feat(dashboard): live STT/LLM model switching (small<->medium, Haiku<->Sonnet)

Adds a "🧠 모델 (STT · LLM)" collapsible panel with two dropdowns:
- STT(귀): tiny/base/small/medium/large-v3. Switching swaps WhisperSTT.model,
  tears the worker down and re-warms it in the BACKGROUND (first switch to a
  not-yet-downloaded size fetches it, so the HTTP call returns immediately and
  the next utterance waits for the reload).
- LLM(두뇌): Haiku 4.5 / Sonnet 4.5. Applied on the next reply (no reload).

Backend: GET /api/models, POST /api/models/stt, POST /api/models/llm; Dashboard
gains models_settings/set_stt_model/set_llm_model. In-memory only (a service
restart reverts to the env defaults small / claude-haiku-4-5).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

wsai/dashboard.py

@@ -87,6 +87,8 @@ def _make_handler(dash: "Dashboard"):
                 self._handle_tts_settings_get()
             elif path == "/api/bot/settings":
                 self._send_json({"ok": True, "settings": dash.bot.get_settings()})
+            elif path == "/api/models":
+                self._send_json(dash.models_settings())
             elif path == "/events":
                 self._stream_events()
             else:
@@ -119,6 +121,10 @@ def _make_handler(dash: "Dashboard"):
                 self._handle_tts_preview()
             elif path == "/api/bot/settings":
                 self._handle_bot_settings_post()
+            elif path == "/api/models/stt":
+                self._handle_model_switch("stt")
+            elif path == "/api/models/llm":
+                self._handle_model_switch("llm")
             else:
                 self._send(404, b"not found", "text/plain; charset=utf-8")
 
@@ -288,6 +294,25 @@ def _make_handler(dash: "Dashboard"):
             monitor.log("info", f"청취 화이트/블랙리스트 업데이트 (guild={guild_id})")
             self._send_json({"ok": True, "guildId": guild_id, "lists": saved})
 
+        def _handle_model_switch(self, which: str) -> None:
+            """Switch the STT size or the LLM model live."""
+            raw = self._read_body()
+            try:
+                data = json.loads(raw.decode("utf-8")) if raw else {}
+                model = data.get("model")
+            except (ValueError, AttributeError):
+                self._send_json({"ok": False, "error": "invalid JSON"}, 400)
+                return
+            try:
+                if which == "stt":
+                    res = dash.set_stt_model(model)
+                else:
+                    res = dash.set_llm_model(model)
+            except Exception as exc:  # noqa: BLE001 — surface the reason to the page
+                self._send_json({"ok": False, "error": f"{type(exc).__name__}: {exc}"}, 400)
+                return
+            self._send_json({"ok": True, which: res})
+
         def _handle_bot_settings_post(self) -> None:
             """Save a bot behaviour toggle (e.g. bargeIn). The bot reads the new
             value on its next report round-trip."""
@@ -548,6 +573,68 @@ class Dashboard:
             pass
         return wav
 
