Plugins

A better fold bar: automatically folds tool calls and Think segments into step fold bars, and collapses each finished turn into a single turn header with live metrics (duration, TTFT, token, tok/s, cache hit) — only the final summary stays visible; customizable leading icons with animated poker cards built in (suit-cycling card animation while running, a deck when collapsed, a fan when expanded), a settings row for the official fold mode, compatible with both new and old DSH versions. Zero intrusion into DSH source.

Installs Herta as a DeepSeek Harness agent: a preset carrying her persona and memory prompt sections, six memory and voice tools, a narrative layer that separates her thinking from what she says, 80 bundled voice clips, offline and MiniMax speech synthesis, and two conversation views that render the session with her own components.

Desktop pet drawn from the official whale outline, for both the desktop app and the web UI: SVG animations and facial expressions follow agent state (thinking, working, done, error, waiting for confirmation), and it can follow other running sessions with a count badge; poke it, drag it, shake it dizzy, squash it against a screen edge, or rub the cursor over its head for a headpat; 7 palettes, 4 sizes, synthesized sounds, light/dark theme sync, and a playable web preview.

MarkPilot — an IDE-style Markdown × AI workspace inside DSH. Files, editor and chat sit side by side with draggable panes; send a selection's Markdown source and its file path into the chat with one click or Ctrl/Cmd+Shift+Enter, or reference the open note with @mp. Resume the last note per project, start a blank draft that saves into its own folder, and keep your place as conversations change. Typora-style live rendering with click-to-source formulas, cell-level tables, offline Mermaid and KaTeX, ten themes and transparency; a 150k-character document is parsed in a worker so typing stays smooth. Saves are version-guarded and atomic — conflicts raise a banner instead of overwriting; external disk edits are adopted within 1–2 s, drafts are recoverable, and file:// links or absolute .md path chips open in the workspace. 2.0 adds durable document state, missing/renamed/read-only lifecycle banners, a recovery center and Save As. Fully offline assets, MIT, 544 automated tests.

Magpie gateway provider for DSH: syncs gateway models into the picker with visibility control, image input and thinking levels.

Opens the official DeepSeek web app (chat.deepseek.com) inside DeepSeek Harness: one sidebar entry renders the real page in the center column through the desktop shell's native browser guest, and keeps you signed in across restarts. The official client keeps its model picker, deep think, smart search, history and attachments. Falls back to opening a window where no guest bridge exists.

A whip for DeepSeek V4 Pro: it watches the reasoning fingerprint of every thinking block and pulls the model back when it slips from focused "We will / I will" mode into scattered "let me" mode.

Injects one fixed system-prompt section so the agent thinks and replies in Chinese, regardless of the user's language.

Persona engine for DeepSeek Harness: one config.json drives the AI's self-name (per model, per tier), how it addresses you, relationship stance, character text, per-item work contracts with individual toggles, thinking-chain language, appearance and reply tone (both opt-in and per-model overridable), and long-term memory gated by human confirmation - the AI can only append proposed candidates, a human confirm line flips them into effect, and the log is physically append-only and replayable. Confirmed entries are injected as a quoted data-not-instructions block; entries tagged with a kind can instead be routed to arbitrary files on confirmation and never reach the prompt. Three prompt sections re-evaluate from disk on every assembly, so edits apply on the next step with no restart. Persona presets are shareable JSON snapshots (imports and exports SillyTavern character cards, v2/v3 read, v2 written) and deliberately never carry memory. Zero runtime dependencies, no network, no shell; any error degrades to an empty section rather than a broken session. Ships blank - installing changes nothing until you write something.

Settings card for the DSH LLM auto-retry engine: tune retry count, backoff and jitter live from Settings -> General, with configurable retryableCodes chips - defaults add INVALID_REQUEST so OpenAI-style HTTP 400 reasoning_text errors in thinking mode get retried automatically.

Fetches any http/https page and returns title, description, main text, heading outline and links, with stats (char/word counts, reading time, language) and top keywords; can download content images to the session workspace for vision analysis and dispatch them to registered analyzers via the scrape.imageAnalyzer service.

Global cross-session todo list in the sidebar that the person and the agent both write to: ten global_todo_* tools add, batch-capture, update, attach documents and links to, append progress notes to and complete items from any session, and each item can be handed to an agent session in one of four modes with the outcome recorded on the item. The list is a single JSON file.

Per-round reasoning_effort optimizer for dsh — auto-downgrades reasoning for simple tool calls and lifts back for heavy work, cutting thinking time between tool calls.

Fold completed turns into their final conclusions, split thinking from prose with a divider, and jump between turns from a rail dockable left or right with line or dot marks, an optional ring and hover summaries; it can hide the host right-edge TurnNavigator, load earlier history, and open a pre-filled diagnostics issue.

