dsh-peak-gate

LIMITATIONS

Known limitations

- **任务量是估算**：按字符类型粗估（ASCII ÷ 3.6，非 ASCII × 0.75）。够分级，不适合对账。 **English:** **Task size is an estimate**: a rough estimate by character class (ASCII ÷ 3.6, non-ASCII × 0.75). Good enough for classifying, not for accounting. - **模型分类要多花一次调用**：每次判定 heavy/light 约几十 token（`maxTokens: 6`）。相比拦下一次 写程序的工具循环（动辄上万 token）可以忽略，但它确实是「每条够长的输入都问一次」的成本模型； 想省就调高 `classifierMinChars` / 调低 `useModelClassifier`。 **English:** **Model classification costs one extra call**: each heavy/light decision is about a few dozen tokens (`maxTokens: 6`). Next to blocking one program-writing tool loop (easily tens of thousands of tokens) that is negligible, but it really is a cost model of "ask once for every input that is long enough"; to save, raise `classifierMinChars` / turn down `useModelClassifier`. - **停下的那一轮没有模型收尾**：第 3 层是「不发请求、直接落提示」，所以用户看到的是进度停住 + 一条说明，而不是模型自己写的总结。要继续就带 `!force` 重发，或把 `peakMaxSteps` 调大。 **English:** **The turn that gets stopped has no model sign-off**: layer 3 is "send no request, drop a notice instead", so what the user sees is progress halting plus an explanation, not a summary written by the model. To carry on, resend with `!force`, or raise `peakMaxSteps`. - **`reject` 方式界面静默**：被拦的消息不进会话记录，会看到"发出去没反应"。默认的 `replace` 没这问题。 **English:** **The `reject` mode is silent in the UI**: the blocked message does not enter the session record, so it looks like "sent but nothing happened". The default `replace` does not have this problem. - **不做记账**：目前只在日志里记拦截次数，没有「今天省了多少钱」的面板。要做需要订阅 `llm/stream` 采真实 usage + 落盘 + client 半边（市场里 `dsh-peak-cost-mode` 是完整形态，可参考）。 **English:** **No accounting**: for now it only records the number of blocks in the log; there is no "how much did I save today" panel. Building one would mean subscribing to `llm/stream` to sample real usage + persisting it + a client half (marketplace plugin `dsh-peak-cost-mode` is a complete form of this, worth referencing). - **精简段会让 KV cache 失效一次**：峰谷切换时 system prompt 变化，会丢一次前缀缓存。一天只切 4 次，可接受。 **English:** **The brevity section invalidates the KV cache once**: switching between peak and off-peak changes the system prompt, so one prefix cache is lost. It only switches 4 times a day, which is acceptable. - 只认 **DeepSeek 官方 API 的计价口径**；走第三方网关时用 `officialProviders` 排除，或自己改 `peakWindows`。 **English:** It only recognises the **pricing rules of the official DeepSeek API**; when going through a third-party gateway, exclude it with `officialProviders`, or change `peakWindows` yourself.