dsh-suite/packages/plugin-deus

One-Sentence Positioning

神模扳机 is a "Minimalist Prompt Trigger Experiment Platform": built-in 5 minimalist trigger presets for one-click injection into conversations, then performs heuristic fingerprint detection on the model's first response, with local Wilson 95% CI trigger rate statistics; includes an arbitrary prompt A/B experiment platform and per-round anchor drift monitoring. The plugin honestly states in multiple places that "Divine Model 3 Mode" is a community observation rather than DeepSeek official confirmation, and does not promise to trigger any hidden capabilities.

Core Capabilities

• One-click injection of 5 built-in minimalist prompt presets (empty input / single character x / we-start guide / Let-me start / pure slash), all editable in the panel (lib/index.js:24-30 / lib/client.js:259-287).
• Two injection paths: "档 A" via Settings panel copy-paste, and "档 B" via the ⚗ chip row above the conversation input box; the latter can write to draft with one click and optionally auto-send (README.md:17-21 / lib/client.js:381-593).
• Heuristic fingerprint detection on the response start for each injection (Pure Region / Chinese / Divine / Undetermined), prioritizing reasoning flow with visible text as fallback; Chinese "我们需要 / 我们应该" in reasoning flow counts as Divine-equivalent (lib/index.js:45-61).
• Local JSONL experiment log + Wilson 95% CI trigger rate panel, supports CSV export, one-click copy Markdown summary for posting, and mini trend chart of last 40 entries (lib/index.js:299-349 / lib/client.js:673-716).
• On install, writes 2 anchor agent presets (narrow anchor ~2 tools / wide anchor ~8 tools), performs per-round fingerprint detection on deus/minimal series sessions and prompts for drift; one-click reanchor or auto-reanchor when drift occurs (lib/index.js:167-218 / lib/index.js:490-531).
• Conversation page dock detects non-anchored sessions and shows gray chip guiding preset switch, implementing the product guidance for the "25 tool session injection ≈ 0%"实测结论 (lib/client.js:603-606 / README.md:29).
• v0.4 Experiment Platform: fill in any two prompt variants + N runs per variant, automatically opens new sessions to isolate context, runs real injection and produces fingerprint distribution comparison table with Wilson CI + two-proportion z-test (lib/client.js:462-528 / lib/index.js:783-813).

Technical Implementation

• Language: JavaScript (host half ESM Node + browser half loaded via __ModuleLoader__.load, no build step; packages/plugin-deus/package.json:5-11 / lib/client.js:14-16)
• Key Dependencies: No third-party runtime dependencies; host half only uses Node built-ins node:fs / node:module / node:path / node:os / node:zlib; browser half depends on host runtime-injected react / slots / locale / inputActions (lib/index.js:13-17 / lib/client.js:719-741 / packages/plugin-deus/package.json:16-24)
• Architecture Pattern: cordis dual-partition plugin — host half apply(ctx) listens to session/event, registers /deus/* HTTP routes (presets/log/stats/version/anchor/bench), and brings matched sessions into "per-round anchoring" state machine; browser half registers settings.section and conversation.input.dock two slots, providing settings panel and chip row above input box (lib/index.js:476-836 / lib/client.js:725-737). cordis.patch.yml has only one line insert: { id: deus-mode } declaration (cordis.patch.yml:7-9)
• Entry Files: host half lib/index.js, browser half lib/client.js, exposed via package.json#exports . and ./client respectively (packages/plugin-deus/package.json:6-11)

Applicable Scenarios

DSH users who want to try "whether minimalist prompts can change model response initialization" in a reproducible, quantifiable way, such as doing prompt A/B experiments, writing prompt research notes, or curious about "whether the minimalist trigger phrases circulating online actually work." Not suitable for scenarios pursuing "out-of-the-box Divine Model" — the plugin repeatedly states this is a heuristic tendency signal rather than a deterministic capability, also suitable as a general minimalist prompt A/B testing platform ("Divine Model is just the first hypothesis tested", experiment platform built-in since v0.4; README.md:6 / README.md:56).

Prerequisites and Compatibility

Installation

dsh plugin --profile web add github:whyihaveyou/dsh-suite/packages/plugin-deus

Configuration

This plugin requires no standalone configuration — no cordis Schema fields, no separate DSH settings group; interaction behavior controlled by 4 indirect switches:

FAQ

Q: Can installation really make the model enter "Divine Model"?

