KiboAI · lulu0119 · Jul 27, 2026 · Jul 27, 2026 · Jul 27, 2026 · Jul 27, 2026
diff --git a/.env.example b/.env.example
@@ -4,7 +4,7 @@
 #
 # 启动顺序：
 #   1) pnpm dev:api          # 读 .env，监听 http://localhost:8787（含 WS /stt-realtime）
-#   2) pnpm dev:playground    # Vite 把 /stt /stt-realtime /llm 代理到 8787
+#   2) pnpm dev:playground    # Vite 把 /stt-realtime /llm 代理到 8787
 
 # ── 代理端口 ─────────────────────────────────────────────
 PORT=8787
@@ -39,39 +39,12 @@ PORT=8787
 # PLAYGROUND_API_ORIGIN=http://localhost:8787
 
 # ── 语音转写 (STT) ────────────────────────────────────────
-# 三个 batch provider（可同时配置，playground 从 GET /stt/providers 切换）：
-#   - openrouter ：云端，JSON → /audio/transcriptions
-#   - openai     ：Whisper 形 multipart /v1/audio/transcriptions（本地 mlx-qwen3-asr 等，见 ADR 0002）
-#   - dashscope  ：阿里云百炼 batch ASR（OpenAI 兼容 chat/completions + input_audio）
-#   - dashscope-realtime：同 key + STT_DASHSCOPE_WS_URL → WS /stt-realtime（ADR 0004）
-# STT_<PROVIDER>_MODEL 不填时回退到注册默认值。
-STT_ACTIVE=openrouter
-# 生产 realtime 与 batch 并行；batch provider 单独冻结。
-# STT_BATCH_ACTIVE=dashscope
-STT_OPENROUTER_BASE_URL=https://openrouter.ai/api/v1
-STT_OPENROUTER_API_KEY=sk-or-xxxxxxxxxxxxxxxx
-STT_OPENROUTER_MODEL=openai/gpt-4o-transcribe
-
-# 本地 Qwen3-ASR（Apple Silicon，mlx-qwen3-asr serve 默认监听 8765）：
-#   pip install "mlx-qwen3-asr[serve]"
-#   mlx-qwen3-asr serve --api-key $(openssl rand -hex 16)
-# 仍经 apps/api /stt 代理转发；浏览器发 /stt?provider=openai，key 留在服务端。
-# 需 apps/api 与 mlx-qwen3-asr 同机（本地 dev）。切到本地：STT_ACTIVE=openai，或浏览器加 ?provider=openai。
-# STT_ACTIVE=openai
-# STT_OPENAI_BASE_URL=http://localhost:8765/v1
-# STT_OPENAI_API_KEY=本地 serve 启动时生成的 key
-# STT_OPENAI_MODEL=Qwen3-ASR-1.7B
-
-# 阿里云百炼 Qwen3-ASR batch（将 {WorkspaceId} 换成业务空间 ID）：
-# STT_ACTIVE=dashscope
-# STT_DASHSCOPE_BASE_URL=https://{WorkspaceId}.cn-beijing.maas.aliyuncs.com/compatible-mode/v1
-# STT_DASHSCOPE_API_KEY=sk-xxxxxxxx
-# STT_DASHSCOPE_MODEL=qwen3-asr-flash
-#
-# Realtime（WS /stt-realtime?provider=dashscope-realtime，ADR 0004）：
-# 复用 STT_DASHSCOPE_API_KEY；另需 WS_URL。浏览器只连同源代理，不直连上游。
-# STT_DASHSCOPE_WS_URL=wss://{WorkspaceId}.cn-beijing.maas.aliyuncs.com/api-ws/v1/realtime
-# STT_DASHSCOPE_REALTIME_MODEL=qwen3-asr-flash-realtime
+# Realtime only（WS /stt-realtime?provider=dashscope-realtime，ADR 0004）。
+# 浏览器只连同源代理，不直连上游。
+STT_ACTIVE=dashscope-realtime
+STT_DASHSCOPE_API_KEY=sk-xxxxxxxx
+STT_DASHSCOPE_WS_URL=wss://{WorkspaceId}.ap-northeast-1.maas.aliyuncs.com/api-ws/v1/realtime
+STT_DASHSCOPE_REALTIME_MODEL=qwen3-asr-flash-realtime
 
 # ── 大语言模型 (LLM) ─────────────────────────────────────
 # 走 xsai（OpenAI 兼容）。provider 与 STT 可不同。

diff --git a/.github/workflows/deploy-production.yml b/.github/workflows/deploy-production.yml
@@ -27,10 +27,13 @@ jobs:
