xsl
f46ecc5eb0
Two defects that only surface against the real DeepSeek API, both of which break the cost model rather than the code: deepseek-v4-flash enables thinking mode by default. The same single-turn question took 13.5s with it on and 1.8s with it off, and the thinking tokens are billed at the output rate. That misses NFR-PERF-02b (15s per turn) and inflates every conversational call. SceneConfig now carries an explicit `thinking` flag, all four V1 scenes set it to false, and the provider always sends the parameter rather than relying on the upstream default. Reasoning tokens are recorded separately in Usage.tokens_reasoning so the cost of ever turning it back on is visible. DeepSeek moves to peak/off-peak pricing on 2026-08-17, with peak (Beijing 09:00-12:00 and 14:00-18:00) at twice the off-peak rate. Reports are generated mostly during peak hours, so a single blended rate would leave ob_ai_call -- the only input to V2 pricing -- off by nearly a factor of two. PeakAwareRateTable resolves the rate at call time; cost is computed and stored then, never recomputed later at whatever the current rate happens to be. Also here: - Replace the live API key in the redaction test with a synthetic one. The test that proves we redact secrets was itself a secret leak waiting for its first commit. - Correct LLM_COST_RATES_JSON, which held placeholder rates. - Make ids.SnowflakeGenerator take an injectable clock so rollover and clock-rollback are testable without patching a slots class. - Rename exceptions to the -Error suffix (N818) and clear the remaining ruff, mypy and import-linter findings, including two dead lines: a duplicate pytestmark assignment and a no-op assert on a None return. Gates: ruff, mypy (96 files incl. tests), 4/4 import contracts, detect-secrets, 166 unit + 3 integration tests, migration up/down round-trip.
packages/llm/ — LLM 网关
架构 §4.3、§8.1。V1 是进程内包，V2 拆成独立服务。 因此接口必须按远程调用来设计：无共享内存假设、明确超时、可序列化的入参出参。
唯一入口
所有 LLM 调用都走 LLMGateway.complete()。业务代码不许自己 httpx.post 到任何
模型供应商 —— 那样会绕过成本归因、预算闸门、故障转移和 Prompt 版本记录，而这四件事
一旦有缺口就补不回来（历史数据是补不出来的）。
已实现
| 能力 | 位置 |
|---|---|
场景化路由（scene → provider/model/温度/超时/预算，全配置化） |
routing.py |
| 故障转移 + 熔断（超时/5xx/限流切备用；连续失败熔断，冷却后半开探测） | gateway.py circuit.py |
| 结构化输出（Pydantic Schema 约束 + 校验失败重试） | gateway.py |
| Prompt 版本（文件化，按 id+version 加载，版本随每次调用落库） | prompts.py /prompts |
| 缓存前缀纪律（固定前缀含时间戳/随机 ID/昵称 → 拒绝或告警） | prompts.py |
成本归因（每次 provider 调用写一行 ob_ai_call，含缓存命中 token） |
cost.py recording.py |
| 全局日成本硬熔断（超预算即拒绝非关键场景，不是只告警） | budget.py |
未实现（按 §8.1 还差）
- 流式输出 SSE（FR-AGT-01 要求首字 p95 < 2s，非流式做不到）
- 审核联动：输入输出审核应在网关内统一完成，业务方无需关心（CR-06）
- 用户级频控：目前只有平台级日预算与业务侧配额
- A/B 分流：按用户哈希选择 prompt 版本
- 智谱 GLM provider：仅结构占位，见
providers/zhipu.py
加一个新场景
- 在
types.Scene加枚举值。 - 在
routing._DEFAULTS加默认配置，想清楚critical该不该为真 ——critical=True意味着日预算耗尽时它仍然会花钱。 - 在
prompts/<scene>/v1.txt写模板。[system]段不许有任何变量。 - 跑
pytest tests/unit/test_prompt_cache_prefix.py，它会扫描仓库里全部 prompt 的固定前缀。
加一个新 provider
- 实现
providers/base.LLMProvider协议（name/default_model/complete/aclose）。 - HTTP 错误必须映射到
errors.py的异常族，retryable决定要不要切备用。 映射错了会导致「明明该切备用却没切」或者「明明是我们请求写错了却把备用也打一遍」。 - usage 字段必须映射出 缓存命中 token 数，否则 NFR-COST-03 的缓存命中率 看板是空的。
- 在
apps/api/deps.build_llm_gateway()注册。
费率
RateTable 来自配置（LLM_COST_RATES_JSON），不是常量。未知模型走 default 费率
而不是 0 —— 成本上的盲区比不精确更危险。上线前必须用供应商实际价目表校准。