@lienze/dsh-subagent-model-policy 0.0.0-stage → 0.3.0

package/CHANGELOG.md

@@ -0,0 +1,47 @@
+# Changelog
+
+本项目遵循 [Keep a Changelog](https://keepachangelog.com/zh-CN/1.1.0/) 的结构，版本号遵循 SemVer。
+
+## [0.3.0] - 2026-11-09
+
+### Changed
+
+- 名称统一为「子智能体模型选择策略」/ `Subagent Model Policy`；npm 包名由 `@lienze/dsh-subagent-policy` 改为 `@lienze/dsh-subagent-model-policy`，安装与升级需使用新包名。
+- 插件显示名与 README 标题同步更新；面板标题此前已使用同一名称。
+- 补充发布元数据：`repository`/`homepage`/`bugs` 指向公开仓库，`keywords` 含插件市场识别用的 `dsh-plugin`，`engines.dsh` 声明最低宿主版本；npm 页面可直接跳转源码仓库。
+- 增加英文说明 `README_en.md` 与顶部语言切换条；`files` 收录英文说明，新增 `npm run verify` 聚合语法检查、测试与打包预检。
+
+### 兼容说明
+
+- 存储域 `lez_subagent_policy`、命令 `/subagent-policy`、工具 `subagent_policy` / `delegate_with_policy`、内部 bundle id 与样式前缀保持不变；已保存策略与执行记录在新包名下继续可读。
+- 旧包 `@lienze/dsh-subagent-policy@0.2.0-rc.1` 保留可查，迁移时改用新包名重新安装。
+
+## [0.2.0-rc.1] - 2026-11-06
+
+### Added
+
+- 三页面板：选择策略、模型池、执行洞察；策略开关与新会话默认开关行。
+- 固定模型与 Agent 自选两类策略；旧版跟随策略只读兼容，显式保存时转换。
+- 模型能力卡、宿主下发思考等级、模型池候选刷新与失效引用清理。
+- 会话级并发子智能体数量与最大递归深度限制，宿主容量实时读取与门禁。
+- 当前启用策略与新会话默认来源的删除保护。
+
+### Changed
+
+- 保存启用中策略、模型池说明与会话限制后，新内容在下一次主模型调用时生效；不再区分「保存并应用」。
+- 设置新会话默认静默更新开关摘要，不弹顶部提示。
+- 详情页提示在返回列表或切换页面后清除。
+- 正常选择策略页隐藏宿主诊断与原生入口提示；诊断数据仍保留在快照。
+- 「刷新候选」与全选/取消全选同一行；移除常驻「刷新服务端状态」按钮，面板重新可见时静默重读。
+
+### Fixed
+
+- 旧版策略生效期间的池说明同步进入下一次注入。
+- 无改动保存启用中策略不再先触发宿主容量门禁。
+- 取消请求后的刷新门禁、错误重试与冲突恢复保持可用。
+- 用户可见容量文案统一为「并发子智能体数量」。
+
+### Limitations
+
+- 预发布版本：AC-12 的 GUI 全项验收未完成；深度/存活真实超限、完整旧数据迁移与策略版本切换等场景仍暂缓。
+- 宿主授权、主模型、官方 profile 与原生工具参数不由本插件修改。

package/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 lez
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

package/README.md

@@ -1,3 +1,82 @@
-# Temporary Holding Version
+# 子智能体模型选择策略 · Subagent Model Policy
 
