發布日期：2026-07-21
趕著回覆訊息時，你會想要快；遇到棘手問題時，則願意多等一下換更深的推理。新的 /effort 指令把這個取捨做成時間軸儀表，可左右選擇，也能直接輸入等級；啟動 CLI 時則用 --effort 選定。/model 現在用穩定的 model slug（模型的固定代號）列出選項，跨工作階段也能可靠指定同一款模型。你替 custom agent（自訂代理，預先寫好工作規則的小幫手）加上 model 後，叫它當子代理時可跑 flash 或 pro；沒指定便沿用父代理。像相機不只選鏡頭，還能轉快門：快拍或細拍，由你在當下決定。
筆電分割畫面時，設定清單不該像過長的海報從視窗底下溢出去。/settings 或 /config 現在是有邊界、可捲動的清單，開關下拉選單也不再閃爍。你查看程式修改處時，/diff 檢視器在換頁、長行折行或展開留言後，不會抖動，更不會把狀態列擠出螢幕。連輸入框也聽得懂你的重綁 Enter：若綁到 prompt.insert_newline，按下去會換行而非送出。這些改動像把凌亂書桌換成有抽屜的工作台；內容還是那些內容，手邊位置終於穩了。
讓 CLI 在背景慢慢更新資料時，最怕其中一件小事出錯，整個工作階段跟著亂掉。現在長時間背景任務共用同一套生命週期，啟動與關閉都有固定順序，啟動時也能安全面對 panic（程式遇到嚴重錯誤而中斷）；單一任務失敗不再拖垮整場，離開前待送的分析資料也會送完。短時間內連續觸發更新，則會合併成一次，少做重複工作。收集自訂規則時的 data race（多個工作同時碰同份資料造成結果不穩）也補上保護。像廚房把各道菜分到獨立爐火，並把連按三次的叫餐鈴合成一張單；一口鍋燒焦，不會讓整間餐廳停擺。
你在 /mcp 面板替遠端 MCP（Model Context Protocol，讓 AI 連接外部工具的協定）伺服器按 Authenticate 時，現在不會再把 CLI 按到當機。更重要的是，MCP 工具回傳的 embedded resources（嵌入式資源，文字或內嵌媒體等隨結果附上的內容）以前會被悄悄丟掉，現在會出現在對話裡。也就是說，接上外部工具後，登入這扇門能打開，工具端帶回來的說明或圖片也不會留在門外。像快遞站以前辦證時會關門，包裹裡附的文件還會被抽走；現在櫃台照常辦理，附檔也完整交到你手上。
你下 --grep="a|b" 這類指令時，管線符號明明只是引號裡的參數，權限檢查卻曾誤以為你真要建立 pipeline（管線，把一個指令輸出交給下一個指令的串接），跳出多餘提示；現在能正確看懂。檢視或搜尋檔案時，若多位元字元剛好被截在中間，也不再因 invalid UTF-8 而失敗。資格不符時，CLI 也重新顯示真正原因，不再只說「unknown reason」。這像機場安檢終於會分辨行李標籤上的豎線和危險物，掃描到半截外文字也不會當成壞行李；拒絕登機時，還會告訴你缺的是哪份證件。
規劃完還想立刻請它挑毛病？現在同一句可排 /plan /grill-me <prompt>，每個開頭斜線指令會依輸入順序解析、啟用並顯示。custom agent 若明確寫 subagent: false，也不會再混進可呼叫的子代理清單。用 -p 或 --print 跑 headless（無頭模式，沒有終端機視窗／視覺介面的執行模式）時，已儲存的 settings.json 政策終於會生效，包括權限、檔案存取、sandbox mode（沙盒模式，把程式限制在安全範圍內的執行方式）、自動執行與成品審查。/btw 的岔題也不再複製成另一筆同標題對話。像火車站的轉乘牌：能依序換線，但標著不接客的月台不會被派車，沒有站務員的夜車仍照既定規則行駛。
When a reply is urgent, you want speed; when a problem is thorny, you may accept a wait for deeper reasoning. The new /effort command turns that tradeoff into a left/right timeline gauge, and also accepts a level directly; --effort makes the choice when launching the CLI. /model now shows stable model slugs, fixed names you can reliably reuse across sessions. A custom agent, a helper with prewritten instructions, can now declare its own model: as a subagent it can use a tier such as flash or pro, otherwise it inherits its parent’s model. It is like setting both a camera lens and shutter speed: choose quick or deliberate for task in front of you.
On a split laptop screen, a settings list should not spill below the window like an overlong poster. /settings and /config are now bounded, scrollable lists, and their dropdowns no longer flicker. While reviewing code changes, the /diff viewer no longer jitters or pushes the status line offscreen when you page, wrapped lines appear, or comments expand. The prompt also honors a remapped Enter: if it is bound to prompt.insert_newline, it inserts a line break instead of submitting. It is like replacing a messy desk with a workbench full of drawers: same tools, but they stay where your hands expect them.
