ManoloRemiddi/dsh-adaptive-reasoning ↗★ 0

dsh-adaptive-reasoning

Conservative per-request reasoning selection for DeepSeek Harness, with no classifier model calls 适合需要优化推理成本和速度的用户,在无需分类器模型的情况下动态控制思考强度。

パッケージ
dsh-adaptive-reasoning
互換性
未検証
Harness ピア範囲
0.1.5-rc.1
Cordis ピア範囲
4.0.2
バージョン
0.2.3
ライセンス
MIT
最終更新
2026/09/23

同名パッケージの別リポジトリ

インストール

$npx -p @deepseek-ai/dsh dsh plugin --profile web add github:ManoloRemiddi/dsh-adaptive-reasoning

ドキュメント

README 全文を読む ↗

Configuration

  • enabled: boolean, default true.
  • textOnly: boolean, default true.
  • presets: exact preset allowlist; empty means inactive.
  • routes: exact provider/model pairs with an efforts mapping for off, low, medium, high. The existing Qwen maps high to xhigh.

Automatic mode owns per-request effort on allowlisted routes. Explicit requests to think deeply remain high. Unsupported effort strings fail DSH's adapter validation before inference. Unrecognised languages retain high reasoning.