生産性ツール · DeepSeek Harness プラグイン — 173 ページ

Drag & drop anything into the DSH composer: local files upload to <workspace>/.dropped/<sessionId>/ and insert absolute-path references; web images / URLs download host-side then reference; plain text inserts into the composer (oversized text becomes a file).

DSH WebUI drag-and-drop file preview: drop files anywhere, full-page preview images and rendered Markdown (inline images resolved from dropped files or auto-searched in the session workspace), zoom/rotate images on click, persistent composer file box, smart desktop screenshot (full screen / free region selection / window picking with live thumbnails, auto-saved into the file box) with preview and external vision model recognition that inserts readable text into the composer for text-only models like DeepSeek.

Offline content-addressed proof for supplied DSH workflow duty separation

DSH 专用记忆插件：跨会话持久记忆（存储领域 + agent 记忆工具 + 系统提示词注入 + AI 建议记忆 + 设置面板卡片）

Edit a sent user message and resend it: fork a new branch from that message's turn and regenerate the reply with the edited text.

Retry any user message, with or without editing it first, and optionally delete the session it came from: forks at the event before the message, opens the child, sends the text, then removes the source log. · 重发任意一条用户消息：可直接重试，也可改完再重试，还可选择重试后删除原会话。

DeepSeek Harness 插件：在 Web GUI 中选择模型的推理程度（reasoning effort）。按当前模型的适配器目录展示可选档位（如 Off / High / Max），一键写入会话的模型选择。

Host-only DSH plugin: remember the reasoning effort that was last in effect per (provider, model) and restore it when a session switches back

DSH Effort Router（模型分流）：按每一轮请求的难度自动选择模型与思考强度——简单问题走便宜快模型，难题才叫强模型。规则判定零 token，灰区交给一次小模型裁定，图片档强度随难度升降（low/high/max）。每轮请求级覆盖，不改你的会话选择与默认模型。

DSH Web 思考强度就地替换插件：把模型座位「推理等级」面板里的原生单选行替换成 Claude Effort 卡片风格的美丽滑块（Off / Low / High / Max，含 WebGL 火焰），选择仍走官方会话链路。

DSH 小精灵插件：页面悬浮一只缓慢游动的半透明 DeepSeek 鲸鱼（官方 favicon 轮廓 + 蓝紫绿渐变），点击打开可拖动/最小化的临时聊天浮窗；默认跟随会话模型，支持插件内单独配置 OpenAI 兼容模型，流式输出、复制、清空。

A browser that lives inside the DSH host: every session gets its own tab, the AI drives it with lazily published browser_embedded_* tools, and a human watches or takes over that same tab in an on-demand right-Sidebar tab — no GUI, no external browser, logins/2FA/QR codes/CAPTCHAs included.

DeepSeek balance wallet button for DSH: one click shows your API balance. · DSH 余额小钱包：点一下就看见本猫的小钱包

One install for SpecFlow, GitFlow, Guardian, and Code Intel on DeepSeek Harness.

Per-session Engram memory for DeepSeek Harness: spawns an engram MCP child per agent session rooted at the session workspace, registers mem_* tools per agent scope, and injects the Memory Protocol as a system-prompt section.

DSH 版 enterworktree：创建（或复用）一个 git worktree，并自动开启一个指向该 worktree 路径的隔离 session，相当于 Claude Code 的 /enterworktree。

入口塑造（payload shaping）：按类别给会话元素排 op 链（reasoning / toolResult / toolArgs / assistantText），在元素被首次发送前定型；代理通道在 wire 层塑形 ⇒ 日志与界面保留原文、前缀单调、零断点税。

Zero-dependency DeepSeek Harness (dsh) plugin: probes the local machine environment (OS / shells / runtimes / tools / disks / proxy) once, caches the report, and injects it into every session's system prompt so the agent starts each conversation already knowing what this computer offers. Adds /env-refresh and /env-suggest slash commands where a command adapter is composed.

Three-mode visible-session advisor plugin for DeepSeek Harness with configurable tool permissions, local-subagent coverage, and task-tree budgets.

Press Escape to stop the current session's running generation, while preserving overlay shortcuts and queued work.

DSH 基础核心全家桶（All-in-one essentials pack）：聊天壁纸 + Token 用量 + 会话管理 + 每轮撤销（带产物回退）+ 文件浏览器/编辑器。零依赖，只注入 CSS/设置项/插槽/只读路由，不覆盖核心模块。

ETeams for DeepSeek Harness: captain-led multi-agent team collaboration with execution chains, task slots, member dialogs and persistent state

ETeams for DeepSeek Harness: captain-led multi-agent team collaboration with execution chains, task slots, member dialogs and persistent state

Condition-driven session wake for DeepSeek Harness: durable file and GitHub watches that fire a configured prompt into a target session (cold-wake included), with transition-only dedupe and throttling