dsh-data-domain-knowledge 0.1.0-alpha.13 → 0.1.0-alpha.15

package/lib/types/index.js

@@ -400,7 +400,7 @@ export function apply(ctx, pluginConfig = {}) {
         }
         const provider = configured.mode === 'fixed' ? configured.provider : defaultProvider;
         const model = configured.mode === 'fixed' ? configured.model : defaultModel;
-        const effectiveReasoningEffort = configured.mode === 'default' ? reasoningEffort : undefined;
+        const effectiveReasoningEffort = configured.reasoningEffort || (configured.mode === 'default' ? reasoningEffort : undefined);
         if (!provider || !model) {
             return { why: `编译模型没有选好（provider=${provider ?? '空'} / model=${model ?? '空'}）` };
         }
@@ -408,7 +408,7 @@ export function apply(ctx, pluginConfig = {}) {
             provider,
             model,
             ...(effectiveReasoningEffort ? { reasoningEffort: effectiveReasoningEffort } : {}),
-            call: async ({ system, prompt, sessionId, onUsage }) => {
+            call: async ({ system, prompt, sessionId, onUsage, onDelta }) => {
                 /**
                  * 先走 DSH 的公开解析接口，再开始发 prompt。这样固定模型过期、默认模型失效时，
                  * 编译会在模型调用边界明确失败，而不是落成一次看似成功的规则抽取。
@@ -417,7 +417,7 @@ export function apply(ctx, pluginConfig = {}) {
                     try {
                         await llm.resolveCallConfig({
                             provider, model,
-                            ...(configured.mode === 'default' && reasoningEffort ? { reasoningEffort } : {}),
+                            ...(effectiveReasoningEffort ? { reasoningEffort: effectiveReasoningEffort } : {}),
                         });
                     }
                     catch (cause) {
@@ -440,13 +440,13 @@ export function apply(ctx, pluginConfig = {}) {
                     // 而不是留空 —— 留空会在 OpenCode 路由上 400。
                     sessionId: sessionId ?? 'knowledge:default',
                 });
-                return collectModelText(llm.stream(request), onUsage);
+                return collectModelText(llm.stream(request), onUsage, onDelta);
             },
         };
     };
     let environment;
     const runtimeOptions = () => ({
-        environmentGuard: () => environment.requireReady(),
+        environmentGuard: kind => environment.requireReady(kind),
         // 0029：家当目录。配置里三个路径为空时，运行时从这里解析模型与解释器。
         resourceRoot: resourceRoot(config.agentRoot),
         // 0032：记忆段。默认库的记忆落在 `<agentRoot>/memory/default`，与知识树平级。
@@ -482,7 +482,7 @@ export function apply(ctx, pluginConfig = {}) {
         // 0044：非默认库的 runtime 也要拿到模型接缝。只传 `config` 不够 —— 默认库走上面
         // 那个 `new KnowledgeRuntime(... runtimeOptions())`，别的库走注册表这条支路，
         // 两边必须都带上，否则「新建一个库 → 导入文档」静默变成规则抽取。
-        runtimeOptions: () => ({ resourceRoot: resourceRoot(config.agentRoot), modelSeam: modelSeamFor, environmentGuard: () => environment.requireReady() }),
+        runtimeOptions: () => ({ resourceRoot: resourceRoot(config.agentRoot), modelSeam: modelSeamFor, environmentGuard: kind => environment.requireReady(kind) }),
     });
     compileModelForBase = id => bases.compileModelOf(id);
     runtime.baseId = bases.defaultId;
@@ -492,14 +492,48 @@ export function apply(ctx, pluginConfig = {}) {
         snapshot: () => ({ config: runtime.config, paths: runtime.resourcePaths() }),
         probe: kind => runtime.environmentProbe(kind),
         reload: () => { runtime.reloadResources(); bases.reloadResources(); },
+        activateSelection: async () => {
+            const keys = ['embeddingProvider', 'embeddingModel', 'rerankProvider', 'rerankModel'];
+            const previous = Object.fromEntries(keys.map(key => [key, config[key]]));
+            const managed = { embeddingProvider: 'wemm-python', embeddingModel: '', rerankProvider: 'bert-python', rerankModel: '' };
+            const write = async (values) => {
+                const group = pluginConfig.knowledge;
+                if (!group)
+                    throw new Error('模型设置不可写');
+                for (const key of keys) {
+                    const reference = group[key];
+                    if (reference && typeof reference.write === 'function')
+                        await reference.write(values[key]);
+                    else
+                        group[key] = values[key];
+                }
+                config = normalizeKnowledge(pluginConfig.knowledge);
+                runtime.updateConfig(config);
+                bases.updateRuntimeConfig(config);
+            };
+            try {
+                await write(managed);
+            }
+            catch (cause) {
+                await write(previous);
+                throw cause;
+            }
+            return () => {
+                const current = normalizeKnowledge(pluginConfig.knowledge);
+                // A later manual setting change belongs to the user and must survive rollback.
+                const restored = Object.fromEntries(keys.map(key => [key, current[key] === managed[key] ? previous[key] : current[key]]));
+                return write(restored);
+            };
+        },
     });
-    // Startup checks imports and resource files; model inference runs on demand.
-    void environment.check(false).catch(() => undefined);
-    const compileModelCatalog = async (kbId) => {
+    // Reuse a matching persisted probe after restart; only stale or missing reports need a resource check.
