release: v2.0.8 dashboard office documents

.gitignore

@@ -23,6 +23,9 @@ recordings/
 .lab-agent/
 .lab-agent/sessions/
 
+# Composer paperclip uploads in a workspace checkout
+ant-code-uploads/
+
 # Local scratch files created during manual verification
 test.txt
 test_file.txt

CHANGELOG.md

@@ -2,6 +2,41 @@
 
 ## Unreleased
 
+## 2.0.8 - 2026-09-06
+
+This is a small Dashboard workflow release on the 2.0 TypeScript runtime.
+The paperclip can attach PDF, docx, xlsx, pptx, and common text files.
+Documents are saved under `ant-code-uploads/` with a bounded preview for
+the model. The right pane can preview those files, and Open launches the
+workspace copy with the system app. Permission mode ids are unchanged.
+
+### Added
+
+- Dashboard paperclip accepts PDF / Office Open XML / text documents in
+  addition to images. Documents are stored in `ant-code-uploads/` and
+  ingested through `document_intake`.
+- PDF text-layer extraction via `unpdf`. Scanned PDFs are not OCR'd;
+  composer-attached PDFs with no text layer send a few page images to
+  the configured vision model.
+- Right-pane preview for PDF, extracted Office text/tables, and
+  clickable attachment chips after refresh.
+- Open uses the OS file association on the workspace copy.
+- `ant-code-uploads/` is added to the project `.gitignore` on send.
+
+### Upgrade
+
+```sh
+git pull
+npm ci
+npm run verify:install
+npm link
+ant-code --version
+```
+
+`ant-code --version` should print `2.0.8`. Restart a running Dashboard
+and hard-refresh the browser. Gateway config and `.lab-agent` sessions
+do not need to be recreated.
+
 ## 2.0.7 - 2026-09-05
 
 This is a dependency security patch on the 2.0 TypeScript runtime.

