You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Added
GPT-6.1 Sol and GPT-6 Luna models: Flex orgs can now pick GPT-6.1 Sol and GPT-6 Luna in the model picker, and Unmetered orgs can pick GPT-6 Luna. Default models are unchanged.
Safer Apex by default: A new Apex Safe Defaults rule, on by default under Toolkit → Rules, steers generated Apex toward with sharing, user mode, and bounded queries. When the agent does write privileged or unbounded Apex, its reply opens with a warning naming the construct and why.
Clearer "Chat unavailable" notice for unlicensed free-tier orgs: When daily usage runs out, the notice shows when it resets and offers View access requirements, which opens your org's live license checklist. The limit notices now sit directly above the chat box and apply to every chat tab.
Fixed
MCP servers no longer stay on "Connecting…" with non-Claude models: A server whose connection fails now shows an error, or "Setup Required" if your org hasn't enabled it, instead of waiting until you reload.
Model picker with a single model: When your org has only one available model, the picker shows it without opening an empty menu.
Chat stays responsive after sending a plan: You can keep typing and use Stop while the agent prepares for a plan that needs a different model. A message sent after Stop is held and sent automatically once the agent is ready.
Experimental
Orchestrate shows failures instead of stalling silently (Experimental): A run that fails to start or stops partway, and any board action that fails, now show an error naming what to fix.
Accept works reliably in Orchestrate (Experimental): Accept and Accept All no longer do nothing after a refusal, a disconnect, or a closed review surface, and a refused task can be retried while the rest land.
Orchestrate can plan edits to org-only metadata (Experimental): A plan that edits a component that exists only in your org, such as a standard page layout, now fetches it from the org and reaches Approve.
Orchestrate offers to install missing tooling (Experimental): When an offline check can't find its tooling, you can review the exact install command, approve it, and the original check retries without rebuilding.
Longer, more reliable verification in Orchestrate (Experimental): Each verification attempt can take up to ten minutes, so larger applications return a verdict. An optional requirement-based verification, off by default, keeps its state across Stop and reload.
Orchestrate recovers after model or entitlement changes (Experimental): Tasks no longer time out after your default model or org entitlements change.
Orchestrate on Windows (Experimental): Acceptance checks and deployments with quoted globs or paths now run as written, and Apex tasks are no longer blocked by long Windows paths.
Clearer Orchestrate progress (Experimental): Plan updates are easier to follow, and duplicate background task progress is hidden when the plan already shows it.