Plugins

Rewrites the composer draft, or only the selected text, with one model call that uses the recent conversation as context, with cancel, revert, fidelity and length checks, and an optional preview.

Plays a character entrance card beside the composer when you switch model: identity name, six scored axes (intelligence, coding, speed, stability, cost, context), quota hint and a dialogue line. 23 curated cards across 15 providers; models without a curated card fall back to a data card derived from public benchmarks, or a generic card when no benchmark data exists. A web client plugin with no dependencies and no hooks.

A Claude Desktop-inspired composer slider for adjusting model reasoning effort with click, drag, or keyboard controls.

Button beside the composer that rewrites the current draft into a more executable prompt in place. Ships four built-in presets spanning light polish to aggressive expansion, accepts user-defined presets, and switches preset, prompt model and reasoning effort directly from the button without leaving the composer. The button appears only when the draft has content, shows a running timer while enhancing, and surfaces failures with a retry.

Windows desktop floating assistant for DeepSeek Harness: an always-on-top chat panel with streaming answers, model switching, and region-screenshot questions.

Recovers a DeepSeek Harness turn that ended with reasoning only, with no text and no tool call, by steering the same turn one more step. Structure-only detection, so it costs no extra model call.

Comment on any passage of a DSH Web conversation and get a brief reply from the session's own model in a small thread next to it, without touching the main chat.

内置 provider / 模型显示名的全量数据库：载入时按环境里实际启用的 provider 与模型覆盖显示名，只影响 UI，不改路由、请求与凭据。

把 API3D（api3d.cn）AI API 网关接成 DSH 的一条模型路由：基础地址、API Key 与出站兼容开关都在设置页。API3D (api3d.cn) as a single DeepSeek Harness model route.

One-click polish, expand and autopilot for the DeepSeek Harness composer: rewrite the draft with the model the session is switched to, expand a one-line wish into a task brief, or let an AI drive a goal round after round.

Read-only Windows AppData analyzer: click the sidebar icon to drill down explorer-style through directories, each row showing recursive size, file count and last-modified time, plus the software that created it and a delete-worthiness verdict built from a knowledge base, mixed-directory splitting, then heuristics, with unmatched guesses marked as such. Scans directories only and never reads file contents; it ships no delete action, and unknown directories can be looked up online with a model summary and cited sources. The online lookup needs a DeepSeek API key (or an equivalent credential); without one the query is only copied to the clipboard.

DSH plugin: detect assistant turns that end with no visible text (zero-width placeholder) and re-prompt the model to write the answer; also injects a visible-closing-answer system prompt rule.

Dual-host coordination for DeepSeek Harness (大肥鱼 bot): hand a task off in local chat, run bounded jobs in a cloud worker, then review results and approve follow-up steps in one workbench. Includes protected six-page onboarding, per-task model selection, a task approval ledger and stepwise controlled delivery.

An isolated translation pipeline where each module is its own model call carrying only its own prompt and its own input — no session history, no tool schemas, no other module's content. Build chains from draggable cards, edit any module's prompt in place, override the model per module, or start from one of two precision presets. The same chain is exposed to the agent as a translate_flow tool.

Manage agent skills and MCP servers across DSH, Claude Code, Codex, Cursor and other tools from one DSH sidebar page: CRUD, cross-tool sharing, import/export, backups and MCP health checks.

Writing council: parallel multi-model drafting, judge consensus & critique synthesis, AI prompt refinement, visual concept illustration, and Markdown file preview.

Runs OpenCode as a DSH subagent over ACP and shows each task live in the right sidebar — its agent and model, tool calls, and the subagents OpenCode spawns itself; permission requests become DSH question cards with allow once, allow always or reject.

Windows x64 bundle with the official Antigravity ACP runtime included. Delegates one-shot text tasks with selectable authentication, model and thinking level; uses host peers for Harness 0.2.1 from rc.3 and 0.2.2 from rc.1, including subsequent RCs and stable releases within both series.

Dormant-by-default MCP for dsh: one mcp_summon tool; summoning a server registers its tools as real native mcp__<server>__<tool> tools for the rest of the session, and release unregisters them and stops the process.

Agent tools that download a Bilibili video, transcribe it with a local whisper.cpp build, then write a structured Markdown note with DSH's own LLM (optionally printing it to PDF). Windows only: missing dependencies are auto-downloaded into the data directory — yt-dlp.exe and the ffmpeg.exe build from GitHub and gyan.dev, the whisper.cpp x64 binary from GitHub, and a ~547 MB model from the hf-mirror mirror; PDF export shells out to a locally installed Edge/Chrome. Two tools: video_to_notes and note_to_pdf. Pure JS, no extra API key.

Long-term memory backed by a local SQLite store: project-scoped memory auto-injection with junk filtering, rule-based or model-assisted memory extraction from recent turns, write-side dedup and feedback-weighted ranking. The Python bridge runs on $NEXUS_PYTHON or the DSH-bundled interpreter at <DSH home>/dsh-runtimes/dsh-primary-runtime/dependencies/python/ (Python 3.8+).

When a live session's context passes a threshold, append older messages to a per-conversation folder on disk, then release the oldest finished turns from the model-visible context; the chat transcript and the session log stay untouched.

Names every new DSH session DATE | CATEGORY | NAME from the session model, with a deterministic local fallback.

Right-click a sent message to edit or withdraw it. The running answer is stopped first, then that message and everything after it are replaced in the conversation: withdrawing leaves a short withdrawn marker and stops there, while editing asks the model again with the new text.