v2.0.1 - Percentage contextLimit support

What's Changed

Features

• Percentage contextLimit support - Configure contextLimit with percentage values like "60%" instead of just absolute token counts

Improvements

• Default contextLimit now set to 60% for optimal context management
• Default compress permission changed to deny, this tool is experimental for now

Breaking Changes

• Removed "model" option from contextLimit configuration - use percentage or absolute values instead

Contributors

• @tobwen
• @spoons-and-mirrors

Full Changelog: v2.0.0...v2.0.1

v2.0.0 - Major Release

What's Changed

New Features

• Compress Tool - New boundary-string matching conversation collapse tool for managing long conversations
• Standardized DCP Tools - Unified Distill, Compress, and Prune as the core context management toolkit
• Prompt System Rewrite - Complete overhaul of the prompt system with build-time generation for better maintainability
• Token Usage Tracking - Real-time session token tracking with configurable contextLimit
• Secure Mode Auth - HTTP Basic Auth support via OPENCODE_SERVER_PASSWORD environment variable
• Toast Notifications - Configurable pruneNotificationType for notification style
• Full Tool Pruning - Edit and write tool outputs are now prunable (no longer protected by default)
• supersedeWrites Enabled by Default - Write tool outputs are automatically superseded
• Message Count Tracking - Message counts now displayed in stats and context info
• apply_patch and multiedit Support - File path extraction for apply_patch and multiedit tools
• XDG Base Directory Support - Respects XDG_CONFIG_HOME and XDG_DATA_HOME environment variables
• Tool Permission Config - New ask/allow/deny permission model replacing boolean config

Fixes

• Opus 4.6 fix for unsupported prefill errors
• Improved error messages across all DCP tools
• Various bugfixes and performance improvements

Other

• Relicensed to AGPL-3.0
• Documentation and provider guidance improvements (cache limits, uniform pricing providers)

Contributors

• @azais-corentin
• @essinghigh
• @spoons-and-mirrors

Full Changelog: v1.2.8...v2.0.0

v1.2.8 - Hybrid injection & empty protectedTools

What's Changed

• fix: new injection strategy for DeepSeek/Kimi models, no more errors for these models
• feat: Default to empty protectedTools for strategies
• docs: Add Ko-fi sponsorship link

Full Changelog: v1.2.7...v1.2.8

v1.2.7 - Context injection and defensive improvements

What's Changed

• Simplify default config to only include schema URL (empty dcp.jsonc by default)
• Add defensive array checks for msg.parts iteration (prevents TypeError: {} is not iterable)
• Inject context as user message when last message is user (prunable tools list on every fetch)
• Show extracted token count in prune notifications

Full Changelog: v1.2.6...v1.2.7

v1.2.6 - Add /dcp sweep command

What's Changed

• Add /dcp sweep command for manual context pruning - allows users to trigger a context sweep on demand

Full Changelog: v1.2.5...v1.2.6

v1.2.5 - Context command UI improvements

What's Changed

• feat: Show tool count in context command UI
• refactor: Optimize token calculation in context command
• feat: Improve /dcp context UI with pruned count and accurate progress bars

Full Changelog: v1.2.4...v1.2.5

v1.2.4 - Unified DCP commands and improved error handling

What's Changed

• New /dcp commands: Added /dcp context and /dcp stats commands for viewing context state and pruning statistics
• Improved error handling: Failed tools now throw errors instead of returning error strings, with prunable tools list shown on failure
• Tool error logging: Log tool error parts for better debugging
• README updates: Added documentation for cache hit rates and new /dcp commands
• Config updates: Added JSON schema support for new commands configuration
• System prompt improvements: Updated injection style to use tool parts instead of user parts, clarified that context_info is not available to the model

Full Changelog: v1.2.3...v1.2.4

v1.2.3 - Prompt loading & Gemini improvements

What's Changed

• Simplify prompt loading by moving prompts from .txt to .ts files for easier maintenance
• Add thoughtSignature bypass for Gemini tool part injection for improved compatibility

Full Changelog: v1.2.2...v1.2.3

v1.2.2 - Injection timing and Gemini compatibility fixes

What's Changed

• Simplified injection timing: Never inject immediately after a user message - wait until assistant has started its turn. This avoids interfering with model reasoning/thinking phases across all providers.
• Gemini compatibility: Use text injection instead of tool parts for Gemini models to avoid thought signature validation errors with Gemini 3+.

Full Changelog: v1.2.1...v1.2.2

v1.2.1 - Fix Claude reasoning block injection

What's Changed

• Fix: require reasoning block before injecting context for Claude models

Full Changelog: v1.2.0...v1.2.1