feat(core): Cap shell output truncation threshold to the remaining context window size (#8379)

packages/core/src/config/config.test.ts

@@ -94,6 +94,9 @@ vi.mock('../telemetry/index.js', async (importOriginal) => {
   return {
     ...actual,
     initializeTelemetry: vi.fn(),
+    uiTelemetryService: {
+      getLastPromptTokenCount: vi.fn(),
+    },
   };
 });
 
@@ -123,8 +126,13 @@ vi.mock('../ide/ide-client.js', () => ({
 }));
 
 import { BaseLlmClient } from '../core/baseLlmClient.js';
+import { tokenLimit } from '../core/tokenLimits.js';
+import { uiTelemetryService } from '../telemetry/index.js';
 
 vi.mock('../core/baseLlmClient.js');
+vi.mock('../core/tokenLimits.js', () => ({
+  tokenLimit: vi.fn(),
+}));
 
 describe('Server Config (config.ts)', () => {
   const MODEL = 'gemini-pro';
@@ -661,6 +669,57 @@ describe('Server Config (config.ts)', () => {
       });
     });
   });
+
+  describe('getTruncateToolOutputThreshold', () => {
+    beforeEach(() => {
+      vi.clearAllMocks();
+    });
+
+    it('should return the calculated threshold when it is smaller than the default', () => {
+      const config = new Config(baseParams);
+      vi.mocked(tokenLimit).mockReturnValue(32000);
+      vi.mocked(uiTelemetryService.getLastPromptTokenCount).mockReturnValue(
+        1000,
+      );
+      // 4 * (32000 - 1000) = 4 * 31000 = 124000
+      // default is 4_000_000
+      expect(config.getTruncateToolOutputThreshold()).toBe(124000);
+    });
+
+    it('should return the default threshold when the calculated value is larger', () => {
+      const config = new Config(baseParams);
+      vi.mocked(tokenLimit).mockReturnValue(2_000_000);
+      vi.mocked(uiTelemetryService.getLastPromptTokenCount).mockReturnValue(
+        500_000,
+      );
+      // 4 * (2_000_000 - 500_000) = 4 * 1_500_000 = 6_000_000
+      // default is 4_000_000
+      expect(config.getTruncateToolOutputThreshold()).toBe(4_000_000);
+    });
+
+    it('should use a custom truncateToolOutputThreshold if provided', () => {
+      const customParams = {
+        ...baseParams,
+        truncateToolOutputThreshold: 50000,
+      };
+      const config = new Config(customParams);
+      vi.mocked(tokenLimit).mockReturnValue(8000);
+      vi.mocked(uiTelemetryService.getLastPromptTokenCount).mockReturnValue(
+        2000,
+      );
+      // 4 * (8000 - 2000) = 4 * 6000 = 24000
+      // custom threshold is 50000
+      expect(config.getTruncateToolOutputThreshold()).toBe(24000);
+
+      vi.mocked(tokenLimit).mockReturnValue(32000);
+      vi.mocked(uiTelemetryService.getLastPromptTokenCount).mockReturnValue(
+        1000,
+      );
+      // 4 * (32000 - 1000) = 124000
+      // custom threshold is 50000
+      expect(config.getTruncateToolOutputThreshold()).toBe(50000);
+    });
+  });
 });
 
 describe('setApprovalMode with folder trust', () => {

packages/core/src/config/config.ts

@@ -39,7 +39,9 @@ import {
   initializeTelemetry,
   DEFAULT_TELEMETRY_TARGET,
   DEFAULT_OTLP_ENDPOINT,
+  uiTelemetryService,
 } from '../telemetry/index.js';
+import { tokenLimit } from '../core/tokenLimits.js';
 import { StartSessionEvent } from '../telemetry/index.js';
 import {
   DEFAULT_GEMINI_EMBEDDING_MODEL,
@@ -916,7 +918,12 @@ export class Config {
   }
 
   getTruncateToolOutputThreshold(): number {
-    return this.truncateToolOutputThreshold;
+    return Math.min(
+      // Estimate remaining context window in characters (1 token ~= 4 chars).
+      4 *
+        (tokenLimit(this.model) - uiTelemetryService.getLastPromptTokenCount()),
+      this.truncateToolOutputThreshold,
+    );
   }
 
   getTruncateToolOutputLines(): number {