When the CLI refreshes data in the background, one small failure should not scramble the whole session. Long-running background tasks now share one lifecycle, with deterministic start and shutdown and panic-safe launching; one failed task no longer disrupts the session, and pending analytics are flushed before exit. Several refresh triggers arriving in a burst are coalesced into one run instead of repeating work. A data race, where concurrent work touches shared data unpredictably, is also guarded while customization rules are collected. It is like giving each dish in a kitchen its own burner and turning three rapid bell rings into one order: one burnt pan does not close the restaurant.
Pressing Authenticate for a remote MCP server in the /mcp panel no longer crashes the CLI. More importantly, embedded resources, extra text or inline media carried inside an MCP tool result, used to be silently discarded; they now appear in the conversation. In other words, once you connect an external tool through MCP, the authentication door opens and its explanation or image can come through with it. It is like a parcel counter that once shut down while checking your ID and removed documents from packages; now the counter stays open and every enclosure reaches your hands.
With a command such as --grep="a|b", a pipe character inside a quoted argument is not a pipeline, a chain that passes one command’s output to another. Permission checks used to misread it and ask unnecessarily; now they interpret it correctly. File viewing and search also stop failing with invalid UTF-8 when a multibyte character is cut at a truncation boundary. And when you are not eligible, the CLI again shows the real reason rather than “unknown reason.” It is like airport screening that can distinguish a vertical mark on a luggage tag from a hazardous item, and can explain which document is missing when it cannot let you board.
Finished planning and want an immediate critique? One prompt can now begin /plan /grill-me <prompt>; each leading slash command is parsed, activated, and displayed in typed order. A custom agent explicitly marked subagent: false no longer sneaks into the callable-subagent list. Headless runs, execution through -p or --print without a terminal window, now honor saved settings.json policies: permissions, file access, sandbox mode, auto-execution, and artifact review. A /btw side question also stops making a duplicate conversation with its parent’s title. It is like a station transfer board: you can change lines in order, closed platforms are not assigned trains, and an unattended night train still follows its rules.
Added a `/effort` command to view and change the current model's reasoning effort, with a left/right timeline-gauge picker and a direct `/effort <level>` form so you can trade latency for depth on the fly.
Added an `--effort` flag to select a model's reasoning-effort variant when launching the CLI.
Added stable, user-facing model slugs that appear in the `/model` picker and are accepted by `--model`, so you can pin a specific model reliably across sessions.
Added a `model` option to custom agent frontmatter so an agent runs at a chosen model tier (such as `flash` or `pro`) when invoked as a subagent, defaulting to `inherit` (the parent's model).
Redesigned the `/model` picker to group models by their base model and choose reasoning effort from a timeline gauge navigable with Left and Right, and added an effort badge to the status line for models that expose multiple effort variants.Improved the `/settings` (`/config`) panel by making it a bounded, scrollable list so it renders correctly in short terminals instead of overflowing, and stopped it from flickering when opening and closing dropdowns.
Improved scrolling in the `/diff` viewer so paging through a diff no longer jitters or pushes the status line off the screen when lines wrap or comments expand.
Fixed the prompt to honor a custom Enter binding to `prompt.insert_newline`, so a remapped Enter inserts a newline instead of submitting.Improved background-task reliability by moving long-running work onto a shared lifecycle with deterministic startup and shutdown and panic-safe launching, so a failure in one background task no longer disrupts the session and pending analytics are flushed on exit instead of dropped.
Improved responsiveness of bursty background refreshes by coalescing rapid repeated triggers into a single run, cutting redundant work.
Fixed a data race when collecting customization rules by guarding the shared structures.Fixed a crash when triggering Authenticate on a remote MCP server in the `/mcp` panel.
Fixed MCP tool results containing embedded resources being silently dropped, so text and inline media returned by MCP servers now surface in the conversation.Fixed permission checks splitting a single command into a pipeline when an argument contained quoted shell metacharacters (such as `--grep="a|b"`), which caused spurious permission prompts.
Fixed the file-view and file-search tools failing with invalid-UTF-8 errors when a multi-byte character was split at a truncation boundary.
Fixed eligibility error messages so the CLI shows the real reason again instead of defaulting to a generic "unknown reason".Added support for stacking multiple leading slash commands in a single prompt, so a chain like `/plan /grill-me <prompt>` parses, activates, and renders every command in the order you typed them.
Fixed custom agents that declare `subagent: false` still appearing in the available-subagents list and being invocable as subagents.
Fixed headless (`-p` / `--print`) runs so they now honor persisted `settings.json` policies, including `permissions`, file access, sandbox mode, auto-execution, and artifact review.
Fixed `/btw` side-questions leaking into the conversation list as duplicate entries that carried the parent conversation's title.