yjh051108/dsh-routing-suite

JavaScript

dsh-routing-suite — injector + router-standard kit: install the runtime injector first, then the task-aware reasoning-mode router preset (measured P1-P23).

ai-agentscordisdeepseek-harnessdshdsh-plugin
Star Growth
Stars
7.1k
Forks
149
Weekly Growth
+147
Issues
35
4k6k
Aug 1Aug 13Aug 25Sep 6
Artifactsnpm
README

dsh-routing-suite — 注入器 × 思维模式路由 × 分级模式 套装

一个仓库装齐三件套:运行时注入器(免重启运行时管理层)+ 思维模式路由预设(任务感知推理模式,P1-P23 实测)+ 分级任务协议(脑暴出题 → 规格化计划 → 打卡制 → 组收官 → 终验,含红队门与审计端点)。

中文 | English

组件入口(快速跳转)

组件 说明 入口
injector 运行时注入器:dev_* 工具全家桶(注入/热重载/卸载/侧挂转正/路由自愈) injector/README.md · 文档
preset 思维模式路由预设(router-standard / router-spec,按模型 persona 路由) preset/README.md · 实验报告
graded 分级模式:会话级两级任务协议(6 工具+三模式+小类模式粒度+红队门+审计端点+超级面板) graded/README.md · 架构 · 理论 · 实测数据

研究与数据(graded)

"模型也会偷懒——我们测到了,还做了对抗" —— 分级模式插件的公开研究位:

一键安装

dsh plugin --profile web add github:yjh051108/dsh-routing-suite

本套装已含上述三组件(injector/preset/graded 均为仓库内普通目录,内容直接入库); graded 发布物:graded/dsh-external-dsh-graded-mode-0.0.1-rc1.tgz(或 Release 附件)。

DSH Target>=0.1.0-rc.6 <0.2.0(已跟进 rc.8 / 0.1.1-rc.2 / 0.1.2-alpha.1)

DSH 目前处于 developer preview,官方明示会有破坏性变更(breaking changes)。 本仓库的版本跟进记录见 preset/CHANGELOG.md

安装链(三步)

# 1. 拉套装(单仓库:injector/preset 内容已直接入库,无需 submodule)
git clone https://github.com/yjh051108/dsh-routing-suite.git
cd dsh-routing-suite

# 2. 一键安装(注入器装配 + 预设复制 + 布局自检 + 提示重启)
.\install.ps1

或手动:

# 步骤 1:装配注入器(官方装配,重启后由 bundles 接管)
dsh plugin --profile web add .\injector
# dsh 不在 PATH 时:npx '@deepseek-ai/dsh' plugin --profile web add .\injector

# 步骤 2:安装 router 预设(每个预设目录平铺复制到 .agent-presets 下,DSH 只扫一级子目录)
$target = Join-Path $env:USERPROFILE '.dsh\.agent-presets\router-standard'
Copy-Item -Recurse .\preset\router-standard $target

$target = Join-Path $env:USERPROFILE '.dsh\.agent-presets\router-spec'
Copy-Item -Recurse .\preset\router-spec $target

# 步骤 3:重启 DSH → 新会话选择 Router Standard / Router Spec (experimental)

注意:不要复制 preset 整目录(会多套一层,DSH 发现不了预设)。

组件

路径 仓库 版本 作用
injector/ dsh-super-injector v0.3.3 运行时注入器:dev_* 工具全家桶(注入/热重载/侧挂转正/卸载/路由自愈);github: 装配由 prepare 钩子自动构建
preset/ dsh-router-standard v0.3.0 … 主线 v1.19.1/v34 思维模式路由预设:router-standard(分类 persona + 完整 sections)/ router-spec(深度思考优先)。router-pro 为规划中(planned),未随 v0.3.0 发布
graded/ dsh-graded-mode v0.0.1-rc1 graded 模式:会话级两级任务协议(脑暴选择题对齐 → 北极星定稿 → 规格化计划 → 按规格注入 → 打卡制 → 组收官 → 终验);6 工具 + 三模式 + 小类模式粒度 + 红队门 + 审计端点;graded/dsh-graded-mode-3.2.0.tgz 可直接 dsh plugin --profile web add

版本号以各组件仓库的 git tag 为准(列内链接直达对应 Release)。

三个组件随本仓库统一演进(injector/preset/graded/ 已是仓库内普通目录,内容直接入库);上游独立仓库保留用于独立发布,后续可转镜像/归档。预设安装目录为 preset/router-standard(已平铺,无额外嵌套)。

router-standard 预设能力(P1-P23 实测摘要)

