See what changed with each update. Historical notes retain their original source, including versions whose installers have been removed.
Added
Added Inspect AutomationBench integration and a dedicated squad, using official scenario state and scoring rules with explicit runtime, call and acceptance evidence.
Changed
Consolidated ten tool entries for todos, task lists/status/boards/schedules, interaction responses and workspace locations into four. Chat declarations decreased from 39 to 33 tools and Mission declarations from 39 to 34.
Updated squad-author guidance, interaction/API contracts and bilingual migration docs. Integrations referring to old tool names or separate todo permissions must adopt the new names and parameters.
Fixed
Desktop startup, return to visibility and runtime check the stable update channel and automatically download/verify new versions; installation still requires restart confirmation. Fixed lost verified-installer state after repeated checks or installation failure, and Windows exiting silently when installer launch failed.
Improved macOS installer-failure diagnostics and prevented Windows dependency-cache cleanup from timing out successfully packaged releases.
New conversations remember the most recently selected model.
Fixed extra allowlists hiding authorized general Chat and Mission tools. Search, memory, planning, scheduling and task controls are directly available; specialized extensions remain on demand.
Fixed Inspect project configuration, long completion-summary reads, streaming observation, poll-limit validation and scoring-policy checks. Each evaluation round fixes its squad source version so later samples cannot read changing configuration.