feat(brain): voice bot answers in 존댓말 (polite Korean)

User requested the voice assistant speak politely. Update the persona to
require 존댓말 on every reply and rewrite the two 반말 output examples
(emotion demo + clarify-question) accordingly. Verified live: replies come
back polite and still one short sentence with emotion tags.

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

dave/bot.mjs

@@ -178,6 +178,16 @@ const speakingSet = new Set();   // userIds currently speaking (for participant
 const activeSubs = new Set();    // userIds with an in-flight receive subscription
 let listsByGuild = {};           // guildId -> {whitelistUsers, blacklistUsers, whitelistRoles, blacklistRoles}
 let botSettings = { bargeIn: true };  // behaviour toggles from the dashboard
+// Barge-in only fires when the Discord "speaking" (green ring) stays on for at
+// least this long — a brief blip (keyboard click, cough) shouldn't cut the bot
+// off, but a genuine ~0.7s of speech should. The dashboard's "유저 음성 인식 시간"
+// (botSettings.bargeInMs) wins when set; WSAI_BARGE_IN_MS / 700 is the fallback.
+const BARGE_IN_MS = Number(process.env.WSAI_BARGE_IN_MS || 700);
+function bargeInMs() {
+  const v = Number(botSettings.bargeInMs);
+  return Number.isFinite(v) && v >= 0 ? v : BARGE_IN_MS;
+}
+const bargeTimers = new Map();  // userId -> pending stop timer
 
 // Attach the bot's audio player (so it can speak) to a fresh connection.
 function setupPlayer(connection) {
@@ -202,15 +212,20 @@ function setupReceiver(connection) {
         return;
       }
     }
-    // Barge-in: the moment an allowed user speaks, stop the bot's current TTS
+    // Barge-in: stop the bot's current TTS only if this user keeps speaking for
-    // so it doesn't talk over them (toggleable from the dashboard, default on).
+    // BARGE_IN_MS (the green ring stays on) — ignores momentary noise blips. The
-    if (botSettings.bargeIn !== false && voicePlayer) {
+    // 'end' handler cancels the pending stop if speaking stops in time.
+    if (botSettings.bargeIn !== false && voicePlayer && !bargeTimers.has(userId)) {
+      bargeTimers.set(userId, setTimeout(() => {
+        bargeTimers.delete(userId);
         try { voicePlayer.stop(true); } catch {}
+      }, bargeInMs()));
     }
     activeSubs.add(userId);
     if (!perUser.has(userId)) perUser.set(userId, { opusPackets: 0, pcmFrames: 0 });
     const opusStream = receiver.subscribe(userId, {
-      end: { behavior: EndBehaviorType.AfterSilence, duration: 800 },
+      // Wait a touch longer after silence so a soft/trailing word isn't clipped.
+      end: { behavior: EndBehaviorType.AfterSilence, duration: 1000 },
     });
     const decoder = new prism.opus.Decoder({ rate: 48000, channels: 2, frameSize: 960 });
     const chunks = [];
@@ -226,7 +241,11 @@ function setupReceiver(connection) {
       handleUtterance(userId, pcm).catch((e) => log(`voice turn error: ${e.message}`));
     });
   });
-  receiver.speaking.on('end', (userId) => speakingSet.delete(userId));
+  receiver.speaking.on('end', (userId) => {
+    speakingSet.delete(userId);
+    const t = bargeTimers.get(userId);   // spoke too briefly -> cancel barge-in
+    if (t) { clearTimeout(t); bargeTimers.delete(userId); }
+  });
 }
 
 // Join (or switch to) a voice channel on command from the dashboard.

