dsh-compaction-micro

In DeepSeek Harness's compaction pipeline, model-backed summarization is the most expensive step: a full model call with significant token cost and latency. When a session accumulates many results from read-only, re-runnable tools like read, glob, or grep, context pressure can hit the ceiling well before the summarize threshold is reached. dsh-compaction-micro fills that gap as a micro-clear tier: at its own microThresholdRatio (default 0.85 of the effective window), it replaces the oldest whitelisted tool results with placeholders, relieving pressure early and reducing how often the costly summarization actually fires.

It is a pure observer by design. It never mutates the model request and makes no model calls. Every rewrite is a single-node content replacement, and each one is preceded by a compaction/prune shadow-price event so the durable log remains the reconstructable source of truth. The whitelist is fully operator-owned: read, glob, grep, and read-only web tools are the intended members, while write, edit, shell, and subprocess tools must stay out. The retainRecent setting keeps the newest N results visible at all times, protecting the working context.

This plugin suits Harness users whose workflows lean heavily on read-only tool calls, such as file reads, code search, and web scraping, who want to cut summarization frequency without losing auditability. It slots in as a lightweight, non-invasive front layer that works alongside the official dsh-compaction-basic, letting the expensive summarize step run only when the cheaper micro-clear has already done its job.