v3.0.0 — Single Compress Tool Architecture
A major release featuring a complete architectural overhaul of the Dynamic Context Pruning system.
Breaking Changes
- The previous 3-tool system (
distill,compress,prune) has been replaced with a singlecompresstool. All context management now flows through one unified interface.
New Features
- Decompress / recompress commands — Built-in commands for managing compression blocks
- Subagent support —
experimental.allowSubAgentsenables compression awareness in sub-agent contexts - Custom prompts —
experimental.customPromptsallows user-defined prompt overrides - Flat schema option —
compress.flatSchemafor simplified tool schema presentation - Protected user messages —
protectUserMessagesprevents user messages from being compressed - Protected tools with glob patterns — Configure which tool outputs are preserved during compression using glob matching
- Visual progress bar — Real-time compression activity indicator
- Configurable turn nudges — Fine-tune when and how compression nudges appear
- Strengthened manual mode — More robust manual compression control
Improvements
- Reduced cache invalidation — Significantly decreases how often DCP causes cache invalidation, fixing Anthropic-related cache issues without needing manual mode
- Infinite conversations — Conversations can now last almost indefinitely; the previous tool pruning approach limited this by leaving user/AI messages untouched
- Simplified model behavior — The model no longer needs to choose between 3 context management tools, reducing decision complexity and improving reliability
Contributors
Migration
This is a major version bump. Users upgrading from v2.x should review the updated README for the new single-tool configuration. The compress tool now handles all context management operations that were previously split across multiple tools.
Full Changelog: v2.1.8...v3.0.0
