DeepSeek Harness plugin

dsh-omni-router

Omni Router - auto-routing DeepSeek Harness preset: simple tasks run directly, complex tasks plan first with approval gate, coding tasks get TDD/delivery-gate guidance.

Jump to install

Source facts

Repository
qwe225380/dsh-omni-router
Latest update
Aug 21, 2026
Category
Models & Providers
GitHub stars
2
Format
plugin
Catalog evidence
Upstream dsh.bundle evidence
Evidence path
package.json#dsh.bundle
Checked against
0.1.0-rc.8
Upstream check date
2026-08-20

This evidence comes from the upstream catalog. This site has not installed, run, or security-reviewed the plugin.

Install

Start with a prompt that asks an agent to review the GitHub repository and source. Switch to the command if you want to install it yourself.

Copy this prompt into DSH, Codex, or another agent and ask it to review the GitHub repository and source first.

Do not install or run any commands yet. Read this plugin's GitHub repository, README, and relevant source code. Then answer the questions below clearly and directly so I can decide whether it fits my needs:

1. What is this plugin, and what problem does it solve?
2. Who is it for, and what are its typical use cases?
3. How is it used after installation? Include one minimal example.
4. What known limitations or privacy, security, compatibility, or maintenance risks does it have?
5. Give a clear recommendation: recommend, conditionally recommend, or do not recommend, with reasons.

Distinguish statements documented by the repository, inferences from source code, and unknowns. If evidence is insufficient, say so explicitly. Do not guess or simply repeat the README.

GitHub: https://github.com/qwe225380/dsh-omni-router
Plugin: dsh-omni-router
Author: qwe225380

Check the source files

Read the README and other files from this plugin directory before installing.

File explorer4 files
README.mdSource · read only
README language

Omni Router

> GitHub: https://github.com/qwe225380/dsh-omni-router > 中文说明

Omni Router 是一个 DeepSeek Harness Agent 预设,它不是一个“单一功能插件”,而是一层智能编排层:把复杂度路由、思维模式路由、任务类型识别、Plan Mode、项目上下文、TDD、交付门、Git 工作流等能力组合成一个统一的、自动运行的工程闭环。

Why Omni Router?

单独使用 dsh-routing-suite、dsh-trio、dsh-doublecheck、superpowers-dsh 时,你需要:

  • 自己决定什么时候用哪个预设
  • 自己记得先出方案还是直接做
  • 自己记得要写测试、要验证、要过质量门
  • 在不同插件/预设之间手动切换

使用 Omni Router 后,这些判断和编排会自动发生:

任务进来
   │
   ├─ 复杂度路由:plan / direct
   ├─ 思维模式路由:spec / react / balanced
   ├─ 任务类型识别:bugfix / feature / refactor / test / review
   │
   ├─ 复杂任务
   │    ├─ 自动进入 Plan Mode
   │    ├─ 自动收集项目上下文
   │    ├─ 生成代码化方案
   │    ├─ 注入 TDD / 交付门 / Git 工作流 / 验收清单提示
   │    ├─ 等你确认
   │    └─ 执行 + 验证
   │
   └─ 简单任务
        └─ 直接执行 + 轻量验证

Features

Core(核心,自带)