tests/latency_ab.py

@@ -0,0 +1,131 @@
+"""Ad-hoc TTFT/total latency A/B across Sonnet versions (+ Haiku reference).
+
+Mirrors the production brain call: Claude Code identity system block, cached
+persona, a short voice-style history, one short user turn. Streams to measure
+time-to-first-token. Interleaves models each round to cancel network drift.
+
+Run: .venv/bin/python tests/latency_ab.py
+"""
+from __future__ import annotations
+
+import json
+import os
+import statistics
+import time
+
+import anthropic
+
+from wsai.backends.claude import ClaudeBrain
+
+CRED = os.environ.get("CLAUDE_CREDENTIALS_PATH")
+if not CRED:
+    raise SystemExit(
+        "Set CLAUDE_CREDENTIALS_PATH to your Claude credentials JSON to run this A/B script."
+    )
+CLAUDE_CODE_ID = "You are Claude Code, Anthropic's official CLI for Claude."
+
+# (label, model, extra_system_line) — extra line appended to persona to force brevity
+BREVITY = "지금부터 답은 무조건 한 문장, 12단어 이내로만. 부연·재확인·군더더기 금지."
+VARIANTS = [
+    ("sonnet-4-5", "claude-sonnet-4-5", None),
+    ("sonnet-5+brev", "claude-sonnet-5", BREVITY),
+    ("haiku-4-5(ref)", "claude-haiku-4-5", None),
+]
+ROUNDS = 16
+
+HISTORY = [
+    ("안녕", "[반가움] 안녕! 뭐 하고 있었어?"),
+    ("그냥 코딩", "[다정] 오 무슨 코딩?"),
+    ("파이썬", "[신남] 좋네, 잘 되고 있어?"),
+    ("응 그럭저럭", "[차분] 다행이다."),
+]
+USER = "지금 몇 시야?"
+
+
+def token() -> str:
+    with open(CRED) as f:
+        return json.load(f)["claudeAiOauth"]["accessToken"]
+
+
+def build(extra: str | None = None):
+    persona = ClaudeBrain.PERSONA
+    if extra:
+        persona = persona + "\n\n" + extra
+    system = [
+        {"type": "text", "text": CLAUDE_CODE_ID},
+        {"type": "text", "text": persona, "cache_control": {"type": "ephemeral"}},
+    ]
+    msgs = []
+    for u, a in HISTORY:
+        msgs.append({"role": "user", "content": u})
+        msgs.append({"role": "assistant", "content": a})
+    msgs.append({"role": "user", "content": "[지금 화면] (아직 못 읽음)\n\n" + USER})
+    return system, msgs
+
+
+def measure(client, model, system, msgs):
+    t0 = time.monotonic()
+    ttft = None
+    out_tokens = 0
+    try:
+        with client.messages.stream(
+            model=model, max_tokens=150, system=system, messages=msgs
+        ) as stream:
+            for ev in stream.text_stream:
+                if ttft is None:
+                    ttft = time.monotonic() - t0
+            total = time.monotonic() - t0
+            final = stream.get_final_message()
+            out_tokens = final.usage.output_tokens
+        return ttft, total, out_tokens, None
+    except Exception as e:  # noqa: BLE001
+        return None, None, None, f"{type(e).__name__}: {str(e)[:120]}"
+
+
+def main():
+    client = anthropic.Anthropic(auth_token=token(), max_retries=1)
+    prebuilt = {label: build(extra) for label, model, extra in VARIANTS}
+    results = {label: {"ttft": [], "total": [], "out": []} for label, _, _ in VARIANTS}
+    dead = set()
+    # one warm-up per variant to prime connection + cache (excluded from stats)
+    for label, model, _ in VARIANTS:
+        system, msgs = prebuilt[label]
+        _, _, _, err = measure(client, model, system, msgs)
+        if err:
+            print(f"[skip] {label} ({model}): {err}")
+            dead.add(label)
+    print(f"\nliving: {[l for l, _, _ in VARIANTS if l not in dead]}")
+    print(f"rounds: {ROUNDS} (interleaved)\n")
+    for r in range(ROUNDS):
+        for label, model, _ in VARIANTS:
+            if label in dead:
+                continue
+            system, msgs = prebuilt[label]
+            ttft, total, out, err = measure(client, model, system, msgs)
+            if err:
+                print(f"  r{r} {label}: ERR {err}")
+                continue
+            results[label]["ttft"].append(ttft)
+            results[label]["total"].append(total)
+            results[label]["out"].append(out)
+        print(f"round {r+1}/{ROUNDS} done")
+
+    print("\n=== median (min–max) over", ROUNDS, "runs ===")
+    print(f"{'variant':<18} {'TTFT s':<18} {'total s':<18} {'out tok'}")
+    for label, _, _ in VARIANTS:
+        d = results[label]
+        if not d["ttft"]:
+            print(f"{label:<18} (no data)")
+            continue
+        tt = d["ttft"]
+        to = d["total"]
+        print(
+            f"{label:<18} "
+            f"{statistics.median(tt):.2f} ({min(tt):.2f}-{max(tt):.2f})   "
+            f"{statistics.median(to):.2f} ({min(to):.2f}-{max(to):.2f})   "
+            f"{statistics.median(d['out']):.0f}"
+        )
+
+
+if __name__ == "__main__":
+    main()

tests/test_monitor.py

@@ -37,7 +37,7 @@ def test_monitor_records_turn_with_timed_steps():
     assert turn["total_ms"] >= 0
     # step-by-step: every stage is named and timed
     names = [s["name"] for s in turn["steps"]]
-    assert names == ["화면 맥락", "두뇌(생각)", "응답(TTS/전송)"]
+    assert names == ["화면 맥락", "LLM(생각)", "응답(TTS/전송)"]
     assert all(s["ok"] is True for s in turn["steps"])
     assert all(s["ms"] >= 0 for s in turn["steps"])
 
