v0.36.0
Gemini CLISummary
This release introduces significant architectural updates, including multi-registry support and tool filtering for subagents, alongside an experimental memory manager agent. Key fixes address issues with context handling, session logging, and UI rendering, while new features enhance plan mode usability and macOS sandboxing security.
New Features
- Introduced multi-registry architecture and tool filtering for subagents.
- Added support for dynamic model Resolution to ModelConfigService.
- Integrated task tracker protocol into the core system prompt.
- Added agent acknowledgment command and enhanced registry discovery.
- Added 'blocked' status to tasks and todos.
- Enabled subagent local execution and tool isolation.
- Formatted multi-line banner warnings with a bold title in the UI.
- Removed the write todo list tool from UI tips.
- Added support for plan mode in non-interactive mode.
- Implemented strict macOS sandboxing using a Seatbelt allowlist.
- Enabled resilient subagent tool rejection with contextual feedback.
- Added content-utils.
- Added partial threading of AgentLoopContext.
- Injected memory and JIT context into subagents.
- Capped JIT context upward traversal at the git root.
- Added experimental memory manager agent to replace the save_memory tool.
- Added token usage metadata to the ACP `send` method's return value.
- Added ModelChain support to ModelConfigService and made ModelDialog dynamic.
Bug Fixes
- Fixed use of theme-aware text colors for console warnings and errors in devtools.
- Fixed CLI to use the active sessionId in useLogger and improved resume robustness.
- Fixed CLI to expand tilde (~) in policy paths read from settings.json.
- Added actionable warnings for terminal fallbacks.
- Fixed CLI to automatically add all VSCode workspace folders to Gemini context.
- Fixed CLI to lazily load settings in onModelChange to prevent stale closure data loss.
- Fixed CLI to resolve subagent grouping and UI state persistence issues.
- Fixed an issue where subagent thoughts were being appended incorrectly.
- Fixed CLI to correctly handle auto-update for standalone binaries.
- Fixed core logic to not persist browser consent sentinel in non-interactive mode.
- Fixed core logic to narrow browser agent description to prevent stealing URL tasks from web_fetch.
- Fixed UI flickering observed on small terminal heights.
- Fixed ACP to provide more meta information in tool_call_update.
- Fixed core logic to explicitly map execution context in LocalAgentExecutor.
- Fixed CLI to resolve duplicate footer appearing when canceling a tool via ESC.
- Fixed core logic to circumvent genai sdk requirement for api key when using gateway auth via ACP.
- Fixed plan mode policies to clarify they are combined with normal mode policies.
Improvements
- Enabled the "Allow all server tools" session policy for the browser agent.
- Removed extra newlines in ShellToolMessage.tsx.
- Simplified keypress and mouse providers and updated related tests.
- Removed references to stale Gemini CLI file structure information in documentation.
- Added FAQ entry for checking Gemini CLI version.
- Fixed logging and virtual list rendering.
- Upgraded ACP SDK from 0.12 to 0.16.1.
- Updated website issue template for label and title.
- Updated callouts to work correctly on GitHub.