@michengai/dsh-agency-agents 1.2.15 → 1.2.16
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +9 -0
- package/CHANGELOG.zh-CN.md +9 -0
- package/README.md +30 -1
- package/README.zh-CN.md +30 -1
- package/assets/icon.svg +26 -0
- package/lib/client.js +34 -21
- package/lib/index.js +1 -1
- package/lib/remote.js +1 -1
- package/lib/{src-D3NK414L.js → src-CSYgxZsN.js} +70 -4
- package/package.json +5 -3
package/CHANGELOG.md
CHANGED
|
@@ -4,6 +4,15 @@
|
|
|
4
4
|
|
|
5
5
|
This file records features and upgrade boundaries. npm and GitHub Releases remain the publication source of truth.
|
|
6
6
|
|
|
7
|
+
## 1.2.16 - 2026-10-26
|
|
8
|
+
|
|
9
|
+
- **Composer Task Recommendations**: Real-time expert and team recommendations above the chat input based on user draft semantics (Issue #27):
|
|
10
|
+
- **Zero-Token Local Scoring Engine**: Client-side stop-word extraction and lightweight scoring (< 3ms) with 250ms debounce, incurring zero LLM token costs and network overhead;
|
|
11
|
+
- **Native Styling and Centered Alignment**: Built with host primitives (`@deepseek-ai/dsh-client-ui-primitives` `Pill` / `Tooltip`) and aligned to composer CSS variables (`--dsh-composer-side-clearance`, `--dsh-composer-card-max-width`);
|
|
12
|
+
- **One-Click Mention and Deduplication**: Click recommendation chips to insert references directly into the input; already mentioned experts are automatically excluded; includes an "Ignore" action to dismiss current suggestions;
|
|
13
|
+
- **Configurable Toggle Switch (Default Off)**: Added dedicated toggle under plugin settings conforming to host volatile configuration rules for live hot-reload and persistence.
|
|
14
|
+
- **Enhanced Test Guarding & Mutation Hardening**: Extracted `resolveRecommendInputs` and introduced regression tests ensuring `ref` priority over `source` and safe fallback handling.
|
|
15
|
+
|
|
7
16
|
## 1.2.15 - 2026-10-26
|
|
8
17
|
|
|
9
18
|
- **Conversational creation for experts and expert teams**: Introduce dual \"Create with AI\" entry points next to creation actions in Experts and Expert Teams views (Issue #16):
|
package/CHANGELOG.zh-CN.md
CHANGED
|
@@ -4,6 +4,15 @@
|
|
|
4
4
|
|
|
5
5
|
本文件记录版本功能与升级边界。正式发布状态以 npm 和 GitHub Releases 为准。
|
|
6
6
|
|
|
7
|
+
## 1.2.16 - 2026-10-26
|
|
8
|
+
|
|
9
|
+
- **输入框上方智能推荐条(Composer Task Recommendations)**:当用户在主会话聊天框键入任务草稿时,基于草稿语义在输入框上方实时推荐最匹配的专家与专家团队(Issue #27):
|
|
10
|
+
- **极致性能与零 Token 损耗**:采用纯客户端本地停用词过滤与轻量打分引擎(`< 3ms`),配合 `250ms` 键入防抖,全程不调用后台大模型,零网络与模型 Token 开销;
|
|
11
|
+
- **原生视觉与居中版心对齐**:使用宿主原生 `@deepseek-ai/dsh-client-ui-primitives`(`Pill` / `Tooltip`)并严格绑定宿主边距 CSS 变量(`--dsh-composer-side-clearance`、`--dsh-composer-card-max-width`),确保推荐条与居中的输入框卡片完美对齐;
|
|
12
|
+
- **一键引用实体与防重机制**:点击推荐胶囊自动通过宿主原生 `insertReference` 将对应专家/团队插入到输入框中;精准提取 `ref` 标识,草稿中已引用的专家自动从推荐结果中排除;支持点击「忽略」临时折叠当前推荐;
|
|
13
|
+
- **独立功能开关(默认关闭)**:在「设置 → 插件」扩展配置中新增「任务智能推荐」独立控制开关,遵循宿主 `volatile` 易失配置规范支持运行期实时热更新与持久化。
|
|
14
|
+
- **强化装配层守护与测试有效性**:抽离 `resolveRecommendInputs` 纯函数并补充变异差分单测,彻底消除测试假绿盲区,确保实体标识提取与降级回退绝对稳健。
|
|
15
|
+
|
|
7
16
|
## 1.2.15 - 2026-10-26
|
|
8
17
|
|
|
9
18
|
- **支持自然语言对话创建专家与专家团队**:在专家与专家团队的新增按钮旁提供「通过对话创建」双入口(Issue #16 需求 2):
|
package/README.md
CHANGED
|
@@ -25,6 +25,8 @@
|
|
|
25
25
|
Choose a specialist in DSH for code review, design, operations, or research. All 331 experts are bundled and ready to use after installation.
