Плагины

Trim idle DeepSeek Harness session logs at startup: keep the last N windows, cut the ones the model can no longer see, and verify the result with the host's own format, session and projection validators before writing. Not context compaction: this shrinks

A panel plugin that keeps mini-games in one place and plays them in the same window. Ships with Tetris: a 10x20 board, seven pieces, a ghost piece and a next-piece preview, line and level scoring, three difficulty levels, and a best score kept in localStorage.

Replaces the chat's "Deep diving" waiting label with rotating fun phrases, editable live in Settings.

Back up and restore DSH plugins through your own private GitHub repository — list, versions, configuration and local sources included.

Exposes a DSH Harness Desktop Web GUI to another machine's browser through a Cloudflare quick or named tunnel, an SSH reverse tunnel, or a custom command, and hands out a tokenised link to open it; a settings panel and CLI start, stop and print the link. Needs DSH Desktop 0.2.1-rc.2 or later and a cloudflared binary.

Blocks the end of a turn while temporary files from that turn are still lying around: it registers a listener on the serial agent/turn-stopping hook and, if anything is left, steers the model with the file list and two ways out — delete them, or justify each one and move it somewhere deliberate. Candidates come from two narrow temp-file sources, plus a backup-retention channel that is on by default and flags excess .bak copies of files touched this turn. The first is the .agent-tmp/ convention directory under the session cwd: whatever is left in it is reported outright, because that directory exists only for throwaway files. The second is the paths this turn actually wrote, captured through the fs/write-intent hook; those are reported only when they sit under that cwd, fall outside .git, node_modules, .tmp-sessions and out, match a temp-file pattern, and the file is still present — which is what keeps junk already sitting in the workspace from being resurrected as a complaint about this turn. One steer per turn, subagent sessions skipped, nothing deleted by the plugin in the default steer mode (only the opt-in mode: prune deletes backups), and the check is readdir plus stat only.

Git tab that covers the session workspace root together with every additional root registered through multi-root-workspace, so sibling checkouts appear as separate sections with their own branch, ahead/behind, line counts, directory tree and commit history; clicking a file opens its unified diff (or whole content, for untracked files) and clicking a commit opens its patch. Each section also has a commit box that stages and commits only the listed files (`git add -- <files>` + `git commit -- <files>`).

Long-form manuscript verification team for DSH: installs 10 skills that run a read-only content/structure check, three-tier revision proposals, a human-selection gate, exact replacements and a final re-check; it never edits the manuscript by itself and requires Agent Teams to spawn teammates.

Turns your local WeChat group chats into a followable todo board: a three-column kanban (todo / doing / needs you) with LLM summaries, automatic flagging of messages that @-mention or reply to you, drag-to-change status, and local-only storage. Needs a local Python 3 with sqlcipher3 and zstandard, plus the existing rion-wechat-reader config.json and keys.json for the WeChat databases; the plugin does not extract the keys itself, and summaries send the quoted chat text to the model you choose.

Manages reusable agent role templates (job prompt, model and reasoning effort, tool allowlist, skills) and starts role members through the native Agent Team interface to delegate and continue tasks.

Scheduled model-news radar: collects official release announcements from LLM vendors through RSS, Hugging Face and GitHub organisation APIs and changelog pages (a 6-vendor sample list is bundled, a user-maintained list is read from ~/.dsh/model-news-radar/sources.json), deduplicates them into daily, weekly or monthly Markdown reports and opens a DSH session that writes the summary. Read-only, no logins, no third-party dependencies.

An on-demand Chinese editing skill that reduces formulaic phrasing while preserving facts, uncertainty, genre, and author voice.

Skill management panel inside DeepSeek Harness settings: lists global and per-project skills, marks whether the running DSH actually loads each one, offers one switch that controls both model and slash-command invocation, previews and edits SKILL.md with the chat UI markdown renderer, and creates, imports (folder or zip) or deletes skills into a recovery folder. It only reads and writes the local skill roots (~/.dsh/skills and <project>/.dsh/skills) through its own Host routes behind the DSH Connection trust fence; no telemetry, no outbound requests, no credentials.

Save conversations, manage notes with images, export files, and sync across devices through your own private GitHub repository.

Synthesises speech through Alibaba Bailian's WebSocket TTS and returns a WAV file plus a preview URL, with an optional per-turn hint that nudges the model towards voice. Needs a Bailian API key and ffmpeg, and the text is sent to Alibaba Cloud; the plugin does not send messages itself — delivery is left to the session's own transport.

Adds a composer microphone button that records audio as MP3 (default), WAV, or WebM and attaches it to the current prompt as a conversation file attachment. The format is chosen on the plugin's own detail page under Settings > Plugins; recordings are passed to the host through the conversation draft-attachment channel and are not written to disk by the plugin.

Image generation through the Volcano Ark / Doubao Seedream API: ark_generate_image for text-to-image, image-to-image and fixed-seed batches, ark_image_gen_status for checking the key, model and output directory. Requires a Volcano Ark API key (ARK_API_KEY).

Image understanding for DSH: describe an image and read the text in it through Zhipu GLM-4V (needs a Zhipu API key, free tier), with optional person identification through Volcengine Ark and anime-only reverse image search through SauceNAO or trace.moe (each needs its own key); ffmpeg must be installed for GIF/WebP input. Keys and the ffmpeg path are set on the plugin's own settings page.

Annotate elements on real Chrome and Edge pages, then send selectors, comments and screenshots to the agent for visual front-end editing.

Drive a real Chrome and bundle the DOM, accessibility tree, HAR with response bodies, console logs, storage, screenshots, PDF and screencast into one self-contained evidence folder.

Zero-dependency browser hand & eye over CDP: drives your own Chrome/Edge with real clicks, Chinese IME typing, scrolling and text/selector element picking, and captures whole-page evidence into one replayable folder — DOM, accessibility tree, screenshots, cookies, storage, console, HAR, resource waterfall and perf metrics.

Electron browser with two faces. For agents it is a runtime: a daemon hands out port leases so several agents drive isolated, profile-backed sessions (24 session tools — navigate, act, observe, screenshot, batch runs — plus network_capture/responses for reading XHR/fetch bodies through a real browser), and it needs no window. For demos it splits a web walkthrough into a Stage window for the audience and a hidden Recorder that produces 1080p30 video with evidence files over CDP, without forking the browser engine; flows are scripted as scene.yaml packages driven from a Director console plus a loopback control API on 127.0.0.1:7317; headless recording shows no window, takes no focus and stays out of the Dock. Requires Node 22.12+ on the 22 line, 24.x, or 26+ (the 23 and 25 lines are excluded).

Exposes Chrome DevTools Protocol to the agent through 17 tools for page summaries, element locators, activity timelines, console, network, storage and downloads, reading either a browser the plugin launches on its own or the tab the user is currently browsing through a small browser extension.

A self-justification gate for outbound claims: a claim needs a checkable source (an http(s) URL plus a quote of at least 8 characters), every number in the text must appear in a quote, and a quote whose vocabulary overlap with the post is under 10% counts as an invalid citation; absolute wording is rejected when the only evidence is second-hand. All verdicts come from deterministic rules — no model call, no network. The guard subcommand refuses to run the publish command when the gate fails (fail-closed), explains what is wrong and how to fix it, and requires an --ack note on the second attempt stating what changed; after three attempts it stops and hands back to a human. Ships check / explain / fix / guard, with 30 tests.