v1.4.0-alpha.20

从 1.3.X 版本更新之前务必备份数据库，自动迁移可能失效！！！

新特性

N/A

修复 & 改进

• core: 修复对话系统的一批问题（#886）
• core: 稳定 prompt 模板中 {pick(...)} 的选择结果。（#878）
• core: 修复压缩 prompt 的 token 预算统计（#883）
• core: 修复 Agent Prompt 创建时可能触发的 prompt injection warning（#880）
• extension-agent: 重构并简化 agent 扩展实现（#881）
• shared-adapter: 扩展图片输入模型白名单，支持识别更多视觉 / OCR 模型模式（#884）
• adapter-openai, adapter-openai-like, adapter-deepseek, adapter-doubao, adapter-hunyuan, adapter-qwen: 修复非 LLM 模型错误声明 ToolCall 能力的问题，Embedding / Reranker 模型不再被识别为可工具调用模型（#882）
• adapter-claude: 支持 Claude 适配器多实例启用（#887）

What's Changed

• [Refactor] improve context compression budgeting by @dingyi222666 in #874
• chore(packages): bump context compression versions by @dingyi222666 in #876
• [Fix] 稳定 prompt pick 选择结果 by @liuwanwan1 in #878
• [Fix] avoid chat prompt injection warning by @dingyi222666 in #880
• [Refactor] simplify agent extension by @dingyi222666 in #881
• [Fix] add vision model patterns to image input whitelist by @Procyon-Nan in #884
• [Fix] count compacted prompt tokens directly by @dingyi222666 in #883
• [Fix] filter tool call capability for non-LLM models by @Procyon-Nan in #882
• fix(adapter-claude): 支持适配器多实例启用 by @Procyon-Nan in #887
• [Fix] 修复 1.4 alpha conversation 回归 by @liuwanwan1 in #886
• chore(packages): bump alpha versions by @dingyi222666 in #888

New Contributors

• @liuwanwan1 made their first contribution in #878

Full Changelog: 1.4.0-alpha.18...v1.4.0-alpha.20