dsh-deepseek-model-router

Readme

# dsh-deepseek-model-router

**Auto-switch DeepSeek models** for [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness). No manual model picking: every model request is routed automatically, and a `switch_model` tool gives you a manual override when you want one.

## How it works

Before every model request, the plugin inspects the conversation and replaces the frozen provider/model:

| Situation | Model used (defaults) |
| --- | --- |
| Conversation contains images | `deepseek-v4-flash-vision-exp` (vision) |
| Complex task — long text, code, math, analysis keywords (score ≥ threshold) | `deepseek-v4-pro` |
| Everything else — chat, short questions, routine work | `deepseek-v4-flash` |

It hooks the same `agent/request` waterfall the built-in model selection uses, installed after it, so the routed model is the final one for the call. The prompt variables (`provider`/`model`) stay in sync with what actually runs.

### Manual override tool

`switch_model(mode, reason?)` — `mode` is one of:

- `auto` — back to automatic routing
- `fast` — force `deepseek-v4-flash`
- `pro` — force `deepseek-v4-pro`
- `vision` — force `deepseek-v4-flash-vision-exp`

The override is per-session and takes effect from the next request.

## Install

From the plugin market (once published), or locally:

```sh
dsh plugin --profile web add dsh-deepseek-model-router
```

The bundle patch activates the router for the whole profile after a restart. Configuration lives in the bundle patch (`cordis.patch.yml` → row `config`): provider, fast/pro/vision models, complexity threshold, tool name, enabled.

## Configuration

All fields optional (defaults shown):

```yaml
config:
  provider: deepseek-official   # only sessions on this provider are routed
  fastModel: deepseek-v4-flash
  proModel: deepseek-v4-pro
  visionModel: deepseek-v4-flash-vision-exp
  complexityThreshold: 2       # task score >= threshold -> pro
  toolName: switch_model       # registered tool name
  enabled: true
```

## Notes

- Only `deepseek-official` sessions are routed; other providers pass through untouched.
- If images are present, the router always prefers an image-capable model (`visionModel`); if that model is not image-capable in your catalog, the current model is kept.
- The composer's model seat still shows the session's default selection; the actual model used per call is decided by the router and recorded in the session's request header. Use `switch_model` to pin a model explicitly.

## License

MIT