@klarkxy/dsh-model-hub 0.1.0-rc.1 → 0.1.0

package/README.md

@@ -1,8 +1,49 @@
 # @klarkxy/dsh-model-hub
 
-Cross-plugin model settings hub for DeepSeek Harness — **development preview**.
-The plugin page lists and edits live model-route fields through the native
-settings service. This build also reports:
+<!-- release-state-note -->
+
+Cross-plugin model settings hub for DeepSeek Harness.
+Open Plugins → Model Hub to search by plugin, purpose or current model and edit
+models from one compact list. Changes save immediately through the owning plugin.
+Open plugin pages automatically reload saved model changes. Unsaved drafts stay
+in place with a conflict notice; reload explicitly discards them. Reconnecting
+reloads the current settings, and subscription failures expose manual refresh.
+
+The hub includes Current Title, Git Commit, Safe Auto and ChatGPT when their
+matching updated plugins are loaded. ChatGPT uses its separate Codex model
+catalog and default.
+
+The Subagent Cast tab manages the host's subagent model allowlist: the enable
+switch and adding, replacing or removing models all write straight into the
+native `subagent-model-selection-settings` settings with revision conflict
+protection, so the host constraint keeps applying. Each model can also
+carry a free-text "task tendency"; tendencies live in
+`<profile>/data/model-hub/cast.json` and are injected into the system prompt as
+read-only advisory hints — the main agent consults them and still makes the final
+model choice when delegating. With this tab in place, the field matrix no longer
+shows a duplicate editor for these native settings.
+
+Add a model before enabling the cast; removing its last model also disables it.
+The allowlist constrains explicit subagent model choices; omitting a model still
+inherits the parent session. Tendencies have their own persisted revision and
+conflict checks across open pages. Existing v1 files migrate on the first
+successful save. A damaged or unreadable tendency file stays untouched and
+shows a recovery error; native cast controls remain available.
+
+Third-party volatile model-route fields are also discovered through native
+`settings.describe()`. Plugins storing settings elsewhere can contribute fields
+using `@klarkxy/dsh-plugin-kit/model-settings`; the hub does not inspect private
+storage. Discovery handles deep schemas, arrays, dictionaries and intersections.
+
+Refresh reloads both settings and model catalogs. Failed sources remain visible,
+and a rejected save requires refresh and a fresh selection. Revision checks and
+owner instance IDs protect against edits from stale pages. Other settings remain
+with their owner, including sibling values inside native model objects.
+A completed save is acknowledged separately from refreshing the page. If that
+refresh fails, the page says the change was saved and offers refresh without
+repeating the write.
+
+The diagnostic report separately answers:
 
 1. Which installed plugins expose model-route fields in their
    `settings.describe()` projection (detected via `@klarkxy/dsh-model-route`,
@@ -14,24 +55,23 @@ settings service. This build also reports:
 The report is logged and written to `<profile>/data/model-hub/probe-report.json`,
 and refreshed on every `settings/document-updated` event.
 
-Existing array items appear as separate rows with numeric indices. Empty
-collections show an informational row; add an item in the owning plugin
-before editing it here. Writes revalidate the target against a fresh native
-projection and use the descriptor revision for CAS. Follow mode preserves
-array elements, and unrelated settings stay with their owning plugin.
+Existing native array items appear as separate rows. Empty collections show an
+informational row; add an item in the owning plugin first. Restoring a native
+field returns to the plugin's inherited/default settings; clearing a direct
+array route preserves the element. Owner fields explain their own default mode.
 
-Both self-test fields are individually volatile, so the native settings
-projection includes them without exposing ordinary configuration. Per-day
-usage metering is not included in this preview.
+Both self-test fields remain in the diagnostic report and are hidden from the
+model selection page. Per-day usage metering is not included.
 
 ## Install
 
-Requires DSH 0.2.0-rc.2 or later. Install the preview channel, restart the profile, then open Plugins → Model Hub:
+Requires DSH 0.2.0-rc.2 or later. Restart the profile, then open Plugins → Model Hub:
 
 ```sh
-dsh plugin --profile web add @klarkxy/dsh-model-hub@next
+dsh plugin --profile web add @klarkxy/dsh-model-hub
 ```
 
