Deepseek Harness Token Saver — Development skill for Claude Code

---

🇬🇧 English

Every turn, DeepSeek Harness injects **all** skill descriptions and **all** tool schemas — needed or not. Measured on a real workspace: **36,870 tokens per step, ~68% of it irrelevant.**

**This skill scopes capabilities to each workspace.**

Per-step overhead	Before	After	Cut
Skill catalog	56 skills	18 skills	−91%
Tool schemas	95 tools	31 tools	−61%
Total	36,870	11,658	−68%

**Install**

curl -L https://github.com/yzsp-studio/deepseek-harness-token-saver/archive/refs/heads/main.tar.gz \
  | tar -xz --strip-components=1 -C "$DSH_HOME/skills/workspace-token-profile"

**Lossless by design** — nothing is deleted: domain skills move to their own workspace (still auto-trigger there), tools use an exact allowlist. No description shortened, no content truncated, no history compressed.

---

🇨🇳 中文

每一轮，DeepSeek Harness 都会把**全部**技能描述和**全部**工具 schema 塞进上下文 —— 不管你用不用。 真实工作区实测：**每步 36,870 token，其中约 68% 与当前工作区无关。**

**本技能按工作区裁剪能力面。**

每步固定开销	之前	之后	砍掉
技能目录	56 个	18 个	−91%
工具 schema	95 个	31 个	−61%
合计	36,870	11,658	−68%

**安装**

curl -L https://github.com/yzsp-studio/deepseek-harness-token-saver/archive/refs/heads/main.tar.gz \
  | tar -xz --strip-components=1 -C "$DSH_HOME/skills/workspace-token-profile"

**为什么无损** —— 一个字都没删：领域技能移到它自己的工作区（在那里照常自动触发），工具用精确白名单。**不缩短描述、不截断内容、不压缩历史。**

---

❤️ Support / 支持本项目

**Free to use. Optional to support.