|
|
26
26
|
|
|
27
27
|
- **Dual-surface access & wide layout**: Access management from **Settings → Experts** or enable the persistent primary sidebar entry to open in a dedicated 960px centered workspace view.
|
|
28
|
+
- **Conversational creation with AI**: Dual-mode creation via \"Create with AI\" jumps to main chat with auto-prefilled prompts. Agents render structured Markdown previews in chat with multi-turn revisions, followed by a native host Ask interaction confirmation before pre-filling into editor modals/drawers for human-in-the-loop review.
|
|
29
|
+
- **Smart bulk batch actions**: Compact native dropdown menu in the filter bar with dynamic filter linkage for \"Enable filtered (N) / Disable filtered (M)\" and \"Enable all (331) / Disable all\", protected by confirmation dialogs and atomic scheduling.
|
|
28
30
|
- **Complete roster of 331 domain experts**: Covers 22 standard divisions from engineering, product, and design to marketing, sales, and legal, with instant prefix matching.
|
|
29
31
|
- **Multi-agent team collaboration**: 5 high-impact built-in review teams and support for up to 100 custom teams with automated handoffs, parallel execution, or debate modes.
|
|
30
32
|
- **Fast invocation in composer**: Summon specialists or teams via the `@` trigger or composer picker button, instantly inserting task examples without accidental sends.
|
|
@@ -71,7 +73,34 @@ Retain `list_experts`, `summon_expert`, and `summon_experts`. Add `list_expert_t
|
|
|
71
73
|
- Team configuration lives in the host `agency-agents` settings namespace. Keep drafts on edit conflicts and review the latest configuration. Disabling or deleting teams does not disable experts or delete historical conversations.
|
|
72
74
|
- Experts cannot summon further experts or automatically add analysis rounds. Cancellation prevents queued starts; completed results and failures are reported separately.
|
|
73
75
|
|
|
74
|
-
|
|
76
|
+
### Conversational Creation with AI
|
|
77
|
+
|
|
78
|
+
Create custom experts and teams directly through natural language in chat via the **Create with AI** entry:
|
|
79
|
+
|
|
80
|
+
1. **Auto-jump with Prompt Prefixes**: Clicking \"Create with AI\" from the creation menu automatically closes the settings drawer and navigates to the main chat, injecting a natural prompt prefix (e.g. `I want to create an expert specializing in ` or `Assemble an expert team for this mission: `) with cursor auto-focused at the end;
|
|
81
|
+
2. **In-Chat Structured Preview**: Specialized tools (`propose_custom_expert` / `propose_expert_team`) generate Markdown preview cards containing division categories, collapsed prompt blocks, and localized roster members with emojis, physically preventing disruptive popups during drafting;
|
|
82
|
+
3. **Fuzzy Expert Slug Resolution**: When assembling teams, the model can specify Chinese names, English names, or short slugs (e.g. `code-reviewer`), which are automatically resolved to valid in-repo expert slugs without schema validation errors;
|
|
83
|
+
4. **Interactive Ask Confirmation**: Once satisfied, the user replies \"confirm\" in chat. The agent triggers a native host **Ask interaction card** for two-step human confirmation;
|
|
84
|
+
5. **Global Pre-filled Modal Launch**: After user submission, the full editor modal or drawer opens with all proposed fields pre-filled and inputs styled to 100% full container width for final human-in-the-loop review and saving;
|
|
85
|
+
6. **Session-Bound Concurrency**: Proposal states are strictly bound to `sessionId`, ensuring multiple concurrent sessions or windows never cross-trigger popups.
