Plugins

Goal mode and autonomous execution plugin for DeepSeek Harness: a sticky top banner keeps the current goal visible across turns while the agent works toward it autonomously.

Browser notifications and a beep when a session turn ends, the agent asks a question or approval is pending; clicking the notification jumps to the session; remote trigger endpoint and SSE channel (npm: `bsfeng-dsh-notify`).

Watches agent execution traces and fires an LLM review at every successful turn end when signals trip, crystallizing reusable workflows into registered skills that improve progressively with later runs.

Roll back to before any of your questions or edit the last one and resubmit, with the workspace files of the discarded turns restored.

Right-hand widget rail for the DSH Web UI: live session stats (turns, LLM/tool time, TTFT, speed, cache, tokens) and OpenCode Go quota via a same-origin host proxy; extensible widget registry for more platform usage widgets, visuals, utility actions and external integrations.

暴露 run_python/run_node 工具，通过 stdin 执行代码返回 stdout/stderr/exit，含解释器路径配置卡 | Exposes run_python/run_node tools that execute code via stdin and return stdout/stderr/exit, with interpreter-path config card

DSH Relay 让你可以通过 Telegram 远程与 DeepSeek Harness 对话，并接收通知。DSH Relay turns Telegram into a remote conversation and notification channel for DeepSeek Harness.

Save English learning cards while interacting with DeepSeek Harness. It adds a line to the system prompt on every step telling the model to check your English and call english_cards_capture first; cards stay on this machine unless you turn on the sync switch (off by default), which after an OAuth login in your browser sends your original and corrected sentences with their cards to aoilang.com.

A ZCode-styled skin for DSH Web: Zai light/dark themes, sidebar workspace session tree, global search, turn navigation with hover wave, and approval-card / message-keycap styling aligned to ZCode — every visual value guarded by contract assertions. A fork of MichengAI/dsh-codex-ui; do not install it alongside the upstream plugin — both repaint the same sidebar and conversation surfaces.

Two gates for outbound AI text. (1) Guard: before any text goes out it blocks the shapes anyone could write — padding verbs plus abstract nouns ("optimised the code structure"), hollow adjectives, universal filler, template openers, git commit/PR/review boilerplate, comments that narrate the code. Covers chat, comments, email, commit messages, PR descriptions, reviews, issues, docs and code comments. Passes with OK, blocks with one line naming what to change. (2) Catch memes: when someone drops a slang term the model does not guess or look it up — it already knows. A built-in 6792-entry meme wiki is queried before the model starts thinking, and the entry usage field ("how to answer this") goes straight into context. Matches only exact title/slug/alias hits, never fuzzy, so ordinary chat never triggers it. Bundled SQLite by node:sqlite, zero dependencies, 1.6ms per query. Settings page: guard on/off, strict mode, prompt section toggle, meme wiki toggle, per-turn auto-lookup, a four-step tone slider (reserved / measured / follow the room / unrestrained) that is behavioural rather than a label, AI-completion switch (off by default, uses the session own model so no API key needed), a paste-and-test box, and three thresholds. No model call for any check.

An always-on working style for DSH agents, packaged as an agent preset: restate the task, read before editing, verify each change with a real command, keep context lean. Ships 29 skills (plan loops, review loops, TDD, unsticking, PR watching) plus a turn guard that pushes back when a turn edits files without a successful check afterwards, and pushes back harder on a second miss.

DSH plugin: smooths running-turn thinking and tool output across all Work details modes, with a capped following window in Compact/Standard and row-level pacing in Detailed/Verbose.

Kifu-style context compression for long DSH sessions: each step folds new turns into a DAG (node = position, edge = attempted fix), injects only a ~417-token hot list (2000-edge fixture, L5 tier), and adds failure memory plus position-repetition detection. Each span is summarized exactly once, so what is retained does not decay with the number of compactions.

