v0.36.0-nightly.20260402.13ccc1645
Gemini CLISummary
This release introduces significant new features, including a behavioral-evals skill, support for multi-turn interactive evals, and a refreshed UX for the Composer layout. Numerous bug fixes address stability, including resolving terminal escape sequence leaks and improving sandbox path resolution in Plan Mode. Performance has also been boosted by parallelizing startup cleanup tasks and backgrounding the IDE client.
New Features
- Introduced PR Guidance workflow for Evals.
- Implemented full "GEMINI CLI" logo display for the logged-out state.
- Added the behavioral-evals skill, including fixing and promoting guides.
- Added LegacyAgentSession support.
- Added TestMcpServerBuilder and support in TestRig for testing utilities.
- Added multi-turn interactive evals for memoryManager.
- Added behavioral evaluations for subagent routing.
- Implemented a refreshed UX for the Composer layout.
- Implemented dynamic macOS sandbox expansion and worktree support.
- Enabled immediate user input during startup, preventing blocking.
- Changed user-facing authentication type from oauth2 to oauth.
- Added maxActionsPerTask setting for the browser agent.
Bug Fixes
- Fixed subagent tool updates not applying configuration overrides immediately.
- Resolved flicker observed at boundaries of the selection list in BaseSelectionList.
- Fixed commands with paths not correctly allowing the "always" option.
- Prevented terminal escape sequences from leaking upon tool exit.
- Fixed minimum height reservation for the selection list in AskUserDialog.
- Hardened AgentSession replay semantics.
- Fixed tool confirmations not taking up the entire terminal height in the UI.
- Prevented redundant remote agent loading when switching models.
- Fixed policy engine not properly supporting allowRedirect.
- Prevented subcommand shadowing and fixed skipping authentication for certain commands.
- Patched a memory leak and enforced logPrompts privacy in telemetry.
- Prevented Ctrl+D from exiting when the input buffer is not empty.
- Separated conversational text from the execute tool command title in ACP.
- Fixed accurately reflecting subagent tool failure in the UI.
- Fixed API key input dialog user interaction when the Gemini API Key was selected.
- Fixed sandbox path resolution in Plan Mode to prevent hallucinations.
- Fixed ACP by passing the cwd to AcpFileSystemService to avoid looping failures when asking for permissions to write the plan md file.
- Fixed core policy configuration requiring explicit specification of toolName.
- Fixed overlapping act() deadlock in app-test-helper during evals.
- Fixed updating telemetry attribute keys and adding a timestamp in plan mode.
- Prevented premature MCP discovery completion.
- Fixed agent loader error formatting when paths are empty.
Improvements
- Parallelized and backgrounded startup cleanup tasks for performance.
- Agents are now disabled by default (though this was later reverted).
- Backgrounded the IDE client to speed up initialization.
- Optimized --version startup time.
- Improved remote agent streaming UI and UX.
- Updated the documentation link for `/mcp refresh` to `/mcp reload`.