A: No. This plugin displays an honest disclaimer fixed at the top of README and UI: "Divine Model 3 Mode" comes from community observation (X @NFT_Chen), not verified by DeepSeek. The plugin only does prompt injection + response start detection + local statistics, does not promise to trigger any hidden capabilities, and observed trigger rates may just be sampling noise.

Q: How to choose between the two injection paths (A / B)?

A: 档 A is one-click copy preset text in Settings panel then manually paste to input box, paired by host via exact text matching with minimal side effects; 档 B is clicking the ⚗ chip row above the conversation input to write to draft (optional auto-send), and actively notifies host for pairing. The same preset library is shared between both sides.

Q: Where are experiment logs stored? Are they uploaded?

A: Local only, path is $DSH_HOME/deus-mode/log.jsonl (default ~/.dsh/deus-mode/log.jsonl), append-write JSONL, no network upload. Can clear via panel "Clear Log" button or DELETE /deus/log, also supports CSV export.

Q: Why is the trigger rate almost 0% for "standard" sessions with 25 tools?

A: Plugin includes honest declaration: experiments show trigger rate ~0% for standard sessions (25 tool directory); triggering requires both "minimalist system prompt + small tool directory" conditions met. This plugin writes 2 agent presets on install (narrow anchor ~2 tools / wide anchor ~8 tools), recommend selecting in Agent preset selector; conversation page dock also gives gray chip guidance to switch on non-anchored sessions.

Q: Why does drift occur even with "anchor maintenance"?

A: Because constant composition doesn't mean constant fingerprint. README cites team experiments: Divine version maintenance rate 0/6 after tool directory completion, and even with identical composition there's 44-89% per-round fluctuation. So plugin detects fingerprint per-round, dock chip turns red on drift, click to send reanchor prompt, also can enable "auto-reanchor on drift" for plugin to automatically resend.

Q: How to use A/B Experiment Platform? How is it different from normal usage?

A: Click 🧪 button in ⚗ chip row on conversation page to open experiment platform, fill variant A, variant B, runs per variant N (1-20), click start. Plugin automatically opens new session for each trial to isolate context, pairs by prompt text exact matching queue attribution, ultimately produces Wilson 95% CI per class, Δ(A−B), two-proportion z-test with significance markers (✱ p<0.05 / ✱✱ p<0.01). "Divine Model" is just the first hypothesis tested, essentially an arbitrary prompt A/B experiment platform.

Q: What residuals are left after uninstalling the plugin?

A: Local log file $DSH_HOME/deus-mode/log.jsonl, presets.json, and agent preset directories installed to $DSH_HOME/.agent-presets/deus-anchored{,-wide}/ (with .deus-managed tags for idempotent identification) are not automatically cleaned; to fully remove manually delete ~/.dsh/deus-mode/ and ~/.dsh/.agent-presets/deus-anchored*.

Learning Curve

Beginner — one command install takes effect, Settings panel provides 5 built-in presets just click to copy; advanced features (drift reanchor / A/B experiment platform) enable as needed. If just want a few quick tries, copy-paste mode 档 A is sufficient.

Known Issues and Limitations

• "3 Mode" not verified by DeepSeek, honest declarations at UI top and README top; observed trigger rates may just be sampling noise, always keep control sampling and do your own statistical inference (README.md:3-10 / lib/client.js:23 / lib/index.js:1-7).
• Anchor maintenance has 44-89% per-round fluctuation (even with unchanged tool directory), Divine version maintenance rate实测 0/6 after tool directory completion — single anchoring cannot be effective throughout, plugin has mitigated via per-round detection + drift reanchor, but "reanchor on drift" also breaks current conversation context (README.md:25-28 / lib/index.js:490-531).
• DSH host version not declared in package.json; runtime auto-detects via /deus/version (candidate package names + filesystem backtrace process.argv[1]), may return 'unknown' if installed in non-typical location (lib/index.js:424-464 / lib/index.js:814-818).
• README cites team N=440 API sampling on deepseek-v4-pro done 2026-08-15, cited file research/deus-mode-matrix.md not found in repo glob, data only from README self-report, treat as tendency signal (README.md:60-62 / Not Found).
• Uninstall does not auto-clean ~/.dsh/deus-mode/ and ~/.dsh/.agent-presets/deus-anchored* directories, manual deletion required (lib/index.js:184-218 / lib/index.js:221-232).
• benchReportFrom returns only the most recent run's report per session (first entry sorted by lastTs), switching runs requires selecting runId in panel dropdown (lib/index.js:374-375 / lib/client.js:353-360).
• A/B Experiment Platform each trial strongly depends on sidebar "New Session" button (via aria-label i18n matching), falls back to same-session consecutive runs with ⚠ warning when not found — context not isolated in this case, detection data may be affected by previous round (lib/client.js:431-450 / lib/client.js:54).