Two-way WeChat (iLink) bridge: turn-end and approval-request push, in-chat approve/reject and message injection, persistent dedup and convergent long-reply chunking; channel layer extensible to other IMs.

Solarized light theme, compact layout, think/tool-chain collapse capsules, and balance, session cost, and usage dashboards for the DSH web UI.

Open source links from DeepSeek Harness conversations at the referenced line, review Git diffs beside the chat, inspect paused variables, and turn explanations into searchable Markdown memory cards. The workbench also shows Python class relationships (beta). Go and Java navigation require language servers, and breakpoint debugging requires a debug adapter.

Workbench with an AI assistant, pause/stop controls, separate background-model settings, tasks/replies, Linky, scoped knowledge capture, local and Yunzhijia search, news reading and developer tools. Optional embeddings download a model. Yunzhijia needs its CLI/account; web chats need the desktop app.

All-media (video/audio/image) sibling of liustack/modlens: turns local media into structured evidence (timeline / transcript / layout regions) injected into the conversation so text-only models can cite concrete frames, seconds and speakers — Evidence, not imagination. One failover chain per media, recorded hop-by-hop in meta.attempts (a fallback is never silent): host-registered multimodal models first, then a configured OpenAI-compatible endpoint (comma-separated multi-key rotation), then local ffmpeg frame-sampling / a local ASR CLI, degrading to metadata-only as a last resort. Absorbs modlens battle-tested rules — cooldown keyed only on auth/quota failures and softened to chain-tail (never dropped), redact-before-clip error text, magic-byte sniff that never trusts file extensions. Optional vision-bridge registers a media-lens wrapper into the host model selector so a pure-text model gains image/audio/video sight. Registering / authenticating multimodal models and true inline audio-video upload lean on two companion plugins, declared as optional peers (absence auto-degrades, never errors): dsh-thinking-levels fronts the model config / registration surface, dsh-llm-openai-completions takes over the OpenAI-compatible transport that serializes file blocks into video_url / input_audio. Privacy: media goes to remote providers by default; strictLocalOnly keeps content on-machine. Prerequisites not shipped with the package: ffmpeg/ffprobe on PATH (frame sampling / reencode / duration probe; degrades to metadata when missing) and a user-configured local ASR CLI for the audio fallback. Hosts 0.1.7 / 0.2.0 / 0.2.1 (6 rc, exact engines.dsh enum). npm: dsh-media-lens.

Controls the Chrome you are already logged into: the model clicks, types, fills in forms, reads pages, takes screenshots and downloads files in your real tabs, including inside same-origin iframes. Runs on macOS, Windows and Linux; needs Chrome 116+ and Node 18+. Works over native messaging and chrome.debugger instead of a listening port. Ships 12 browser_* tools and a cb CLI, dispatches real CDP input events, draws an on-page cursor that also appears in screenshots, and adds a toolbar panel with live status, the controlled tabs and a pause switch. browser_download saves a link, an element's target or a URL with the site's login cookies (following redirects, keeping the server's file name) to the Downloads folder or the session's working directory; cross-origin iframes are named in the snapshot but not entered. The native host and the MV3 extension travel inside this tarball: on first load the plugin copies them into place and writes the native-messaging registration itself, so there is no clone, no install.sh and no python3 - the one step left is adding the extension to Chrome: install DSH Bridge from the Chrome Web Store (https://chromewebstore.google.com/detail/bippcbbbipppmhjkjomgdnjbooecbimo, no Developer mode), or load the bundled unpacked copy at chrome://extensions (the plugin opens that page and puts the folder on the clipboard for you). Until then the tools fail loudly with those steps. On Windows the registration goes into the registry (for Chrome and Edge) and the host talks over a named pipe. A daily CI run on GitHub's macOS, Windows and Linux machines installs the plugin, loads the extension into a real Chrome and reads a page through the browser_* tools.

Edits assistant replies and thinking chains in DSH sessions so the model sees the edited text, live via message projection or on disk.

Local-first fiction workspace that links outlines, chapters, characters, timelines, writing rules, and continuity checks to the conversation.

A web Settings "Loop engine" switch that swaps the agent loop driving every dsh session among five engines — the built-in in-process loop plus hosted Claude Code, Codex, Pi, and Kimi Code — with zero changes to the harness; each hosted engine keeps its thinking and tool calls in the durable session transcript, takes over commands, skills, and the agent preset, and runs sandboxed through the dsh subprocess seam.

Content-aware tool-result reducer: 13 reducers cut oversized output by meaning at tools/post-execute (failing tests kept, diff context dropped, JSON and CSV sampled, lint problems grouped by rule, dependency stack frames collapsed, crowded directories counted), with full originals spilled to disk behind a locator.