perf: reduce idle CPU/memory with lazy loading (#3)

## Changed
- Make insight polling non-overlapping and disable when poll interval is <= 0
- Lazy-load search backend client and chat adapters to avoid loading unused SDKs
- Prune expired in-memory conversation entries on save (rate-limited)

Author: kgrubb

src/agent/tools/registry.ts

@@ -48,7 +48,7 @@ export async function createToolRegistry(opts: {
   policy: AgentPolicy;
 }): Promise<ToolRegistry> {
   const log = getLogger({ component: 'agent.tools' });
-  const client = createSearchClient(opts.config.search);
+  const client = await createSearchClient(opts.config.search);
   const fields = await waitForOpenSearchFieldMapping({
     client,
     indexPattern: opts.config.search.indexPattern,

src/insights/engine.ts

@@ -37,7 +37,13 @@ export async function startInsightEngine(opts: { config: KaytooConfig; insightSi
   const { config } = opts;
   const log = getLogger({ component: 'insights' });
   const controller = new AbortController();
-  const client = createSearchClient(config.search);
+
+  if (config.behavior.pollIntervalSeconds <= 0) {
+    log.info({ pollSeconds: config.behavior.pollIntervalSeconds }, 'insight polling disabled');
+    return { stop: () => controller.abort() };
+  }
+
+  const client = await createSearchClient(config.search);
   const fields = await waitForOpenSearchFieldMapping({
     client,
     indexPattern: config.search.indexPattern,
@@ -55,9 +61,17 @@ export async function startInsightEngine(opts: { config: KaytooConfig; insightSi
   });
   const dedupe = new DedupeStore(config.behavior.dedupeTtlSeconds * 1000);
   const warnAt = new Map<string, number>();
-  const timer = setInterval(() => void pollOnce(), config.behavior.pollIntervalSeconds * 1000);
+  let timer: NodeJS.Timeout | undefined;
+  let inFlight = false;
+
+  const scheduleNext = (): void => {
+    if (controller.signal.aborted) return;
+    timer = setTimeout(() => void pollOnce(), config.behavior.pollIntervalSeconds * 1000);
+  };
 
   async function pollOnce(): Promise<void> {
+    if (inFlight) return;
+    inFlight = true;
     return runWithLogContextAsync({ pollId: randomUUID() }, async () => {
       try {
         if (controller.signal.aborted) return;
@@ -167,6 +181,9 @@ export async function startInsightEngine(opts: { config: KaytooConfig; insightSi
         await postFindings(novel);
       } catch (e) {
         log.error({ ...logErr(e) }, 'poll failed');
+      } finally {
+        inFlight = false;
+        scheduleNext();
       }
     });
   }
@@ -195,7 +212,7 @@ export async function startInsightEngine(opts: { config: KaytooConfig; insightSi
   return {
     stop: () => {
       controller.abort();
-      clearInterval(timer);
+      if (timer) clearTimeout(timer);
     },
   };
 }

src/main.ts

@@ -1,19 +1,10 @@
 import { parseKaytooArgv } from './cli/argv.js';
 import { getConfig } from './config.js';
-import { createSlackNotifierWithRetry } from './notify/slack.js';
 import { createConsoleInsightSink } from './notify/consoleInsightSink.js';
 import { startInsightEngine } from './insights/engine.js';
-import { ChatRouter } from './chat/router.js';
-import { startSlackSocketAdapter } from './chat/adapters/slackSocket.js';
-import { startMatrixAdapter } from './chat/adapters/matrix.js';
-import { startMattermostAdapter } from './chat/adapters/mattermost.js';
-import { createAgentRuntime } from './agent/runtime.js';
-import { createSlackChatNotifier } from './notify/slackNotifier.js';
-import { createMatrixNotifier } from './notify/matrixNotifier.js';
-import { createMattermostNotifier } from './notify/mattermostNotifier.js';
-import { createMultiNotifier, createPromiseBackedNotifier } from './notify/multiNotifier.js';
 import { getLogger, initLogging } from './logging/logger.js';
 import { startE2eHttpChatServer } from './e2eHttpChatServer.js';
+import type { ChatRouter as ChatRouterT } from './chat/router.js';
 
 const argv = parseKaytooArgv();
 const config = getConfig(process.env, argv);
@@ -28,7 +19,13 @@ log.info(
 const insightSink =
   config.output === 'console'
     ? createConsoleInsightSink(getLogger({ component: 'insights.out' }))
-    : createSlackNotifierWithRetry({ botToken: config.slack!.botToken!, log: getLogger({ component: 'notify.slack' }) });
+    : await (async () => {
+        const { createSlackNotifierWithRetry } = await import('./notify/slack.js');
+        return createSlackNotifierWithRetry({
+          botToken: config.slack!.botToken!,
+          log: getLogger({ component: 'notify.slack' }),
+        });
+      })();
 
