feat(wx-poc): 会话存档按小程序卡区分源平台,支持京东

原 poller 把所有 weapp 卡片写死成 source="meituan",分享京东也会让手机弹/开
美团。现按 weapp 结构判源:
- poller 新增 _resolve_source(): 按 weapp username/appid/title 关键词判平台
  (jingdong/京东 → jd, meituan/美团 → meituan);认不出默认 meituan + 打
  WARNING 并日志全量 weapp 结构,便于照真实 username/appid 精确补规则。
- app-server _SOURCE_PACKAGES 加 jd → com.jingdong.app.mall。

客户端无需改动(只认 source_package 反查平台名 + launch),纯服务端可热部署。

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

alembic/versions/comparison_llm_cost.py

@@ -0,0 +1,33 @@
+"""comparison_record: llm_cost_yuan + llm_price_snapshot(比价 LLM 调用成本 + 当时单价快照)
+
+回填 llm_calls 时按「当时的价」逐模型算出本次比价 LLM 总成本(元),连同所用单价快照一起冻结到
+记录上;admin 比价记录详情展示实际成本(旧记录 NULL → 前端回退估算)。见 services/llm_cost.py。
+
+Revision ID: comparison_llm_cost
+Revises: ad_ecpm_trace_id
+Create Date: 2026-07-13
+"""
+from collections.abc import Sequence
+
+import sqlalchemy as sa
+from sqlalchemy.dialects import postgresql
+
+from alembic import op
+
+revision: str = "comparison_llm_cost"
+down_revision: str | Sequence[str] | None = "ad_ecpm_trace_id"
+branch_labels: str | Sequence[str] | None = None
+depends_on: str | Sequence[str] | None = None
+
+_JSONB = sa.JSON().with_variant(postgresql.JSONB(), "postgresql")
+
+
+def upgrade() -> None:
+    # 均可空、无索引;SQLite 原生支持 ADD COLUMN,无需 batch_alter_table(同 comparison_debug_fields)。
+    op.add_column("comparison_record", sa.Column("llm_cost_yuan", sa.Float(), nullable=True))
+    op.add_column("comparison_record", sa.Column("llm_price_snapshot", _JSONB, nullable=True))
+
+
+def downgrade() -> None:
+    op.drop_column("comparison_record", "llm_price_snapshot")
+    op.drop_column("comparison_record", "llm_cost_yuan")

app/admin/schemas/comparison.py

@@ -36,6 +36,8 @@ class AdminComparisonListItem(BaseModel):
     retry_count: int | None = None
     input_tokens: int | None = None      # Σ usage.prompt_tokens(server 派生)
     output_tokens: int | None = None     # Σ usage.completion_tokens(server 派生)
+    # 本次比价 LLM 总成本(元,按当时价冻结);旧记录/未回填为 None → 前端「成本」列回退估算。见 services/llm_cost.py。
+    llm_cost_yuan: float | None = None
     device_model: str | None = None
     rom_vendor: str | None = None
     rom_name: str | None = None
@@ -72,3 +74,5 @@ class AdminComparisonDetail(AdminComparisonListItem):
     # 原始上报全量;「卡在哪一步」从 raw_payload.platform_results[*].status 读
     # (store_not_found/items_not_found/below_minimum/unsupported = 卡在 找店/加菜/起送/读价)。
     raw_payload: dict | None = None
+    # 算成本所用单价快照 {mode, prices:{model:{...}}}(llm_cost_yuan 继承自列表项)。见 services/llm_cost.py。
+    llm_price_snapshot: dict | None = None

app/api/internal/wx_finance.py

@@ -0,0 +1,58 @@
+"""会话存档独立 poller → app-server 的内部信号端点(server→server, 非客户端接口)。
+
+会话存档轮询在【独立进程】跑(WeWorkFinanceSdk 是 Go c-shared 库, 嵌进 app-server 会 segfault
+把主进程带崩, 见 scripts/wx_finance_poller.py)。poller 拉到美团卡片/截图后 POST 到这里, 由
+app-server 打比价信号 —— set_pending 在 app-server 进程内存, 心跳才 pop 得到, 故必须回到本进程。
+靠共享密钥头 X-Internal-Secret(== settings.INTERNAL_API_SECRET)校验;未配置 → 503。
+"""
+from __future__ import annotations
+
+import hmac
+import logging
+from typing import Annotated
+
+from fastapi import APIRouter, Header, HTTPException, status
+
+from app.core import wx_poc_signal
+from app.core.config import settings
+from app.schemas.wx_finance import WxFinancePendingIn, WxFinancePendingOut
+
+logger = logging.getLogger("shagua.internal.wx_finance")
+
+router = APIRouter(prefix="/internal", tags=["internal"])
+
+
+def _check_secret(x_internal_secret: str | None) -> None:
+    """共享密钥校验。未配置 → 503(挡裸奔);不匹配 → 401(常量时间比较)。"""
+    configured = settings.INTERNAL_API_SECRET
+    if not configured:
+        raise HTTPException(
+            status_code=status.HTTP_503_SERVICE_UNAVAILABLE,
+            detail="internal api not configured",
+        )
+    if not x_internal_secret or not hmac.compare_digest(x_internal_secret, configured):
+        raise HTTPException(
+            status_code=status.HTTP_401_UNAUTHORIZED, detail="invalid internal secret"
+        )
+
+
+@router.post(
+    "/wx-finance/pending",
+    response_model=WxFinancePendingOut,
+    summary="会话存档 poller 触发比价信号(独立进程→app-server, 打 set_pending)",
+)
+def wx_finance_pending(
+    payload: WxFinancePendingIn,
+    x_internal_secret: Annotated[str | None, Header()] = None,
+) -> WxFinancePendingOut:
+    _check_secret(x_internal_secret)
+    device_id = settings.WX_POC_TEST_DEVICE_ID
+    if not device_id:
+        logger.warning("wx_finance pending: WX_POC_TEST_DEVICE_ID 未配置, 忽略信号")
+        return WxFinancePendingOut(ok=False)
+    wx_poc_signal.set_pending(device_id, payload.source)
+    logger.info(
+        "wx_finance PoC: 已给测试设备 %s 打比价信号 source=%s (触发=%s)",
+        device_id, payload.source, payload.kind,
+    )
+    return WxFinancePendingOut(ok=True)

app/api/v1/compare_record.py

@@ -27,6 +27,7 @@ from app.schemas.compare_record import (
     ComparisonRecordOut,
     ComparisonRecordPage,
 )
+from app.services.llm_cost import compute_llm_cost, get_llm_prices
 from app.services.pricebot_llm_calls import fetch_llm_calls
 
 logger = logging.getLogger("shagua.compare_record")
@@ -81,6 +82,8 @@ def _backfill_llm_calls(record_id: int, trace_id: str) -> None:
         # error 的调用 usage 可能为 None,or {} 兜底)
         rec.input_tokens = sum((c.get("usage") or {}).get("prompt_tokens") or 0 for c in calls)
         rec.output_tokens = sum((c.get("usage") or {}).get("completion_tokens") or 0 for c in calls)
+        # 本次比价 LLM 成本(元)+ 当时单价快照:按 app_config 现价逐模型算好冻结(services/llm_cost.py)。
+        rec.llm_cost_yuan, rec.llm_price_snapshot = compute_llm_cost(calls, get_llm_prices(db))
         db.commit()
         logger.info(
             "backfill llm_calls trace=%s n=%d in_tok=%d out_tok=%d",

app/api/v1/device.py

@@ -14,18 +14,28 @@ import logging
 from fastapi import APIRouter
 
 from app.api.deps import CurrentUser, DbSession
+from app.core import wx_poc_signal
 from app.repositories import device as device_repo
 from app.schemas.device import (
     DeviceOut,
     DeviceRegisterRequest,
     HeartbeatRequest,
+    HeartbeatResponse,
     LivenessAckRequest,
     LivenessOut,
     OkResponse,
+    PendingCompare,
 )
 
 logger = logging.getLogger("shagua.device")
 
+# PoC:源平台代号 → Android 包名(前端 launch 用)。poller 按分享的小程序卡区分平台。
+# 客户端只认 source_package:用它反查平台名(弹窗"原平台")+ launch 对应 App,故新增平台只需在此登记包名。
+_SOURCE_PACKAGES = {
+    "meituan": "com.sankuai.meituan",
+    "jd": "com.jingdong.app.mall",  # 京东主 App(含京东秒送/外卖),后端 intent skill=jd_waimai
+}
+
 router = APIRouter(prefix="/api/v1/device", tags=["device"])
 
 
@@ -52,12 +62,12 @@ def register_device(
     return DeviceOut.model_validate(device)
 
 
-@router.post("/heartbeat", response_model=OkResponse, summary="上报心跳")
+@router.post("/heartbeat", response_model=HeartbeatResponse, summary="上报心跳")
 def report_heartbeat(
     req: HeartbeatRequest,
     user: CurrentUser,
     db: DbSession,
-) -> OkResponse:
+) -> HeartbeatResponse:
     device_repo.touch_heartbeat(
         db,
         user_id=user.id,
@@ -65,7 +75,16 @@ def report_heartbeat(
         accessibility_enabled=req.accessibility_enabled,
         registration_id=req.registration_id,
     )
-    return OkResponse()
+    # PoC:该设备有"待比价"信号则带回(只对写死的测试设备生效, 取走即清、只弹一次)
+    source = wx_poc_signal.pop_pending(req.device_id)
+    pending = None
+    if source:
+        pending = PendingCompare(
+            source_platform=source,
+            source_package=_SOURCE_PACKAGES.get(source, ""),
+        )
+        logger.info("heartbeat 下发比价信号 device=%s source=%s", req.device_id, source)
+    return HeartbeatResponse(pending_compare=pending)
 
