dsh-tabbit 0.3.0 → 0.3.3

package/CHANGELOG.md

@@ -1,5 +1,46 @@
 # Changelog
 
+## 0.3.3
+
+- Restored web-client compatibility with DSH `0.1.1-rc.2`. The client no
+  longer hard-injects either generation of the conversation service: it uses
+  `uiConversation.events` on DSH `0.1.2-alpha.1+` and falls back to
+  `conversationEvents` on `0.1.1`, so neither host remains pending while
+  waiting for the other generation's service name.
+
+## 0.3.2
+
+- `tabbit_browser` gains `list_tasks: true`: a zero-side-effect readback of
+  the browser tasks this session already has open (which one an omitted
+  `task` would target next), read from the in-memory session registry only.
+  A memory jog for when the model has lost track after a gap in the
+  conversation, instead of guessing a task name or opening a redundant one.
+- `claim_tabs` now works when the task already exists, not only when the
+  same call creates it. Runtime Service silently drops a creation-time
+  `--claim-tab` on a reused task, so previously the plugin threw
+  `CLAIM_REQUIRES_NEW_TASK` and forced starting a fresh task just to attach
+  a tab the user pointed at mid-conversation. It now falls back to the
+  standalone `claim` CLI command for tasks the session already knows about,
+  so an already-open tab can join ongoing browser state instead of always
+  starting over.
+- Fixed the bundled `SKILL.md`'s frontmatter parser assuming LF-only line
+  endings: a copy checked out or published with CRLF endings would have had
+  its raw YAML header served to the model as skill content instead of being
+  stripped.
+
+## 0.3.1
+
+- Fixed the web client plugin never activating on current DSH hosts — the web
+  UI showed `dsh-tabbit: pending (waiting for service: conversationEvents)`.
+  The DSH client runtime refactor (already part of `0.1.2-alpha.1`) renamed the
+  `conversationEvents` service to `uiConversation` and moved node registration
+  from `service.register()` to `service.events.register()`. The client plugin
+  now injects `uiConversation`; the `/tabbit-info` status card, the `@tab`
+  composer source, and the watching-instance hint all had been blocked behind
+  that pending inject and work again. The `conversation.chat.node` keyed slot
+  registration and the Definition contract (`match`/`start`/`update`/
+  `buildViewNode`) are unchanged.
+
 ## 0.3.0
 
 - Major upgrade: browser automation now runs through the native

package/README.en.md

@@ -1,8 +1,8 @@
 # dsh-tabbit — Tabbit's Official DeepSeek Harness Plugin
 
-**English** | [简体中文](README.md)
+**English** | [简体中文](README.md) | [Changelog](CHANGELOG.md)
 
-![Tabbit Browser for DeepSeek Harness](tabbit-for-dsh.png?v=2)
+![Tabbit Browser for DeepSeek Harness](assets/dsh-tabbit-banner.png)
 
 A DeepSeek Harness (dsh) plugin bundle for Tabbit Browser. Through this
 plugin, dsh calls on Tabbit to complete agent tasks: real pages, real login
@@ -48,6 +48,12 @@ dsh plugin --profile web add link:/path/to/dsh-tabbit   # local development
 > upgrade notice from the daily update check, and simply re-running the
 > install command upgrades in place.
 
+## Community & Support
+
+Scan the QR code below to join the **dsh-tabbit Developer Group** to share feedback, ask questions, and discuss new features:
+
+![dsh-tabbit Developer Group](assets/dsh-tabbit-developer-community-qr.png.jpg)
+
 ## Settings
 