           CN_SERVER_HOST: ${{ secrets.CN_SERVER_HOST }}
           CN_SERVER_USER: ${{ secrets.CN_SERVER_USER }}
           CN_RELAY_ORIGIN: ${{ secrets.CN_RELAY_ORIGIN }}
+          JP_STT_DASHSCOPE_API_KEY: ${{ secrets.JP_STT_DASHSCOPE_API_KEY }}
+          JP_STT_DASHSCOPE_WS_URL: ${{ secrets.JP_STT_DASHSCOPE_WS_URL }}
         run: |
           for secret_name in \
             DEPLOY_SSH_KEY SERVER_HOST SERVER_USER \
-            CN_DEPLOY_SSH_KEY CN_SERVER_HOST CN_SERVER_USER CN_RELAY_ORIGIN; do
+            CN_DEPLOY_SSH_KEY CN_SERVER_HOST CN_SERVER_USER CN_RELAY_ORIGIN \
+            JP_STT_DASHSCOPE_API_KEY JP_STT_DASHSCOPE_WS_URL; do
             if [ -z "${!secret_name}" ]; then
               echo "::error::Missing Actions secret: ${secret_name}"
               exit 1
@@ -49,26 +52,44 @@ jobs:
           chmod 600 ~/.ssh/id_ed25519_primary ~/.ssh/id_ed25519_cn
           ssh-keyscan -H "$SERVER_HOST" >> ~/.ssh/known_hosts
           ssh-keyscan -H "$CN_SERVER_HOST" >> ~/.ssh/known_hosts
-      - name: Configure relay origin on primary
+      - name: Configure primary routes and providers
         env:
           SERVER_HOST: ${{ secrets.SERVER_HOST }}
           SERVER_USER: ${{ secrets.SERVER_USER }}
           CN_RELAY_ORIGIN: ${{ secrets.CN_RELAY_ORIGIN }}
+          JP_STT_DASHSCOPE_API_KEY: ${{ secrets.JP_STT_DASHSCOPE_API_KEY }}
+          JP_STT_DASHSCOPE_WS_URL: ${{ secrets.JP_STT_DASHSCOPE_WS_URL }}
         run: |
           case "$CN_RELAY_ORIGIN" in
             http://*|https://*) ;;
             *) echo "::error::CN_RELAY_ORIGIN must be an absolute HTTP(S) origin"; exit 1 ;;
           esac
-          printf '%s\n' "$CN_RELAY_ORIGIN" \
+          case "$JP_STT_DASHSCOPE_WS_URL" in
+            wss://*.ap-northeast-1.maas.aliyuncs.com/api-ws/v1/realtime) ;;
+            *) echo "::error::JP_STT_DASHSCOPE_WS_URL must use the Alibaba Cloud Tokyo realtime endpoint"; exit 1 ;;
+          esac
+          printf '%s\n%s\n%s\n' \
+            "$CN_RELAY_ORIGIN" \
+            "$JP_STT_DASHSCOPE_API_KEY" \
+            "$JP_STT_DASHSCOPE_WS_URL" \
             | ssh -i ~/.ssh/id_ed25519_primary "${SERVER_USER}@${SERVER_HOST}" '
                 set -eu
                 IFS= read -r origin
-                escaped=$(printf "%s" "$origin" | sed "s/[|&]/\\\\&/g")
-                if grep -q "^RELAY_CN_ORIGIN=" /opt/kibotalk/.env; then
-                  sed -i "s|^RELAY_CN_ORIGIN=.*|RELAY_CN_ORIGIN=${escaped}|" /opt/kibotalk/.env
-                else
-                  printf "RELAY_CN_ORIGIN=%s\n" "$origin" >> /opt/kibotalk/.env
-                fi
+                IFS= read -r stt_key
+                IFS= read -r stt_url
+                set_env() {
+                  name=$1
+                  value=$2
