Available Oct 27 for Pro and Enterprise, you describe a change in plain language, the agent edits the connected repo with a live preview of your running app beside the chat, then opens a PR or MR that follows your team's skills and contribution guidelines. When a human or agent reviewer asks for changes, it makes them and updates the request, so engineers still decide what ships.
Announced Oct 26, four open Agent Skills ( iam helper for troubleshooting , policy management , policy simulator , privileged access management ) sit on read only Google managed MCP servers so Gemini CLI, Claude Code, Cursor and ADK agents can trace a 403, simulate a policy change against 90 days of access logs, or request a time bound PAM grant. Diagnostics run on their own, but any policy change pauses for human approval, and risky bindings like allUsers or roles/owner get flagged; install with npx skills add google/skills…
Released Oct 27, /diff now opens as soon as tracked changes are ready, loads untracked files and older turns in the background, and skips reading binaries or untracked files over 1 MB, while Auto mode's sandbox escape review gets more time so it asks you to approve safe commands less often. Failed or canceled tool calls now reach the agent as errors it can recover from, and forks or /btw side questions no longer inherit an active /goal they can't stop…
Released Oct 27, OpenCode lets you create, install and share experimental desktop plugins, and plugins can now add keyed session instructions that survive compaction without busting the prompt cache. Heads up: the OpenCode GitHub Action is no longer published, @opencode/ai now exports the Promise API at its root (Effect moved to @opencode/ai/effect ), and TUI plugins need OpenTUI 0.7…
Launched Oct 27, you point your agent's base URL at Governor, which picks a model and reasoning level per request from the models your team allows, holds that choice for the whole task, and answers code searches itself with file and line results. Bito reports one customer's Claude Code cost per task fell 48% ($4.12 to $2.14) with success held at 100%; access is by request…
Added Oct 26 with no BYOK proxy needed, the open weight 1M context model turns on under Cursor Settings Models, and Cursor says all four effort tiers (44.9% to 47.7% on CursorBench 4.0) sit on its cost frontier. It still trails Opus 5.5 Max's 57.8%, and as a third party model it bills outside your plan's included usage…
Spotted something we missed? Start a thread.