dsh-soul 0.7.2 → 0.8.0

package/README.md

@@ -2,6 +2,8 @@
 
 [简体中文](./README.md) | [English](./README_EN.md)
 
+[![Listed on dsh-plugin.org](https://dsh-plugin.org/badges/listed.svg)](https://dsh-plugin.org/plugins/aliuyanfeng/dsh-soul)
+
 DeepSeek Harness 个性化设置插件，用于配置 Agent 的昵称、回复风格、语调和自定义指令。
 
 ## 功能
@@ -22,7 +24,7 @@ DeepSeek Harness 个性化设置插件，用于配置 Agent 的昵称、回复
 - Agent 可调用工具 `set_persona`，让模型在对话中直接调整人设
 - 人设预设：7 个内置人设（苏格拉底式提问者 / 极简主义者 / 资深架构师 / 教学型讲解者 / 严格代码审阅者 / 头脑风暴伙伴 / 自驱型协作者，随插件提供、不可删除）+ 多套自建人设一键切换（保存 / 使用 / 列表 / 删除，Web UI 与命令双入口）；每行摘要显示风格与偏离默认的维度
 - 预设只管 Agent 的人格：**切换预设不会动「关于你」**（昵称 / 职业 / 介绍是你本人的资料，与「Agent 怎么说话」是两码事）。内置与自建一视同仁——保存预设时只快照风格 / 特质 / 回复长度 / 输出语言 / 自定义指令，不包含身份字段；早期版本存下的完整快照会在读取时自动剥离
-- `set_persona` 确认模式（`requireToolConfirmation`）：Agent 的人设修改需经 `/soul confirm` 确认后才生效
+- `set_persona` 确认模式（`requireToolConfirmation`）：Agent 的人设修改先在输入框处弹出**宿主原生审批面板**，由你点「允许一次」或「拒绝」（`Enter` 允许 / `Esc` 拒绝）后才会生效；审批通道不可用时（如会话审批策略为 `never`）退化为待确认提议，需 `/soul confirm` 处理
 - 输入框光轨：Agent 回复中时，输入框边框显示沿边循环流动的光轨（默认颜色 `#679EFE`）
   - 颜色：预设色板、取色器或十六进制输入（如 `#679EFE`）
   - 流动速度：慢 / 中 / 快；光带粗细：细 / 中 / 粗
@@ -37,10 +39,23 @@ DeepSeek Harness 个性化设置插件，用于配置 Agent 的昵称、回复
 - 提示词插值已关闭：自定义指令里的 `{{…}}` 按字面保留，不会触发宿主提示词装配失败
 - 配置更新后同步到所有活动 Agent
 - 变更检测：仅在影响 Agent 行为的配置实际变化时刷新提示词并注入会话；纯外观配置（输入框光轨）只落盘，不产生注入消息
+- 变更记录与回滚（**插件详情页**：插件管理 → 已安装插件 → dsh-soul；设置页只在「人设预设」展开区给出指引，保留条数下拉也在这个区块内）：每次配置改动都追加一条记录（时间、**来源**、字段级 `原值 → 新值`），可**撤销某一次改动及其之后的全部改动**；删掉的预设可单点恢复。记录写在 `$DSH_HOME/soul-config.history.jsonl`（追加式，一行一条，人能直接读），保留条数可调（50 / 100 / 200 / 500，默认 200，改完即时生效）；存不下时**只影响记录、不影响配置写入**，并如实告诉你「记录可能不完整」。详情页给时间线与撤销，也可直接看文件
 - 插件图标：插件管理列表与侧栏入口显示专属图标（`assets/icon.svg`，36×36，随 npm 包发布）
 - 设置导航图标：设置面板左侧「个性化」那一项显示与插件图标同一套几何的图标——同一条光轨与火花，颜色与线宽则跟随邻居（单色 `currentColor`）；DSH 的导航图标按 section id 硬编码、插件声明不了，故由客户端做 DOM 替换
 - 提示词预览（只读，设置页底部折叠区）：无未保存编辑时显示**当前生效**的提示词；表单一有改动就转为预览**草稿**——展示保存后真正会注入的内容，边改边看（停止输入 400ms 后自动重新编译），并在保存 / 重置 / 应用预设后自动刷新。未通过校验的字段会单独列出且不计入预览
 
