發布日期：2026-06-22
現在輸入 /usage 不只能看用量，還能直接把辛苦賺來的「次數重置抵用券」換掉；系統也新增了「代幣預算（token budgets）」，能幫你跨代理人追蹤花費，快沒錢時會提醒你，一旦燒光就會自動踩煞車。在控制代理人上，你現在可以設定是要把它們通通關掉、叫它們才動，還是讓它們主動幫忙。最有趣的是，系統現在有時間觀念了，能接收排程定時提醒、主動查詢現在幾點。就像你請了一個帶有碼表的會計師，他隨時能幫你把折價券兌現，幫你盯著所有員工的開銷不讓預算超支，時間一到還會主動提醒你該開會了。
找工具跟找資料不再像大海撈針。現在輸入 /plugins，遠端外掛會整齊地分類成「官方精選」、「工作區專用」和「與我共用」，而且系統會在對話中主動推薦你當下最適合的外掛。安裝外掛時也變得更強大，就算碰到奇怪的商城結構或是下載錯誤，系統都會優雅地幫你處理好，並在安裝完立刻把新工具拿出來用。另外，我們還加了一個「索引式網頁搜尋」模式，它允許 AI 幫你即時搜尋網路，但會嚴格限制它只能連上伺服器批准過的安全網站。就像你去逛一間重新裝潢好的五金行，工具都分門別類擺好，店員會看你正在修水管而主動遞上板手；如果你要查水管的修理手冊，它會幫你上網查，但絕不會不小心點進奇怪的詐騙網站。
讓遠端開發變得跟在本地端一樣順手又安全。現在，系統跑在遠端時能完美保留那個作業系統原汁原味的路徑格式、命令列外殼（shell）以及安全沙盒機制，甚至會自動在那裡找出 AGENTS.md 檔案來讀。更棒的是，負責執行的後台引擎跟工具連線（MCP）現在具備了「九命怪貓」的體質：就算網路突然閃斷，它們也能熬過短暫的失聯，靠著重新刷新網址或安全地重試輸入來原地復活。就像你用遠端遙控機器人去別的星球出任務，它不只完美適應了當地的重力跟大氣層，就算通訊中斷了幾秒，它也不會直接當機，而是會乖乖在原地等你重新連上線，然後若無其事地繼續工作。
現在父代理人（主管 AI）不會再被騙了，如果子代理人（部屬 AI）做到一半當機或嚴重出錯，主管會確實收到錯誤報告，而不是看到一個「成功完成但裡面是空的」假報告。為了讓你感覺更順暢，系統減少了開機時不必要的網路查詢、先把模型熱機好、平行讀取資料，也砍掉了會狂塞硬碟的重複連線紀錄；如果你在 Linux 上用 Ctrl+Z 暫停畫面再用 fg 回來，畫面也不會壞掉。另外，原本失蹤的「目標優先（Goal-first）」對話，現在也終於能乖乖存檔並在歷史清單裡找到了。就像公司抓出了總是交白卷卻說自己做完的薪水小偷，順便優化了整個開機流程，讓每天早上的打卡跟開會時間少了一半，而且昨天暫停的企劃案今天打開依然完好如初。
The /usage command now doesn't just show stats, but lets you actively redeem your hard-earned usage-limit reset credits; we've also introduced configurable "token budgets" that track spending across agent threads, warn you when you're running low, and automatically hit the brakes if your budget runs out. When managing agents, you can now configure them to be fully disabled, explicit-request-only, or proactively helpful. Most interestingly, the system now has a sense of time: it can receive scheduled UTC reminders and proactively check the current time. Like hiring an accountant with a stopwatch: they help you cash in your discount vouchers, watch over every employee's expenses so you never go over budget, and actively remind you when it's time for a meeting.
