deepseek-harness-compaction-ui

DeepSeek Harness ships with a context compaction engine, but its native configuration is ratio-based and YAML-driven, making it hard to reason about actual token budgets at a glance. This plugin adds a visual configuration card to the Web settings page, styled consistently with native Harness plugin cards, so users can tune compaction without opening a terminal or editing YAML files.

The core capabilities revolve around two things. First, both the trigger threshold and the recent-content retention can be set as a percentage of the model context window or as an absolute token count, with automatic conversion between the two when the unit is switched. Second, configuration is hot-reloaded on save, so there is no need to recreate a session. A small progress ring sits next to the input box at all times; hovering over it reveals the current context usage, the auto-compaction threshold, and the occupancy ratio. The ring turns red once the threshold is reached, while a lighter outer ring shows Harness's own context-window occupancy. Independent strategies can also be configured per provider or model.

This plugin is best suited for teams or individuals running long, multi-turn conversations in Harness and needing fine-grained control over the context window, especially when switching between models with different context sizes. It is important to note that this plugin is purely a visual configuration layer. Summary generation, message replacement, checkpointing, persistence, and all other compaction engine logic remain the responsibility of the official @deepseek-ai/dsh-compaction-basic package.