@dsh-libing/knowledge-base 0.1.0

package/CHANGELOG.md

@@ -0,0 +1,16 @@
+# Changelog
+
+## 0.1.0
+
+- 图片复核改用宽幅双栏与窄屏上下布局，普通文字编辑描述和标识，按需展开类型及技术字段；原图可放大查看，补充各项操作影响，保持标题与底部按钮可见。
+
+- 详情及任务字段提供含义说明，最近操作一条一行、无空白行；长质量报告使用独立滚动区、13px 文字和问题分类，保留完整明细。
+- 待复核任务增加明确的草稿发布与缺口确认入口，区分提交和完成；页头使用说明提供建库流程、按库名的检索指令、普通 query 示例和六个工具的用途及问法。
+
+- 首页打开及弹窗操作按钮统一为官方 32px 控件；新建与回收确认弹窗使用 Portal 独立样式、上下排列标签及铺满内容区的官方 Input，输入框自动获得焦点。
+
+- 侧边栏复用包内透明线条 logo 与宿主尺寸，采用随主题变化的单色；工作台采用官方插件页的 960px 内容区、标题、留白、32px 页头按钮和简洁知识库列表。
+
+- 从旧 knowledge-base 迁移到 strict TypeScript 和 DSH rc.2 服务，拆分版本存储、Word/PDF 转换、模型预算、持久任务、原文检索与客户端工作台。
+- 提供六个 libing_kb 工具、三个内置技能，以及建库、上传/范围预检、任务/草稿、固定版本阅读、图片复核和回收管理页面。
+- 使用独立数据根、原件/片段完整性检查、可取消 PDF worker、持久检查点及显式预算/缺口确认；卸载等待工作退出后释放数据根。