@@ -66,7 +66,7 @@ def test_monitor_marks_errors():
 
     turn = snap["turns"][0]
     assert turn["status"] == "error"
-    brain_step = next(s for s in turn["steps"] if s["name"] == "두뇌(생각)")
+    brain_step = next(s for s in turn["steps"] if s["name"] == "LLM(생각)")
     assert brain_step["ok"] is False
     assert "boom" in brain_step["error"]
     assert snap["status"]["errors_total"] >= 1

wsai/__main__.py

@@ -82,11 +82,11 @@ def _run_stt_test(host: str, port: int) -> None:
     monitor.set_components({"source": "none", "vision": "none", "stt": "whisper",
                             "brain": "none", "tts": "none"})
     monitor.set_status(running=True, listening=False)
-    monitor.log("info", "STT 인식 테스트 서버 시작 — GPU 워밍업 중…")
+    monitor.log("info", "STT 인식 테스트 서버 시작 — GPU 워밍업 중…", cat="READY")
     print("\n  STT 워밍업 중… (모델 로드 + CUDA 예열)")
     dash.warm()  # load + warm the GPU worker so the first recognition is instant
     dev = getattr(stt, "resolved_device", None) or "?"
-    monitor.log("info", f"STT 준비 완료 (device={dev}). 녹음/파일 업로드로 인식하세요.")
+    monitor.log("info", f"STT 준비 완료 (device={dev}). 녹음/파일 업로드로 인식하세요.", cat="READY")
 
     shown = host if host not in ("0.0.0.0", "") else _lan_ip()
     print(f"\n  음성 인식 테스트 사이트:  http://{shown}:{port}   (STT device: {dev})")
@@ -100,6 +100,28 @@ def _run_stt_test(host: str, port: int) -> None:
         dash.stop()
 
 
+def _apply_persisted_state(stt, tts, brain) -> None:
+    """Apply dashboard settings saved to the state store (STT/LLM model + TTS
+    controls) onto freshly-built backends, so they survive a restart."""
+    from . import state_store
+    st = state_store.load()
+    models = st.get("models") or {}
+    if models.get("stt") and stt is not None:
+        stt.model = str(models["stt"])
+    if models.get("llm") and brain is not None:
+        brain.model = str(models["llm"])
+    tj = st.get("tts") or {}
+    base = tj.get("base") or {}
+    for k in ("speed", "word_gap", "sentence_gap", "pitch"):
+        if base.get(k) is not None and tts is not None:
+            try:
+                setattr(tts, k, float(base[k]))
+            except (TypeError, ValueError):
+                pass
+    if tts is not None and isinstance(tj.get("overrides"), dict):
+        tts.emotion_overrides = {k: dict(v) for k, v in tj["overrides"].items() if isinstance(v, dict)}
+
+
 def _run_voice_server(host: str, port: int) -> None:
     """Serve the STT+TTS voice-turn endpoint that the Discord bot (dave/bot.mjs)
     calls: it POSTs a captured utterance wav and gets back the reply wav to play
@@ -120,7 +142,7 @@ def _run_voice_server(host: str, port: int) -> None:
     if os.environ.get("WSAI_BRAIN", "claude").lower() not in ("none", "echo"):
         try:
             from .backends.claude import ClaudeBrain
-            model = os.environ.get("WSAI_BRAIN_MODEL", "claude-sonnet-4-5")
+            model = os.environ.get("WSAI_BRAIN_MODEL", "claude-sonnet-5")
             brain = ClaudeBrain(model=model)
             brain_name = "claude"
         except Exception as exc:  # noqa: BLE001
@@ -129,20 +151,24 @@ def _run_voice_server(host: str, port: int) -> None:
     monitor = Monitor()
     stt = WhisperSTT()
     tts = MeloTTS()
+    # Restore persisted dashboard settings (models + TTS controls) BEFORE warmup
+    # so the worker loads the last-chosen model. Bot lists/toggles are restored
+    # inside BotControl. Applied before dash.warm() so nothing reloads twice.
+    _apply_persisted_state(stt, tts, brain)
     dash = Dashboard(monitor, host=host, port=port, stt=stt, tts=tts, brain=brain)
     dash.start()
     monitor.set_components({"source": "none", "vision": "none", "stt": "whisper",
                             "brain": brain_name, "tts": "melo"})
     monitor.set_status(running=True, listening=False)
-    monitor.log("info", "디스코드 음성 서버 시작 — STT+TTS GPU 워밍업 중…")
+    monitor.log("info", "디스코드 음성 서버 시작 — STT+TTS GPU 워밍업 중…", cat="READY")
     print("\n  STT+TTS 워밍업 중… (모델 로드 + CUDA 예열)")
     dash.warm()
     sdev = getattr(stt, "resolved_device", None) or "?"
     monitor.set_status(listening=True)
-    monitor.log("info", f"음성 서버 준비 완료 (STT device={sdev}). 디스코드 봇 연결 대기.")
+    monitor.log("info", f"음성 서버 준비 완료 (STT device={sdev}). 디스코드 봇 연결 대기.", cat="READY")
 