-This version is a temporary placeholder for this package. An operational version to replace this has been submitted for review and is awaiting a staged release.
+简体中文 · [English](https://github.com/ChenMuDX/dsh-subagent-model-policy/blob/main/README_en.md)
+
+作者：lez · 版本 `0.3.0` · 目标宿主 DeepSeek Harness `0.2.2-rc.1`。
+
+> 本版本使用数字版本号，但 AC-12 的 GUI 全项验收仍未完成，深度/存活超限等真实场景仍暂缓；不要把本版本表述为全项验收版。
+
+在当前会话中管理子智能体的模型选择策略、候选模型池、会话级并发/递归限制，并记录创建时的策略与实际响应证据。默认关闭；安装不会改变主模型、其他会话的模型选择或宿主授权。
+
+## 安装
+
+在 DeepSeek Harness 的插件管理器中安装：
+
+```text
+@lienze/dsh-subagent-model-policy@0.3.0
+```
+
+安装后按当前 Harness 的方式重载插件与页面；源码修改不等于运行实例已更新。若当前 Harness 使用 CLI 管理 profile，按该版本文档把同一包说明符加入对应 profile。
+
+包名中的 `@lienze` 是发布账号的 scope；插件内部的存储域、命令与工具名沿用历史前缀 `lez-`，不影响安装、升级或已保存数据。
+
+## 入口
+
+会话标题栏的「子智能体模型选择策略」按钮，或右侧栏页签中的同名入口。面板包含三个页面：
+
+- **选择策略**：策略开关、新会话默认开关、策略池与页内编辑。
+- **模型池**：候选勾选、刷新候选、能力卡、会话并发/递归限制。
+- **执行洞察**：本插件记录的真实创建与响应证据。
+
+## 策略语义
+
+- **固定模型**：委派只使用选中的模型；不匹配时报错，不静默替换。
+- **Agent 自选**：从会话池候选子集中选择，并在委派时显式填写精确 provider/model。
+- 旧版「跟随父模型」策略只读保留；保存为固定或 Agent 自选时显式转换。
+- 保存启用中的策略后，新内容在下一次主 Agent 发起模型调用时生效；保存非启用策略只更新策略库，启用后才生效。
+- 无改动重复保存不制造新版本；界面不显示内部版本号。
+- 当前启用策略和新会话默认的来源策略不可删除；需先关闭策略或更改/清除默认。
+- 设置为新会话默认后不弹顶部提示，只更新「新会话默认」开关摘要。
+- 关闭策略不会删除策略、不改主模型，也不追溯终止已创建的子智能体。
+
+## 模型池与限制
+
+- 模型池只列出当前宿主允许的候选；勾选不会扩大宿主授权，也不管理 API Key、连接或官方授权。
+- 当前启用策略引用的候选受保护，移除会被拒绝并提示先调整或关闭。
+- 「刷新候选」按发起瞬间的宿主授权与可用性移除失效候选，并同步清理策略中的失效引用；当前启用内容受保护。
+- 模型行显示宿主提供的思考等级与能力卡；未提供时不猜测等级名。
+- 并发子智能体数量与最大递归深度为会话公共配置，切换策略不改变；实际委派同时受宿主上限和本会话限制约束。
+- 保存启用中策略、启用策略和保存限制共用宿主容量门禁；无法读取宿主容量时拒绝写入并说明原因，保留已启用内容。
+
+## Agent 工具
+
+- `subagent_policy`：读取策略、模型说明、注入预览与运行回读。
+- `delegate_with_policy`：创建本地 fresh/fork 子智能体，返回子会话 id；自选模式必须显式提供精确 provider/model 和一句选择依据。完成后由原生 subagent 运行时投递结果。
+
+策略保存不自动创建子任务；是否委派仍由用户任务与 Agent 决定。
+
+## 原生入口边界
+
+原生 `subagent`、`subagent_fork`、`workflow`、`spawn_teammate` 的本地创建只做准入校验，不代填或改写模型参数。固定/自选模式下，原生入口不能指定目标模型时可能被拒绝，此时应改用 `delegate_with_policy`。插件不 monkey-patch、不替换官方工具，也不保证外部后端或非本地创建。
+
+## 证据语义
+
+- `assistant/message`：从子会话自身后缀的模型消息读取实际响应路由。
+- `request/header`：只说明拟请求路由，不显示为实际响应。
+- `pending`：尚无证据。
+- 子 Agent 自述、模型说明和请求配置不能替代实际响应证据。执行洞察保留创建时的策略名，不随之后重命名或删除改写。
+
+## 范围与暂缓
+
+- 覆盖当前会话内的策略库、模型池、默认内容与会话限制；不做跨会话共享策略库或项目级默认。
+- 不修改宿主授权、主模型、官方 profile 配置或原生工具参数。
+- 宿主撤销授权后，已绑定子 Agent 的后续请求由宿主拒绝，插件不静默回退。
+- 预发布阶段未完成 AC-12 全项 GUI 验收；深度/存活真实超限、完整旧数据迁移与策略版本切换等场景仍暂缓。合成测试通过不等于这些场景已实测。
+
+## 卸载
+
+通过 Harness 插件管理器禁用或移除 bundle。插件保存的会话策略与执行记录保留在宿主 storageDomain；重新启用后仍可读取。不要手动编辑 profile 生成的 `package.json`、`cordis.patch.yml` 或 `cordis.yml`。
+
+## 许可证
+
+MIT。见 [LICENSE](LICENSE)。

