Files
gemini-cli/packages/core/src/context/system-tests/SimulationHarness.ts
T
Your Name 63e8b825a7 burndown
2026-04-07 03:24:40 +00:00

163 lines
6.1 KiB
TypeScript

/**
* @license
* Copyright 2026 Google LLC
* SPDX-License-Identifier: Apache-2.0
*/
import type { BaseLlmClient } from "../../core/baseLlmClient.js";
/**
* @license
* Copyright 2026 Google LLC
* SPDX-License-Identifier: Apache-2.0
*/
import { ContextManager } from '../contextManager.js';
import { AgentChatHistory } from '../../core/agentChatHistory.js';
import type { Content } from '@google/genai';
import type { SidecarConfig } from '../sidecar/types.js';
import { ContextEnvironmentImpl } from '../sidecar/environmentImpl.js';
import { ContextTracer } from '../tracer.js';
import { ContextEventBus } from '../eventBus.js';
import { PipelineOrchestrator } from '../sidecar/orchestrator.js';
import { registerBuiltInProcessors } from '../sidecar/builtins.js';
import { debugLogger } from "../../utils/debugLogger.js";
export interface TurnSummary {
turnIndex: number;
tokensBeforeBackground: number;
tokensAfterBackground: number;
}
export class SimulationHarness {
readonly chatHistory: AgentChatHistory;
contextManager!: ContextManager;
env!: ContextEnvironmentImpl;
orchestrator!: PipelineOrchestrator;
readonly eventBus: ContextEventBus;
config!: SidecarConfig;
private tracer!: ContextTracer;
private currentTurnIndex = 0;
private tokenTrajectory: TurnSummary[] = [];
static async create(config: SidecarConfig, mockLlmClient: BaseLlmClient, mockTempDir = '/tmp/sim'): Promise<SimulationHarness> {
const harness = new SimulationHarness();
await harness.init(config, mockLlmClient, mockTempDir);
return harness;
}
private constructor() {
this.chatHistory = new AgentChatHistory();
this.eventBus = new ContextEventBus();
}
private async init(
config: SidecarConfig,
mockLlmClient: BaseLlmClient,
mockTempDir: string
) {
this.config = config;
// Register all standard processors
registerBuiltInProcessors();
this.tracer = new ContextTracer({ targetDir: mockTempDir, sessionId: 'sim-session' });
// Using real token calculator instead of mock, so we test actual string sizes
const InMemoryFS = (await import('../system/InMemoryFileSystem.js')).InMemoryFileSystem;
const DetIdGen = (await import('../system/DeterministicIdGenerator.js')).DeterministicIdGenerator;
this.env = new ContextEnvironmentImpl(
mockLlmClient,
'sim-prompt',
'sim-session',
mockTempDir,
mockTempDir,
this.tracer,
4, // 4 chars per token average
this.eventBus,
new InMemoryFS(),
new DetIdGen()
);
this.orchestrator = new PipelineOrchestrator(config, this.env, this.eventBus, this.tracer);
this.contextManager = ContextManager.create(config, this.env, this.tracer, this.orchestrator);
this.contextManager.subscribeToHistory(this.chatHistory);
}
/**
* Simulates a single "Turn" (User input + Model/Tool outputs)
* A turn might consist of multiple Content messages (e.g. user prompt -> model call -> user response -> model answer)
*/
async simulateTurn(messages: Content[]) {
// 1. Append the new messages
const currentHistory = this.chatHistory.get();
this.chatHistory.set([...currentHistory, ...messages]);
// 2. Measure tokens immediately after append (Before background processing)
const tokensBefore = this.env.tokenCalculator.calculateEpisodeListTokens(
this.contextManager.getWorkingBufferView()
);
debugLogger.log(`[Turn ${this.currentTurnIndex}] Tokens BEFORE: ${tokensBefore}`);
// 3. Yield to event loop to allow internal async subscribers and orchestrator to finish
await new Promise(resolve => setTimeout(resolve, 50));
// 3.1 Simulate what projectCompressedHistory does with the sync handlers
let currentView = this.contextManager.getWorkingBufferView();
const currentTokens = this.env.tokenCalculator.calculateEpisodeListTokens(currentView);
if (this.config.budget && currentTokens > this.config.budget.maxTokens) {
debugLogger.log(`[Turn ${this.currentTurnIndex}] Sync panic triggered! ${currentTokens} > ${this.config.budget.maxTokens}`);
const syncPipelines = this.config.pipelines.filter(p => p.execution === 'blocking');
const orchestrator = this.orchestrator;
for (const pipe of syncPipelines) {
await orchestrator.executePipeline(pipe.name, currentView, {
currentTokens,
maxTokens: this.config.budget.maxTokens,
retainedTokens: this.config.budget.retainedTokens,
isBudgetSatisfied: false,
deficitTokens: currentTokens - this.config.budget.maxTokens,
protectedEpisodeIds: new Set()
});
currentView = this.contextManager.getWorkingBufferView();
}
// Inject the truncated view back into the graph
for (let i = 0; i < currentView.length; i++) {
const ep = currentView[i];
if (!this.contextManager.getWorkingBufferView().find(c => c.id === ep.id)) {
this.eventBus.emitVariantReady({
targetId: ep.id,
variantId: 'v-emergency',
variant: {
status: 'ready',
type: 'masked', // Truncation is technically a mask
text: ep.yield?.text || '',
recoveredTokens: 0,
}
});
}
}
// Wait for variant propagation
await new Promise(resolve => setTimeout(resolve, 50));
}
// 4. Measure tokens after background processors have (hopefully) emitted variants
const tokensAfter = this.env.tokenCalculator.calculateEpisodeListTokens(
this.contextManager.getWorkingBufferView()
);
debugLogger.log(`[Turn ${this.currentTurnIndex}] Tokens AFTER: ${tokensAfter}`);
this.tokenTrajectory.push({
turnIndex: this.currentTurnIndex++,
tokensBeforeBackground: tokensBefore,
tokensAfterBackground: tokensAfter,
});
}
async getGoldenState() {
const finalProjection = await this.contextManager.projectCompressedHistory();
return {
tokenTrajectory: this.tokenTrajectory,
finalProjection
};
}
}