 ### Basic configuration
@@ -103,6 +109,26 @@ gets denied — that combination gets a denial message pointing at the
 
 ## Development
 
+Developing and testing this project depends on a local
+[deepseek-harness](https://github.com/deepseek-ai/deepseek-harness) checkout.
+The `@deepseek-ai/*` packages published on npm generally lag too far behind
+to install directly as dependencies, so you need to build a harness checkout
+locally first, then point the repo's `.dsh-harness` symlink at it with the
+script below (that path is already ignored in `.gitignore`). That's the only
+place you configure it — neither `tsconfig.json` nor `package.json` needs
+any changes.
+
+Get and build deepseek-harness (skip if you already have one):
+```bash
+git clone https://github.com/deepseek-ai/deepseek-harness && cd deepseek-harness && pnpm install && pnpm build && cd ..
+```
+
+Point this repo at it — replace `/path/to/deepseek-harness` with your actual checkout path:
+```bash
+npm run link-harness -- /path/to/deepseek-harness   # or set DSH_HARNESS_PATH
+```
+
+Install and build:
 ```bash
 pnpm install && pnpm build   # tsc → lib/
 npm test                     # build + node --test tests/
@@ -111,6 +137,5 @@ npm test                     # build + node --test tests/
 ## Known limitations / Roadmap
 
 - Mentioning bookmarks/favorites isn't supported yet.
-- `claim_tabs` isn't usable end-to-end.
 - Screenshots entering context require a model route that accepts image input.
 - Windows regression testing is limited.

package/README.md

@@ -1,8 +1,8 @@
 # dsh-tabbit — Tabbit 的 DeepSeek Harness 官方插件
 
-[English](README.en.md) | **简体中文**
+[English](README.en.md) | **简体中文** | [Changelog](CHANGELOG.md)
 
-![Tabbit Browser for DeepSeek Harness](tabbit-for-dsh.png?v=2)
+![Tabbit Browser for DeepSeek Harness](assets/dsh-tabbit-banner.png)
 
 Tabbit Browser 的 DeepSeek Harness（dsh）插件包（bundle）。dsh 可以通过此插件调用 Tabbit 完成 Agent 任务：真实页面、真实登录态、真实交互，经原生 code-first 工具驱动（不走 shell 转发）。适用于网页自动化、信息提取、QA 与评测。
 
@@ -41,6 +41,12 @@ dsh plugin --profile web add link:/path/to/dsh-tabbit   # 本地开发
 
 > 本包取代早期的 `tabbit-browser` skill-only 插件；npm 上的 0.2.x 版本也由本版本接续——0.2.x 用户经每日更新检查会收到升级提示，重跑安装命令即可原地升级。
 
+## 社区与交流
+
+欢迎扫描下方二维码加入 **dsh-tabbit 开发者交流群**，交流使用心得、反馈问题与探讨新特性：
+
+![dsh-tabbit 开发者交流群](assets/dsh-tabbit-developer-community-qr.png.jpg)
+
 ## 设置
 
 ### 基本配置
@@ -74,6 +80,21 @@ Full access 说明：dsh 的 `danger-full-access` 权限预设会给会话写入
 
 ## 开发
 
+项目的开发和测试依赖本地的 [deepseek-harness](https://github.com/deepseek-ai/deepseek-harness) 环境。
+由于 `@deepseek-ai/*` 系列包在 npm 上的发布版本普遍滞后，不能直接作为依赖安装，因此需要先在本地构建好一份 harness 检出，再通过下方脚本将仓库内的 .dsh-harness 符号链接指向它（该路径已在 .gitignore 中忽略）。
+整个流程只需配置这一处，`tsconfig.json` 和 `package.json` 均无需改动。
+
+Deepseek Harness 代码获取及安装（如果已安装请忽略）
+```bash
+git clone https://github.com/deepseek-ai/deepseek-harness && cd deepseek-harness && pnpm install && pnpm build && cd ..
+```
+
+将本机的 Deepseek 环境路径修改并 ln 到本项目，将 `/path/to/deepseek-harness` 改为你的实际目录。
+```bash
+npm run link-harness -- /path/to/deepseek-harness   # 或设置环境变量 DSH_HARNESS_PATH
+```
+
+安装和编译
 ```bash
 pnpm install && pnpm build   # tsc → lib/
 npm test                     # 构建 + node --test tests/
@@ -82,6 +103,5 @@ npm test                     # 构建 + node --test tests/
 ## 已知限制 / 路线图
 
