feat(core): Support auto-distillation for tool output.

evals/app-test-helper.ts

@@ -13,7 +13,12 @@ import {
 } from './test-helper.js';
 import fs from 'node:fs';
 import path from 'node:path';
-import { DEFAULT_GEMINI_MODEL } from '@google/gemini-cli-core';
+import {
+  DEFAULT_GEMINI_MODEL,
+  type ScriptItem,
+  extractFakeResponses,
+  extractUserPrompts,
+} from '@google/gemini-cli-core';
 
 export interface AppEvalCase {
   name: string;
@@ -23,6 +28,12 @@ export interface AppEvalCase {
   files?: Record<string, string>;
   setup?: (rig: AppRig) => Promise<void>;
   assert: (rig: AppRig, output: string) => Promise<void>;
+  /**
+   * Optional script to "prime the pump" before the main prompt.
+   * A sequential array interleaving MockUserTurn (e.g., userText('hello'))
+   * and FakeResponse (e.g., mockGenerateContentStreamText('hi')).
+   */
+  script?: ScriptItem[];
 }
 
 /**
@@ -31,11 +42,23 @@ export interface AppEvalCase {
  */
 export function appEvalTest(policy: EvalPolicy, evalCase: AppEvalCase) {
   const fn = async () => {
+    const configOverrides = {
+      model: DEFAULT_GEMINI_MODEL,
+      ...evalCase.configOverrides,
+    };
+
+    let userPrompts: string[] = [];
+
+    if (evalCase.script) {
+      // Extract fake model responses to inject into the ContentGenerator
+      configOverrides.fakeResponses = extractFakeResponses(evalCase.script);
+
+      // Extract the sequence of user prompts for the Mock User driver
+      userPrompts = extractUserPrompts(evalCase.script);
+    }
+
     const rig = new AppRig({
-      configOverrides: {
-        model: DEFAULT_GEMINI_MODEL,
-        ...evalCase.configOverrides,
-      },
+      configOverrides,
     });
 
     const { logDir, sanitizedName } = await prepareLogDir(evalCase.name);
@@ -67,6 +90,11 @@ export function appEvalTest(policy: EvalPolicy, evalCase: AppEvalCase) {
       // Wait for initial ready state
       await rig.waitForIdle();
 
+      // Execute priming script if requested
+      if (userPrompts.length > 0) {
+        await rig.driveMockUser(userPrompts, evalCase.timeout);
+      }
+
       // Send the initial prompt
       await rig.sendMessage(evalCase.prompt);
 

evals/auto_distillation.eval.ts

@@ -0,0 +1,82 @@
+/**
+ * @license
+ * Copyright 2026 Google LLC
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+import { describe, expect } from 'vitest';
+import path from 'node:path';
+import fs from 'node:fs';
+import { appEvalTest } from './app-test-helper.js';
+import {
+  userText,
+  mockGenerateContentStreamText,
+} from '@google/gemini-cli-core';
+
+describe('Auto-Distillation Behavioral Evals', () => {
+  appEvalTest('USUALLY_PASSES', {
+    name: 'Agent successfully navigates truncated output using the structural map to extract a secret',
+    timeout: 180000,
+    configOverrides: {},
+    setup: async (rig) => {
+      const testDir = rig.getTestDir();
+
+      const mockData: any = {
+        system_info: {
+          version: '1.0.0',
+          uptime: 999999,
+          environment: 'production',
+        },
+        active_sessions: [],
+        quarantined_payloads: [
+          { id: 'Subject-01', status: 'cleared' },
+          {
+            id: 'Subject-89',
+            secret_token: 'the_cake_is_a_lie',
+            status: 'held_for_review',
+          },
+          { id: 'Subject-99', status: 'cleared' },
+        ],
+        archived_metrics: [],
+      };
+
+      for (let i = 0; i < 300; i++) {
+        mockData.active_sessions.push({
+          session_id: `sess_${i.toString().padStart(4, '0')}`,
+          duration_ms: Math.floor(Math.random() * 10000),
+          bytes_transferred: Math.floor(Math.random() * 50000),
+        });
+      }
+
+      for (let i = 0; i < 2000; i++) {
+        mockData.archived_metrics.push({
+          timestamp: Date.now() - i * 1000,
+          cpu_load: parseFloat(Math.random().toFixed(4)),
+          mem_usage: parseFloat(Math.random().toFixed(4)),
+        });
+      }
+
+      const massiveString = JSON.stringify(mockData, null, 2);
+
+      fs.writeFileSync(
+        path.join(testDir, 'server_state_dump.json'),
+        massiveString,
+      );
+    },
+    script: [
+      userText('We have a critical error in production. Are you ready to help?'),
+      mockGenerateContentStreamText(
+        'I am ready. Please provide the details of the error.',
+      ),
+    ],
+    prompt: `My application crashed with: "FATAL: Subject-89 held for review in quarantine". \n\nPlease run \`cat server_state_dump.json\` to investigate. The file is massive, so your tool output will be automatically truncated and you will receive a structural map instead. Use that structural map to determine the right command to extract the \`secret_token\` for Subject-89. Please state the exact secret token when you find it.`,
+    assert: async (rig) => {
+      await rig.waitForIdle(120000);
+
+      const finalOutput = rig.getStaticOutput();
+
+      // Ensure the agent correctly extracted the secret token after navigating the distilled output
+      expect(finalOutput).toContain('the_cake_is_a_lie');
+    },
+  });
+});

evals/hybrid_handoff.eval.ts

@@ -0,0 +1,39 @@
+import { describe, expect } from 'vitest';
+import {
+  appEvalTest,
+} from './app-test-helper.js';
+import {
+  userText,
+  mockGenerateContentStreamText,
+} from '@google/gemini-cli-core';
+
+describe('Hybrid Handoff (Mock User to Live Model)', () => {
+  appEvalTest('ALWAYS_PASSES', {
+    name: 'Mock User successfully primes AppRig using a scripted history and hands off to live model',
+    timeout: 120000,
+    script: [
+      userText('Start priming'),
+      mockGenerateContentStreamText(
+        "Hello! I am a fake response. Let's prime the pump.",
+      ),
+      userText('Continue priming'),
+      mockGenerateContentStreamText(
+        'Pump primed successfully. Ready for handoff.',
+      ),
+    ],
+    prompt: 'What is 2 + 2? Please answer with exactly the number "4".',
+    assert: async (rig) => {
+      // The Mock User has automatically driven the script before sending the final prompt.
+      // So the history now has the 2 fake turns in it, and the final prompt was just sent to the LIVE model.
+
+      await rig.waitForIdle(60000);
+
+      const liveOutput = rig.getStaticOutput();
+
+      // Ensure the handoff was successful
+      expect(liveOutput).toContain('4');
+
+      await rig.drainBreakpointsUntilIdle(undefined, 10000);
+    },
+  });
+});