新功能
新增 /skills reload 指令，免重啟工作階段即時重載技能
在終端機對話中途擴充技能或自訂指令，現在完全不必打斷當前思路。
當你正在跟 AI 協同開發，臨時新增了一項技能檔案或自訂斜線指令，想立刻在對話中使用卻不想結束當前對話，這正是鍵入 /skills reload 的時刻。這項新增的子指令會在背景非同步重新載入所有偵測到的技能與斜線指令，既不用重新啟動工作階段（session），也不會暫停或鎖定你的終端機鍵盤輸入，讓你一邊繼續提問、一邊立即取用最新擴充。
繁體中文
新功能
新增 /model 行內自動補齊與互動選單模糊搜尋列
切換對話所使用的運算模型時，現在不必再死記複雜的型號代碼。
當你需要針對複雜任務切換運算模型，這正是輸入 /model 的時刻。現在只要在終端機輸入 /model 時，輸入列就會即時彈出行內參數的自動補齊下拉選單；若進入互動式選單介面，上方還新增了即時模糊搜尋列，讓你能直接輸入模型名稱或型號 ID 快速過濾篩選，挑出最合適的目標。
新功能
OAuth 權限不足時明確提示執行 /logout 與 /login 指令
登入驗證遇到權限不足時，錯誤提示現在會給出明確的修復指引。以前當你的 OAuth token（授權令牌）缺少系統所需的存取權限範圍時，終端機只會拋出晦澀的授權失敗訊息，讓人不知該如何處理。現在系統改進了身分驗證錯誤提示，會直接在畫面中列出清晰的修復步驟，主動指引你依序執行 /logout 登出與 /login 重新登入指令以更新權限。
問題修復
修復切換至 Gemini 思考模型與聯網搜尋摘要的解析異常
無論中途切換至具備推理能力的 Gemini 思考模型，或是執行聯網搜尋，運作流程現在更加順暢。以前若前幾輪對話是由非 Gemini 模型所產生，中途切換為 Gemini 思考模型時，系統會因為找不到前序對話的 thought_signature（思考簽章）而跳出格式驗證錯誤。此外，使用 search_web 聯網搜尋工具時，若負責產出摘要的模型在開頭吐出了思考過程片段，也會導致搜尋以 "no summary returned from GenerateContent" 錯誤失敗。這兩項思考模型解析問題現在都已經修好，模型切換與搜尋摘要皆能如常運作。
問題修復
工具呼叫格式錯誤時代理迴圈會回傳錯誤並自我修正
讓 AI 自動調用工具時，若參數格式不合規，流程不再直接斷線罷工。以前當工具呼叫未通過結構驗證（schema validation）時，對話輪次會過早以 NO_TOOL_CALL 錯誤直接終止，導致工作被迫中斷。現在這個問題已經修好，代理迴圈會將驗證錯誤訊息回傳給 AI，讓模型能即時理解錯誤並自動修正參數重新呼叫。
問題修復
修復 MCP 工具定義原地修改與 hooks.json 遭預算截斷問題
自訂擴充設定與外部工具整合，現在運作更加穩定可靠。
以前系統在為 MCP（Model Context Protocol，讓 AI 連接外部工具的協定）擴充工具結構時，會直接在記憶體中就地修改工具定義，導致後續觸發嚴格參數驗證失敗。同時，當自訂內容啟動 token（AI 處理文字的計量單位）預算截斷機制時，原本在 hooks.json 設定的掛鉤規則還會無預警被整份丟棄。這兩項設定遺失與參數驗證問題現在都已修好。
問題修復
修復自訂子代理 exclude 技能過濾遺失與終止狀態被誤重設問題
指派專門分工的自訂子代理（subagent，負責特定分工的輔助 AI）協同工作時，權限過濾與狀態追蹤更加精確。以前當系統在執行階段套用技能過濾時，自訂子代理會遺失原本設定的 exclude 技能排除清單。同時，在發布訂閱機制的批次合併更新下，已被終止的子代理也容易因晚到的步驟更新而被錯誤重設為活躍狀態。這兩項技能過濾與狀態追蹤問題現在都已經修好。
問題修復
修復重啟後對話與背景常駐工作無法自動復原及日誌遺漏問題
終端機重新啟動後的對話復原與除錯記錄，現在更加完整可靠。以前當對話中仍有正在進行的發言輪次、或是背景有正在執行的 daemon（背景常駐程式）工作時，重新啟動後往往無法順利自動恢復對話。同時，終端機中發生的 ERROR 錯誤步驟（例如工具執行失敗或逾時），過去也常在 transcript.jsonl 對話記錄檔中被意外漏掉。這兩項工作階段自動復原與日誌記錄問題現在都已經修好。
問題修復
SSH 連線於 Apple_Terminal 停用不支援的超連結跳脫序列
透過 SSH 遠端連線在 macOS 原生終端機操作時，畫面文字不再出現多餘的亂碼符號。以前使用 Apple_Terminal 建立 SSH 連線時，由於終端機不支援特定的超連結格式，畫面上會直接印出未解析的 OSC 8 原始跳脫序列（escape sequences）。現在系統已主動停用不支援的超連結序列，這個顯示問題已經修好，遠端文字呈現恢復乾淨整潔。
English
FEATURE
Added /skills reload Command to Instantly Reload Skills Without Restarting Session
Expanding skills or custom commands mid-conversation in the terminal no longer requires interrupting your flow.
When collaborating with AI and adding a new skill file or custom slash command that you want to use right away without ending the ongoing session, this is the exact moment to type /skills reload. This newly added subcommand asynchronously reloads all discovered skills and slash commands in the background without restarting the session or blocking terminal keyboard input, allowing you to keep asking questions while immediately utilizing the latest extensions.