 - 收藏夹/书签的提及暂不支持。
-- `claim_tabs` 端到端暂不可用。
 - 截图进上下文要求当前模型路由支持图像输入。
 - Windows 版本回归测试有限。

package/client/client.js

@@ -35,8 +35,11 @@ var module = { exports: {} }; var exports = module.exports;
  * dsh 客户端插件协议与服务端 Cordis 如出一辙：导出 {name, inject, apply(ctx)}；
  * 这里注入的 'inputTriggers' 是 dsh-client-ui-input-trigger 提供的"输入框
  * 触发菜单"服务（package.json dsh.client.inject 里声明了要它一起装）；
- * 'conversationEvents'/'slots' 是 dsh-client-runtime（web 包核心）恒定
- * 提供的服务，不用额外声明。
+ * 'slots' 是 web 客户端核心恒定提供的服务；会话节点注册服务在宿主
+ * 0.1.2-alpha.1 从 'conversationEvents' 政名为 'uiConversation'，注册入口
+ * 也从 service.register() 移到了 service.events.register()。两者不能作为
+ * 硬注入，否则另一代宿主会让整个客户端插件永久 pending，故在 apply
+ * 中按新接口优先、旧接口兜底的顺序探测。
  * 共享的触发菜单负责渲染候选列表 UI——@tab 部分只提供数据，不需要 React。
  *
  * 功能三：把服务端 /tabbit-info 命令落下的 tabbit/status 会话事件折成
@@ -310,8 +313,8 @@ function TabbitStatusNodeView({ node }) {
 }
 
 /*
- * tabbit/status 事件 → 聊天节点的 Definition（conversationEvents 服务的
- * 契约形状，同 dsh 文档 adding-a-conversation-node 的单事件业务写法）。
+ * tabbit/status 事件 → 聊天节点的 Definition（uiConversation 服务的契约
+ * 形状，同 dsh 文档 adding-a-conversation-node 的单事件业务写法）。
  * 每次执行落一条整值事件，以 event.seq 为节点身份各自独立成行——重复
  * 执行 /tabbit-info 天然多行并存、不互相覆盖；anchorSeq 用事件自身的
  * seq，位置正好落在命令行节点（command/run）之后。start/update 都是纯
@@ -345,7 +348,7 @@ const tabbitStatusNode = {
 
 module.exports = {
   name: 'dsh-tabbit-client',
-  inject: ['inputTriggers', 'conversationEvents', 'slots'],
+  inject: ['inputTriggers', 'slots'],
   apply(ctx) {
     // 观看实例打点：告诉服务端"是哪个浏览器在看这个页面"。fire-and-forget
     // （不 await、双层吞错）——纯属锦上添花的提示，任何失败都不能影响页面。
@@ -370,14 +373,19 @@ module.exports = {
       }
     }
 
-    // 功能三：/tabbit-info 状态卡。conversationEvents/slots 已在 inject 里
-    // 声明（apply 只会在它们就绪后运行），这里仍按本文件惯例做形状防御。
-    // 两者内部都是 ctx.effect 式注册（随插件 fiber 卸载自动回收），直接调。
+    // 功能三：/tabbit-info 状态卡。新宿主使用 uiConversation.events，
+    // 0.1.1 宿主使用 conversationEvents；只要任一接口存在就注册状态卡。
+    // 服务内部都是 ctx.effect 式注册（随插件 fiber 卸载自动回收），直接调。
+    const uiConversation = ctx.get('uiConversation');
     const conversationEvents = ctx.get('conversationEvents');
     const slots = ctx.get('slots');
-    if (conversationEvents && typeof conversationEvents.register === 'function'
-        && slots && typeof slots.inject === 'function') {
-      conversationEvents.register(tabbitStatusNode);
+    const registerStatusNode = typeof uiConversation?.events?.register === 'function'
+      ? (definition) => uiConversation.events.register(definition)
+      : typeof conversationEvents?.register === 'function'
+        ? (definition) => conversationEvents.register(definition)
+        : undefined;
+    if (registerStatusNode && slots && typeof slots.inject === 'function') {
+      registerStatusNode(tabbitStatusNode);
       slots.inject('conversation.chat.node', () => slots.register(
         { name: 'conversation.chat.node', key: 'tabbit-status' },
         TabbitStatusNodeView,

