Usage 0.29.1

同一場 Council（讓多個 AI 分座討論的功能）裡，現在能替每個座位挑不同專長。想請一位看合約、一位找網頁問題、另一位做安全檢查時，模型選擇器旁的角色選擇器會列出已安裝人才包裡的專家。角色（persona，先替 AI 戴上的專業帽子）會以提示詞前綴加入，所以同一個 CLI（從命令列啟動的工具）可在兩張椅子分別扮演兩種人。角色會按人才包分組並可收合，五十個選項也不會把畫面捲到找不到底。角色文字則在執行時向 talent-market CLI 取得，不會塞進 usage 本身；像替一場會議現場選角，每張椅子都能坐上合適的專家。

一群 AI 反覆討論時，你可以先決定這場會議要當腦力激盪，還是挑錯大會。Council 從第二輪起可替每位參與者從五種辯論風格中選一種：找共識、挑漏洞、補缺口、蘇格拉底式追問假設，或唱反調的魔鬼代言人。無論選哪種，每人回應開頭的立場標籤都會保留，並統計在摘要上方。若打開 end on consensus（全員同意便停止的開關），某一輪人人標示同意時，討論會立刻結束，不會耗完原先預留輪數；只要有人沒完成該輪，便不算共識。像會議主席先訂好規則、再看全體舉牌，牌沒全舉起就不宣布散會。

要主持人替多個 AI 的意見下結論時，現在它先看不到發言者姓名。交給主持人的逐字稿（transcript，完整保留每句話的討論記錄）已匿名化，避免它偏愛自己那個 CLI（從命令列啟動的工具）寫出的答案。摘要送回你的畫面或匯出成 Markdown（能直接排版的純文字格式）前，名字會恢復，所以你仍會讀到「Claude 主張……」，而非「參與者 A」。只有說話者的標記被遮住，原話不會改動；即使討論題目正是 Claude 或 Codex，內容也照樣看得懂。這像盲飲評酒：評審先喝內容、不看酒標，公布結果時才告訴你是哪一瓶。

In one Council (a feature where several AIs discuss an issue from separate seats), each seat can now take a different specialty. Need one person to review a contract, one to spot front-end trouble, and another to audit security? A role picker beside the model picker lists specialists from your installed talent packs. A persona (an expert hat placed on an AI before it speaks) is injected as a prompt prefix, so the same CLI (a tool launched from a command line) can wear two different hats in two seats. Roles are grouped by talent pack in a collapsible picker, keeping even 50 choices manageable. Their text is fetched from the talent-market CLI at runtime instead of bundled into usage: like casting a meeting, each chair gets the right specialist.

When a group of AIs keeps debating, you can now decide whether the meeting should brainstorm together or hunt for flaws. From round two onward, Council lets each participant take one of five debate styles: constructive consensus-seeking, adversarial fault-finding, collaborative gap-filling, Socratic questioning of assumptions, or devil's advocate. Whatever style you choose, the stance label opening each reply stays visible and is counted above the summary. Enable end on consensus (a switch that stops once everyone agrees), and the discussion ends as soon as every participant marks agreement in a round instead of using all budgeted rounds. If anyone fails to finish that round, it does not count. Like a chair setting meeting rules, then checking every raised hand before calling the room settled.

When a moderator is asked to draw a conclusion from several AIs, it now sees the opinions before it sees their authors. The transcript (a record preserving every spoken line) handed to the moderator is anonymized, so it cannot favor answers written by its own CLI (a tool launched from a command line). Names return before the summary reaches your screen or is exported as Markdown (plain text that can carry formatting), so you still read “Claude argued...” rather than “Participant A.” Only author labels are hidden; the words themselves remain untouched, even when the discussion is about Claude or Codex. It is like a blind wine tasting: judges taste what is in the glass before learning its label.