DeepSeek Harness Plugin Hub

Publish and manage complete Harness Profiles. Discover Plugins for your next setup.

Explore

PluginsPresetsDocsNews

Community

Publish a pluginContactReport an issue

Resources

Plugin Hub on GitHubDeepSeek HarnessSystem statusPrivacy notice
© 2026 DeepSeek Harness Plugin HubPowered byPaxTech

Independent and unofficial. Not affiliated with, authorized by, or endorsed by DeepSeek.

Omni Router — DSH Plugin for DeepSeek Harness
← Plugins
O

dsh-omni-router

Omni Router

Omni - an Agent Reliability Kernel for DeepSeek Harness. Decide. Prepare. Verify: DSH owns execution, Omni owns reliability.

The plugin will be installed here. Keep web if you are unsure.

npx -y @deepseek-ai/dsh plugin --profile web add github:qwe225380/dsh-omni-router#e35054e29225072afaad693b2245316de1cc0818
READMECompatibilityVersions

Compatibility and provenance

Omni Router is published as dsh-omni-router and currently resolves to version 3.0.0-rc.29. The Hub verifies its manifest and preserves the exact installation source for reproducible installs.

DSH compatibility
*
Runtime surfaces
any
Release source
github
Registry updated
8/31/2026

Versions

3.0.0-rc.29prerelease
8/31/2026
2.4.0stable
8/24/2026
1.53.0stable
8/21/2026
1.52.0stable
8/21/2026
Show 26 more versionsCollapse versions
3.0.0-rc.27prerelease
8/30/2026
3.0.0-rc.25prerelease
8/30/2026
3.0.0-rc.24prerelease
8/30/2026
3.0.0-rc.23prerelease
8/30/2026
3.0.0-rc.22prerelease
8/30/2026
3.0.0-rc.20prerelease
8/30/2026
3.0.0-rc.19prerelease
8/29/2026
3.0.0-rc.18prerelease
8/29/2026
3.0.0-rc.17prerelease
8/29/2026
3.0.0-rc.14prerelease
8/29/2026
3.0.0-rc.13prerelease
8/26/2026
3.0.0-rc.12prerelease
8/26/2026
3.0.0-rc.11prerelease
8/26/2026
3.0.0-rc.10prerelease
8/26/2026
3.0.0-rc.8prerelease
8/25/2026
3.0.0-rc.7prerelease
8/25/2026
3.0.0-rc.5prerelease
8/25/2026
3.0.0-rc.3prerelease
8/25/2026
3.0.0-rc.2prerelease
8/24/2026
3.0.0-rc.1prerelease
8/24/2026
1.45.0stable
8/21/2026
1.44.0stable
8/21/2026
1.40.0stable
8/21/2026
1.30.0stable
8/21/2026
1.25.0stable
8/21/2026
1.23.0stable
8/20/2026

Related plugins

Loading related plugins…

Latest
3.0.0-rc.29
DSH
*
HMR
Process restart
Tree shaking
Safe tree shaking not declared
Unpacked size
Unavailable
Files
Unavailable
Surface
any
License
MIT
Source
github
GitHub
★ 4
Weekly downloads
0
Last push
8/31/2026
View source ↗
README badge

Click the badge to copy Markdown for your README.

Do you maintain this Plugin?Claim benefit · Priority security scan

Verify the GitHub repository declared in package.json to manage this listing. After you claim it, Hub will prioritize a security scan of the current version and publish the result when it passes.

Claim this Plugin →
Report an issue
DeepSeek Harness Plugin Hub
ProfilesPluginsCategoriesNewsDocsSign inManage Profiles
ProfilesPluginsCategoriesNewsDocsSign in

Related plugins

More verified plugins in agents-orchestration.

Headless@deepseek-ai/dsh-headlessThe dsh one-shot bundle: a direct core Agent/Session runner over dsh-base with no Host, HTTP, or browser layerExperimental Agent Team Web Profile@deepseek-ai/dsh-experimental-agent-team-web-profileExperimental Web profile layer for Agent Teams Remote and UI pluginsSubagent Codex@deepseek-ai/dsh-subagent-codexOne-shot Codex subagent provider over the official app-server protocolSubagent Claude Code@deepseek-ai/dsh-subagent-claude-codeOne-shot Claude Code subagent provider over the official Agent SDK

README

Omni

English

Omni 给 DeepSeek Harness 增加一层任务级可靠性: 判断何时介入、准备正确上下文、约束完成条件、验证真实证据,并在失败时改变策略。 它优先复用 DSH 已有的 Skills、Tools 和 Plugins,而不是重新实现它们。

DSH 给 Agent 能力,Omni 让结果可靠。

默认快。必要时聪明。完成必须有证据。

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

Why Omni

Coding Agent 最常见的六个问题:

  • 找错代码和上下文
  • 复杂任务做着做着跑偏
  • 忘掉验收条件
  • 测试没有真正跑
  • 失败后重复同一种办法
  • 最后说“完成了”,实际并没有完成