+                  escaped=$(printf "%s" "$value" | sed "s/[|&]/\\\\&/g")
+                  if grep -q "^${name}=" /opt/kibotalk/.env; then
+                    sed -i "s|^${name}=.*|${name}=${escaped}|" /opt/kibotalk/.env
+                  else
+                    printf "%s=%s\n" "$name" "$value" >> /opt/kibotalk/.env
+                  fi
+                }
+                set_env RELAY_CN_ORIGIN "$origin"
+                set_env STT_DASHSCOPE_API_KEY "$stt_key"
+                set_env STT_DASHSCOPE_WS_URL "$stt_url"
                 chmod 600 /opt/kibotalk/.env
               '
       - name: Validate server environments
@@ -84,7 +105,8 @@ jobs:
             for name in \
               RELAY_PRIMARY_ORIGIN RELAY_CN_ORIGIN RELAY_CN_ENABLED \
               RELAY_TOKEN_PRIVATE_KEY RELAY_TOKEN_PUBLIC_KEY RELAY_NODE_SECRET \
-              STT_ACTIVE STT_BATCH_ACTIVE LLM_ACTIVE; do
+              STT_ACTIVE STT_DASHSCOPE_API_KEY STT_DASHSCOPE_WS_URL \
+              LLM_ACTIVE; do
               grep -q "^${name}=." /opt/kibotalk/.env
             done
           '
@@ -94,7 +116,7 @@ jobs:
             for name in \
               RELAY_PRIMARY_ORIGIN RELAY_ACCEPT_NEW_SESSIONS \
               RELAY_TOKEN_PUBLIC_KEY RELAY_NODE_SECRET \
-              STT_ACTIVE STT_BATCH_ACTIVE LLM_ACTIVE; do
+              STT_ACTIVE LLM_ACTIVE; do
               grep -q "^${name}=." /opt/kibotalk-relay/.env
             done
             relay_port=$(sed -n "s/^RELAY_HTTP_PORT=//p" /opt/kibotalk-relay/.env)

diff --git a/AGENTS.md b/AGENTS.md
@@ -31,15 +31,14 @@ update the spec — don't silently drift.
 Client orchestration + thin proxy (ADR 0001). The browser runs the pipeline;
 `apps/api` is a stateless Hono proxy that hides provider keys and forwards STT/LLM.
 
-**STT modes** (ADR 0004): **batch** (`POST /stt`) and **realtime**
-(`WS /stt-realtime`) run in parallel—not a migration. Local VAD + speaker
-verification own turn boundaries (`pauseMs`); realtime uses Manual commit
-(no server VAD). Timeline may show partial drafts; formal turns + LLM fire
-only on finalized transcript.
+**STT** (ADR 0004): **realtime only** via `WS /stt-realtime` (DashScope
+`qwen3-asr-flash-realtime`). Local VAD + speaker verification own turn
+boundaries (`pauseMs`); upstream uses Manual commit (no server VAD). Timeline
+may show partial drafts; formal turns + LLM fire only on finalized transcript.
 
 ```
 apps/
-  api/        Hono proxy: /stt, /stt-realtime (WS), /llm (SSE). Keys server-side only.
+  api/        Hono proxy: /stt-realtime (WS), /llm (SSE). Keys server-side only.
   playground/ Vite + React dev panel (Chinese UI) for testing each layer
               (声纹页 covers enrollment + free-speech verify / threshold tuning).
   web/        PWA shell (not yet built).
