Preface¶
The plugin ecosystem of DeepSeek Harness provides Tool, Skill, and Plugin capabilities for Agents to use; however, in real-world tasks, executing an action does not necessarily mean the result is reliable. dsh-omni-router complements this layer: it is positioned as “Omni - an Agent Reliability Kernel for DeepSeek Harness,” dividing task processing into three steps: Decide, Prepare, and Verify.
DSH continues to handle execution, while Omni handles reliability. Below is an introduction to its positioning, verified features, installation commands, and typical usage.
What This Is¶
qwe225380/dsh-omni-router is maintained by qwe225380, with the package name dsh-omni-router. The current package.json version is 3.0.0-rc.27, and it is licensed under MIT.
Its one-line positioning is:
Omni - an Agent Reliability Kernel for DeepSeek Harness. Decide. Prepare. Verify: DSH owns execution, Omni owns reliability.
In other words, it does not replace DSH; rather, it adds task-level judgment, context preparation, capability auditing, evidence verification, and failure recovery on top of DSH’s existing execution capabilities.
Core Features¶
Task Interception¶
Simple tasks are truly NOOP; only complex tasks enter Assist / Guard mode. This prevents all tasks from going through complex workflows.
Task Contract and Context¶
It uses Task Contract + Precise Context + Capability Sufficiency. Tasks first establish acceptance criteria and capability boundaries before entering execution.
Evidence and Recovery¶
It focuses on Evidence / Freshness / Acceptance coverage / Recovery, treating acceptance, evidence freshness, acceptance coverage, and failure recovery as checkpoints before and after task completion.
Capability Reuse¶
The design principle is to prioritize reusing existing DSH Tool / Skill / Plugin capabilities rather than reinventing the wheel; capability gaps are only filled when they genuinely block tasks.
Failure Handling¶
When retries are ineffective, the strategy shifts to repair, expand context, or change hypotheses, rather than repeating the same attempt.
Capability Gap Installation¶
The default is recommend mode: identify gaps, search, recommend, and install only after user/Host approval; auto-installation only occurs when auto-trusted is explicitly enabled.
Combination with Router Standard¶
When Router Standard is detected, it automatically generates combined presets and delegates persona/reasoning routing.
Benchmarking¶
It provides the OmniBench multi-arm benchmark and release gates for comparison and release threshold checks.
Installation and Enablement¶
The installation command is as follows:
dsh plugin --profile web add dsh-omni-router
This command adds dsh-omni-router to the current DSH web profile.
Notes before use:
- The current package.json version is
3.0.0-rc.27, which is in RC status. - It is licensed under MIT.
- The plugin runs with the current dsh process permissions; before installation, review the source code, configuration options, and license before deciding whether to enable it in important environments.
Typical Usage¶
The following examples come from plugin documentation and can be understood by task type.
Fixing Concurrent Refresh Issues¶
Scenario: Fixing an occasional concurrent refresh issue where refreshSession is called three times.
Workflow:
- Generate a Task Contract, including C1-C3.
- Organize Relevant Context.
- DSH executes.
- Provide evidence: concurrency test 28/28, full auth suite 143/143.
- Output Proof: 3/3.
This example demonstrates the combination of Task Contract, execution evidence, and Proof.
Frontend UI Fix¶
Workflow:
- Determine this is a frontend task.
- Check current DSH capabilities.
- Reuse Browser and UI/Verification Skill.
- Generate UI acceptance criteria.
- DSH executes.
- Collect screenshot/tool/test evidence.
- Verify.
When Browser and UI/Verification Skill already exist, similar capabilities are not reinstalled.
Complex Refactoring¶
All existing capabilities are reused; only gaps are filled as needed.
Combination with dsh-router-standard¶
Router Standard continues to handle DeepSeek persona/attention/workflow routing; Omni handles capability, context, evidence, and reliability.
Running OmniBench¶
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>
These commands are used to generate the manifest, run the benchmark, generate the matrix, and check gates, respectively.
Applicable Scenarios and Notes¶
Suitable for scenarios where you need to execute tasks on DSH while also focusing on task contracts, capability reuse, evidence, and failure recovery. Especially relevant when tasks involve acceptance criteria, tool evidence, or cross-step recovery.
Notes:
- In default recommend mode, installing gap capabilities requires user/Host approval; auto-installation only occurs when auto-trusted is explicitly enabled, and
.omni/capability-audit.jsonis recorded. - If DSH currently lacks reliable event hooks, Omni will output Verified, Partially Verified, or Unverified.
- 3.0 RC benchmark validation is in progress; frontier parity is not claimed until gates are passed.
- The current version
3.0.0-rc.27is an RC version; for production environments, it is recommended to verify behavior and permissions first.
Conclusion¶
The value of dsh-omni-router lies in separating “execution” from “reliability”: DSH provides execution and plugin capabilities, while Omni adds the task-level constraints of Decide, Prepare, and Verify. It does not aim to replace DSH but rather to reuse existing Tool, Skill, and Plugin capabilities as much as possible and provide clear status when evidence is insufficient.
GitHub: https://github.com/qwe225380/dsh-omni-router
Directory page (provided via plugin leads, not found in previously captured materials; please verify before use): https://www.skillhub.cn/plugins/qwe225380/dsh-omni-router