v0.36.0-preview.0
Gemini CLISummary
This release introduces significant architectural updates, including multi-registry support and enhanced subagent capabilities like local execution and memory injection. Key fixes address issues with path expansion, session handling, and UI rendering stability across various terminal sizes. Users will benefit from improved context awareness, better sandboxing on macOS, and new features like plan mode support in non-interactive sessions.
New Features
- Introduced multi-registry architecture and tool filtering for subagents.
- Enabled dynamic model resolution support within the ModelConfigService.
- Integrated task tracker protocol into the core system prompt.
- Added agent acknowledgment command and enhanced registry discovery for agent-to-agent (A2A) interactions.
- Added 'blocked' status capability for tasks and todos.
- Enabled subagent local execution and tool isolation.
- Formatted multi-line banner warnings with a bold title for better visibility.
- Removed the 'write todo list' tool from UI tips.
- Added support for plan mode in non-interactive command-line usage.
- Implemented strict macOS sandboxing using a Seatbelt allowlist.
- Added content-utils for core functionality.
- Added partial threading of AgentLoopContext.
- Enabled the 'Allow all server tools' session policy for the browser agent.
- Injected memory and JIT context into subagents.
- Capped JIT context upward traversal at the Git root directory.
- Added an experimental memory manager agent to replace the save_memory tool.
- Added token usage metadata to the return value of the ACP 'send' method.
- Added ModelChain support to ModelConfigService and made ModelDialog dynamic.
Bug Fixes
- Fixed console warnings and errors to use theme-aware text colors in devtools.
- Improved resume robustness by ensuring the active sessionId is used in the logger.
- Fixed issue where tilde (~) in policy paths from settings.json was not expanded.
- Added actionable warnings for terminal fallbacks.
- Fixed automatic inclusion of all VSCode workspace folders into Gemini context.
- Fixed UI state persistence and subagent grouping issues in the CLI.
- Fixed lazy loading of settings in onModelChange which could lead to stale closure data loss.
- Fixed issue where subagent thoughts were incorrectly appended to output.
- Fixed duplicate footer appearing when canceling a tool via the ESC key.
- Fixed flickering issues observed on small terminal heights.
- Fixed providing insufficient meta information in tool_call_update for ACP.
- Fixed auto-update mechanism for standalone binaries.
- Fixed circumventing the genai SDK requirement for an API key when using gateway authentication via ACP.
- Fixed persisting the browser consent sentinel in non-interactive mode.
- Fixed the browser agent description being too broad, preventing it from stealing URL tasks meant for web_fetch.
- Fixed logging and virtual list rendering issues.
- Fixed plan mode policies being incorrectly combined with normal mode policies.
- Fixed an issue where copied extensions were not writable in the user's local directory.
Improvements
- Improved Docker image reference for the GitHub MCP server.
- Refactored UI to extract SessionBrowser search and navigation components.
- Simplified keypress and mouse providers and updated associated tests.
- Removed extra newlines from ShellToolMessage.tsx.
- Refactored subagent trajectory deletion and switched to native filesystem testing.
- Simplified createMockSettings calls in CLI tests.
- Removed references to stale Gemini CLI file structure information in documentation.
- Updated website issue template for better label and title handling.
- Updated callouts to function correctly on GitHub.
- Upgraded ACP SDK from version 0.12 to 0.16.1.
- Added minor style updates based on an initial documentation audit.