1.0.0-beta.3

published 6 September 2026

• Added GPT-6 Astra as a first-class model: it appears once a connected OpenAI account serves it, effort reaches max, and each turn is priced at the published rate
• Updated GPT-5.6 pricing to the current published rates, including the long-context tier
• Added a sixth effort level, ultra, above max, wherever the account serves it; a level the provider refuses is remembered and the nearest served level runs, with a note in the transcript
• Fixed a model switch on a stopped runner doing nothing; it now restarts the runner on the requested model, and an agent paused by a rate limit resumes when the window resets
• Fixed /model, /effort, /status, the strip, the picker, /config and /accounts disagreeing about the model being served
• Fixed the preserved-thinking notice appearing after a resume or restart even though the session kept its reasoning
• Changed the notice for dropped older reasoning after a model switch or a context trim to one quiet line instead of a warning on every message
• Fixed usage meters going stale without saying so; they now refresh themselves, show their age, and say when a read failed
• Fixed the Anthropic usage panel getting stuck on "Failed to load … HTTP 429"; it now honours the server's retry wait and says how long to wait
• Fixed a refused API key or token reading "Not logged in"; the error now names the setting to fix
• Fixed removing or switching an account being announced more than once, and added c on the sign-in cards to copy the code
• Added image paste in the composer on every platform (ctrl+v; alt+v on Windows) at any size; a large image is resized to the provider's limits and the composer says what it attached
• Changed Esc to interrupt the chat's own turn only: sub-agents and workflows keep running and the transcript says how many; in the Crew view x x stops one agent and r resumes it
• Added up arrow on an empty composer to take back a message sent while the turn was busy, before the model reads it; once read, the up arrow says so
• Changed the text before a tool call to a one-line working note that the final answer never restates
• Fixed deepthink and supercode glowing in a rainbow or mixed theme colours; they now glow in your critter's colour, in the composer and in the sent message
• Fixed selected text not following the view while scrolling, and the highlight painting outside the chat pane
• Fixed mouse drags repainting on every report, and made /mouse off stick across boots
• Fixed a mis-bound shortcut flashing "Unknown command" for a line you never typed, and a split key sequence leaking characters into the input
• Changed the chat's title bar to show the session's name instead of a ticking clock, and added how long your sub-agents have worked under the composer
• Fixed /rename being refused in every session; it now renames the session everywhere it is shown
• Added compaction progress in the chat (the stage, a bar, a clock); Esc cancels it cleanly, and a stalled or cancelled compaction keeps no half summary
• Fixed Mercury re-spawning or forgetting agents, workflows and shells after a compaction
• Fixed a false "rewrote sent history" warning when old tool results are trimmed to fit the window; it is now named as routine cleanup
• Changed the seat ceiling to one number everywhere: /seats, the Boot Menu and /config set it, idle agents no longer hold a seat, and the reading counts available memory
• Fixed a waiting agent's row not saying what it waits for, and a throttled agent's retries each spending their own budget
• Fixed sub-agents ending without saying how; they now report the real cause and what they wrote once, and a timeout is never blamed on you
• Fixed a background agent's completion arriving late in the chat; it now shows the moment it arrives, marked queued until the model reads it
• Fixed the workflow panel, the board, the crew view and the chip disagreeing about what is running, and reading "idle" right after a start instead of "not reported yet"
• Added "waiting for your answer" in the crew view and the tasks list for a sub-agent waiting on your permission
• Fixed a blocked team task not reading blocked on every screen, and creating a team hiding earlier tasks
• Fixed the Concourse close chord's hint going stale; it now does exactly what it says (stop, then archive, then delete), and a parked session reopens on one Enter
• Fixed a silent workflow stream waiting on the provider's guard instead of being cut at the stall budget
• Changed Sovereign mode to ask nothing: sensitive-file edits, ask rules and a server's ask ceiling run without a card, and the transcript says what would have asked
• Fixed the mode band showing a mode the runner did not hold; it now paints the real mode from the first frame, and a refused mode change reaches the screen with its reason
• Fixed a consent answer switching Apollo mode off; it now ends only through the review's approval or shift+tab, and writes outside the spec directory are refused
• Fixed Flow's safety check denying an action when it could not read its own verdict; it now keeps the evidence and says so
• Changed a Flow block in a session you are not looking at to wait for your answer there instead of being denied on the spot; a sub-agent's ask waits the same way
• Fixed a hook's deny decision leaving no audit row in the transcript
• Fixed consent cards overflowing the screen; a long edit or command is cut with a "+N more lines" line and ctrl+f shows the whole thing
• Fixed web search failing on a throttled request; it now cools down, answers a repeated query from the cache, and names what refused in one line
• Added Godot runtime pause, step and resume to drive a game frame by frame, and fixed the editor not being found where it lives and action presses not reaching the game's callbacks
• Fixed Unity test results landing where the editor could not write them; a bridged test run is no longer refused
• Fixed the interrupt not ending a running Bash command, and a request accepted with nothing arriving hanging instead of being named and ended
• Added /bug to show the exact report and file it through your own signed-in GitHub CLI, with every provider's key shape and GitHub tokens masked; /feedback opens the three-form chooser
• Added the Intel Mac build (macos-x64) to the release; mercury update names it on an Intel Mac
• Added release signing: archives are signed and verified before they publish, an unsigned build is announced once per install, and the archive carries the licence documents
• Changed mercury update to read the release channel anonymously, so no GitHub CLI or sign-in is needed; an update whose signature does not verify is refused
• Changed mercury install to put the command on your PATH once, and to say so when the version is already in place
• Fixed the doctor judging a headless run as a broken boot and reading the boot's rungs out of order
• Fixed a git init offer for the home directory, a drive root or a huge folder without a project marker; the doctor now names a home-directory repository and how to remove it
• Changed the project's .mercury folder to hold shared config only; machine-local files live in your Mercury home, and the doctor names any leftovers
• Fixed git and process spawns in the idle loop and at boot, and the daemon polling its own files instead of watching them
• Fixed a piped doctor --json printing more than one record or exiting before the reader drained
• Fixed a headless stream-json run not opening with its init line
• Fixed a fresh /logins not reaching hosted sessions until the daemon restarted
• Changed the companion's tips so every one is true today and teaches one thing