|
|
86
|
+
|
|
87
|
+
### Bulk Batch Actions
|
|
88
|
+
|
|
89
|
+
A native compact **Batch Actions** dropdown menu is available on the right of the filter toolbar:
|
|
90
|
+
|
|
91
|
+
- **Dynamic Filter Linkage**: Dynamically adapts based on active source, category, status, and search filters. Offers **Enable filtered (N) / Disable filtered (M)** when filters are active, and smoothly switches to **Enable all (331) / Disable all** when unfiltered;
|
|
92
|
+
- **Confirmation & Conflict Protection**: Includes modal confirmation dialogs indicating affected expert counts, powered by atomic scheduling (`performBulkCommit`) to eliminate dirty writes caused by concurrent changes during review. Only affects availability and never auto-summons.
|
|
93
|
+
|
|
94
|
+
### Composer Task Recommendations
|
|
95
|
+
|
|
96
|
+
Real-time expert and expert team suggestions displayed directly above the chat composer based on draft semantics (Issue #27):
|
|
97
|
+
|
|
98
|
+
- **Zero-Token Local Scoring Engine**: Client-side tokenizer, stop-word filtering, and lightweight scoring algorithm (< 3ms) debounced at 250ms, incurring zero network overhead and zero LLM token consumption;
|
|
99
|
+
- **Native Visuals & Centered Clearance**: Built with host primitives (`@deepseek-ai/dsh-client-ui-primitives` `Pill` / `Tooltip`) and mapped to composer CSS clearance variables, guaranteeing horizontal alignment with the centered input box;
|
|
100
|
+
- **One-Click Mention & Deduplication**: Click candidate chips to insert references directly into the input via host `insertReference`; already referenced experts are automatically filtered out from suggestions; includes an "Ignore" button to collapse suggestions;
|
|
101
|
+
- **Dedicated Extension Toggle**: Toggle under **Settings → Plugins** (default off), supporting host volatile hot-reloading and persistence.
|
|
102
|
+
|
|
103
|
+
> This README describes version 1.2.16. **This is a breaking upgrade: DeepSeek Harness 0.1.x is no longer supported.** Installation on an older host is rejected by the version gate. npm and GitHub Releases determine publication status. See the [changelog](CHANGELOG.md) for the full boundary. Expert and team settings are retained; back up settings before downgrading to 1.0.x.
|
|
75
104
|
## Screenshots
|
|
76
105
|
|
|
77
106
|
The screenshots below show the 1.0.0 interface in Chinese. Experts and expert teams share the same settings page.