 
 @router.get("/liveness", response_model=LivenessOut, summary="查询本机掉线告警(后置检测)")

app/api/v1/wx_kf.py

@@ -0,0 +1,159 @@
+"""微信客服(企业微信)消息接收回调 /wx/kf/callback(裸路径, 非 /api/v1)。
+
+对应企业微信「微信客服」→ 客服账号 → 接收消息回调里填的 URL。是服务号版(wx_mp.py)的
+平行实现:下游(set_pending → 心跳下发 → 前端弹窗)完全一致, 只有上游"如何拿到用户消息"不同 ——
+服务号是微信把消息直接 POST 给你;微信客服是 POST 一个事件通知, 你再用 <Token> 主动 sync_msg 拉。
+
+  GET  : URL 接入验证 → msg_signature 验签(含密文 echostr) → 解密 echostr → 返回明文。
+         ⚠️ 与公众号明文模式不同:企业微信 echostr 是密文, 必须解密后返回。
+  POST : 收事件(密文) → msg_signature 验签 → AES 解密 → 解析 XML。
+         Event=kf_msg_or_event → 取 <Token>/<OpenKfId> → sync_msg 增量拉消息 →
+         客户(origin=3)发来的图片(image) → set_pending(测试设备, "meituan") + 下载落盘。
+         其余消息/事件记日志忽略。一律返回 "success"(异常也吞掉, 不让企业微信重试轰炸)。
+
+回调加密与公众号「安全模式」同源, 直接复用 integrations/wx_mp_crypto(验签 + AES-256-CBC 解密),
+差别只在 receiveid 用 corpid(公众号是 appid)。MVP 只做"收图打信号 + 落盘证明链路通"。
+"""
+from __future__ import annotations
+
+import logging
+import xml.etree.ElementTree as ET
+from pathlib import Path
+from typing import Any
+
+from fastapi import APIRouter, Request
+from fastapi.responses import PlainTextResponse
+
+from app.core import wx_poc_signal
+from app.core.config import settings
+from app.integrations import wx_kf_client, wx_mp_crypto
+
+logger = logging.getLogger("shagua.wx_kf")
+
+router = APIRouter(prefix="/wx/kf", tags=["wx-kf"])
+
+# 收到的截图暂存目录(MVP 验证用;接上识别后可改为不落盘或定期清理)
+_INBOX = Path(settings.MEDIA_ROOT) / "wx_kf_inbox"
+
+
+@router.get("/callback", include_in_schema=False)
+async def verify(
+    msg_signature: str = "", timestamp: str = "", nonce: str = "", echostr: str = ""
+) -> PlainTextResponse:
+    """企业微信 URL 接入验证:验签(密文 echostr 参与)→ 解密 echostr → 返回明文。"""
+    if not settings.wx_kf_callback_configured:
+        logger.warning("wx_kf verify: 回调凭证未配齐(CorpId/Token/AESKey/Secret)")
+        return PlainTextResponse("", status_code=503)
+    if not wx_mp_crypto.verify_msg_signature(
+        settings.WX_KF_TOKEN, timestamp, nonce, echostr, msg_signature
+    ):
+        logger.warning("wx_kf verify: signature 校验失败 ts=%s nonce=%s", timestamp, nonce)
+        return PlainTextResponse("invalid signature", status_code=403)
+    try:
+        plain = wx_mp_crypto.decrypt_message(
+            settings.WX_KF_AES_KEY, settings.WX_KF_CORP_ID, echostr
+        )
+    except Exception:
+        logger.exception("wx_kf verify: echostr 解密失败")
+        return PlainTextResponse("decrypt failed", status_code=403)
+    return PlainTextResponse(plain)
+
+
+@router.post("/callback", include_in_schema=False)
+async def receive(request: Request) -> PlainTextResponse:
+    """收事件通知(安全模式)。任何异常都吞掉返回 success, 不让企业微信重试轰炸, 靠日志排查。"""
+    if not settings.wx_kf_callback_configured:
+        logger.warning("wx_kf receive: 回调凭证未配齐")
+        return PlainTextResponse("success")
+    try:
+        body = (await request.body()).decode("utf-8")
+        qp = request.query_params
+        # 安全模式外层 XML 只有 ToUserName + Encrypt。来源=企业微信服务器(HTTPS)+ 下面验签,
+        # 且 stdlib ET 不扩展外部实体, XXE 不适用。
+        encrypt = ET.fromstring(body).findtext("Encrypt") or ""
+        if not wx_mp_crypto.verify_msg_signature(
+            settings.WX_KF_TOKEN,
+            qp.get("timestamp", ""),
+            qp.get("nonce", ""),
+            encrypt,
+            qp.get("msg_signature", ""),
+        ):
+            logger.warning("wx_kf receive: msg_signature 校验失败")
+            return PlainTextResponse("success")
+        xml = wx_mp_crypto.decrypt_message(
+            settings.WX_KF_AES_KEY, settings.WX_KF_CORP_ID, encrypt
+        )
+        await _handle_event(ET.fromstring(xml))
+    except Exception:
+        logger.exception("wx_kf receive: 处理异常")
+    return PlainTextResponse("success")
+
+
+async def _handle_event(root: ET.Element) -> None:
+    """客服事件:MsgType=event 且 Event=kf_msg_or_event → 用 Token 增量拉消息。"""
+    event = root.findtext("Event") or ""
+    if event != "kf_msg_or_event":
+        logger.info("wx_kf 收到事件(暂忽略): event=%s", event)
+        return
+    token = root.findtext("Token") or ""
+    open_kfid = root.findtext("OpenKfId") or ""
+    msgs = await wx_kf_client.sync_messages(token, open_kfid)
+    logger.info("wx_kf sync_msg 拉到 %d 条 open_kfid=%s", len(msgs), open_kfid)
+    for msg in msgs:
+        await _handle_message(msg)
+
+
+async def _handle_message(msg: dict[str, Any]) -> None:
+    """处理单条客服消息。只关心【客户发来的】(origin=3)图片 / 小程序卡片:打比价信号(+ 落盘)。
+
+    origin:3=客户发的 / 4=系统推送 / 5=接待人员发的。接待人员或系统消息一律不触发比价。
+    两种触发物对应场景:image=用户发结算页截图;miniprogram=用户分享美团小程序卡片(见需求截图)。
+    """
+    if msg.get("origin") != 3:
+        return
+    msgtype = msg.get("msgtype") or ""
+    external_userid = msg.get("external_userid") or ""
+    if msgtype == "image":
+        media_id = (msg.get("image") or {}).get("media_id", "")
+        logger.info(
+            "wx_kf 收到图片: external_userid=%s media_id=%s", external_userid, media_id
+        )
+        _fire_pending("image")
+        await _save_media(media_id, external_userid)
+    elif msgtype == "miniprogram":
+        mp = msg.get("miniprogram") or {}
+        logger.info(
+            "wx_kf 收到小程序卡片: external_userid=%s title=%s appid=%s",
+            external_userid,
+            mp.get("title"),
+            mp.get("appid"),
+        )
+        # 卡片无可下载媒体, 只打信号;源平台/订单识别(可据 appid 判平台)留后续, 现写死美团。
+        _fire_pending("miniprogram")
+    else:
+        logger.info(
+            "wx_kf 收到消息(暂忽略): external_userid=%s type=%s", external_userid, msgtype
+        )
+
+
+def _fire_pending(kind: str) -> None:
+    """PoC:给写死的测试设备打"从美团比价"信号(下次心跳带回前端弹窗, 与服务号版完全一致)。"""
+    poc_dev = settings.WX_POC_TEST_DEVICE_ID
+    if not poc_dev:
+        return
+    wx_poc_signal.set_pending(poc_dev, "meituan")
+    logger.info("wx_kf PoC: 已给测试设备 %s 打比价信号 source=meituan (触发=%s)", poc_dev, kind)
+
+
+async def _save_media(media_id: str, external_userid: str) -> None:
+    """图片消息只给 media_id, 走临时素材下载接口拉回落盘(PoC 验证用)。"""
+    content = await wx_kf_client.download_media(media_id)
+    if not content:
+        return
+    try:
+        _INBOX.mkdir(parents=True, exist_ok=True)
+        dest = _INBOX / f"{external_userid[:16]}_{media_id[:16]}.jpg"
+        dest.write_bytes(content)
+        logger.info("wx_kf 图片已落盘: %s (%d bytes)", dest, len(content))
+    except Exception:
+        logger.exception("wx_kf 图片落盘失败 media_id=%s", media_id)

app/api/v1/wx_mp.py

@@ -0,0 +1,112 @@
+"""微信服务号消息接收回调 /wx/mp/callback(裸路径, 非 /api/v1)。
+
+对应公众平台"设置与开发 → 基本配置 → 服务器配置"里填的 URL。
+  GET  : URL 接入验证 → 校验 signature → 原样返回 echostr(明文)
+  POST : 收用户消息(安全模式密文) → msg_signature 验签 → AES 解密 → 解析 XML
+         图片消息(MsgType=image): 打日志(openid/PicUrl/MediaId) + 用 PicUrl 下载落盘
+         其余消息/事件: 记一条日志忽略。一律返回 "success"(微信据此不再重试)。
+
+MVP 只做"收到图片并落盘"证明链路通;后续接:识别平台+订单 → 打 pending_compare 标记
+→ 心跳带回前端弹窗,以及 openid↔device 绑定。任何异常都返回 "success"、不让微信重试轰炸。
+"""
+from __future__ import annotations
+
+import logging
+import xml.etree.ElementTree as ET
+from pathlib import Path
+
+import httpx
+from fastapi import APIRouter, Request
+from fastapi.responses import PlainTextResponse
+
+from app.core import wx_poc_signal
+from app.core.config import settings
+from app.integrations import wx_mp_crypto
+
+logger = logging.getLogger("shagua.wx_mp")
+
+router = APIRouter(prefix="/wx/mp", tags=["wx-mp"])
+
+# 收到的截图暂存目录(MVP 验证用;接上识别后可改为不落盘或定期清理)
+_INBOX = Path(settings.MEDIA_ROOT) / "wx_inbox"
+
+
+@router.get("/callback", include_in_schema=False)
+async def verify(
+    signature: str = "", timestamp: str = "", nonce: str = "", echostr: str = ""
+) -> PlainTextResponse:
+    """微信 URL 接入验证:校验 signature 通过则原样返回 echostr。"""
+    if not settings.WX_MP_TOKEN:
+        logger.warning("wx_mp verify: WX_MP_TOKEN 未配置")
+        return PlainTextResponse("", status_code=503)
+    if wx_mp_crypto.verify_url_signature(
+        settings.WX_MP_TOKEN, timestamp, nonce, signature
+    ):
+        return PlainTextResponse(echostr)
+    logger.warning("wx_mp verify: signature 校验失败 ts=%s nonce=%s", timestamp, nonce)
+    return PlainTextResponse("invalid signature", status_code=403)
+
+
+@router.post("/callback", include_in_schema=False)
+async def receive(request: Request) -> PlainTextResponse:
+    """收用户消息(安全模式)。任何异常都吞掉返回 success,不让微信重试轰炸,靠日志排查。"""
+    if not settings.wx_mp_callback_configured:
+        logger.warning("wx_mp receive: 回调凭证未配齐(Token/AESKey/AppID)")
+        return PlainTextResponse("success")
+    try:
+        body = (await request.body()).decode("utf-8")
+        qp = request.query_params
+        # 安全模式外层 XML 只有 ToUserName + Encrypt。来源=微信服务器(HTTPS)+ 下面验签,
+        # 且 stdlib ET 不扩展外部实体, XXE 不适用。
+        encrypt = ET.fromstring(body).findtext("Encrypt") or ""
+        if not wx_mp_crypto.verify_msg_signature(
+            settings.WX_MP_TOKEN,
+            qp.get("timestamp", ""),
+            qp.get("nonce", ""),
+            encrypt,
+            qp.get("msg_signature", ""),
+        ):
+            logger.warning("wx_mp receive: msg_signature 校验失败")
+            return PlainTextResponse("success")
+        xml = wx_mp_crypto.decrypt_message(
+            settings.WX_MP_AES_KEY, settings.WX_MP_APPID, encrypt
+        )
+        await _handle_message(ET.fromstring(xml))
+    except Exception:
+        logger.exception("wx_mp receive: 处理异常")
+    return PlainTextResponse("success")
+
+
+async def _handle_message(root: ET.Element) -> None:
+    openid = root.findtext("FromUserName") or ""
+    msg_type = root.findtext("MsgType") or ""
+    if msg_type == "image":
+        pic_url = root.findtext("PicUrl") or ""
+        media_id = root.findtext("MediaId") or ""
+        logger.info(
+            "wx_mp 收到图片: openid=%s media_id=%s pic_url=%s", openid, media_id, pic_url
+        )
+        # PoC: 写死"该测试设备要从美团比价", 下次心跳带回前端弹选平台窗
+        poc_dev = settings.WX_POC_TEST_DEVICE_ID
+        if poc_dev:
+            wx_poc_signal.set_pending(poc_dev, "meituan")
+            logger.info("wx_mp PoC: 已给测试设备 %s 打比价信号 source=meituan", poc_dev)
+        await _download(pic_url, openid, media_id)
+    else:
+        logger.info("wx_mp 收到消息(暂忽略): openid=%s type=%s", openid, msg_type)
+
+
+async def _download(pic_url: str, openid: str, media_id: str) -> None:
+    """用 PicUrl 直接下载图片落盘。PicUrl 是临时公网链接,无需 access_token / IP 白名单。"""
+    if not pic_url:
+        return
+    try:
+        _INBOX.mkdir(parents=True, exist_ok=True)
+        async with httpx.AsyncClient(timeout=15) as client:
+            resp = await client.get(pic_url)
+            resp.raise_for_status()
+        dest = _INBOX / f"{openid[:12]}_{media_id[:16]}.jpg"
+        dest.write_bytes(resp.content)
+        logger.info("wx_mp 图片已落盘: %s (%d bytes)", dest, len(resp.content))
+    except Exception:
+        logger.exception("wx_mp 图片下载失败 pic_url=%s", pic_url)