+    # -- live model switching (STT size / LLM model) --------------------- #
+    STT_OPTIONS = ["tiny", "base", "small", "medium", "large-v3"]
+    LLM_OPTIONS = ["claude-haiku-4-5", "claude-sonnet-4-5"]
+
+    def models_settings(self) -> dict:
+        stt = self.stt
+        brain = self.brain
+        return {
+            "ok": True,
+            "stt": {
+                "enabled": stt is not None,
+                "current": getattr(stt, "model", None),
+                "ready": bool(getattr(stt, "_ready", False)),
+                "device": getattr(stt, "resolved_device", None),
+                "options": self.STT_OPTIONS,
+            },
+            "llm": {
+                "enabled": brain is not None,
+                "current": getattr(brain, "model", None),
+                "options": self.LLM_OPTIONS,
+            },
+        }
+
+    def set_stt_model(self, model: str) -> dict:
+        """Switch the whisper model size live. Tears down the current worker and
+        warms the new one in the BACKGROUND so the HTTP call returns fast (the
+        first switch to a not-yet-downloaded size fetches it, which can take a
+        while); the next utterance waits for the reload if it isn't warm yet."""
+        import asyncio
+        if self.stt is None:
+            raise RuntimeError("STT not enabled")
+        model = str(model).strip()
+        if not model:
+            raise ValueError("model required")
+        if model not in self.STT_OPTIONS:
+            raise ValueError(f"unknown STT model: {model}")
+        if model != self.stt.model:
+            self.stt.model = model
+            self._submit(self.stt.aclose())  # drop old worker (fast)
+            # Reload+warm in the background; don't block the HTTP response.
+            asyncio.run_coroutine_threadsafe(self._warm_stt_bg(), self._loop)
+        return self.models_settings()["stt"]
+
+    async def _warm_stt_bg(self) -> None:
+        try:
+            await self.stt.warmup()
+            self.monitor.log("info", f"STT 모델 로드 완료: {self.stt.model} "
+                             f"(device={getattr(self.stt, 'resolved_device', '?')})")
+        except Exception as exc:  # noqa: BLE001
+            self.monitor.log("error", f"STT 모델 로드 실패({self.stt.model}): {exc}")
+
+    def set_llm_model(self, model: str) -> dict:
+        """Switch the Claude model live — applied on the next reply (no reload)."""
+        if self.brain is None:
+            raise RuntimeError("LLM(brain) not enabled — echo 모드입니다")
+        model = str(model).strip()
+        if model not in self.LLM_OPTIONS:
+            raise ValueError(f"unknown LLM model: {model}")
+        self.brain.model = model
+        self.monitor.log("info", f"LLM 모델 변경: {model} (다음 답변부터 적용)")
+        return self.models_settings()["llm"]
+
     def voice_turn(self, audio_bytes: bytes, speaker: str = "",
                    guild: str = "", channel: str = "") -> dict:
         """One Discord voice turn: decode the uploaded utterance, recognise it
@@ -944,6 +1031,24 @@ PAGE = r"""<!DOCTYPE html>
         <span>잡음 로그 표시 안 함 <b class="cfghint">(들음: (빈 결과) 또는 답변: [잡음]) · 기본: 켜짐</b></span></label>
     </div>
   </section>
+  <section class="sttbox" id="modelbox">
+    <h2 id="modelToggle" class="collap"><span id="modelCaret">▸</span> 🧠 모델 (STT · LLM)</h2>
+    <div id="modelBody" style="display:none">
+      <div class="sttrow">
+        <label style="font-size:12.5px;color:var(--muted)">STT(귀)
+          <select id="mSTT" class="ttssel"></select></label>
+        <button id="mSTTapply" class="btn">적용</button>
+        <span id="mSTTstat" class="sttstat"></span>
+      </div>
+      <div class="sttrow" style="margin-top:8px">
+        <label style="font-size:12.5px;color:var(--muted)">LLM(두뇌)
+          <select id="mLLM" class="ttssel"></select></label>
+        <button id="mLLMapply" class="btn">적용</button>
+        <span id="mLLMstat" class="sttstat"></span>
+      </div>
+      <p class="cfghint" style="margin:8px 0 0">STT는 전환 시 모델을 다시 로드합니다(처음 medium/large-v3는 다운로드로 수 분 걸릴 수 있어요). LLM은 다음 답변부터 즉시 적용됩니다. 서비스 재시작 시 기본값(small · Haiku)으로 돌아갑니다.</p>
+    </div>
+  </section>
   <div class="demobar" id="demobar" style="display:none"></div>
   <div class="comp" id="comp"></div>
   <div class="tfilter" id="tfilter">
@@ -1213,7 +1318,46 @@ function wireCollapse(toggleId, bodyId, caretId){
     localStorage.setItem('wsai_hideNoise', hideNoise ? '1':'0');
     applyTurnFilter();
   };
+  // 모델 설정: STT 크기 / LLM 모델 실시간 전환.
+  wireCollapse('modelToggle','modelBody','modelCaret');
+  initModels();
 })();
+const STT_LABEL = {tiny:'tiny (가장 빠름)', base:'base', small:'small (기본)',
+  medium:'medium (정확도↑)', 'large-v3':'large-v3 (최고 정확도)'};
+const LLM_LABEL = {'claude-haiku-4-5':'Haiku 4.5 (가장 빠름)',
+  'claude-sonnet-4-5':'Sonnet 4.5 (고품질·조금 느림)'};
+function fillSel(sel, options, current, labels){
+  sel.innerHTML='';
+  options.forEach(o=>{ const el=document.createElement('option');
+    el.value=o; el.textContent=(labels[o]||o); sel.appendChild(el); });
+  if(current) sel.value=current;
+}
+async function initModels(){
+  let j; try{ j=await (await fetch('/api/models')).json(); }catch(e){ return; }
+  if(!j||!j.ok) return;
+  if(j.stt.enabled){ fillSel($('mSTT'), j.stt.options, j.stt.current, STT_LABEL);
+    $('mSTTstat').textContent='현재: '+(j.stt.current||'?')+(j.stt.device?(' · '+j.stt.device):''); }
+  else { $('mSTT').disabled=$('mSTTapply').disabled=true; $('mSTTstat').textContent='STT 비활성'; }
+  if(j.llm.enabled){ fillSel($('mLLM'), j.llm.options, j.llm.current, LLM_LABEL);
+    $('mLLMstat').textContent='현재: '+(LLM_LABEL[j.llm.current]||j.llm.current||'?'); }
+  else { $('mLLM').disabled=$('mLLMapply').disabled=true; $('mLLMstat').textContent='LLM 비활성(echo 모드)'; }
+  $('mSTTapply').onclick = async ()=>{
+    const m=$('mSTT').value; $('mSTTstat').textContent='전환 중… (첫 다운로드면 수 분 걸릴 수 있어요)';
+    try{ const r=await fetch('/api/models/stt',{method:'POST',headers:{'Content-Type':'application/json'},
+        body:JSON.stringify({model:m})}); const jj=await r.json();
+      if(jj.ok){ toast('STT 모델 전환: '+m); $('mSTTstat').textContent='전환됨: '+m+' · 로딩은 백그라운드로 진행됩니다'; }
+      else { $('mSTTstat').textContent='실패: '+(jj.error||''); }
+    }catch(e){ $('mSTTstat').textContent='오류: '+e; }
+  };
+  $('mLLMapply').onclick = async ()=>{
+    const m=$('mLLM').value;
+    try{ const r=await fetch('/api/models/llm',{method:'POST',headers:{'Content-Type':'application/json'},
+        body:JSON.stringify({model:m})}); const jj=await r.json();
+      if(jj.ok){ toast('LLM 모델 변경: '+(LLM_LABEL[m]||m)); $('mLLMstat').textContent='현재: '+(LLM_LABEL[m]||m)+' · 다음 답변부터'; }
+      else { $('mLLMstat').textContent='실패: '+(jj.error||''); }
+    }catch(e){ $('mLLMstat').textContent='오류: '+e; }
+  };
+}
 
 // --- 봇 목소리(TTS) 조절: 감정별 슬라이더 → 미리듣기 → 봇 적용 -------------- #
 let ttsInited = false;