+    if (environment.view().report.state !== 'ready')
+        void environment.check(false).catch(() => undefined);
+    const compileModelCatalog = async (kbId, queryModelRoute = {}) => {
         const service = ctx.get('agentDefaultModel');
         const defaultRoute = service?.currentSelection?.() ?? {};
         const llm = ctx.get('llm');
-        const providers = [...new Set([defaultRoute.provider, compileModelForBase(kbId).provider].filter((value) => Boolean(value)))];
+        const providers = [...new Set([...(llm?.listProviders?.() ?? []).map(provider => provider.id), defaultRoute.provider, compileModelForBase(kbId).provider].filter((value) => Boolean(value)))];
         const models = [];
         const failures = [];
         if (!llm?.listModels) {
@@ -528,6 +562,21 @@ export function apply(ctx, pluginConfig = {}) {
             }
         }
         const selection = compileModelForBase(kbId);
+        const effortProvider = selection.mode === 'fixed' ? selection.provider : defaultRoute.provider;
+        const effortModel = selection.mode === 'fixed' ? selection.model : defaultRoute.model;
+        const requestedProvider = queryModelRoute.provider || effortProvider;
+        const requestedModel = queryModelRoute.model || effortModel;
+        if (requestedProvider && requestedModel && llm?.resolveModelInfo) {
+            const item = models.find(model => model.provider === requestedProvider && model.id === requestedModel);
+            if (item) {
+                try {
+                    item.reasoning = (await llm.resolveModelInfo(requestedProvider, requestedModel)).reasoning;
+                }
+                catch (cause) {
+                    failures.push(`思考强度目录：${cause instanceof Error ? cause.message : String(cause)}`);
+                }
+            }
+        }
         const selected = selection.mode === 'fixed' && selection.provider && selection.model
             ? models.find(item => item.provider === selection.provider && item.id === selection.model)
             : undefined;
@@ -688,24 +737,59 @@ export function apply(ctx, pluginConfig = {}) {
     // 意图只保留固定的会话审计契约；可编辑词表不再属于插件配置。
     applyIntent(ctx, { intentRoot: () => intentRoot(config.agentRoot) });
     /**
-     * 需要本轮用户审批的知识写工具。`images.publish` 只写入本地素材目录，按用户要求免审批。
+     * 需要本轮用户审批的知识写工具。`image_publish` 只写入本地素材目录，按用户要求免审批。
      *
      * 0044 换过一批名字：`write_knowledge_card`（只接受已确认提议）退役了，因为正文现在
      * 只读、卡片由编译直接写入；`list_card_proposals` 也退役了，因为待办队列不存在了。
      * 新增的两个是**入库之后**的动作：加备注、review 抬分。
      */
-    const writeTools = new Set(['add_source', 'import_sources', 'propose_wiki_cards', 'annotate_wiki_card', 'review_wiki_card', 'set_knowledge_status', 'rebuild_knowledge', 'record_feedback']);
+    ctx.tools.register(tool('run_knowledge_pipeline', 'Run one independent knowledge stage. Import only saves sources by default. Stages: extract Knowledge/Event, normalize existing results, materialize selected complete card groups, build retrieval indexes, or all. Returns a durable background task; inspect it with get_knowledge_pipeline.', {
+        stage: { type: 'string', required: true, enum: ['extract', 'normalize', 'materialize', 'build', 'all'] },
+        kb: { type: 'string' }, source_ids: { type: 'array', description: 'Optional explicit source IDs; omit for all sources. An empty selection is rejected.' },
+        group_keys: { type: 'array', description: 'Optional full card group keys returned by get_knowledge_pipeline.' }, force: { type: 'boolean', description: 'Explicitly rerun the selected stage despite an unchanged checkpoint.' },
+    }, async (args) => withRuntime(args.kb, async (target) => {
+        const pipeline = await target.pipeline();
+        const { task } = pipeline.start(args.stage, { sourceIds: args.source_ids, groupKeys: args.group_keys, force: args.force === true });
+        return success({ task, url: `/knowledge/pages/tasks?kb=${encodeURIComponent(target.baseId)}&task=${encodeURIComponent(task.id)}` });
+    })));
+    ctx.tools.register(tool('get_knowledge_pipeline', 'Read pending stages, materializable card groups, persistent task history, or incremental public model output. Use the returned next cursor as after on the next call.', {
+        kb: { type: 'string' }, task_id: { type: 'string' }, after: { type: 'integer' },
+    }, async (args) => withRuntime(args.kb, async (target) => {
+        const pipeline = await target.pipeline();
+        return args.task_id
+            ? success({ task: pipeline.store.task(args.task_id) ?? null, ...pipeline.store.events(args.task_id, Math.max(0, args.after ?? 0)) })
+            : success({ status: await target.pipelineStatus(), ...pipeline.store.tasks() });
+    })));
+    const writeTools = new Set(['add_source', 'import_sources', 'propose_wiki_cards', 'annotate_wiki_card', 'review_wiki_card', 'set_knowledge_status', 'rebuild_knowledge', 'run_knowledge_pipeline', 'record_feedback']);
     ctx.on('tools/pre-execute', (exec, next) => {
         if (writeTools.has(exec.name))
             return Promise.resolve({ kind: 'ask', reason: '该操作会写入 Tita Knowledge Store，需要本轮用户审批。' });
         return next();
     });