package/README.md

@@ -0,0 +1,80 @@
+# @dsh-libing/knowledge-base
+
+适配 DSH `0.2.0-rc.2` 的 TypeScript 知识库插件。将 Word/PDF 转为分章节的原文片段，保留表格、图片、来源和不可变版本；在 DSH 侧边栏「知识库」管理导入、构建任务、检索、图片复核与回收站。
+
+## 安装与开始
+
+```sh
+dsh plugin --profile web add --save-exact @dsh-libing/knowledge-base
+dsh web
+```
+
+尚未发布时，在仓库执行 `pnpm dev:market --port 18786`，从 LiBing 市场安装本地构建。页面先新建知识库，再上传 `.docx` 或 `.pdf` 开始构建。`.doc` 需额外配置本机 LibreOffice 可执行文件的绝对路径。纯文字导入无需模型；图片理解和扫描 PDF 的版面恢复需要在 DSH 配置支持图片的模型，并在导入选项或插件配置中选定。摘要模型未配置时优先使用宿主默认模型；没有模型则使用原文摘要。
+
+PDF 可以先预检：只抽样核对页数、书签及 OCR 可能性，不调用模型。页范围按原件实际页序、从 1 开始，包含结束页。相同文档 ID 替换当前内容；改变已有 PDF 的部分导入范围必须明确确认，历史版本仍保留。默认按文件名生成文档 ID，替换其他名称的文档可在高级选项填写原文档 ID。
+
+## 任务与复核
+
+任务、逐页收据、模型调用用量和已完成检查点保存在磁盘。取消后可重试；卸载时取消并等待任务退出，重载后继续队列。模型调用和已报告 tokens 受到预算限制，耗尽后暂停，只有明确提高额度才继续。重试复用可用缓存；供应商未报告的用量不能精确计入 tokens。
+
+图片失败、扫描页无法恢复或文档覆盖缺口默认进入「待复核」，原来的已发布版本继续可用。草稿只允许在管理页面预览，Agent 工具不检索草稿。没有阻塞问题时自动发布，任务显示「已发布」；普通解析提示不一定阻止发布。「待复核」任务提供「查看待发布草稿」与「发布草稿」两个入口，后者须在弹窗中确认接受报告中的覆盖缺口。提交后仍需等待任务变为「已发布」；接受缺口不会补回遗漏的图片、扫描页或复杂对象，正式检索仍提示缺口。人工图片描述、装饰图确认、忽略一次引用、恢复和重新识别都生成新任务与新版本；保存时核对当前版本及旧复核哈希，防止并发覆盖。图片描述是生成内容，应结合原图核对。
+
+章节导航和摘要用于定位，证据来自原文片段。检索按关键词、原始标识符与中文词组匹配，提供固定 revision、原件哈希、文档版本、片段路径及 PDF 页/区域。大段原文使用有来源行号的语义窗口与继续读取位置。复制引用后可长期按同一 revision 读取；当前库更新不会让已经打开的阅读器自动切换。回收知识库后该库不再向 Agent 开放，恢复后版本重新可用。
+
+## 工具与技能
+
+| 工具 | 用途 |
+| --- | --- |
+| `libing_kb_build` | 从本机文件或上传收据创建持久构建任务；可显式新建库 |
+| `libing_kb_job` | 查询、取消、重试；明确调整预算或接受缺口 |
+| `libing_kb_catalog` | 列库、章节与分页导航 |
+| `libing_kb_search` | 在指定正式版本检索原文与来源 |
+| `libing_kb_read` | 阅读片段、匹配窗口或接着上次位置读取 |
+| `libing_kb_read_image` | 将该正式版本中的图片作为真实附件交给支持图片的 Agent |
+
+包内提供 `libing-knowledge-base`、`libing-kb-build`、`libing-kb-search` 三个模型/用户均可调用的技能，无需复制到用户技能目录。页头「使用说明」包含建库流程、按知识库名称生成的检索命令，以及 6 个工具的用途和普通问法；未发布库不生成检索命令。知识库 ID 留作内部版本与引用追溯，日常使用只需库名。
+
+将指令粘贴到 **DSH 会话输入框**（不是终端）：
+
+```text
+/libing-kb-search "文档中有哪些关键要求？请引用原文来源。" --kb "接口知识库"
+/libing-kb-build "/绝对路径/手册.pdf" --kb "接口知识库" --pages 11-23
+/libing-knowledge-base 请列出可用知识库及文档目录
+```
+
+问题、库名和路径含空格时加双引号。导入路径是 DSH 所在电脑上的真实路径；也可从页面上传文档。首次建库需明确要求“新建知识库”；`--pages` 指实际 PDF 页序，从 1 开始且包含结束页。工具由 Agent 调用，参数 `kbId` 已同时接受唯一库名与内部 ID，用户不必填写工具 JSON。