Omni 就是为了解决这些工程问题,而不是再做一个“Router”。


What Omni does

核心价值Omni 实际做什么
🧠 Intervention Intelligence简单任务真正 NOOP,复杂任务才进入 Assist / Guard
🎯 Context & Task IntelligenceTask Contract + 精准 Context + Capability Sufficiency
✅ Proof & RecoveryEvidence / Freshness / Acceptance coverage / Recovery
♻️ Reuse first设计原则:优先复用 DSH 已有 Tool / Skill / Plugin,不重复造轮子
🧩 Capability Gap Management只在确实阻塞任务时补齐缺口
🔄 失败纪律retry 不奏效时 repair / expand context / change hypothesis

30 秒 Demo

Without Omni

User:
修复 refreshSession 偶发并发刷新三次的问题

Agent:
修改 session.ts
运行了一些检查
Done ✅

问题:

  • 真实并发测试没跑
  • 调用方漏了一个
  • 没有 regression test
  • 没有证据证明“真的完成了”

With Omni

Task Contract
C1 one refresh request per session
C2 concurrent callers behave correctly
C3 regression test passes

Relevant Context
session.ts
middleware/auth.ts
websocket/reconnect.ts
session-concurrency.test.ts

DSH executes...

Evidence
✓ concurrency test: 28/28
✓ full auth suite: 143/143
✓ relevant files verified

Proof: 3/3

一个插件,复用整个 DSH 生态

这不是 Omni 的核心,而是它的生态策略:Reuse first. Don't reimplement host capabilities.

                    Omni
          Capability + Reliability
                     │
      ┌──────────────┼──────────────┐
      │              │              │
      ▼              ▼              ▼
    Skills         Plugins         Tools
      │              │              │
      ├── TDD        ├── Browser    ├── Shell
      ├── Git        ├── Router     ├── Editor
      └── Verify     └── Others     └── Tests
      │              │              │
      └──────────────┼──────────────┘
                     ▼
              DeepSeek Harness
                     │
                     ▼
                  Execute
                     │
                     ▼
          Evidence → Verify

Omni 不追求自己拥有最多能力,而是尽可能让已经存在的优秀能力发挥作用。

示例:前端 UI 修复

用户:
帮我检查这个网页的实现,并修复 UI 问题。

Omni:
1. 判断这是前端任务
2. 检查当前 DSH 能力
3. 已有 Browser → 直接复用
4. 已有 UI/Verification Skill → 继续由 DSH 使用
5. 不再安装重复能力
6. 生成 UI acceptance criteria
7. DSH 执行
8. 收集 screenshot / tool / test evidence
9. Verify

示例:复杂重构

当前环境:
✓ Git
✓ Testing
✓ Skills
✗ 某个任务要求的特殊能力

Omni:
已有能力 → 全部复用
缺失能力 → 只补这个 gap

Under the hood

Omni 的用户界面很简单,但内部并不是一个 Prompt 模板。

Task
 │
 ▼
Task Classification
 │
 ▼
Intervention Gate ───────────→ NOOP
 │
 ▼
Task Contract
 ├── Acceptance Criteria
 ├── Risk
 ├── Context Budget
 └── Capability Requirements
 │
 ▼
Context Engine
 ├── multilingual tokenizer
 ├── lexical retrieval
 ├── symbol retrieval
 ├── dependency graph
 └── freshness tracking
 │
 ▼
Capability Audit
 ├── existing tools
 ├── native skills
 ├── installed plugins
 └── missing capabilities
 │
 ▼
DeepSeek Harness
 │
 ▼
DSH Events
 │
 ▼
Evidence Trust
 ├── T0 model claim
 ├── T1 observation
 ├── T2 host/tool
 ├── T3 deterministic execution
 └── T4 independent verifier
 │
 ▼
Criterion Verification
 ├── PASS → Verified
 └── FAIL → Recovery
              │
              └──→ DSH

Reliability Engineering

Omni 不只是功能堆叠,它专门处理这些 Agent failure modes:

✓ 模型自己声称测试通过
  → model evidence 只能 T0/T1

✓ Tool 执行结果和模型文字混淆
  → provenance-aware Evidence

✓ 文件修改后旧测试结果仍被当有效
  → Evidence freshness / invalidation

✓ permission / dependency / timeout 被误认为成功
  → success whitelist + failure taxonomy

✓ 中文任务检索不到英文 symbol
  → multilingual query tokenizer + semantic expansion

✓ 相同失败无限 retry
  → failure-aware recovery

✓ 简单任务被复杂框架拖慢
  → true L0 NOOP

✓ 已经有插件还重复造能力
  → capability reuse first

✓ 缺能力时用户需要手动安装
  → approval-gated provisioning (recommend by default, auto-trusted only on explicit opt-in)

✓ 两个 Router 同时控制模型
  → compatibility delegation

Omni 不只是 Router

普通 Router
Task → choose mode/model/agent

Omni
Task
→ 判断是否需要介入
→ 复用已有能力
→ 补齐缺失能力
→ 准备正确 Context
→ 定义 Acceptance
→ Host 执行
→ 收集真实 Evidence
→ 验证完成
→ 失败后恢复

Routing is one decision. Reliability is the whole lifecycle.


