Mika 0.1.29

变化 / What's new

• 长任务更不容易被上下文挤爆，也不再白白浪费空间。 什么时候压缩历史改为按 token 算，并用服务商报的真实用量校准：大窗口的模型能用上更多空间，小窗口的本地模型也不会每一步都在压缩。
• Claude 新模型上任务不再中途被拒。 新的 Claude 模型会校验回传的思考内容，Mika 展开工具、压缩历史后会被拒绝；现在照原样回传，被拒时自动去掉思考内容重发一次。
• 找不到模型时如实转达服务端的说法。 自建网关、中转服务回 404 时，提示里带上服务端自己的原话，不再一律说「模型名不对」。
• 引用和附件不再干扰回答语言。 引用 Mika 之前的英文回答、再用中文说一句时，回答不会再被误判语言而退回改写；引用的 Mika 原话也不会被当成你说过的话记进记忆。
• 带截屏或网页说明的插话只给撤回，不再在输入框里露出内部标记。
• 配对失效时，通知里也能直接填配对码。
• 有的系统不让建「受信任的后台屏」时，退回普通后台屏继续用（后台操作可能影响主屏焦点，自检里会如实提示）。

感谢 @JoeyWu-tech、@ChasePassion、@johnhldiver-beep 的贡献！

• Long tasks fit the context better. When to compact history is now measured in tokens and calibrated with the provider's real usage: large-window models use more of their space, and small-window local models no longer compact on every step.
• Tasks on newer Claude models no longer get rejected midway. Newer Claude models verify returned thinking blocks, which failed after Mika expanded tools or compacted history; blocks are now returned exactly as received, and if rejected Mika drops them and retries once.
• "Model not found" now relays the server's own message, e.g. from self-hosted gateways, instead of always blaming the model name.
• Quotes and attachments no longer confuse the reply language, and a quoted Mika answer is no longer remembered as something you said.
• Steering messages that carry a screenshot or page context can be withdrawn but not edited, so internal markers never show up in the input box.
• When pairing has expired, the notification lets you type the pairing code directly.
• On systems that don't allow a trusted background screen, Mika falls back to a regular one (it may affect focus on your main screen; the self-check says so).

Thanks to @JoeyWu-tech, @ChasePassion and @johnhldiver-beep for their contributions!

从旧版本直接覆盖安装即可，对话和设置都保留；也可以在 关于 › 检查更新 里直接下载安装。Install over the previous version; your chats and settings are kept. You can also update from About › Check for updates.