dongsheng123132/dsh-cache-stabilizer ↗★ 1
dsh-cache-stabilizer
DeepSeek Harness 的缓存前缀稳定和基于证据的缓存指标
AI 分析
核心用途是通过优化工作目录定位和工具规范化来提高大模型提示词缓存(Prompt Cache)的复用率。适合需要降低 Token 消耗、优化响应延迟的 DSH 用户。
安装
npx -p @deepseek-ai/dsh dsh plugin --profile web add github:dongsheng123132/dsh-cache-stabilizer说明文档
阅读完整 README ↗dsh-cache-stabilizer
An MIT-licensed DeepSeek Harness plugin that improves the chance of provider prompt-cache reuse without hiding stale state.
It makes two semantics-preserving changes:
- Moves the working directory out of DSH's known default persona sentence and into the runtime-context snapshot. Different projects can then share the same system-prompt prefix while each request still receives the correct
cwd. - Canonicalizes object-key order inside tool schemas. Tool order itself is already deterministic in DSH.
It also adds /cache, a human-only command that reports the provider's durable cacheReadTokens, uncached inputTokens, and cache-write tokens. It never invents a cache hit.
Install
dsh plugin --profile web add dsh-cache-stabilizer
Restart DSH, send a few messages, then enter /cache in a command-capable client.
For a custom profile, replace web with its profile name. To disable either optimization in a profile patch:
- id: dsh-cache-stabilizer
config:
relocateCwd: false
canonicalizeTools: false
Safety boundary
Only the exact sentence used by DSH's standard/headless coding persona is relocated. A custom persona that mentions {{cwd}} in another form is left unchanged because blindly moving arbitrary prose can change meaning. The plugin does not freeze tool catalogs, reuse stale context, proxy model responses, or implement a second cache.
DeepSeek's provider cache is automatic and depends on an exact prefix match from token zero. Storage and eviction remain provider-controlled.
Development
npm test
npm run check