package/lib/core/index.js

@@ -284,6 +284,14 @@ export class TabbitService {
         this.defaultTaskNames.set(agentId, name);
         return name;
     }
+    /*
+     * 只读查询当前已定型的默认任务名——不存在就返回 undefined，绝不像
+     * defaultTaskFor 那样顺手创建并定型一个新名字（给 list_tasks 之类的查询
+     * 路径用：查询本身不该有"顺带开一个任务"的副作用）。
+     */
+    currentDefaultTask(agentId) {
+        return this.defaultTaskNames.get(agentId);
+    }
     /* 登记共享 fetch 任务在某实例上被用过（供 releaseAll 精准清理）。 */
     markFetchTaskUsed(instanceId) {
         this.fetchTaskInstances.add(instanceId);
@@ -455,15 +463,21 @@ const SKILL_FALLBACK_DESCRIPTION = "Recipes for operating the user's Tabbit Brow
  * （正文才是要进模型上下文的内容）。
  * 只认我们自己写的这一层【扁平 key: value】（值可带引号），认不出的行直接
  * 跳过——不为一个几行的 frontmatter 引入 YAML 依赖，也不假装能解析嵌套结构。
+ * 开合分隔符按开头是 \n 还是 \r\n 定档、全程沿用同一种换行符（含内部按行
+ * split 取字段）——容忍 Windows checkout（git autocrlf）产出的 CRLF 文件，
+ * 不强行改写正文本身的换行风格。
  */
-function parseSkillDocument(source) {
-    if (!source.startsWith('---\n'))
+export function parseSkillDocument(source) {
+    const newline = source.startsWith('---\r\n') ? '\r\n' : '\n';
+    const opening = `---${newline}`;
+    if (!source.startsWith(opening))
         return { fields: {}, body: source };
-    const end = source.indexOf('\n---\n', 4);
+    const closing = `${newline}---${newline}`;
+    const end = source.indexOf(closing, opening.length);
     if (end === -1)
         return { fields: {}, body: source };
     const fields = {};
-    for (const line of source.slice(4, end).split('\n')) {
+    for (const line of source.slice(opening.length, end).split(newline)) {
         const match = /^([A-Za-z][\w-]*):\s*(.*)$/u.exec(line);
         if (match === null)
             continue;
@@ -472,7 +486,7 @@ function parseSkillDocument(source) {
         if (value !== '')
             fields[match[1]] = value;
     }
-    return { fields, body: source.slice(end + 5) };
+    return { fields, body: source.slice(end + closing.length) };
 }
 /*
  * 读盘 + 解析，组装 dsh 要的两样东西：候选卡片（list 用）与正文（get 用）。