app/core/config.py

@@ -136,6 +136,40 @@ class Settings(BaseSettings):
     # (不跳授权、不拿 openid),整套微信代码保留。审核通过后 .env 置 true + 重启即启用,无需改代码。
     WX_MP_OAUTH_ENABLED: bool = False
 
+    # ===== 微信服务号(消息接收回调) =====
+    # 用户发消息给服务号 → 微信 POST 到 /wx/mp/callback(安全模式:Token 验签 + AESKey 解密)。
+    # 区别于上面的网页授权(那是落地页拿 openid);这里是被动收用户主动发来的消息(截图比价入口)。
+    WX_MP_TOKEN: str = ""            # 公众平台"服务器配置"的 Token(URL 验签 + 消息验签)
+    WX_MP_AES_KEY: str = ""          # EncodingAESKey(43 位, 消息 AES-256-CBC 加解密)
+
+    # ===== 微信截图比价 PoC(临时验证链路, 验证后删) =====
+    # 只对这台写死的测试设备下发"从美团比价"信号;空=不触发任何设备(部署上线零风险)。
+    # 收图 → 给此 device_id 打 pending → 该设备下次心跳带回 → 前端弹选平台窗。
+    WX_POC_TEST_DEVICE_ID: str = ""
+
+    # ===== 微信客服(企业微信;截图比价上游入口 —— 服务号版的平行实现) =====
+    # 用户在「微信客服」会话发消息 → 企业微信 POST 事件到 /wx/kf/callback → 后端 sync_msg 拉消息。
+    # 回调加密与公众号「安全模式」同源(复用 wx_mp_crypto), 但 receiveid 用 corpid 且 echostr 需解密。
+    # 拉消息 / 下载图片走 qyapi, 需 access_token(corpid + secret 换取)。全空=回调端点拒收, 零风险。
+    WX_KF_CORP_ID: str = ""          # 企业微信 corpid(回调解密 receiveid + 换 access_token)
+    WX_KF_SECRET: str = ""           # 「微信客服」Secret(换 access_token)
+    WX_KF_TOKEN: str = ""            # 客服「接收消息」回调的 Token(URL 验签 + 消息验签)
+    WX_KF_AES_KEY: str = ""          # 客服「接收消息」回调的 EncodingAESKey(AES-256-CBC 解密)
+
+    # ===== 微信会话内容存档(企业微信;"发给成员"那条路的截图/美团卡片比价上游入口) =====
+    # 用户把美团小程序卡片/截图发给企业微信【成员】(非客服) → 会话存档 GetChatData 轮询拉取 →
+    # RSA+AES 解密 → 识别 weapp/image → 给测试设备打比价信号。下游(心跳→弹窗→比价)复用。
+    # 需后台开通「会话内容存档」+ 上传 RSA 公钥 + 配可信IP + 拿存档 Secret;.so 放服务器。
+    # corpid 复用上面的 WX_KF_CORP_ID(同一企业)。ENABLED 默认关 → 未配好前零影响。
+    WX_FINANCE_ENABLED: bool = False           # 总开关(worker 是否启动)
+    WX_FINANCE_SECRET: str = ""                # 会话存档【专用】Secret(≠微信客服/自建应用)
+    WX_FINANCE_SDK_PATH: str = "./libWeWorkFinanceSdk_C.so"                 # WeWorkFinanceSdk C 库(Linux .so)
+    WX_FINANCE_PRIVATE_KEY_PATH: str = "./secrets/wx_finance_private.pem"   # RSA 私钥(解 encrypt_random_key)
+    WX_FINANCE_POLL_INTERVAL_SEC: float = 0.5  # GetChatData 轮询间隔秒(无回调、自己定频;支持亚秒如 0.5)
+    WX_FINANCE_SEQ_FILE: str = "./data/wx_finance_seq.txt"                  # seq 游标持久化(防丢/重复)
+    WX_FINANCE_RECEIVER_USERID: str = ""       # 接收成员 userid(万朗杰);只处理【非他发】的消息, 空=不过滤
+    WX_FINANCE_INTERNAL_URL: str = "http://127.0.0.1:8770"   # poller 通知 app-server 打信号的内部地址(同机)
+
     @property
     def wx_mp_configured(self) -> bool:
         """服务号网页授权凭证齐全(缺则落地页不发起授权,降级为无 openid)。"""
@@ -146,6 +180,26 @@ class Settings(BaseSettings):
         """落地页是否真正发起微信授权 = 凭证齐全 且 总开关开。"""
         return self.wx_mp_configured and self.WX_MP_OAUTH_ENABLED
 
+    @property
+    def wx_mp_callback_configured(self) -> bool:
+        """消息接收回调凭证齐全(Token + AESKey + AppID)。缺则回调端点拒绝处理消息。"""
+        return bool(self.WX_MP_TOKEN and self.WX_MP_AES_KEY and self.WX_MP_APPID)
+
+    @property
+    def wx_kf_callback_configured(self) -> bool:
+        """微信客服回调凭证齐全(CorpId + Token + AESKey + Secret)。缺则回调端点拒绝处理。"""
+        return bool(
+            self.WX_KF_CORP_ID
+            and self.WX_KF_TOKEN
+            and self.WX_KF_AES_KEY
+            and self.WX_KF_SECRET
+        )
+
+    @property
+    def wx_finance_configured(self) -> bool:
+        """会话存档 worker 是否该启动 = 总开关开 且 corpid + 存档 Secret 齐全。"""
+        return bool(self.WX_FINANCE_ENABLED and self.WX_KF_CORP_ID and self.WX_FINANCE_SECRET)
+
     # ===== 微信支付(商家转账到零钱 / 提现)=====
     # 真实凭证放 .env(已 gitignore),证书 .pem 放 secrets/。WECHAT_APP_ID 同时用于
     # 微信登录(code 换 openid)与转账,必须与 App 端开放平台 appid 一致。

app/core/config_schema.py

@@ -96,4 +96,19 @@ CONFIG_DEFS: dict[str, dict[str, Any]] = {
         "group": "首页轮播", "type": "enum", "hidden": True,
         "help": "mixed=真实优先+种子补位(默认);real=只用真实比价记录;seed=只用种子/合成(演示)。",
     },
+    # 比价 LLM 调用成本计价。值是嵌套 JSON(非 str→int),借 dict_str_int 类型在配置页走原始 JSON
+    # 编辑框;set_value 不校验类型,嵌套 JSON 照存。
+    "llm_token_price": {
+        "default": {
+            "per_model": {"qwen3.5-flash": {"input_per_1m": 0.8, "output_per_1m": 2.0}},
+            "default": {"input_per_1m": 3.0, "output_per_1m": 15.0},
+            "currency": "CNY", "unit": "per_1m_tokens",
+        },
+        "label": "LLM 模型单价(元/百万 token)",
+        "group": "LLM 成本", "type": "dict_str_int",
+        "help": (
+            "比价 LLM 调用成本计价。JSON:per_model 按模型配 input/output 单价(元/1M token),"
+            "default 兜底未登记的模型。改价只影响之后回填的新记录,历史记录用当时价格快照。"
+        ),
+    },
 }

app/core/wx_poc_signal.py

@@ -0,0 +1,27 @@
+"""微信截图比价 PoC:进程内存的"待比价"信号(单 worker 够用, 重启即失效, PoC 可接受)。
+
+收图端点 set_pending(device_id, source) → 该设备下次心跳 pop_pending 取走并清除 →
+心跳响应带回 → 前端弹选平台窗。只对 settings.WX_POC_TEST_DEVICE_ID 写入, 不碰其他设备。
+后续接真识别 / openid↔device 绑定时整体替换本模块。
+"""
+from __future__ import annotations
+
+import threading
+
+_lock = threading.Lock()
+_pending: dict[str, str] = {}  # device_id -> source_platform(如 "meituan")
+
+
+def set_pending(device_id: str, source_platform: str) -> None:
+    if not device_id:
+        return
+    with _lock:
+        _pending[device_id] = source_platform
+
+
+def pop_pending(device_id: str) -> str | None:
+    """取出并清除该设备的待比价信号;无则 None。心跳每帧调, 取到即消费(只弹一次)。"""
+    if not device_id:
+        return None
+    with _lock:
+        return _pending.pop(device_id, None)

