v0.43.0-preview.0
Gemini CLISummary
This release introduces significant new capabilities, including the ability to export and import sessions, and the addition of Local and Remote Subagent Protocols. Numerous bug fixes address stability, including resolving chat corruption, preventing hangs during OAuth, and improving tool execution reliability. Users will also benefit from UX refinements like hiding read-only settings and improved handling of untrusted folders.
New Features
- Added support for exporting the current session to a file and importing it via a flag.
- Added Machine Hostname to the CLI interface display.
- Introduced LocalSubagentProtocol behind the AgentProtocol interface.
- Added RemoteSubagentProtocol behind the AgentProtocol interface.
- Enabled the model to use the edit tool specifically for surgical edits.
- Added shell command safety evaluations.
- Prefix tool call IDs with tool names to support tool rendering in ACP compliant IDEs.
Bug Fixes
- Fixed a typo related to model steering for surgical edits.
- Rejected numeric project IDs when setting GOOGLE_CLOUD_PROJECT.
- Removed unsafe type assertion suppressions in error utilities.
- Allowed redirection in YOLO and AUTO_EDIT modes without sandboxing.
- Fixed a chat corruption bug within the context manager.
- Provided JSON output for AgentExecutionStopped events in non-interactive mode.
- Handled invalid custom plans directories gracefully.
- Resolved a race condition when waiting for tool completion in the a2a-server.
- Randomized sandbox container names to prevent conflicts.
- Fixed hysteresis issues in asynchronous context management pipelines.
- Preserved the system PATH in the Git environment to resolve ENOENT errors.
- Fixed an argument mismatch in resolveClassifierModel within NumericalClassifierStrategy.
- Fixed an issue where transcribed text would not show after releasing the spacebar.
- Prevented false command conflicts when launching the tool from the home directory.
- Fixed an issue where the tool explanation was incorrectly placed in the thought stream instead of tool call content.
- Fixed an issue where the tool call streaming ID could collide during parallel tool calls.
- Threw explicit errors when tool responses were dropped.
- Resolved the "function response turn must come immediately after function call" error.
- Treated HTTP 404 responses as 405 in StreamableHTTPClientTransport for MCP.
- Prevented duplicate SessionStart systemMessage rendering.
- Prevented an infinite thought loop in ACP mode by disabling nextSpeakerCheck.
- Used a static tool name in the confirmation prompt to avoid parsing errors.
- Fixed tool turn handling in NumericalClassifierStrategy to prevent 400 Bad Request errors.
- Handled malformed projects.json files in the ProjectRegistry.
Improvements
- Clarified that Auto Memory proposes memory updates and skills.
- Tightened the private Auto Memory patch allowlist.
- Hid read-only settings scopes from the CLI.
- Improved the user experience for listing MCPs in untrusted folders.
- Prevented silent hangs during OAuth authentication on headless Linux systems.
- Fixed an issue where executable bits were not preserved for macOS binaries during CI builds.
- Migrated core tools to use the native ToolDisplay property, fixing UI rendering.
- Removed unnecessary wrapping of arguments.
- Fixed issues in triage workflows related to "Argument list too long".
- Fixed JSON parsing issues in the scheduled triage workflow.
- Hid the /memory add subcommand when memoryV2 is enabled.
- Cached model routing decisions in LocalAgentExecutor.
- Added quotes around the session ID in the resume tip for better clarity.
- Removed noisy theme registration logs from the terminal.
- Implemented codebase-aware effort level triage in CI.
- Improved the snapshotter logic for context management.
- Fixed the snapshotter model configuration.
- Allowed installing extensions from SSH repositories.
- Fixed an issue where transcribed text was not showing after releasing space.
- Added a gutter width calculation to the input prompt width calculation.