v0.37.0
Gemini CLISummary
This release introduces significant new capabilities, including experiment-gated support for gemini flash 3.1 lite and new skills for code duplication checking and automated failure replication. Numerous bug fixes stabilize the CLI experience, particularly around browser agent security, TTY handling in headless environments, and improved error reporting for agent loading.
New Features
- Added support for experiment-gated use of gemini flash 3.1 lite.
- Implemented forbiddenPaths configuration for OS-specific sandbox managers.
- Introduced a new skill specifically designed to look for duplicated code during PR reviews.
- Enabled support for inline agentCardJson configuration for remote agents.
- Added a 'ci' skill to facilitate automated failure replication.
- Implemented Windows sandbox dynamic expansion (Phase 1 and 2.1).
- Added dynamic discovery of read-only tools for the browser agent.
- Added reliability harvester and 500/503 retry support for evaluations.
- Added development worker infrastructure for GCP integration.
- Enabled dynamic Linux sandbox expansion and worktree support.
- Added maxActionsPerTask setting for the browser agent.
Bug Fixes
- Fixed an issue where the auto-update spinner was shown even when no update was in progress.
- Stabilized copy mode to prevent flickering and cursor resets in the CLI.
- Fixed flaky ctrl-c-exit test by moving it to the non-blocking test suite.
- Fixed an issue where the shell tool would conditionally expose additional_permissions incorrectly.
- Prevented over-eager slash subcommand completion in the UI.
- Fixed dynamic model routing for gemini 3.1 pro to correctly use the customtools model.
- Fixed console log/info messages appearing in headless mode.
- Fixed browser agent bypassing domain restrictions by correctly detecting embedded URLs in query parameters.
- Fixed browser agent's domain restriction system prompt to include proxy bypass constraints.
- Fixed policy issue where write_file argsPattern in plan mode incorrectly required the session ID.
- Fixed an issue where attachments were not allowed by adding a required permission prompt.
- Threaded AbortSignal to chat compression requests to prevent hangs.
- Resolved TTY hang in headless environments by unconditionally resuming process.stdin before React Ink launch.
- Fixed UI composer by cleaning up estimated string length hacks.
- Prevented the escape key from cancelling ongoing requests while in shell mode.
- Fixed CLI to show helpful guidance when no skills are available.
- Fixed chat logs and errors to correctly handle tail tool calls.
- Fixed CLI to prioritize primary name matches when searching for slash commands.
- Fixed browser agent input blocker to remain active across page navigations.
- Fixed plan mode to switch the model back to a flash model after exiting plan mode.
- Fixed core logic to strip leading underscores from error types reported in telemetry.
- Fixed an issue where non-interactive ASK_USER denials were hardcoded instead of relying on explicit policy rules.
Improvements
- Centralized test agents into test-utils for reuse.
- Reverted disabling agents by default.
- Updated telemetry attribute keys and added timestamps for planning operations.
- Improved agent loader error formatting when paths are empty.
- Refined onboarding metrics to explicitly log duration and use the tier name.
- Reduced logging verbosity by adding toJSON methods to tools and invocations.
- Standardized OS-specific sandbox tests and extracted Linux helper methods.
- Fixed grammar in CONTRIBUTING documentation and numbering in sandbox documentation.
- Added documentation notes regarding root privileges in sandbox documentation.
- Documented agent_card_json string literal options for remote agents.
- Cleaned up UI composer by removing estimated string length hacks.
- Clarified policy requirement for general.plan.directory in the settings schema.
- Merged examples of use into quickstart documentation.