  • 复杂度自动路由direct / balanced / plan 三态路由,简单任务直接做,复杂/模糊任务先方案。
  • 思维模式路由:自动选择 spec(方案优先)/ react(直接执行)/ balanced(自动)。
  • 任务类型识别bugfix / feature / refactor / test / review / other
  • Plan Mode 集成:复用 DSH 内置 Plan Mode,方案必须经你确认后才执行。
  • 项目上下文收集:按任务类型 + 语义关键词发现相关文件,生成有界上下文摘要。
  • Risk Model:独立评估风险等级,高风险(数据库/auth/生产配置)即使复杂度低也强制 plan + approval。
  • Adaptive Rerouting:执行中可通过 omni_reroute / /omni reroute 在 direct ↔ plan 间动态切换。
  • 代码化方案模板:目标、范围、涉及文件、步骤、接口/数据变更、测试计划、风险、兼容性、回滚、验收标准。
  • TDD 技能路由:编码任务提示加载 test-driven-development / red-green-tdd,不重复实现 TDD。
  • 交付门技能路由:完成前提示加载 verification-before-completion / delivery-proof,不重复实现质量门。
  • Git 技能路由:编码任务提示加载 using-git-worktrees / git-discipline
  • 验收清单技能路由:方案批准后提示加载 executing-plans / writing-plans
  • 轻量验证技能路由:简单任务提示加载 verification-loop / verification-before-completion
  • 状态持久化:分类/思维模式/任务类型写入会话事件,resume 可恢复。
  • 降级保护:Plan Mode 不可用时自动限制为只读工具。
  • Hybrid 分类:启发式置信度 + 低置信度时 LLM 结构化分类。
  • Policy EnginebuildPolicyDecision 统一输出完整决策对象(taskType/complexity/risk/executionMode/approval/verification/gitPolicy)。
  • Intent EnginebuildIntent 提取 desired outcome / constraints / acceptance criteria。
  • Context BudgetbuildContextBudget 按复杂度/风险分配上下文预算。
  • Next Best ActiondecideNextAction 作为 Agent Runtime 核心 API,根据状态返回下一步行动。
  • Repository SnapshotbuildRepositorySnapshot 识别包管理器、测试框架、框架、入口点。
  • Project Brain v1/v2/v3buildProjectBrain 聚合仓库快照、符号索引、依赖/测试映射、工程约定;v2 用 SQLite + 轻量 AST + git graph 持久化到 .omni/project-brain.db,并解析真实 import/call/extends/implements 图边;可选安装 web-tree-sitter + tree-sitter-wasms 后自动升级为 Tree-sitter AST 解析(indexProjectBrainV3WithAst / omni_ast_scan);v3 在 SQLite 之上叠加 Hybrid Retrieval 沿真实图双向扩展查询。
  • Mission Planner v1buildMission 把任务组织成 Mission → Phase → Task 骨架,支持动态 Replan。
  • Agent Runtimeomni_mission_run 用真实 subagents 执行 Observe → Think → Act → Replan 循环,带 maxGlobalSteps / maxReplans / maxSameActionRetries / maxRepairs / maxTokens / maxCost / maxToolCalls / maxWallClockMs 预算。
  • 统一主链(优化5 Integration)omni_mission_run 已接入 Task Compiler → Capability Brain → Mission DAG → bindCapabilitiesToDag → DAG Runtime → Evidence,并统一 ExecutionBudget;DAG 失败会保留 failed → Repair → Retry lineage。
  • P1 能力:Planner-generated DAG(可并行分支)、Dynamic Context Expansion(按 uncertainty 自动升级)、Capability Manifest(插件/skill 声明式能力)、Failure-aware DAG mutation(按失败分类生成 repair)。
  • P1/P2 能力:Evidence Store(harness-captured 证据记录,omni_mission_run 自动采集)、Memory v3(skill distillation、execution policies、historical failures、cross-session strategies)。
  • P1/P3 能力:Real capability sandbox(已接入 omni_mission_run,按 role + capability 生成 toolFilter)、Capability Resolver v2(risk/reliability 综合打分)、OmniBench v2 scaffold(manifest schema + runner stub + prepare script + DAG-to-plan compile)。
  • TaskDecisioncreateTaskDecision 生成唯一决策对象,Policy/Runtime 统一消费,避免重复 classify 分叉。
