v0.7.0 Release Notes

Released 2026-02-27

v0.6.0 -> v0.7.0

Summary

Major architectural refactor, LLM-driven agent routing, OpenAI-compatible endpoint support, and enhanced session reliability through file locking and atomic writes.

翻譯

大型架構重構，LLM 驅動的 Agent 路由、支援 OpenAI 相容端點，以及透過檔案鎖與原子寫入提升 session 可靠性。

Changes

FEAT

• Add LLM-driven automatic agent routing mechanism to dynamically select the best provider per request
• Add compat provider supporting any OpenAI-compatible endpoint (defaults to Ollama at localhost:11434)
• Add search_history tool for searching session conversation history with time-range filtering (1d/7d/1m/1y)
• Add duplicate tool-call result caching to avoid redundant API requests within a session
• Add post-iteration automatic summary when MaxToolIterations is exceeded instead of returning an error
• Add Unicode parameter normalization (normalizeArgs) to handle \uXXXX escape sequences from LLM outputs

翻譯

• 新增 LLM 驅動的自動 Agent 路由機制，依請求動態選擇最佳 provider
• 新增 compat provider，支援任意 OpenAI 相容端點（預設 Ollama localhost:11434）
• 新增 search_history 工具，支援時間範圍過濾（1d/7d/1m/1y）的 session 歷史搜尋
• 實作重複工具呼叫結果快取，避免 session 內冗餘 API 請求
• 超過 MaxToolIterations 後自動呼叫摘要，不再直接回傳錯誤
• 新增 normalizeArgs 處理 LLM 輸出的 \uXXXX Unicode 轉義序列

FIX

• Fix concurrent session race condition by acquiring an exclusive flock on config.json before read/write
• Fix config.json initialization with atomic O_EXCL write to prevent duplicate session creation under concurrency
• Fix corrupted config.json detection and explicit error propagation

翻譯

• 修正並發 session 競態，讀寫 config.json 前取得排他 flock 鎖
• 使用 O_EXCL 原子寫入初始化 config.json，防止並發時建立重複 session
• 修正損毀 config.json 的偵測並明確傳遞錯誤

UPDATE

• Force agent selector to always return a valid agent from the available list; remove NONE fallback response
• Replace dynamic limit parameter in search_web and search_history tools with a fixed cap of 10
• Reduce MaxToolIterations default from 32 to 8; introduce MaxSkillIterations (128) for skill-mode execution
• Add default_model field support in agent config to control fallback agent ordering

翻譯

• 強制 Agent 選擇器必須從可用清單回傳有效代理，移除 NONE fallback
• 將 search_web 和 search_history 的動態 limit 參數改為固定 10 筆上限
• 預設 MaxToolIterations 從 32 降至 8；skill 模式改用 MaxSkillIterations（128）
• Agent 設定檔新增 default_model 欄位控制 fallback 順序

REFACTOR

• Split monolithic agents/exec.go into exec/ sub-package with responsibility-separated files (execute, run, toolCall, getSession, selectAgent, selectSkill, writeHistory, extractSummary)
• Extract shared types (Agent, Message, Session, AgentRegistry, AgentEntry, Output, OutputChoices, ToolCall) into internal/agents/types package
• Simplify tool result message format; remove redundant Tool 'x'\nresult: prefix

翻譯

• 將單體 agents/exec.go 拆分為 exec/ 子套件，按職責分離各檔案
• 將共用型別（Agent、Message、Session、AgentRegistry 等）抽離至 internal/agents/types 套件
• 簡化 tool result 訊息格式，移除冗餘的 Tool 'x'\nresult: 前綴

PERF

• Replace int64 arithmetic in calculator with big.Int for exact integer precision
• Reduce Google RSS, Yahoo Finance, and general API timeouts from 30s to 5s
• Cap Google RSS deduplicated results at 10 entries

翻譯

• 計算器整數運算改用 big.Int 確保精確度
• Google RSS、Yahoo Finance 及一般 API 超時從 30 秒縮短至 5 秒
• Google RSS 去重後結果上限設為 10 筆

---

Files Changed

---

Generated by SKILL