FEATURE
Added Inline Autocomplete for /model and Fuzzy Search Bar in Interactive Picker
When switching compute models during a conversation, you no longer need to memorize complex model IDs.
When you need to switch compute models for complex tasks, this is the moment to enter /model. Typing /model in the terminal now instantly pops up an inline argument autocomplete dropdown; entering the interactive menu interface provides a live fuzzy search bar at the top, enabling you to filter quickly by model name or ID and pick the most suitable option.
FEATURE
Explicit Instructions to Run /logout and /login When OAuth Permissions Are Insufficient
Error messages now provide explicit remediation instructions when encountering insufficient permissions during authentication. Previously, when an OAuth token lacked required access scopes, the terminal merely threw an obscure authorization failure message, leaving users unsure how to resolve it. Authentication error prompts have now been improved to display clear recovery steps directly on screen, actively guiding you to run /logout and /login sequentially to refresh permissions.
BUGFIX
Fixed Parsing Errors When Switching to Gemini Thinking Models and Web Search Summaries
Workflows now run much smoother whether switching mid-conversation to Gemini thinking models equipped with reasoning capabilities or executing web searches. Previously, when prior conversation turns were generated by non-Gemini models, switching mid-session to a Gemini thinking model would fail with format validation errors due to missing thought_signature from previous turns. In addition, when using the search_web tool, if the summarization model emitted leading thought parts, the search would fail with a "no summary returned from GenerateContent" error. Both parsing issues with thinking models have now been fixed, allowing model switching and search summaries to function as expected.
BUGFIX
Agent Loop Returns Errors and Self-Corrects on Tool Call Validation Failures
When letting AI invoke tools automatically, invalid argument formats no longer abruptly terminate the workflow. Previously, when a tool call failed schema validation, the conversation turn would terminate prematurely with a NO_TOOL_CALL error, forcing work to halt. This issue has now been resolved: the agent loop returns the validation error message to the AI, allowing the model to promptly understand the error, correct its arguments, and retry the call.
BUGFIX
Fixed In-Place MCP Tool Definition Mutation and hooks.json Token Budget Truncation
Custom extension configurations and external tool integrations now operate with greater stability and reliability.