-    ctx.tools.register(tool('recall_knowledge', 'Recall compiled knowledge through keyword and vector candidate routes, rerank wiki sections, then return the top wiki cards with their matched sections and provenance.', { query: { type: 'string', required: true }, limit: { type: 'integer' }, top_k: { type: 'integer' }, top_n: { type: 'integer' }, rerank: { type: 'boolean' }, context_max_chars: { type: 'integer', description: '这次召回的上下文字符上限；省略则用设置里的值（默认 20000，0 = 不限）。超出部分落在系统临时目录，响应里的 budget.overflow_path 就是它的绝对路径 —— 需要更多信息时直接读那个文件。' }, include_problems: { type: 'boolean' }, mode: { type: 'string', enum: ['fts', 'vector', 'hybrid'] }, kb: { type: 'string' }, session_key: { type: 'string' }, task_id: { type: 'string' }, subtask_id: { type: 'string' }, parent_subtask_id: { type: 'string' }, run_id: { type: 'string' }, turn_id: { type: 'string' }, trace_id: { type: 'string' }, domain: { type: 'string' } }, async (args, exec) => withRuntime(args.kb, target => target.recall(args.query, args.top_k ?? args.limit, args.include_problems ?? false, args.mode, lineageOf(args, exec), { contextTopN: args.top_n, rerank: args.rerank, ...(typeof args.context_max_chars === 'number' ? { contextMaxChars: args.context_max_chars } : {}) })), {
+    ctx.tools.register(tool('recall_knowledge', 'Recall compiled Knowledge through keyword and vector candidate routes, rerank WikiCard sections, then return cards with matched sections, provenance, and Event reference IDs. Use get_knowledge_events to read referenced Events.', { query: { type: 'string', required: true }, limit: { type: 'integer' }, top_k: { type: 'integer' }, top_n: { type: 'integer' }, rerank: { type: 'boolean' }, context_max_chars: { type: 'integer', description: '这次召回的上下文字符上限；省略则用设置里的值（默认 20000，0 = 不限）。超出部分落在系统临时目录，响应里的 budget.overflow_path 就是它的绝对路径 —— 需要更多信息时直接读那个文件。' }, include_problems: { type: 'boolean' }, mode: { type: 'string', enum: ['fts', 'vector', 'hybrid'] }, kb: { type: 'string' }, session_key: { type: 'string' }, task_id: { type: 'string' }, subtask_id: { type: 'string' }, parent_subtask_id: { type: 'string' }, run_id: { type: 'string' }, turn_id: { type: 'string' }, trace_id: { type: 'string' }, domain: { type: 'string' } }, async (args, exec) => withRuntime(args.kb, target => target.recall(args.query, args.top_k ?? args.limit, args.include_problems ?? false, args.mode, lineageOf(args, exec), { contextTopN: args.top_n, rerank: args.rerank, ...(typeof args.context_max_chars === 'number' ? { contextMaxChars: args.context_max_chars } : {}) })), {
         call: recallCallView,
         // 查询词从参数来，命中从结果来 —— 两者拼成一张能看懂的账本。
         meta: (args, value) => recallMetaOf({ ...dataOf(value), query: args?.query }),
         result: resultViewOf(recallResultMarkdown),
     }));
+    ctx.tools.register(tool('get_knowledge_events', 'Read Event records referenced by recalled WikiCards. Pass event IDs from card metadata.event_refs; the result includes full timelines, evidence, outcomes, and missing fields.', {
+        event_ids: { type: 'array', required: true, description: 'Event IDs from WikiCard metadata.event_refs; maximum 50 IDs.' },
+        kb: { type: 'string' }, session_key: { type: 'string' }, task_id: { type: 'string' }, subtask_id: { type: 'string' },
+        run_id: { type: 'string' }, turn_id: { type: 'string' }, trace_id: { type: 'string' }, domain: { type: 'string' },
+    }, async (args) => withRuntime(args.kb, async (target) => {
+        const ids = [...new Set((Array.isArray(args.event_ids) ? args.event_ids : [])
+                .filter((id) => typeof id === 'string' && id.trim() !== '').map(id => id.trim()))];
+        if (ids.length === 0)
+            return failure({ events: [], missing_ids: [] }, 'EVENT_IDS_REQUIRED', '请传入 WikiCard 的 event_refs 中的 Event ID。', false);
+        if (ids.length > 50)
+            return failure({ events: [], missing_ids: [] }, 'EVENT_IDS_LIMIT', '一次最多读取 50 个 Event ID；请分批读取该 WikiCard 的 event_refs。', false);
+        const byId = new Map((await target.loadEvents()).map(event => [event.id, event]));
+        const events = ids.flatMap(id => {
+            const event = byId.get(id);
+            return event ? [event] : [];
+        });
+        return success({ events, missing_ids: ids.filter(id => !byId.has(id)) });
+    })));
     ctx.tools.register(tool('check_knowledge_embeddings', 'Inspect the local embedding provider and binary/int8 vector index without loading the model.', { kb: { type: 'string' }, session_key: { type: 'string' } }, async (args) => withRuntime(args.kb, target => target.embeddingHealth())));
     ctx.tools.register(tool('query_knowledge_graph', 'Query entities, neighbors, paths, lists and relation types in the compiled knowledge graph.', { action: { type: 'string', enum: ['entity', 'neighbors', 'path', 'list', 'types'], required: true }, entity_id: { type: 'string' }, id: { type: 'string' }, entity_type: { type: 'string' }, from_id: { type: 'string' }, to_id: { type: 'string' }, type: { type: 'string' }, status: { type: 'string' }, include_problem: { type: 'boolean' }, limit: { type: 'integer' }, depth: { type: 'integer' }, kb: { type: 'string' }, session_key: { type: 'string' }, task_id: { type: 'string' }, subtask_id: { type: 'string' }, parent_subtask_id: { type: 'string' }, run_id: { type: 'string' }, turn_id: { type: 'string' }, trace_id: { type: 'string' }, domain: { type: 'string' } }, async (args, exec) => withRuntime(args.kb, target => target.graphQuery(args.action, args.entity_id ?? args.id, args.entity_type ?? args.type, args.depth ?? 1, { fromId: args.from_id, toId: args.to_id, status: args.status, includeProblem: args.include_problem ?? false, limit: args.limit ?? 50 }, lineageOf(args, exec)))));
     ctx.tools.register(tool('query_source_truth', 'Read SourceRaw (原始可追溯文档) with access audit. Each call returns up to `limit` matched sources and `max_chars` total characters. If `hasMoreInCurrentPage` is true, continue with `offset += returnedChars`. When it is false and `hasMoreSources` is true, continue with `source_offset = nextSourceOffset` and reset `offset` to 0. `truncated` stays true until both the current source page and all matching source pages are complete. Do not answer from partial content.', { query: { type: 'string', required: true }, limit: { type: 'integer' }, max_chars: { type: 'integer' }, offset: { type: 'integer' }, source_offset: { type: 'integer', description: '来源分页游标；当前页字符已读完后才使用 nextSourceOffset，并将 offset 重置为 0' }, kb: { type: 'string' }, session_key: { type: 'string' }, task_id: { type: 'string' }, subtask_id: { type: 'string' }, run_id: { type: 'string' }, turn_id: { type: 'string' }, trace_id: { type: 'string' }, domain: { type: 'string' } }, async (args, exec) => withRuntime(args.kb, target => target.sourceTruth(args.query, args.limit ?? 20, args.max_chars ?? 40000, lineageOf(args, exec), args.offset ?? 0, args.source_offset ?? 0))));
