繁體中文

新功能

移除 AI 人才市場面板,簡化安裝與發版流程

挑選現成的 AI 角色卡,不再需要獨立面板。

專案正式移除了 AI 人才市場面板。該面板原本用來瀏覽精選的 subagent(子代理)角色卡並安裝至 ~/.claude/agents/ 目錄。隨著 Claude Code 內建角色選擇,加上 GPT Store 示範了平台自帶功能後的目錄命運,三篇論文也證實角色前綴會削弱推理。該面板還夾帶私有專案編譯出的唯一外部二進位檔 vendor/instate-cli,發版密鑰 INSTATE_CLI_TOKEN 曾兩度過期,悄悄出貨少了 23 MB 且面板全空的程式。移除該功能一併拆掉了校驗檔、指紋防呆、發版上傳下載流程,以及面板一致性與主題計數腳本中專屬 macOS 的特例。

新功能

移除 AI 圓桌討論視窗,回歸專注保護用量額度

想並排比較不同 AI 的回答,交給兩個終端機視窗處理得更快。專案正式移除了 AI 圓桌討論視窗。這項功能過去會在獨立視窗中,讓 Claude Code、Codex 與 Antigravity 進行多輪討論,開始前提供 token(AI 處理與計費的文字單位)預估、過程中計票共識,並支援唯讀資料夾存取。它是專案最大的子系統,在 discussion/ 目錄佔了約 3,600 行程式碼。但它產出的成果,只是由兩款工具並排回答同一題且一次結束,開兩個終端機視窗早就更快。用量監控工具不該拿來跑對話,花掉的額度正是它存在所要保護的額度。若日後需要精簡版本,零件都仍完整留在 git 歷史紀錄中。

English

FEATURE

Remove AI Talent Market panel and simplify installation and release workflows

Choosing ready-made AI persona cards no longer requires a dedicated panel.

The project officially removed the AI Talent Market panel. The panel was originally used to browse curated subagent persona cards and install them into the ~/.claude/agents/ directory. As Claude Code built persona selection directly into the platform, and the GPT Store demonstrated the fate of catalogs once the host platform ships its own features, three papers also confirmed that persona prefixes degrade reasoning. The panel also carried the project's only external binary compiled from a private project, vendor/instate-cli; its release secret INSTATE_CLI_TOKEN expired twice, silently shipping an app that was 23 MB lighter with an empty panel. Removing this feature also eliminated the checksum file, fingerprint guards, release upload and download workflows, and macOS-specific carve-outs in the panel parity and theme-count scripts.

FEATURE

Remove AI Council window and refocus on protecting usage quota

Comparing answers from different AI tools side by side is much faster when left to two terminal windows. The project officially removed the AI Council window. This feature previously ran multi-round discussions between Claude Code, Codex, and Antigravity in a separate window, providing upfront token estimates, consensus tallying during the process, and read-only folder access. It was the project's largest subsystem, accounting for roughly 3,600 lines of code across the discussion/ directory. Yet the output it produced was merely two tools answering the same question side by side in a single pass, which opening two terminal windows already achieved faster. A usage monitor should not be used to run conversations; the quota it consumed was the exact quota it exists to protect. If a stripped-down version is ever needed in the future, the parts remain fully intact in git history.

原始 CHANGELOG

Original changelog移除 AI 人才市場面板,簡化安裝與發版流程
**AI Talent Market.** The panel browsed a curated set of subagent personas and installed them into `~/.claude/agents/`. Claude Code has since built persona selection in itself, and the evidence against the pattern accumulated from the other side: the GPT Store showed what happens to a persona catalogue once the host platform ships its own, and three papers now report that a persona prefix measurably degrades a model's reasoning rather than sharpening it. The panel also carried the project's only vendored binary — `vendor/instate-cli`, built from a separate private project, fetched during release through a `INSTATE_CLI_TOKEN` secret that expired twice and silently shipped a 23 MB-lighter app with an empty panel. Removing the feature removes that whole chain: the checksum file, the fingerprint guard, the release upload/download dance, and the macOS-only carve-outs in the panel parity and theme-count scripts.
Original changelog移除 AI 圓桌討論視窗,回歸專注保護用量額度
**AI Council.** A separate window ran a multi-round discussion between Claude Code, Codex and Antigravity, with a token estimate up front, consensus tallying, and read-only folder access. It was the largest subsystem in the project at roughly 3,600 lines across `discussion/`, and the thing it produced — the same question answered by two tools, side by side, ending in one pass — is something two terminal windows already do faster. A usage monitor is the wrong place to run conversations; the quota it spent was the quota it exists to protect. The parts are in git history if a stripped-down version is ever wanted.