GoAgent v0.4.1
GoAgent v0.4.1 is a hotfix release for v0.4.0. It unblocks freeform teacher chat that was raising "棋盘图证据不完整" for prompts the backend classifier interpreted as vision-required, and gives KataGo whole-game analysis a larger timeout budget so heavy networks (e.g. the bundled zhizi b28) can finish a long game without aborting partway through.
QQ 群:1030632742,欢迎一起交流、提建议、完善 GoAgent。
GoAgent-0.4.1-win-x64-nvidia-portable.zip is not attached to this release because its packaged size exceeds GitHub's 2 GB single-asset upload limit; if you need an NVIDIA portable build, run pnpm dist:local:win locally.
中文
下载前先选版本
| 平台 / 场景 |
推荐下载 |
| macOS Apple Silicon |
GoAgent-0.4.1-mac-arm64.dmg |
| macOS Intel |
GoAgent-0.4.1-mac-x64.dmg |
| Windows 普通版,OpenCL 推荐包 |
GoAgent-0.4.1-win-x64.exe 或 GoAgent-0.4.1-win-x64-portable.zip |
| Windows NVIDIA 专版 |
GoAgent-0.4.1-win-x64-nvidia.exe(GoAgent-0.4.1-win-x64-nvidia-portable.zip 超出 GitHub 2 GB 上传限制,请用 pnpm dist:local:win 本机打包) |
| 校验文件 |
SHA256SUMS.txt |
本版重点
- 修复手动给老师输入问题时,遇到「为什么这里」「这里不好」「分析一下」等关键词被分类器错判为需要棋盘图的任务,从而抛出「棋盘图证据不完整」错误而无法发送。现在棋盘图是否必需仅由前端显式
request.mode 决定(current-move / move-range),freeform 聊天即使分类器推断出 current-move 也不再阻断,agent 收到 boardImageAttached=false 的上下文后会改用 SGF / KataGo 工具回答。
- 修复 KataGo 整盘胜率分析跑到一半超时、曲线只画出前几手的问题。把单局查询预算从 2.5 秒/手提到 5 秒/手(最低 180 秒),让 zhizi b28 等重网络也有足够时间跑完一整盘。
繁體中文
下載前先選版本
| 平台 / 使用情境 |
建議下載 |
| macOS Apple Silicon |
GoAgent-0.4.1-mac-arm64.dmg |
| macOS Intel |
GoAgent-0.4.1-mac-x64.dmg |
| Windows 一般版,OpenCL 推薦包 |
GoAgent-0.4.1-win-x64.exe 或 GoAgent-0.4.1-win-x64-portable.zip |
| Windows NVIDIA 專版 |
GoAgent-0.4.1-win-x64-nvidia.exe(GoAgent-0.4.1-win-x64-nvidia-portable.zip 因 GitHub 單檔 2 GB 上限未上傳) |
| 校驗檔 |
SHA256SUMS.txt |
本版重點
- 修復手動向老師輸入問題時,因關鍵詞被分類器判定為需要棋盤圖而拋出「棋盤圖證據不完整」的問題。
- 修復整盤胜率分析跑到一半超時的問題,分析預算提升到 5 秒/手。
English
Pick the right package before downloading
| Platform / use case |
Recommended download |
| macOS Apple Silicon |
GoAgent-0.4.1-mac-arm64.dmg |
| macOS Intel |
GoAgent-0.4.1-mac-x64.dmg |
| Standard Windows x64, OpenCL recommended |
GoAgent-0.4.1-win-x64.exe or GoAgent-0.4.1-win-x64-portable.zip |
| Windows NVIDIA edition |
GoAgent-0.4.1-win-x64-nvidia.exe (the GoAgent-0.4.1-win-x64-nvidia-portable.zip artifact exceeds GitHub's 2 GB per-asset limit; build it locally with pnpm dist:local:win) |
| Checksums |
SHA256SUMS.txt |
Why update
- Manual freeform prompts to the teacher no longer raise "棋盘图证据不完整 (board image evidence incomplete)" when the question matches keywords like
为什么这里 / 这里不好 / 分析一下 that the backend classifier interpreted as a vision-required intent. Image requirement is now driven only by the renderer's explicit request.mode; freeform chats route as freeform regardless of inferred intent, and the agent receives the same boardImageAttached=false context so it falls back to SGF / KataGo tools instead of refusing.
- KataGo whole-game analysis no longer times out partway through long games. The per-query budget is raised from 2.5 s to 5 s (and the floor from 120 s to 180 s), giving the bundled zhizi b28 network enough room to finish a 200-move sweep without leaving the winrate curve stuck on the opening moves.