@@ -735,13 +819,14 @@ export function apply(ctx, pluginConfig = {}) {
             buildVersion: base.buildVersion,
         })),
     })));
-    ctx.tools.register(tool('images.publish', 'Publish one local image into a knowledge base assets bucket without user approval. Pass exactly one of `file_path` or `content_base64` (standard Base64 or a data URI); optional `filename` overrides the stored name, and optional `mime_type` must match the file bytes. Supports PNG, JPEG, GIF, WebP, AVIF, BMP and SVG up to 20 MiB. Returns the local asset URL and Markdown image reference.', {
+    const imagePublishParameters = {
         file_path: { type: 'string', description: '宿主可读取的本地图片文件路径；与 content_base64 二选一' },
         content_base64: { type: 'string', description: '图片的标准 Base64，或 data:image/...;base64,...；与 file_path 二选一' },
         filename: { type: 'string', description: '可选文件名；默认使用 file_path 的文件名，Base64 输入默认 image-<hash>.<ext>' },
         mime_type: { type: 'string', description: '可选 MIME 类型；必须与图片二进制内容一致' },
         kb: { type: 'string', description: '知识库 id；省略时写入默认库' },
-    }, async (args) => withRuntime(args.kb, async (target) => {
+    };
+    const publishImage = async (args) => withRuntime(args.kb, async (target) => {
         try {
             return success(await publishKnowledgeImage(target.rootDir, target.baseId, {
                 filePath: args.file_path,
@@ -755,7 +840,9 @@ export function apply(ctx, pluginConfig = {}) {
                 return failure({}, cause.code, cause.message, false, cause.suggestion);
             return failure({}, 'IMAGE_PUBLISH_FAILED', `图片写入失败：${cause instanceof Error ? cause.message : String(cause)}`, true);
         }
-    })));
+    });
+    const imagePublishDescription = 'Publish one local image into a knowledge base assets bucket without user approval. Pass exactly one of `file_path` or `content_base64` (standard Base64 or a data URI); optional `filename` overrides the stored name, and optional `mime_type` must match the file bytes. Supports PNG, JPEG, GIF, WebP, AVIF, BMP and SVG up to 20 MiB. Returns the local asset URL and Markdown image reference.';
+    ctx.tools.register(tool('image_publish', imagePublishDescription, imagePublishParameters, publishImage));
     ctx.tools.register(tool('add_source', 'Add or update a SourceRaw document (原始可追溯文档). DSH requests one user approval before execution.', { title: { type: 'string' }, content: { type: 'string', required: true }, external_id: { type: 'string' }, metadata: { type: 'json' }, source_id: { type: 'string' }, allow_duplicate_title: { type: 'boolean', description: '默认 false：标题与已有来源重复时拒收（SOURCE_TITLE_DUPLICATED），要求改标题。传 true 才允许并存。' }, kb: { type: 'string' }, session_key: { type: 'string' }, task_id: { type: 'string' }, subtask_id: { type: 'string' }, run_id: { type: 'string' }, turn_id: { type: 'string' }, trace_id: { type: 'string' }, domain: { type: 'string' } }, async (args, exec) => withRuntime(args.kb, target => target.addSource({ title: args.title, content: args.content, externalId: args.external_id, metadata: args.metadata, sourceId: args.source_id, ...(args.allow_duplicate_title === true ? { allowDuplicateTitle: true } : {}), lineage: lineageOf(args, exec) })), {
         call: addSourceCallView,
         meta: (args, value) => addSourceMetaOf(args, dataOf(value)),
@@ -775,7 +862,7 @@ export function apply(ctx, pluginConfig = {}) {
      * `compile` 默认开：批量导入的意图几乎总是「进来就能用」。关掉它的场景是「先全部
      * 落盘、稍后统一编译」——那时卡片还没生成，工具返回里会明说 `compile: null`。
      */
-    ctx.tools.register(tool('import_sources', 'Import many SourceRaw documents and compile them into Wiki cards in ONE call (one approval, one compile, per-source failures reported instead of aborting the batch). Use this instead of calling add_source repeatedly. DSH requests one user approval before execution.', { sources: { type: 'array', required: true, description: `来源数组，每项 { title?, content, external_id?, metadata?, allow_duplicate_title? }；一次最多 ${MAX_IMPORT_SOURCES} 份，超了请分批。标题与已有来源重复的那一份会被拒（SOURCE_TITLE_DUPLICATED），逐条报出来而不是整批失败` }, compile: { type: 'boolean', description: '默认 true：写入后立刻编译成 WikiCard。false 表示只落盘' }, kb: { type: 'string' }, session_key: { type: 'string' }, task_id: { type: 'string' }, subtask_id: { type: 'string' }, run_id: { type: 'string' }, turn_id: { type: 'string' }, trace_id: { type: 'string' }, domain: { type: 'string' } }, async (args, exec) => withRuntime(args.kb, target => target.importSources({ sources: (Array.isArray(args.sources) ? args.sources : []).map(item => ({ ...(item?.title ? { title: item.title } : {}), content: item?.content, ...(item?.external_id ? { externalId: item.external_id } : {}), ...(item?.metadata ? { metadata: item.metadata } : {}), ...(item?.allow_duplicate_title === true ? { allowDuplicateTitle: true } : {}) })), ...(args.compile === false ? { compile: false } : {}), lineage: lineageOf(args, exec) })), { call: importCallView, meta: (_args, value) => importMetaOf(dataOf(value)), result: resultViewOf(importResultMarkdown) }));