🌐 中文 · English

别再翻 dsh-plugin topic 了，这里都是还能跑的插件。

dsh-suite 是 DeepSeek Harness（DSH）插件的活目录——每小时自动刷新、每日兼容实测——外加 DSH 内置插件商店、create-dsh-plugin 脚手架、几个自研插件，以及一个**一体化工作台**（🚧 开发中）。

🧭 生态 / Ecosystem

• 🌐 目录网站 — 在线逛全部插件
• 📖 DSH 中文文档 · dsh-docs.com — 中文教程与文档，开箱即用无需翻墙
• 📕 DSH 插件开发指南 — 中英双语成书
• 🎨 皮肤画廊 — 151 款皮肤昼夜双预览

---

为什么做 dsh-suite

DSH 发布时没有官方插件 registry。现在找插件只能翻 GitHub 的 dsh-plugin topic（50+ 个零散小插件）和当天冒出来的几个静态 awesome-list——而 DSH 自己还在发破坏性变更（breaking changes）。

所以我们做了四件事：

1. 一个「活」目录——880+ 精选插件，CI 每小时刷新数据、每天把收录的包真实装进临时 profile 重测兼容性。
2. 一个内置插件商店——@dsh-suite/plugin-manager 在 DSH Web UI 的设置页里加一个 Store 标签：逛目录、搜索、看徽章、一键安装，全程不用离开 DSH。
3. 一个脚手架——npm create dsh-plugin 一条命令生成可跑的 dsh.bundle + Cordis 骨架。官方没给脚手架，而「怎么迁移我的插件」是社区呼声最高的需求之一。
4. 几个自研插件——不是纯搬运，有第一方产出。
5. 一个一体化工作台——dsh-workstation，开箱即用版（🚧 开发中）。

快速开始

# 1. 逛目录网站
open https://whyihaveyou.github.io/dsh-suite/zh.html

# 2. 把插件商店装进你的 DSH Web UI
npx @deepseek-ai/dsh plugin --profile web add @dsh-suite/plugin-manager
#    → 重启 Web UI，然后 设置 → Plugins → Store

# 3.（开发者）造一个自己的插件
npm create dsh-plugin@latest my-plugin

📚 插件目录

⭐ 精选

🧰 工具

🧩 技能

🎨 界面

🐋 皮肤

💬 会话

🧠 模型

🛡️ 沙箱

🎛️ 编排

🔌 桥接

📦 预设

🧷 其他

徽章含义：🟢 兼容 · 🔴 不兼容 · ⚪ 未实测 · ⚫ 弃坑。 共 1676 个条目，按分类分表、类内按 ⭐ 降序。收录 / 字段词典见 docs/catalog-schema.md。

兼容性

每个条目带一枚徽章，CI 每天对照最新 DSH 版本重测（静态 peer 比对 → 安装检查 → 配置组装，三层均无需 key）：

机器可读的每日结果见 data/compat-report.json，工作流见 .github/workflows/compat.yml。

🧩 自研插件（第一方）

版本号由 npm 徽章实时拉取 latest，永不过时；目录表格则由 hourly bot 每小时刷新。

🛠 create-dsh-plugin

npm create dsh-plugin@latest

交互式脚手架，多套模板 + Claude Code / MCP 迁移指南，见 docs/migration-guide.zh-CN.md。

贡献

收录目录条目、提交插件，或阅读 15 条插件设计准则。收录标准与提交流程见 CONTRIBUTING.md；想提名新条目，用收录申请模板开 issue。

路线图

• 现在（MVP）——目录 + 兼容性 CI（第一层）+ 脚手架 + 2 个自研插件。
• 接下来——安装 / 配置组装两层兼容验证、star 自动刷新、迁移指南、plugin-team-board。

License

MIT © 2026 whyihaveyou

---

📚 文档

• 贡献指南——收录标准、提交流程与 15 条插件设计准则
• 文档——字段词典、分类体系与迁移指南