  • Evidence Protocolsrc/evidence.mjs 用结构化 command/file/test/finding 证据判定 PASS/FAIL;isQaPass 已支持从 QA 输出解析 JSON evidence,不再只信文本。
  • Task CompilercompileTask 生成 objective / constraints / non-goals / acceptance / hidden assumptions / ambiguities / invariants / risk / artifacts;compileTaskWithLLM 可用 LLM 增强。
  • Capability Brainsrc/capability-brain.mjs 以 provider-agnostic 方式注册/解析/选择能力,缺失时优雅降级;支持从工具名自动发现能力。
  • Role capability sandbox:QA / Reviewer / Judge 通过 toolFilter 禁止 edit/write/shell,权限来自 Runtime 而非模型自觉。
  • Mission DAGsrc/mission-dag.mjs 用 Task 依赖图代替固定模板,支持插入任务、就绪调度、并行批次、失败时插入 repair。
  • DAG-driven RuntimerunDagLoop 真正按 DAG 执行 ready tasks,支持 maxParallel 并行。
  • Capability-aware DAGbindCapabilitiesToDag 把 Task 的 requiredCapabilities 解析为具体 provider id。
  • Progressive Context Expansionsrc/context-expansion.mjs 按 repo map → symbols → implementations → callers → tests/configs 逐步展开上下文。
  • Hybrid Retrievalsrc/hybrid-retrieval.mjs 组合 lexical + symbol + graph/test 扩展,返回排序候选。
  • Failure Taxonomysrc/failure-taxonomy.mjs 稳定分类失败(test/build/dependency/scope/permission/timeout)并给出恢复路径。
  • Memory v2omni_memory 维护 project/decision/failure/trajectory/learnedSkills;支持 recordLearnedSkill / retrieveLearnedSkill / learnFromTrajectory,自动持久化到 .omni/memory.json
  • Experience-based Routersrc/experience-router.mjs 根据 learned skill 偏置后续路由。
  • Capability performance learningrecordCapabilityOutcome 更新 successRate / lastUsed,能力解析优先历史表现。
  • OES False Completion:评分加入 honesty 维度,惩罚“说 DONE 但实际没完成”。
  • Policy/State Orchestration:用 workflowPolicy 状态机代替大量重复 prompt。
  • Agent 选择selectAgentForTask 推荐 frontend/backend/db/browser/security/review agent。
  • Fable 风格子代理链omni_delegate 可运行 builder → qa-verifier → (repair → qa-verifier)* → code-reviewer → judge,独立验证 + 冷读对抗审查 + 证据交接 + 最终裁决。
  • 技能建议:根据任务类型/文本自动提示加载相关技能(skill 工具),避免重复实现已有能力。
  • 工程原则:轻量注入 Omni 原创的工程原则、Integrity Rules、独立 Verifier/Repair 证据链(methodologyDirectives: false 可关闭)。
  • dsh-routing-suite 兼容:检测到 router-standard 时自动让出 reasoning-mode 路由,避免重复开发;未检测到时保留 Omni 自身轻量路由。
  • Router Benchmark:内置 benchmark/run.mjs(556 个任务),可评估 accuracy / false-direct rate。
  • Engineering Benchmark v1npm run benchmark:engineering 用 OES 评分评估端到端工程交付质量(L1-L10)。
  • Real agent-run comparisonomni_benchmark / omni_benchmark_all 收集 raw/omni 真实运行结果(real-tasks.json 10 个 L1-L10 任务;workspace 没有该文件时自动使用 bundle 内任务集;omni_benchmark_all 支持 maxParallel 并发采集),omni_benchmark_status 查看采集覆盖率,omni_benchmark_import 可导入外部会话结果,npm run benchmark:compare 输出 OES 对比,npm run benchmark:continuous 输出趋势并自动检测回退(支持 --json)。
  • Visual QA 硬编排:前端/UI 任务在 omni_mission_run 的 validate 阶段强制 browser_screenshot + omni_visual_check,未 PASS 则 Replan;可 autoVisualQA: false 关闭。
  • Benchmark 分析benchmark/analyze.mjs 输出混淆矩阵和最高频误判词。
  • LLM 对比评估benchmark/llm-eval.mjs 可对比 Heuristic vs LLM 路由效果。
  • 手动控制/omni 命令 + omni_* 工具。