package/lib/runtime/client.js

@@ -23,12 +23,15 @@
  *  - 任务【只能看到自己打开的或被显式 claim（认领）的标签页】，无法枚举
  *    用户的其它标签页——这是浏览器侧的安全边界；
  *  - 整机最多 8 个并发任务；单次求值最长 120 秒；
- *  - 本客户端用到的 CLI 动词：`nodejs`（求值）、`finish`（结束任务，
- *    keep 语义两代有别，见 finishTask 注释）、`receipt`（查回执）、
- *    `checkpoint`（检查点）、`resource`（分块读资源）、`tasks`（列任务）。
- *    新代 CLI（1.11.16+）另有 tabs/claim/resume/screenshot/inspect/paste
- *    和 persistent 持久模式（JSON 帧协议）——本客户端尚未使用，见浏览器
- *    共享 skill（~/.agents/skills/tabbit）与 TabbitDance 源码。
+ *  - 本客户端用到的 CLI 动词：`nodejs`（求值，创建时可带 --claim-tab）、
+ *    `finish`（结束任务，keep 语义两代有别，见 finishTask 注释）、
+ *    `receipt`（查回执）、`checkpoint`（检查点）、`resource`（分块读资源）、
+ *    `tasks`（列任务）、`claim`（对已存在的任务追加认领标签页，见
+ *    claimTabs()——真机 `--help` 确认过是独立顶层子命令，不是只存在于
+ *    persistent 帧协议里）。新代 CLI（1.11.16+）另有 tabs/resume/
+ *    screenshot/inspect/paste 和 persistent 持久模式（JSON 帧协议）——
+ *    本客户端尚未使用，见浏览器共享 skill（~/.agents/skills/tabbit）与
+ *    TabbitDance 源码。
  *
  * ─── 一次 evaluate 的完整旅程 ─────────────────────────────────────────
  *
@@ -371,6 +374,26 @@ export class TabbitClient {
     async checkpoint(task) {
         return await this.invoke(['checkpoint', '--task', task], '', CONTROL_TIMEOUT_MS);
     }
+    /*
+     * 把一个或多个可用（available）标签页追加认领进一个【已经存在】的任务
+     * （`claim` 命令）。与 evaluate() 的 claimTabs 只在任务创建那一刻生效不同，
+     * 这是独立的一次性 CLI 子命令——真机 `tabbit-cli --help` 的 usage 行确认过
+     * 它是顶层动词（`claim --task <name> --tab <id>...`），不是只存在于 persistent
+     * 帧协议里；真机验证过对一个此前已创建好的任务追加 claim 可行。
+     * 批量原子：文档明确"duplicate, stale, busy, unsupported, or cross-window
+     * inputs do not partially claim"——要么全部认领成功，要么整批失败并抛错，
+     * 调用方不需要处理"部分成功"的中间态。
+     */
+    async claimTabs(task, tabIds) {
+        const argv = ['claim', '--task', task];
+        for (const tab of tabIds)
+            argv.push('--tab', String(tab));
+        const response = (await this.invoke(argv, '', CONTROL_TIMEOUT_MS));
+        return {
+            ...(typeof response.groupId === 'string' ? { groupId: response.groupId } : {}),
+            ...(typeof response.ownedPageCount === 'number' ? { ownedPageCount: response.ownedPageCount } : {}),
+        };
+    }
     /* 列出当前实例上的所有任务（`tasks` 命令）。返回值形状异常时兜底为空数组。 */
     async listTasks() {
         const value = await this.invoke(['tasks'], '', CONTROL_TIMEOUT_MS);

