1.0.0-beta.5

published 8 September 2026

• Changed the sign-in list to open on the right row
• Changed /model to list real models only, in the order OpenAI, Anthropic, OpenRouter, then the rest
• Changed commands run from a session to see your own environment, not Mercury's internal stamps; the proof suites refuse to start under such stamps
• Added append and section modes to the Edit tool, and Eval cells that keep their state and report every nested call when something fails
• Added a write scope for sibling worktrees through --add-dir, a reviewer agent that works on a frozen copy of a commit, and a held state for a message sent during a compaction
• Improved the checks Mercury runs on itself: a map of generated files, per-proof exit codes, a preflight for terminal captures, cost accounting for agent fleets and a readable session state
• Fixed a GPT reply cut off after a tool call breaking every later request of the session, and automatic compaction retrying a summary the provider had refused; the refusal is now reported once with the way out
• Changed the line shown when sub-agents are still running to say how to stop one