+## 路线图
+
+| 版本 | 方向 | 状态 |
+| --- | --- | --- |
+| `0.7.2` | 设置页样式对齐宿主、卡片堆叠轮播、人设变更走宿主原生审批面板 | 已发布 |
+| `0.8.0` | **变更审计与回滚** —— 记录每次配置改动（时间戳、来源、字段级 from→to），插件详情页给时间线与撤销（设置页只留去向指引） | [设计文档](./DESIGN-audit-rollback.md) 已完成 |
+| `0.9.0` | **三级隔离** —— Global → Workspace → Session，工作区决定人设、会话可临时覆盖 | [设计文档](./DESIGN-persona-scope.md) 设计完成待实现 |
+| `0.8.2` | 自演化提议 + diff 审批 —— Agent 提议精炼人设，改动以 diff 走审批面板 | 规划中 |
+| `0.9.0` | 记忆层 —— 跨会话长期记忆，按人设隔离、按需召回 | 规划中 |
+
+设计与取舍写在对应的设计文档里；已发布的改动见 [RELEASE_NOTES.md](./RELEASE_NOTES.md)，发版流程见 [PUBLISHING.md](./PUBLISHING.md)。路线图是方向而非承诺，顺序可能调整。
+
 ## 安装
 
 ```powershell
@@ -115,7 +130,7 @@ dsh plugin --profile <profile> update dsh-soul
 
 配置保存后会同步到所有活动 Agent，当前会话下一次请求即可使用最新配置；无实际变化的保存不产生注入消息。
 
-Agent 也可以通过工具 `set_persona` 在对话中直接调整你的人设（昵称、回复风格和语调、特质、回复语言、自定义指令）。模型只会在明确请求改变称呼、语气、风格或语言时调用该工具。开启「人设变更需确认」（`requireToolConfirmation`）后，该工具的修改会以待确认提议返回，需使用 `/soul confirm` 确认或 `/soul reject` 拒绝后才会生效。
+Agent 也可以通过工具 `set_persona` 在对话中直接调整你的人设（昵称、回复风格和语调、特质、回复语言、自定义指令）。模型只会在明确请求改变称呼、语气、风格或语言时调用该工具。开启「人设变更需确认」（`requireToolConfirmation`）后，该工具的修改会先在输入框处弹出**宿主原生审批面板**（与沙箱提权、模型切换用的是同一套交互），由你点「允许一次」或「拒绝」——`Enter` 允许、`Esc` 拒绝，改动会写入会话审计日志。审批通道不可用时（例如会话审批策略为 `never`、或没有可用的 Web UI 应答者）则退化为待确认提议，需 `/soul confirm` 确认或 `/soul reject` 拒绝。若你处在完全权限的会话中（审批策略 `never`），面板不会出现、每次变更都要手动输入命令，此时建议关闭该开关。
 
 ## 配置文件
 

package/README_EN.md

@@ -2,6 +2,8 @@
 
 # dsh-soul
 
+[![Listed on dsh-plugin.org](https://dsh-plugin.org/badges/listed.svg)](https://dsh-plugin.org/plugins/aliuyanfeng/dsh-soul)
+
 A personalization plugin for DeepSeek Harness (DSH). Configure your agent's nickname, reply style, tone, and custom instructions.
 
 ## Features
@@ -22,7 +24,7 @@ A personalization plugin for DeepSeek Harness (DSH). Configure your agent's nick
 - Agent-callable tool `set_persona` to let the model adjust persona during a conversation
 - Named personas: 7 built-in personas (Socratic Questioner / Minimalist / Senior Architect / Teaching Explainer / Strict Code Reviewer / Brainstorm Partner / Self-Driven Collaborator — shipped with the plugin, cannot be deleted) plus your own named personas, switched with one click (save / use / list / delete, in both the Web UI and slash commands); each row summarises the style and every dimension that differs from its default
 - Personas cover the **agent's personality only**: switching a persona never touches **"About you"** (nickname / occupation / bio describe *you*, not how the agent talks). Built-in and self-saved personas behave identically — saving a persona snapshots style / traits / reply length / reply language / custom instructions, never the identity fields; full-config snapshots written by earlier versions are stripped on read
-- `set_persona` confirmation mode (`requireToolConfirmation`): agent persona changes take effect only after `/soul confirm`
+- `set_persona` confirmation mode (`requireToolConfirmation`): agent persona changes first raise the **host's native approval panel** above the composer, taking effect only after you click "Allow once" or "Reject" (`Enter` allows, `Esc` rejects); when the approval channel is unavailable (e.g. the session approval policy is `never`) it degrades to a pending proposal handled with `/soul confirm`
 - Composer light trail: while the agent is replying, a light trail loops around the composer border (default color `#679EFE`)
   - Color: preset swatches, a color picker, or a hex input (e.g. `#679EFE`)
   - Flow speed: slow / normal / fast (default slow, 4.8s / 3.6s / 2.4s); trail thickness: thin / normal / thick (default thin, 1.5 / 2.5 / 4 px)
@@ -37,10 +39,23 @@ A personalization plugin for DeepSeek Harness (DSH). Configure your agent's nick
 - Prompt interpolation is disabled: `{{…}}` in custom instructions stays literal and can no longer break system-prompt assembly
 - Configuration synced to all active agents after every update
 - Change detection: the prompt is refreshed and sessions injected only when a behavior-affecting field actually changed — appearance-only config (composer light trail) is persisted without injecting anything
+- Change history and rollback (**plugin details page**: Plugin manager → Installed plugins → dsh-soul; the settings page only keeps a pointer, and the retention dropdown lives in that same block): every config change appends a record (time, **source**, field-level `old → new`), and you can **undo one change together with everything after it**; deleted personas can be restored individually. The log lives in `$DSH_HOME/soul-config.history.jsonl` (append-only, one record per line, readable by eye), and the retention count is configurable (50 / 100 / 200 / 500, default 200, applied immediately). When it cannot be written, **only the log suffers — config writes still go through**, and you are told the history may be incomplete
 - Plugin icon: a dedicated icon in the plugin manager list and the sidebar entry (`assets/icon.svg`, 36×36, shipped with the npm package)
 - Settings nav icon: the "Personalization" entry in the settings panel draws the same geometry as the plugin icon — same trail and spark, but coloured and weighted like its neighbours (monochrome `currentColor`). DSH hardcodes nav glyphs by section id and offers no way for a plugin to declare one, so the client swaps the glyph in the DOM
 - Prompt preview (read-only, collapsible section at the bottom of the settings page): shows the **currently active** prompt while the form is untouched; as soon as any edit is pending it compiles the **draft** instead — what will actually be injected once you save — refreshing live (400 ms after you stop typing) and again after save / reset / applying a persona. Fields that fail validation are listed separately and excluded from the preview
 
+## Roadmap
+
+| Version | Direction | Status |
+| --- | --- | --- |
+| `0.7.2` | Settings-page styling aligned with the host, stacked card carousel, persona changes routed through the host's native approval panel | Released |
+| `0.8.0` | **Change audit and rollback** — every config change recorded (timestamp, source, field-level from→to), with a timeline and undo on the plugin details page (the settings page keeps only a pointer) | [Design doc](./DESIGN-audit-rollback.md), done |
+| `0.9.0` | **Three-level scoping** — Global → Workspace → Session; the workspace picks the persona, a session may override it | [Design doc](./DESIGN-persona-scope.md), designed |
+| `0.8.2` | Self-evolution proposals with diff approval — the Agent proposes persona refinements and the diff goes through the approval panel | Planned |
+| `0.9.0` | Memory layer — long-term memory across sessions, isolated per persona, recalled on demand | Planned |
+
+Design rationale lives in the corresponding design document; shipped changes are in [RELEASE_NOTES.md](./RELEASE_NOTES.md) and the release process in [PUBLISHING.md](./PUBLISHING.md). The roadmap states direction, not a commitment — the order may change.
+
 ## Installation
 
