feat(compare): 比价记录 LLM token 成本落库与展示(按当时价冻结) (#133)

- comparison_record 加 llm_cost_yuan(元/float)+ llm_price_snapshot(JSON)两列
- _backfill_llm_calls 回填时按 app_config 当时单价逐模型算成本、冻结成本+快照到记录
- app_config 新增 llm_token_price 配置(per_model + default 兜底,运营在系统配置页可改)
- services/llm_cost.py:compute_llm_cost 纯函数(按 model 分桶、error/无 usage 跳过、
  脏价格当 unpriced 不抛异常以免连累 token 回填)+ get_llm_prices reader
- admin schema 暴露成本:列表项带 llm_cost_yuan,详情另带价格快照
- tests/test_llm_cost.py(10 测试);scripts/seed_mock_llm_cost.py(mock seeder)

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

---------

Co-authored-by: guke <guke@autohome.com.cn>
Reviewed-on: #133

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/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/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/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/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}

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()

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