dsh-next-input：助手回复后的下一句输入建议
dsh-next-input 是 DeepSeek Harness 插件，在助手本轮全部完成后于空输入框占位符中显示可能的下一句回复，Tab 填入草稿后编辑发送。沿用当前会话模型与凭据，支持桌面图形安装与 CLI Web 安装（0.2.2 / DSH 0.2.1-rc.2）。可配置启用、重试、超时与建议字数；长会话用滚动摘要，辅助调用计入模型费用。
Read Moredsh-custom-reasoning: Auto-Completion of Reasoning Intensity for Third-Party Models
`dsh-custom-reasoning` is a plugin specifically designed for DeepSeek Harness (DSH) to address the issue that third-party models—especially those routed through `llm-pi-ai`—cannot use the official reasoning feature due to missing reasoning effort metadata. By automatically completing a five-level `reasoningEfforts` declaration and the `compat` field for eligible models, the plugin enables reasoning capabilities with zero configuration. It supports multiple formats, including OpenAI, and persists these declarations, allowing developers to seamlessly and consistently use the official reasoning selector without manual intervention.
Read Moredsh-suggest-ghost: Input Prediction and Autocomplete Plugin for DeepSeek Harness
`dsh-suggest-ghost` is a client plugin developed for the DeepSeek Harness (DSH) web interface, designed to address the issue of repetitive input in long conversations. It provides two core capabilities: predicting the next sentence with an LLM (ghost suggestions) and performing history-based prefix completion (similar to zsh). The plugin supports Chinese input methods, cross-session hotness management, and data de-identification, and is compatible with the DSH-specified kernel version. Users can install it via the CLI and configure it in the settings panel, with options including adoption shortcuts, model routing, and word-by-word adoption, making it suitable for scenarios that require improved development efficiency.
Read Moredsh-suggest-reply: Sidebar Candidate Reply Generation Tool
`dsh-suggest-reply` is a sidebar plugin based on DeepSeek Harness, designed to help users generate candidate lines for AI replies. Users define tasks and JSON format through a custom System Prompt; the plugin automatically reads the latest message and optional director requirements to generate a style list. The generation process is isolated from the main session and supports real-time streaming debugging, parameter configuration, and one-click insertion of results. The plugin depends on `dsh-better-sidebar`, providing a clean, standalone generation experience.
Read Moredsh-model-meta-autofill: Automatically Completing Metadata for DSH Custom Vendor Models
dsh-model-meta-autofill is a DSH plugin designed to automatically fill in missing metadata for custom provider models in the `llm-pi-ai` settings area, including context length, maximum output, display name, and input modalities. By default, the plugin fetches data from models.dev, supports building with an offline index bundled at build time for offline usability, and allows merging other directory sources via the `sources` configuration. Its core principle is to "only fill missing fields with evidence," never overwriting user-existing settings, and it does not make guesses about newly released or niche models not included in the directory. The plugin provides a rich set of operation interfaces, including graphical options on the settings page (such as single-model filling, batch override, and data refresh) as well as a native loopback HTTP API, facilitating invocation by developers through scripts or consoles. Installation requires Node.
Read More