
Security News
Deno Goes the Way of the Dinosaur
Cloudflare is acquiring Deno’s team, which will provide another year of bug fixes and security updates before ending support for the runtime.
dsh-omni-router
Advanced tools
Omni - a reliability and orchestration control plane for DeepSeek Harness, designed to turn fast models into dependable long-horizon software-engineering agents.
Fast by default. Smart when needed. Proven when done. Decide · Compose · Verify Install it. Keep coding.
Omni 是一个 DeepSeek Harness 的 reliability and orchestration control plane,定位为 Intelligence Runtime:把复杂度路由、任务编译、项目上下文、能力编排、Mission DAG、Runtime 预算、证据验证、Memory/Learning、Benchmark 组合成一个统一的、自动运行的工程闭环。
单独使用 dsh-routing-suite、dsh-trio、dsh-doublecheck、superpowers-dsh 时,你需要:
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
direct / balanced / plan 三态路由,简单任务直接做,复杂/模糊任务先方案。spec(方案优先)/ react(直接执行)/ balanced(自动)。bugfix / feature / refactor / test / review / other。omni_reroute / /omni reroute 在 direct ↔ plan 间动态切换。test-driven-development / red-green-tdd,不重复实现 TDD。verification-before-completion / delivery-proof,不重复实现质量门。using-git-worktrees / git-discipline。executing-plans / writing-plans。verification-loop / verification-before-completion。buildPolicyDecision 统一输出完整决策对象(taskType/complexity/risk/executionMode/approval/verification/gitPolicy)。buildIntent 提取 desired outcome / constraints / acceptance criteria。buildContextBudget 按复杂度/风险分配上下文预算。decideNextAction 作为 Agent Runtime 核心 API,根据状态返回下一步行动。buildRepositorySnapshot 识别包管理器、测试框架、框架、入口点。buildProjectBrain 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。buildMission 把任务组织成 Mission → Phase → Task 骨架,支持动态 Replan。omni_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。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。omni_mission_run 自动采集)、Memory v3(skill distillation、execution policies、historical failures、cross-session strategies)。src/memory-engine.mjs 统一 Memory v1/v2/v3,主 Runtime 已切换为单一 Memory Engine API。src/evidence-engine.mjs 统一证据协议与存储;DAG scheduler 通过 selectReadyBatch 避免 write overlap 并行冲突。compileDagToWorkflow 将 Mission DAG 编译为 DSH-native workflow outline。omni_mission_run,按 role + capability 生成 toolFilter)、Capability Resolver v2(risk/reliability 综合打分)、OmniBench v2 runner(manifest schema + prepare + prompt generation + --exec 本地执行/结果采集 + DAG-to-plan compile)。omni_mission_resume 跨 session 继续 DAG,runDagLoop 支持 onProgress 周期保存)、Harness evidence extraction(extractHarnessEvidence 从 commands/tests/toolCalls/EVIDENCE_JSON 提取机器可验证证据)、context token budget(maxContextTokens 真正限制检索输出)、semantic failure retrieval + evidence-backed skill distillation、OmniBench v2 real runner。omni_explain / omni_doctor 用户入口。createTaskDecision 生成唯一决策对象,Policy/Runtime 统一消费,避免重复 classify 分叉。src/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。compileTask 生成 objective / constraints / non-goals / acceptance / hidden assumptions / ambiguities / invariants / risk / artifacts;compileTaskWithLLM 可用 LLM 增强。src/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。toolFilter 禁止 edit/write/shell,权限来自 Runtime 而非模型自觉。src/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。runDagLoop 真正按 DAG 执行 ready tasks,支持 maxParallel 并行。bindCapabilitiesToDag 把 Task 的 requiredCapabilities 解析为具体 provider id。src/context-expansion.mjs 按 repo map → symbols → implementations → callers → tests/configs 逐步展开上下文。src/hybrid-retrieval.mjs 组合 lexical + symbol + graph/test 扩展,返回排序候选。src/failure-taxonomy.mjs 稳定分类失败(test/build/dependency/scope/permission/timeout)并给出恢复路径。omni_memory 维护 project/decision/failure/trajectory/learnedSkills;支持 recordLearnedSkill / retrieveLearnedSkill / learnFromTrajectory,自动持久化到 .omni/memory.json。src/experience-router.mjs 根据 learned skill 偏置后续路由。recordCapabilityOutcome 更新 successRate / lastUsed,能力解析优先历史表现。workflowPolicy 状态机代替大量重复 prompt。selectAgentForTask 推荐 frontend/backend/db/browser/security/review agent。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。skill 工具),避免重复实现已有能力。methodologyDirectives: false 可关闭)。benchmark/run.mjs(556 个任务),可评估 accuracy / false-direct rate。npm run benchmark:engineering 用 OES 评分评估端到端工程交付质量(L1-L10)。omni_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)。omni_mission_run 的 validate 阶段强制 browser_screenshot + omni_visual_check,未 PASS 则 Replan;可 autoVisualQA: false 关闭。benchmark/analyze.mjs 输出混淆矩阵和最高频误判词。benchmark/llm-eval.mjs 可对比 Heuristic vs LLM 路由效果。/omni 命令 + omni_* 工具。| 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-dsh | TDD、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。
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.
npm i dsh-omni-router
cd node_modules/dsh-omni-router
node scripts/install-preset.mjs
# 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.
/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.omni_status, omni_explain, omni_doctor, omni_plan, omni_direct, omni_mode, omni_reroute, omni_delegate, omni_memory, omni_benchmark, omni_mission_run, omni_mission_resume, omni_capability_audit, omni_capability_provision, omni_capability_probe, omni_capability_performance, omni_visual_check.- 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 for details.
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
npm test
# 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.omni_benchmark tool inside a DSH session.Current router baseline (556 tasks):
Current OmniBench real baseline (2026-08-21, 10 L1-L10 tasks):
docs/OmniBench-2026-08-21.mdMIT
FAQs
Omni - a reliability and orchestration control plane for DeepSeek Harness, designed to turn fast models into dependable long-horizon software-engineering agents.
We found that dsh-omni-router demonstrated a healthy version release cadence and project activity because the last version was released less than a year ago. It has 1 open source maintainer collaborating on the project.

Security News
Cloudflare is acquiring Deno’s team, which will provide another year of bug fixes and security updates before ending support for the runtime.

Security News
Anthropic’s free OSS Scanner uses Mythos and other Claude models to find open source vulnerabilities and send maintainers reports with proposed fixes.

Security News
On Front Lines’ Early Adopters: CISO podcast, Andrew Becherer shares lessons from Datadog, weighs the long-term costs of AI-built tools, and explains how founders can actually gain a CISO’s trust.