+
+普通 query 也可这样说：「请使用知识库‘接口知识库’回答……；只依据已发布原文，注明文档版本、章节和片段，资料不足时明确说明。」常见操作直接说明意图：列出知识库、查找关键要求、补读刚才结果的前后文、核对引用原图，或查看刚才的构建任务。会话需要提供相应技能与 `libing_kb_*` 工具，并配置可用模型；原图核对需要支持图片的模型，技能中心可查看是否已选用技能。
+
+会话检索默认由 Agent 查找已发布版本，再固定同一次回答的引用与续读。页面「浏览与搜索」的关键词检索不调用模型；其中「在所选会话检索」通过官方命令接口发给明确选择的现有会话，指令按库名表达，并附带 `--revision` 保持当前阅读的正式版本，草稿不能发送。更新后如需最新内容，应先选择「打开最新版本」。此前的 JSON 检索指令继续兼容。
+
+工具和技能使用 `libing` 前缀，避免与旧插件同时启用时重名。默认数据根为 `$DSH_HOME/libing/knowledge-base`（未设置 `DSH_HOME` 时位于 `~/.dsh`），不自动移动或改写旧插件数据。迁移既有库建议从保留原件重新导入；不要把未经备份/验证的旧数据目录直接作为新插件的数据根。
+
+## 配置与边界
+
+在官方插件管理中配置，主要字段：
+
+| 配置 | 默认值 |
+| --- | --- |
+| `dataRoot` | 独立数据根，必须为绝对路径 |
+| `chunkTargetChars` / `evidenceMaxChars` | 6000 / 12000 字符（含来源和结构信息） |
+| `visionProvider` / `visionModel` | 空；两者成对配置 |
+| `summaryProvider` / `summaryModel` | 空；两者成对配置 |
+| `maxModelCalls` / `maxModelTokens` | 每任务 1000 / 2000000 |
+| `modelTimeoutMs` / `modelAttempts` | 90000 / 3 |
+| `maxInputBytes` / `maxUncompressedBytes` | 64 MiB / 256 MiB |
+| `maxArchiveEntries` | 10000 |
+| `pdfMaxPages` / `pdfMaxPagePixels` / `pdfPageTimeoutMs` | 2000 / 12000000 / 60000 |
+| `libreOfficePath` | 空，只有 `.doc` 需要 |
+
+DOCX 压缩包、上传流、图片和 PDF 渲染均有限制；PDF 在可取消的独立 worker 中解析。复杂版面、扫描文字和生成图片描述仍需人工核对；加密 PDF 需先解除密码。切分保留合并单元格、表头及图文组合，不强行截断不可分的超长结构，报告会列出超出目标的片段。
+
+同一个数据根只能运行一个任务服务，避免两个 profile 同时写入。浏览器通过 DSH 官方 Connection 认证，上传后只提交收据，不接收任意本机路径。原件/图片读取核对版本、路径边界和内容哈希。彻底删除仅限回收站中的库，需要完整名称和不可恢复确认；会删除该库原件、历史与关联任务，上传收据和其他库不受影响。
+
+本插件不需要向量数据库、Python、外部 PDF 命令或 CI。实际验证平台为 macOS、Node.js 24、DSH rc.2；Windows、LibreOffice `.doc` 转换与真实供应商识别质量需另行验收。
+
+工作台复用官方插件管理的内容宽度、留白、标题和 32px 操作按钮；新建与回收确认弹窗用官方 Input，在标签下方铺满弹窗内容区。导航与包 logo 共用 `icon.svg`：导航使用随主题变化的单色，市场与官方插件管理保留资源的识别颜色。
+
+首页详情显示带说明的知识库 ID、创建时间和当前正式版本；最近操作一条一行，行间不再留空行。任务中模型调用与 tokens 为累计用量 / 预算上限，页、图片、片段分别表示对应阶段进度；未报告的 tokens 不能精确计入。质量报告区可用滚轮或键盘上下滚动，保留全部问题和技术明细，区分待处理问题与普通提示。
+
+图片复核使用双栏弹窗：左侧原图可在新窗口放大查看，右侧直接编辑描述和图中文字（每行一个标识）；窄屏改为上下排列。图片类型、不确定项与完整结构化信息按需展开，识别出的关联和技术字段在保存时保留。重新识别、装饰图确认、忽略及恢复分别提供影响说明；它们直接重建且不保存正在编辑的人工描述，未保存修改会提示确认放弃。弹窗正文可滚动，页头和 32px 底部按钮保持可见。

