DeepSeek Harness 플러그인 — 484 페이지

让主会话模型从已配置模型里自主选择子代理模型

超级模组注入器：运行时注入任意本地 DSH 插件包（junction 链接 + loader.create，不碰 patch/package.json/不重启），热重载全家桶 + 开发侧挂区一键转正 + 一键卸载 + 路由自愈 + 插件管理 UI（设置页：列表/卸载/拖入内化），清单持久化重启自动恢复——DSH 生态的 BepInEx 式模组注入入口

Taffy Live Atelier — candy acrylic DSH Web UI theme with light/dark stage art, character overlays, and a Taffy agent preset.

Task delegation router for DSH: main model plans/verifies, sub-models search/browse/code/verify.

thincoder 四功能移植：advisor 评审收敛 / 工程模式 eng / 飞刀 escalate / 会诊 consult（host-only 纯 JS 免构建）

Detect thinking-chain loops at agent/turn-stopping and steer to break them (DSH plugin, no proxy)

Fixes malformed tool-call arguments (nested {"arguments":{…}} wrappers, wrapper-key aliases, param-key aliases) at the llm/stream seam — repairs execution, persistence and history replay in one place. 修复模型产出的畸形工具参数。

Windows 系统托盘插件：托盘图标 + 菜单（打开界面/状态/退出）+ 气泡通知，基于 trayicon（exe 宿主，无 native 编译）

Codex-style turn folding: collapse each group of reasoning + tool-call nodes (between visible outputs) into a single disclosure unit

Custom text and colour for the DSH chat running-status line — the 「深度求索中，用时 12秒 ···」 row — configured from the Plugins page.

DSH-UI-Manager: manage installed UI/skin plugins from the web settings — enable one at a time, restore the default UI, rename display names, open plugin homepages, and uninstall skins.

会话行「标记未读 / 取消标记」：小红点主动未读标记（用户钉住待处理），与系统绿点区分；绿点存在时两项禁用，红点仅手动取消才消失，服务端持久化

在设置页检查 npm 上 @deepseek-ai/dsh 官方最新版本，并与本地安装版本对比（host: webServer API + node fetch；client: settings.section 设置页）

DeepSeek 用量面板：余额（API Key）+ Token 用量（平台会话令牌）仪表盘，注册在 设置 → DeepSeek 用量

DSH 用量页：本机跨会话 Token 使用统计（Codex 风格），纯只读、不出本机 | Local cross-session token usage statistics page for the DSH Web GUI (Codex-style), read-only and machine-local

DeepSeek Harness 使用统计：Token 汇总、热力图、每日趋势、模型占比（Settings 页面）

用量统计：聚合所有会话的 token 使用总量与模型分布，设置页显示

DSH 用量与工具调用统计（token 分桶、费用估算、工具调用次数排行榜、按天趋势）

DeepSeek Harness 智能体刘海灵动岛：Mac/Win 双端兼容，实时显示 Agent 思考、工具执行与子代理状态的全局灵动岛状态栏

DeepSeek Harness vision bridge: session screenshots route by model capability. Image-capable models receive attachments inline; text-only models see the standard placeholder, and the bridge reads the image through a vision-language model — on demand via the vision_bridge_read tool or automatically via agent/pre-step context injection. 会话截图按模型能力自动分流：多模态模型原样直读；纯文本模型经 VLM 代读（tool/auto 双模式）。

DSH profile bundle: auto-bridge images for text-only models (DeepSeek etc.) — instead of the native "model does not support images" rejection, the image is materialized into the workspace and the model is instructed to run the configured vision tool (vision_glance, fuzzy-selectable) and answer from its analysis. Works on WSL/Ubuntu and native Windows; sessions on excluded presets (default minimal) keep the original blocking behavior.

Lightweight DSH upload bridge: images and common documents become workspace paths so a text-only DeepSeek model can read them with Qwen-MM-Plugins vision tools.

辅助视觉模型：图片→文本描述，供文本模型、浏览器截图兜底与聊天贴图降级使用

DeepSeek Harness-native integration for agent-vision-toolkit: image Q&A, OCR, grounding, UI restoration, pixel diff, Artifacts, and Web UI.