fix(code-review): use configured model for title/aux calls (BYOK)

apps/web/src/lib/cloud-agent-next/cloud-agent-client.ts

@@ -96,6 +96,8 @@ export type PrepareSessionInput = {
   initialPayload?: SendMessagePayload;
   mode: AgentMode;
   model: string;
+  /** Optional cheap same-vendor model for title/aux calls (Code Reviewer). */
+  smallModel?: string;
   variant?: string;
   // GitHub-specific params
   githubRepo?: string;

apps/web/src/lib/code-reviews/core/model-selection.test.ts

@@ -1,5 +1,9 @@
-import { resolveEffectiveModel } from './model-selection';
-import type { CodeReviewAgentConfig } from '@kilocode/db/schema-types';
+import {
+  catalogPricesFromStoredModels,
+  resolveCheapSameVendorSmallModel,
+  resolveEffectiveModel,
+} from './model-selection';
+import type { CodeReviewAgentConfig, StoredModel } from '@kilocode/db/schema-types';
 
 const FALLBACK = 'anthropic/claude-sonnet-4.6';
 
@@ -15,6 +19,15 @@ function baseConfig(
   };
 }
 
+function priced(id: string, prompt: string): StoredModel {
+  return {
+    id,
+    name: id,
+    type: 'language',
+    endpoints: [{ pricing: { prompt, completion: prompt } }],
+  };
+}
+
 describe('resolveEffectiveModel', () => {
   it('uses the global model when there are no overrides', () => {
     const result = resolveEffectiveModel(baseConfig(), 'acme/api', FALLBACK);
@@ -119,3 +132,43 @@ describe('resolveEffectiveModel', () => {
     expect(result.source).toBe('global');
   });
 });
+
+describe('resolveCheapSameVendorSmallModel', () => {
+  const catalog = catalogPricesFromStoredModels({
+    'anthropic/claude-sonnet-4.6': priced('anthropic/claude-sonnet-4.6', '0.000003'),
+    'anthropic/claude-haiku-4.5': priced('anthropic/claude-haiku-4.5', '0.0000008'),
+    'anthropic/claude-opus-4.6': priced('anthropic/claude-opus-4.6', '0.000015'),
+    'openai/gpt-5': priced('openai/gpt-5', '0.00001'),
+    'openai/gpt-5-nano': priced('openai/gpt-5-nano', '0.0000001'),
+  });
+
+  it('picks the cheapest strictly-cheaper same-vendor sibling', () => {
+    expect(resolveCheapSameVendorSmallModel('anthropic/claude-sonnet-4.6', catalog)).toBe(
+      'anthropic/claude-haiku-4.5'
+    );
+  });
+
+  it('does not cross vendors', () => {
+    expect(resolveCheapSameVendorSmallModel('openai/gpt-5', catalog)).toBe('openai/gpt-5-nano');
+  });
+
+  it('leaves managed models unset when no cheaper sibling exists', () => {
+    expect(resolveCheapSameVendorSmallModel('anthropic/claude-haiku-4.5', catalog)).toBeUndefined();
+  });
+
+  it('falls back to the primary for sole-model direct-BYOK vendors', () => {
+    expect(resolveCheapSameVendorSmallModel('neuralwatt/glm-5.2-short', catalog)).toBe(
+      'neuralwatt/glm-5.2-short'
+    );
+  });
+
+  it('picks a cheaper BYOK sibling when catalog prices exist', () => {
+    const byokCatalog = catalogPricesFromStoredModels({
+      'neuralwatt/glm-5.2-short': priced('neuralwatt/glm-5.2-short', '0.000002'),
+      'neuralwatt/glm-tiny': priced('neuralwatt/glm-tiny', '0.0000001'),
+    });
+    expect(resolveCheapSameVendorSmallModel('neuralwatt/glm-5.2-short', byokCatalog)).toBe(
+      'neuralwatt/glm-tiny'
+    );
+  });
+});

apps/web/src/lib/code-reviews/core/model-selection.ts

@@ -12,14 +12,113 @@
  * or Bitbucket). See `RepositoryModelOverrideSchema`.
  */
 
-import type { CodeReviewAgentConfig } from '@kilocode/db/schema-types';
+import type { CodeReviewAgentConfig, StoredModel } from '@kilocode/db/schema-types';
+import { DIRECT_BYOK_PROVIDERS_META } from '@/lib/ai-gateway/providers/direct-byok/direct-byok-meta';
+import { getOpenRouterModelsMetadataFromDatabase } from '@/lib/ai-gateway/providers/gateway-models-cache';
 
 export type EffectiveModelSelection = {
   modelSlug: string;
   thinkingEffort: string | null;
   source: 'repository_override' | 'global';
 };
 
