Google Gemini CLI — Fixing Stream Cancellation & Memory Leaks (PR #20778)

1. The Problem: Detached In-Flight LLM Invocations

In multi-turn terminal chat sessions with Gemini, pressing Ctrl+C or issuing an abort command cancelled the active UI stream, but failed to cancel the background chat history compression service.

Because the compression routine was invoked on a detached promise chain without receiving the active turn's AbortSignal, background network requests continued executing against Google Cloud APIs — causing billing spikes, memory leaks, and race conditions on subsequent prompt dispatches.

2. The Solution: End-to-End AbortSignal Threading

I refactored the execution pipeline across LocalAgentExecutor, GeminiClient, and the ChatCompressionService:

// Refactored async execution boundary with AbortSignal threading
async executeTurn(prompt: string, options: { signal?: AbortSignal }) {
  const { signal } = options;
  
  // Guaranteed signal propagation through compression middleware
  if (this.shouldCompressHistory()) {
    await this.compressionService.compress({
      history: this.chatHistory,
      signal, // Propagated abort controller
    });
  }

  return this.client.streamGenerateContent(prompt, { signal });
}

• Graceful Cleanup: Guaranteed that all HTTP sockets and background workers terminate immediately upon user cancellation.
• Deterministic State: Eliminated out-of-order history compression race conditions when rapid consecutive prompts are entered.

3. Code Review & Upstream Merge

The pull request underwent review by Google engineers, verified with unit test suites and integration tests across Linux, macOS, and Windows runtime environments, and was merged into main.