Finding tools and information is no longer like searching for a needle in a haystack. Running /plugins now neatly organizes remote plugins into "OpenAI Curated", "Workspace", and "Shared with me" sections, and the system will proactively recommend relevant plugins right in the middle of your conversation. Installing plugins is also much more robust: it gracefully handles weird marketplace layouts, manifest fallbacks, and download errors, making new tools available immediately after installation. We've also added an "indexed web-search" mode that lets the AI search the web live, while strictly restricting its direct page access to only server-approved safe URLs. Like visiting a newly renovated hardware store: all the tools are perfectly categorized, the clerk proactively hands you a wrench when they see you fixing a pipe; and if you need to look up a repair manual online, they'll search it for you while making sure never to accidentally click on sketchy websites.
Remote development is now just as seamless and secure as working locally. When running remotely, the system now perfectly preserves the host operating system's native path formats, shell environments, and security sandboxes; it even correctly discovers and reads AGENTS.md on the remote side. Even better, the underlying execution engine and tool connections (MCP) now have "nine lives": if your network drops unexpectedly, they can survive transient disconnects by refreshing signed URLs and safely retrying inputs to resume exactly where they left off. Like sending a remote-controlled rover to another planet: not only does it perfectly adapt to the local gravity and atmosphere, but if your communications link cuts out for a few seconds, it doesn't just crash. It patiently waits for you to reconnect and casually resumes its work.
Parent agents (manager AIs) will no longer be fooled: if a subagent (employee AI) crashes or fails fatally, the manager now correctly receives an error report instead of a fake "successfully completed but empty" report. To make everything feel snappier, the system skips unnecessary network lookups at startup, warms up the model cache, parallelizes data reads, and stops spamming your hard drive with duplicate connection logs. If you suspend the UI in Linux using Ctrl+Z and resume with fg, the rendering will no longer break. Additionally, those missing "Goal-first" threads are finally being saved and returned in your history list correctly. Like catching that one lazy employee who always turned in blank sheets while claiming the work was done, while simultaneously optimizing the whole morning routine to cut commute and meeting times in half, and ensuring yesterday's paused projects are still right there when you return.
`/usage` can now show and redeem earned usage-limit reset credits, with confirmation, retry, and refreshed availability states. (#28154, #28793)
Configurable rollout token budgets track usage across agent threads, provide remaining-budget reminders, and abort turns when exhausted. (#28746, #28494, #28707, #29423)
App-server clients can configure multi-agent delegation as disabled, explicit-request-only, or proactive at the thread and turn level. (#28685, #28792, #29324)
Codex can now receive scheduled UTC time reminders and query the current time directly, including through client-provided app-server clocks. (#28822, #28824, #28835, #29011)`/plugins` now organizes remote plugins into OpenAI Curated, Workspace, and Shared with me sections, while eligible turns can recommend and install relevant plugins. (#26703, #28399, #28400, #27704, #28403)
Added an indexed web-search mode that permits live searches while restricting direct page access to server-approved URLs. (#28489)
Plugin loading and installation now handle root marketplace layouts, manifest fallbacks, multiple skill paths, actionable download errors, and immediate tool refreshes. (#28771, #28789, #28790, #28863, #28951)Exec-server processes and stdio MCP sessions now survive transient disconnects, including signed-URL refresh and retry-safe stdin writes. (#28512, #28374, #28546, #28895)
Remote environments now preserve executor-native paths, shells, `AGENTS.md` discovery, and sandbox behavior across operating systems. (#28146, #28152, #28958, #28983, #29099, #29108, #29113, #29424)Parent agents now receive terminal subagent errors instead of seeing failed work as an empty successful completion. (#28375)
Goal-first threads are once again persisted and returned by `thread/list` and `thread/search`. (#28808)
Reduced startup and session latency by deferring unnecessary DNS work, warming the model cache, reusing parsed plugin skills, parallelizing skill metadata reads, and skipping redundant catalog synchronization. (#28542, #28699, #28844, #29326, #29005)
Reduced persistent-log churn by removing per-event WebSocket payload logging and filtering duplicated telemetry records. (#29432, #29457)
Restored reliable Linux TUI rendering after suspending with `Ctrl+Z` and resuming with `fg`. (#28342)