package/cordis.patch.yml

@@ -0,0 +1,3 @@
+- insert:
+    - id: libing-knowledge-base
+      name: '@dsh-libing/knowledge-base'

package/icon.svg

@@ -0,0 +1,6 @@
+<svg xmlns="http://www.w3.org/2000/svg" width="64" height="64" viewBox="0 0 16 16" fill="none" stroke="#58766b" stroke-width="1.25" stroke-linecap="round" stroke-linejoin="round">
+  <path d="M8.7 14.4H3.65a1.9 1.9 0 0 1-1.9-1.9V3.5a1.9 1.9 0 0 1 1.9-1.9h7.1a1.9 1.9 0 0 1 1.9 1.9v3.1"/>
+  <path d="M4.65 5.2h5.1M4.65 8.1h2.3M4.65 11h1.1"/>
+  <circle cx="10.85" cy="10.75" r="3.1"/>
+  <path d="m13.05 12.95 1.7 1.7"/>
+</svg>

package/lib/chunks.d.ts

@@ -0,0 +1,8 @@
+import type { Block, Chapter, ContentChunk, Document, DocInfo, Source } from './data.ts';
+import type { Settings } from './config.ts';
+export declare function sourceLocations(blocks: Block[]): Source[];
+/** Split enriched, source-preserving blocks; each fragment retains its original block identity. */
+export declare function createChunks(document: Document, docInfo: DocInfo, config: Settings): {
+    nodes: Chapter[];
+    chunks: ContentChunk[];
+};

