seainml/dsh-token-slim

README 摘要

dsh-token-slim Token optimization plugin suite for DeepSeek Harness (DSH) . Three composable Cordis plugins that apply the techniques researched in docs/RESEARCH.md to the extension points DSH already provides — without touching the shipped compaction engine. Plugin DSH extension point Research basis Default dsh-token-slim/noise-filter tools/post-execute "quiet flags / output limits" (Claude blog), rtk / squeez on dsh-token-slim/context-audit tokenMeter.measure + tools.register + session/event /context habit, token-cost awareness on dsh-token-slim/selective-context agent/pre-step Selective Context / memory-compaction papers off (experimental) Why In agentic coding tools every token that enters the context is re-read on every later turn. The highest-leverage optimizations are therefore: (1) keep noisy command output out of the context, (2) know how much context you are burning and what to do about it, (3) when under pressure, keep only the high-value history. DSH already ships the compression engine ( tokenMeter , compaction , toolResultPruner ); this suite adds the content-aware and user-facing layers around it. See docs/RESEARCH.md for the full report. Install The plugins run insi…

分类