dsh-model-router

Deterministic model routing for DeepSeek Harness. The plugin ships a small static model catalog and exposes it to the model as three pure tools: recommend a model for a task, compare named models, and price a token budget. No network calls, no subprocesses, no provider registry — every answer comes from the embedded catalog and fixed arithmetic, so repeated calls return byte-identical results.

Package: @qingshanjiluo/dsh-model-router · Type: host tool plugin (Cordis, inject: ['tools'])

Install

npx -y @deepseek-ai/dsh plugin --profile web add @qingshanjiluo/dsh-model-router

Tools

Tool	Parameters	Returns
model_recommend	task (chat | code | analysis | summarize | agents | vision), needVision (boolean), budget (max USD per task, 0 = configured default)	Best-first ranked list with score, estCostUsd, context, capabilities, and per-dimension reasons; plus considered, every excluded model with its reason, and notes.
model_compare	ids (array of model ids)	Resolved rows sorted by reference price (context, costIn, costOut, refCostUsd, capabilities), winners per dimension, deltas cost/context ratios, and missing ids.
model_cost	id, tokensIn, tokensOut	inputCostUsd / outputCostUsd / totalCostUsd for exactly that workload, plus contextTokens, fitsInContext, and notes. Unknown id or negative tokens ⇒ ok: false with error.

All three tools are concurrency-safe and validated against a declared output schema. Argument mistakes are rejected by the tool registry before the body runs.

How the score is built

Each task kind carries a profile: required capabilities, a target context size, a typical token mix, and three weights.

score = w_capability * capabilityFit + w_cost * costFit + w_context * contextFit

• capabilityFit — share of the task's required capabilities the model has.
• costFit — the model's price for the task's typical token mix, normalized linearly across the surviving candidates (cheapest = 1).
• contextFit — min(1, context / targetContext).

The score is rounded to 4 decimals and ties break on cheaper, then on catalog id, so rankings never depend on argument order or locale. needVision: true is a hard filter (non-vision models are excluded, not just penalized), as is budget. With preferLowCost, the cost weight doubles and the three weights renormalize.

Built-in catalog

Prices are USD per 1 000 000 tokens. model_compare prices a fixed reference workload of 9000 input + 1000 output tokens so ratios stay comparable.

id	context	cost in /M	cost out /M	capabilities
deepseek-chat	128 000	0.27	1.10	text, reasoning, tools, code
deepseek-reasoner	64 000	0.27	1.10	text, reasoning, code
gpt-4o-mini	128 000	0.15	0.60	text, vision, tools, code
gpt-4o	128 000	2.50	10.00	text, vision, tools, code
claude-sonnet-4	200 000	3.00	15.00	text, vision, reasoning, tools, code
claude-haiku-3.5	200 000	0.80	4.00	text, vision, tools, code
gemini-flash	1 000 000	0.10	0.40	text, vision, tools, code
llama-3.3-70b	128 000	0.60	0.60	text, tools
qwen2.5-coder-32b	32 768	0	0	text, code (self-hosted, unmetered)

Ids are matched case- and separator-insensitively (GPT-4o-Mini resolves).

Configuration

Field	Type	Default	Meaning
maxResults	number	5	Ranked entries returned at most (clamped to 1..catalog size).
defaultBudgetUsd	number	0	Per-task USD ceiling applied when a call passes budget: 0; 0 = unlimited.
preferLowCost	boolean	false	Double the cost dimension of the score, then renormalize the weights.

Development

npm install --no-audit --no-fund
npx tsc --noEmit          # types
npm run build             # lib/index.js + lib/index.d.ts
npx vitest run            # behavior tests (export face + >=2 cases per tool)
node scripts/load-smoke.mjs   # loads lib/index.js and registers the tools

License

MIT