日本語
ダウンロード前に選ぶもの
| 環境 |
推奨ファイル |
| macOS Apple Silicon |
GoAgent-0.4.1-mac-arm64.dmg |
| macOS Intel |
GoAgent-0.4.1-mac-x64.dmg |
| Windows 標準版、OpenCL 推奨 |
GoAgent-0.4.1-win-x64.exe または GoAgent-0.4.1-win-x64-portable.zip |
| Windows NVIDIA 版 |
GoAgent-0.4.1-win-x64-nvidia.exe(GoAgent-0.4.1-win-x64-nvidia-portable.zip は GitHub の 2 GB 制限を超えるため未配布) |
| チェックサム |
SHA256SUMS.txt |
主な変更
- フリーフォームの質問が分類器によって画像必須と誤判定され、「棋盘图证据不完整」で送信できなくなる不具合を修正しました。
- 整局勝率分析が途中でタイムアウトする不具合を修正し、1 手あたりの予算を 5 秒に拡大しました。
한국어
다운로드 전 선택
| 환경 |
권장 다운로드 |
| macOS Apple Silicon |
GoAgent-0.4.1-mac-arm64.dmg |
| macOS Intel |
GoAgent-0.4.1-mac-x64.dmg |
| Windows 표준 x64, OpenCL 권장 |
GoAgent-0.4.1-win-x64.exe 또는 GoAgent-0.4.1-win-x64-portable.zip |
| Windows NVIDIA 에디션 |
GoAgent-0.4.1-win-x64-nvidia.exe(GoAgent-0.4.1-win-x64-nvidia-portable.zip은 GitHub의 2GB 단일 자산 한도를 초과하여 업로드되지 않았습니다) |
| 체크섬 |
SHA256SUMS.txt |
이번 버전
- 사용자가 자유롭게 입력한 질문이 분류기에 의해 이미지 필수 작업으로 잘못 분류되어 "棋盘图证据不完整" 오류가 발생하던 문제를 수정했습니다.
- 전체 대국 승률 분석이 도중에 타임아웃되는 문제를 수정하고 수당 분석 예산을 5초로 늘렸습니다.
ภาษาไทย
เลือกไฟล์ก่อนดาวน์โหลด
| แพลตฟอร์ม |
ไฟล์ที่แนะนำ |
| macOS Apple Silicon |
GoAgent-0.4.1-mac-arm64.dmg |
| macOS Intel |
GoAgent-0.4.1-mac-x64.dmg |
| Windows x64 มาตรฐาน แนะนำ OpenCL |
GoAgent-0.4.1-win-x64.exe หรือ GoAgent-0.4.1-win-x64-portable.zip |
| Windows รุ่น NVIDIA |
GoAgent-0.4.1-win-x64-nvidia.exe (ไฟล์ GoAgent-0.4.1-win-x64-nvidia-portable.zip ใหญ่เกินขีดจำกัด 2 GB ของ GitHub จึงไม่ถูกอัปโหลด) |
| Checksums |
SHA256SUMS.txt |
จุดสำคัญของรุ่นนี้
- แก้ปัญหาที่คำถามอิสระบางคำถูกตีความว่าเป็นงานที่ต้องใช้ภาพกระดาน จนทำให้แสดงข้อผิดพลาด "棋盘图证据不完整"
- แก้ปัญหาที่การวิเคราะห์อัตราชนะตลอดทั้งเกมหมดเวลากลางทาง โดยขยายงบประมาณต่อหมากจาก 2.5 วินาทีเป็น 5 วินาที
Tiếng Việt
Chọn gói tải xuống
| Nền tảng |
Gói khuyến nghị |
| macOS Apple Silicon |
GoAgent-0.4.1-mac-arm64.dmg |
| macOS Intel |
GoAgent-0.4.1-mac-x64.dmg |
| Windows x64 tiêu chuẩn, khuyến nghị OpenCL |
GoAgent-0.4.1-win-x64.exe hoặc GoAgent-0.4.1-win-x64-portable.zip |
| Windows phiên bản NVIDIA |
GoAgent-0.4.1-win-x64-nvidia.exe (gói GoAgent-0.4.1-win-x64-nvidia-portable.zip vượt giới hạn 2 GB của GitHub nên chưa được tải lên) |
| Checksums |
SHA256SUMS.txt |
Điểm mới
- Khắc phục lỗi khi câu hỏi tự do bị bộ phân loại hiểu là tác vụ cần ảnh bàn cờ, dẫn đến lỗi "棋盘图证据不完整".
- Khắc phục lỗi phân tích winrate cả ván bị hết thời gian giữa chừng, tăng ngân sách mỗi nước lên 5 giây.
Quality baseline
This release keeps the existing top-quality baseline: grounded shape recognition engine, local pattern matcher, knowledge source-policy gates, optimized move-range review, quality checks and eval gates, Real Eval / engine silver fixture gate, KataGo engine pool telemetry, Release artifact smoke, student level, student age, teacher persona style settings with evidence boundary, teacher sessions, selective PR #6 integration, strict selected-provider TTS, offline synthesis validation for Kokoro, Vision Evidence Chain, KataGo Trace Translator, Volcengine / Doubao TTS, and multilingual release guidance.
Thanks to layiku and wimi321.