|
package/README.zh-CN.md
CHANGED
|
@@ -25,6 +25,8 @@
|
|
|
25
25
|
遇到代码审查、设计、运营或研究任务时,在 DSH 中选一位合适的专家,让它从专业角度协助你完成工作。内置 331 名专家,安装后即可选择使用。
|
|
26
26
|
|
|
27
27
|
- **多端与双入口访问**:支持原有「设置 → 专家」弹窗管理,亦可在插件详情页开启主导航栏常驻「专家」大面板(960px 居中版心,彻底摆脱弹窗视口滚动限制)。
|
|
28
|
+
- **自然语言对话创建(Create with AI)**:新建专家与专家团队支持「通过对话创建」双入口,自动跳转主会话预填提示词;智能体在对话中出具结构化图文预览,支持多轮自然微调,满意后经宿主原生 Ask 卡片把关,全量回填反显至编辑弹窗/抽屉(Human-in-the-Loop)。
|
|
29
|
+
- **专家管理智能批量操作(Batch Actions)**:在管理筛选条右侧引入原生紧凑批量菜单。与来源、分类、状态及关键词检索动态联动,自适应提供「启用筛选 (N) / 停用筛选 (M)」及「启用全部 (331) / 停用全部」;具备二次确认与原子调度保护。
|
|
28
30
|
- **331 位全领域专家名册**:涵盖工程、设计、产品、营销、法务、金融等 22 个标准分区,支持中英文关键词与拼音前缀秒级即时检索。
|
|
29
31
|
- **专家团(Teams)多角色协同**:内置 5 大高频业务团队,支持自定义最多 100 个协作团队;支持根据任务复杂度自动选择接力推进、多路并行或深度评审协作模式。
|
|
30
32
|
- **会话输入框快捷召唤**:支持对话框左侧机器人按钮快速点选,或在正文中输入 `@` 快捷呼出专家及团队名册,无缝填入任务示例。
|
|
@@ -71,7 +73,34 @@
|
|
|
71
73
|
- 团队配置保存在宿主 `agency-agents` 命名空间。编辑冲突保留草稿并提示核对最新配置;停用或删除团队不会停用专家,也不会删除历史会话。
|
|
72
74
|
- 专家不得继续召唤专家或自动追加分析轮次。取消会停止未启动任务;已有结果和失败原因分别报告。
|
|
73
75
|
|
|
74
|
-
|
|
76
|
+
### 自然语言对话创建(Create with AI)
|
|
77
|
+
|
|
78
|
+
在专家与专家团队列表的新增按钮旁均提供「**通过对话创建**」入口:
|
|
79
|
+
|
|
80
|
+
1. **自动跳转与提示词引导**:点击后自动关闭管理面板并平滑跳转至主聊天框,预填自然语言提示前缀(如 `我想创建一个专家,专门负责 ` 或 `帮我组建一个专家团队,任务目标是 `),光标自动聚焦至末尾;
|
|
81
|
+
2. **对话流结构化预览(只读防打扰)**:智能体通过专职工具 `propose_custom_expert` / `propose_expert_team` 在聊天气泡中出具 Markdown 预览卡片,包含所属分类、折叠提示词与带国际化中文名及 Emoji 的成员阵容;在草案构想与多轮微调阶段物理阻断意外弹窗;
|
|
82
|
+
3. **成员智能模糊别名匹配(Fuzzy Expert Resolver)**:大模型组建团队时无论输入中文全称(如“代码审查工程师”)、英文名还是短 slug(如“code-reviewer”),均自动解析映射为库内合法的真实专家 slug,彻底避免名称报错;
|
|
83
|
+
4. **原生 Ask 卡片交互把关**:用户对预览满意并在对话框中回复「确认」后,模型唤起宿主原生 **Ask 交互卡片**进行二次防误触把关;
|
|
84
|
+
5. **全局动态反显弹窗**:用户在卡片确认提交后,系统自动呼出编辑抽屉或中心模态弹窗,所有字段全量回填反显且单行输入框 100% 满宽整齐排版,由用户最终人工审查并保存;
|
|
85
|
+
6. **会话 ID 隔离**:草案生命周期与当前会话严格绑定,多窗口或分屏并发使用时各收各的草案,绝不跨窗口串台弹窗。
|
|
86
|
+
|
|
87
|
+
### 专家管理智能批量操作(Batch Actions)
|
|
88
|
+
|
|
89
|
+
在专家管理界面的顶部筛选条右侧提供了原生「**批量操作**」下拉菜单:
|
|
90
|
+
|
|
91
|
+
- **动态筛选联动**:与来源、分类、状态及关键词搜索实时联动。当有筛选条件时,自适应提供「**启用筛选 (N)**」与「**停用筛选 (M)**」,精准控制当前目标子集;无筛选时,自适应切换为「**启用全部 (331)**」与「**停用全部**」;
|
|
92
|
+
- **二次确认与并发保护**:点击后弹出模态确认对话框并显示受影响专家数量;底层采用原子调度器(`performBulkCommit`),防范弹窗停留期间外部并发写入导致的旧快照覆盖风险;仅改变专家可用范围,不自动召唤、不产生子代理。
|
|
93
|
+
|
|
94
|
+
### 输入框任务智能推荐(Task Recommendations)
|
|
95
|
+
|
|
96
|
+
在主会话聊天框键入任务内容时,系统基于草稿语义在输入框正上方提供实时的专家与专家团队智能推荐(Issue #27):
|
|
97
|
+
|
|
98
|
+
