dsh-compaction-meta

Most harness frameworks today manage context through passive threshold compaction: once token counts cross a preset line, the system mechanically trims history with no input from the model. dsh-compaction-meta tackles this by returning the compaction decision to the model itself, letting its meta-cognitive judgment determine which spans to keep and which to fold, rather than triggering on a single number.

The plugin plans to deliver three core capabilities: context_list exposes the current context folded by turn and tool call, with numbering and token estimates so the model can locate specific memory intervals; context_compact lets the model replace a chosen interval with self-authored summary text, routed through the official compaction transaction to preserve checkpoint markers, tool-call pairing balance, and log rebuildability; and it disables threshold-based passive compaction entirely, supplemented by a companion skill that teaches the model when and how to use these actions.

Ideal for developers who care about model meta-cognition and long-term learning, especially those running long conversations, multi-tool-chain invocations, or scenarios requiring fine-grained control over context lifecycle in DSH. The project is currently in skeleton form and not yet implemented, making it most relevant to early adopters who wish to track or contribute to its progression.