1.0.0-beta.6

published 11 September 2026

• Added computer use behind `MERCURY_COMPUTER_USE`: the model takes screenshots and drives the mouse and keyboard on every provider route that carries images; the first act in each application asks by name, Esc stops it, one session drives at a time, and screenshots never enter the saved conversation
• Added stop, pause and kill controls to the workflows views, from any session and after a compaction or a clear
• Changed the cockpit on a small or resized terminal to a compact layout: the session's critter and facts in a band at the top, the chat takes the space, one line under the composer for the sessions and agents, the Session Concourse as a split view, and no minimum window size
• Changed rarely used tools to load on demand; a first request carries a third less definition text, and on the OpenAI route a third of its former bytes
• Changed the child-memory guard to be on by default and to park a session instead of killing it
• Changed the model's guidance to batch independent reads
• Fixed preserved reasoning being dropped after an idle hour, a large prune, a resume, a credential refresh, a reconnect, or a mid-conversation tool mount
• Fixed the streamed reply hiding for a moment when its reasoning lands before its words
• Fixed a finished agent's report arriving more than once
• Fixed a shell command with an ampersand and a glob ending with exit 127
• Fixed planning without an interactive approval channel trapping a headless run; it is refused instead
• Fixed tool descriptions directing a sub-agent to a tool it does not have
• Fixed the Edit tool refusing an edit from a model that sends every field, empty ones included
• Fixed a spurious input-stream error shown while recording voice
• Fixed the doctor's bug-report link
• Removed internal checks that no longer applied