+    ctx.tools.register(tool('import_sources', 'Import many SourceRaw documents in one call. By default only stores sources. Explicit compile:true runs extract, normalize, materialize, and index publication for the imported documents. Per-source import failures are reported without aborting the batch. DSH requests one user approval before execution.', { sources: { type: 'array', required: true, description: `来源数组，每项 { title?, content, external_id?, metadata?, allow_duplicate_title? }；一次最多 ${MAX_IMPORT_SOURCES} 份，超了请分批。标题与已有来源重复的那一份会被拒（SOURCE_TITLE_DUPLICATED），逐条报出来而不是整批失败` }, compile: { type: 'boolean', description: '默认 false：只保存来源；显式 true 才运行完整抽取、归一与物化' }, kb: { type: 'string' }, session_key: { type: 'string' }, task_id: { type: 'string' }, subtask_id: { type: 'string' }, run_id: { type: 'string' }, turn_id: { type: 'string' }, trace_id: { type: 'string' }, domain: { type: 'string' } }, async (args, exec) => withRuntime(args.kb, target => target.importSources({ sources: (Array.isArray(args.sources) ? args.sources : []).map(item => ({ ...(item?.title ? { title: item.title } : {}), content: item?.content, ...(item?.external_id ? { externalId: item.external_id } : {}), ...(item?.metadata ? { metadata: item.metadata } : {}), ...(item?.allow_duplicate_title === true ? { allowDuplicateTitle: true } : {}) })), compile: args.compile === true, lineage: lineageOf(args, exec) })), { call: importCallView, meta: (_args, value) => importMetaOf(dataOf(value)), result: resultViewOf(importResultMarkdown) }));
     /**
      * `propose_wiki_cards` —— 把一批卡片**直接写进 `02_wiki`**（工具名不变，语义在 0044 改了）。
      *
@@ -829,7 +916,7 @@ export function apply(ctx, pluginConfig = {}) {
         lineage: lineageOf(args, exec),
     }))));
     ctx.tools.register(tool('set_knowledge_status', 'Change a Wiki card status. This status belongs to the Card lifecycle; it does not change Knowledge truth status. problem requires problem_note and recovery must pass through unreviewed. DSH requests one user approval before execution.', { entity_id: { type: 'string', required: true }, status: { type: 'string', enum: ['unreviewed', 'draft', 'reviewed', 'active', 'deprecated', 'problem'], required: true }, problem_note: { type: 'string' }, kb: { type: 'string' }, session_key: { type: 'string' }, task_id: { type: 'string' }, subtask_id: { type: 'string' }, run_id: { type: 'string' }, turn_id: { type: 'string' }, trace_id: { type: 'string' }, domain: { type: 'string' } }, async (args, exec) => withRuntime(args.kb, target => target.setStatus(args.entity_id, args.status, args.problem_note, lineageOf(args, exec)))));
-    ctx.tools.register(tool('rebuild_knowledge', 'Rebuild schema/wiki/compiled knowledge indexes. DSH requests one user approval before execution.', { only: { type: 'string', enum: ['all', 'schema', 'wiki', 'embeddings'] }, kb: { type: 'string' }, session_key: { type: 'string' }, task_id: { type: 'string' }, subtask_id: { type: 'string' }, run_id: { type: 'string' }, turn_id: { type: 'string' }, trace_id: { type: 'string' }, domain: { type: 'string' } }, async (args, exec) => withRuntime(args.kb, target => target.rebuild(args.only ?? 'all', lineageOf(args, exec))), {
+    ctx.tools.register(tool('rebuild_knowledge', 'Rebuild schema/wiki/compiled knowledge indexes from saved cards without document extraction. Use run_knowledge_pipeline for extract, normalize, materialize, or the full chain. DSH requests one user approval before execution.', { only: { type: 'string', enum: ['all', 'schema', 'wiki', 'embeddings'] }, kb: { type: 'string' }, session_key: { type: 'string' }, task_id: { type: 'string' }, subtask_id: { type: 'string' }, run_id: { type: 'string' }, turn_id: { type: 'string' }, trace_id: { type: 'string' }, domain: { type: 'string' } }, async (args, exec) => withRuntime(args.kb, target => target.rebuild(args.only ?? 'all', lineageOf(args, exec), { extract: false })), {
         call: rebuildCallView,
         meta: (args, value) => rebuildMetaOf(args?.only ?? 'all', dataOf(value)),
         result: resultViewOf(rebuildResultMarkdown),
@@ -911,7 +998,7 @@ export function apply(ctx, pluginConfig = {}) {
                     const id = kbIdOf(url.searchParams.get('kb'));
                     if (!bases.get(id))
                         return sendEnvelope(res, failure({}, 'BASE_NOT_FOUND', `知识库不存在: ${id}`, true, '先 GET /knowledge/bases 拿到可用 id。'));
-                    return send(res, 200, success(await compileModelCatalog(id)));
+                    return send(res, 200, success(await compileModelCatalog(id, { provider: url.searchParams.get('provider') ?? undefined, model: url.searchParams.get('model') ?? undefined })));
                 } },
             { kind: 'exact', path: '/knowledge/stats', handler: async (req, res) => {
                     const resolved = resolveBase(query(req).searchParams.get('kb'));
@@ -1010,6 +1097,67 @@ export function apply(ctx, pluginConfig = {}) {
                         return sendEnvelope(res, resolved.failure);
                     return send(res, 200, success({ progress: resolved.runtime.compileProgressView(url.searchParams.get('runId') ?? undefined) }));
                 } },
+            { kind: 'exact', path: '/knowledge/pipeline', handler: async (req, res) => {
+                    try {
+                        const body = req.method === 'POST' ? await readBody(req, 100_000) : {};
+                        const resolved = resolveBase(body.kb ?? query(req).searchParams.get('kb'));
+                        if ('failure' in resolved)
+                            return sendEnvelope(res, resolved.failure);
+                        if (req.method === 'GET')
+                            return send(res, 200, success(await resolved.runtime.pipelineStatus()));
+                        if (req.method !== 'POST')
+                            return send(res, 405, envelopeError('METHOD_NOT_ALLOWED', '知识处理只接受 GET 或 POST'));
