@dsh-external/dsh-auto-continue

Readme

# dsh-auto-continue

**Automatically resumes replies that get cut off at the output token limit — built for self-hosted Ollama users.** · **回答被输出上限截断时自动续写——为自部署 Ollama 用户打造。**

[![dsh-plugin](https://img.shields.io/badge/dsh-plugin-4B8BBE?logo=deepseek&logoColor=white)](https://github.com/topics/dsh-plugin)

When a model hits the per-request output token cap (`max-tokens`), DeepSeek Harness ends the turn and shows *"Send 'continue' to let the model resume"*. This plugin watches for exactly that event and sends the continuation **automatically** — no manual typing, no babysitting long answers.

当模型达到单次请求的输出 token 上限（`max-tokens`）时，DeepSeek Harness 会结束这一轮并提示"发送'继续'可让模型接着输出"。本插件监听该事件并**自动**发送续写指令——不用手动打字，不用守着长回答。

> 🦙 **Who is this for?** Self-hosted Ollama deployments. Local models on consumer hardware usually run with a deliberately small output cap (4K–8K) to keep the GPU/CPU stable — so long answers get truncated constantly. This plugin is the band-aid that makes truncation painless without raising the cap. It works with any provider, but it is *designed for* Ollama.
>
> 🦙 **这个插件是给谁的？** 自部署 Ollama 的用户。本地模型跑在消费级硬件上，通常刻意把输出上限设得很小（4K–8K）以保证硬件稳定——于是长回答频繁被截断。本插件就是让截断不再痛苦的"创可贴"，**不需要调高上限**。任何 provider 都能用，但它是*为 Ollama 设计的*。

## Features / 功能

- 🔁 **Automatic continuation** — on a `max-tokens` turn end, queues a "继续" prompt via `agent.followup()`; the model resumes in a new turn. / 检测到 `max-tokens` 截断自动队列"继续"，模型在新一轮接着输出。
- 🛡️ **Loop protection** — at most `maxAutoContinues` consecutive auto-continues per session (default 5); any normally-finished turn resets the counter. / 同一会话连续截断最多自动继续 5 次（默认），防死循环；正常结束的轮次自动重置计数。
- 🙅 **Never fights your input** — skips when you already have queued messages. / 你已有排队消息时自动跳过，不与手动输入打架。
- 📝 **Audit log** — every auto-continue is recorded. / 每次自动继续都有日志可查。

## Install / 安装

```sh
dsh plugin --profile web add github:zhou1736948757-cpu/dsh-auto-continue
```

Or browse it in the [dsh-market](https://github.com/dsh-market/dsh-market#readme) plugin market inside Settings. / 或在设置里的 dsh-market 插件市场中浏览安装。

## How it works / 工作原理

- Listens to each agent session's `session/event` feed for `turn/end` with `reason.kind === "max-tokens"`.
- Queues a user message (default text `继续`, tagged `source.kind = "auto-continue"`) through `agent.followup()` after a short delay so the turn fully settles.
- The truncated output stays in the conversation, so the model continues from where it stopped — exactly like typing "继续" yourself, but automatic.

## Configuration / 配置

Add to `settings.yaml` (or accept the defaults): / 在 `settings.yaml` 中添加（或直接用默认值）：

```yaml
dsh-auto-continue:
  maxAutoContinues: 5      # max consecutive auto-continues per session (1-20) / 连续截断最多自动继续次数
  continueText: 继续        # the continuation prompt text / 续写指令文本
  delayMs: 300             # delay before sending, lets turn/end settle (ms) / 发送前延迟
  logFile: ''              # log path, default $DSH_HOME/super-injector/dsh-auto-continue.log / 日志路径
```

## Log / 日志

Every auto-continue is appended to `$DSH_HOME/super-injector/dsh-auto-continue.log` (default). If a reply stops without resuming, the log says whether it hit the cap or failed. / 每次自动继续都会写入 `$DSH_HOME/super-injector/dsh-auto-continue.log`（默认路径）。如果某次回答没续上，看日志就知道是"达到上限停止"还是"发送失败"。

## Notes / 说明

- This is a **safety net**, not a fix for the root cause. If your hardware allows a larger cap, raising the model's `maxTokens` (Settings → Models) and/or lowering `reasoningEffort` reduces truncation itself. / 这是**兜底方案**，不是治本。如果硬件允许，调大模型 `maxTokens`（设置 → 模型）或降低 `reasoningEffort` 能减少截断本身。
- With `reasoningEffort: max`, a truncation may happen with zero visible text (reasoning consumed the budget). The plugin still continues — the model keeps reasoning until it produces output. / 推理强度为 max 时，截断可能发生在正文出现之前（额度被思考耗尽）。插件照常续写，模型会继续思考直到出正文。
- Each continuation is an ordinary new request — no extra load on your hardware beyond what a manual "继续" would cause. / 每次续写就是一次普通请求，对硬件的压力与你手动发"继续"完全一样。

## License / 许可证

BSD-3-Clause