 ```powershell
@@ -117,7 +132,7 @@ Slash commands are also available:
 
 Once saved, the configuration is synced to all active agents and takes effect on the next request in the current session; no-op saves inject nothing.
 
-The agent can also use the `set_persona` tool to adjust your persona (nickname, style & tone, traits, reply language, custom instructions). The model only invokes this tool when you explicitly ask to change how it addresses or responds to you. With **Require confirmation for persona changes** (`requireToolConfirmation`) enabled, tool changes come back as a pending proposal and take effect only after `/soul confirm` (or are discarded by `/soul reject`).
+The agent can also use the `set_persona` tool to adjust your persona (nickname, style & tone, traits, reply language, custom instructions). The model only invokes this tool when you explicitly ask to change how it addresses or responds to you. With **Require confirmation for persona changes** (`requireToolConfirmation`) enabled, tool changes first raise the **host's native approval panel** above the composer (the same interaction the host uses for sandbox escalation and model switching), where you click "Allow once" or "Reject" — `Enter` allows, `Esc` rejects, and the decision is written to the session audit log. When the approval channel is unavailable (for example the session approval policy is `never`, or no web UI answerer is present) it degrades to a pending proposal handled with `/soul confirm` or `/soul reject`. Under a full-permission session (approval policy `never`) the panel never appears and every change requires a manual command, so turning this switch off is recommended.
 
 ## Configuration File