app/integrations/wx_finance_sdk.py

@@ -0,0 +1,100 @@
+"""微信会话内容存档 SDK 封装(WeWorkFinanceSdk C 库的 ctypes 绑定 + 消息解密)。
+
+企业微信「会话内容存档」是拿"成员↔外部用户私聊消息正文"(含美团小程序卡片)的官方途径:
+主动轮询 GetChatData(seq) 拉取。每条消息两段密文:
+  - encrypt_random_key:用【企业自持 RSA 私钥】解出 AES 密钥
+  - encrypt_chat_msg :SDK DecryptData 用该密钥解出明文消息 JSON
+本模块封装:加载 .so → Init → GetChatData → RSA 解密随机密钥 → DecryptData。纯外部依赖,
+不含业务逻辑(识别 weapp/image → 打信号在 core/wx_finance_worker.py)。
+
+.so 需放服务器(Linux `libWeWorkFinanceSdk_C.so`),路径 settings.WX_FINANCE_SDK_PATH。
+⚠️ CDLL 在实例化时才加载(非 import 期),故本模块在无 .so 的机器上也能安全 import。
+"""
+from __future__ import annotations
+
+import base64
+import ctypes
+import json
+import logging
+from typing import Any
+
+from cryptography.hazmat.primitives.asymmetric import padding
+from cryptography.hazmat.primitives.serialization import load_pem_private_key
+
+logger = logging.getLogger("shagua.wx_finance")
+
+
+class WxFinanceError(RuntimeError):
+    """SDK 调用返回非 0 / 业务 errcode 时抛。"""
+
+
+class WxFinanceSdk:
+    """WeWorkFinanceSdk 薄封装。实例化即加载 .so + Init(需 corpid/secret 正确 + 服务器 IP 在可信IP)。"""
+
+    def __init__(self, sdk_path: str, corpid: str, secret: str, private_key_pem: bytes) -> None:
+        self._priv = load_pem_private_key(private_key_pem, password=None)
+        self._lib = ctypes.CDLL(sdk_path)  # 缺 .so / 缺 libssl 依赖会在此抛 OSError
+        self._bind()
+        self._sdk = self._lib.NewSdk()
+        ret = self._lib.Init(self._sdk, corpid.encode(), secret.encode())
+        if ret != 0:
+            raise WxFinanceError(f"Init 失败 ret={ret}(检查 corpid / 会话存档 Secret / 可信IP)")
+
+    def _bind(self) -> None:
+        lib = self._lib
+        lib.NewSdk.restype = ctypes.c_void_p
+        lib.Init.argtypes = [ctypes.c_void_p, ctypes.c_char_p, ctypes.c_char_p]
+        lib.Init.restype = ctypes.c_int
+        lib.GetChatData.argtypes = [
+            ctypes.c_void_p, ctypes.c_ulonglong, ctypes.c_uint,
+            ctypes.c_char_p, ctypes.c_char_p, ctypes.c_int, ctypes.c_void_p,
+        ]
+        lib.GetChatData.restype = ctypes.c_int
+        # ⚠️ DecryptData 不吃 sdk 句柄(与 GetChatData 不同, 它是纯解密函数)——只有 3 个参数,
+        # 多传 sdk 会让参数错位、encrypt_key 收到 sdk 指针 → DecryptData 返 10008(解析 encrypt_key 出错)。
+        lib.DecryptData.argtypes = [ctypes.c_char_p, ctypes.c_char_p, ctypes.c_void_p]
+        lib.DecryptData.restype = ctypes.c_int
+        lib.NewSlice.restype = ctypes.c_void_p
+        lib.FreeSlice.argtypes = [ctypes.c_void_p]
+        lib.GetContentFromSlice.argtypes = [ctypes.c_void_p]
+        lib.GetContentFromSlice.restype = ctypes.c_void_p
+        lib.GetSliceLen.argtypes = [ctypes.c_void_p]
+        lib.GetSliceLen.restype = ctypes.c_int
+        lib.DestroySdk.argtypes = [ctypes.c_void_p]
+
+    def _slice_bytes(self, slc: int) -> bytes:
+        ptr = self._lib.GetContentFromSlice(slc)
+        length = self._lib.GetSliceLen(slc)
+        return ctypes.string_at(ptr, length)
+
+    def get_chat_data(self, seq: int, limit: int = 1000, timeout: int = 10) -> list[dict[str, Any]]:
+        """拉 seq 之后的消息(返回从 seq+1 起)。返回 chatdata 列表(每项含 seq/encrypt_random_key/encrypt_chat_msg)。"""
+        slc = self._lib.NewSlice()
+        try:
+            ret = self._lib.GetChatData(self._sdk, int(seq), int(limit), None, None, int(timeout), slc)
+            if ret != 0:
+                raise WxFinanceError(f"GetChatData 失败 ret={ret}")
+            data = json.loads(self._slice_bytes(slc).decode("utf-8"))
+        finally:
+            self._lib.FreeSlice(slc)
+        if data.get("errcode"):
+            raise WxFinanceError(f"GetChatData errcode={data.get('errcode')} {data.get('errmsg')}")
+        return data.get("chatdata", [])
+
+    def decrypt(self, encrypt_random_key_b64: str, encrypt_chat_msg: str) -> dict[str, Any]:
+        """RSA 私钥解 encrypt_random_key → 得 AES 密钥 → DecryptData 解 encrypt_chat_msg → 明文 dict。"""
+        # 企业微信用你上传的 RSA 公钥(PKCS1)加密随机密钥, 这里用对应私钥解出, 原样交给 DecryptData。
+        aes_key = self._priv.decrypt(base64.b64decode(encrypt_random_key_b64), padding.PKCS1v15())
+        slc = self._lib.NewSlice()
+        try:
+            ret = self._lib.DecryptData(aes_key, encrypt_chat_msg.encode(), slc)
+            if ret != 0:
+                raise WxFinanceError(f"DecryptData 失败 ret={ret}")
+            return json.loads(self._slice_bytes(slc).decode("utf-8"))
+        finally:
+            self._lib.FreeSlice(slc)
+
+    def close(self) -> None:
+        if getattr(self, "_sdk", None):
+            self._lib.DestroySdk(self._sdk)
+            self._sdk = None

app/integrations/wx_kf_client.py

@@ -0,0 +1,126 @@
+"""微信客服(企业微信)API 客户端:access_token 缓存 + sync_msg 拉消息 + 临时素材下载。
+
+对应 api/v1/wx_kf.py:收到 `kf_msg_or_event` 回调后, 用这里的函数增量拉取用户消息。
+  - access_token(7200s)进程内缓存, 提前 300s 刷新(单 worker 够用, 重启即失效)。
+  - sync_msg 游标 next_cursor 按 open_kfid 进程内存续(PoC 重启后从最近 3 天重拉, 可接受)。
+凭证来自 settings(WX_KF_CORP_ID / WX_KF_SECRET);纯外部 HTTP, 不含业务逻辑。
+"""
+from __future__ import annotations
+
+import asyncio
+import logging
+import time
+from typing import Any
+
+import httpx
+
+from app.core.config import settings
+
+logger = logging.getLogger("shagua.wx_kf")
+
+_QYAPI = "https://qyapi.weixin.qq.com/cgi-bin"
+
+# access_token 进程内缓存(单 worker)。token 空 / 未到刷新点直接用。
+_token_lock = asyncio.Lock()
+_token_cache: dict[str, Any] = {"token": "", "expire_at": 0.0}
+
+# sync_msg 游标:open_kfid -> next_cursor(PoC 进程内存;重启从最近 3 天重拉)
+_cursors: dict[str, str] = {}
+
+
+async def _get_access_token() -> str:
+    """取企业微信 access_token, 进程内缓存, 提前 300s 过期刷新。失败返回空串(调用方降级)。"""
+    now = time.time()
+    if _token_cache["token"] and _token_cache["expire_at"] - 300 > now:
+        return _token_cache["token"]
+    async with _token_lock:
+        now = time.time()  # 拿锁后复检, 避免并发重复刷新
+        if _token_cache["token"] and _token_cache["expire_at"] - 300 > now:
+            return _token_cache["token"]
+        try:
+            async with httpx.AsyncClient(timeout=10) as client:
+                resp = await client.get(
+                    f"{_QYAPI}/gettoken",
+                    params={
+                        "corpid": settings.WX_KF_CORP_ID,
+                        "corpsecret": settings.WX_KF_SECRET,
+                    },
+                )
+                data = resp.json()
+        except Exception:
+            logger.exception("wx_kf gettoken 请求失败")
+            return ""
+        if data.get("errcode"):
+            logger.warning(
+                "wx_kf gettoken errcode=%s errmsg=%s", data.get("errcode"), data.get("errmsg")
+            )
+            return ""
+        _token_cache["token"] = data.get("access_token", "")
+        _token_cache["expire_at"] = now + int(data.get("expires_in", 7200))
+        return _token_cache["token"]
+
+
+async def sync_messages(token: str, open_kfid: str) -> list[dict[str, Any]]:
+    """收到 kf_msg_or_event 回调后调用:用回调带的一次性 token 增量拉消息, has_more 循环拉完。
+
+    token:    回调事件里的 <Token>(消息拉取凭证, 非配置 Token;不传会有严格频控)。
+    open_kfid:客服账号 id(回调里的 <OpenKfId>)。
+    返回本次新拉到的 msg_list(已合并多页);游标按 open_kfid 进程内存续。
+    """
+    access_token = await _get_access_token()
+    if not access_token:
+        return []
+    out: list[dict[str, Any]] = []
+    cursor = _cursors.get(open_kfid, "")
+    try:
+        async with httpx.AsyncClient(timeout=15) as client:
+            for _ in range(20):  # 最多 20 页护栏, 防异常时死循环
+                body: dict[str, Any] = {"cursor": cursor, "token": token, "limit": 1000}
+                if open_kfid:
+                    body["open_kfid"] = open_kfid
+                resp = await client.post(
+                    f"{_QYAPI}/kf/sync_msg",
+                    params={"access_token": access_token},
+                    json=body,
+                )
+                data = resp.json()
+                if data.get("errcode"):
+                    logger.warning(
+                        "wx_kf sync_msg errcode=%s errmsg=%s",
+                        data.get("errcode"),
+                        data.get("errmsg"),
+                    )
+                    break
+                out.extend(data.get("msg_list", []))
+                cursor = data.get("next_cursor", cursor)
+                _cursors[open_kfid] = cursor
+                if not data.get("has_more"):
+                    break
+    except Exception:
+        logger.exception("wx_kf sync_msg 请求失败 open_kfid=%s", open_kfid)
+    return out
+
+
+async def download_media(media_id: str) -> bytes | None:
+    """临时素材下载(图片消息只给 media_id)。成功返回二进制, 失败返回 None。"""
+    if not media_id:
+        return None
+    access_token = await _get_access_token()
+    if not access_token:
+        return None
+    try:
+        async with httpx.AsyncClient(timeout=20) as client:
+            resp = await client.get(
+                f"{_QYAPI}/media/get",
+                params={"access_token": access_token, "media_id": media_id},
+            )
+            resp.raise_for_status()
+            # 成功=二进制文件流;失败=JSON(errcode)。据 Content-Type 区分。
+            ctype = resp.headers.get("Content-Type", "")
+            if "application/json" in ctype or "text/plain" in ctype:
+                logger.warning("wx_kf media/get 非文件响应: %s", resp.text[:200])
+                return None
+            return resp.content
+    except Exception:
+        logger.exception("wx_kf media/get 下载失败 media_id=%s", media_id)
+        return None

