Skills Manager v1.34.1
Skills Manager v1.34.1
发布日期:2026-08-16
递交范围:v1.34.0...v1.34.1
发布概览
- 修复 1.34.0 引入的回归:如果你此前已经手动把 DeepSeek Harness 添加为自定义 Agent,它的 skills 路径会变得无法修改。
用户可见更新
- 被新增内置 Agent 遮住的自定义 Agent,路径可以再次修改(#378)—— 自定义 Agent 的 key 是从显示名派生的,所以名为「DeepSeek Harness」的自定义 Agent 会被存成
deepseek_harness——恰好是 1.34.0 新增的那个内置 key。此后应用解析和显示的都是内置的那份,而路径修改却被写进了被遮住的自定义定义里:保存提示成功,路径纹丝不动。现在两个路径都能正常修改,你不需要做任何清理。任何手动添加的 Agent 只要 key 撞上以后新增的内置 Agent,都会遇到同样的问题。
开发者与治理更新
find_adapter_with_store在同 key 时优先解析内置适配器,但两个路径写入函数都先查自定义工具,于是把修改存进了没人会读的地方。现在两者都改为先查内置;真正的自定义 Agent 仍然把路径存在自己的定义上。- 两个命令的 store 侧逻辑抽成
apply_tool_skills_dir和apply_tool_project_skills_dir,这样写入才能对着真实 store 测——命令本身是 async 且带 TauriState,这正是它们此前一个测试都没有的原因。 - 3 项回归测试,已反向验证:把任一写入函数改回旧优先级,对应测试即变红,而「真正的自定义 Agent」那条对照仍然通过。
当前校验状态
- ✅ 跨平台构建通过(macOS Intel / macOS ARM / Windows x64 / Linux x64)
- ✅ TypeScript typecheck 通过
English release notes
Release date: 2026-08-16
Commit range: v1.34.0...v1.34.1
Release Overview
- Fixes a regression in 1.34.0: if you had already added DeepSeek Harness as a custom agent by hand, its skills paths became impossible to change.
User-facing
- A custom agent shadowed by a new built-in can have its paths edited again (#378) — A custom agent's key is derived from its display name, so an agent added as "DeepSeek Harness" was stored under
deepseek_harness— the same key 1.34.0 shipped as a built-in. From then on the built-in was what the app resolved and displayed, while path edits were written to the hidden custom definition: the save reported success and the path never moved. Both paths are editable again, and no cleanup is needed. The same would have happened to anyone whose hand-added agent shared a key with any future built-in.
Developer & Governance
find_adapter_with_storeresolves a built-in ahead of a custom tool of the same key, but both path writers checked custom tools first and stored the edit where nothing reads it. Both now consult built-ins first; a genuine custom agent still keeps its paths on its own definition.- The store side of each command is split into
apply_tool_skills_dirandapply_tool_project_skills_dirso the writes can be exercised against a real store — the commands are async and take TauriState, which is why neither had a test. - 3 regression tests, reverse-verified: restoring the old precedence in either writer fails its test while the genuine-custom-agent case keeps passing.
Current Verification
- ✅ Cross-platform build passed (macOS Intel / macOS ARM / Windows x64 / Linux x64)
- ✅ TypeScript typecheck passed
完整变更:v1.34.0...v1.34.1