dsh-compaction-cache: Enabling DSH Compaction Summaries to Hit Provider Prefix Caches
The `dsh-compaction-cache` plugin for DeepSeek Harness (DSH) is designed to resolve the issue where provider prefix caches cannot be reused during compaction summaries because `reasoningEffort`, `temperature`, and `maxTokens` are missing. The plugin is mounted on the `llm/stream` stream and only takes effect for requests marked with `purpose: 'compaction'`. By mirroring routing header parameters and enforcing `maxTokens` alignment rules, it completes the summary envelope to keep the invocation environment consistent with the previous turn, significantly reducing computation costs. The plugin has zero dependencies, requires no build, and is distributed only via GitHub. All configuration is optional and enabled by default. Once active, logs will
Read More