app/integrations/wx_mp_crypto.py

@@ -0,0 +1,69 @@
+"""微信服务号消息接收回调的验签与解密(安全模式)。
+
+服务号"服务器配置"选安全模式后:
+  - URL 接入验证(GET): sha1(sort(token, timestamp, nonce)) == signature → 原样返回 echostr
+  - 消息(POST): body 里 <Encrypt> 是密文;
+      msg_signature = sha1(sort(token, timestamp, nonce, encrypt))
+      密文 AES-256-CBC 解出: random(16B) + msg_len(4B big-endian) + msg + appid, 去 PKCS7(块=32) padding
+
+只做验签 + 解密(收消息)。被动回复(加密)MVP 不需要 —— 收到后走客服消息异步回执。
+密钥/口令来自 settings(WX_MP_TOKEN / WX_MP_AES_KEY / WX_MP_APPID),本模块只做纯算法、不读配置。
+"""
+from __future__ import annotations
+
+import base64
+import hashlib
+import hmac
+
+from cryptography.hazmat.primitives.ciphers import Cipher, algorithms, modes
+
+
+def verify_url_signature(token: str, timestamp: str, nonce: str, signature: str) -> bool:
+    """GET 接入验证:token/timestamp/nonce 三者字典序排序拼接后 sha1。"""
+    return _consteq(_sha1(token, timestamp, nonce), signature)
+
+
+def verify_msg_signature(
+    token: str, timestamp: str, nonce: str, encrypt: str, msg_signature: str
+) -> bool:
+    """POST 消息验签:四者(含密文 encrypt)字典序排序拼接后 sha1。"""
+    return _consteq(_sha1(token, timestamp, nonce, encrypt), msg_signature)
+
+
+def decrypt_message(aes_key_b64: str, expected_appid: str, encrypt_b64: str) -> str:
+    """解密 <Encrypt> 密文, 返回明文消息 XML。appid 不符抛 ValueError。
+
+    aes_key_b64: EncodingAESKey(43 位, 不含结尾 '='), 补 '=' 后 base64 解出 32 字节 AES-256 key。
+    """
+    aes_key = base64.b64decode(aes_key_b64 + "=")  # 43 → 32 bytes
+    iv = aes_key[:16]
+    decryptor = Cipher(algorithms.AES(aes_key), modes.CBC(iv)).decryptor()
+    plain = decryptor.update(base64.b64decode(encrypt_b64)) + decryptor.finalize()
+    plain = _pkcs7_unpad(plain)
+
+    # random(16) + msg_len(4, big-endian) + msg(msg_len) + from_appid
+    content = plain[16:]
+    msg_len = int.from_bytes(content[:4], "big")
+    msg = content[4 : 4 + msg_len]
+    from_appid = content[4 + msg_len :].decode("utf-8")
+    if expected_appid and from_appid != expected_appid:
+        raise ValueError(f"appid mismatch: {from_appid!r} != {expected_appid!r}")
+    return msg.decode("utf-8")
+
+
+def _sha1(*parts: str) -> str:
+    return hashlib.sha1("".join(sorted(parts)).encode("utf-8")).hexdigest()
+
+
+def _consteq(a: str, b: str) -> bool:
+    return hmac.compare_digest(a, b)
+
+
+def _pkcs7_unpad(data: bytes) -> bytes:
+    """微信用块大小 32 的 PKCS7,末字节即 padding 长度(1..32)。越界则原样返回(容错)。"""
+    if not data:
+        return data
+    pad = data[-1]
+    if pad < 1 or pad > 32:
+        return data
+    return data[:-pad]

app/main.py

@@ -19,6 +19,7 @@ from app.api.internal.app_version import router as internal_app_version_router
 from app.api.internal.launch_confirm import router as internal_launch_confirm_router
 from app.api.internal.price import router as internal_price_router
 from app.api.internal.store import router as internal_store_router
+from app.api.internal.wx_finance import router as internal_wx_finance_router
 from app.api.v1.ad import router as ad_router
 from app.api.v1.analytics import router as analytics_router
 from app.api.v1.auth import router as auth_router
@@ -39,6 +40,8 @@ from app.api.v1.signin import router as signin_router
 from app.api.v1.tasks import router as tasks_router
 from app.api.v1.user import router as user_router
 from app.api.v1.wallet import router as wallet_router
+from app.api.v1.wx_kf import router as wx_kf_router
+from app.api.v1.wx_mp import router as wx_mp_router
 from app.api.v1.wxpay import router as wxpay_router
 from app.core.config import settings
 from app.core.daily_exchange_worker import (
@@ -137,9 +140,15 @@ app.include_router(internal_price_router)
 app.include_router(internal_store_router)
 app.include_router(internal_app_version_router)
 app.include_router(internal_launch_confirm_router)
+# 会话存档独立 poller → 打比价信号(server→server, X-Internal-Secret;轮询进程在 app 外, 见 scripts/wx_finance_poller.py)
+app.include_router(internal_wx_finance_router)
 app.include_router(platform_router)
 # CPS 群发短链跳转 /c/{code}(公网无鉴权:记点击 → 302 跳美团)
 app.include_router(cps_redirect_router)
+# 微信服务号消息接收回调 /wx/mp/callback(公网无鉴权:微信服务器验签, 截图比价入口)
+app.include_router(wx_mp_router)
+# 微信客服(企业微信)消息接收回调 /wx/kf/callback(服务号版的平行实现, 收事件→sync_msg 拉图)
+app.include_router(wx_kf_router)
 
 # 用户上传文件(头像)静态服务。生产可改由 nginx 直接 serve MEDIA_ROOT。
 _media_root = Path(settings.MEDIA_ROOT)

app/models/comparison.py

@@ -137,6 +137,12 @@ class ComparisonRecord(Base):
     # 每次 LLM 调用明细 [{scene,model,input_messages,output,usage,latency_ms,error}];
     # server 收上报后按 trace_id 同机拉 pricebot 落库(见 compare_record 端点)。旧记录/未采集为 None。
     llm_calls: Mapped[list | None] = mapped_column(_JSON, nullable=True)
+    # 本次比价 LLM 总成本(元):回填时按「当时的价」逐模型算好冻结(见 services/llm_cost.py)。
+    # 单次亚分级 → float「元」(不用 *_cents)。旧记录/未回填为 None,前端回退「估算成本」。
+    llm_cost_yuan: Mapped[float | None] = mapped_column(Float, nullable=True)
+    # 算成本所用单价快照 {mode, prices:{model:{input_per_1m,output_per_1m,_source}}}:app_config 只存
+    # 当前价、不留历史,故把当时价冻结进来供审计/复算。
+    llm_price_snapshot: Mapped[dict | None] = mapped_column(_JSON, nullable=True)
 
     created_at: Mapped[datetime] = mapped_column(
         DateTime(timezone=True), server_default=func.now(), index=True, nullable=False

app/schemas/device.py

@@ -36,6 +36,20 @@ class OkResponse(BaseModel):
     ok: bool = True
 
 
+class PendingCompare(BaseModel):
+    """PoC:后端通过心跳下发的"用户要从某源平台比价"信号。前端据此弹选平台窗 + launch 源平台。"""
+    source_platform: str            # pricebot 源平台代号(PoC 写死 "meituan")
+    source_package: str             # 源平台 Android 包名(前端 launch 用)
+
+
+class HeartbeatResponse(BaseModel):
+    """心跳响应。常规只回 ok;PoC 期带回 pending_compare 触发比价(exclude_none 省略 null)。"""
+    model_config = ConfigDict()
+
+    ok: bool = True
+    pending_compare: PendingCompare | None = None
+
+
 class LivenessOut(BaseModel):
     """本机掉线告警状态(后置检测 pull 版)。客户端只需这一个布尔判断要不要弹「开启自启动」引导,
     故只返回 kill_alert_pending(不暴露设备详情 / 内部 liveness_state 等)。从未注册过 → 默认 False(无告警)。"""

app/schemas/wx_finance.py

@@ -0,0 +1,13 @@
+"""会话存档独立 poller → app-server 内部信号端点的 schema。"""
+from __future__ import annotations
+
+from pydantic import BaseModel
+
+
+class WxFinancePendingIn(BaseModel):
+    source: str = "meituan"   # 源平台代号(pop 后前端 launch 用;PoC 写死 meituan)
+    kind: str = ""            # 触发物类型 weapp / image(仅日志用)
+
+
+class WxFinancePendingOut(BaseModel):
+    ok: bool = True

app/services/llm_cost.py

@@ -0,0 +1,53 @@
+"""LLM 调用成本计算(纯逻辑,无 DB):按 model 分桶累加 token × 单价,返回总成本(元)+ 价格快照。
+
+用量取自 comparison_record.llm_calls[].usage(pricebot 已归一为 prompt/completion_tokens);
+error / 无 usage 的调用跳过。price_cfg = {per_model:{model:{input_per_1m,output_per_1m}}, default:{...}}。
+成本单位「元」——单次亚分级,用 float(不用 *_cents);snapshot 只含本次用到的模型的价(审计用,
+不存整张价表)。用到但没配价(既无 per_model 又无 default)的模型 → 快照标 unpriced,成本按 0 计。
+"""
+from __future__ import annotations
+
+_PRICE_KEY = "llm_token_price"
+
+
+def get_llm_prices(db) -> dict:
+    """读 LLM 单价配置(app_config;表内无则回退 CONFIG_DEFS 默认)。返回 compute_llm_cost 的 price_cfg。"""
+    from app.repositories import app_config  # 延迟 import:compute_llm_cost 纯逻辑不牵连 DB 层
+    return app_config.get_value(db, _PRICE_KEY)
+
+
+def compute_llm_cost(calls: list[dict], price_cfg: dict) -> tuple[float | None, dict | None]:
+    """遍历 calls 按 model 分桶,cost = Σ(入/1e6*入价 + 出/1e6*出价);无有效调用 → (None, None)。"""
+    if not calls:
+        return None, None
+    per_model = price_cfg.get("per_model") or {}
+    default = price_cfg.get("default")
+    buckets: dict[str, list[int]] = {}  # model -> [Σprompt_tokens, Σcompletion_tokens]
+    for c in calls:
+        if c.get("error"):
+            continue
+        usage = c.get("usage") or {}
+        model = c.get("model") or "unknown"
+        b = buckets.setdefault(model, [0, 0])
+        b[0] += usage.get("prompt_tokens") or 0
+        b[1] += usage.get("completion_tokens") or 0
+    if not buckets:  # 全是 error / 无 usage
+        return None, None
+    total = 0.0
+    prices: dict[str, dict] = {}
+    for model, (tin, tout) in buckets.items():
+        price = per_model.get(model, default)
+        in_p = price.get("input_per_1m") if isinstance(price, dict) else None
+        out_p = price.get("output_per_1m") if isinstance(price, dict) else None
+        # 没配价 / 无 default / 单价残缺或非法(配置页手改 JSON 可能存出脏数据)→ 标记待补价、
+        # 不计入成本;绝不抛异常,以免连累同一回填里的 token/llm_calls 落库。
+        if not isinstance(in_p, (int, float)) or not isinstance(out_p, (int, float)):
+            prices[model] = {"input_per_1m": in_p, "output_per_1m": out_p, "unpriced": True}
+            continue
+        total += tin / 1e6 * in_p + tout / 1e6 * out_p
+        prices[model] = {
+            "input_per_1m": in_p,
+            "output_per_1m": out_p,
+            "_source": "per_model" if model in per_model else "default",
+        }
+    return round(total, 6), {"mode": "per_model", "prices": prices}

