Plugins

Model-facing display_file tool that renders 36 file extensions inline in the web UI: images with a click-to-zoom lightbox, seekable video, audio, PDF, Office and OpenDocument files converted to PDF on the host via LibreOffice, and local HTML in a sandboxed frame; bytes stream from a signed HTTP route with range support, and PNG/JPEG/WebP/GIF also enter model context on image-capable routes.

Terminal coding agent for DeepSeek Harness with a full-screen alternate-screen UI, turn-aware session navigation, and a gated /ship workflow that takes one-line requirements through specification, TDD, and rerun acceptance checks.

In-session question navigator for the DSH Web UI: a dot minimap on the conversation edge (left or right, configurable in settings), one dot per user question — hover shows the full question text, click jumps to it.

Drives a real Chrome through a local MCP server so the agent can search Google/Bing/Sogou, read any page including JS-rendered and login-only content, automate click/scroll/type/key-press/select in a background window, and run platform search/detail/comment/download for Zhihu, Xiaohongshu, Douyin, Bilibili, X and Reddit; resumable comments, SHA-256-verified downloads, cookies never returned, localhost only.

Detects which conversation model an account is being served: launches concurrent throwaway probe conversations, scores each streaming chain-of-thought by how its paragraphs open, and cancels probes that read as the older model within seconds. Includes an offline self-check that replays 13 hand-labelled chains-of-thought through the same classifier.

Playwright/CDP web-fetch provider for the built-in web_fetch tool: renders pages in a real browser, denoises them (Readability + DOMPurify) and returns markdown.

Scroll-painting message rail for DeepSeek Harness: a slim right-edge rail with one indicator per user message, scroll-spy follows the reading position, hover unfolds a canvas with full message previews, and clicking jumps to any history message; animates in sync with better-sidebar panels.

A marquee mood-light ring around the DSH web UI edge that reacts to session state, turning green while the agent works, pink when done, and yellow when something needs attention, with colors, ring width, rotation speed, and flash frequency configurable from the settings panel.

Compact Codex coding mode for DSH: light prompt guardrails against GPT over-explaining, with an editable prompt and native Codex-style tools.

Runs work as a crew of role agents: a product manager writes the PRD or DoD, waits for your confirmation, then starts architect, engineer, QA and reviewers whose tool sets are locked per role, and shares work through files on disk.

Codex Pets-style floating pixel pet with 10 built-in Petdex-compatible pets, a 9-state agent-status engine (idle/running/waiting/review/failed plus client-side waving/jumping/directions), draggable overlay, pet picker/zoom settings, notification bubbles, and a `pet_say` tool for the agent to speak through the pet.

Use DSH from iPhone/iPad: an explicit-opt-in LAN reverse proxy (Host/Origin rewrite through the loopback trust fence, WebSocket upgrades included), iOS PWA chrome (home-screen icon, standalone meta, viewport-fit), and touch/mobile CSS (safe areas, keyboard lifting, composer row fit).

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.

Model high-availability failover (quarantine, circuit breaking, probe recovery) and subagent orchestration (fanout/pipeline/supervisor) with bilingual settings UI.

Perlica (Arknights: Endfield) themed tiered sound notifications: plan-ready, task-done, needs-your-input, and error tones; silent for plain chat, system-level playback (works in background), cross-platform (Windows/macOS/Linux), custom TTS sounds.

Pause / resume / cancel the in-flight conversation task from the composer area: forced pause interrupts immediately and remembers the interrupted tool, safe pause lands after the tool or reasoning finishes, resume asks for confirmation, and the default pause granularity is configurable in the settings.

LLM provider that runs the locally installed Claude Code CLI as the model backend, so requests go through an existing Claude subscription instead of a metered API key; native tool calls are bridged over MCP.

Neumorphic DSH Web theme with light and dark palettes, ambient lighting, material shadows, grain texture, frosted-glass surfaces, and micro-interactions.

Customizable brand area for the Web UI: replace the whale logo and DeepSeek wordmark with local images, and edit the HARNESS badge text (double-click to change, right-click to reset).

Highlights parts of speech (nouns green, verbs red, adjectives/adverbs purple) in rendered DSH Web Markdown, with light/dark palettes, toggles, and stream-aware updates.

Piano-key in-conversation navigation for the Web UI: one key per user message with hover ladder, message-preview tooltip, and active-message highlight; click to smooth-scroll.

Office workspace & session dashboard: a floating 6-column sprite panel visualizing workspaces, sessions, token usage and subagents, plus Agent Mail and an office log tab. Feishu/Lark views (message feed, meeting schedules, transcripts) additionally require the `@larksuite/cli` tool installed and authenticated separately.

Auto thinking-effort routing for dsh: a locally fine-tuned 7-question classifier (ModernBERT, 842MB, <1s CPU inference) picks low/high/max for no-intent messages; in-session retries escalate the tier (low to high to max) under exclusion constraints; the judge service fails safe to low if unreachable. A tier chip in the input bar shows the live tier and the last 20 judgments. All inference is local - task text never leaves the machine.

Chrome and Edge browser control that drives the real logged-in profile, tabs and cookies through an MV3 extension and a loopback Bridge, exposing browser_* tools, a Skill and /chrome commands.