OpenViking v0.3.23 Release Notes / 发布说明
Release date / 发布日期: 2026-06-03
Full Changelog / 完整变更记录: v0.3.22...v0.3.23
Compare range / 对比范围: v0.3.22...v0.3.23 (42 commits)
中文
版本概览
v0.3.23 是一次偏体验与可靠性的版本:原生 ov CLI 改为更完整的交互式配置与诊断入口,Web Studio 增加 Playground 与连接身份管理,VikingBot 的经验召回改为配置驱动,同时补齐一批资源导入、会话、插件和 VLM 链路的可靠性修复。
主要更新
- 原生
ovCLI 体验重构:ov config现在是配置管理入口,可交互添加、编辑、删除、切换配置;ov config show、ov config validate、ov config switch保留为显式子命令。新增ov language/ov lang选择显示语言,ov status [--verbose]提供聚合诊断视图,ov health与错误提示改为更可读的渲染。 - Web Studio Playground 与身份管理:Studio 侧边栏新增 Playground,可在同一工作区内查看上下文树、运行 Terminal 操作并与 Agent 面板交互;Connection & Identity 页面支持保存连接、选择 account/user/agent、创建 account/user、复制或重新生成 API key。
- VikingBot 经验召回配置化:新增
bot.ov_server.recall_exp_first_round_only、exp_recall_limit、exp_recall_max_chars,用于单任务/评测场景中只在第一轮注入 agent experience;本地与远端模式都按传入agent_id做经验命名空间隔离。 - 资源 Watch 更易用:
add_resource设置watch_interval > 0时不再强制要求显式to;如果导入结果返回稳定root_uri,watch task 会自动绑定到该 URI。CLI/MCP/文档示例同步更新。 - Agent 插件与会话可靠性:Claude Code / OpenClaw 插件改为向 OpenViking 写入结构化 tool parts,工具调用与结果不再只能内联到文本;CJK-aware token 估算覆盖 Python 与插件侧,降低中文、日文、韩文会话的预算低估风险。
新功能用法
CLI 新入口:
ov language zh-CN
ov config
ov config validate
ov status --verboseov config setup-cli 已移除;新配置请使用裸 ov config。
导入远端资源并绑定自动刷新到本次导入产生的 root_uri:
ov add-resource https://github.com/example/repo.git --watch-interval 60如果资源类型无法返回稳定 root_uri,服务端会要求显式传入 --to。
VikingBot 只在第一轮注入经验召回:
{
"bot": {
"ov_server": {
"recall_exp_first_round_only": true,
"exp_recall_limit": 2,
"exp_recall_max_chars": 10000
}
}
}体验与兼容性改进
ovcli.conf默认 URL 统一到http://127.0.0.1:1933,配置文件序列化会跳过默认值和空字段,减少配置噪音。- 首次使用新 CLI 时,如果尚未选择显示语言,交互环境会弹出语言选择;非交互环境应先运行
ov language en或ov language zh-CN。 ov status默认展示整理后的诊断视图;需要原始组件数据时使用ov status --verbose或-o json。- 语义处理默认并发从 100 调整/校准为 64,避免大目录或多 root 同时处理时过度占用 VLM。
- 本地目录上传会跳过 symlink,避免递归、重复导入或越界读取。
ls/ 资源浏览的摘要批量读取与 Web Studio 资源浏览导航、保存延迟、查找面板二级调色板均有优化。
修复
- 修复加密存储
mv过程中可能二次删除的问题。 - 修复 VikingDB 批量 upsert/fetch/delete 的单批数量限制。
- 修复 Codex VLM Responses 流事件解析,在 completed response 没有 output 时使用 text deltas 兜底。
- 修复显式 VLM provider 经过 VikingBot /
ovcli配置序列化后没有按 VLM adapter 路由的问题。 - 修复 Web Studio 打包产物未进入 release wheel、Playground bot 流程、随机 UUID fallback 和资源浏览细节问题。
- 修复 server tool context 中 proxy namespace policy 传播问题。
- 修复 LangChain tool schema、OpenClaw auto-recall health precheck 鉴权、
memory_store零提取时缺少模型可见反馈等问题。 - Docker runtime stage 增加
git,避免运行期依赖 Git 的资源导入失败。
文档、测试与安全
- 文档新增通用 Agent 集成指南,并补充 Cursor、Trae、Codex、Claude Code、OpenClaw、OpenCode、SDK、MCP 等图文配置材料。
- MCP / Resource API 文档更新
add_resourcewatch 语义,read()目录 URI 错误现在记录结构化 details。 - 配置文档修正
vlm.max_concurrent默认值为 64。 - 新增 CJK token 估算测试、session message/commit 测试和 CLI config/status/doctor 回归测试。
- OpenClaw auto-recall health precheck 现在会带认证信息,避免误判受保护的 OpenViking 服务。
English
Overview
v0.3.23 focuses on operator experience and reliability. The native ov CLI now has a fuller interactive configuration and diagnostics flow, Web Studio adds Playground plus connection and identity management, VikingBot experience recall becomes config-driven, and several resource import, session, plugin, and VLM reliability issues are fixed.