deploy/wx-finance-poller.service

@@ -0,0 +1,18 @@
+[Unit]
+# 会话存档轮询 poller —— 独立进程跑 WeWorkFinanceSdk(Go c-shared .so),与 app-server 隔离:
+# 嵌进 app-server 会 segfault 把主进程带崩,拆出来后崩了 systemd 只重启本服务、不影响 app-server。
+Description=Shaguabijia 会话存档轮询 poller (WeWorkFinanceSdk, isolated from app-server)
+After=network.target shaguabijia-app-server.service
+
+[Service]
+Type=simple
+User=root
+WorkingDirectory=/opt/shaguabijia-app-server
+Environment="PATH=/opt/shaguabijia-app-server/.venv/bin:/usr/bin:/bin"
+EnvironmentFile=/opt/shaguabijia-app-server/.env
+ExecStart=/opt/shaguabijia-app-server/.venv/bin/python scripts/wx_finance_poller.py
+Restart=on-failure
+RestartSec=5
+
+[Install]
+WantedBy=multi-user.target

docs/database/comparison_record.md

@@ -40,6 +40,8 @@
 | `raw_payload` | JSON(PG: JSONB) | nullable | 客户端原始上报全量(calibration + done.params),取数兜底 |
 | `input_tokens` | Integer | nullable | 本次 LLM 累计输入 token = Σ `llm_calls[].usage.prompt_tokens`(server 收上报后从 `llm_calls` 累加;旧记录/未采集为 null) |
 | `output_tokens` | Integer | nullable | 本次 LLM 累计输出 token = Σ `llm_calls[].usage.completion_tokens`(同上) |
+| `llm_cost_yuan` | Float | nullable | 本次比价 LLM 总成本(元),回填时按「当时价」逐模型算好冻结(见 `services/llm_cost.py`);旧记录/未回填为 null → 前端回退「估算成本」 |
+| `llm_price_snapshot` | JSON(PG: JSONB) | nullable | 算成本所用单价快照 `{mode, prices:{model:{input_per_1m,output_per_1m,_source}}}`;`app_config` 只存当前价、不留历史,故冻结当时价供审计/复算 |
 | `created_at` | DateTime(tz) | server_default now(), index | 时间 |
 
 > `ordered`(已下单)是**瞬态字段**,不在表里:`list_records` 读取时按 `store_name ∈ 该用户 source='compare' 的 savings_record.shop_name 集合` 现挂到实例上供出参用。

scripts/seed_mock_llm_cost.py

@@ -0,0 +1,248 @@
+"""一次性 mock:造带 LLM token 成本的比价记录 + 配好 app_config 模型单价,用于测「管理后端」LLM 成本展示。
+
+覆盖 admin「比价记录」详情抽屉的「LLM 成本」展示分支:
+  • app_config.llm_token_price ← 写一条多模型单价(= 配置页「LLM 成本」卡片「已改」态,get_llm_prices 读它)
+  • comparison_record ← 造 5 条,逐条**复用生产的 compute_llm_cost + 与 _backfill_llm_calls 同款派生**
+    (llm_call_count/retry_count/input_tokens/output_tokens/llm_cost_yuan/llm_price_snapshot),
+    确保 mock 行 = 真实回填产出。5 条刻意覆盖:
+      ① 单模型真实样本(qwen3.5-flash ×4) → ¥0.006184(核对精确值)
+      ② 多模型(flash + plus)          → 快照含两个模型、各自 _source=per_model
+      ③ 未登记模型(deepseek-v3)        → 走 default,快照 _source=default
+      ④ 旧记录(有 token、无 cost)       → llm_cost_yuan=NULL → 前端回退「估算成本」
+      ⑤ 含 error 调用                    → error 那次跳过计费、retry_count+1
+
+记录挂到库里第一个真实用户(admin 列表能显示手机号);无用户则 user_id=NULL(孤儿行,admin 照样全看)。
+created_at 用北京 naive、最近几分钟内错开,详情列表倒序即 ①→⑤ 置顶。
+
+幂等:重跑先按 trace_id 前缀「MOCKLLM-」清旧再建。app_config 单价是 upsert(不随 --clean-only 删,
+因该 key 本就是本需求新增、无历史真实值;要改价直接去配置页或重跑本脚本)。
+
+  python -m scripts.seed_mock_llm_cost              # 造价格 + 5 条记录
+  python -m scripts.seed_mock_llm_cost --clean-only # 只清 MOCKLLM- 记录(保留单价)
+
+验收:admin「比价记录」→ 找 trace「MOCKLLM-」的 5 条 → 点开详情看「LLM 成本」:
+  ①②③⑤ 显示「实际·当时价」+ 价格快照;④ 显示「估算」。
+"""
+from __future__ import annotations
+
+import argparse
+import sys
+from datetime import datetime, timedelta, timezone
+
+from sqlalchemy import delete, select
+
+from app.db.session import SessionLocal
+from app.models.comparison import ComparisonRecord
+from app.models.user import User
+from app.repositories import app_config
+from app.services.llm_cost import compute_llm_cost
+
+if hasattr(sys.stdout, "reconfigure"):
+    sys.stdout.reconfigure(encoding="utf-8")  # Windows 控制台输出中文/¥
+
+_BJ = timezone(timedelta(hours=8))
+ID_PREFIX = "MOCKLLM-"
+
+# ── 写进 app_config 的模型单价(get_llm_prices 读它;配置页「LLM 成本」卡片可再改)──
+PRICE_CFG = {
+    "per_model": {
+        "qwen3.5-flash": {"input_per_1m": 0.8, "output_per_1m": 2.0},
+        "qwen3.5-plus": {"input_per_1m": 4.0, "output_per_1m": 12.0},
+    },
+    "default": {"input_per_1m": 3.0, "output_per_1m": 15.0},
+    "currency": "CNY",
+    "unit": "per_1m_tokens",
+}
+
+
+def _c(scene: str, model: str, pin: int, cout: int, error: str | None = None) -> dict:
+    """一条 llm_calls 明细,结构对齐真实 pricebot 归一后契约:
+    {scene, model, input_messages:[{role,content}], output, usage:{prompt/completion/total_tokens},
+     latency_ms, error}(详情抽屉会遍历 input_messages,缺了会崩)。error 的调用无 usage/output。"""
+    return {
+        "scene": scene,
+        "model": model,
+        "error": error,
+        "input_messages": [
+            {"role": "system", "content": f"你是比价助手,负责 {scene} 环节。"},
+            {"role": "user", "content": f"[mock] 请处理本次比价的 {scene} 任务。"},
+        ],
+        "output": None if error else f"[mock] {scene} 环节完成。",
+        "usage": None if error else {
+            "prompt_tokens": pin, "completion_tokens": cout, "total_tokens": pin + cout,
+        },
+        "latency_ms": 780,
+    }
+
+
+# ── 5 条记录蓝本:calls 决定成本;freeze=False 模拟旧记录(有 token 无 cost)──
+RECORDS = [
+    {
+        "label": "①单模型·真实样本",
+        "source": ("美团外卖", 4280), "best": ("京东秒送", 3680),
+        "store": "肯德基(建国路店)", "product": "疯狂星期四全家桶",
+        "info": "在京东秒送找到同款,到手价 ¥36.80,省 ¥6.00",
+        "freeze": True,
+        "calls": [
+            _c("store_match", "qwen3.5-flash", 1512, 22),
+            _c("dish_match", "qwen3.5-flash", 2111, 160),
+            _c("dish_match", "qwen3.5-flash", 1940, 142),
+            _c("summary", "qwen3.5-flash", 1325, 13),
+        ],
+    },
+    {
+        "label": "②多模型·flash+plus",
+        "source": ("淘宝闪购", 5900), "best": ("美团外卖", 5200),
+        "store": "瑞幸咖啡(国贸店)", "product": "生椰拿铁×2、丝绒拿铁",
+        "info": "在美团外卖找到同款,到手价 ¥52.00,省 ¥7.00",
+        "freeze": True,
+        "calls": [
+            _c("store_match", "qwen3.5-flash", 2000, 50),
+            _c("dish_match", "qwen3.5-flash", 1800, 40),
+            _c("reasoning", "qwen3.5-plus", 3000, 500),
+        ],
+    },
+    {
+        "label": "③未登记模型走 default",
+        "source": ("京东秒送", 3100), "best": ("美团外卖", 2650),
+        "store": "麦当劳(soho店)", "product": "麦辣鸡腿堡套餐",
+        "info": "在美团外卖找到同款,到手价 ¥26.50,省 ¥4.50",
+        "freeze": True,
+        "calls": [
+            _c("store_match", "deepseek-v3", 5000, 800),
+        ],
+    },
+    {
+        "label": "④旧记录·有token无成本(回退估算)",
+        "source": ("美团外卖", 3600), "best": ("淘宝闪购", 3200),
+        "store": "华莱士(双井店)", "product": "全鸡汉堡套餐",
+        "info": "在淘宝闪购找到同款,到手价 ¥32.00,省 ¥4.00",
+        "freeze": False,  # 模拟本需求上线前的老记录:llm_cost_yuan=NULL → 前端回退估算
+        "calls": [
+            _c("store_match", "qwen3.5-flash", 2000, 100),
+        ],
+    },
+    {
+        "label": "⑤含 error 调用(跳过计费)",
+        "source": ("淘宝闪购", 4100), "best": ("京东秒送", 3750),
+        "store": "海底捞(合生汇店)", "product": "番茄锅底、肥牛卷",
+        "info": "在京东秒送找到同款,到手价 ¥37.50,省 ¥3.50",
+        "freeze": True,
+        "calls": [
+            _c("store_match", "qwen3.5-flash", 0, 0, error="timeout"),
+            _c("store_match", "qwen3.5-flash", 1500, 30),
+        ],
+    },
+]
+
+_PLATFORM_ID = {  # 展示名 → 平台代号(comparison_results / source/best 列用)
+    "美团外卖": "meituan", "京东秒送": "jd", "淘宝闪购": "taobao",
+}
+
+
+def _naive_bj_now() -> datetime:
+    return datetime.now(_BJ).replace(tzinfo=None)
+
+
+def clean(db) -> int:
+    n = db.execute(
+        delete(ComparisonRecord).where(ComparisonRecord.trace_id.like(f"{ID_PREFIX}%"))
+    ).rowcount or 0
+    db.commit()
+    return n
+
+
+def _build_record(spec: dict, owner_id: int | None, created_at: datetime) -> tuple[ComparisonRecord, float | None]:
+    """按蓝本造一条记录,LLM 派生完全对齐 _backfill_llm_calls;返回 (记录, 冻结成本或 None)。"""
+    calls = spec["calls"]
+    src_name, src_cents = spec["source"]
+    best_name, best_cents = spec["best"]
+
+    # —— 与 _backfill_llm_calls 同款派生 ——
+    llm_call_count = len(calls)
+    retry_count = sum(1 for c in calls if c.get("error"))
+    input_tokens = sum((c.get("usage") or {}).get("prompt_tokens") or 0 for c in calls)
+    output_tokens = sum((c.get("usage") or {}).get("completion_tokens") or 0 for c in calls)
+    if spec["freeze"]:
+        cost, snapshot = compute_llm_cost(calls, PRICE_CFG)  # 复用生产纯函数
+    else:
+        cost, snapshot = None, None  # 旧记录:回填这段代码上线前就有,只有 token 没成本
+
+    rec = ComparisonRecord(
+        user_id=owner_id,
+        device_id=f"{ID_PREFIX.lower()}dev",
+        business_type="food",
+        trace_id=f"{ID_PREFIX}{spec['label'][0]}",  # ①..⑤ 各一,唯一
+        status="success",
+        source_platform_id=_PLATFORM_ID.get(src_name), source_platform_name=src_name,
+        source_price_cents=src_cents,
+        best_platform_id=_PLATFORM_ID.get(best_name), best_platform_name=best_name,
+        best_price_cents=best_cents,
+        saved_amount_cents=src_cents - best_cents,
+        is_source_best=False,
+        store_name=spec["store"],
+        product_names=spec["product"],
+        information=spec["info"],
+        items=[{"name": spec["product"], "qty": 1}],
+        comparison_results=[
+            {"platform_id": _PLATFORM_ID.get(src_name), "platform_name": src_name,
+             "price": src_cents / 100, "is_source": True, "rank": 2},
+            {"platform_id": _PLATFORM_ID.get(best_name), "platform_name": best_name,
+             "price": best_cents / 100, "is_source": False, "rank": 1},
+        ],
+        total_ms=90_000 + llm_call_count * 1000,
+        step_count=llm_call_count * 3,
+        llm_call_count=llm_call_count,
+        retry_count=retry_count,
+        input_tokens=input_tokens,
+        output_tokens=output_tokens,
+        llm_calls=calls,
+        llm_cost_yuan=cost,
+        llm_price_snapshot=snapshot,
+        created_at=created_at,
+    )
+    return rec, cost
+
+
+def seed(db) -> list[tuple[str, float | None]]:
+    app_config.set_value(db, "llm_token_price", PRICE_CFG, admin_id=None)  # upsert 单价
+    owner_id = db.execute(select(User.id).order_by(User.id).limit(1)).scalar()
+    base = _naive_bj_now()
+    out: list[tuple[str, float | None]] = []
+    for i, spec in enumerate(RECORDS):
+        rec, cost = _build_record(spec, owner_id, base - timedelta(minutes=i * 3))
+        db.add(rec)
+        out.append((spec["label"], cost))
+    db.commit()
+    return out, owner_id
+
+
+def main() -> None:
+    parser = argparse.ArgumentParser(description="造带 LLM 成本的比价记录 + app_config 模型单价(测管理后端)")
+    parser.add_argument("--clean-only", action="store_true", help="只清 MOCKLLM- 记录,不重建(保留单价)")
+    args = parser.parse_args()
+
+    db = SessionLocal()
+    try:
+        removed = clean(db)
+        if removed:
+            print(f"🧹 已清理旧 mock 记录 {removed} 条")
+        if args.clean_only:
+            print("✅ 仅清理,已完成(app_config 单价保留)。")
+            return
+
+        results, owner_id = seed(db)
+        print(f"\n✅ 已写入 app_config.llm_token_price(单价)+ {len(results)} 条比价记录"
+              f"(挂 user_id={owner_id or 'NULL(孤儿行)'})")
+        print("\n📋 每条冻结成本(admin 详情「LLM 成本」应显示):")
+        for label, cost in results:
+            shown = "NULL → 前端回退「估算」" if cost is None else f"¥{cost}"
+            print(f"  {label:<20} {shown}")
+        print("\n👉 验收:admin「比价记录」→ trace 搜「MOCKLLM-」→ 点开详情核对 LLM 成本 + 价格快照。")
+        print("   配置页「系统配置」→「福利页」Tab →「LLM 成本」卡片,单价应为「已改」态。")
+    finally:
+        db.close()
+
+
+if __name__ == "__main__":
+    main()