package/lib/chunks.js

@@ -0,0 +1,197 @@
+import { required } from "./util.js";
+import { posix } from 'node:path';
+import { blockMarkdown } from "./document/convert.js";
+import { figureMarkdown } from "./figures/process.js";
+import { chars, escapeHtml, hash } from "./util.js";
+function tableText(rows) {
+    return '<table>\n' + rows.map((row, i) => '<tr>' + row.map(c => `<${i ? 'td' : 'th'}${c.colspan > 1 ? ` colspan="${c.colspan}"` : ''}${c.rowspan > 1 ? ` rowspan="${c.rowspan}"` : ''}>${escapeHtml(c.text).replace(/\n/g, '<br>')}</${i ? 'td' : 'th'}>`).join('') + '</tr>').join('\n') + '\n</table>';
+}
+function splitParagraph(block, target) {
+    if (chars(block.text) <= target)
+        return [block];
+    const sentences = block.text.match(/[^。！？!?\n]+[。！？!?\n]*|[。！？!?\n]+/gu) || [block.text];
+    const parts = [];
+    let pending = '';
+    for (const sentence of sentences) {
+        if (pending && chars(pending + sentence) > target) {
+            parts.push({ ...block, text: pending, part: parts.length + 1 });
+            pending = '';
+        }
+        pending += sentence;
+    }
+    if (pending)
+        parts.push({ ...block, text: pending, part: parts.length + 1 });
+    return parts;
+}
+function splitTable(block, target) {
+    if (chars(block.markdown ?? '') <= target || block.rows.length < 3)
+        return [block];
+    const header = required(block.rows[0], '表头');
+    // A spanning header cannot safely be repeated or detached from its rows.
+    if (header.some(c => c.rowspan > 1))
+        return [{ ...block, oversizedReason: '表头合并行不能安全重复' }];
+    const groups = [];
+    let start = 1;
+    while (start < block.rows.length) {
+        let end = start + 1;
+        for (let i = start; i < end; i++)
+            for (const c of block.rows[i] || [])
+                end = Math.min(block.rows.length, Math.max(end, i + c.rowspan));
+        groups.push(block.rows.slice(start, end));
+        start = end;
+    }
+    const parts = [];
+    let rows = [];
+    const push = () => {
+        if (rows.length) {
+            const actual = [header, ...rows];
+            parts.push({ ...block, rows: actual, markdown: tableText(actual), text: actual.map(r => r.map(c => c.text).join(' | ')).join('\n'), part: parts.length + 1 });
+            rows = [];
+        }
+    };
+    for (const group of groups) {
+        if (rows.length && chars(tableText([header, ...rows, ...group])) > target)
+            push();
+        rows.push(...group);
+    }
+    push();
+    return parts;
+}
+export function sourceLocations(blocks) {
+    const result = [];
+    for (const b of blocks)
+        for (const location of (b.type === 'table' && b.part && b.rows.flat().some(c => c.source) ? b.rows.flat().flatMap(c => c.source ? [c.source] : []) : b.source?.locations || (b.source ? [b.source] : [])))
+            if (location.page) {
+                const source = { ...location };
+                delete source.locations;
+                const old = result.find(s => s.page === source.page && s.method === source.method && s.caption === (b.caption || ''));
+                if (old && old.bbox && source.bbox)
+                    old.bbox = [Math.min(old.bbox[0], source.bbox[0]), Math.min(old.bbox[1], source.bbox[1]), Math.max(old.bbox[2], source.bbox[2]), Math.max(old.bbox[3], source.bbox[3])];
+                else if (!old)
+                    result.push({ ...source, caption: b.caption || '' });
+            }
+    return result;
+}
+function metadata(chunk) {
+    const fields = ['chunkId', 'docId', 'sourceHash', 'sourceVersion', 'headingPath', 'blockRange', 'assetRefs', 'tableRefs', 'relatedChunks', 'reviewStatus', 'oversizedReason', 'sourceFormat', 'pageRange', 'selectedPages'];
+    return '---\n' + fields.filter(k => chunk[k] !== undefined).map(k => `${k}: ${JSON.stringify(chunk[k])}`).join('\n') + '\n---\n\n';
+}
+/** Split enriched, source-preserving blocks; each fragment retains its original block identity. */
+export function createChunks(document, docInfo, config) {
+    const nodes = [{ nodeId: 'root', title: docInfo.title, level: 0, parentId: null, headingPath: [], children: [], blocks: [] }], stack = [nodes[0]];
+    const ignored = new Set(document.blocks.filter(b => b.type === 'figure' && b.figure?.status === 'ignored').map(b => b.occurrenceId));
+    let number = 0;
+    for (const block of document.blocks) {
+        // Exclude occurrences, not asset IDs: an identical image can be kept elsewhere.
+        if (ignored.has(block.occurrenceId) || Boolean(block.figureRef && ignored.has(block.figureRef)))
+            continue;
+        if (block.type === 'heading') {
+            while (stack.length > 1 && required(stack.at(-1), '章节').level >= block.level)
+                stack.pop();
+            const parent = required(stack.at(-1), '父章节'), node = { nodeId: `s-${String(++number).padStart(4, '0')}`, title: block.text, level: block.level, parentId: parent.nodeId, headingPath: [...parent.headingPath, block.text], children: [], blocks: [], anchor: block.anchor };
+            parent.children.push(node.nodeId);
+            nodes.push(node);
+            stack.push(node);
+        }
+        else
+            required(stack.at(-1), '章节').blocks.push(block);
+    }
+    const chunks = [];
+    for (const node of nodes) {
+        const units = [];
+        for (let i = 0; i < node.blocks.length; i++) {
+            const block = required(node.blocks[i], '正文块');
+            const budget = Math.max(128, config.chunkTargetChars - chars(metadata({ docId: docInfo.docId, headingPath: node.headingPath })) - 400);
+            const parts = block.type === 'table' ? splitTable(block, budget) : ['paragraph', 'list', 'note'].includes(block.type) ? splitParagraph(block, budget) : [block];
+            for (const part of parts) {
+                const unit = { blocks: [part] };
+                if (part.type === 'figure' && node.blocks[i + 1]?.figureRef === part.occurrenceId) {
+                    unit.blocks.push(required(node.blocks[++i], '关联图注'));