+                        const pipeline = await resolved.runtime.pipeline();
+                        const started = pipeline.start(body.stage, { sourceIds: body.sourceIds, groupKeys: body.groupKeys, force: body.force === true });
+                        return send(res, 202, success({ task: started.task }));
+                    }
+                    catch (cause) {
+                        return send(res, 400, envelopeError('PIPELINE_INVALID', cause instanceof Error ? cause.message : String(cause)));
+                    }
+                } },
+            { kind: 'exact', path: '/knowledge/pipeline-tasks', handler: async (req, res) => {
+                    try {
+                        if (req.method !== 'GET')
+                            return send(res, 405, envelopeError('METHOD_NOT_ALLOWED', '任务历史只接受 GET'));
+                        const url = query(req);
+                        const resolved = resolveBase(url.searchParams.get('kb'));
+                        if ('failure' in resolved)
+                            return sendEnvelope(res, resolved.failure);
+                        const pipeline = await resolved.runtime.pipeline();
+                        const limit = Math.min(100, Math.max(1, Number(url.searchParams.get('limit')) || 20));
+                        const offset = Math.max(0, Number(url.searchParams.get('offset')) || 0);
+                        return send(res, 200, success({ ...pipeline.store.tasks(limit, offset), extractions: await resolved.runtime.listExtractionRuns({ limit, offset }) }));
+                    }
+                    catch (cause) {
+                        return send(res, 400, envelopeError('PIPELINE_HISTORY_FAILED', cause instanceof Error ? cause.message : String(cause)));
+                    }
+                } },
+            { kind: 'exact', path: '/knowledge/pipeline-task', handler: async (req, res) => {
+                    try {
+                        const url = query(req);
+                        const body = req.method === 'POST' ? await readBody(req, 20_000) : {};
+                        const resolved = resolveBase(body.kb ?? url.searchParams.get('kb'));
+                        if ('failure' in resolved)
+                            return sendEnvelope(res, resolved.failure);
+                        const pipeline = await resolved.runtime.pipeline();
+                        const id = typeof body.id === 'string' ? body.id : url.searchParams.get('id') ?? '';
+                        const task = pipeline.store.task(id);
+                        if (!task)
+                            return send(res, 404, envelopeError('PIPELINE_NOT_FOUND', '任务不存在'));
+                        if (req.method === 'GET')
+                            return send(res, 200, success({ task, ...pipeline.store.events(id, Math.max(0, Number(url.searchParams.get('after')) || 0)) }));
+                        if (req.method !== 'POST')
+                            return send(res, 405, envelopeError('METHOD_NOT_ALLOWED', '任务只接受 GET 或 POST'));
+                        if (body.action === 'retry')
+                            return send(res, 202, success({ task: pipeline.start(task.stage, { ...task.selection, force: false }).task }));
+                        if (!['pause', 'resume', 'cancel'].includes(body.action))
+                            throw new Error('不支持的任务操作');
+                        return send(res, 200, success({ task: pipeline.control(id, body.action) }));
+                    }
+                    catch (cause) {
+                        return send(res, 409, envelopeError('PIPELINE_TASK_REJECTED', cause instanceof Error ? cause.message : String(cause)));
+                    }
+                } },
             { kind: 'exact', path: '/knowledge/extraction-tasks', handler: async (req, res) => {
                     if (req.method !== 'GET')
                         return send(res, 405, envelopeError('METHOD_NOT_ALLOWED', '抽取任务只接受 GET'));
@@ -1083,7 +1231,7 @@ export function apply(ctx, pluginConfig = {}) {
                         if ('failure' in resolved)
                             return sendEnvelope(res, resolved.failure);
                         return send(res, 202, success(resolved.runtime.startCompile(body.only ?? 'all', lineageOf(body), {
-                            ...(typeof body.extract === 'boolean' ? { extract: body.extract } : {}),
+                            extract: body.extract === true,
                             ...(body.forceExtract === true ? { forceExtract: true } : {}),
                         })));
                     }
@@ -1113,7 +1261,7 @@ export function apply(ctx, pluginConfig = {}) {
                                 ...(item?.external_id ? { externalId: String(item.external_id) } : {}),
                                 ...(item?.metadata && typeof item.metadata === 'object' ? { metadata: item.metadata } : {}),
                             })),
-                            ...(body.compile === false ? { compile: false } : {}),
+                            compile: body.compile === true,
                             lineage: lineageOf(body),
                         }));
                     }
@@ -1206,7 +1354,8 @@ export function apply(ctx, pluginConfig = {}) {
                          * 任何一条轨迹 —— 「这个实体是哪次抽出来的」就答不上了。
                          */
                         const runId = extractionRunId();
-                        const seam = modelSeamFor();
+                        const seam = modelSeamFor(resolved.runtime.baseId);
+                        const draft = await (await resolved.runtime.axisRegistry())?.draft();
                         const axesBlock = await resolved.runtime.axisPromptBlockView();
                         const note = resolved.runtime.extractionNote();