  • 三行为带 + weak 内路由:spec(计划-集体)/ react(执行者)/ mixed(陷阱,回避)/ weak(模型自分类)
  • 按模型选 persona:Pro=spec 句+few-shot(区分度 +5.0);Flash=neutral+classify(+5.7)
  • 近距离引导:每轮用户消息后注入固定引导(缓存 92-94% 命中),路由 96% + 收敛 100% + 反稀释
  • 单任务三锚(persona 静态):回顾 + 收敛 + 反跑题 —— 开放任务完成率 0% → 100%
  • plan-mode 保留:只替换 persona section,plan 边界不失忆
  • AI 自优化工具dev_router_status / dev_router_mode / dev_mode_subagent

v0.3.0 变更(真实装配链路修复)

  • 首轮路由真实生效#13):经 agent/inbox/claimed 在装配前捕获首条真实用户消息,首个请求即按任务分类(此前所有会话首轮无条件 weak)
  • 近距离引导改走 agent/pre-step#34/#36/#55):引导与用户消息同请求注入,真实链路上可达,且不再产生额外的第二次 API 调用(此前每轮多 1 次调用 = 费用 2×)
  • 缺导入修复#11)、preset.yml YAML 引号(#53)、promoted 后完整回归(#44)、安装脚本与文档修正(#35/#42/#41)、injector git 装配自动构建(#40

文档

  • 注入器引导(规范铁律 10 条):injector/README.md
  • 路由预设论文与实验:preset/docs/paper.md + preset/docs/experiments.md(P1-P23)
  • 仓库结构迁移(submodule → 直接文件):docs/FLATTEN-MIGRATION.md

许可证

MIT。致谢:xiaobright/modeltest(V4.1b 评测)、xiaobright/dsh-anchored-standard(锚定机制)。

Related repositories
affaan-m/ECC

The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.

JavaScriptnpmMIT Licenseai-agentsanthropic
ecc.tools
252.5k37.9k
NousResearch/hermes-agent

The agent that grows with you

PythonPyPIMIT Licenseaiai-agent
hermes-agent.nousresearch.com
242.6k49.9k
deepseek-ai/deepseek-harness

DeepSeek Harness: Everything is a Plugin.

TypeScriptnpmMIT Licensecordisdsh
deepseek.com/harness
214.2k25.2k
firecrawl/firecrawl

The context API to search, scrape, and interact with the web at scale. 🔥

TypeScriptnpmGNU Affero General Public License v3.0aicrawler
firecrawl.dev
177.3k9.7k
langchain-ai/langchain

The agent engineering platform.

PythonPyPIlibraryMIT Licenseaianthropic
docs.langchain.com/langchain/
145.8k24.4k
DietrichGebert/ponytail

Makes your AI agent think like the laziest senior dev in the room. The best code is the code you never wrote.

JavaScriptnpmskillMIT Licenseagent-skillsai-agents
ponytail.dev
129.6k6.9k
Graphify-Labs/graphify

Turn any codebase, with its docs, SQL schemas, configs, and PDFs, into a queryable knowledge graph. A /graphify skill for Claude Code, Cursor, Codex, and Gemini CLI: local deterministic AST parsing, every edge explained, no vector store.

PythonPyPIskillApache License 2.0claude-codegraphrag
graphify.com
115.4k11.2k
browser-use/browser-use

🌐 Make websites accessible for AI agents. Automate tasks online with ease.

PythonPyPIMIT Licensellmai-agents
browser-use.com
112.8k12.4k
google-gemini/gemini-cli

An open-source AI agent that brings the power of Gemini directly into your terminal.

TypeScriptnpmcliApache License 2.0geminigemini-api
geminicli.com
106.8k14.5k
thedotmack/claude-mem

Persistent Context Across Sessions for Every Agent – Captures everything your agent does during sessions, compresses it with AI, and injects relevant context back into future sessions. Works with Claude Code, OpenClaw, Codex, Gemini, Hermes, Copilot, OpenCode + More

JavaScriptnpmskillApache License 2.0aiai-agents
claude-mem.ai
93.4k8.2k
infiniflow/ragflow

RAGFlow is a leading open-source Retrieval-Augmented Generation (RAG) engine that fuses cutting-edge RAG with Agent capabilities to create a superior context layer for LLMs

GoGo ModulesApache License 2.0aiai-agents
ragflow.io
90.2k10.6k
bytedance/deer-flow

An open-source long-horizon SuperAgent harness that researches, codes, and creates. With the help of sandboxes, memories, tools, skill, subagents and message gateway, it handles different levels of tasks that could take minutes to hours.

PythonPyPIMIT Licenseagentagentic
deerflow.tech
81.7k11.3k