Orchestrated(编排复用,推荐安装)

插件提供能力Omni Router 如何用
dsh-trio浏览器自动化、MCP、GitHub/GitLab需要浏览器/PR/远程操作时自动可用
dsh-doublecheck交付质量门、验证、返工检测复杂编码任务完成前提示走质量门
superpowers-dshTDD、调试、规划、协作技能编码任务自动提示加载对应技能

> 这些不是强制依赖。装了 → 完整能力;没装 → Omni Router 核心功能依然可用。

Install

One-command via DSH plugin add

dsh plugin --profile web add dsh-omni-router

This installs the bundle and automatically copies the Omni Router preset into ~/.dsh/.agent-presets/omni-router on first load. Restart DSH, then select Omni Router.

Install via npm

npm i dsh-omni-router
cd node_modules/dsh-omni-router
node scripts/install-preset.mjs

One-command install

# Linux / macOS
./install.sh

# Windows PowerShell
./install.ps1

# Cross-platform
node scripts/install-preset.mjs

Then restart DSH and select Omni Router in a new session.

Manual overrides

  • Say “直接做” / “直接执行” → force direct execution.
  • Say “先出方案” / “先设计方案” → force plan-first.
  • /omni status — show current routing state.
  • /omni plan — enter plan-first mode.
  • /omni direct — enter direct mode.
  • /omni mode spec|react|balanced — set thinking mode.
  • /omni reroute plan|direct — adaptively reroute current task.
  • Model tools: omni_status, omni_plan, omni_direct, omni_mode, omni_reroute, omni_delegate, omni_memory, omni_benchmark, omni_mission_run, omni_visual_check.

Configuration

- id: omni-router
  name: ./src/omni-router.mjs
  config:
    requireConfirmation: true
    useLLMClassification: false
    # planFirstKeywords: [自定义, 关键词]
    # directKeywords: [直接跑, 马上改]
    # Visual QA (OpenAI-compatible vision endpoint)
    # autoVisualQA: true
    # visionApiUrl: https://api.openai.com/v1/chat/completions
    # visionApiKey: sk-...
    # visionModel: gpt-4o-mini

See [docs/CONFIGURATION.md](./docs/CONFIGURATION.md) for details.

Repository structure

omni-router/
├── agent.cordis.yml
├── preset.yml
├── src/
│   ├── omni-router.mjs
│   ├── agent-chain.mjs
│   ├── agent-runtime.mjs
│   ├── benchmark-runner.mjs
│   ├── compat.mjs
│   ├── engineering-benchmark.mjs
│   ├── judge.mjs
│   ├── memory.mjs
│   ├── methodology.mjs
│   ├── mission-planner.mjs
│   ├── project-brain.mjs
│   ├── project-brain-v2.mjs
│   └── skill-suggest.mjs
├── test/
├── benchmark/
│   ├── run.mjs
│   ├── engineering-run.mjs
│   ├── real-tasks.json
│   ├── compare.mjs
│   └── tasks.json
├── scripts/install-preset.mjs
├── docs/
├── .github/workflows/ci.yml
├── README.md
├── README.zh-CN.md
└── LICENSE

Test

npm test

Benchmark

# Router accuracy (556 tasks)
npm run benchmark

# Engineering OES demo (sample data)
npm run benchmark:engineering

# Real agent-run OES comparison: Omni vs raw Flash
npm run benchmark:compare

# Show collected OmniBench coverage / missing pairs
npm run benchmark:status
  • benchmark/run.mjs — router accuracy / false-direct / false-plan on benchmark/tasks.json.
  • benchmark/engineering-run.mjs — OES framework on benchmark/engineering-tasks.json.
  • benchmark/compare.mjs — reads benchmark/results/{raw,omni}/*.json and prints OES comparison.
  • Collect real runs with the omni_benchmark tool inside a DSH session.

Current router baseline (556 tasks):

  • complexity accuracy: 73.7%
  • taskType accuracy: 95.9%
  • thinkingMode accuracy: 71.4%
  • false-direct rate: 0.0%
  • false-plan rate: 0.0%

Current OmniBench real baseline (2026-08-21, 10 L1-L10 tasks):

  • raw: 10/10 PASS, avg OES 1.000
  • omni: 10/10 PASS, avg OES 1.000
  • missing pairs: none
  • regression alerts: none
  • Full report: docs/OmniBench-2026-08-21.md

License

MIT