+/** Catalog entry used to pick a cheap same-vendor small model. */
+export type CatalogModelPrice = {
+  id: string;
+  /** USD per input token; null when the catalog has no usable prompt price. */
+  promptPrice: number | null;
+};
+
+const DIRECT_BYOK_VENDORS = new Set(Object.keys(DIRECT_BYOK_PROVIDERS_META));
+
+export function modelVendorId(modelId: string): string | undefined {
+  const vendor = modelId.split('/')[0]?.trim();
+  return vendor || undefined;
+}
+
+export function isDirectByokVendor(vendor: string | undefined): boolean {
+  return vendor != null && DIRECT_BYOK_VENDORS.has(vendor);
+}
+
+/**
+ * Build catalog price rows from the OpenRouter `StoredModel` map.
+ * Uses the cheapest endpoint prompt price per model.
+ */
+export function catalogPricesFromStoredModels(
+  models: Record<string, StoredModel>
+): CatalogModelPrice[] {
+  return Object.values(models)
+    .filter(model => (model.type ?? 'language') === 'language' && model.endpoints.length > 0)
+    .map(model => {
+      const prices = model.endpoints
+        .map(endpoint =>
+          endpoint.pricing?.prompt != null ? Number.parseFloat(endpoint.pricing.prompt) : Number.NaN
+        )
+        .filter(price => Number.isFinite(price) && price >= 0);
+      return {
+        id: model.id,
+        promptPrice: prices.length > 0 ? Math.min(...prices) : null,
+      };
+    });
+}
+
+/**
+ * Pick a cheap same-vendor model for Code Reviewer title/aux calls.
+ *
+ * - Prefer the lowest-priced same-vendor sibling that is strictly cheaper than the
+ *   primary when the primary's price is known.
+ * - When the primary has no catalog price, pick the cheapest same-vendor sibling.
+ * - Direct-BYOK vendors with no cheaper sibling fall back to the primary (user's
+ *   key) so aux calls do not fall through to kilo-auto/small → Gemma on Kilo credits.
+ * - Managed vendors with no cheaper sibling leave small_model unset.
+ */
+export function resolveCheapSameVendorSmallModel(
+  primaryModelId: string,
+  catalog: readonly CatalogModelPrice[]
+): string | undefined {
+  const vendor = modelVendorId(primaryModelId);
+  if (!vendor) return undefined;
+
+  const primaryPrice = catalog.find(entry => entry.id === primaryModelId)?.promptPrice ?? undefined;
+
+  const siblings = catalog.filter(
+    entry =>
+      entry.id !== primaryModelId &&
+      modelVendorId(entry.id) === vendor &&
+      entry.promptPrice != null &&
+      Number.isFinite(entry.promptPrice)
+  );
+
+  const cheaper =
+    primaryPrice != null && Number.isFinite(primaryPrice)
+      ? siblings.filter(entry => (entry.promptPrice as number) < primaryPrice)
+      : siblings;
+
+  if (cheaper.length === 0) {
+    return isDirectByokVendor(vendor) ? primaryModelId : undefined;
+  }
+
+  cheaper.sort((a, b) => {
+    const priceDelta = (a.promptPrice as number) - (b.promptPrice as number);
+    return priceDelta !== 0 ? priceDelta : a.id.localeCompare(b.id);
+  });
+  return cheaper[0]?.id;
+}
+
+/**
+ * Resolve the Code Reviewer aux/title small model for a primary review model.
+ * Soft-fails to BYOK-primary or unset when the gateway catalog cannot be loaded.
+ */
+export async function resolveReviewSmallModel(primaryModelId: string): Promise<string | undefined> {
+  const vendor = modelVendorId(primaryModelId);
+  try {
+    const models = await getOpenRouterModelsMetadataFromDatabase();
+    return resolveCheapSameVendorSmallModel(primaryModelId, catalogPricesFromStoredModels(models));
+  } catch {
+    return isDirectByokVendor(vendor) ? primaryModelId : undefined;
+  }
+}
+
 /**
  * Resolve the effective model for a review's repository.
  *

