發布日期：2026-09-15
一邊等待 AI 處理主要工作，一邊透過 /btw 提出旁支問題，取用解答文字現在更加俐落。
當你交代助理撰寫長篇企劃，趁空檔用 /btw 查詢專有名詞，讀完後想把整段答覆貼進工作報告，這正是輸入 /copy btw 的時刻。這項新增的斜線子指令，能將當前進行中的 /btw 旁支回答全文直接複製至剪貼簿，就算回覆卡片已經收合或被捲動移出可見畫面，也完全不影響複製。此外，當你在終端機鍵入 /copy 時，指令輸入列還會顯示參數的幽靈提示（輸入指令時以淡色顯示的待填參數建議），提醒你能接續輸入哪些可用選項。
指揮 AI 配合特定流程呼叫工具時，模型遵循呼叫限制的表現現在更加精準可靠。以前在每次請求中指定工具選用約束，例如限制必須呼叫任意工具的 any、強制調用工具的 required，或是指定特定具名函式，若未深入底層，執行時偶爾仍可能偏離指示。現在系統會將每筆請求的工具選用約束，直接轉換為後端原生的函式呼叫設定（由模型服務底層介面直接控管工具挑選的機制），確保工具呼叫的保真度徹底落實。
檢查自動化流程的掛鉤設定時，外掛隨附的規則不再無端隱形。以前在終端機執行 /hooks 指令或使用掛鉤檢查工具列出使用中的 hooks.json 設定時，系統會漏掉已啟用外掛套件裡自帶的 hook（掛鉤，特定事件發生時被觸發的小程式），導致你難以確認外掛功能是否如期攔截；這個遺漏問題現在已經修好，清單會如實列出已啟用外掛隨附的所有掛鉤設定。
配置專門分工的自訂子代理時，工具連線與檔案路徑的兩項問題已獲修正。以前在建立自訂子代理時就算指定了 enable_mcp_tools: true，子代理也只會收到一份空的伺服器清單，無法順利繼承父代理原先設定妥當的 MCP（Model Context Protocol，讓 AI 連接外部工具的協定）伺服器；同時，宣告式子代理設定若使用 relative_path_to_config 指定相對設定檔路徑，系統也無法以父代理所在的目錄為基準完成解析。這兩項路徑與設定繼承問題現在都已經修好，子代理會確實繼承父代理的外部工具伺服器，相對路徑也能順利對齊。
在終端機展開新對話或使用工具時，評分詢問不再無端跳出打擾節奏。以前只要工作階段剛開始的前幾輪對話、或者工具遇到短暫發生但隨後已順利復原的暫態錯誤，系統的主動回饋提示就會過早觸發，在任務尚未明朗時突然向你索取意見；這個過早觸發的問題現在已經修好，不會再在對話初期或暫態錯誤復原時過早彈出提示打斷思路。
While waiting for AI to handle primary tasks, asking side questions via /btw and retrieving the answer text is now much smoother.
When you ask an assistant to draft a long-form proposal and use /btw during downtime to look up technical terms, and then want to paste the full answer into your report after reading, this is the exact moment to type /copy btw. This newly added slash subcommand copies the full text of the active /btw side-question response directly to your clipboard, even if the response card is collapsed or scrolled out of view. In addition, when you type /copy in the terminal, the command input bar displays argument ghost hints (faded suggestions for parameters to fill in while typing commands) to remind you of available options you can enter next.
When directing AI to call tools for specific workflows, the model's adherence to calling restrictions is now more precise and reliable. Previously, when specifying per-request tool-choice constraints—such as any requiring calling any tool, required enforcing tool invocation, or specifying named functions—executions could still occasionally deviate from instructions if not integrated into the underlying layer. Now, the system directly translates per-request tool-choice constraints into native backend function-calling configurations (a mechanism where the model service's underlying interface directly manages tool selection), ensuring complete fidelity in tool calling.
When inspecting hook configurations in automated workflows, rules bundled with plugins no longer invisibly disappear. Previously, when running the /hooks command in the terminal or using hook inspection utilities to list active hooks.json configurations, the system omitted hooks bundled inside enabled plugin packages (small programs triggered when specific events occur), making it difficult to verify whether plugin features were intercepting as expected. This omission issue is now fixed, and listings faithfully include all hook configurations bundled with enabled plugins.
When configuring custom subagents for specialized division of labor, two issues with tool connections and file paths have been fixed. Previously, even if enable_mcp_tools: true was specified when creating a custom subagent, the subagent only received an empty server list and failed to inherit the configured MCP (Model Context Protocol, a protocol allowing AI to connect to external tools) servers from the parent agent. At the same time, if a declarative subagent configuration specified a relative configuration file path using relative_path_to_config, the system failed to resolve it against the parent agent's directory. Both path and configuration inheritance issues are now fixed, ensuring subagents properly inherit the parent agent's external tool servers and relative paths align smoothly.
When starting a new conversation in the terminal or using tools, feedback rating prompts no longer pop up unexpectedly to disrupt your flow. Previously, during the first few turns of a session or when tools encountered transient errors that quickly recovered, proactive feedback prompts triggered prematurely, soliciting feedback before tasks were clear. This premature triggering issue is now fixed, preventing pop-up prompts from interrupting your train of thought during early conversations or recovered transient errors.
Added the `/copy btw` slash subcommand to copy the full text of the active `/btw` side-question response to the clipboard even when the response card is collapsed or scrolled, along with argument ghost hints when typing `/copy`.Improved tool-calling fidelity by translating per-request tool-choice constraints (`any`, `required`, and named functions) into native backend function-calling configurations.Fixed the `/hooks` command and hook inspection utilities omitting hooks bundled inside enabled plugins when listing active `hooks.json` configurations.Fixed custom subagents created with `enable_mcp_tools: true` receiving an empty MCP server list instead of inheriting the parent agent's configured MCP servers, and fixed declarative subagent configurations failing to resolve relative config file paths (`relative_path_to_config`) against the parent agent's directory.Fixed proactive feedback prompts triggering prematurely during the first few turns of a session or on transient, recovered tool errors.