Plugins

Identity-gated Tailscale access to the Web GUI with dsh still bound to loopback: a 127.0.0.1 proxy published by `tailscale serve` admits a request only when the Tailscale login it stamps is allowed, with an optional per-device allowlist, so the privileged /api methods behind the loopback check — Settings, Models, Plugins — answer without dsh being told to trust a remote host.

Agent-driven kanban board: one board per workspace, per-task agent sessions, and a guarded human/agent workflow (Backlog → To Do → In Progress → Review → Done).

Host-side unattended job scheduler: runs one-shot agent tasks or shell commands on cron, interval, or one-time schedules, with persisted at-most-once dispatch state, overlap/misfire policies, and bounded run history.

Creates and inspects RecoWork local-agent-project workspaces through two bounded DSH tools and a web workspace card.

A governed swarm of six specialist DSH agents — orchestrator, planner, knowledge-base bridge, developer and two reviewers — turns a requirement into a strict phase pipeline where every delivery must prove itself with structured evidence before the next phase may start; reviewers are read-only, code merges only after a review pass, and an append-only event log gives full replay and audit on a live kanban board. Its design draws inspiration from the Hermes Agent kanban.

Browse the public skillhub.cn catalog from Settings and install selected SKILL.md bundles into ~/.dsh/skills.

Embedded MCP server letting external agents (Hermes, Claude Code, etc.) drive DeepSeek Harness over StreamableHTTP: sync/async task execution, a three-tier file sandbox, a Web-backed approval bridge, session management and search, and path-jailed file tools.

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.

Right-hand widget rail for DeepSeek Harness Web UI: live session stats, task/context/heatmap/quote widgets, OpenCode Go quota, multi-column grid layouts (1/2/4 columns) with 2x4 long tiles, a continuous magnification wave, and per-instance sizing; widgets are installable, reorderable and resizable from Settings > Components.

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.

Manage Git-backed DSH plugin markets, installed agent skills and optional Codex hooks from the web settings page.

AI can write resume text, but templates may still look wrong, pages may overflow, whitespace may remain, and small edits can destabilize the layout. dsh-resume provides an end-to-end DSH workflow for truthful content refinement, template selection, A4 layout tuning, visual review, and tailored resume versions for different job-search scenarios.

Control a HarmonyOS phone from the DSH web UI with AI: live H.264 screen mirroring, mouse touch and system keys, hilog streaming, and agentic tools that let the model read the screen, locate UI controls, then tap, long-press, press keys, or type.

Rewrites the Chinese copy of a market-installed plugin into English using a model already configured in the harness — automatically when the plugin is installed, then live-reloaded. Each file is translated in one whole-file pass, validated by a real parser where one exists (node --check for JavaScript, JSON and YAML parsers otherwise) and handed back to the model with the error for up to three repair attempts; files above 48 KB, or that never parse, fall back to a line-range pass that checks every returned line against the original for structure. Trigger phrases and locale tables can optionally be preserved rather than translated, and the untouched original is kept beside the file as a .zh.bak.

Hardening guard for mcp__omni__parse in DeepSeek Harness: a tools/pre-execute listener that denies private/reserved host URLs (SSRF), enforces an allow-list or ask (consent), and is fail-closed when allowedRoots is empty.

Named subagent roster: manage role entries (model, persona, tool filter, depth, background mode) from a hot-reloaded settings page; the model picks one with list_subagents and dispatches by id with delegate — foreground, background, or continuable.

Deep research as an agent mode: one reviewed loop that plans, researches in adaptive parallel rounds driven by the gaps the researchers themselves declare, synthesises a cited report and then reviews it adversarially, instead of a script the model rewrites per call. Questions the round budget never reached are named in the report rather than dropped.

Config-driven lifecycle hooks, including context injection: declare event, handler and outcome in YAML, where the handler is an in-harness tool (MCP included), a shell command or an HTTP endpoint, and the outcome is model-visible context, a denied tool call, or fire-and-forget.

Points web_search at a self-hosted SearXNG instance instead of the native route: pages until maxResults is satisfied, and reports degraded coverage when engines fail so a thin result set is not read as a thin web.

A guarded local fetch provider for the ctx.web seam — the backend dsh does not ship for its own web_fetch tool. Every redirect hop is re-resolved and checked against the non-routable ranges, with a streamed size cap and a content-type allowlist.

Selectable SearXNG, Brave, Tavily, Gemini Search Grounding, or Wikipedia backend for the native web_search tool, with Web settings and no implicit fallback.

Cue Omni Reader for DeepSeek Harness: registers the audited omni-reader-mcp MCP server so its URL/file parse tools surface as mcp__omni__ (parse, get_parse_status, read_result, read_outline, save_result, discard_result, cancel_parse).

Progressive tool disclosure: a deferred tool group costs one catalog line instead of its full schemas until the model loads it with tool_search, with one switch per group in Settings and each group measured against the live registry.

A thinking level stored per provider and model, so reasoning effort outlives the session it was set in and also reaches subagents, which dsh never asks. The in-session picker still wins, and a model with no row keeps its provider default.