feat(context): implement observation masking for tool outputs (#18389)

packages/core/src/core/client.test.ts

@@ -213,6 +213,7 @@ describe('Gemini Client (client.ts)', () => {
       getGlobalMemory: vi.fn().mockReturnValue(''),
       getEnvironmentMemory: vi.fn().mockReturnValue(''),
       isJitContextEnabled: vi.fn().mockReturnValue(false),
+      getToolOutputMaskingEnabled: vi.fn().mockReturnValue(false),
       getDisableLoopDetection: vi.fn().mockReturnValue(false),
 
       getSessionId: vi.fn().mockReturnValue('test-session-id'),

packages/core/src/core/client.ts

@@ -54,6 +54,7 @@ import { handleFallback } from '../fallback/handler.js';
 import type { RoutingContext } from '../routing/routingStrategy.js';
 import { debugLogger } from '../utils/debugLogger.js';
 import type { ModelConfigKey } from '../services/modelConfigService.js';
+import { ToolOutputMaskingService } from '../services/toolOutputMaskingService.js';
 import { calculateRequestTokenCount } from '../utils/tokenCalculation.js';
 import {
   applyModelSelection,
@@ -84,6 +85,7 @@ export class GeminiClient {
 
   private readonly loopDetector: LoopDetectionService;
   private readonly compressionService: ChatCompressionService;
+  private readonly toolOutputMaskingService: ToolOutputMaskingService;
   private lastPromptId: string;
   private currentSequenceModel: string | null = null;
   private lastSentIdeContext: IdeContext | undefined;
@@ -98,6 +100,7 @@ export class GeminiClient {
   constructor(private readonly config: Config) {
     this.loopDetector = new LoopDetectionService(config);
     this.compressionService = new ChatCompressionService();
+    this.toolOutputMaskingService = new ToolOutputMaskingService();
     this.lastPromptId = this.config.getSessionId();
 
     coreEvents.on(CoreEvent.ModelChanged, this.handleModelChanged);
@@ -562,6 +565,8 @@ export class GeminiClient {
     const remainingTokenCount =
       tokenLimit(modelForLimitCheck) - this.getChat().getLastPromptTokenCount();
 
+    await this.tryMaskToolOutputs(this.getHistory());
+
     // Estimate tokens. For text-only requests, we estimate based on character length.
     // For requests with non-text parts (like images, tools), we use the countTokens API.
     const estimatedRequestTokenCount = await calculateRequestTokenCount(
@@ -1056,4 +1061,20 @@ export class GeminiClient {
 
     return info;
   }
+
+  /**
+   * Masks bulky tool outputs to save context window space.
+   */
+  private async tryMaskToolOutputs(history: Content[]): Promise<void> {
+    if (!this.config.getToolOutputMaskingEnabled()) {
+      return;
+    }
+    const result = await this.toolOutputMaskingService.mask(
+      history,
+      this.config,
+    );
+    if (result.maskedCount > 0) {
+      this.getChat().setHistory(result.newHistory);
+    }
+  }
 }