Codex 0.142.4

如果你公司網路要求所有連線都要先過一道代理伺服器（proxy，幫你的網路請求轉手繼承送出去的中介站，常見於公司防火牆環境）才能出去，Windows 上的 Codex 登入流程以前不認得這個設定，現在可以選擇性開啟，支援 PAC（自動設定腳本）、WPAD（自動偵測代理位置）、固定代理位址，以及「哪些網址不要走代理」的例外規則。這就像你家信箱以前規定信一定要先送到社區警衛室才能轉交，現在系統終於認得這條規定，不會把信卡在半路。

當你接了很多 MCP（Model Context Protocol，讓 AI 連接外部工具的協定）外掛伺服器時，AI 以前得把所有工具的說明書整本塞進腦袋裡才知道有什麼可以用。現在如果環境支援，AI 改成用「搜尋」（tool search，AI 用查詢的方式臨時找出需要的工具，而不是整本背起來）的方式找工具，找得更準，舊型號的模型跟舊版外掛也都還能正常運作。這就像以前你進圖書館得把整層樓的書名都背下來才知道哪本有用，現在改成直接用搜尋系統查「我要找什麼」，馬上指給你看在哪一格。

外掛（plugin，幫 Codex 擴充額外功能的附加套件）的開發者現在可以額外準備一張「深色模式」（dark mode，背景偏黑、適合夜間使用的顯示模式）專用的圖示，透過外掛自己的設定檔或遠端外掛商店提供，當你切到深色介面時就會自動換成這張，不會再出現亮色圖示在暗背景上突兀刺眼的狀況。

你如果丟一個網路上的圖片連結（HTTP/HTTPS image，存在遠端伺服器上、用網址連過去才看得到的圖片）給 AI 看，以前格式或來源有問題時可能整個流程默默卡住或失敗，原因說不清楚。現在系統會產生清楚的驗證錯誤訊息，連 AI 自己都讀得懂「為什麼這張圖不能用」，而直接貼上的圖片資料（data URL）或本機檔案的圖片不受影響、照樣能用。

安全分類器（safety classifier，自動判斷一條指令安不安全的系統，決定要不要事先問你同意）原本碰到 PowerShell（Windows 的指令稿語言）指令裡有些它解析不出來、但確實可以被執行的程式片段（AST，abstract syntax tree，把程式碼拆解成結構樹方便電腦理解的格式）時，可能會誤判成安全、直接放行。現在只要分類器看不懂、判斷不了，一律先攔下來要你親自核准，不會因為「看不懂所以放行」而留下安全漏洞。像海關以前看不懂的可疑物品有時會放過，現在改成看不懂就一律先攔下來開箱檢查。

Code Mode（讓 AI 直接用寫程式的方式呼叫工具、而非一個個指令慢慢喊的模式）需要模型提供特定的中繼資料（metadata，描述這個模型支援哪些功能的附加資訊）才能正常運作。以前選到不支援的模型，可能就默默出問題讓你一頭霧水；現在系統會主動跳出警告，讓你在出狀況前就知道「這個模型可能不適合這樣用」。

If your company network routes everything through a proxy (proxy, a middleman server your network requests get handed off through, common on corporate firewalls) before it can reach the internet, Codex's Windows login flow used to not understand that setup. Now you can opt in to support for PAC (auto-config scripts), WPAD (auto-discovering the proxy), static proxy addresses, and bypass rules for which destinations skip the proxy. It's like a mailbox rule that always required mail to go through the building's security desk first — the system finally understands that rule, so your mail doesn't get stuck halfway.

When you've hooked up a lot of MCP servers (Model Context Protocol, the standard that lets the AI talk to external tools), the AI used to have to memorize every tool's full instruction manual just to know what was available. Now, where supported, the AI uses tool search instead (tool search, looking up the specific tool it needs on demand rather than memorizing the whole catalog), finding tools more accurately, while older models and plugins keep working fine. It's like having to memorize every book title on a whole library floor just to find one useful book — now you just search for what you want and get pointed straight to the shelf.

Plugin (plugin, an add-on package that extends Codex with extra features) developers can now ship a dedicated dark-mode (dark mode, the dark-background display mode meant for low-light use) icon, supplied through the plugin's own manifest or a remote plugin store, which swaps in automatically when you switch to a dark interface — no more glaring light-colored icons sitting on a dark background.

If you point the AI at a remote image link (HTTP/HTTPS image, a picture that lives on a server elsewhere and only loads via its URL), problems with the format or source used to silently stall or fail the whole flow with no clear reason. Now the system produces a clear validation error the AI itself can read and understand — 'here's why this image didn't work' — while inline pasted image data (data URLs) and local files are unaffected and keep working.

The safety classifier (safety classifier, the system that automatically judges whether a command is safe and decides whether to ask your permission first) used to sometimes wave through PowerShell (Windows's scripting language) commands containing executable code regions (AST, abstract syntax tree, a structured breakdown of code that computers can parse) it couldn't actually parse — treating 'can't read it' as 'must be safe.' Now anything the classifier can't read gets stopped for your manual approval instead, closing that 'unreadable so let it through' loophole. It's like customs that used to sometimes wave through suspicious items it couldn't identify; now anything it can't identify gets pulled aside and opened up first.

Code Mode (a mode where the AI calls tools by writing code directly instead of issuing commands one at a time) needs the model to supply specific metadata (metadata, extra information describing which features a model supports) to work properly. Picking an unsupported model used to just fail silently and leave you confused. Now the system proactively warns you up front: 'this model might not work well this way.'