apps/web/src/lib/code-reviews/triggers/prepare-review-payload.test.ts

@@ -15,9 +15,11 @@ const mockFindPreviousCompletedReview = jest.fn();
 const mockUpdatePreviousReviewSummary = jest.fn();
 const mockUpdateRepositoryReviewInstructionsMetadata = jest.fn();
 const mockGenerateReviewPrompt = jest.fn();
+const mockResolveReviewSmallModel = jest.fn();
 
 import type { CodeReviewAgentConfig } from '@/lib/agent-config/core/types';
 import type * as CodeReviewsDb from '@/lib/code-reviews/db/code-reviews';
+import type * as ModelSelection from '@/lib/code-reviews/core/model-selection';
 
 jest.mock('@/lib/integrations/platforms/github/adapter', () => ({
   generateGitHubInstallationToken: (...args: unknown[]) =>
@@ -62,6 +64,16 @@ jest.mock('@sentry/nextjs', () => ({
   captureException: jest.fn(),
 }));
 
+jest.mock('@/lib/code-reviews/core/model-selection', () => {
+  const actual = jest.requireActual<typeof ModelSelection>(
+    '@/lib/code-reviews/core/model-selection'
+  );
+  return {
+    ...actual,
+    resolveReviewSmallModel: (...args: unknown[]) => mockResolveReviewSmallModel(...args),
+  };
+});
+
 import { db } from '@/lib/drizzle';
 import { insertTestUser } from '@/tests/helpers/user.helper';
 import {
@@ -249,6 +261,8 @@ describe('prepareReviewPayload', () => {
     mockUpdatePreviousReviewSummary.mockReset();
     mockUpdateRepositoryReviewInstructionsMetadata.mockReset();
     mockGenerateReviewPrompt.mockReset();
+    mockResolveReviewSmallModel.mockReset();
+    mockResolveReviewSmallModel.mockResolvedValue(undefined);
   });
 
   afterAll(async () => {
@@ -979,4 +993,39 @@ describe('prepareReviewPayload', () => {
       upstreamBranch: 'refs/pull/1235/head',
     });
   });
+
+  it('includes resolved smallModel on GitHub session input when present', async () => {
+    const [review] = await db
+      .insert(cloud_agent_code_reviews)
+      .values(defineReview(testUser.id, integration.id))
+      .returning();
+    mockResolveReviewSmallModel.mockResolvedValueOnce('anthropic/claude-haiku-4.5');
+
+    const payload = await prepareReviewPayload({
+      reviewId: review.id,
+      owner: { type: 'user', id: testUser.id, userId: testUser.id },
+      agentConfig: { config: baseAgentConfig },
+      platform: 'github',
+    });
+
+    expect(mockResolveReviewSmallModel).toHaveBeenCalledWith('test-model');
+    expect(payload.sessionInput.smallModel).toBe('anthropic/claude-haiku-4.5');
+  });
+
+  it('omits smallModel from session input when no cheap sibling is resolved', async () => {
+    const [review] = await db
+      .insert(cloud_agent_code_reviews)
+      .values(defineReview(testUser.id, integration.id))
+      .returning();
+
+    const payload = await prepareReviewPayload({
+      reviewId: review.id,
+      owner: { type: 'user', id: testUser.id, userId: testUser.id },
+      agentConfig: { config: baseAgentConfig },
+      platform: 'github',
+    });
+
+    expect(mockResolveReviewSmallModel).toHaveBeenCalledWith('test-model');
+    expect(payload.sessionInput).not.toHaveProperty('smallModel');
+  });
 });