                         const report = await runWikiTrial({
@@ -1216,7 +1365,7 @@ export function apply(ctx, pluginConfig = {}) {
                             ...(typeof body.subject === 'string' && body.subject.trim() ? { subject: body.subject.trim() } : {}),
                             /** 与编译相同：先查注册表，再把有候选的坐标合并成一次模型补判。 */
                             resolveAxisSet: axes => resolved.runtime.resolveTrialAxisSet({
-                                axes,
+                                registry: draft?.registry, axes,
                                 axesBlock,
                                 runId,
                                 call: seam.call,
@@ -1239,7 +1388,7 @@ export function apply(ctx, pluginConfig = {}) {
                             },
                             ...('call' in seam ? { call: seam.call } : {}),
                             ...('why' in seam ? { callWhy: seam.why } : {}),
-                        });
+                        }).finally(() => draft?.close());
                         const summary = trialSummary(report);
                         /**
                          * 落库。写不进去**不改这次试跑的结论** —— 报告照样返回，只是多一句
@@ -1253,24 +1402,7 @@ export function apply(ctx, pluginConfig = {}) {
                             report: { ...report, summary },
                             lineage: lineageOf(body),
                         });
-                        /**
-                         * 把这次抽出来的 Knowledge 收进真源（0063 §2）。
-                         *
-                         * **只在跑一份已入库的文档时收。** 理由有两条：来源要指得回一份**真实的**
-                         * 文档（临时正文的 `documentId` 是一个标题字符串，存下来就是悬空引用）；
-                         * 而「这条知识是哪份文档说的」正是它可验证的前提。
-                         *
-                         * **只收引文核对得上的那些。** 引文对不上原文 = 模型改写了，那条判断不是从
-                         * 原文推出来的（§5.7 的引文核验）。把它当知识存进真源，等于让一句编出来的话
-                         * 看起来像有出处 —— 而报告里已经把它报成例外了，那里才是它该待的地方。
-                         */
-                        const ingest = docId
-                            ? await resolved.runtime.ingestTrialKnowledge({
-                                runId, knowledge: report.knowledge,
-                                objects: report.objects, actions: report.actions, scenes: report.scenes,
-                            })
-                            : { saved: 0, merged: 0, skipped: 0 };
-                        return send(res, 200, success({ ...report, summary, run, ingest }));
+                        return send(res, 200, success({ ...report, summary, run }));
                     }
                     catch (cause) {
                         return send(res, 400, envelopeError('REQUEST_INVALID', cause instanceof Error ? cause.message : String(cause)));
@@ -1523,7 +1655,10 @@ export function apply(ctx, pluginConfig = {}) {
                     return send(res, 400, envelopeError('REQUEST_INVALID', `不认识的 action: ${action || '（空）'}`));
                 } },
             /**
-             * 抽取轨迹（`extraction_run`）：按文档列、按 id 读全文。
+             * 轨迹（`extraction_run`）：按文档或阶段列、按 id 读全文。
+             *
+             * 2026-10-18 起三段共用这张表（`stage` 列：`extract` / `normalize` / `materialize`）。
+             * `stage` 省略 = 三段一起 —— 面板里那份「这份文档的轨迹」就是这么读的。
              *
              * 列表**不返回报告本体** —— 一条报告几十到几百 KB，翻一页列表不该按全文付费。
              * 点开某一条走 `/knowledge/extraction-run?id=…`（单独一条路由，不是路径参数：
@@ -1537,8 +1672,11 @@ export function apply(ctx, pluginConfig = {}) {
                     if ('failure' in resolved)
                         return sendEnvelope(res, resolved.failure);
                     const docId = url.searchParams.get('doc_id');
+                    const stageParam = url.searchParams.get('stage');
+                    const stage = stageParam === 'extract' || stageParam === 'normalize' || stageParam === 'materialize' ? stageParam : undefined;
                     const page = await resolved.runtime.listExtractionRuns({
                         ...(docId === null ? {} : { docId }),
+                        ...(stage === undefined ? {} : { stage }),
                         limit: clampInt(url.searchParams.get('limit'), 20, 1, 100),
                         offset: clampInt(url.searchParams.get('offset'), 0, 0, Number.MAX_SAFE_INTEGER),
                     });
@@ -1665,10 +1803,8 @@ export function apply(ctx, pluginConfig = {}) {
                     return send(res, 200, success({ cases: (await storeOf(resolved.runtime).snapshot()).cases }));
                 } },
             /**
-             * 一次排查结束：落一条 `agent_diagnosis_case` 的 SourceRaw，并写入一张案例卡片。
-             *
-             * 0044：卡片不再只提议 —— 它直接入库并标成 `unreviewed`。响应给 `sourceId` 与
-             * `entityId`，界面据此说清「记录留下了，卡片也建了，等人 review」。
+             * 一次排查结束只落一条 SourceRaw。Case 视图从记录派生；Knowledge / Event 的抽取、
+             * 归一和 WikiCard 物化由知识处理阶段分别执行。
              */
             { kind: 'exact', path: '/diagnosis/run', handler: async (req, res) => {
                     if (req.method !== 'POST')
@@ -1683,9 +1819,9 @@ export function apply(ctx, pluginConfig = {}) {
                     }
                     const outcome = await storeOf(resolved.runtime).recordRun(record);
                     if (outcome.problems.length > 0 && outcome.sourceId === undefined) {
-                        return send(res, 200, failure({ sourceId: '', entityId: '' }, 'RUN_WRITE_FAILED', outcome.problems.join('；'), true));
+                        return send(res, 200, failure({ sourceId: '' }, 'RUN_WRITE_FAILED', outcome.problems.join('；'), true));
                     }
-                    return send(res, 200, success({ sourceId: outcome.sourceId ?? '', entityId: outcome.entityId ?? '' }, outcome.problems));
+                    return send(res, 200, success({ sourceId: outcome.sourceId ?? '' }, outcome.problems));
