@dsh-libing/knowledge-base 0.1.3 → 0.1.4

package/CHANGELOG.md

@@ -1,5 +1,9 @@
 # Changelog
 
+## 0.1.4
+
+- 支持 Solo Team 长期 Agent 的受管资产选择；在原 owner 检查作用域，不复制存储。普通会话保持现有行为。
+
 ## 0.1.3
 
 - 增加有容量上限的 manifest / 完整正文检索缓存，保持正文子串匹配与评分；每次检查文件变化，返回证据仍复读原文核验哈希。

package/README.md

@@ -84,3 +84,7 @@ DOCX 压缩包、上传流、图片和 PDF 渲染均有限制；PDF 在可取消
 图片预处理由独立 worker 执行，同一进程全部知识库与插件实例共用一个名额，最多 32 个等待者；等待和运行均可取消。每张图最多 32 MiB / 5000 万像素，单张输出最多 8 MiB；总览长边与裁片边长最多 4096，裁片最多 64（默认 12）。`imageTimeoutMs` 默认 60 秒、最多 120 秒，超时或卸载会终止 worker 后释放名额。worker 逐张输出并等待官方附件保存完成，裁片仅复制所需区域。该限制约束图片预处理的位图并发，不能作为整个宿主进程的 RSS 上限；PDF 引擎和模型调用仍有各自的生命周期与限制。
 
 检索保留原有完整正文的子串匹配与评分，包括只在正文出现的词。每个 runtime 使用 LRU 缓存（最多 256 个文件、估算载荷 32 MiB），复用 manifest 与完整正文的小写检索视图；超大文件不驻留，超出容量按需重新读盘。每次仍验证知识库可访问性、根路径及文件身份/大小/纳秒修改和变更时间；首次读正文核验哈希，返回候选证据时另外复读原文核验。解析后的 manifest 冻结，构建继承文档使用独立副本。当前 revision 指针不缓存，发布、回收、恢复、彻底删除和关闭会清空缓存。冷检索仍扫描所有片段，缓存载荷预算不是宿主进程总内存限制，也不替代可信本机权限边界。
+
+## 受管 Agent 的知识选择（0.1.4）
+
+公开可选 `libingKnowledgeCatalog` 提供库 ID/名称，不读取其他插件私有存储。Solo Team 0.2.0 提供 `libingKnowledgeAccess` 的实际 Session 绑定；本 owner 在每次调用检查库 ID（含唯一名称解析），过滤清单，拒绝未选库和模型建库/任务管理。已发布版本及原文证据规则保持。失去受管选择服务时不静默开放所有库；普通会话沿用原功能。此处不是 OS 文件 ACL。

package/lib/agent-access.d.ts

@@ -0,0 +1,23 @@
+/** Optional managed-Agent selection is enforced by the knowledge owner on every read. */
+import { type Context } from '@deepseek-ai/cordis';
+import type { Agent } from '@deepseek-ai/dsh-agent';
+import type { Runtime } from './runtime.ts';
+declare module '@deepseek-ai/cordis' {
+    interface Context {
+        libingKnowledgeAccess: {
+            resolveSession(sessionId: string): readonly string[] | undefined;
+        };
+        libingKnowledgeCatalog: {
+            readonly version: 1;
+            list(): Promise<{
+                kbId: string;
+                name: string;
+            }[]>;
+        };
+    }
+}
+export declare function registerKnowledgeCatalog(ctx: Context, runtime: Runtime): void;
+export declare function selectedKnowledge(runtime: Runtime, agent: Agent | undefined, tool: string, raw: Record<string, unknown>): Promise<{
+    allowed?: readonly string[];
+    kbId?: string;
+}>;

package/lib/agent-access.js

@@ -0,0 +1,28 @@
+/** Optional managed-Agent selection is enforced by the knowledge owner on every read. */
+import { Service } from '@deepseek-ai/cordis';
+export function registerKnowledgeCatalog(ctx, runtime) {
+    class Catalog extends Service {
+        version = 1;
+        constructor() { super(ctx, 'libingKnowledgeCatalog'); }
+        async list() { return (await runtime.store.list()).map(item => ({ kbId: item.kbId, name: item.name })); }
+    }
+    new Catalog();
+}
+export async function selectedKnowledge(runtime, agent, tool, raw) {
+    const allowed = agent ? runtime.models.ctx.get('libingKnowledgeAccess')?.resolveSession(agent.id) : undefined;
+    if (allowed === undefined) {
+        if (agent && /^solo:(?:chat|standing-run):/u.test(agent.session.header.agentPreset ?? ''))
+            throw new Error('Managed knowledge selection is unavailable');
+        return {};
+    }
+    if (!['kb_catalog', 'kb_search', 'kb_read', 'kb_read_image'].includes(tool))
+        throw new Error('This Agent cannot administer knowledge libraries');
+    if (raw.kbId === undefined && tool === 'kb_catalog')
+        return { allowed };
+    if (typeof raw.kbId !== 'string')
+        throw new Error('A bound library is required');
+    const kb = await runtime.store.kb(raw.kbId);
+    if (!allowed.includes(kb.kbId))
+        throw new Error('Knowledge library is not bound to this Agent');
+    return { allowed, kbId: kb.kbId };
+}

package/lib/chunks.js

@@ -78,7 +78,8 @@ function metadata(chunk) {
 }
 /** 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 nodes = [{ nodeId: 'root', title: docInfo.title, level: 0, parentId: null, headingPath: [], children: [], blocks: [] }];
+    const 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) {
@@ -88,7 +89,8 @@ export function createChunks(document, docInfo, config) {
         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-${String(++number).padStart(4, '0')}`, title: block.text, level: block.level, parentId: parent.nodeId, headingPath: [...parent.headingPath, block.text], children: [], blocks: [], anchor: block.anchor };
+            const parent = required(stack.at(-1), '父章节');
+            const node = { nodeId: `s-${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);
@@ -102,7 +104,18 @@ export function createChunks(document, docInfo, config) {
         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];
+            let parts;
+            if (block.type === 'table') {
+                parts = splitTable(block, budget);
+            }
+            else {
+                if (['paragraph', 'list', 'note'].includes(block.type)) {
+                    parts = splitParagraph(block, budget);
+                }
+                else {
+                    parts = [block];
+                }
+            }
             for (const part of parts) {
                 const unit = { blocks: [part] };
                 if (part.type === 'figure' && node.blocks[i + 1]?.figureRef === part.occurrenceId) {
@@ -131,9 +144,18 @@ export function createChunks(document, docInfo, config) {
             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 blocks = group.flatMap(u => u.blocks);
+            const 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] : []) };
+            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: (() => {
+                    if (blocks.some(b => b.needsReview) || figureResults.some(f => !f || ['failed', 'skipped'].includes(f.status) || f.needsReview && !f.humanReviewed)) {
+                        return 'needs-review';
+                    }
+                    if (figureResults.some(f => !f.humanReviewed)) {
+                        return 'generated';
+                    }
+                    return '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);