zhangzhangco/dsh-tier-router ↗★ 0

dsh-tier-router

Automatic tier-based model routing for DeepSeek Harness: a virtual `smart` model classifies each request by difficulty (hard / normal / easy) and by vision need, then delegates it to the models you already configured. 三级难度 + 视觉自动路由,虚拟 smart 模型零配置接入。 适合已配置多模型,想通过虚拟smart模型实现自动分流的用户。

패키지
dsh-tier-router
호환성
미검증
Harness peer 범위
^0.1.5-rc.2
Cordis peer 범위
^4.0.2
버전
0.2.1
라이선스
NOASSERTION
최근 업데이트
2026. 9. 16.

같은 패키지 이름의 다른 저장소

설치

$npx -p @deepseek-ai/dsh dsh plugin --profile web add github:zhangzhangco/dsh-tier-router

Configuration

Settings live in the tier-router namespace as flat fields. Edit them in the settings card or write them directly:

tier-router:
  enabled: true
  classifier: heuristic          # heuristic | llm
  hardProvider: codex-local
  hardModel: gpt-6-astra
  hardEffort: ''                 # e.g. low / high / max; empty = unspecified
  normalProvider: codex-local
  normalModel: gpt-5.5
  normalEffort: ''
  easyProvider: gpudev
  easyModel: qwen3.8-27b-q5
  easyEffort: ''
  visionProvider: codex-local
  visionModel: gpt-6-astra
  visionEffort: ''
  visionMode: replace            # replace (structured evidence, default) | route (whole turn)
  visionCacheTtl: 3600           # seconds of vision-evidence cache; 0 disables
  visionFallbacks: []            # [{provider, model}] explicit vision fallbacks
  fallbackProvider: ''           # last resort; empty = the session default model
  fallbackModel: ''
  llmClassifierProvider: ''      # classifier: llm; empty = reuse the easy tier
  llmClassifierModel: ''
FieldDefaultMeaning
enabledtrueMaster switch; when off, requests go to the session default model.
classifierheuristicheuristic (built-in scoring) or llm (a model decides the tier).
hardProvider / hardModel / hardEffortcodex-local / gpt-6-astra / ''Hardest tier.
normalProvider / normalModel / normalEffortcodex-local / gpt-5.5 / ''Everyday tier.
easyProvider / easyModel / easyEffortgpudev / qwen3.8-27b-q5 / ''Cheapest tier.
visionProvider / visionModel / visionEffortcodex-local / gpt-6-astra / ''Image tier.
visionModereplaceImage handling; see below.
visionCacheTtl3600Vision-evidence cache in seconds.
visionFallbacks[]Explicit vision fallbacks before the default model.
fallbackProvider / fallbackModel''Route used when no tier is configured; empty = session default.
llmClassifierProvider / llmClassifierModel''Classifier model for classifier: llm.
classifierTimeoutMs4000Budget for the LLM classifier. On timeout the heuristic decides immediately and the slow answer is cached for later requests.
visionTimeoutMs60000Budget for one vision-sidecar call, so a hung vision provider cannot stall the turn.
contextGuardtrueSkip routes whose known context window cannot hold the request.