發布日期：2026-07-01 ~ 07-08
為了避免 AI 運作時在背景瘋狂寫入日誌縮短你的硬碟壽命，系統移除了過於頻繁的資料庫記錄。以前系統在處理資料庫日誌時（TRACE，最詳細、連每一秒傳遞了什麼細微資料都記錄下來的偵錯等級），會把巨量的繁雜日誌事件反覆寫入本機 ~/.codex/logs_2.sqlite 資料庫，產生嚴重的寫入放大（write amplification，寫入硬碟的資料量遠大於實際檔案大小的現象），一年下來可能寫入高達 640 TB 的巨量資料，嚴重威脅 SSD（固態硬碟，電腦儲存資料的硬碟）的壽命；現在把這些囉唆的通訊過程從資料庫裡拿掉，停止了這類冗長事件的橋接，以降低寫入放大與 SSD 耗損。這就像以前助理把跟你講的每一字、呼吸的每一口氣都拿筆刻在木板上，一小時就堆滿一整個倉庫；現在改成只記重要結論，省下了大片森林。
為了保護你的程式碼私密性與個人敏感資料不外洩，偵錯日誌現在不會再記錄原始連線內容。以前系統在進行偵錯時，會把對話中傳輸的 WebSocket payloads（透過網路即時傳輸資料時承載的原始內容封包，包括你輸入的程式碼和對話）一字不漏地寫入本機明文追蹤日誌（trace logs，追蹤日誌，記錄程式執行詳細過程的文字檔）中，有安全隱憂；現在已加強安全防護、將這些原始資訊自動遮蔽過濾。這就像快遞公司的包裹追蹤系統以前會把箱子裡面的所有財物明細貼在箱子外面，現在改成只顯示包裹單號，保護你的隱私不被路人看光。
為了讓你能立即使用最新的強大人工智慧，系統已經將最新版模型納入支援清單。當你透過 Amazon Bedrock（亞馬遜雲端託管各大品牌 AI 模型的服務平台）串接模型時，現在可以直接選用最新的 GPT-5.6 變體模型（包括 Sol、Terra、Luna 等不同大小規格的版本），免去你自己手動寫設定來新增模型名稱的麻煩。這就像是你在串流平台上，系統自動幫你把最新上映的電影分類列在首頁推薦，你點一下就能直接播放，不用自己大費周章去搜尋片源。
為了提高大任務自動拆解與多工協作的成功率，系統最佳化了多個代理人之間的派發邏輯。現在系統在進行 multi-agent delegation（一個主 AI 為了完成工作，分頭派工作給多個小助手 AI 合作的機制）時，會套用重新整理過、最穩健的第一版引導方針，並清理了可能阻礙子代理人協作的規則，防止助手們互相干擾卡死。這就像是辦公室主管在給實習生分配任務時，重新整理了「分工守則」，講清楚誰做什麼、怎麼回報，讓小助手們不會因為權限打架而呆立在現場、不知道該聽誰的。
System stopped saving noisy connection events. Previously, the database logging (TRACE, the most verbose debugging level that records every tiny bit of activity second-by-second), continuously saving verbose log events to ~/.codex/logs_2.sqlite. This created massive write amplification (write amplification, a phenomenon where the actual volume of data written to physical storage is far larger than the logical file size), racking up to 640 TB of disk writes a year and risking premature solid-state drive (SSD, the high-speed storage drive in your computer) failure. These verbose bridge events are now ignored, stopping the bridging of these verbose log events to reduce write amplification and SSD wear. It's like an assistant writing down every single word you speak and breath you take by carving them into wooden blocks, filling a whole warehouse in an hour; now he only logs key decisions, saving a forest.
Debug logs no longer save raw connection messages. Previously, the app recorded full WebSocket payloads (WebSocket payloads, the raw data packets transmitted over the network containing your code and messages) directly inside local plain-text trace logs (trace logs, detailed text files that record the application's step-by-step runtime operations) during debug sessions; now these are automatically stripped. It's like a courier company's package tracking system that used to print the exact contents of your box on the shipping label; now it only shows a tracking number to keep your items private.
Model selector now supports the latest GPT-5.6 variants. If you access models via Amazon Bedrock (Amazon Bedrock, Amazon's cloud hosting platform for top-tier AI models), you can now choose GPT-5.6 (including the Sol, Terra, and Luna sizes) directly from the list, without having to manually edit config files to map them. It's like a video streaming app automatically listing the newest blockbuster movie directly on your home screen, ready to play with one click, instead of making you hunt down the raw file path yourself.
Delegation manager has streamlined its dispatch rules. During multi-agent delegation (multi-agent delegation, a workflow where one main AI assigns and divides work among several assistant AIs working together), the app now enforces the robust v1 layout instructions and sweeps away redundant guardrail blocks, keeping helpers moving without gridlock. It's like an office manager updating the team handbook to clarify exactly who does what and how to report back, so assistants don't end up standing around frozen because of conflicting instructions.
Fixed critical SQLite TRACE-level database logging causing severe write amplification and high SSD wear (up to 640TB/year) by stopping bridging verbose log events.Security: Prevented full 'Responses' WebSocket request payloads from being written to trace logs.Added support for GPT-5.6 variants (Sol, Terra, Luna) to the static Amazon Bedrock catalog.Restored v1 delegation guidance and cleaned up multi-agent delegation guardrail configurations.