scripts/wx_finance_poller.py

@@ -0,0 +1,152 @@
+"""会话存档【独立进程】轮询 poller —— 单独 systemd 服务跑, 与 app-server 隔离。
+
+为什么独立进程:WeWorkFinanceSdk 是 Go 编译的 c-shared 库, 自带 Go runtime。嵌进 app-server(uvicorn)
+后台线程里跑会 segfault、把主进程带崩(2026-07-17 实测 core-dump 每 5s 循环)。独立进程让 Go runtime
+独占进程 + 主线程, 崩了 systemd 只重启本 poller、不动 app-server。拉到美团卡片(weapp)/截图(image)
+→ POST app-server 内部端点 /internal/wx-finance/pending 打比价信号(下游心跳→弹窗→比价 完全复用)。
+
+  手动跑:  cd /opt/shaguabijia-app-server && .venv/bin/python scripts/wx_finance_poller.py
+  生产:    deploy/wx-finance-poller.service 常驻(systemctl enable --now wx-finance-poller)
+"""
+from __future__ import annotations
+
+import json
+import logging
+import signal
+import time
+from pathlib import Path
+
+import httpx
+
+from app.core.config import settings
+from app.core.logging import setup_logging
+from app.integrations.wx_finance_sdk import WxFinanceError, WxFinanceSdk
+
+logger = logging.getLogger("shagua.wx_finance_poller")
+
+_stop = False
+
+
+def _on_signal(_signum: int, _frame: object) -> None:
+    global _stop
+    _stop = True
+
+
+def _load_seq() -> int:
+    try:
+        return int(Path(settings.WX_FINANCE_SEQ_FILE).read_text().strip() or "0")
+    except Exception:
+        return 0
+
+
+def _save_seq(seq: int) -> None:
+    try:
+        p = Path(settings.WX_FINANCE_SEQ_FILE)
+        p.parent.mkdir(parents=True, exist_ok=True)
+        p.write_text(str(seq))
+    except Exception:
+        logger.exception("seq 持久化失败")
+
+
+def _notify(client: httpx.Client, source: str, kind: str) -> None:
+    """POST 内部端点让 app-server 打信号(set_pending 在 app-server 进程内存)。"""
+    url = f"{settings.WX_FINANCE_INTERNAL_URL.rstrip('/')}/internal/wx-finance/pending"
+    try:
+        resp = client.post(
+            url,
+            json={"source": source, "kind": kind},
+            headers={"X-Internal-Secret": settings.INTERNAL_API_SECRET},
+            timeout=10,
+        )
+        logger.info("notify app-server (%s) → %s", kind, resp.status_code)
+    except Exception:
+        logger.exception("notify app-server 失败")
+
+
+def _resolve_source(msg: dict) -> str:
+    """从 weapp 小程序卡判源平台代号(= app-server _SOURCE_PACKAGES 的 key)。
+
+    会话存档 weapp 结构里 username(gh_id)/appid/pagepath/title 任一含平台特征即判定:
+    先按关键词启发式(覆盖大多数);认不出 → 默认 meituan(过渡期不阻断)+ WARNING,
+    并把整个 weapp 结构打进日志,便于照真实 username/appid 精确补规则。
+    """
+    weapp = msg.get("weapp") or {}
+    blob = json.dumps(weapp, ensure_ascii=False).lower()
+    # 排障日志:打全 weapp 结构(顺带 msg 顶层 key,防结构不在 weapp 下),便于补/改判定规则
+    logger.info("weapp 判源: msg_keys=%s weapp=%s", list(msg.keys()), blob[:1500])
+    if any(k in blob for k in ("jingdong", "jd.com", "jddj", "京东")):
+        return "jd"
+    if any(k in blob for k in ("meituan", "sankuai", "美团")):
+        return "meituan"
+    logger.warning("weapp 源平台未识别, 暂默认 meituan;见上一行结构日志补 _resolve_source 规则")
+    return "meituan"
+
+
+def _handle(client: httpx.Client, msg: dict) -> None:
+    """明文消息:外部用户(from≠接收成员)发来的 weapp/image → 通知 app-server 打信号。"""
+    msgtype = msg.get("msgtype") or ""
+    frm = msg.get("from") or ""
+    receiver = settings.WX_FINANCE_RECEIVER_USERID
+    if receiver and frm == receiver:
+        return  # 成员自己发的, 跳过
+    if msgtype not in ("weapp", "image"):
+        return
+    # weapp 卡片按小程序判源平台;image(截图)判不出平台,沿用 meituan。
+    source = _resolve_source(msg) if msgtype == "weapp" else "meituan"
+    logger.info("命中触发 from=%s type=%s source=%s", frm, msgtype, source)
+    _notify(client, source, msgtype)
+
+
+def main() -> None:
+    setup_logging(debug=settings.APP_DEBUG)
+    if not settings.wx_finance_configured:
+        logger.warning("WX_FINANCE 未配齐(ENABLED/corpid/存档Secret), poller 退出")
+        return
+    if not settings.INTERNAL_API_SECRET:
+        logger.error("INTERNAL_API_SECRET 未配置, poller 无法通知 app-server, 退出")
+        return
+    signal.signal(signal.SIGTERM, _on_signal)
+    signal.signal(signal.SIGINT, _on_signal)
+
+    sdk = WxFinanceSdk(
+        settings.WX_FINANCE_SDK_PATH,
+        settings.WX_KF_CORP_ID,
+        settings.WX_FINANCE_SECRET,
+        Path(settings.WX_FINANCE_PRIVATE_KEY_PATH).read_bytes(),
+    )
+    seq = _load_seq()
+    logger.info(
+        "wx_finance poller 启动, 从 seq=%d 轮询, 间隔 %ss", seq, settings.WX_FINANCE_POLL_INTERVAL_SEC
+    )
+    with httpx.Client() as client:
+        while not _stop:
+            try:
+                msgs = sdk.get_chat_data(seq, limit=1000)
+                if msgs:
+                    logger.info("拉到 %d 条(seq>%d)", len(msgs), seq)
+                for item in msgs:
+                    try:
+                        plain = sdk.decrypt(item["encrypt_random_key"], item["encrypt_chat_msg"])
+                        _handle(client, plain)
+                    except Exception:
+                        logger.exception("解密/处理单条失败 seq=%s", item.get("seq"))
+                    seq = max(seq, int(item.get("seq", seq)))
+                if msgs:
+                    _save_seq(seq)
+            except WxFinanceError:
+                logger.exception("GetChatData 出错")
+            except Exception:
+                logger.exception("轮询异常")
+            # 可打断的 sleep:按 ≤0.5s 分片, 支持亚秒间隔(如 0.5s), 收到 SIGTERM 尽快退
+            interval = max(0.1, settings.WX_FINANCE_POLL_INTERVAL_SEC)
+            step = min(0.5, interval)
+            slept = 0.0
+            while slept < interval and not _stop:
+                time.sleep(step)
+                slept += step
+    sdk.close()
+    logger.info("wx_finance poller 已停止")
+
+
+if __name__ == "__main__":
+    main()

