction(trace_id 空)按 ref_id==client_event_id 从 ad_feed_reward_record 补 trace_id;
只补比价/领券两类,无关类型与无匹配的不动SQL 与迁移 coin_transaction_trace_id 保持同步"""
phone = "13800002002"
_login(client, phone)
with SessionLocal() as db:
user = get_user_by_phone(db, phone)
assert user is not None
db.add(AdFeedRewardRecord(
client_event_id="evt-cmp", user_id=user.id, reward_date="2026-08-07",
duration_seconds=10, unit_count=1, ecpm_raw="1000",
feed_scene="comparison", trace_id="trace-CMP", coin=5, status="granted",
))
db.add(AdFeedRewardRecord(
client_event_id="evt-cpn", user_id=user.id, reward_date="2026-08-07",
duration_seconds=10, unit_count=1, ecpm_raw="1000",
feed_scene="coupon", trace_id="trace-CPN", coin=7, status="granted",
))
# 广告行存在但 trace_id 为空(2026-07-15 前的老比价广告):回填必须跳过、保持 NULL(EXISTS 守护)。
db.add(AdFeedRewardRecord(
client_event_id="evt-null", user_id=user.id, reward_date="2026-08-07",
duration_seconds=10, unit_count=1, ecpm_raw="1000",
feed_scene="comparison", trace_id=None, coin=5, status="granted",
))
db.commit()
_, c1 = crud_wallet.grant_coins(db, user.id, 5, biz_type="feed_ad_reward_comparison", ref_id="evt-cmp", remark="比价奖励")
_, c2 = crud_wallet.grant_coins(db, user.id, 7, biz_type="feed_ad_reward_coupon", ref_id="evt-cpn", remark="领券奖励")
_, c3 = crud_wallet.grant_coins(db, user.id, 30, biz_type="signin", remark="每日签到奖励")
_, c4 = crud_wallet.grant_coins(db, user.id, 5, biz_type="feed_ad_reward_comparison", ref_id="evt-null", remark="比价奖励")
db.commit()
assert c1.trace_id is None and c2.trace_id is None
db.execute(text(
"""
UPDATE coin_transaction SET trace_id = (
SELECT r.trace_id FROM ad_feed_reward_record r
WHERE r.client_event_id = coin_transaction.ref_id)
WHERE biz_type IN ('feed_ad_reward_comparison', 'feed_ad_reward_coupon')
AND trace_id IS NULL
AND EXISTS (SELECT 1 FROM ad_feed_reward_record r2
WHERE r2.client_event_id = coin_transaction.ref_id
AND r2.trace_id IS NOT NULL)
"""
))
db.commit()
db.refresh(c1)
db.refresh(c2)
db.refresh(c3)
db.refresh(c4)
assert c1.trace_id == "trace-CMP"
assert c2.trace_id == "trace-CPN"
assert c3.trace_id is None # 无关类型不动
assert c4.trace_id is None # 广告行 trace_id 为空 → EXISTS 守护跳过、保持 NULL
def test_grant_feed_reward_sets_coin_trace_id(client) -> None:
"""grant_feed_reward(comparison) 把 trace_id 透传给 grant_coins,coin_transaction 带上本场 trace_id。"""
phone = "13800002003"
_login(client, phone)
with SessionLocal() as db:
user = get_user_by_phone(db, phone)
assert user is not None
rec = crud_feed.grant_feed_reward(
db, user.id,
client_event_id="evt-fr-1", ecpm="1000", duration_seconds=10,
feed_scene="comparison", trace_id="trace-FR", display_coin=5,
)
assert rec.status == "granted", rec.status
txn = db.execute(
select(CoinTransaction).where(
CoinTransaction.user_id == user.id,
CoinTransaction.ref_id == "evt-fr-1",
)
).scalar_one()
assert txn.biz_type == "feed_ad_reward_comparison"
assert txn.trace_id == "trace-FR"
def _seed_coin(db, user_id, amount, biz_type, *, trace_id=None, ref_id=None, remark=None):
crud_wallet.grant_coins(
db, user_id, amount, biz_type=biz_type, ref_id=ref_id, remark=remark, trace_id=trace_id
)
def test_coin_transactions_aggregate_by_trace(client) -> None:
"""一次比价的多条广告金币聚合成一条:金额合计、merged_count=条数、balance_after 取最后一条。"""
phone = "13800002004"
token = _login(client, phone)
with SessionLocal() as db:
user = get_user_by_phone(db, phone)
assert user is not None
_seed_coin(db, user.id, 5, "feed_ad_reward_comparison", trace_id="t1", ref_id="e1", remark="比价奖励")
_seed_coin(db, user.id, 3, "feed_ad_reward_comparison", trace_id="t1", ref_id="e2", remark="比价奖励")
_seed_coin(db, user.id, 4, "feed_ad_reward_comparison", trace_id="t1", ref_id="e3", remark="比价奖励")
db.commit()
r = client.get("/api/v1/wallet/coin-transactions", headers=_auth(token))
assert r.status_code == 200, r.text
items = r.json()["items"]
assert len(items) == 1
row = items[0]
assert row["biz_type"] == "feed_ad_reward_comparison"
assert row["amount"] == 12
assert row["merged_count"] == 3
