Plugins

Multi-model orchestrator: the main brain decomposes complex tasks and dispatches subtasks to GLM / Kimi / Qwen sub-agents by each model's strengths, with per-model token usage.

Prompt enhancer for the DSH web UI with basic/standard/expert modes: standard/expert modes rewrite drafts with project context and recent session history, retry gracefully when gateways reject reasoning-effort settings, and show a before/after review with undo.

Blocks edit/write writes that contain characters matching configurable regex rules, executed before the tool runs, with configurable messages and same-turn pwsh propagation; default rule matches full-width parentheses.

DSH Desktop enhancement plugin: it puts verifiable facts on the table. Constraints are injected only when relevant, so idle chat costs no extra tokens; injection is layered (session-stable text in the system prompt, volatile content as a tail snapshot) and the prompt cache stays valid. Discipline: each turn is classified (question / research / discussion / diagnosis / execution) with boundaries enforced - no file edits on question turns, no unsolicited fixes on diagnosis turns; citing code lines not opened this session, negative claims about unseen symbols, requirements that do not match the deliverable, and dressing up “I did not check” as “waiting for your decision” are all called out; replies are written for humans - a code, field name or id must be explained in plain words the first time it appears. Memory and knowledge: 75%/90% context warnings plus an exported session recap (say "continue" in a new session); project knowledge accumulated across sessions by working directory, auto-captured and deletable by id. Inspectable artifacts: task contract / change record (auto-logged, unverified items flagged at delivery) / hypothesis record (conclusions must pass an evidence gate) / design decisions. Dashboard and personas: routing, trigger hits and block assembly are written to lume-metrics.jsonl and queryable via lume_metrics; personas are distilled from chat logs, novels, scripts or character sheets, with long-term and temporary memory, a memory star map and persona card export/import.

Governed long-term memory for DeepSeek Harness: append-only evidence ledger with write guards, 7-tier authority claims, supersede chains, and budgeted pre-step injection with per-domain quotas.

DeepSeek Harness plugin that aggressively saves tokens without affecting task quality: lossless head/tail truncation of tool results (default), grep-result folding, pressure-triggered dsh-compaction-basic delegation, and an LLM-summary CompactionEngine subclass for the aggressive mode. Saves 97.6% of input tokens on an 8-scenario synthetic suite, with all reductions routed through dsh event flow so model-visible = logged stays intact.

Live token stats under the DSH Web composer: streamed tokens/s, output token and first-byte latency, plus per-step estimate-vs-actual reconciliation using the official BPE tokenizer, to tell a slow stream from a stalled one.

Personalization for the agent: a Web UI panel (about-you, reply style & tone, traits for headings/lists and emoji, output language, custom instructions) and `/soul` commands, with nickname/occupation/bio/style/traits/language compiled into a system prompt and re-injected into active sessions. Exposes `soulConfig` for other plugins and a `set_persona` tool the agent can call.

Switch the language of the agent's thinking/reasoning process (chain-of-thought): 简体中文, English, Русский, Français, Deutsch, Español, 日本語, 한국어 and more, or auto-follow the system locale. Adds a Settings → General row, a system-prompt instruction (per-prompt assembly, applies to new sessions), and the /thinking-language command.

Give DeepSeek (text-only) models vision in Claude Code, Codex, and Agent Plugins clients: describe images via any OpenAI-compatible vision endpoint.

Research-debt guard for DeepSeek Harness: nudges coding agents to search docs, GitHub and the web before repeated local trial-and-error.

Learns what you leave unsaid in your prompts from messy turns and your own corrections, and injects the learned directives into the system prompt of every new conversation; the directives are editable in Settings and a composer button rewrites the current draft with them.

LLM adapter plugin for locally deployed Qwen models behind a vLLM OpenAI-compatible endpoint, with per-model multimodal switch, fully configurable reasoning efforts, and a web settings page for editing the deployment from the frontend.

FormatForge for DeepSeek Harness: forge any file format (pdf/docx/xlsx/pptx/eml/toml/...) into AI-readable structured data. Thin bundle that spawns the local Python CLI (python -m formatforge) and returns protocol JSON; enhancement is delegated to the cur

Register locally-defined custom agents as callable subagents in dsh: the main conversation invokes any agent by name via use_agent, each running as a real dsh subagent with its own persona. Supports breakpoint resume for interrupted runs.

Let your AI agent see and operate a real Android phone: phone_look (vision + UI-tree fusion), tap/swipe/type, screenshot — over adb, for any MCP client.

A-share market dashboard: realtime quotes and indices, millisecond full-market search, market sentiment thermometer, candlestick K-line detail popups, watchlist, ETF/convertible-bond/LOF boards, screener, and a 7x24 aggregated news feed.

Wraps globalThis.fetch to route DeepSeek Harness model-API requests through an HTTP/SOCKS proxy by target domain, leaving all other requests on the direct connection, without modifying Harness source.

Makes plain Enter insert a newline in the composer and Ctrl/Cmd+Enter submit, toggleable in settings.

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.

Manages OpenCode Zen free-tier and Go-tier models in settings.yaml — fetches the live model listings from opencode.ai, reports configured-versus-live drift per route, and adds or removes entries through four agent tools and an OpenCode Models settings section.

User Message Timeline - floating pill rail with hover preview, click/drag jump and pagination, switchable left/right.

Windows system and custom desktop notifications for DeepSeek Harness events: task completed, failed, running, awaiting approval, and awaiting answer. Click-to-jump to the session.

Aggregated web search provider that fans one query out to 11 engines (Exa, Parallel, DuckDuckGo, SearXNG, Brave, Tavily, Serper, Mojeek, DeepSeek, Anthropic, OpenAI) concurrently, URL-dedupes and merges the results. Ships a settings card for toggling backends, storing API keys in the credential store, and tuning concurrency and timeouts.