# dsh-model-controller

[中文](../README.md) | English

Install this plugin into a DSH profile and it will pick a model once per user turn: regex keywords first, then a semantic classifier if nothing matches.

The current release requires DeepSeek Harness `0.1.1-rc.2` or a compatible version.

## Default routes

| When                                       | Model               | Reasoning |
| ------------------------------------------ | ------------------- | --------- |
| Complex structure (`v4pm`, `重构`, `设计`) | `deepseek-v4-pro`   | `max`     |
| Review (`review`)                          | `deepseek-v4-pro`   | `high`    |
| Small script (`v4fm`)                      | `deepseek-v4-flash` | `max`     |
| Small method (`v4fh`)                      | `deepseek-v4-flash` | `high`    |
| Active Plan Mode                           | `deepseek-v4-pro`   | `max`     |
| Semantic classifier (fallback)             | `deepseek-v4-flash` | `off`     |

Unmatched requests keep the Agent's original model. The route lasts for the whole turn.

## Settings

Open **Settings → Plugins → Plugin configuration**. Saved changes apply from the next user turn.

- **Global matching mode** — Mixed (regex, then semantic), Semantic only, or Regex only.
- **Route rules** — Add, edit, reorder, or delete routes; restore the built-in table.
- **Semantic classifier** — Model used only when mixed/semantic mode needs a classification step.
- **Plan Mode** — Force the dedicated Pro/max model while Plan Mode is active.
- **Routing metadata** — Show provider, model, effort, and route id at the start of reasoning.

## Install

```sh
dsh plugin --profile web add dsh-model-controller@next
dsh web
```

From this repository, build first and install the checkout:

```sh
pnpm install
pnpm run build
dsh plugin --profile web add .
```

Remove it with `dsh plugin --profile web remove dsh-model-controller`.