Плагины

Записывает локальные сводки JSONL по количеству вызовов инструментов на каждый ход и грубым сигналам верификации, не сохраняя промпты, аргументы инструментов или текст результатов.

Генерация функций, инструментов MCP, навыков рабочего процесса и офлайн-пакетов тестов из исходного кода, одобренного пользователем.

Мост канала Feishu (Lark) с отказом в закрытое состояние: пишите боту — получайте в ответ ходы агента dsh. Интеграция только через официальный Python SDK (жёстко зафиксированная версия); белый список deny-by-default, проверка подписи/времени/повторов вебхука, липкие сессии по каждому чату; двуязычная документация.

Единый слой памяти для любого AI-инструмента и агента: инъекция Context Bundle, воспроизведение в момент промпта, инструменты MCP и захват треда DSH по завершении хода.

Импортируйте историю чатов claude-code / codex / reasonix / zcode в сессии dsh: привязка к рабочему пространству, сохранение вызовов инструментов, защита от слишком больших сессий, восстановление сжатия zcode.

Нативное окно рабочего стола Windows (WebView2) для DSH: установка одной командой, автоматическая загрузка zip-архива приложения из GitHub Releases, создание ярлыка на рабочем столе и добавление инструмента desktop_launch для запуска из чата.

DeepSeek Harness × OpenCode TUI 前端：把 dsh 的 Agent、会话、工具接入官方 opencode 终端。OpenCode TUI frontend for the DeepSeek Harness (dsh): agents, sessions and tools in the official opencode terminal.

compact reasoning and tool activity groups for DeepSeek Harness Web and DeepSeek Harness Desktop.

Vets third-party plugins before you trust them: static scan for malicious patterns (exfiltration, credential access, obfuscation, persistence) and over-privileged path use, transitive-dependency coverage, official-package hash baseline for supply-chain tamper detection, and an optional plugin-tool call gate.

Fine-grained permission gateway: per-category tool-call review (outside-workspace directories, commands, file read/write, subagents, repeated actions) with global & per-project allow/deny exceptions, quick-tool defaults, custom rules, a bilingual approval modal with inline diff details, custom rejection reasons and a sandbox-upgrade flow.

Agent security guardrail: intercepts and audits every tool call, requiring human confirmation on sensitive operations.

Health checks for DSH plugins: manifest/patch/entry/build/pack/install verification, model-callable plugin_check, profile host-shadowing + BOM detection, environment diagnostics, and poison preflight.

Continual self-refinement loop: persistent memory, periodic review-and-refine, cross-session shared knowledge, and automatic rollback through a model-callable harness_refine tool.

Feishu (Lark) bridge for DeepSeek Harness: one-command install, streaming progress cards with a tool panel, ask/approval button cards, attachments, and proactive feishu_send pushes.

Bridge DSH agent sessions, tool calls, and approvals to the CodeIsland macOS notch panel over a Unix socket, with in-panel allow/deny.

Workspace-safe Office tools for agents: create/read Word, create/read/update Excel, and create/read PowerPoint decks with PNG/JPG/GIF image placement.

Vision MCP and DSH bundle for text-only DeepSeek: analyze_image, analyze_clipboard, compare_images and vision_status tools, a visual settings page, free GLM-4.6V-Flash by default, result caching and rate-limit tolerance; keys stay out of logs.

MiniMax-powered multimodal plugin: real-time voice call mode (streaming conversation, floating dock UI), voice mode and mic voice input, plus image/video/music/speech generation and vision inspection tools.

Zero-dependency browser control: drive a real headless Edge/Chrome over CDP with 8 `pilot_*` tools (navigate/click/type/keys/eval/screenshot) plus a live draggable cockpit panel in the Web GUI - text-first snapshots for text-only models, no Playwright, no API key.

Seven video tools (probe/cut/concat/encode/subtitle/extract/gif) on the official subprocess service: shell-free argv, zero runtime dependencies.

Optimize a raw instruction into a professional prompt through the harness llm service — four-section (Role / Task / Context / Format) or a headerless plain-text style. Context-aware by default: the recent conversation is injected as background and may enrich the ## Context section. Long outputs resume from the truncation point with auto-expanded budgets; validated results are cached (identical requests cost zero model calls); a unified call budget bounds worst-case cost. Composer ✨ one-click optimize, undo, cancel and a token-cost hint; prompt_optimize tool (optimize + iterate), auto-optimize hook, auto-detected zh/en role document.

Zotero library tools via the local API (no API key): search items/collections, read metadata and abstracts, list attachments, read full text (parses PDFs on the fly when Zotero cache misses), download PDFs, and manage notes.

Dedicated model routes, tools, and system guidance for vision, compaction, reviews, subagents, titles, and image generation.

Connect dsh to a local Obsidian vault: search, read, write, move, and trash notes through `obsidian_*` tools.