@@ -49,8 +48,8 @@ packages/
   prompts/    Reply-suggestion prompts (Velin TSX → markdown).
   speaker/    Speaker verification (wavlm-base-plus-sv, WASM + IndexedDB);
               `verify` returns raw `similarity` plus label `confidence`.
-  stt/        Provider-agnostic STT: batch adapters + DashScope realtime mapper
-              helpers (server-side). Providers declare mode batch|realtime.
+  stt/        DashScope realtime mapper helpers (server-side); `apps/api` relays
+              `WS /stt-realtime` only.
   pipeline/   Conversation store + turn state machine.
   ui/         shadcn/ui primitives on Tailwind v4 (shared).
   app-shared/ Shared client types/config shell.
@@ -62,29 +61,28 @@ These are spec-named choices. **Do not rewrite or substitute them** with hand-ro
 
 - **LLM → `xsai`** (`@xsai/stream-text`). Never hand-roll fetch/SSE for LLM. An official `xsai` skill is installed at `.agents/skills/xsai/` — read its `SKILL.md` + `references/` before touching `packages/llm`.
 - **Prompts → Velin** (`@velin-dev/core-react`). Render TSX components to markdown strings; don't template with raw string concat.
-- **STT → provider-agnostic factory** in `packages/stt`, reached via the `apps/api` `/stt` proxy. Local ASR (`mlx-qwen3-asr`) also goes through the proxy (`?provider=openai`), **never browser-direct** (ADR 0002).
+- **STT → DashScope realtime** via `packages/stt` mapper helpers, reached through the `apps/api` `WS /stt-realtime` relay. **Never browser-direct** to upstream (ADR 0004).
 - **VAD → Silero** via `@huggingface/transformers`. v6.2 needs a 64-sample context prepended to each 512-sample chunk (576 input), carried across calls; v5 takes 512 raw. See `apps/playground/src/audio/silero-vad.ts`.
 - **Speaker → `Xenova/wavlm-base-plus-sv`** via transformers.js in a Web Worker; embeddings persist in IndexedDB.
 
 ## Audio pipeline specifics
 
-- **Sample rate**: 16 kHz mono PCM everywhere (VAD, STT uploads, speaker embeddings).
+- **Sample rate**: 16 kHz mono PCM everywhere (VAD, realtime STT uplink, speaker embeddings).
 - **VAD chunk size**: 512 samples (32 ms) per `processAudio` call (`packages/audio/src/vad.ts` `newBufferSize`). Silero v6.2 prepends 64-sample context → 576 input; v5 takes 512 raw. See `docs/solutions/silero-vad-v6-context-frame.md`.
-- **STT upload format**: WAV 16 kHz mono PCM via `encodeWav` (`packages/audio/src`). The `/stt` proxy forwards the WAV body as-is.
 - **Speaker embeddings**: computed in a Web Worker (`apps/playground/src/audio/speaker-worker.ts`), persisted in IndexedDB. Don't run the WASM model on the main thread.
-- **Segment aggregation / TurnGate** (`packages/audio/src/aggregator.ts`): sits between VAD (+ speaker verification) and batch `ingestSegment` or realtime `append`/`commit`. Accumulates same-speaker speech and flushes on `pauseMs`, `maxMs` (speech only), or speaker change. PCM is direct-concatenated (no gap fill). Spec §2.4 / ADR 0004. Pipeline fires LLM per finalized turn and does NOT wait on pause itself.
+- **Segment aggregation / TurnGate** (`packages/audio/src/aggregator.ts`): sits between VAD (+ speaker verification) and realtime `append`/`commit` → `ingestFinalizedTurn`. Merges same-speaker fragment transcripts and flushes on `pauseMs`, `maxMs` (speech only), or speaker change. Spec §2.4 / ADR 0004. Pipeline fires LLM per finalized turn and does NOT wait on pause itself.
