Claude Code dynamically injects <system-reminder> content blocks into every API request that are invisible in the UI, not stored in conversation history, and actively hidden from the model. This report documents all 15+ injection categories, their triggers, and estimated token costs by reverse-engineering the cli.js source and analyzing session-level JSONL logs. It is a companion to Report #1, focusing on the injection mechanism itself rather than cost impact.
Claude Code 動態地將 <system-reminder> 內容區塊注入每個 API 請求，這些區塊在 UI 中不可見、未儲存在對話歷史記錄中，並且被主動隱藏在模型中。此報告透過逆向工程 cli.js 原始碼並分析會話級 JSONL 日誌，記錄所有 15 以上的注入類別、觸發器和預估代幣成本。它是報告 #1 的配套報告，專注於注入機制本身而不是成本影響。
-
Claude Code power users — understand what hidden instructions are in your requests and why
-
Claude Code 進階使用者 — 了解您的請求中有哪些隱藏指令以及原因
-
Cost analysts — learn what comprises the "hidden token tax" on every conversation turn
-
成本分析師 — 了解每個對話轉折上「隱藏代幣稅」包含什麼
-
Security researchers — understand how system prompts are managed and what risks that implies
-
安全研究人員 — 了解系統提示如何管理及其隱含的風險
-
The complete catalogue of 15+ injection types and their triggers
-
15 個以上注入類型及其觸發器的完整目錄
-
How system-reminder blocks are assembled and injected invisibly
-
系統提醒區塊如何組裝和隱形注入
-
Token cost breakdown per injection type
-
每個注入類型的代幣成本明細
-
Which operations (like Read calls) carry the highest hidden overhead
-
哪些操作（如 Read 呼叫）具有最高隱藏開銷
-
Why some injections fire every turn while others are conditional
-
為什麼某些注入在每個轉折上触发，而其他的是條件性的