README.md

@@ -113,7 +113,7 @@ ant-code doctor
 ant-code
 ```
 
-`ant-code --version` should print `2.0.7`. If you previously linked a
+`ant-code --version` should print `2.0.8`. If you previously linked a
 JavaScript install, run `npm link` again so the global command points at
 `src/cli/index.ts`.
 

THIRD_PARTY_NOTICES.md

@@ -12,6 +12,7 @@ runtime assets from the following third-party packages:
 | `react` | UI runtime dependency | MIT |
 | `katex` | math rendering runtime dependency; Dashboard CSS and font assets | MIT |
 | `mermaid` | Dashboard diagram rendering bundle input | MIT |
+| `unpdf` | PDF text-layer extraction for document_intake and Dashboard paperclip | MIT |
 | `yaml` | Dashboard YAML parsing bundle input | ISC |
 | `esbuild` | development/build tool for Dashboard assets | MIT |
 | `postject` | development/build tool for optional executable packaging | MIT |

config/skills/document-intake/SKILL.md

@@ -15,8 +15,10 @@ argument_hint: 提供文档路径、需要提取的内容类型、摘要粒度
 2. 优先调用 `document_intake`：
    - 支持 txt/md/json/csv/xml/html。
    - 支持轻量解析 docx/pptx/xlsx。
-   - PDF 在核心实现中只给出边界提示。
-3. 如果用户允许，并且本机安装了 MarkItDown，可通过受权限控制的 shell 命令做额外转换。
+   - PDF 抽取内嵌文本层，默认每次最多 20 页；需要后续页时传 `pageStart`。
+   - Dashboard 回形针也可直接附上这些文件；发送时会写入 `ant-code-uploads/` 并抽文本。
+   - 扫描件没有 OCR。回形针附上的无文本层 PDF 会把前几页内嵌图送给视觉模型。
+3. 只有文本层为空、且用户允许时，才考虑本机 MarkItDown 或其他 OCR 转换器。
 4. 对大文件只提取目录、标题、表格预览和关键段落，不输出全文。
 
 ## 输出要求

docs/releases/2.0.8-dashboard-office-documents_zh.md

@@ -0,0 +1,41 @@
+# Ant Code v2.0.8：Dashboard 回形针办公文档与本地预览
+
+发布日期：2026-09-06
+
+这是 2.0 TypeScript 运行时上的小版本。Dashboard 回形针可以附上 PDF、docx、xlsx、pptx 和常见文本，发送后写入当前项目的 `ant-code-uploads/`，并抽出有界预览给模型。右侧栏可预览这些文件，Office/PDF 的「打开」会调用本机系统应用。权限模式 id 不变。本说明处于公开发布候选阶段，不表示 CI、tag 或 GitHub Release 已经完成。
+
+## 对用户工作流的直接变化
+
+- Dashboard 回形针同时接受图片和 PDF / docx / xlsx / pptx / txt / Markdown / CSV / JSON / HTML。图片仍走视觉附件；文档写入工作区 `ant-code-uploads/`，提示里只放短预览和保存路径，避免把全文塞进网关请求。
+- 无文本层的扫描 PDF 不做 OCR。回形针附上的这类 PDF 最多把 2 页缩小后的页面图送给已配置的视觉模型。
+- 点对话里的附件芯片会在右侧栏预览。PDF 用浏览器内嵌查看器；docx / xlsx / pptx 显示抽出的文本或表格。预览栏的「打开」用系统关联应用打开工作区里那份副本，而不是再下载一遍。
+- 刷新或重开任务后，新发送的附件芯片仍可点。上传目录会写入项目 `.gitignore` 的 `ant-code-uploads/`，避免大文件被提交。
+- 老格式 `.doc` / `.xls` / `.ppt` 仍不支持。TUI 没有回形针。
+
+## 配置、数据与安全边界
+
+- 配置格式、凭据存储和网关协议没有迁移。现有配置和 `.lab-agent` 会话可继续使用。
+- 回形针文档落在当前工作区的 `ant-code-uploads/`。这是项目内副本，不是你在文件选择器里选中的原路径。「打开」编辑并保存时，改的是这份副本。
+- 这次没有改权限模式 id。Dashboard 仍只绑定本机回环地址。
+- 自动化测试不能代表所有私有研究数据、外部模型供应商或现场网络环境都已得到验证。
+
+## 升级与产物位置
+
+需要 Node.js 22.18+。从源码更新：
+
+```sh
+git pull
+npm ci
+npm run verify:install
+npm link
+ant-code --version
+```
+
+`ant-code --version` 应显示 `2.0.8`。若以前 link 过旧安装，需要重新 `npm link`。正在跑的 Dashboard 需要重启并硬刷新浏览器。现有网关配置和会话不必重建。本次新增运行时依赖 `unpdf`（MIT），用于 PDF 文本层。
+
+## 验证、限制与待完成门禁
+
+- 开发仓 `npm test`（`node:test` 口径）：tests 1283，pass 1282，fail 1。失败项是 `full access lets fork skills use the base profile toolset instead of the skill allowlist`（TLS `ECONNRESET`），与本次回形针/预览改动无关。Dashboard 文件、服务、运行时与 UI 相关单测已通过。完整 `npm run verify:release` 交给 GitHub CI。
+- 扫描件 PDF 没有 OCR；Office 预览是抽出文本，不是排版还原。Windows 用系统关联应用打开本地文件。
+- 未把 TaxaMask 内嵌主题、`dashboard-embed` 或源码写入守卫带入独立版。独立 Ant Code 发布不会自动更新 TaxaMask 内嵌拷贝。
+- 候选阶段最终结论以正式仓同一提交的 CI 和 Release 门禁为准。

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ant-code/cli",
-  "version": "2.0.7",
+  "version": "2.0.8",
   "description": "AGPL-licensed local coding agent with a terminal UI, local dashboard, tool permissions, skills, MCP integration, and model gateway adapters.",
   "type": "module",
   "private": true,
@@ -96,6 +96,7 @@
     "mermaid": "^11.17.2",
     "react": "^19.2.5",
     "typescript": "6.0.3",
+    "unpdf": "^1.8.1",
     "yaml": "^2.9.0"
   },
   "devDependencies": {

scripts/verify-release-seal.ts

@@ -41,6 +41,12 @@ const REVIEWED_RUNTIME_DEPENDENCIES = new Map([
     license: "Apache-2.0",
     provenanceMarker: "open_source: TypeScript compiler and language service, Apache-2.0 licensed, public npm package"
   }],
+  ["unpdf", {
+    versionSpec: "^1.8.1",
+    installedVersion: "1.8.1",
+    license: "MIT",
+    provenanceMarker: "open_source: unpdf PDF text extraction library, MIT licensed, public npm package"
+  }],
   ["yaml", {
     versionSpec: "^2.9.0",
     installedVersion: "2.9.0",

src/context/builder.ts

@@ -111,7 +111,7 @@ export async function buildInitialContext(options: { cwd: string; config: import
       "- /compact is a context-window operation that prefers a separate model summarization request through the configured lab gateway; if the gateway is unavailable or summarization fails, it falls back to local deterministic compaction.",
       "- Compaction keeps recent messages exactly and stores a bounded redacted summary of older messages for future turns.",
       "- web_fetch is the stable public tool name, but URL fetching is configured MCP-first by default: prefer the fetch MCP when available, then fall back to the built-in bounded fetcher. web_search uses built-in DuckDuckGo HTML only; do not expect Bing/Brave scrapers or search MCP. A self-hosted SearXNG is the only stable no-key search backend. Use network tools only when the task needs current or external information, and cite URLs in user-facing conclusions.",