- **零 Token 本地纯计算**:采用纯客户端切词、停用词过滤与轻量打分算法(耗时 `< 3ms`),配合 `250ms` 键入防抖,全程不调用后台大模型,零网络延迟与 Token 消耗;
|
|
99
|
+
- **原生视觉与居中对齐**:使用宿主原生 `@deepseek-ai/dsh-client-ui-primitives`(`Pill` / `Tooltip`)并严格遵循宿主 CSS 边距变量,确保推荐胶囊条与居中的主输入卡片完全对齐;
|
|
100
|
+
- **一键引用与智能排除**:点击推荐项自动通过宿主原生 `insertReference` 将实体插入输入框草稿中,已引用的专家自动从推荐列表中排除,支持点击「忽略」快速折叠;
|
|
101
|
+
- **独立配置开关**:在「设置 → 插件」扩展配置中提供「任务智能推荐」独立开关(默认关闭),支持实时热更新并持久化保存。
|
|
102
|
+
|
|
103
|
+
> 本 README 描述 1.2.16。**这是破坏性升级:不再支持 DeepSeek Harness 0.1.x。** 在旧宿主上安装会被版本门拒绝。正式发行以 npm 和 GitHub Releases 为准。完整边界见[更新日志](CHANGELOG.zh-CN.md)。专家和专家团配置仍保留;回退 1.0.x 前先备份设置。
|
|
75
104
|
## 界面预览
|
|
76
105
|
|
|
77
106
|
以下为 1.0.0 中文界面截图,专家与专家团共用同一个设置页面。
|
package/assets/icon.svg
ADDED
|
@@ -0,0 +1,26 @@
|
|
|
1
|
+
<svg width="36" height="36" viewBox="0 0 36 36" fill="none" xmlns="http://www.w3.org/2000/svg">
|
|
2
|
+
<defs>
|
|
3
|
+
<linearGradient id="dsh_svg_agency_cap" x1="4" y1="6" x2="32" y2="20" gradientUnits="userSpaceOnUse">
|
|
4
|
+
<stop stop-color="#8B5CF6"/>
|
|
5
|
+
<stop offset="1" stop-color="#6366F1"/>
|
|
6
|
+
</linearGradient>
|
|
7
|
+
<linearGradient id="dsh_svg_agency_suit" x1="8" y1="18" x2="28" y2="30" gradientUnits="userSpaceOnUse">
|
|
8
|
+
<stop stop-color="#6366F1"/>
|
|
9
|
+
<stop offset="1" stop-color="#4F46E5"/>
|
|
10
|
+
</linearGradient>
|
|
11
|
+
</defs>
|
|
12
|
+
<!-- Scholar / Mortarboard Cap Top Diamond (Diamond Top) -->
|
|
13
|
+
<polygon points="18,6 31,12 18,17.5 5,12" fill="url(#dsh_svg_agency_cap)"/>
|
|
14
|
+
<!-- Cap Skull Base -->
|
|
15
|
+
<path d="M10 14.5V18.5C10 21 13.5 22.5 18 22.5C22.5 22.5 26 21 26 18.5V14.5" fill="#4338CA" fill-opacity="0.3" stroke="#4338CA" stroke-width="1.5"/>
|
|
16
|
+
<!-- Cap Tassel Hanging on the Right -->
|
|
17
|
+
<path d="M25 14L28.5 18V23.5" stroke="#F59E0B" stroke-width="1.8" stroke-linecap="round"/>
|
|
18
|
+
<circle cx="28.5" cy="24" r="1.5" fill="#F59E0B"/>
|
|
19
|
+
|
|
20
|
+
<!-- Professional Suit / Collar Outline -->
|
|
21
|
+
<path d="M10 24L18 29L26 24" stroke="url(#dsh_svg_agency_suit)" stroke-width="2.4" stroke-linecap="round" stroke-linejoin="round"/>
|
|
22
|
+
<!-- Professional Tie / Expert Badge in Center -->
|
|
23
|
+
<path d="M17 26.5L19 26.5L18.5 31L18 31.5L17.5 31L17 26.5Z" fill="#F59E0B"/>
|
|
24
|
+
<!-- Glowing Expert Sparkle Star on Forehead -->
|
|
25
|
+
<path d="M18 10L18.8 12L21 12.8L18.8 13.6L18 15.5L17.2 13.6L15 12.8L17.2 12L18 10Z" fill="#FFFFFF"/>
|
|
26
|
+
</svg>
|