Usage 0.29.0

Ask one AI and you get one opinion. The new AI Council opens a menu-bar window where Claude Code, Codex and Antigravity discuss a topic you set, in rounds. You choose which ones take part and what model each runs (Opus, Sonnet or Haiku for Claude; Terra, Luna or Sol for Codex; a choice of Gemini models for Antigravity), set anywhere from one to five rounds, and can nominate one participant to act as moderator and write the closing summary. Weighing up whether to switch databases? Put the question in and let the three of them push back on each other. Once a discussion starts, the setup panel collapses into a status bar so the transcript and summary get the full window height.

Letting three AIs argue for five rounds sounds fun, but the bill does not grow in a straight line. The estimate shown before you start now reports a token total — what you are actually billed on — rather than only a count of CLI calls. Push the round count to three or more and the figure turns amber, with a plain comparison against what two rounds would cost. The reason is simple: every extra round has each participant re-read everything said so far, and the transcript only gets thicker. Photocopying the minutes of a meeting works the same way — each additional session means more pages next time, not one more sheet.

Describing a broken layout in words defeats even the person looking at it. You can now drag images straight into the topic box or paste them with Cmd+V, and every participating CLI gets to read them. You can also attach a working directory so they can consult the real files in your project before speaking, instead of guessing at what your code looks like. That directory is read-only: they can see it, they cannot change it. It is like handing out the mockups and a full case file in a meeting room — the material is open for discussion, but nobody walks off having edited the original.