+                }
+                if (part.type === 'table' && /^(?:注[:：]?|Note\b)/i.test(node.blocks[i + 1]?.text || ''))
+                    unit.blocks.push(required(node.blocks[i + 1], '表注'));
+                units.push(unit);
+            }
+            if (block.type === 'table' && /^(?:注[:：]?|Note\b)/i.test(node.blocks[i + 1]?.text || ''))
+                i++;
+        }
+        const heading = `# ${docInfo.title}\n\n${node.headingPath.length ? '## ' + node.headingPath.join(' / ') + '\n\n' : ''}`;
+        let group = [];
+        const ancestors = [];
+        let ancestor = node;
+        while (ancestor.parentId) {
+            ancestor = required(nodes.find(n => n.nodeId === ancestor.parentId), '父章节');
+            if (ancestor.nodeId !== 'root')
+                ancestors.unshift(ancestor.nodeId);
+        }
+        const dir = posix.join('wiki/documents', docInfo.docId, ...ancestors, node.nodeId === 'root' ? 'intro' : node.nodeId);
+        const sourcePrefix = posix.relative(dir, posix.join('sources', docInfo.docId));
+        const render = (unit) => unit.blocks.map(b => blockMarkdown(b, document.assets, sourcePrefix) + (b.type === 'figure' ? '\n\n' + figureMarkdown(b) : '')).join('\n\n');
+        const make = () => {
+            if (!group.length)
+                return;
+            const chunkId = `c-${docInfo.docId}-${String(chunks.length + 1).padStart(5, '0')}`;
+            const blocks = group.flatMap(u => u.blocks), body = heading + group.map(render).join('\n\n') + '\n';
+            const figureResults = blocks.filter(b => b.type === 'figure').map(b => required(b.figure, '图片处理结果'));
+            const value = { content: '', fullChars: 0, bodyChars: 0, contentHash: '', summary: '', keywords: [], chunkId, docId: docInfo.docId, sourceHash: docInfo.sourceHash, sourceVersion: docInfo.sourceVersion || '', title: node.title, nodeId: node.nodeId, headingPath: node.headingPath, blockRange: [required(blocks[0], '起始块').blockId, required(blocks.at(-1), '终止块').blockId], assetRefs: [...new Set(blocks.flatMap(b => [b.assetId, b.sourcePreviewAssetId, ...b.sourcePreviewAssetIds || []].filter((x) => typeof x === 'string' && Boolean(x))))], tableRefs: [...new Set(blocks.filter(b => b.type === 'table').map(b => b.blockId))], relatedChunks: [], reviewStatus: blocks.some(b => b.needsReview) || figureResults.some(f => !f || ['failed', 'skipped'].includes(f.status) || f.needsReview && !f.humanReviewed) ? 'needs-review' : figureResults.some(f => !f.humanReviewed) ? 'generated' : 'source', path: posix.join(dir, chunkId + '.md'), body, segments: group.map(u => ({ text: render(u), blockIds: u.blocks.map(b => b.blockId), type: required(u.blocks[0], '语义块').type, ...document.sourceFormat === 'pdf' ? { sourceLocations: sourceLocations(u.blocks) } : {} })), anchors: blocks.flatMap(b => b.anchor ? [b.anchor] : []) };
+            if (document.sourceFormat === 'pdf') {
+                value.sourceFormat = 'pdf';
+                value.sourceLocations = sourceLocations(blocks);
+                const pages = required(value.sourceLocations).map(s => s.page);
+                value.pageRange = [Math.min(...pages), Math.max(...pages)];
+                value.selectedPages = document.selectedPages;
+            }
+            if (chars(body + metadata(value)) > config.chunkTargetChars)
+                value.oversizedReason = blocks.find(b => b.oversizedReason)?.oversizedReason || '完整图文、合并表行或单句超过目标，保留语义块';
+            chunks.push(value);
+            group = [];
+        };
+        for (const unit of units) {
+            if (group.length && chars(heading + [...group, unit].map(render).join('\n\n')) > config.chunkTargetChars - 800)
+                make();
+            group.push(unit);
+        }
+        make();
+    }
+    for (let i = 0; i < chunks.length; i++) {
+        const c = required(chunks[i]);
+        c.relatedChunks = [chunks[i - 1], chunks[i + 1]].flatMap(x => x ? [x.chunkId] : []);
+        c.content = metadata(c) + c.body;
+        c.fullChars = chars(c.content);
+        c.bodyChars = chars(c.body);
+        c.contentHash = hash(c.content);
+    }
+    const anchorMap = {};
+    for (const chunk of chunks)
+        for (const anchor of chunk.anchors)
+            anchorMap[anchor] = chunk.chunkId;
+    for (const node of nodes)
+        if (node.anchor)
+            anchorMap[node.anchor] = chunks.find(c => c.nodeId === node.nodeId || c.headingPath.join('/').startsWith(node.headingPath.join('/')))?.chunkId;
+    for (const chunk of chunks) {
+        for (const ref of document.internalRefs || [])
+            if (chunk.body.includes(`](#${ref.target})`) && anchorMap[ref.target])
+                chunk.relatedChunks.push(required(anchorMap[ref.target]));
+        chunk.relatedChunks = [...new Set(chunk.relatedChunks)].filter(x => x !== chunk.chunkId);
+        const rewrite = (value) => value.replace(/\]\(#([^)]+)\)/g, (match, anchor) => {
+            const target = chunks.find(c => c.chunkId === anchorMap[anchor]);
+            return target ? `](${posix.relative(posix.dirname(chunk.path), target.path)})` : match;
+        });
+        chunk.body = rewrite(chunk.body);
+        for (const segment of chunk.segments)
+            segment.text = rewrite(segment.text);
+        chunk.content = metadata(chunk) + chunk.body;
+        chunk.fullChars = chars(chunk.content);
+        chunk.contentHash = hash(chunk.content);
+        let cursor = chunk.content.indexOf(chunk.body);
+        for (const segment of chunk.segments) {
+            const start = chunk.content.indexOf(segment.text, cursor);
+            if (start < 0)
+                throw new Error('片段语义块无法定位');
+            segment.startLine = chunk.content.slice(0, start).split('\n').length;
+            segment.endLine = segment.startLine + segment.text.split('\n').length - 1;
+            cursor = start + segment.text.length;
+        }
+    }
+    return { nodes: nodes.map(({ blocks: _blocks, ...node }) => node), chunks };
+}

