Deploy: migrate k3s → Docker; search logic → master-agents-svc

- deploy/k3s удалён, deploy/docker добавлен (Caddyfile, docker-compose, searxng)
- chat-svc: agents/models/prompts удалены, использует llm-svc (LLMClient, EmbeddingClient)
- master-agents-svc: SearchOrchestrator, classifier, researcher, actions, widgets
- web-svc: ChatModelSelector, Optimization, Sources удалены; InputBarPlus; UnregisterSW
- geo-device-svc, localization-svc: Dockerfiles
- docs: 02-k3s-services-spec.md, RUNBOOK/TELEMETRY/WORKING удалены

Co-authored-by: Cursor <cursoragent@cursor.com>

services/master-agents-svc/src/lib/agent/searchOrchestrator.ts

@@ -0,0 +1,156 @@
+import type { LlmClient } from '../llm-client.js';
+import SessionManager from '../session.js';
+import type { TextBlock } from '../types.js';
+import type { ClassifierOutput } from '../actions/types.js';
+import { getClassifierPrompt } from '../prompts/classifier.js';
+import { getWriterPrompt } from '../prompts/writer.js';
+import { classify } from './classifier.js';
+import { research } from './researcher.js';
+import { executeAllWidgets } from '../widgets/index.js';
+
+export type SearchOrchestratorConfig = {
+  llm: LlmClient;
+  mode: 'speed' | 'balanced' | 'quality';
+  sources: ('web' | 'discussions' | 'academic')[];
+  fileIds: string[];
+  systemInstructions: string;
+  locale?: string;
+  memoryContext?: string;
+  answerMode?: import('../prompts/writer.js').AnswerMode;
+  responsePrefs?: { format?: string; length?: string; tone?: string };
+  learningMode?: boolean;
+};
+
+export type SearchOrchestratorInput = {
+  chatHistory: { role: string; content: string }[];
+  followUp: string;
+  config: SearchOrchestratorConfig;
+};
+
+export async function runSearchOrchestrator(
+  session: SessionManager,
+  input: SearchOrchestratorInput,
+): Promise<void> {
+  const { chatHistory, followUp, config } = input;
+
+  const classification = await classify({
+    chatHistory,
+    query: followUp,
+    llm: config.llm,
+    locale: config.locale,
+    enabledSources: config.sources,
+  });
+
+  const widgetPromise = executeAllWidgets({
+    chatHistory,
+    followUp,
+    classification,
+    llm: config.llm,
+  }).then((outputs) => {
+    for (const o of outputs) {
+      session.emitBlock({
+        id: crypto.randomUUID(),
+        type: 'widget',
+        data: { widgetType: o.type, params: o.data ?? {} },
+      });
+    }
+    return outputs;
+  });
+
+  let searchPromise: Promise<{ searchFindings: import('../types.js').Chunk[] }> | null = null;
+  if (!classification.classification.skipSearch) {
+    searchPromise = research(session, config.llm, {
+      chatHistory,
+      followUp,
+      classification,
+      config: {
+        mode: config.mode,
+        sources: config.sources,
+        fileIds: config.fileIds,
+        locale: config.locale,
+      },
+    });
+  }
+
+  const [widgetOutputs, searchResults] = await Promise.all([widgetPromise, searchPromise ?? Promise.resolve({ searchFindings: [] })]);
+
+  session.emit('data', { type: 'researchComplete' });
+
+  const MAX_RESULTS_FOR_WRITER = 15;
+  const MAX_CONTENT_PER_RESULT = 180;
+  const findingsForWriter = (searchResults?.searchFindings ?? []).slice(0, MAX_RESULTS_FOR_WRITER);
+  const finalContext =
+    findingsForWriter
+      .map((f, index) => {
+        const content = f.content.length > MAX_CONTENT_PER_RESULT ? f.content.slice(0, MAX_CONTENT_PER_RESULT) + '…' : f.content;
+        return `<result index=${index + 1} title="${String(f.metadata?.title ?? '').replace(/"/g, "'")}">${content}</result>`;
+      })
+      .join('\n') || '';
+
+  const widgetContext = widgetOutputs
+    .map((o) => `<result>${o.llmContext}</result>`)
+    .join('\n-------------\n');
+
+  const finalContextWithWidgets =
+    `<search_results note="These are the search results and assistant can cite these">\n${finalContext}\n</search_results>\n` +
+    `<widgets_result noteForAssistant="Its output is already showed to the user, assistant can use this information to answer the query but do not CITE this as a source">\n${widgetContext}\n</widgets_result>`;
+
+  const writerPrompt = getWriterPrompt(
+    finalContextWithWidgets,
+    config.systemInstructions,
+    config.mode,
+    config.locale,
+    config.memoryContext,
+    config.answerMode,
+    config.responsePrefs,
+    config.learningMode,
+  );
+
+  const answerStream = config.llm.streamText({
+    messages: [
+      { role: 'system', content: writerPrompt },
+      ...chatHistory,
+      { role: 'user', content: followUp },
+    ],
+    options: { maxTokens: 4096 },
+  });
+
+  let responseBlockId = '';
+  let hasContent = false;
+
+  for await (const chunk of answerStream) {
+    if (!chunk.contentChunk && !responseBlockId) continue;
+    if (!responseBlockId) {
+      const block: TextBlock = {
+        id: crypto.randomUUID(),
+        type: 'text',
+        data: chunk.contentChunk ?? '',
+      };
+      session.emitBlock(block);
+      responseBlockId = block.id;
+      if (chunk.contentChunk) hasContent = true;
+    } else {
+      const block = session.getBlock(responseBlockId) as TextBlock | null;
+      if (block) {
+        block.data += chunk.contentChunk ?? '';
+        if (chunk.contentChunk) hasContent = true;
+        session.updateBlock(block.id, [{ op: 'replace', path: '/data', value: block.data }]);
+      }
+    }
+  }
+
+  if (!hasContent && findingsForWriter.length > 0) {
+    const lines = findingsForWriter.slice(0, 10).map((f, i) => {
+      const title = (f.metadata?.title as string) ?? 'Без названия';
+      const excerpt = f.content.length > 120 ? f.content.slice(0, 120) + '…' : f.content;
+      return `${i + 1}. **${title}** — ${excerpt}`;
+    });
+    session.emitBlock({
+      id: crypto.randomUUID(),
+      type: 'text',
+      data: `## По найденным источникам\n\n${lines.join('\n\n')}\n\n*Ответ LLM недоступен. Проверьте модель в Settings.*`,
+    });
+  }
+
+  session.emit('end', {});
+}