Works better with good plugins

Omni 不要求生态围绕自己重写。

例如安装 dsh-router-standard 后,Omni 可以自动生成组合预设:

Omni Router + Router Standard

Router Standard 继续负责 DeepSeek persona / attention / workflow routing;Omni 负责 capability、context、evidence 和 reliability。


Built like infrastructure, not a prompt pack

Omni 对 routing、Context、Evidence、Capability、Recovery、Host Adapter、Installer 和 OmniBench 均有独立测试覆盖(约 60 个测试文件)。

每个 RC 都要求完整 npm test 回归;Benchmark 支持 multi-arm(raw / omni / mid / frontier / stack / stack_omni…)。

Measured Reliability

Evidence Federation  第三方结果 → EvidenceRecord v1(provider 自报 trust 无效)
Context Metrics      Recall@5/10 · Precision · Token 开销 · irrelevant ratio
Intervention KPIs    NOOP Precision · missed/false intervention rate
Recovery Telemetry   failure fingerprint · retry/repair/hypothesis funnel · success rate
Multi-arm Benchmark  comparePair(baseline, candidate) + release gates

Proof of Completion

每个验收标准都有 id(如 C1、C2)。每个标准必须至少有一条 fresh evidence 记录并满足要求的信任等级:

TrustSource
T0model claim
T1agent observation / model-written JSON
T2host/tool output
T3deterministic command/test execution
T4independent / hidden verifier
Completed.
✓ C1 根因已修复
✓ C2 已添加回归测试
✓ C3 143 个测试通过
Proof: 3/3

如果 DSH 当前没有可靠事件 hook,Omni 会诚实地输出:

Verified
Partially Verified
Unverified

Compatibility

Project / CapabilityOmni strategy
dsh-router-standardDelegate persona/reasoning routing
DSH native SkillsReuse automatic activation
Browser pluginReuse browser capability
Testing/verification SkillsReuse evidence/results
Git pluginsReuse native workflow
MCP serversTreat as host capabilities
Missing capabilityProvision only when necessary

Evolution

1.x
Router
↓
2.x
Planner / Context / Capability / Evidence exploration
↓
2.4
Reliability Kernel
↓
3.0
Convergence:
Decide · Prepare · Verify

3.0 intentionally removed or hid overlapping runtime/router features instead of continuing feature growth.


OmniBench

3.0 RC benchmark validation is in progress.

Target:
≥50 repos
≥100 tasks
≥3 paired runs/task
hidden verifier 100%
Medium/Hard uplift ≥10pp
False completion <3%
Cost ≤2.5×

在数据达到这些 Gate 之前,不会宣称 frontier parity。

npm run omnibench:v2:generate -- benchmark/omnibench-v2/manifest.local.example.json
node benchmark/omnibench-v2/run.mjs benchmark/omnibench-v2/manifest.local.example.json --exec
node benchmark/omnibench-v2/matrix.mjs <results.json>
node benchmark/omnibench-v2/gates.mjs <results.json>

Architecture

kernel/          task-contract · intervention-gate · omni-event · evidence-trust
context/         query-tokenizer · capsule · freshness
recovery/        failure-taxonomy · strategy-shift · recovery-policy
capability/      auditor · quality · solver · provisioner · performance
mission/         mission-ir · planner-dag · mission-dag(兼容层)
host/            interface · dsh-adapter
benchmark/       omnibench-v2 runner · matrix · gates

FAQ

Q: 它会拖慢简单任务吗?

A: 不会。L0 是真正的 NOOP,Omni 不注入任何 prompt。

Q: 它会替换 DeepSeek Harness 吗?

A: 不会。DSH 仍然拥有执行、工具、Skills、插件、Runtime。

Q: 它会接管我的插件吗?

A: 不会。Omni 只定义“任务达到什么状态才算成功”,不调度具体插件。

Q: 它能自动安装技能/插件吗?

A: 能。默认 recommend 模式:识别缺口 → 搜索 → 推荐 → 用户/Host 审批后才安装;只有显式开启 auto-trusted 才自动安装,并记录 .omni/capability-audit.json。

Q: 它和 Router Standard 冲突吗?

A: 不冲突。检测到 Router Standard 时,Omni 自动生成组合预设,并让出 reasoning/persona routing。


Credits / Inspirations

Omni 的设计受到 DeepSeek Harness 社区多个项目的启发,尤其是:

  • dsh-router-standard / dsh-routing-suite
    • persona routing
    • attention engineering
    • progressive disclosure
    • real Harness assembly-chain research

Omni 不替代这些能力,而是让它们在同一任务可靠性闭环中协同工作。


License

MIT