Repository navigation
0.2.6
[0.2.6] - 2025-12-09
Change
- Welcome screen Change: The original welcome screen, which prompted with '/model list', has been changed to '/model'.
- System Notification: Added a new notification system that sends a desktop notification and plays a sound alert when AI tasks are completed.
- Enhanced
file_readTool: Added support for reading specific line ranges.- New optional parameters:
start_lineandend_line. - Useful for reading large files in chunks to save tokens.
- Displays
(Lines X-Y)header when a range is selected.
- New optional parameters:
- Add LSP support for outline and index(Other Branch)
- Added LSP client implementation using async-lsp-client
- Refactored file outline and indexer to prioritize Tree-sitter with LSP fallback
- Added configuration support for custom LSP servers via lsp.json
- Added --ts and --lsp flags to /index commands
- Added LSP configuration documentation
- Separate Shorekeeper model configuration: Add the /model sk command to support setting up Shorekeeper models independently.
- Added Jury mode (--jury): Introduces a three-model independent jury mechanism, requiring a 2/3 majority vote to execute sensitive operations.
- Hooks Support: Provides a hooks system that allows you to set up automated execution of custom scripts or commands at specific stages.
- A built-in Lua interpreter was introduced.
- See
docs\HOOKS.mdfor details.
- MCP
- Deep MCP Tool Integration
- Dynamic Tool Discovery: Friendev can now automatically discover and register all tools provided by connected MCP servers.
- Namespace Isolation: To prevent naming conflicts, MCP tools are automatically named in the format server/tool.
- AI-Native Support: AI assistants can now "see" and directly call these MCP tools to complete tasks without manual intervention.
- Resource Access System (Resources)
- Resource List: Added the mcp_resource_list tool to view all available resources.
- Resource Reading: Added the mcp_resource_read tool to support reading the content of resource URIs.
- Interactive Prompt System
- Command: Added the /prompt interactive command.
- Workflow: Implemented a 5-step fully automated workflow: Select Server → Select Prompt → Fill Input Parameters → Execute Prompt → Automatically Send Results to AI.
- Added the /mcp Command
- Deep MCP Tool Integration
- Add todo management tools and command
- Implement
todo_writeandtodo_readtools for AI task management - Add
/todocommand for users to view current task list - Support status visualization (pending, in_progress, completed)
- Implement
- /send.md
- The
/send.mdcommand has been added. - After using the
/send.mdcommand, the system will automatically read the send.md file in the project root directory and then send it to AI.
- The
What's Changed
- Feat/helloaixiaoji/user experience optimization scope view by @HelloAIXIAOJI in #7
- System Notification by @HelloAIXIAOJI in #8
- Feat/helloaixiaoji/lsp outline by @HelloAIXIAOJI in #9
- Feat/helloaixiaoji/lsp by @HelloAIXIAOJI in #10
- Separate Shorekeeper model configuration by @HelloAIXIAOJI in #11
- Feat/jury by @HelloAIXIAOJI in #12
- Feat/hok by @HelloAIXIAOJI in #13
- Update .gitignore to include and exclude patterns by @HelloAIXIAOJI in #14
- upload dot friendev by @HelloAIXIAOJI in #15
- Ign by @HelloAIXIAOJI in #16
- Feat/helloaixiaoji/todo by @HelloAIXIAOJI in #17
- Feat/mcp/client by @HelloAIXIAOJI in #18
- Feat/helloaixiaoji/user experience optimization by @HelloAIXIAOJI in #19
- fix a big error by @HelloAIXIAOJI in #20
- refactor(commands): Handling to support full argument array by @HelloAIXIAOJI in #21
- 0.2.6Chagelog by @HelloAIXIAOJI in #22
- /send.md by @HelloAIXIAOJI in #23
New Contributors
- @HelloAIXIAOJI made their first contribution in #7
Full Changelog: v0.2.5...v0.2.6