dsh-routing-suite
Run the following command in DeepSeek Harness:
dsh plugin install yjh051108/dsh-routing-suite
Paste the following prompt into your AI chat to install this plugin:
Run dsh plugin install yjh051108/dsh-routing-suite inside DeepSeek Harness to install the plugin; full source repository: https://github.com/yjh051108/dsh-routing-suite
About this plugin
In long-horizon reasoning runs, the most visible failure mode is not the ceiling of model capability but the decay of diligence: verification steps get skipped, tool calls degrade, and chart-reading simply vanishes in the latter half of a session. dsh-routing-suite is a three-in-one package built to address exactly this problem, bundling a runtime hot-injection layer, task-aware thinking-mode routing, and a graded session protocol with red-team gates into a single repository so that DSH pins every request to the correct reasoning track without a process restart.
The three components complement each other. The injector ships a full set of dev_* tools for hot reload, side-mount promotion, unmount, and self-healing routing, eliminating the need to restart the runtime. The routing preset, validated across P1-P23 experimental rounds, introduces a three-band behavior model with weak-band internal routing that selects spec, react, or neutral personas per model, achieving 96 percent routing accuracy and 100 percent convergence on open-ended task completion. The graded mode decomposes a session into a two-level protocol spanning brainstorm alignment, specification planning, specification-driven injection, checkpoint execution, group close-out, and final verification, complete with six tools, three modes, fine-grained mode control, a red-team gate, audit endpoints, and a sanitised measurement script anyone can run against their own session JSONL to reproduce diligence metrics.
It is aimed at developers and researchers who use DSH for long-horizon inference, multi-turn tool orchestration, or structured task pipelines. If you are tired of first-turn routing silently falling back to the weak band, of paying for an extra API call every round, or of watching model diligence drift as sessions stretch, this suite speaks with measured data, is licensed under MIT, and is ready with a single install command.
Editor's note
Written for AI Agent Hub. This is separate from the publisher description above.
What problem it solves
On a long reasoning run the failure is often skipped checks and sloppy tool use, not the ceiling of one model. This routing suite lets you send steps to different models, for example a cheaper draft and a stronger check. The problem is diligence across a long trace, and the side effect is a more complicated bill.
Who it is for
Harness users who already know which model should do which step and who will test on a public sample first. You can read a routing config.
Who should skip it
Skip it if you have not pinned a single model for a week. The first week is one model on purpose. Skip it if you wanted image input. That is ModLens, a different inference plugin. Skip it if you cannot say which host receives the code. Routing is how a private repo ends up on an endpoint you did not budget. The inference category intro tells you to treat the table as production config.
First three steps
- Write a two-step route on paper: who drafts, who checks. Include the model names.
- Run it on a public snippet with a deliberate mistake the checker should catch.
- Read the bill or the logs and confirm both models were the ones you named. If a third model appeared, stop and fix the config before any client code is in the prompt.
Alternatives on this site
A single reasoning model may be simpler if you only needed more thinking, not a committee. A workflow plugin checks a diff after the fact. It does not choose the model. Use the simpler one until the long run actually decays.
Questions
Will this raise quality automatically? No. It raises the number of calls. Quality depends on the route you wrote.
Can I leave it on for every chat? Only after the public-sample test. Default-on routing is how short questions get expensive.
Use Cases
- Sustain model diligence across long multi-turn reasoning sessions via enforced routing and protocol constraints
- Hot-swap DSH plugins and routing presets without restart to rapidly iterate and validate inference strategies
- Quantitatively audit session-level task quality with red-team gates, compliance checks, and sanitised metric reproduction
Best For
- Developers building long-horizon reasoning or multi-turn tool orchestration on DSH
- Researchers who need to monitor and quantify model inference diligence in real time
- Engineering teams requiring auditable and reproducible session-level task protocols
Related Plugins
Install in dsh with no login or API key for unlimited free access to frontier models such as DeepSeek V4.1 Flash and Kimi K3, plus a local OpenAI-compatible gateway.
ModLens is a vision plugin for DeepSeek Harness that gives text-only models sight by reading images pasted directly into chat, with zero-config setup and multiple vision engines.
Improves token efficiency in DSH via low-intrusion OPS tools, selective hiding of official tools, and on-demand dynamic loading for Native and PTC modes.
On-demand vision for text-only DeepSeek Harness agents: built-in free keyless vision chain and 14 vision tools, routing image turns as tool calls to vision models with pixel fidelity, no Python needed, one-command install.
Reader reviews
No approved reader reviews on this listing yet. Yours can be the first.