scripts/wx_finance_probe.py

@@ -0,0 +1,45 @@
+"""会话存档 SDK 连通性探针 —— 上 worker 前,在【云服务器】单独验证整条链路是否打通。
+
+  在 /opt/shaguabijia-app-server 下跑:  .venv/bin/python scripts/wx_finance_probe.py
+
+依次验证:加载 .so → NewSdk/Init(corpid+存档Secret+可信IP)→ GetChatData(0)→ 试解第一条。
+只读、不写任何比价信号。任一步报错都会明确指出卡在哪(便于逐项排:.so 缺依赖 / Init 失败 /
+可信IP 没放行 / 私钥不匹配)。全绿了再把 WX_FINANCE_ENABLED 置 true 上 worker。
+"""
+from __future__ import annotations
+
+from pathlib import Path
+
+from app.core.config import settings
+from app.integrations.wx_finance_sdk import WxFinanceSdk
+
+
+def main() -> None:
+    print(f"[cfg] sdk_path   = {settings.WX_FINANCE_SDK_PATH}")
+    print(f"[cfg] corpid     = {settings.WX_KF_CORP_ID}")
+    print(f"[cfg] secret_set = {bool(settings.WX_FINANCE_SECRET)}")
+    print(f"[cfg] privkey    = {settings.WX_FINANCE_PRIVATE_KEY_PATH}")
+
+    sdk = WxFinanceSdk(
+        settings.WX_FINANCE_SDK_PATH,
+        settings.WX_KF_CORP_ID,
+        settings.WX_FINANCE_SECRET,
+        Path(settings.WX_FINANCE_PRIVATE_KEY_PATH).read_bytes(),
+    )
+    print("[1] 加载 .so + Init 成功")
+
+    msgs = sdk.get_chat_data(0, limit=100)
+    print(f"[2] GetChatData(0) 成功, 拉到 {len(msgs)} 条")
+    if msgs:
+        first = msgs[0]
+        print(f"    第一条 seq={first.get('seq')} publickey_ver={first.get('publickey_ver')}")
+        plain = sdk.decrypt(first["encrypt_random_key"], first["encrypt_chat_msg"])
+        print(f"[3] 解密成功, msgtype={plain.get('msgtype')} from={plain.get('from')}")
+    else:
+        print("[3] 暂无历史消息(先让用户加接收成员好友、发条消息, 再跑一次)")
+    sdk.close()
+    print("探针完成:链路全通 ✓")
+
+
+if __name__ == "__main__":
+    main()

tests/test_llm_cost.py

@@ -0,0 +1,184 @@
+"""LLM 调用成本计算 compute_llm_cost:按模型分桶累加 token × 单价;error/无 usage 跳过。"""
+from __future__ import annotations
+
+from app.services.llm_cost import compute_llm_cost
+
+_PRICE = {
+    "per_model": {"qwen3.5-flash": {"input_per_1m": 0.8, "output_per_1m": 2.0}},
+    "default": {"input_per_1m": 3.0, "output_per_1m": 15.0},
+}
+
+
+def test_sums_per_model_single_model():
+    # 真实样本:4 次 qwen3.5-flash;Σprompt=6888、Σcompletion=337
+    calls = [
+        {"model": "qwen3.5-flash", "error": None, "usage": {"prompt_tokens": 1512, "completion_tokens": 22}},
+        {"model": "qwen3.5-flash", "error": None, "usage": {"prompt_tokens": 2111, "completion_tokens": 160}},
+        {"model": "qwen3.5-flash", "error": None, "usage": {"prompt_tokens": 1940, "completion_tokens": 142}},
+        {"model": "qwen3.5-flash", "error": None, "usage": {"prompt_tokens": 1325, "completion_tokens": 13}},
+    ]
+    cost, snapshot = compute_llm_cost(calls, _PRICE)
+    # 6888/1e6*0.8 + 337/1e6*2.0 = 0.0055104 + 0.000674 = 0.0061844 → round(6)
+    assert cost == 0.006184
+    assert snapshot == {
+        "mode": "per_model",
+        "prices": {
+            "qwen3.5-flash": {"input_per_1m": 0.8, "output_per_1m": 2.0, "_source": "per_model"},
+        },
+    }
+
+
+def test_multi_model_prices_each_bucket_separately():
+    calls = [
+        {"model": "qwen3.5-flash", "error": None, "usage": {"prompt_tokens": 1_000_000, "completion_tokens": 0}},
+        {"model": "gpt-x", "error": None, "usage": {"prompt_tokens": 0, "completion_tokens": 1_000_000}},
+    ]
+    price = {
+        "per_model": {
+            "qwen3.5-flash": {"input_per_1m": 0.8, "output_per_1m": 2.0},
+            "gpt-x": {"input_per_1m": 10.0, "output_per_1m": 30.0},
+        },
+        "default": {"input_per_1m": 3.0, "output_per_1m": 15.0},
+    }
+    cost, snap = compute_llm_cost(calls, price)
+    assert cost == 30.8  # qwen 1M入×0.8=0.8 + gpt-x 1M出×30=30.0
+    assert set(snap["prices"]) == {"qwen3.5-flash", "gpt-x"}
+
+
+def test_unknown_model_falls_back_to_default():
+    calls = [{"model": "mystery", "error": None, "usage": {"prompt_tokens": 1_000_000, "completion_tokens": 0}}]
+    price = {"per_model": {}, "default": {"input_per_1m": 3.0, "output_per_1m": 15.0}}
+    cost, snap = compute_llm_cost(calls, price)
+    assert cost == 3.0
+    assert snap["prices"]["mystery"]["_source"] == "default"
+
+
+def test_unpriced_model_marked_and_zero_cost():
+    calls = [{"model": "mystery", "error": None, "usage": {"prompt_tokens": 1_000_000, "completion_tokens": 999}}]
+    cost, snap = compute_llm_cost(calls, {"per_model": {}})  # 无 default
+    assert cost == 0.0
+    assert snap["prices"]["mystery"]["unpriced"] is True
+
+
+def test_error_and_missing_usage_calls_skipped():
+    calls = [
+        {"model": "qwen3.5-flash", "error": "boom", "usage": {"prompt_tokens": 9_999_999, "completion_tokens": 9_999_999}},
+        {"model": "qwen3.5-flash", "error": None, "usage": None},  # 无 usage
+        {"model": "qwen3.5-flash", "error": None, "usage": {"prompt_tokens": 1_000_000, "completion_tokens": 0}},
+    ]
+    cost, _ = compute_llm_cost(calls, _PRICE)
+    assert cost == 0.8  # 只有第 3 条计入
+
+
+def test_empty_or_all_error_returns_none():
+    assert compute_llm_cost([], _PRICE) == (None, None)
+    assert compute_llm_cost(None, _PRICE) == (None, None)
+    all_error = [{"model": "x", "error": "boom", "usage": {"prompt_tokens": 100, "completion_tokens": 100}}]
+    assert compute_llm_cost(all_error, _PRICE) == (None, None)
+
+
+def test_malformed_price_entry_is_treated_as_unpriced_not_raised():
+    # 手改配置页可能存出残缺/非法单价(缺 output_per_1m、非 dict);不能抛异常连累 token 回填。
+    calls = [
+        {"model": "bad-a", "error": None, "usage": {"prompt_tokens": 1_000_000, "completion_tokens": 5}},
+        {"model": "bad-b", "error": None, "usage": {"prompt_tokens": 1_000_000, "completion_tokens": 5}},
+        {"model": "ok", "error": None, "usage": {"prompt_tokens": 1_000_000, "completion_tokens": 0}},
+    ]
+    price = {
+        "per_model": {
+            "bad-a": {"input_per_1m": 0.8},   # 缺 output_per_1m
+            "bad-b": 5,                        # 非 dict
+            "ok": {"input_per_1m": 3.0, "output_per_1m": 15.0},
+        },
+    }
+    cost, snap = compute_llm_cost(calls, price)  # 不得抛异常
+    assert cost == 3.0  # 只有 ok(1M 入 × 3.0)计入;两个残缺项按 unpriced
+    assert snap["prices"]["bad-a"].get("unpriced") is True
+    assert snap["prices"]["bad-b"].get("unpriced") is True
+
+
+def test_get_llm_prices_falls_back_to_default_then_uses_override():
+    from app.db.session import SessionLocal
+    from app.models.app_config import AppConfig
+    from app.repositories import app_config
+    from app.services.llm_cost import get_llm_prices
+
+    db = SessionLocal()
+    try:
+        # 无 override → CONFIG_DEFS 默认(含 per_model / default)
+        prices = get_llm_prices(db)
+        assert "per_model" in prices and "default" in prices
+        # 有 override → 用 DB 值
+        app_config.set_value(
+            db, "llm_token_price",
+            {"per_model": {"m": {"input_per_1m": 1.0, "output_per_1m": 2.0}},
+             "default": {"input_per_1m": 0.0, "output_per_1m": 0.0}},
+            admin_id=1,
+        )
+        assert get_llm_prices(db)["per_model"]["m"]["input_per_1m"] == 1.0
+    finally:
+        row = db.get(AppConfig, "llm_token_price")
+        if row is not None:
+            db.delete(row)
+            db.commit()
+        db.close()
+
+
+def test_backfill_llm_calls_stores_cost_and_snapshot(monkeypatch):
+    from datetime import UTC, datetime
+
+    from app.api.v1 import compare_record
+    from app.db.session import SessionLocal
+    from app.models.app_config import AppConfig
+    from app.models.comparison import ComparisonRecord
+    from app.repositories import app_config
+
+    sample = [
+        {"model": "qwen3.5-flash", "error": None, "usage": {"prompt_tokens": 1512, "completion_tokens": 22}},
+        {"model": "qwen3.5-flash", "error": None, "usage": {"prompt_tokens": 2111, "completion_tokens": 160}},
+        {"model": "qwen3.5-flash", "error": None, "usage": {"prompt_tokens": 1940, "completion_tokens": 142}},
+        {"model": "qwen3.5-flash", "error": None, "usage": {"prompt_tokens": 1325, "completion_tokens": 13}},
+    ]
+    monkeypatch.setattr(compare_record, "fetch_llm_calls", lambda trace_id: sample)
+
+    db = SessionLocal()
+    try:
+        app_config.set_value(
+            db, "llm_token_price",
+            {"per_model": {"qwen3.5-flash": {"input_per_1m": 0.8, "output_per_1m": 2.0}},
+             "default": {"input_per_1m": 3.0, "output_per_1m": 15.0}},
+            admin_id=1,
+        )
+        rec = ComparisonRecord(
+            trace_id="llmcost-bf-1", status="success",
+            created_at=datetime.now(UTC).replace(tzinfo=None),
+        )
+        db.add(rec)
+        db.commit()
+        rid = rec.id
+    finally:
+        db.close()
+
+    compare_record._backfill_llm_calls(rid, "llmcost-bf-1")  # 独立 session 内回填
+
+    db = SessionLocal()
+    try:
+        rec = db.get(ComparisonRecord, rid)
+        assert rec.llm_cost_yuan == 0.006184
+        assert rec.llm_price_snapshot["prices"]["qwen3.5-flash"]["input_per_1m"] == 0.8
+        assert rec.input_tokens == 6888  # 现有 token 派生仍在
+    finally:
+        db.delete(db.get(ComparisonRecord, rid))
+        row = db.get(AppConfig, "llm_token_price")
+        if row is not None:
+            db.delete(row)
+        db.commit()
+        db.close()
+
+
+def test_admin_detail_schema_exposes_llm_cost_fields():
+    from app.admin.schemas.comparison import AdminComparisonDetail
+
+    fields = AdminComparisonDetail.model_fields
+    assert "llm_cost_yuan" in fields
+    assert "llm_price_snapshot" in fields