package/lib/client-types/client/BrowsePanel.d.ts

@@ -0,0 +1,13 @@
+import type { LibraryView } from '../contract.ts';
+import type { Actions } from './common.ts';
+export interface SearchSession {
+    id: string;
+    label: string;
+}
+export declare function BrowsePanel({ library, draftId, t, busy, act, dirty, operationError, go, sessions, send }: {
+    library: LibraryView;
+    draftId?: string | undefined;
+    go: (action: () => void) => void;
+    sessions: SearchSession[];
+    send: (id: string, line: string) => Promise<boolean>;
+} & Actions): import("react/jsx-runtime").JSX.Element;

package/lib/client-types/client/Icon.d.ts

@@ -0,0 +1,4 @@
+/** Reuse the package silhouette in the shell's theme-aware navigation color. */
+import { type ReactNode } from 'react';
+import type { PropsRuntime } from '@deepseek-ai/dsh-client-ui-slots';
+export declare function KnowledgeIcon({ size }: Pick<PropsRuntime<'sidebar.panellist'>, 'size'>): ReactNode;

package/lib/client-types/client/ImageReview.d.ts

@@ -0,0 +1,20 @@
+import type { Figure, FigureData } from '../data.ts';
+import type { ModelDirectory, TreeDocument } from '../contract.ts';
+import type { T } from './common.ts';
+export type ReviewAction = 'save' | 'ignore' | 'decoration' | 'restore' | 'reidentify';
+export declare function ImageReview({ t, doc, figure, imageUrl, body, setBody, vision, setVision, directory, busy, stale, error, close, action }: {
+    t: T;
+    doc: TreeDocument;
+    figure: Figure;
+    imageUrl: string;
+    body: FigureData;
+    setBody: (value: FigureData) => void;
+    vision: string;
+    setVision: (value: string) => void;
+    directory: ModelDirectory | undefined;
+    busy: boolean;
+    stale: boolean;
+    error: string;
+    close: () => void;
+    action: (kind: ReviewAction) => void;
+}): import("react/jsx-runtime").JSX.Element;