Lets the model turn an answer into a web page that lives right inside the chat. The show_page tool either fills one of 8 built-in templates (report, compare, checklist, dashboard, timeline, flashcards, quiz, flow), where the model writes only the content and the plugin does layout, interaction and saving, or saves a full page the model wrote itself (calculators, games, custom layouts). Hand-written pages get the same design as base styles inside a CSS @layer, so the page's own CSS still wins. The page appears under the reply at its content height, interactive and following DSH's light/dark theme. The page's localStorage is stored in DSH per page, so what users tick or type survives reopening and new versions, and the same page open in the chat and the right sidebar stays in sync. Each edit is saved as a new version you can flip back to. After saving, the page is checked in a hidden local Chrome or Edge (script errors, horizontal overflow, dark mode not applied) and problems go back to the model. The right sidebar shows the page full-size with versions, a gallery of all pages and export to PNG, PDF or HTML. Pages run in a sandboxed, separate-origin frame with a strict CSP: no access to DSH data and no fetch, XHR or WebSocket. Works on phones with dsh-mobile.

Sidebar panel for Agnes image and video generation that also ships a node canvas and a short-drama mode. Canvas nodes are chained by extracting the previous shot's tail frame with ffmpeg and submitting it as the next shot's first frame; short-drama mode turns an imported script into storyboard shots. Generation and polling run in the Host process, so tasks survive a page reload or a closed panel, and results plus an 80-entry history are stored on disk.

Nothing stays in the system prompt: a hit returns titles only, a body loads only when read, and project rules can grow themselves.

Windows native toast notifications for DSH desktop sessions: turn ends, approvals, questions and a reminder for an unanswered wait — each with its own sound and popup text (up to three named templates per situation), three delivery modes, previews, and imported custom sounds. Clicking a toast raises the DSH window.

Game asset production pipeline: generate images, video and audio, turn a video into a coherent sprite animation, derive recoloured variants, adapt assets to an engine schema, install them into the project and verify the code actually references them. Requires ffmpeg for the video-to-sprite path, and an image/video API key (Agnes AI / OpenAI / fal) for generation.

Ctrl+F find-in-session for the conversation, powered by the session store rather than the DOM - labels every hit by source (you / assistant / reasoning / tool call / tool result / system), groups results by turn, merges identical snippets, offers prose-vs-everything scope, regex / case / whole-word options, highlights all matches, and copies or exports the result set as Markdown.

Makes the agent narrate its process: every turn opens with how it reads the message and what it will do first, then short interim notes between tool-call batches, with in-app settings for style, sentence cap and toggles.

A governed registry for MCP connectors (stdio and streamable-http): four-level fail-closed authorization, mcpServers JSON import that rejects malformed args, credential redaction at every output boundary, an agent bridge that returns an honest review verdict instead of a fake approval, and a /mcp-reg operator console plus a settings page.

Multi-agent roundtable discussion: each member runs as a subagent on its own model and speaks in turn, a host summarizes each round, and termination writes meeting-minutes Markdown.

Finishes 头歌 (EduCoder) homework over the platform JSON API, with no browser automation. educoder_sweep scans every course for homeworks that still have unfinished challenges — complete homeworks are skipped without reading a single challenge, so a full-account scan costs no model tokens — and can finish them; multiple-choice challenges also cost no model tokens, because the platform returns the answer key in any submission response, while code challenges are generated by the model, uploaded with evaluate:0 and graded once per challenge, then retried using the error report the platform returns. Three tools: educoder_sweep, educoder_homework, educoder_account.

Keeps project decisions in plain markdown instead of chat history: one main document plus per-module documents, every entry carrying a file:line citation. The agent interviews you to turn open questions into settled decisions, and a five-dimension score (task complexity, extensibility, maintainability, code quality, reusability) is derived from the documents and shown on a button beside the model picker. A session can bind several projects; projects are sized small/medium/large, which widens the entry limits, and a reusable-pipeline section in the main document is injected when its trigger words appear.