                 } },
             /**
              * 入库之后的两个人工动作（0044）：加备注、review 抬分。
@@ -1842,6 +1978,31 @@ export function apply(ctx, pluginConfig = {}) {
                             if (rest.length === 2) {
                                 if (method === 'PATCH') {
                                     const body = await readBody(req);
+                                    if (body.compileModel && typeof body.compileModel === 'object') {
+                                        const selection = body.compileModel;
+                                        const catalog = await compileModelCatalog(baseId);
+                                        const provider = selection.mode === 'fixed' ? selection.provider : catalog.default.provider;
+                                        const model = selection.mode === 'fixed' ? selection.model : catalog.default.model;
+                                        if (selection.mode !== 'fixed' && selection.mode !== 'default')
+                                            return send(res, 400, envelopeError('MODEL_INVALID', '请选择跟随默认或固定模型'));
+                                        if (selection.mode === 'fixed' && (!provider || !model))
+                                            return send(res, 400, envelopeError('MODEL_INVALID', '固定模型需要 provider 和 model'));
+                                        const llm = ctx.get('llm');
+                                        if (provider && model && llm?.resolveCallConfig) {
+                                            try {
+                                                const info = selection.reasoningEffort ? await llm.resolveModelInfo(provider, model) : undefined;
+                                                const effort = info?.reasoning?.efforts.find(item => item.id === selection.reasoningEffort);
+                                                if (selection.reasoningEffort && !effort)
+                                                    throw new Error('所选模型不支持这个思考强度，请重新选择');
+                                                await llm.resolveCallConfig({ provider, model, ...(effort ? { reasoningEffort: effort.id } : {}) });
+                                            }
+                                            catch (cause) {
+                                                return send(res, 400, envelopeError('MODEL_INVALID', cause instanceof Error ? cause.message : String(cause)));
+                                            }
+                                        }
+                                        else if (selection.reasoningEffort)
+                                            return send(res, 400, envelopeError('MODEL_INVALID', '模型不可用，无法校验思考强度'));
+                                    }
                                     const updated = await bases.update(baseId, {
                                         name: body.name, description: body.description,
                                         extractionNote: typeof body.extractionNote === 'string' ? body.extractionNote : undefined,
@@ -2089,6 +2250,27 @@ export function apply(ctx, pluginConfig = {}) {
                                 const search = (url.searchParams.get('query') ?? '').trim().toLocaleLowerCase();
                                 if (search)
                                     documents = documents.filter(item => `${item.docId}\n${item.title}\n${item.externalId}`.toLocaleLowerCase().includes(search));
+                                /**
+                                 * 文件类型的计数对着**搜索之后、类型筛选之前**的集合算。
+                                 *
+                                 * 下拉要能切换类型，所以每个类型的数量必须是「当前搜索下有多少份」，
+                                 * 而不是「当前这一页里有多少份」—— 后者会让选项随翻页变来变去。
+                                 */
+                                const extCounts = {};
+                                for (const item of documents) {
+                                    const key = item.ext || '—';
+                                    extCounts[key] = (extCounts[key] ?? 0) + 1;
+                                }
+                                /**
+                                 * 文件类型筛选**在服务端**做（0070）。
+                                 *
+                                 * 原来是「客户端取前 500 份、再在这一页里筛」：超过 500 份的库，第 501 份起
+                                 * 在面板上根本不可达，而表头把这种截断显示成「筛出 500 / 600 个」——
+                                 * 读起来像用户自己开了筛选。既然这里已经支持 `offset`，筛选就该和分页同层。
+                                 */
+                                const ext = (url.searchParams.get('ext') ?? '').trim();
+                                if (ext)
+                                    documents = documents.filter(item => item.ext === ext);
                                 const total = documents.length;
                                 const offset = clampInt(url.searchParams.get('offset'), 0, 0, Number.MAX_SAFE_INTEGER);
                                 const limit = clampInt(url.searchParams.get('limit'), 50, 1, 500);
@@ -2099,7 +2281,7 @@ export function apply(ctx, pluginConfig = {}) {
                                  * 两件事。放成并列的一张表，前端按 `docId` 查，文档类型不用为审计字段让步。
                                  */
                                 const extractionRuns = await resolved.runtime.extractionRunCounts();
-                                return send(res, 200, success({ documents: documents.slice(offset, offset + limit), total, extractionRuns }));
+                                return send(res, 200, success({ documents: documents.slice(offset, offset + limit), total, extCounts, extractionRuns }));
                             }
                             const docId = rest[1];
                             const resolved = resolveBase(kbParam);