package/lib/tool-browser/index.js

@@ -56,7 +56,8 @@ const DESCRIPTION = [
     '`globalThis` persists across calls in the same task — store state instead of re-scraping. `console.*` output is discarded; communicate via the return value only.',
     'Return small JSON-serializable values. To show yourself a screenshot, save it with `page.screenshot({ path: artifactPath("x.png") })` and return `{ screenshots: [artifactPath("x.png")] }`.',
     'Pass `finish: true` on your last call once this piece of browser work is done, so the task and its tab group close instead of sitting open indefinitely; add `keep_tabs: true` alongside it if the tabs should stay open regardless (e.g. tabs the user pointed you at).',
-    "Pass `list_tabs: true` (no code needed) to list every tab currently open in the user's browser — all windows, including tabs you did not open; metadata only, zero side effects. Use it to find a tab the user referred to, then attach that tab via `claim_tabs` when creating a new task.",
+    "Pass `list_tabs: true` (no code needed) to list every tab currently open in the user's browser — all windows, including tabs you did not open; metadata only, zero side effects. Use it to find a tab the user referred to, then attach that tab via `claim_tabs` (works whether the task is new or already exists).",
+    'Pass `list_tasks: true` (no code needed) to list the browser tasks this session already has open — a memory jog for when you have lost track after a gap in the conversation, before deciding whether to reuse one or start fresh.',
     'Call `tabbit_browser_install` once before your first browser call in a session, and load the `tabbit` skill for recipes before non-trivial work.',
 ].join(' ');
 /* 渲染进上下文的文本上限（30k 字符），防止巨型结果撑爆模型上下文。 */
@@ -94,7 +95,7 @@ export function apply(ctx) {
             // ctx.tabbit.listAllTabs() 而非求值——所以离线时如实报离线。
             list_tabs: {
                 type: 'boolean',
-                description: "List every tab of the user's running browser instead of running code: returns { tabCount, tabs: [{ tabId, windowId, index, title, url, active, state, group? }], ... }. state 'available' means the tab can be attached via claim_tabs (new task only); 'busy' means another automation task owns it. tabCount is always the full total; at most 100 entries are returned (listTruncated: true when clipped) — narrow with tabs_filter instead of re-listing. Zero side effects; the browser is never launched for a listing. When set, code/task/finish are ignored.",
+                description: "List every tab of the user's running browser instead of running code: returns { tabCount, tabs: [{ tabId, windowId, index, title, url, active, state, group? }], ... }. state 'available' means the tab can be attached via claim_tabs (new task only); 'busy' means another automation task owns it. tabCount is always the full total; at most 100 entries are returned (listTruncated: true when clipped) — narrow with tabs_filter instead of re-listing. Zero side effects; the browser is never launched for a listing. Takes precedence over list_tasks if both are set. When set, code/task/finish are ignored.",
             },
             // tabs_filter：大浏览器（真机 170+ 标签页）上全量清单会顶到渲染截断线，
             // 模型必须能按关键词收窄而不是反复重列。
@@ -102,6 +103,14 @@ export function apply(ctx) {
                 type: 'string',
                 description: 'With list_tabs: case-insensitive substring matched against tab title and url, to find a specific tab in a large browser (e.g. the site name the user mentioned).',
             },
+            // list_tasks：查这个会话自己已经开过哪些任务——纯内存读取会话任务
+            // 登记表（sessionTaskRegistry/defaultTaskNames），不碰 CLI 也不碰
+            // 浏览器。模型隔了几轮容易忘记自己还有任务开着，给它一个零成本的
+            // 记忆点，别每次都靠猜测复用任务名或者重开一个。
+            list_tasks: {
+                type: 'boolean',
+                description: "List the browser tasks this session currently has open, without touching the browser: returns { taskCount, defaultTask?, tasks: [{ task, isDefault }] }. defaultTask is the task an omitted `task` argument would target next. Use this when you've lost track of what you left open earlier in the conversation. Ignored if list_tabs is also set. When set, code/task/finish are ignored.",
