chore(sbot): sync release notes for v0.3.0

Author: linyuan.yang

packages/backend/sbot/package.json

@@ -2,8 +2,8 @@
   "name": "@qingfeng346/sbot",
   "version": "0.3.0",
   "description": "Self-hosted AI agent server. Open-source alternative to OpenAI, Claude, and Gemini interfaces. Run LLM agents locally with MCP tool support, memory, multi-channel integrations (Lark/Feishu), and a built-in web UI.",
-  "releasenoteEn": "This release includes the following main changes:\n\n### Added\n\n1. Added the Tencent Yuanbao channel for connecting sbot to Yuanbao bots, with reliable text, image, and file exchange in conversations.\n2. Long-term memory can now use Git to manage version history, making knowledge changes easier to track and restore.\n3. Added server process lifecycle logs for startup, normal exits, unexpected exits, and leftover running instances. The Admin UI now shows application and lifecycle logs separately.\n\n### Improved\n\n1. Memory now keeps high-priority entries and adds summaries that are relevant to the current query, avoiding unrelated context while leaving full entries available on demand.\n2. Agent capabilities such as Memory, Agenda, Notes, Wiki, and Skills are now composable plugins with clearer extension and configuration boundaries.\n3. Model, embedding, and conversation-storage capabilities are separated into independent providers for more flexible configuration and future extensions.\n4. Docker environments now prepare Git automatically, so Git-managed memory works without extra runtime setup.\n\n### Fixed\n\n1. Fixed model streams with no renderable content or tool calls being treated as successful responses. They now return a clear error that points to thinking or token settings.\n2. Improved Tencent Yuanbao token handling, reconnection, and media transfer to reduce message interruptions caused by transient network or authentication failures.",
-  "releasenoteZh": "本次主要更新如下：\n\n### 新增\n\n1. 新增腾讯元宝渠道，支持将 sbot 接入元宝机器人；文本、图片和文件可在对话中稳定收发。\n2. 长期记忆支持通过 Git 管理历史版本，便于追踪和回溯知识内容的变更。\n3. 新增服务进程生命周期日志，可记录启动、正常退出、异常退出和遗留运行实例，管理后台可分别查看普通日志与生命周期日志。\n\n### 优化\n\n1. 记忆会保留优先级高的条目，并根据当前问题补充最相关的记忆摘要；避免无关内容占用上下文，完整正文仍按需读取。\n2. 记忆、日程、笔记、知识库和技能等 Agent 能力改为可组合的插件，扩展和配置边界更清晰。\n3. 模型、嵌入模型和会话存储能力拆分为独立 Provider，配置和后续扩展更灵活。\n4. Docker 环境会自动准备 Git，启用 Git 管理的记忆无需额外处理运行环境。\n\n### 修复\n\n1. 修复模型流式响应未返回可显示内容或工具调用时被静默当作成功的问题；现在会返回明确错误，并提示检查 thinking 或 Token 配置。\n2. 优化腾讯元宝渠道的令牌、重连和媒体传输处理，降低临时网络或鉴权失败导致消息中断的概率。",
+  "releasenoteEn": "This release includes the following main changes:\n\n### Added\n\n1. Added image input (vision) support: models that accept images can now exchange them directly in conversations. Support is detected automatically from the model catalog and can be overridden per model in its configuration; unsupported models gracefully fall back to text handling with a notice.\n2. The model list in the Admin UI now shows capability badges (such as image input), making model support easy to see at a glance.\n\n### Improved\n\n1. Reworked the model integration layer: OpenAI, Gemini, Anthropic, and Ollama channels now use the official native SDKs, reducing intermediate dependencies and aligning behavior with the official APIs.\n2. MCP tools now connect through the official native SDK with more stable connections and automatic recovery after disconnects.\n3. Better gateway compatibility: supports OpenAI endpoints that only accept the new token parameter, as well as various OpenAI and Anthropic gateways, so third-party proxies work out of the box.\n4. Updated the built-in model catalog so new models are recognized automatically; model and embedding lists and their configuration options have been tidied up accordingly.\n5. Channel logs now carry a unified prefix, making it easier to tell sources apart when multiple channels run at once.\n6. Improved the upgrade command's execution flow and feedback.\n\n### Fixed\n\n1. Fixed duplicated conversation content when using models via the OpenAI Responses API.\n2. Fixed request failures caused by the new OpenAI token parameter not being accepted.\n3. Fixed errors when an image was sent to a model without image support; it now falls back gracefully with a notice.",
+  "releasenoteZh": "本次主要更新如下：\n\n### 新增\n\n1. 新增图片输入（视觉）支持：支持图片的模型可直接在对话中收发图片；是否支持会按模型目录自动判断，也可在模型配置中显式声明，未收录的模型会自动降级为文本处理并给出提示。\n2. 管理后台模型列表展示模型能力标识（如图片输入），模型支持情况一目了然。\n\n### 优化\n\n1. 模型接入层重构：OpenAI、Gemini、Anthropic、Ollama 等渠道改为官方原生 SDK 直连，减少中间层依赖，行为与各官方接口保持一致。\n2. MCP 工具改为官方原生 SDK 接入，连接更稳定，并支持断线自动恢复。\n3. 更好的网关兼容性：兼容仅支持新版 Token 参数的 OpenAI 接口及各类 OpenAI / Anthropic 网关，第三方中转服务开箱可用。\n4. 更新内置模型库，新增模型可被自动识别；模型与嵌入模型列表及配置项同步整理。\n5. 渠道日志统一加前缀，多渠道运行时更易区分日志来源。\n6. 优化 upgrade 命令的执行流程与提示。\n\n### 修复\n\n1. 修复使用 OpenAI Responses 接口的模型对话内容重复的问题。\n2. 修复新版 OpenAI 接口 Token 参数不兼容导致请求失败的问题。\n3. 修复不支持图片输入的模型收到图片时报错的问题，现会自动降级并提示。",
   "license": "MIT",
   "repository": {
     "type": "git",