v1.3.2

v1.3.1 -> v1.3.2

Summary

Shrinks the tool surface around live data and privileges: the agent now calls public APIs directly and escalates through one password-gated command path instead of dedicated tools. Background work is reshaped into two purpose-built tools that always watch their output.

翻譯

收斂即時資料與權限相關的工具：agent 改為直接呼叫公開 API，提權統一走單一條需要密碼確認的指令路徑，不再依賴專用工具。背景工作重整為兩個專用工具，且一律監看輸出。

Changes

FEAT

• Add built-in public_api_list and route live-data requests to read and call a fitting keyless public API directly, building a script tool only when none fits or a user-supplied API needs a key (@pardnchiu) [74769ee, afa20d8]
• Support sudo argv in run_command / background_command, run as sudo -n outside the sandbox after a system-password confirmation, and prompt for the password before a TUI ! shell command that uses sudo (@pardnchiu) [26ba436]
• Add background_subagents for long subagent legs, writing each leg's progress to the session background log and never prompting from a background leg (@pardnchiu) [f8e9c4f]
• Watch every background_command for error keywords, resuming the task with the latest output while the command runs; a decision model, when set, drops benign hits and repeats (@pardnchiu) [b92cb8b, f8e9c4f]

翻譯

• 新增內建 public_api_list，即時資料請求改為讀文件後直接呼叫合適的免金鑰公開 API，只有找不到合適的 API 或使用者提供的 API 需要金鑰時才建立 script tool
• run_command／background_command 支援 sudo argv，經系統密碼確認後以 sudo -n 在沙箱外執行；TUI 的 ! shell 指令使用 sudo 時會先跳出密碼框
• 新增 background_subagents 處理耗時的 subagent，每個 leg 的進度寫入 session 背景 log，背景 leg 不會跳出確認
• 所有 background_command 都會監看錯誤關鍵字，指令執行中即帶最新輸出接續任務；設定 decision model 時會先濾掉誤判與重複回報

UPDATE

• Enable decision routing whenever a decision model is selected (@pardnchiu) [040067d]
• Offer background_command and background_subagents to TUI requests regardless of the decision model (@pardnchiu) [b92cb8b, f8e9c4f]
• Return a hint to re-run with sandbox: false when the sandbox blocks a command with no path to grant, and tell the agent in both permission modes to escalate and continue instead of stopping to ask (@pardnchiu) [afa20d8]
• Include the live log path in the background_command start result so progress questions read the log instead of guessing (@pardnchiu) [afa20d8]
• Move the restricted-path approval notice from the TUI to the debug log (@pardnchiu) [26ba436]
• Clear background log directories together with background tasks (@pardnchiu) [f8e9c4f]
• Update bundled skills to install dependencies through run_command and drop API tool references (@pardnchiu) [74769ee, afa20d8]

翻譯

• 只要選了 decision model 就啟用 decision 分派
• background_command 與 background_subagents 對 TUI 請求開放，不再依賴 decision model
• 沙箱擋下指令且找不到可放行的路徑時，回傳改用 sandbox: false 重跑的提示；兩種權限模式都要求 agent 直接提權並繼續，不停下來詢問
• background_command 啟動結果附上即時 log 路徑，詢問進度時讀 log 而非推測
• 受限路徑核准通知從 TUI 改寫進 debug log
• 清除背景任務時一併清除背景 log 目錄
• 內建 skill 改以 run_command 安裝依賴，並移除 API tool 相關描述

REMOVE

• Remove the user-defined API tool adapter and its directories (@pardnchiu) [74769ee]
• Remove pkg_manage (@pardnchiu) [afa20d8]
• Remove the dispatcher_beta toggle (@pardnchiu) [040067d]

翻譯

• 移除使用者自訂 API tool 的 adapter 與相關目錄
• 移除 pkg_manage
• 移除 dispatcher_beta 開關

Scope

• internal/runtime/toolAdapter/api/, extensions/apis/, extensions/extensions.go — REMOVE
• internal/tools/external/ — FEAT (publicAPIList.go, toolAdapter.go)
• internal/tools/ — FEAT, UPDATE, REMOVE (runCommand.go, runCommandBackground.go, watchOutput.go, legacyAPITool.go)
• internal/tools/subagent/, internal/agents/exec/subagent/ — FEAT, UPDATE
• internal/tools/interactive/ — UPDATE, REMOVE (pkgManage.go, background.go, backgroundStart.go)
• internal/tools/file/variant/, internal/tools/searcher/, internal/tools/history/action/ — UPDATE
• internal/agents/exec/ — UPDATE (multiple files)
• internal/runtime/mcp/ — UPDATE
• internal/runtime/tui/ — FEAT, UPDATE
• internal/session/config/, internal/runtime/routes/handler/, page/ — REMOVE
• internal/filesystem/, internal/utils/ — UPDATE
• configs/ — UPDATE
• extensions/skills/ — UPDATE
• doc/ — DOC
• package-lock.json — CHORE

---

Generated by SKILL