dsh-smart-router · Smart model routing for DeepSeek Harness

English · 简体中文

  

Why this exists

In August 2026, DeepSeek raised prices across the V4 family and introduced peak/off-peak pricing. Flash output tokens jumped from ¥2 to ¥4.5 (¥9 at peak); Pro went up 3–5×. My monthly bill went from ¥120 to ¥700+.

I tried everything. OpenCode Go's "affordable subscription" turned out to be the same official price with a monthly cap. MiMo-V2.5 was cheap but too weak for real coding. GLM-5.1 and Kimi K3 cost more than Flash. Subscription plans didn't help either — Xiaomi's Token Plan Pro covered my usage but couldn't match the quality; Zhipu's Coding Plan was too tight; Google AI Pro and ChatGPT subscriptions don't offer API access.

No single option could deliver both "no quality loss" and "affordable price."

887M tokens/month, 80% cached input — that was my usage profile with DS Flash. The price hike hit output tokens hardest, and most of my output was thinking tokens I couldn't just turn off.

Then it clicked: instead of finding one cheaper model to replace everything, let the same system call different models for different tasks. Simple file operations, formatting, and Q&A go to the cheap model; real reasoning and complex coding stay on Flash. Distribute on demand, not one-size-fits-all.

That's how dsh-smart-router was born. It's not an architectural breakthrough — just a practical tool born from sticker shock, routing each request to the best-fit model so you can find your own balance between capability and cost.

> If AI API bills have ever stung you, hopefully this small tool saves you some money.

What it solves

• Performance tiering: hard tasks (architecture, refactoring, tricky debugging) go to a strong model; easy tasks (small talk, wrap-ups, tiny edits) go to a cheap fast model; everything else hits the middle tier — like cc-switch's per-tier model slots for Claude Code, but for DSH.
• Vision built in: many users add a separate vision plugin; this router integrates vision routing into itself. Requests with images → vision tier automatically; plain text → difficulty tiers. Out of the box it ships a free anonymous vision model (OVHcloud Qwen2.5-VL-72B-Instruct, no API key), replaceable with any vision-capable model you configured.
• Cache friendly: pick all three tiers from the same vendor family (e.g. pro and flash editions of one series) for higher prefix-cache hit rates and lower cost (hint shown in the settings card).

Install

dsh plugin --profile web add dsh-smart-router

Restart dsh web, then:

1. Open Settings → Smart Router: pick a model for each of the four rows (hard / normal / easy / vision). Dropdowns list the models you already configured; the vision row only lists models that declare image input. 2. Back in chat, open the model picker and choose Smart Router (auto route). 3. Chat normally — every request is classified and routed automatically.

> Zero configuration also works: requests fall back to your default model (fail-open, never silently broken).

How it works

You pick the virtual model "smart" (declares text+image, passes DSH image admission)
        │
        ▼
SmartRouterAdapter.stream(request)
        │
        ├─ has image? ──────→ vision tier (free anonymous OVH by default) ──┐
        │                                                                │
        └─ else: classify ──→ hard → hard tier                             │
          (heuristic by    → normal → normal tier                          │
           default / LLM)  → easy → easy tier                              │
        │                                                                │
        ▼                                                                ▼
   ctx.llm.prepareCall({provider, model}).stream(request)  ← passthrough stream
        │
        └─ missing tier → hard→normal→easy→default ladder; error only when all fail

• Routing happens at the LLM adapter layer (architecture from llm-adaptive / dsh-vision-mix): a virtual provider is registered; stream() decides per request and delegates via prepareCall().stream() — no host patches, no conflicts with other vision plugins.
• Classifier modes: heuristic (default; zero cost — keywords + code volume + file references) or LLM (more accurate; one small call on the easy-tier model, 120s cache).
• Image admission: the virtual model declares inputModalities: ['text', 'image'], so DSH's preflight (MODEL_DOES_NOT_SUPPORT_IMAGES) lets images through — no host patching (compare dsh-easyvision's patch approach).

Settings (Settings → Smart Router)

Reasoning effort two-tier hierarchy: The reasoning effort in the chat input model selector (Off / High) is the master switch:

• Off: reasoning disabled for all tiers (no extended thinking)
• High: each tier uses the effort configured above (e.g. hardEffort=high, normalEffort=max, etc.)

A hint explaining this is shown in the tier models section of the settings page.

Manual config (~/.dsh/profiles/web/settings.yaml):

smart-router:
  enabled: true
  classifier: heuristic        # heuristic | llm
  hardProvider: deepseek-official
  hardModel: deepseek-v4-pro
  normalProvider: deepseek-official
  normalModel: deepseek-chat
  easyProvider: deepseek-official
  easyModel: deepseek-flash
  visionProvider: zhipu-vision # free quota; add GLM_API_KEY under Settings → Models
  visionModel: glm-4v-flash
  visionFallbacks: []
  fallbackProvider: ''
  fallbackModel: ''

On install the plugin idempotently seeds two free vision routes into llm-pi-ai (missing keys only, your config is never touched):

> The vision default works out of the box; for more stable/stronger vision, point the vision row at any vision model you configured.

Coexistence with other vision plugins

The router only participates when the session model is smart; it never takes over existing provider routes and never touches the host, so modlens, dsh-vision-router and friends can stay installed side by side.

Development

npm test          # node --test tests/ (80 cases: classifier / chain / vision sidecar / config API / seed idempotency / schema)

Local install for debugging:

dsh plugin --profile web add C:\path\to\dsh-smart-router
dsh --profile web --dump-config | grep smart-router

Acknowledgements (References)

This plugin directly references the following open-source projects and DSH internals:

License

MIT