Release version 0.47.0
This release lets you add pi-ai provider models to DeepSeek from the settings page, updates Codex to 2.8.0 with the GPT-6.2 models and Claude Code to 0.109.0, keeps Claude Code's plan visible after a turn ends, makes OpenCode more reliable while it is starting, and further reduces memory use and background work.
New
- You can now set up pi-ai provider routes for DeepSeek on the settings page. Each provider's models show up as their own group in the model picker. (#1188, @ker2xu)
Improved
- Updated Codex (codex-acp) to 2.8.0 and Claude Code (claude-agent-acp) to 0.109.0. Codex now offers GPT-6.2 Sol (the new default) and GPT-6.2 Astra, and the model placeholder in the input box shows GPT-6.2. Switching a Codex conversation's model provider now shows a clear message instead of quietly starting a new conversation and losing its history.
- When OpenCode is still starting up, codeg now waits briefly and retries opening the conversation instead of showing a generic error. If it is still not ready, the original conversation is kept rather than replaced by an empty one. (#1190, @ker2xu)
- Less background work while the app is open: the "time ago" labels in the sidebar update without re-rendering the whole list, multiple desktop windows connected to a remote server no longer process each other's events, and connection keepalive checks no longer pile up when a reply is slow. (#1191, #1199, #1201, @kogekiplay)
Fixed
- Claude Code's plan checklist no longer disappears from Session Details after a turn ends or when you reopen a conversation. (#1203 reported by @xhaozhuo-eng)
- A Claude sub-agent still running in the background no longer gets cut off when the main agent ends its turn, so automated tasks wait for it to finish. (#1198 reported by @ttxhxz)
- Reduced memory retained when you scroll long conversations, close terminals, browse the pet marketplace or change custom agent icons, and after replies with large tool outputs. (#1189, #1192, #1193, #1195, #1197, #1205, @kogekiplay)
Thanks to @kogekiplay and @ker2xu for contributing to this release, and to @ttxhxz and @xhaozhuo-eng for the reports.
发布版本 0.47.0
这一版支持在设置页为 DeepSeek 添加 pi-ai 供应商的模型;Codex 更新至 2.8.0,新增 GPT-6.2 模型,Claude Code 更新至 0.109.0;Claude Code 的计划在回合结束后不再消失;OpenCode 启动期间打开会话更稳定;同时继续减少内存占用和后台开销。
新增
改进
- Codex(codex-acp)更新至 2.8.0,Claude Code(claude-agent-acp)更新至 0.109.0。 Codex 新增 GPT-6.2 Sol(新默认模型)和 GPT-6.2 Astra,输入框的模型占位提示也换成了 GPT-6.2。给 Codex 会话切换模型供应商时,现在会明确提示,不会再悄悄新开一个会话、丢掉原来的历史。
- OpenCode 还在启动时,codeg 会稍等片刻再重新打开会话,不再显示笼统的错误;如果仍未就绪,会保留原会话,不会被一个空会话替换。(#1190,@ker2xu)
- 减少应用运行时的后台开销:侧边栏的「多久以前」标签更新时不再重新渲染整个列表;多个桌面窗口连接同一远程服务器时,不再互相处理对方的事件;回复较慢时,连接保活检查也不会再越积越多。(#1191、#1199、#1201,@kogekiplay)
修复
- 修复 Claude Code 的计划清单在回合结束或重新打开会话后,从会话详情中消失的问题。(#1203 由 @xhaozhuo-eng 反馈)
- 修复主智能体结束回合时,后台仍在运行的 Claude 子智能体被一起中断的问题,自动化任务现在会等它完成。(#1198 由 @ttxhxz 反馈)
- 减少滚动长会话、关闭终端、浏览宠物市场、更换自定义智能体图标后,以及工具输出较大的回复之后残留的内存。(#1189、#1192、#1193、#1195、#1197、#1205,@kogekiplay)
感谢 @kogekiplay 和 @ker2xu 为本次发布做出的贡献,也感谢 @ttxhxz 和 @xhaozhuo-eng 的反馈。