package/README_en.md

@@ -0,0 +1,82 @@
+# Subagent Model Policy
+
+[简体中文](https://github.com/ChenMuDX/dsh-subagent-model-policy/blob/main/README.md) · English
+
+Author: lez · Version `0.3.0` · Target host DeepSeek Harness `0.2.2-rc.1`.
+
+> This release uses a numeric version, but the full AC-12 GUI acceptance is still incomplete, and real scenarios such as depth/liveness overflow remain deferred. Do not describe it as fully accepted.
+
+Manage the model selection policy for subagents in the current session: candidate model pool, session-level concurrency and recursion limits, and the recorded policy and actual response evidence at creation time. Disabled by default; installing it does not change the main model, other sessions' model selection, or host authorization.
+
+## Install
+
+Install from the DeepSeek Harness plugin manager:
+
+```text
+@lienze/dsh-subagent-model-policy@0.3.0
+```
+
+After installing or updating, reload the plugin and page the way your Harness does; a source change is not a running instance. If your Harness manages profiles through the CLI, add the same package specifier to that profile.
+
+The `@lienze` scope is the publishing account's scope. Internal identifiers keep the historical `lez-` prefix; this does not affect installation, upgrades or stored data.
+
+## Entry points
+
+The "Subagent Model Policy" button in the session title bar, or the entry with the same name in the right sidebar tabs. The panel has three pages:
+
+- **Choose strategy**: policy switch, new-session default switch, strategy library and in-page editing.
+- **Model pool**: candidate selection, refresh candidates, ability cards, session concurrency/recursion limits.
+- **Execution insights**: real creations and response evidence recorded by this plugin.
+
+## Policy semantics
+
+- **Fixed model**: delegation only uses the selected model; a mismatch fails instead of silently switching.
+- **Agent choice**: the agent picks from the session pool's candidate subset and must supply an exact provider/model when delegating.
+- The legacy "follow parent model" policy is kept read-only; saving as fixed or agent choice converts it explicitly.
+- After saving an enabled policy, the new content takes effect on the next main-agent model call; saving a disabled policy only updates the strategy library.
+- Saving without changes does not create a new revision; the UI does not show internal revision numbers.
+- The enabled policy and the source strategy of the new-session default cannot be deleted; close the policy or change/clear the default first.
+- Setting a new-session default does not raise a top banner; only the "New-session default" summary updates.
+- Closing the policy does not delete it, does not change the main model, and does not terminate subagents already created.
+
+## Model pool and limits
+
+- The model pool only lists candidates the host currently allows; selecting them does not widen host authorization, and the plugin does not manage API keys, connections or official authorization.
+- Candidates referenced by the enabled policy are protected; removing them is refused with a prompt to adjust or close the policy first.
+- "Refresh candidates" drops candidates that became invalid at the moment of the request and cleans their stale references in policies; the currently enabled content stays protected.
+- Model rows show the thinking levels and ability cards provided by the host; when the host provides none, no level names are guessed.
+- Concurrent subagent count and maximum recursion depth are shared session settings and do not change when switching policies; actual delegation is constrained by both the host ceiling and this session's limits.
+- Saving an enabled policy, enabling a policy and saving limits share the host capacity gate; when host capacity cannot be read, writes are refused with the reason and the enabled content is preserved.
+
+## Agent tools
+
+- `subagent_policy`: reads the policy, model notes, injection preview and runtime readback.
+- `delegate_with_policy`: creates a local fresh/fork subagent and returns its id; agent-choice mode must pass an exact provider/model and a one-line reason. Results are delivered by the native subagent runtime.
+
+Saving a policy never creates subagent tasks; whether to delegate is still decided by the user's task and the agent.
+
+## Native entry boundaries
+
+Local creations through native `subagent`, `subagent_fork`, `workflow` and `spawn_teammate` get admission checks only; the plugin neither fills in nor rewrites model parameters. Under fixed/agent-choice policies, a native entry that cannot name the target model may be refused; use `delegate_with_policy` in that case. The plugin does not monkey-patch or replace official tools, and does not cover external backends or non-local creations.
+
+## Evidence semantics
+
+- `assistant/message`: the actual response route read from the sub-session's own trailing model message.
+- `request/header`: only the intended request route, never shown as an actual response.
+- `pending`: no evidence yet.
+- A sub-agent's self-report, model notes and request configuration cannot replace actual response evidence. Execution insights keep the strategy name captured at creation time and do not rewrite it after later renames or deletions.
+
+## Scope and deferred items
+
+- Covers the strategy library, model pool, defaults and session limits inside the current session; no cross-session shared library and no project-level defaults.
+- Does not modify host authorization, the main model, official profile configuration or native tool parameters.
+- After the host revokes authorization, subsequent requests of already-bound sub-agents are refused by the host; the plugin does not fall back silently.
+- The pre-release has not completed full AC-12 GUI acceptance; real depth/liveness overflow, complete legacy data migration and strategy-version switching remain deferred. Passing synthetic tests does not mean these scenarios were measured.
+
+## Uninstall
+
+Disable or remove the bundle through the Harness plugin manager. Stored session policies and execution records stay in the host storageDomain and remain readable after re-enabling. Do not hand-edit the profile-generated `package.json`, `cordis.patch.yml` or `cordis.yml`.
+
+## License
+
+MIT. See [LICENSE](LICENSE).