 // Bind e2e HTTP chat before the insight engine: the engine awaits the first poll, which can take a long time
 // (OpenSearch + LLM) and would otherwise block /health and /chat until that poll finishes.
@@ -49,14 +46,33 @@ if (config.output === 'console') {
     'console output mode: insight polling (Slack/Matrix/Mattermost adapters off unless KAYTOO_HTTP_CHAT_BIND set)',
   );
 } else {
+  const [{ ChatRouter }, { createAgentRuntime }, notify, slackSocket, matrixAdapter, mattermostAdapter] =
+    await Promise.all([
+      import('./chat/router.js'),
+      import('./agent/runtime.js'),
+      import('./notify/multiNotifier.js'),
+      import('./chat/adapters/slackSocket.js'),
+      import('./chat/adapters/matrix.js'),
+      import('./chat/adapters/mattermost.js'),
+    ]);
+
+  const { createMultiNotifier, createPromiseBackedNotifier } = notify;
+  const { startSlackSocketAdapter } = slackSocket;
+  const { startMatrixAdapter } = matrixAdapter;
+  const { startMattermostAdapter } = mattermostAdapter;
+
+  const { createSlackChatNotifier } = await import('./notify/slackNotifier.js');
+  const { createMatrixNotifier } = await import('./notify/matrixNotifier.js');
+  const { createMattermostNotifier } = await import('./notify/mattermostNotifier.js');
+
   const slackCfg = config.slack!;
   const slack = insightSink;
   const slackNotifier = createSlackChatNotifier({ slack });
   // Adapters call `onEvent` as soon as they connect; the router is created later, so we
   // resolve the router through a promise until `ChatRouter` is constructed below.
-  const routerCtl = Promise.withResolvers<ChatRouter>();
+  const routerCtl = Promise.withResolvers<ChatRouterT>();
 
-  const onEvent = async (evt: Parameters<ChatRouter['handleEvent']>[0]) => {
+  const onEvent = async (evt: Parameters<ChatRouterT['handleEvent']>[0]) => {
     await (await routerCtl.promise).handleEvent(evt);
   };
 

src/search/client.ts

@@ -1,10 +1,9 @@
-import { Client as ElasticClient } from '@elastic/elasticsearch';
-import { Client as OpenSearchClient } from '@opensearch-project/opensearch';
 import type { KaytooConfig } from '../config.js';
 import type { SearchClient } from './types.js';
 
-export function createSearchClient(config: KaytooConfig['search']): SearchClient {
+export async function createSearchClient(config: KaytooConfig['search']): Promise<SearchClient> {
   if (config.backend === 'elasticsearch') {
+    const { Client: ElasticClient } = await import('@elastic/elasticsearch');
     const client = new ElasticClient({
       node: config.url,
       auth: { username: config.username, password: config.password },
@@ -17,6 +16,7 @@ export function createSearchClient(config: KaytooConfig['search']): SearchClient
     return adapted as unknown as SearchClient;
   }
 
+  const { Client: OpenSearchClient } = await import('@opensearch-project/opensearch');
   const client = new OpenSearchClient({
     node: config.url,
     auth: { username: config.username, password: config.password },

src/storage/conversationStore.ts

@@ -118,6 +118,16 @@ export function createFileConversationStore(opts: {
 /** Non-persistent in-memory store (single process). */
 export function createMemoryConversationStore(opts: { ttlMs: number }): ConversationStore {
   const map = new Map<string, StoredConversation>();
+  const gc = { nextAtMs: 0 };
+  const maybePrune = (): void => {
+    const now = Date.now();
+    if (now < gc.nextAtMs) return;
+    gc.nextAtMs = now + 60_000;
+    const cutoff = now - opts.ttlMs;
+    for (const [k, e] of map) {
+      if (!isStoredConversation(e) || e.updatedAtMs < cutoff) map.delete(k);
+    }
+  };
   return {
     async load(key) {
       const e = map.get(key);
@@ -133,6 +143,7 @@ export function createMemoryConversationStore(opts: { ttlMs: number }): Conversa
       return e;
     },
     async save(key, data) {
+      maybePrune();
       const merged: StoredConversation = { ...data, updatedAtMs: Date.now() };
       if (!isStoredConversation(merged)) throw new Error('invalid conversation payload');
       map.set(key, merged);

test/opensearch/queries.test.ts

@@ -13,7 +13,7 @@ describe('opensearch helpers', () => {
     const { createSearchClient } = await import('../../src/search/client.js');
     const { Client } = (await import('@opensearch-project/opensearch')) as unknown as { Client: ReturnType<typeof vi.fn> };
 
-    createSearchClient({
+    await createSearchClient({
       backend: 'opensearch',
       url: 'https://os.example.com',
       username: 'u',
@@ -27,7 +27,7 @@ describe('opensearch helpers', () => {
     });
 
     Client.mockClear();
-    createSearchClient({
+    await createSearchClient({
       backend: 'opensearch',
       url: 'https://os.example.com',
       username: 'u',

test/searchClient.test.ts

@@ -28,7 +28,7 @@ describe('createSearchClient', () => {
       };
     });
     const { createSearchClient } = await import('../src/search/client.js');
-    const sc = createSearchClient({
+    const sc = await createSearchClient({
       backend: 'elasticsearch',
       url: 'https://es.example:9200',
       username: 'u',
@@ -52,7 +52,7 @@ describe('createSearchClient', () => {
       return { search: vi.fn(), fieldCaps: vi.fn() };
     });
     const { createSearchClient } = await import('../src/search/client.js');
-    createSearchClient({
+    await createSearchClient({
       backend: 'opensearch',
       url: 'https://os.example:9200',
       username: 'u',
@@ -74,7 +74,7 @@ describe('createSearchClient', () => {
       return { search: vi.fn(), fieldCaps: vi.fn() };
     });
     const { createSearchClient } = await import('../src/search/client.js');
-    createSearchClient({
+    await createSearchClient({
       backend: 'opensearch',
       url: 'https://os.example',
       username: 'u',