-      "- document_intake extracts bounded local text from common text/HTML/Office files inside the workspace. It does not fully parse PDFs unless a skill workflow provides an external converter.",
+      "- document_intake extracts bounded local text from common text/HTML/Office files and the text layer of PDFs inside the workspace. PDFs default to a 20-page window; pass pageStart/maxPages to continue. Dashboard paperclip can attach those files directly: images stay vision attachments, documents are saved under ant-code-uploads/ and ingested. Scanned PDFs are not OCR'd; composer-attached PDFs with no text layer send a few embedded page images to the configured vision model.",
       "- MCP is optional. Use mcp_list to inspect configured servers/tools; mcp_call works only when a local or lab-approved MCP server is configured, and missing MCP servers do not disable built-in local tools.",
         "- Recommended no-key MCP servers are enabled by default when self-contained; service-bound entries such as SearXNG/SQLite may remain disabled until configured. Use /mcp doctor before relying on MCP, and /mcp doctor --live when you need tool discovery.",
       "- Skills are local instruction resources discovered from project/configured skill directories. Use skill_list, skill_read, or skill_run before applying a specialized workflow.",

src/core/session-messages.ts

@@ -285,16 +285,57 @@ export function normalizeUserTurnMessage(message: SessionMessage | string | Reco
 
 export function persistableUserTurnMessage(prompt: string, attachments: unknown = []): SessionMessage {
   const normalized = normalizeInputAttachments(attachments);
-  if (normalized.length === 0) {
-    return { role: "user", content: prompt };
+  const documents = Array.isArray(attachments)
+    ? attachments.filter((item): item is Record<string, unknown> => Boolean(item) && typeof item === "object" && item.type === "document")
+    : [];
+  const documentBlocks = documents.map((item) => ({
+    type: "text",
+    text: `[文档附件：${String(item.name ?? "document")}]`
+  }));
+  const chips = persistableAttachmentChips(attachments);
+  const message: SessionMessage = normalized.length === 0 && documentBlocks.length === 0
+    ? { role: "user", content: prompt }
+    : {
+        role: "user",
+        content: [
+          ...(String(prompt ?? "").trim() ? [{ type: "text", text: String(prompt ?? "") }] : []),
+          ...documentBlocks,
+          ...normalized.map(imageAttachmentSummaryBlock)
+        ]
+      };
+  if (chips.length > 0) {
+    message.attachments = chips;
   }
-  return {
-    role: "user",
-    content: [
-      ...(String(prompt ?? "").trim() ? [{ type: "text", text: String(prompt ?? "") }] : []),
-      ...normalized.map(imageAttachmentSummaryBlock)
-    ]
-  };
+  return message;
+}
+
+export function persistableAttachmentChips(attachments: unknown = []): NonNullable<SessionMessage["attachments"]> {
+  if (!Array.isArray(attachments)) {
+    return [];
+  }
+  const chips: NonNullable<SessionMessage["attachments"]> = [];
+  for (const item of attachments) {
+    if (!item || typeof item !== "object") {
+      continue;
+    }
+    const record = item as Record<string, unknown>;
+    const type = record.type === "document" ? "document" : record.type === "image" ? "image" : null;
+    if (!type) {
+      continue;
+    }
+    const storedPath = String(record.path ?? "").trim().replace(/\\/g, "/");
+    chips.push({
+      type,
+      name: String(record.name ?? type).trim().slice(0, 160) || type,
+      mimeType: String(record.mimeType ?? record.mime_type ?? "").trim(),
+      size: nonNegativeInteger(record.size ?? record.bytes ?? record.sizeBytes, 0) ?? 0,
+      ...(storedPath ? { path: storedPath } : {})
+    });
+    if (chips.length >= 10) {
+      break;
+    }
+  }
+  return chips;
 }
 
 
@@ -366,7 +407,7 @@ export function messagesForModelContext(messages: unknown = []): SessionMessage[
     if (!message || typeof message !== "object") {
       return [];
     }
-    const { interruptedDraft: _interruptedDraft, ...rest } = message;
+    const { interruptedDraft: _interruptedDraft, attachments: _attachments, ...rest } = message;
     return [rest];
   });
 }

src/core/session-persist.ts

@@ -52,7 +52,8 @@ import type {
   SessionTurnMetadata
 } from "./session-types.ts";
 import {
-  imageAttachmentSummaryBlock
+  imageAttachmentSummaryBlock,
+  persistableAttachmentChips
 } from "./session-messages.ts";
 import {
   isPlainObject
@@ -326,6 +327,7 @@ export function persistableTranscriptMessages(messages: unknown, session: AgentS
   return persistableMessagesWithOptions(messages, {
     includeThinking: true,
     includeToolCalls: false,
+    includeAttachments: true,
     stripGoalStatus: session?.goal?.enabled === true
   });
 }
@@ -700,6 +702,12 @@ export function persistableMessage(message: unknown, options: Record<string, unk
       persisted.name = message.name;
     }
   }
+  if (options.includeAttachments === true && role === "user") {
+    const chips = persistableAttachmentChips(message.attachments);
+    if (chips && chips.length > 0) {
+      persisted.attachments = chips;
+    }
+  }
   return persisted;
 }