Highlights
- Native
ovCLI refresh:ov configis now the configuration manager for adding, editing, deleting, and switching saved configs.ov config show,ov config validate, andov config switchremain explicit subcommands.ov language/ov langchooses display language,ov status [--verbose]provides system diagnostics, andov healthplus runtime errors now render with clearer guidance. - Web Studio Playground and identity management: Studio now includes Playground with a context tree, Terminal, and Agent panel. The Connection & Identity page can save connection state, select account/user/agent, create accounts and users, copy keys, and regenerate API keys.
- Config-driven VikingBot experience recall:
bot.ov_server.recall_exp_first_round_only,exp_recall_limit, andexp_recall_max_charslet task-style runners inject agent experience only on the first turn. Local and remote modes both isolate experience namespaces by incomingagent_id. - Simpler resource watches:
add_resourceno longer requires explicittowhenwatch_interval > 0; if the import returns a stableroot_uri, the watch task binds to that URI. CLI, MCP, and docs examples were updated. - Agent plugin and session reliability: Claude Code and OpenClaw plugins now write structured tool parts to OpenViking instead of flattening calls/results into text only. CJK-aware token estimation is shared across Python and plugin code, reducing budget underestimation for Chinese, Japanese, and Korean sessions.
New Feature Usage
New CLI entry points:
ov language en
ov config
ov config validate
ov status --verboseov config setup-cli has been removed; use bare ov config for new setup.
Import a remote resource and bind automatic refresh to the imported root_uri:
ov add-resource https://github.com/example/repo.git --watch-interval 60If the resource type cannot return a stable root_uri, the server asks for explicit --to.
VikingBot first-turn experience recall:
{
"bot": {
"ov_server": {
"recall_exp_first_round_only": true,
"exp_recall_limit": 2,
"exp_recall_max_chars": 10000
}
}
}Improvements
ovcli.confdefaults tohttp://127.0.0.1:1933, and config serialization omits default/empty fields to reduce noise.- On first use, the new CLI prompts for a display language in interactive environments. Non-interactive automation should run
ov language enorov language zh-CNbefore other commands. ov statusdefaults to a curated diagnostic view. Useov status --verboseor-o jsonfor raw component data.- Semantic processing concurrency is calibrated to 64 by default instead of 100, preventing excessive VLM pressure during large directory or multi-root processing.
- Local directory uploads skip symlinks to avoid recursive, duplicate, or out-of-scope ingestion.
ls/ resource browsing now batch-fetches abstracts more efficiently; Web Studio resource navigation, save latency, and the find palette L2 flow were polished.
Fixes
- Fixed encrypted storage
mvpaths that could delete twice. - Fixed VikingDB batch limits for upsert, fetch, and delete operations.
- Fixed Codex VLM Responses stream parsing by falling back to text deltas when the completed response has no output.
- Fixed explicit VLM provider routing through the VikingBot /
ovcliVLM adapter path. - Fixed Web Studio release-wheel asset bundling, Playground bot flow, random UUID fallback, and resource-browser details.
- Fixed proxy namespace policy propagation in server tool context.
- Fixed LangChain tool schemas, authenticated OpenClaw auto-recall health prechecks, and model-visible feedback when
memory_storeextracts zero memories. - Added
gitto the Docker runtime stage so Git-backed resource imports work in runtime images.
Docs, Tests, and Security
- Added general Agent integration guides and image-backed setup docs for Cursor, Trae, Codex, Claude Code, OpenClaw, OpenCode, SDK, and MCP usage.
- Updated MCP / Resource API docs for
add_resourcewatch semantics;read()directory-URI errors now document structured details. - Corrected the documented
vlm.max_concurrentdefault to 64. - Added CJK token estimation tests, session message/commit tests, and CLI config/status/doctor regressions.
- OpenClaw auto-recall health precheck now authenticates against protected OpenViking services.