 
 ## Conventions
 
-- **Keys in env, never client.** All provider keys/config live in `.env` (see `.env.example`), loaded by `apps/api`. Naming: `<SCOPE>_<PROVIDER>_<FIELD>` (e.g. `LLM_OPENROUTER_API_KEY`, `STT_OPENAI_BASE_URL`).
+- **Keys in env, never client.** All provider keys/config live in `.env` (see `.env.example`), loaded by `apps/api`. Naming: `<SCOPE>_<PROVIDER>_<FIELD>` (e.g. `LLM_OPENROUTER_API_KEY`, `STT_DASHSCOPE_API_KEY`).
 - **Pure functions; no new classes** unless the framework/API requires it. Imports at top. TS unions: exhaustive `switch`.
 - **Full words.** No obscure abbreviations; only use ones common in software.
 - **Smallest correct diff.** Only change what was asked. No drive-by refactor, tests, or docs unless asked. When you touch code, small progressive refactors alongside the change are welcome.
 - **Reuse and extend** existing functions/modules; do not duplicate similar logic. Before implementing a feature (a selector, a hook, a helper, a card), grep the repo for it first — chances are someone already wrote it. The moment you notice a second copy of something, extract it into a shared module/component and have both callers use it. Examples already in the repo: `SttProviderSelect` + `useSttProviders` + `sttUrl` (`apps/playground/src/SttProviderSelect.tsx`, used by both the VAD panel and the direct-API panel), `padBuffer` / `encodeWav` (`packages/audio`), `createSegmentAggregator` (`packages/audio/aggregator`).
 - **No backward-compatibility guards.** If a rename/breakage is needed, do it directly and update callers in the same change.
 - **Playground UI is Chinese** — labels, examples, and sample content in Chinese.
 - **Tailwind v4 + shadcn/ui** for all UI (playground included). Shared primitives in `packages/ui`（Button、Card、Badge、Input、Textarea、Label、Tabs、Separator、Accordion、Collapsible、Dialog、DropdownMenu、Popover、Progress、ScrollArea、Select、Sheet、Skeleton、Slider、Switch、Tooltip、Toaster）。缺组件先 `shadcn add` 进该包并导出；**改样式改源组件 / token，不要平行重造**。
-- **Shared playground config lives in one Zustand store** (`apps/playground/src/config-store.ts`, `useConfig`) — the React analog of a Pinia store. VAD/ASR/merge/speaker knobs, language prefs (`uiLang` / `conversationLang` / `level` / `languagesConfirmed`, persisted; `meaningLang` is derived from `uiLang` in the session snapshot), and selectors (provider, VAD model, transcribe mode) are shared across the VAD panel and the live session: change one on a tab and it's already aligned on the other. Subscribe per-field (`useConfig(s => s.field)`); in async callbacks read `useConfig.getState()`. Stage-grouped field components live in `apps/playground/src/components/ConfigFields.tsx` (`VadParamsFields`, `AsrPadFields`, `MergeParamsFields`, `LanguagePrefsFields`, `VadModelSelect`, `TranscribeModeSelect`, `TranscribeProviderSelect`, `NumberField`) — reuse these instead of re-declaring the same knobs.
+- **Shared playground config lives in one Zustand store** (`apps/playground/src/config-store.ts`, `useConfig`) — the React analog of a Pinia store. VAD/merge/speaker knobs, language prefs (`uiLang` / `conversationLang` / `level` / `languagesConfirmed`, persisted; `meaningLang` is derived from `uiLang` in the session snapshot), and selectors (VAD model, STT provider) are shared across the VAD panel and the live session: change one on a tab and it's already aligned on the other. Subscribe per-field (`useConfig(s => s.field)`); in async callbacks read `useConfig.getState()`. Stage-grouped field components live in `apps/playground/src/components/ConfigFields.tsx` (`VadParamsFields`, `MergeParamsFields`, `LanguagePrefsFields`, `VadModelSelect`, `TranscribeProviderSelect`, `NumberField`) — reuse these instead of re-declaring the same knobs.
 
 ## Commands