Smart Compaction: Letting the Model Decide When to Compress Context

Smart Compaction is a plugin for DeepSeek Harness (DSH) designed to address the potential information loss caused by the original compaction mechanism, which is triggered based on a fixed threshold. It provides the model with `compact_now` and `context_status` tools, allowing the model to proactively assess and use safe points—such as task completion—to trigger compaction. The plugin can also automatically trigger compaction when a task is completed and context usage exceeds a configured threshold, and the threshold can be adjusted via configuration. Once installed, it relies only on DSH's public APIs and makes no invasive modifications.