assert row["balance_after"] == 12
def test_coin_transactions_distinct_traces_stay_separate(client) -> None:
"""不同 trace(两次比价 / 一次领券)各成一条;不同会话不合并。"""
phone = "13800002005"
token = _login(client, phone)
with SessionLocal() as db:
user = get_user_by_phone(db, phone)
assert user is not None
_seed_coin(db, user.id, 5, "feed_ad_reward_comparison", trace_id="cmpA", ref_id="a1", remark="比价奖励")
_seed_coin(db, user.id, 6, "feed_ad_reward_comparison", trace_id="cmpB", ref_id="b1", remark="比价奖励")
_seed_coin(db, user.id, 6, "feed_ad_reward_comparison", trace_id="cmpB", ref_id="b2", remark="比价奖励")
_seed_coin(db, user.id, 7, "feed_ad_reward_coupon", trace_id="cpnC", ref_id="c1", remark="领券奖励")
db.commit()
items = client.get("/api/v1/wallet/coin-transactions", headers=_auth(token)).json()["items"]
assert len(items) == 3
assert sorted(i["amount"] for i in items) == [5, 7, 12]
cpn = next(i for i in items if i["biz_type"] == "feed_ad_reward_coupon")
assert cpn["amount"] == 7 and cpn["merged_count"] == 1
def test_coin_transactions_non_session_rows_stay_per_row(client) -> None:
"""签到 / 无 trace 的通用信息流各自一行,不被聚合;夹在比价广告中间的签到不影响比价聚合。"""
phone = "13800002006"
token = _login(client, phone)
with SessionLocal() as db:
user = get_user_by_phone(db, phone)
assert user is not None
_seed_coin(db, user.id, 5, "feed_ad_reward_comparison", trace_id="t1", ref_id="e1", remark="比价奖励")
_seed_coin(db, user.id, 30, "signin", remark="每日签到奖励")
_seed_coin(db, user.id, 4, "feed_ad_reward_comparison", trace_id="t1", ref_id="e2", remark="比价奖励")
_seed_coin(db, user.id, 8, "feed_ad_reward", trace_id=None, ref_id="w1", remark="信息流广告奖励")
db.commit()
items = client.get("/api/v1/wallet/coin-transactions", headers=_auth(token)).json()["items"]
assert len(items) == 3
cmp_row = next(i for i in items if i["biz_type"] == "feed_ad_reward_comparison")
assert cmp_row["amount"] == 9 and cmp_row["merged_count"] == 2
signin_row = next(i for i in items if i["biz_type"] == "signin")
assert signin_row["amount"] == 30 and signin_row["merged_count"] == 1
feed_row = next(i for i in items if i["biz_type"] == "feed_ad_reward")
assert feed_row["amount"] == 8 and feed_row["merged_count"] == 1
def test_coin_transactions_pagination_no_phantom_regroup(client) -> None:
"""交错跨游标不产生残组:会话广告成员被其它记录隔开、rep_id 在游标上、成员在游标下时,
翻到下一页该会话不得以残组重复出现(锁死 spec §11 反面优化警示)"""
phone = "13800002007"
token = _login(client, phone)
with SessionLocal() as db:
user = get_user_by_phone(db, phone)
assert user is not None
_seed_coin(db, user.id, 5, "feed_ad_reward_comparison", trace_id="t1", ref_id="e1", remark="比价奖励") # id=n+1
_seed_coin(db, user.id, 30, "signin", remark="每日签到奖励") # id=n+2
_seed_coin(db, user.id, 40, "signin", remark="每日签到奖励") # id=n+3
_seed_coin(db, user.id, 4, "feed_ad_reward_comparison", trace_id="t1", ref_id="e2", remark="比价奖励") # id=n+4 = t1 的 rep
db.commit()
t1_ids = db.execute(
select(CoinTransaction.id)
.where(CoinTransaction.user_id == user.id, CoinTransaction.trace_id == "t1")
.order_by(CoinTransaction.id)
).scalars().all()
# 第 1 页(limit=2):按 rep 降序 = [t1(rep=n+4, 合计 9), signin(n+3, 40)]
p1 = client.get("/api/v1/wallet/coin-transactions?limit=2", headers=_auth(token)).json()
assert len(p1["items"]) == 2
assert p1["items"][0]["biz_type"] == "feed_ad_reward_comparison"
assert p1["items"][0]["amount"] == 9 and p1["items"][0]["merged_count"] == 2
assert p1["items"][1]["biz_type"] == "signin" and p1["items"][1]["amount"] == 40
assert p1["next_cursor"] is not None
# 前置条件:游标(第1页末条 rep)必须严格落在 t1 两成员 id 之间,才是真正的「跨游标」场景;
# 否则用例会退化成非跨游标、悄悄测错形状仍通过,失去回归守护意义。
assert t1_ids[0] < p1["next_cursor"] < t1_ids[-1]
# 第 2 页:只剩另一条 signin(30);t1 的 rep 在游标上,成员虽在游标下也不得成残组重复
p2 = client.get(
f"/api/v1/wallet/coin-transactions?limit=2&cursor={p1['next_cursor']}",
headers=_auth(token),
).json()
assert len(p2["items"]) == 1
assert p2["items"][0]["biz_type"] == "signin" and p2["items"][0]["amount"] == 30
assert all(i["biz_type"] != "feed_ad_reward_comparison" for i in p2["items"])
assert p2["next_cursor"] is None