dsh-force-compact

When serving large models via a local llama.cpp endpoint, the context window is often only a few dozen K tokens, and a long conversation quickly hits the hard ceiling. dsh-force-compact takes the opposite approach: instead of fighting a small cap, it actively compacts the conversation—summarizing older history into a tight compressed head while retaining the latest N tokens verbatim, so the agent always reasons over a compact, high-signal prompt with deep memory preserved.

Key capabilities include two compaction engines (official service first, builtin persistent transaction as automatic fallback) that switch transparently; thinking mode disabled only for the plugin's own summarization call while all other requests remain untouched; a live UI badge showing compaction state (compressing / done / working); and hot-reloadable threshold and retention settings that take effect on the next request without a restart. All processing stays local with zero API cost and zero data egress.

Ideal for developers and self-hosters running local llama.cpp or OpenAI-compatible endpoints with limited context windows, who value privacy, cost efficiency, and the ability to extend effective session length without sacrificing inference quality.