fix: load review backend managers only when selected

server/lib/importScoping.test.js

@@ -147,6 +147,12 @@ describe('Tailcat shared owners stay independent of forwarding (#6570)', () => {
 // [entry, target, why, specifier] — same first three columns as NARROWED above,
 // plus the specifier the call site must still name in its `await import()`.
 const DEFERRED = [
+  ['services/codeReview.js', 'services/lmStudioManager.js',
+    'reads the live endpoint only for a selected LM Studio review', './lmStudioManager.js'],
+  ['services/codeReview.js', 'services/ollamaManager.js',
+    'reads endpoints and model capabilities only for an Ollama review', './ollamaManager.js'],
+  ['services/codeReview.js', 'services/mtplxServerManager.js',
+    'resolves the managed daemon only for an MTPLX review', './mtplxServerManager.js'],
   ['services/agentManagement.js', 'lib/privateSecuritySandbox.js',
     'loads sandbox cleanup only for private assessments', '../lib/privateSecuritySandbox.js'],
   ['services/cos.js', 'services/persistentMindAdapter.js',

server/services/codeReview.backendLoading.test.js

@@ -0,0 +1,28 @@
+import { afterEach, expect, it, vi } from 'vitest';
+import { mockJsonResponse } from '../lib/testHelper.js';
+
+vi.mock('./settings.js', () => ({
+  getSettings: async () => ({}),
+  settingsEvents: { on: vi.fn() },
+}));
+vi.mock('./lmStudioManager.js', () => { throw new Error('Manager unavailable'); });
+vi.mock('./ollamaManager.js', () => { throw new Error('Manager unavailable'); });
+
+import { getCodeReviewDefaults, runLocalCodeReview } from './codeReview.js';
+
+afterEach(() => vi.unstubAllGlobals());
+
+it('reads defaults without managers and reviews an explicit endpoint when capability loading fails', async () => {
+  expect(await getCodeReviewDefaults()).toMatchObject({ reviewers: [] });
+  const fetchMock = vi.fn().mockResolvedValue(mockJsonResponse({
+    choices: [{ message: { content: 'No findings.' } }],
+  }));
+  vi.stubGlobal('fetch', fetchMock);
+
+  expect(await runLocalCodeReview({
+    backend: 'ollama', model: 'example-coder', effort: 'high',
+    diff: 'example diff', baseUrl: 'http://localhost:11434',
+  })).toMatchObject({ ok: true, findings: 'No findings.' });
+  expect(fetchMock).toHaveBeenCalledTimes(1);
+  expect(JSON.parse(fetchMock.mock.calls[0][1].body)).toMatchObject({ reasoning_effort: 'high' });
+});

server/services/codeReview.js

@@ -51,11 +51,6 @@ import {
   resolveGoalFidelityConfig,
 } from '../lib/goalFidelity.js'
 import { getSettings, settingsEvents } from './settings.js'
-import { getBaseUrl as getLmStudioBaseUrl } from './lmStudioManager.js'
-import {
-  getBaseUrl as getOllamaBaseUrl,
-  getModelCapabilities as getOllamaModelCapabilities,
-} from './ollamaManager.js'
 
 // LM Studio (`:1234`), Ollama (`:11434`) and MTPLX (`:8000/v1`) all ship
 // OpenAI-compatible `/v1/chat/completions`. Resolve through each manager's live
@@ -64,16 +59,13 @@ import {
 // otherwise the catalog UI and the reviewer would silently desync when a user
 // relocates their install.
 //
-// Every entry is awaited at the call site, which lets MTPLX's stay a DYNAMIC
-// import. That is deliberate: `mtplxServerManager.js` pulls in the managed-daemon
-// watcher and its PM2/filesystem graph, and this module is imported by the agent
-// spawn path — a static import would put that whole graph behind every one of its
-// importers (and did break suites that partially mock `lib/fileUtils.js`). The
-// review request is a one-off HTTP call, so paying the resolve lazily costs
-// nothing.
+// Every entry is awaited at the call site. Keep all manager imports lazy:
+// defaults-only callers (agent prompting, task generation and cleanup) must not
+// load model download/install or daemon-management dependencies. Each manager
+// remains the owner of its live endpoint; only a selected backend loads it.
 const BACKEND_BASE_URLS = {
-  lmstudio: () => getLmStudioBaseUrl(),
-  ollama: () => getOllamaBaseUrl(),
+  lmstudio: async () => (await import('./lmStudioManager.js')).getBaseUrl(),
+  ollama: async () => (await import('./ollamaManager.js')).getBaseUrl(),
   mtplx: async () => (await import('./mtplxServerManager.js')).getMtplxServerEndpoint(),
 }
 
@@ -368,7 +360,9 @@ async function modelRejectsThinking(backend, model) {
   const cacheKey = thinkingCacheKey(backend, model)
   if (thinkingUnsupportedModels.get(cacheKey) === true) return true
   if (backend !== 'ollama') return false
-  const capabilities = await getOllamaModelCapabilities(model).catch(() => null)
+  const capabilities = await import('./ollamaManager.js')
+    .then(({ getModelCapabilities }) => getModelCapabilities(model))
+    .catch(() => null)
   if (!Array.isArray(capabilities) || capabilities.length === 0) return false
   if (capabilities.includes('thinking')) return false
   thinkingUnsupportedModels.set(cacheKey, true)