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
Added j/k navigation aliases to the non-filterable /subagents-stop selector and clarify step list while preserving text input in editor and search modes. Thanks to @magoz for #686.
Added the optional versioned fleetStatus RPC capability with bounded, current-session child roles, goals, model/effort, split token usage, elapsed timestamps, stable opaque reconciliation keys, and explicit overflow counts. Thanks to @neumie for #682.
Fixed
Enabled the advertised j/k navigation aliases after activating the persistent FleetView while leaving printable editor input untouched before activation. Thanks to @magoz for #685.
Added opt-in subagents.projectRootResolution: "git-root" so monorepos and git worktrees can keep the default nearest-root behavior unless they choose to resolve project packages and agentOverrides from the git root. Thanks to @klajdo-f for #677.
Recognized structurally compatible custom editors in FleetView focus detection, restoring FleetView arrow-key activation and navigation when a custom editor has focus. Thanks to @magoz for #679.
Scoped foreground fleet records to their originating parent session and propagated resolved model, thinking effort, and split input/output usage through live foreground controls.
Matched fleet RPC filtering to the canonical session-file identity used by live async and foreground state.
Kept pi-intercom stable IDs from leaking into child sessions and used the current intercom runtime ID for unnamed supervisor targets.
Improved acceptance policy validation errors and tool-schema guidance for invalid evidence kinds. Thanks to @atimofeev for #672.
Tolerated temporary steering inbox scan failures so pending steer requests can be retried on the next poll. Thanks to @hughcars for #670.
Retried short, zero-activity child startup exits on the same model with bounded backoff, reducing concurrent subagent launch races without replaying model or tool work. Thanks to @felipeteodorocw for #671.
Bounded streamed subagent progress snapshots so a long or deeply nested fan-out no longer emits a tool_execution_update line above the child-stdout protocol cap and gets the child killed with protocol_output_limit. Streamed onUpdate snapshots now carry compact tool-call summaries instead of the full message transcript, cap recentTools, and truncate recentOutput line length; the returned result and detached-exit recovery keep the full transcript. Thanks to @shaharmor for #680/#681.