+            },
             // task：显式指定任务名（不同名字 = 完全独立的浏览器状态）。
             // 一般不传，用会话默认任务即可。
             task: {
@@ -125,15 +134,18 @@ export function apply(ctx) {
                 type: 'integer',
                 description: 'Per-call evaluation timeout in milliseconds, max 120000 (default 120000).',
             },
-            // claim_tabs：把用户【明确指给模型】的已有标签页认领进任务。
-            // Runtime Service 的规则是"只在任务创建那一刻生效"，复用任务时它把
-            // claim【静默忽略】——静默成功最坑，所以 client.ts 的 evaluateOnce 看到
-            // 回执里 reused=true 就自己抛 CLAIM_REQUIRES_NEW_TASK（报错的是我们，
-            // 不是服务端），让模型知道要换个新任务名。
+            // claim_tabs：把用户【明确指给模型】的已有标签页认领进任务。两条路径：
+            //  - 任务是本次调用【新建】的：走 evaluate() 自带的 --claim-tab（创建时
+            //    生效）；
+            //  - 任务【已经存在】（本会话此前用过，见 sessionTasks 判定）：改走
+            //    独立的 claim 一次性子命令（client.ts 的 claimTabs()）——真机
+            //    确认过这是顶层命令，不是只在任务创建那一刻才生效。
+            // 模型不需要关心走的是哪条路径，两边结果都会体现在返回值里
+            // （claimedTabCount/ownedPageCount 或 claimError）。
             claim_tabs: {
                 type: 'array',
                 items: { type: 'integer' },
-                description: 'Tab ids EXPLICITLY provided by the user to attach their existing tabs to this task. Only honored when the task is first created.',
+                description: 'Tab ids EXPLICITLY provided by the user to attach their existing tabs to this task — works whether this call creates the task or the task already exists.',
             },
             // finish：本次调用后关闭任务（从任务列表消失；除非 keep_tabs，标签组
             // 也一起关）。让模型按任务性质自己判断：一次性查询干完就关；用户可能
@@ -160,6 +172,9 @@ export function apply(ctx) {
         timeoutMs: 160_000,
         async execute(args, exec) {
             const tabbit = ctx.tabbit;
+            // 极端情况下工具可能无所属 agent（如某些编排形态），用 'shared' 兜底。
+            // list_tasks 分支也要用它，故提到最前面统一算一次。
+            const agentId = exec.agent !== undefined ? String(exec.agent.id) : 'shared';
             // list_tabs 分支：直连清单，与求值路径完全无关（不碰任务/agent/CLI）。
             if (args.list_tabs === true) {
                 try {
@@ -202,19 +217,53 @@ export function apply(ctx) {
                     throw error;
                 }
             }
+            // list_tasks 分支：会话任务登记表快照，纯内存读取（sessionTasks/
+            // currentDefaultTask 都不做 I/O），比 list_tabs 还更轻——连 Runtime
+            // Service 的被动端点都不连。
+            if (args.list_tasks === true) {
+                const tasks = tabbit.sessionTasks(agentId);
+                const defaultTask = tabbit.currentDefaultTask(agentId);
+                return {
+                    status: 'succeeded',
+                    taskCount: tasks.length,
+                    ...(defaultTask !== undefined ? { defaultTask } : {}),
+                    tasks: tasks.map((task) => ({ task, isDefault: task === defaultTask })),
+                };
+            }
             // 常规求值必须有代码（schema 层已放开 required，这里补明确报错）。
             if (args.code === undefined || args.code === '') {
                 return {
                     status: 'failed',
-                    error: 'code is required unless list_tabs is true',
+                    error: 'code is required unless list_tabs or list_tasks is true',
                     errorCode: 'MISSING_CODE',
                 };
             }
-            // 极端情况下工具可能无所属 agent（如某些编排形态），用 'shared' 兜底。
-            const agentId = exec.agent !== undefined ? String(exec.agent.id) : 'shared';
             // 任务名决策：显式 task 优先，否则取会话默认任务（首调可被 label 定名）。
             const task = args.task !== undefined && args.task !== '' ? args.task : tabbit.defaultTaskFor(agentId, args.label);
             const client = tabbit.client();
+            const requestedClaims = args.claim_tabs !== undefined && args.claim_tabs.length > 0 ? args.claim_tabs : undefined;
+            // 本会话登记表里已经有这个任务名 → 一定是复用，创建时的 --claim-tab
+            // 语义上不会生效（evaluate 会撞见 reused=true 直接抛错）。改走独立的
+            // claim 子命令，成功/失败都不让整个调用失败，只把结果软性附加到
+            // 最终返回值——求值本身该跑还是照跑（模型可能只是顺手多认领一个
+            // 标签页，不代表这次调用的主要目的就是认领）。
+            let claimOutcome = {};
