Skip to content

README.md

Latest commit

 

History

History
40 lines (25 loc) · 2.48 KB

README.md

File metadata and controls

40 lines (25 loc) · 2.48 KB

Reverse-Engineering the Claude Agent SDK: Root Cause and Fix for the 2–3% Credit Burn Per Message

逆向工程 Claude Agent SDK：每條訊息 2–3% 額度燃燒的根本原因與修復

English Version | 繁體中文版

What This Report Covers / 這篇報告在講什麼

This report investigates why multi-agent systems using the Claude Agent SDK consume 2–3% of the 5-hour credit quota per message, despite minimal message content. By reverse-engineering the SDK's 12MB minified source, the investigation reveals that the SDK spawns a new Node.js process on every call, which rebuilds the entire conversation history and completely misses the prompt cache. The report documents the root cause discovery process and the implemented solution.

這份報告調查為什麼使用 Claude Agent SDK 的多代理系統儘管訊息內容很少，卻消耗 5 小時額度配額的 2–3%。透過逆向工程 SDK 的 12MB 最小化程式碼，調查發現 SDK 在每次呼叫時都會生成新的 Node.js 程序，完整重建對話歷史記錄，完全錯過提示快取。報告記錄了根本原因的發現過程和實施的解決方案。

Who Should Read This / 適合誰讀

  • SDK users running multi-agent systems — understand the true cost of each API call and why it's higher than expected

  • 運行多代理系統的 SDK 使用者 — 了解每次 API 呼叫的真實成本及其為何高於預期

  • DevOps/Cost optimization engineers — learn why persistent sessions dramatically reduce per-message costs

  • DevOps/成本最佳化工程師 — 了解持久化會話如何將每訊息成本從 2–3% 降低到 1% 以下

  • Anthropic SDK maintainers — understand the performance implications of the current architecture

  • Anthropic SDK 維護人員 — 了解當前架構的效能影響

What You'll Learn / 你會學到什麼

  • Why the Agent SDK spawns a new process on every call

  • 為什麼 Agent SDK 在每次呼叫時都會生成新程序

  • How the cli.js engine rebuilds conversation history on each invocation

  • 如何在每次呼叫時重建 cli.js 引擎的對話歷史記錄

  • The impact of cache misses on overall credit consumption

  • 快取未命中對整體額度消耗的影響

  • How persistent sessions reduce cost from 2–3% to below 1% per message

  • 持久化會話如何將成本從 2–3% 降低到每訊息 1% 以下

  • The specific architectural changes needed for optimization

  • 優化所需的特定架構變更