v0.38.0
Gemini CLISummary
This release introduces several new features, including enabling 'TerminalBuffer' mode to reduce flicker, adding support for default environment variable values, and implementing a background memory service for skill extraction. Numerous bug fixes address stability issues across the CLI and core, particularly around sandbox reliability on Linux and Windows, and improved chat history recording. Users will also benefit from UI refinements like compact tool output being enabled by default and better handling of input prompts.
New Features
- Enabled 'TerminalBuffer' mode to solve flicker in the UI.
- Added support for default values for environment variables in the CLI.
- Introduced a background memory service for skill extraction.
- Implemented context-aware persistent policy approvals.
- Enabled compact tool output by default in the CLI.
- Added a scrollbar for the input prompt.
- Implemented background process monitoring and inspection tools.
- Added an experimental setting: experimental.adk.agentSessionNoninteractiveEnabled.
- Enabled the 'Other' option for the yes/no question type.
- Introduced ContextCompressionService in the core.
- Scoped subagent workspace directories via AsyncLocalStorage.
- Implemented a high-signal PR regression check for evaluations.
- Migrated nonInteractiveCli to LegacyAgentSession.
- Added minimalist sandbox status labels in the CLI.
- Enabled browser agent metrics.
- Prioritized summary for topics in core and CLI.
- Changed default loadingPhrases to 'off' to hide tips in the CLI.
Bug Fixes
- Fixed crash when the vim editor is not found in PATH on Windows.
- Fixed unwanted vertical spacing around various tool output treatments in the UI.
- Fixed crash in AnsiOutputText when handling non-array data.
- Fixed shell output display issues.
- Fixed table styling in the UI.
- Fixed restoration of topic headers.
- Fixed crash related to ARG_MAX when refactoring the Linux sandbox.
- Fixed issue where complete_task tool calls were not being recorded in chat history.
- Fixed agent stopping when all declinable tools are cancelled.
- Fixed issue where stale retry/loading state was not cleared after cancellation.
- Fixed issue where unsafe type assertions occurred in the Core File System.
- Fixed issue where model quota was incorrectly shown in /stats; refactored quota display.
- Fixed broken PowerShell translation and native __write issues in the Windows sandbox.
- Fixed issue where sandbox approvals were not correctly persisted and matched for proactive expansions.
- Fixed issue where stale slash command list was shown after /skills reload in the CLI.
- Fixed issue where sandbox usability was poor and build errors occurred.
- Fixed issue where computer-use model detection failed for analyze_screenshot in the browser component.
- Fixed handling of all InvalidStreamError types gracefully in prompts.
- Fixed issue where skill system instructions were not injected into subagent prompts if activated.
- Improved Windows sandbox reliability and fixed related integration tests.
- Fixed issue where the global temp directory was not correctly moved under the tmp directory.
Improvements
- Updated README.md for links.
- Refactored Linux sandbox to fix ARG_MAX crashes.
- Enhanced sandbox usability and fixed build errors.
- Optimized Terminal Serializer.
- Auto configured memory.
- Updated ink version to 6.6.7.
- Updated stale browser agent documentation.
- Moved project memory dir under tmp directory.
- Clarified release coordination documentation.
- Added instructions for how to start react in prod and force react to prod mode.
- Fixed Windows CI execution and resolved exposed platform failures.
- Removed unused error variables in catch blocks.
- Explicitly allowed web_fetch in plan mode with ask_user.
- Fixed issue where the input box and footer visibility were incorrectly hidden in copy mode (revert).
- Fixed issue where non-array data caused a crash in AnsiOutputText.
- Fixed issue where sandbox configuration did not pass includeDirectories.
- Discouraged the use of the update topic tool for simple tasks.
- Hid model quota in /stats and refactored quota display.
- Moved agent disabling instructions and updated remote agent status documentation.
- Did not run pr-eval workflow when no steering changes were detected.