v1.3.16 Release Notes
Released
v1.3.15 -> v1.3.16
Summary
Speech and image generation move out of the tool layer into shared agent services with one consistent shape, and stale model selections are now cleared when their credentials go away. Model routing also learns the new Gemini generation and regains its intended output format.
翻譯
語音與生圖從工具層移入共用的 agent 服務,兩者結構一致,且憑證移除後會自動清掉失效的模型選擇。模型路由也認得新一代 Gemini,並恢復原本應有的輸出格式。
Changes
FIX
- Fix the dispatcher dropping its output-format wrapper, so the model saw only the tier table without the two-line answer contract; the selector prompt is restored as the outer template and its work-kind list is now rendered from
tier.goinstead of being hand-written (@pardnchiu) [44810797] - Clear stale speech-to-text and text-to-speech selections the same way image generation already did: deleting a key or provider, opening
/model stt//model tts, or loading model routing in the WebUI now resets a model that is no longer available; the TUI routing tab also runs the cleanup before reading config, so it no longer shows the just-cleared value (@pardnchiu) [142cfd6f, 84cd1210]
翻譯
- 修正 dispatcher 遺失輸出格式外框,模型只看到 tier 表、沒有兩行回覆的規範;selector prompt 恢復為外層模板,工作類型清單改由
tier.go渲染,不再手寫 - 語音轉文字與文字轉語音比照生圖清理失效的模型選擇:刪除 key/provider、開啟
/model stt//model tts、WebUI 載入模型路由時,已不可用的模型會被重設;TUI 路由分頁也改為先清理再讀設定,不再顯示剛被清掉的值
UPDATE
- Add Gemini 4 tiers (
-carbonS,-argonA,-hydrogenB) and match DeepSeek by-pro/-flashfamily instead of prefix, with-flashplaced above other DeepSeek models (@pardnchiu) [44810797] generate_audionow honorsoutput_fileas a path likegenerate_image: absolute is used as-is and relative joins under the working directory, where it previously kept only the file name in the output directory (@pardnchiu) [142cfd6f]generate_imagefillsaspect_ratio1:1,size1k andqualitymedium when omitted instead of leaving them to the provider, and always replaces the file extension with the returned format (@pardnchiu) [84cd1210]- Rewrite agent-facing errors for missing keys, unset or mismatched speech, audio and image models, and empty generation results so each states the cause and the next step, pointing users to the TUI command first (@pardnchiu) [af814a19, 142cfd6f, 84cd1210]
翻譯
- 新增 Gemini 4 的 tier(
-carbonS、-argonA、-hydrogenB),DeepSeek 改以-pro/-flash系列比對而非前綴,-flash排在其他 DeepSeek 模型之前 generate_audio的output_file比照generate_image視為路徑:絕對路徑直接使用、相對路徑接在工作目錄下;原本只保留檔名並存到輸出目錄generate_image未指定時補上aspect_ratio1:1、size1k、qualitymedium,不再交給 provider 決定,且副檔名一律依回傳格式替換- 改寫給 agent 的錯誤訊息(缺 key、語音/生圖模型未設定或類型不符、生成結果為空),每則都說明原因與下一步,並優先引導使用者使用 TUI 指令
REFACTOR
- Move speech and image logic into
internal/agents/audioandinternal/agents/imagewith matching model, options, prune and agent-construction functions, keeping only tool registration ininternal/tools/external(@pardnchiu) [af814a19, 142cfd6f, 84cd1210] - Centralize the provider model-list cache key and its invalidation in torii, replacing four per-package copies with one call; the transcription prompt moves to an embedded prompt file (@pardnchiu) [af814a19]
翻譯
- 語音與生圖邏輯移至
internal/agents/audio與internal/agents/image,兩者的模型、選項、清理與建立 agent 函式結構一致,internal/tools/external只保留工具註冊 - provider 模型清單的快取 key 與失效處理集中到 torii,四個 package 各自的實作合併為一次呼叫;轉錄 prompt 改為嵌入的 prompt 檔
Scope
internal/agents/audio/,internal/agents/image/— FIX, UPDATE, REFACTORinternal/tools/external/— UPDATE, REFACTOR (audio.go,image.go)internal/agents/exec/model/,configs/prompts/system_prompt/— FIX, REFACTORinternal/session/config/— UPDATE (tier.go)internal/agents/keychain/— UPDATEinternal/runtime/tui/,internal/runtime/routes/handler/— FIX, REFACTORinternal/runtime/torii/,internal/agents/decision/,internal/agents/probe/— REFACTORdoc/— DOC
Generated by SKILL