v0.24.0-preview.0
Gemini CLISummary
This release introduces significant enhancements to Agent Skills, including infrastructure setup, autonomous activation tools, and system prompt awareness. New features also include experimental in-CLI extension management and support for remote agents. Several stability improvements were made, such as fixing crashes related to I/O errors and enhancing policy engine granularity for shell command allowlisting.
New Features
- Added support for Folder Trust when using Hooks.
- Introduced dynamic mode-aware policy evaluation.
- Enabled automatic persistence of the last used /model across Gemini CLI sessions.
- Implemented unified secrets sanitization and environment variable redaction options.
- Added experimental in-CLI subcommands for installing and uninstalling extensions.
- Added support for A2A Client Manager.
- Enabled terse transformations of image paths within the text buffer.
- Added security warnings and improved layout for the Hooks list.
- Added project-level hook warnings for security context.
- Added modifyOtherKeys protocol support for tmux integration.
- Enabled granular shell command allowlisting within the policy engine.
- Added support for remote agents and multi-agent TOML files.
- Implemented tool input modification support.
- Added instructions to the extensions update info notification.
- Added extension settings information to the /extensions list command output.
- Implemented Core Skill Infrastructure and Tiered Discovery for Agent Skills.
- Added remote agents to the agent registry.
- Implemented STOP_EXECUTION functionality and enhanced hook decision handling.
- Implemented Autonomous Activation Tool and Context Injection for Agent Skills.
- Added folder suggestions to the /dir add command.
- Implemented Agent Integration and System Prompt Awareness for Agent Skills.
- Added Status Bar Integration to display Skill Counts for Agent Skills.
- Implemented Unification of Representation and Centralized Loading for Agent Skills.
- Restored MessageBus optionality for soft migration (Phase 1).
- Standardized Tool and Agent Invocation constructors (Phase 2).
Bug Fixes
- Fixed a crash that occurred when cleaning up unhandled EIO errors during stdin reading.
- Fixed missing `type` field in MCPServerConfig.
- Resolved an unhandled promise rejection in ide-client.ts.
- Fixed failure in GitService.initialize when checking if the current directory is a repository.
- Fixed an issue where the `enableShellOutputEfficiency` setting was missing from the settings schema.
- Fixed the policy engine by enabling granular shell command allowlisting.
- Fixed deduplication issues for agent hooks and added cross-platform integration tests.
- Fixed a build issue caused by a year-specific linter rule.
- Fixed an unhandled promise rejection in mcp-client-manager.
- Improved shell command detection when handling redirection.
- Fixed mocking of powershell output in shell-utils tests.
Improvements
- Improved error messages displayed when using the `--resume` flag.
- Made schema validation errors non-fatal, preventing crashes.
- Limited the display of MCP resources to 10 by default.
- Implemented exponential back-off retries for retryable errors without a specified back-off duration.
- Unified shell security policy and removed legacy logic.
- Enforced mandatory MessageBus injection (Phase 3 Hard Migration).