Skip to content

1.0.31

Choose a tag to compare

View all tags
@deskwand deskwand released this 23 Aug 12:51
· 6 commits to main since this release
1.0.31
4a78d07

🎉 1.0.31

中文

✨ 新功能

  • 用户级 AGENTS.md 规则: 新增对用户级 AGENTS.md 规则的支持，并内置可视化设置编辑器，让你无需手动修改文件即可为 Agent 配置全局规则。

  • 实用模型运行环境: 新增顶层 utilityRuntime 配置，用于会话标题生成与记忆 LLM 提取，替代旧的 memoryRuntime.llm。API 设置新增「实用模型」标签页，支持继承开关与提供方/模型下拉选择。

  • 充值记录查看: 新增充值历史记录视图与分页浏览，方便回顾每一笔充值明细。

  • UI 字体大小可调: 新增界面字号设置（默认 14，范围 12–20，步进 1），带数字输入框、步进器与 PX 单位。调整后通过根字号与 --ui-font-scale 变量全局缩放所有 rem 文本，并持久化到现有设置链路。

  • DeepSeek 视觉模型预设: 为 DeepSeek 提供方预设新增 deepseek-v4-flash-vision-exp 模型，可直接选用具备视觉能力的模型。

🐛 修复

  • 图片直传视觉模型: 修复附件图片未能直接发送给视觉模型的问题，现在图片会作为图像输入直达具备视觉能力的模型。

  • 逐条分发助手消息: 修复同一回合内多次工具调用产生多条助手消息时，因共享同一幂等键只有首条被投递的问题。现在每条回合消息都会作为独立的频道回复送达。

  • 目标用时统计: 修正目标跨暂停/恢复/重启时的耗时计算，避免暂停与离线时间被重复计入，并在各边界正确记录时间。

  • 自动技能学习默认关闭: 自动技能学习功能默认改为关闭，避免无意中触发学习消耗。

  • 全局记忆开关默认关闭: 全局记忆开关默认改为关闭，降低误开启带来的资源消耗。

  • 聊天输入禁用拼写检查: 聊天输入框禁用拼写检查，避免中文等场景下的拼写红波干扰输入体验。

♻️ 重构

  • 记忆设置页精简: 将记忆设置页简化为「开关」与「清空」两个入口，移除调试、嵌入与评估相关代码和接口，保留搜索/读取、摄入与存储等核心能力。

🔧 工程

  • 版本号更新至 1.0.31，并同步构建各平台更新通道产物。

English

✨ New Features

  • User-level AGENTS.md rules: Added support for user-level AGENTS.md rules along with a visual settings editor, so you can configure global rules for your Agent without manually editing files.

  • Utility model runtime: Introduced a top-level utilityRuntime config used for session title generation and memory LLM extraction, replacing the old memoryRuntime.llm. Added a "Utility Model" tab in API settings with an inherit toggle and provider/model dropdowns.

  • Top-up history view: Added a top-up history records view with pagination, making it easy to review every recharge entry.

  • Adjustable UI font size: Added a UI font size setting (default 14, range 12–20, step 1) with a numeric input, stepper, and PX unit control. Adjusting the value scales all rem-based text globally via the root font-size and --ui-font-scale, persisted through the existing settings chain.

  • DeepSeek vision model preset: Added the deepseek-v4-flash-vision-exp model to the DeepSeek provider preset so you can pick a vision-capable model.

🐛 Bug Fixes

  • Images sent directly to vision models: Fixed attached images not reaching vision-capable models — images are now passed as image input straight to models that support vision.

  • Per-message delivery for assistant turns: Fixed a bug where multiple assistant messages from multi-step tool use in a single turn shared one idempotency key, so only the first was delivered. Now every round is delivered as its own channel reply.

  • Goal elapsed-time accounting: Fixed elapsed-time calculation across pause/resume/restart so paused and offline time are no longer double-counted, with proper checkpoints at each boundary.

  • Auto skill learning off by default: Auto skill learning now defaults to disabled to avoid triggering unintended learning costs.

  • Global memory switch off by default: The global memory switch now defaults to disabled to reduce resource usage from accidental toggles.

  • Disabled spellcheck on chat input: Disabled spellcheck on the chat input textarea to avoid red squiggly interference, especially with Chinese text.

♻️ Refactoring

  • Simplified memory settings: Simplified the memory settings page to a toggle and a "clear all" button; removed debug, embedding and evaluation code and IPC while keeping search/read, ingestion, and stores.

🔧 Engineering

  • Version bumped to 1.0.31, with update-channel artifacts built for all platforms.

68 files changed · +2833 −4458
Full Changelog: 1.0.29...1.0.31

Assets 6
Loading