apps/web/src/lib/code-reviews/triggers/prepare-review-payload.ts

@@ -60,6 +60,7 @@ import {
   type ReviewScope,
 } from '../db/code-reviews';
 import { DEFAULT_CODE_REVIEW_MODEL, DEFAULT_CODE_REVIEW_MODE } from '../core/constants';
+import { resolveReviewSmallModel } from '../core/model-selection';
 import type { Owner } from '../core';
 import { generateReviewPrompt } from '../prompts/generate-prompt';
 import type { CodeReviewAgentConfig } from '@/lib/agent-config/core/types';
@@ -113,6 +114,11 @@ export type SessionInput = {
   prompt: string;
   mode: 'code';
   model: string;
+  /**
+   * Optional cheap same-vendor model for kilo title/aux calls. When omitted, the
+   * CLI falls through to its default small-model list (kilo-auto/small → Gemma).
+   */
+  smallModel?: string;
   /** Thinking effort variant name (e.g. "high", "max") — undefined means model default */
   variant?: string;
   upstreamBranch: string;
@@ -302,12 +308,14 @@ export async function prepareReviewPayload(
         // Single source for the standard reviewer's model so the session input and the
         // forward-shaped `reviewAgents[0]` can never drift apart.
         const standardModel = config.model_slug || DEFAULT_CODE_REVIEW_MODEL;
+        const smallModel = await resolveReviewSmallModel(standardModel);
         const sessionInput: SessionInput = {
           gitUrl: `https://bitbucket.org/${workspaceSlug.data}/${repositorySlug.data}.git`,
           kilocodeOrganizationId: owner.id,
           prompt,
           mode: DEFAULT_CODE_REVIEW_MODE as 'code',
           model: standardModel,
+          ...(smallModel ? { smallModel } : {}),
           variant: config.thinking_effort ?? undefined,
           upstreamBranch: review.head_ref,
           platform: PLATFORM.BITBUCKET,
@@ -758,6 +766,7 @@ export async function prepareReviewPayload(
     // Single source for the standard reviewer's model so the session input and the
     // forward-shaped `reviewAgents[0]` can never drift apart.
     const standardModel = config.model_slug || DEFAULT_CODE_REVIEW_MODEL;
+    const smallModel = await resolveReviewSmallModel(standardModel);
     const gateThreshold = config.gate_threshold ?? 'off';
     const githubCheckoutRef = getGitHubPullRequestCheckoutRef(review.pr_number);
     const sessionInput: SessionInput =
@@ -771,6 +780,7 @@ export async function prepareReviewPayload(
             prompt,
             mode: DEFAULT_CODE_REVIEW_MODE as 'code',
             model: standardModel,
+            ...(smallModel ? { smallModel } : {}),
             variant,
             upstreamBranch: review.head_ref,
           }
@@ -784,6 +794,7 @@ export async function prepareReviewPayload(
               prompt,
               mode: DEFAULT_CODE_REVIEW_MODE as 'code',
               model: standardModel,
+              ...(smallModel ? { smallModel } : {}),
               variant,
               upstreamBranch: review.head_ref,
               ...(gateThreshold !== 'off' ? { gateThreshold } : {}),
@@ -797,6 +808,7 @@ export async function prepareReviewPayload(
               prompt,
               mode: DEFAULT_CODE_REVIEW_MODE as 'code',
               model: standardModel,
+              ...(smallModel ? { smallModel } : {}),
               variant,
               upstreamBranch: githubCheckoutRef,
               ...(gateThreshold !== 'off' ? { gateThreshold } : {}),

packages/worker-utils/src/cloud-agent-next-client.ts

@@ -39,6 +39,8 @@ export type CloudAgentPrepareSessionInput = {
   prompt: string;
   mode: string;
   model: string;
+  /** Optional cheap same-vendor model for title/aux calls (Code Reviewer). */
+  smallModel?: string;
   variant?: string;
   githubRepo?: string;
   githubToken?: string;

services/cloud-agent-next/src/execution/types.ts

@@ -85,6 +85,8 @@ export function renderExecutionTurnContent(turn: AcceptedExecutionTurn): string
 export type ModelChoice = {
   model: string;
   variant?: string;
+  /** Optional cheap same-vendor model for title/aux calls (Code Reviewer). */
+  smallModel?: string;
 };
 
 /** Fully resolved agent selection. */

services/cloud-agent-next/src/persistence/session-metadata.ts

@@ -167,6 +167,7 @@ const MetadataAgentSchema = z
   .object({
     mode: z.string().optional(),
     model: z.string().optional(),
+    smallModel: z.string().optional(),
     variant: z
       .string()
       .max(50)

services/cloud-agent-next/src/router/handlers/session-prepare.ts

@@ -262,6 +262,7 @@ export function prepareInputToSessionCreateRequest(input: PrepareInput): Session
     agent: {
       mode: input.mode,
       model: input.model,
+      ...(input.smallModel ? { smallModel: input.smallModel } : {}),
       variant: input.variant,
     },
     repository,

services/cloud-agent-next/src/router/schemas.ts

@@ -383,6 +383,11 @@ export const PrepareSessionInput = z
       'Kilo Code execution mode (built-in or custom slug from runtimeAgents)'
     ),
     model: modelIdSchema.describe('AI model to use'),
+    smallModel: modelIdSchema
+      .optional()
+      .describe(
+        'Optional cheap same-vendor model for title/aux calls (Code Reviewer). When omitted, CLI defaults apply.'
+      ),
     variant: z
       .string()
       .max(50)