 ## License
 
 SEE LICENSE IN LICENSE
+

package/cordis.patch.yml

@@ -1,4 +1,4 @@
-# Verification probe stage: detection + projection questions first, UI later.
+# Model settings hub; self-test fields remain diagnostic only.
 - insert:
     - id: model-hub
       name: '@klarkxy/dsh-model-hub'

package/docs/README.zh-CN.md

@@ -1,7 +1,38 @@
-# @klarkxy/dsh-model-hub
+# @klarkxy/dsh-model-hub
 
-DeepSeek Harness 跨插件模型设置枢纽——**开发预览**。插件页通过宿主设置服务
-列出并编辑可实时更新的模型路由字段，同时报告：
+<!-- release-state-note -->
+
+DeepSeek Harness 跨插件模型设置枢纽。打开「插件 → 模型枢纽」，
+按插件、用途或当前模型搜索，在紧凑列表中直接选择模型，改动立即由所属插件保存。
+已经打开的插件页会自动刷新已保存的模型设置。有未保存草稿时保留草稿并提示冲突，
+显式重新加载才丢弃草稿。断线重连后重读最新设置；订阅不可用时保留手动刷新入口。
+
+对应更新后的插件加载时，枢纽覆盖会话标题、Git 提交、安全审核以及 ChatGPT。
+ChatGPT 使用独立的 Codex 模型目录和默认模型。
+
+「子代理阵容」页签管理宿主的子代理模型白名单：启用开关、添加、替换或移除模型
+都直接写入原生 `subagent-model-selection-settings` 设置（带版本冲突保护），
+宿主约束照常生效。每个模型还可以标注一句「任务倾向」，倾向保存在
+`<profile>/data/model-hub/cast.json`，并作为只读提示注入系统提示——主智能体
+委派子代理时参考这些倾向自行选模型，倾向不做硬约束。该页签接管后，字段总览
+不再重复显示这组原生设置。
+
+先添加模型才能启用阵容，移除最后一个模型时会同时关闭开关。白名单约束显式选择的
+子代理模型；不指定模型时仍继承父会话。任务倾向有独立的持久修订号，多个页面同时
+修改时检查冲突。旧 v1 文件在首次成功保存时迁移；损坏或不可读时保留原文件并显示
+恢复提示，原生阵容控制仍可使用。
+
+第三方插件的 volatile 模型字段仍通过原生 `settings.describe()` 发现，支持深层
+schema、数组、字典和组合字段。使用其他存储方式的插件可以通过
+`@klarkxy/dsh-plugin-kit/model-settings` 提供模型字段；枢纽不读取其私有存储。
+
+刷新会同时更新设置和模型列表。读取失败的来源保留错误说明；保存发生冲突后，
+刷新并重新选择。设置版本和插件实例标识防止旧页面覆盖新设置；提示词、回退开关
+等其他字段保持原值，包括原生模型对象内部的附加字段。
+保存确认与页面刷新分别处理；写入成功而刷新失败时，页面明确提示已保存，可单独
+刷新，不会自动重复写入。
+
+诊断报告另行记录：
 
 1. 已安装插件的 `settings.describe()` 投影里有哪些模型路由字段
    （通过 `@klarkxy/dsh-model-route` 检测，标记或裸形状均可）？
@@ -11,21 +42,21 @@ DeepSeek Harness 跨插件模型设置枢纽——**开发预览**。插件页
 报告会写日志，并落盘到 `<profile>/data/model-hub/probe-report.json`，
 每次 `settings/document-updated` 事件后刷新。
 
-实际数组元素按数字索引显示为独立行。空集合显示提示行，需要先在所属插件中
-添加元素再编辑。写入前重新检测宿主投影中的目标字段，并用描述符版本进行 CAS
-校验。跟随默认模型保留数组元素，其他设置仍由所属插件维护。
+实际原生数组元素显示为独立行。空集合显示提示行，需要先在所属插件中添加元素。
+恢复原生字段时使用所属插件的继承值或默认值；清除直接数组路由时保留该元素。
+各插件提供的模型字段会说明各自的默认模式。
 
-两个自检字段各自声明 volatile，让宿主投影包含它们而不扩大普通配置的可见范围。
-此预览尚不包含按天用量统计。
+两个自检字段保留在诊断报告中，不出现在模型选择页。尚不包含按天用量统计。
 
 ## 安装
 
-需要 DSH 0.2.0-rc.2 或更高版本。安装预发布通道，重启 profile 后打开「插件 → 模型枢纽」：
+需要 DSH 0.2.0-rc.2 或更高版本。重启 profile 后打开「插件 → 模型枢纽」：
 
 ```sh
-dsh plugin --profile web add @klarkxy/dsh-model-hub@next
+dsh plugin --profile web add @klarkxy/dsh-model-hub
 ```
 
 ## 许可证
 
 SEE LICENSE IN LICENSE
+