Previously, when the system augmented tool schemas for MCP (Model Context Protocol), it mutated tool definitions in place in memory, triggering subsequent strict argument validation failures. Meanwhile, when customization token budget truncation was activated, hook rules configured in hooks.json were silently dropped without warning. Both configuration loss and parameter validation issues have now been resolved.
BUGFIX
Fixed Custom Subagent exclude Skill Filter Loss and Accidental Reset of Termination Status
Permission filtering and status tracking are now more precise when delegating tasks to specialized custom subagents. Previously, when runtime skill filtering was applied, custom subagents would lose their configured exclude skill filter lists. Additionally, under pubsub batch coalescing updates, killed subagents could be mistakenly reset to active status by late step updates. Both skill filtering and status tracking issues have now been fixed.
BUGFIX
Fixed Conversation and Daemon Background Task Auto-Resume Failures and Omitted Logs After Restart
Conversation recovery and debug logging following terminal restarts are now more complete and dependable. Previously, when a conversation had active turns or daemon background tasks running, auto-resuming the session smoothly after a restart often failed. Furthermore, terminal ERROR steps (such as failed tool executions or timeouts) were frequently omitted from transcript.jsonl conversation logs. Both session auto-recovery and logging issues have now been resolved.
BUGFIX
Disabled Unsupported Hyperlink Escape Sequences in Apple_Terminal Over SSH Connections
Terminal text no longer displays stray garbled characters when operating over SSH connections in native macOS terminals. Previously, when establishing SSH connections using Apple_Terminal, unsupported hyperlink formats caused raw, unparsed OSC 8 escape sequences to print directly to the screen. The system now proactively disables unsupported hyperlink sequences, resolving this display glitch and restoring clean remote text rendering.
原始 CHANGELOG
Original changelog新增 /skills reload 指令，免重啟工作階段即時重載技能
Added a `/skills reload` subcommand to asynchronously reload discovered skills and slash commands without restarting the session or blocking user input.Original changelog新增 /model 行內自動補齊與互動選單模糊搜尋列
Added an inline argument autocomplete dropdown when typing `/model` and a live fuzzy search bar inside the interactive `/model` picker to filter models by name or ID.Original changelogOAuth 權限不足時明確提示執行 /logout 與 /login 指令
Improved authentication error messages when OAuth tokens lack required scopes by including explicit instructions to run `/logout` and `/login`.Original changelog修復切換至 Gemini 思考模型與聯網搜尋摘要的解析異常
Fixed mid-conversation model switching to Gemini thinking models failing with `thought_signature` validation errors when prior turns were generated by non-Gemini models.
Fixed web search (`search_web`) failing with `"no summary returned from GenerateContent"` when the summarization model emits leading thought parts.Original changelog工具呼叫格式錯誤時代理迴圈會回傳錯誤並自我修正
Fixed agent turns terminating prematurely with `NO_TOOL_CALL` when a tool call fails schema validation, allowing the agent loop to return the validation error and self-correct.Original changelog修復 MCP 工具定義原地修改與 hooks.json 遭預算截斷問題
Fixed MCP tool schema augmentation mutating tool definitions in place, preventing strict argument validation failures.
Fixed `hooks.json` configurations being silently dropped when customization token budget truncation is active.Original changelog修復自訂子代理 exclude 技能過濾遺失與終止狀態被誤重設問題
Fixed custom subagents losing configured `exclude` skill filters when runtime skill filtering is applied.
Fixed subagent status and termination tracking under pubsub batch coalescing so killed subagents are not reset to active by late step updates.Original changelog修復重啟後對話與背景常駐工作無法自動復原及日誌遺漏問題
Fixed conversations with active turns or running daemon background tasks failing to auto-resume after a restart.
Fixed terminal `ERROR` steps (such as failed tool executions or timeouts) being omitted from `transcript.jsonl` conversation logs.Original changelogSSH 連線於 Apple_Terminal 停用不支援的超連結跳脫序列
Fixed raw OSC 8 escape sequences rendering in `Apple_Terminal` over SSH connections by disabling unsupported hyperlink sequences.