+            let claimAtCreation = requestedClaims;
+            if (requestedClaims !== undefined && tabbit.sessionTasks(agentId).includes(task)) {
+                claimAtCreation = undefined; // 已经在这里处理了，别再让 evaluate 撞 CLAIM_REQUIRES_NEW_TASK
+                try {
+                    const claimed = await client.claimTabs(task, requestedClaims);
+                    claimOutcome = {
+                        claimedTabCount: requestedClaims.length,
+                        ...(claimed.ownedPageCount !== undefined ? { ownedPageCount: claimed.ownedPageCount } : {}),
+                    };
+                }
+                catch (error) {
+                    claimOutcome = {
+                        claimError: error instanceof TabbitCliError ? friendlyCliError(error) : String(error?.message ?? error),
+                    };
+                }
+            }
             let outcome;
             try {
                 outcome = await client.evaluate({
@@ -222,7 +271,7 @@ export function apply(ctx) {
                     code: args.code,
                     ...(args.read_only === true ? { readOnly: true } : {}),
                     ...(args.timeout_ms !== undefined ? { timeoutMs: args.timeout_ms } : {}),
-                    ...(args.claim_tabs !== undefined && args.claim_tabs.length > 0 ? { claimTabs: args.claim_tabs } : {}),
+                    ...(claimAtCreation !== undefined ? { claimTabs: claimAtCreation } : {}),
                     signal: exec.signal, // 用户取消 → 一路传到 CLI 子进程被杀
                 });
             }
@@ -268,6 +317,7 @@ export function apply(ctx) {
                     error: outcome.errorMessage ?? 'evaluation failed',
                     ...(outcome.taskWasReset ? { taskWasReset: true } : {}),
                     ...(outcome.notes.length > 0 ? { notes: outcome.notes } : {}),
+                    ...claimOutcome,
                     ...finishOutcome,
                 };
             }
@@ -281,6 +331,7 @@ export function apply(ctx) {
                 task,
                 ...(outcome.taskWasReset ? { taskWasReset: true } : {}),
                 ...(outcome.notes.length > 0 ? { notes: outcome.notes } : {}),
+                ...claimOutcome,
                 ...finishOutcome,
             };
             if (decoded === undefined) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-tabbit",
-  "version": "0.3.0",
+  "version": "0.3.3",
   "description": "Tabbit Browser bundle for DeepSeek Harness: a code-first tabbit_browser tool, browser-backed web_fetch, @tab composer mentions, a dedicated page-access permission, an environment preflight with background installer download, and a daily plugin update check.",
   "repository": {
     "type": "git",
@@ -44,7 +44,8 @@
   "scripts": {
     "build": "tsc -p tsconfig.json",
     "prepack": "npm run build",
-    "test": "npm run build && node --test"
+    "test": "npm run build && node --test",
+    "link-harness": "node scripts/link-harness.mjs"
   },
   "engines": {
     "node": "^22.19.0 || >=24.0.0"
@@ -62,15 +63,16 @@
     "@deepseek-ai/dsh-system-prompt": ">=0.1.1-rc.2 || >=0.1.2-alpha.1",
     "@deepseek-ai/dsh-tools": ">=0.1.1-rc.2 || >=0.1.2-alpha.1",
     "@deepseek-ai/dsh-user-approval": ">=0.1.1-rc.2 || >=0.1.2-alpha.1",
+    "@deepseek-ai/dsh-util-values": ">=0.1.1-rc.2 || >=0.1.2-alpha.1",
     "@deepseek-ai/dsh-web": ">=0.1.1-rc.2 || >=0.1.2-alpha.1",
     "@deepseek-ai/schemastery": ">=3.18.0"
   },
   "devDependencies": {
-    "@deepseek-ai/dsh-llm": "0.1.1-rc.2",
-    "@deepseek-ai/dsh-session": "0.1.1-rc.2",
-    "@deepseek-ai/dsh-tools": "0.1.1-rc.2",
-    "@deepseek-ai/dsh-web": "0.1.2-alpha.2",
-    "@deepseek-ai/schemastery": "link:../deepseek-harness/vendor/schemastery",
+    "@deepseek-ai/dsh-llm": "link:./.dsh-harness/packages/llm/llm",
+    "@deepseek-ai/dsh-session": "link:./.dsh-harness/packages/core/session",
+    "@deepseek-ai/dsh-tools": "link:./.dsh-harness/packages/core/tools",
+    "@deepseek-ai/dsh-web": "link:./.dsh-harness/packages/web/web",
+    "@deepseek-ai/schemastery": "link:./.dsh-harness/vendor/schemastery",
     "@types/node": "^24.0.0",
     "typescript": "^5.7.0"
   },