package/lib/client-types/client/ImportPanel.d.ts

@@ -0,0 +1,19 @@
+import type { State, ModelDirectory } from '../contract.ts';
+import type { Actions } from './common.ts';
+export declare function ModelSelect({ value, set, directory, label, image, defaultLabel }: {
+    value: string;
+    set: (value: string) => void;
+    directory: ModelDirectory | undefined;
+    label: string;
+    image?: boolean;
+    defaultLabel?: string;
+}): import("react/jsx-runtime").JSX.Element;
+export declare function route(value: string): {
+    provider: string;
+    model: string;
+} | null;
+export declare function ImportPanel({ kbId, state, t, busy, act, dirty, submitted }: {
+    kbId: string;
+    state: State;
+    submitted: () => void;
+} & Actions): import("react/jsx-runtime").JSX.Element;

package/lib/client-types/client/JobsPanel.d.ts

@@ -0,0 +1,6 @@
+import type { PublicJob } from '../contract.ts';
+import type { Actions } from './common.ts';
+export declare function JobsPanel({ jobs, t, busy, act, operationError, preview }: {
+    jobs: PublicJob[];
+    preview: (jobId: string) => void;
+} & Omit<Actions, 'dirty'>): import("react/jsx-runtime").JSX.Element;

package/lib/client-types/client/KnowledgePage.d.ts

@@ -0,0 +1,7 @@
+import type { PropsLocale } from '@deepseek-ai/dsh-client-ui-slots';
+import type { UseSessions } from '@deepseek-ai/dsh-client-ui-session/client';
+export interface PageProps extends PropsLocale<'libingKnowledgeBase'> {
+    useSessions: UseSessions;
+    send: (id: string, line: string) => Promise<boolean>;
+}
+export declare function KnowledgePage({ t, useSessions, send }: PageProps): import("react/jsx-runtime").JSX.Element;

package/lib/client-types/client/Metadata.d.ts

@@ -0,0 +1,10 @@
+import type { ReactNode } from 'react';
+export interface MetadataField {
+    label: string;
+    hint: string;
+    value: ReactNode;
+}
+/** Explain public values without requiring users to interpret an unlabeled JSON dump. */
+export declare function Metadata({ fields }: {
+    fields: MetadataField[];
+}): import("react/jsx-runtime").JSX.Element;

package/lib/client-types/client/QualityReport.d.ts

@@ -0,0 +1,6 @@
+import type { PublicJob } from '../contract.ts';
+import type { T } from './common.ts';
+export declare function QualityReport({ report, t }: {
+    report: NonNullable<PublicJob['report']>;
+    t: T;
+}): import("react/jsx-runtime").JSX.Element;

package/lib/client-types/client/SettingsPanel.d.ts

@@ -0,0 +1,6 @@
+import type { LibraryView } from '../contract.ts';
+import type { Actions } from './common.ts';
+export declare function SettingsPanel({ library, t, busy, act, dirty, remove }: {
+    library: LibraryView;
+    remove: () => void;
+} & Actions): import("react/jsx-runtime").JSX.Element;

package/lib/client-types/client/SourceText.d.ts

@@ -0,0 +1,3 @@
+export declare function SourceText({ text }: {
+    text: string;
+}): import("react/jsx-runtime").JSX.Element;

package/lib/client-types/client/UsageGuide.d.ts

@@ -0,0 +1,8 @@
+import type { LibraryView } from '../contract.ts';
+import type { T } from './common.ts';
+export declare function UsageGuide({ libraries, preferred, t, close }: {
+    libraries: LibraryView[];
+    preferred: string;
+    t: T;
+    close: () => void;
+}): import("react/jsx-runtime").JSX.Element;