dsh-routing-suite: model routing DSH plugin for DeepSeek

dsh-routing-suite is a class of community DSH plugin that adds routing to DeepSeek Harness: it dispatches requests to different models or presets by task, cost or reasoning mode (source). This article covers what it solves, how it plugs into DeepSeek Harness, and the trade-off between cutting cost and keeping quality.

What the DSH plugin dsh-routing-suite is: a dispatch layer in front of your requests

The core job of a routing plugin is to insert a decision layer between "request" and "model", so not every task goes to the same model. It targets the waste that comes from a one-size-fits-all setup:

1. Dispatch by task — light tasks go to a lightweight model or a lightweight reasoning mode, heavy ones to a strong model. Expected: heavy work is not skimped on, light work is not expensive.
2. Dispatch by cost — set cost thresholds per model and steer cost-insensitive tasks to the more economical provider. Expected: overall spend goes down.
3. Dispatch by capability — tasks needing long context, images or a specific tool call go to a model that has that capability. Expected: capability matches, fewer mistakes.

There is more than one implementation named dsh-routing-suite, and the routing dimensions differ slightly: some lean towards a "lightweight smart routing mode" that automatically picks between check-first and direct execution based on the first task; others are more of a "routing suite" bundling a runtime injector, a task-aware thinking-mode routing preset and a tiered task protocol (source). What they share: all of them sit on top of the official everything-is-a-plugin model integration and add one more layer of on-demand dispatch.

How it plugs into DeepSeek Harness

Routing is a layer on top of the official model configuration: configure providers first, then let the routing plugin choose which one to use. The official docs state that models are managed under Settings → Models, supporting built-in third-party providers (such as anthropic, openai and others) as well as custom model APIs (supply a Provider ID, base URL, protocol, credentials and model list); changes take effect on the next request, with no service restart (source). Plugging in a routing plugin takes four steps:

1. Configure the model providers — add every model you will use under Settings → Models and fill in its key. Expected: each provider works on its own and can hold a conversation directly. The details of this step are in configure multiple model providers.
2. Install the routing plugin — open DSH Plugin Hub and search with keywords such as routing or model scheduling, or run the command listed on the entry. Expected: the plugin joins the plugin tree of the web profile.
3. Write the routing rules — specify the target model for each class of request, by task type / cost / capability. Expected: the rules cover your main task scenarios.
4. Set a fallback model — name the one to use when no rule matches. Expected: the chain survives a missed rule.
5. Validate on a small sample — run a few real task types and compare results against spend. Expected: you can tell whether the tiers are sensible.

Note that the routing plugin and the official model configuration are additive, not alternative: with no providers configured, routing has nothing to dispatch to; with no routing plugin installed, the official configuration still works.

DSH plugin fit and cost: do the savings and the quality gains hold up

Whether "cheaper" and "better" hold at the same time depends on your task mix and how accurate the rules are — the plugin makes no such guarantee. Do the arithmetic before deciding:

The trade-offs in detail:

1. Savings are conditional — money is only saved when lightweight tasks are the majority and the routing judgment is reliable; if the rules misjudge a hard task as simple, the savings come back as a quality loss.
2. The configuration complexity is real — you maintain several providers, a target model per task class, and a failure fallback; the more models and the finer the rules, the larger the maintenance surface.
3. A fallback is mandatory — the chain becomes "request → routing → model", and a mistake anywhere can surface as an off-target answer or an error; without a fallback model, a missed rule takes the whole chain down.

DSH plugin notes

These are the DSH plugin points to settle before wiring routing into your config. Item by item:

1. Configure providers before adding routing: routing builds on the official model configuration — do not reverse the order.
2. Rules must be verifiable: run a real-task comparison before launch; do not set tiers by feel.
3. Do not skip the fallback: always set a default model so a missed rule does not mean total failure.
4. Remember the middle layer when debugging: besides checking the model, also check the routing rules themselves.
5. Match the versions: for community plugins, verify the compatible core version and maintenance status; the sources are in configuration and configuration layers.

Sources: Configure models (official docs), dsh-routing-suite (dragonbaba), dsh-routing-suite (yjh051108), dsh-routing-suite (npm)