     shown = host if host not in ("0.0.0.0", "") else _lan_ip()
-    print(f"\n  음성 서버 준비 완료 (STT device: {sdev}, 두뇌: {brain_name})")
+    print(f"\n  음성 서버 준비 완료 (STT device: {sdev}, LLM: {brain_name})")
     print(f"  대시보드/상태:   http://{shown}:{port}")
     print(f"  봇 연결 엔드포인트: http://127.0.0.1:{port}/api/voice-turn\n")
     try:

wsai/backends/claude.py

@@ -37,6 +37,10 @@ _CLAUDE_CODE_ID = "You are Claude Code, Anthropic's official CLI for Claude."
 # still fails fast rather than leaving the bot silent for many seconds.
 _MAX_RETRIES = int(os.environ.get("WSAI_BRAIN_MAX_RETRIES", "4"))
 
+# Cap the reply length. Voice replies must be short (1 sentence), and a smaller
+# cap also means fewer tokens to generate -> lower latency. Tunable via env.
+_MAX_TOKENS = int(os.environ.get("WSAI_BRAIN_MAX_TOKENS", "150"))
+
 
 def _load_oauth_token() -> str | None:
     path = os.environ.get("CLAUDE_CREDENTIALS_PATH")
@@ -123,23 +127,36 @@ class ClaudeVision:
 
 
 class ClaudeBrain:
+    # Injected as an always-on, cached system block on top of the (editable)
+    # persona. Sonnet 5 answers correctly but more verbosely than 4.5 for the
+    # same voice prompt (measured 46-54 vs 28 output tokens), which erased its
+    # ~0.4s time-to-first-token advantage in total turn time. This hard brevity
+    # rule pulls Sonnet 5 back to ~30 tokens, so the faster first token actually
+    # translates into a faster (and lower-variance) whole reply. Kept separate
+    # from PERSONA so a dashboard persona edit can never drop it.
+    BREVITY = "지금부터 답은 무조건 한 문장, 12단어 이내로만. 부연·재확인·군더더기 금지."
+
     PERSONA = (
         "너는 디스코드를 이용해 사용자와 실시간으로 대화하는 AI 인공지능이야.\n\n"
         "1. 역할\n"
         "- 사용자의 말을 듣고 자연스럽게 대답한다.\n"
-        "- 음성 대화에 어울리게 짧고 빠르게 반응한다.\n"
+        "- 음성 대화에 어울리게 아주 짧고 빠르게 반응한다.\n"
-        "- 친구처럼 편하게, 무례하거나 과하게 장난치진 않는다.\n\n"
+        "- 다정하고 친근하게 대하되 항상 존댓말로 답한다. 무례하거나 과하게 장난치진 않는다.\n\n"
         "2. 언어\n"
         "- 음성 출력은 무조건 한국어다. 어떤 경우에도 한국어로만 답한다.\n"
+        "- 항상 존댓말로 답한다. 어떤 경우에도 반말을 쓰지 않는다.\n"
         "- 사용자가 다른 언어로 말하거나 \"영어로 해줘\"처럼 다른 언어를 요청해도 한국어로 답한다.\n"
         "- 다른 언어를 요청받으면 한국어로 짧게 그렇게는 못 한다고 말한다.\n\n"
-        "3. 답변 방식\n"
+        "3. 답변 길이 (가장 중요)\n"
+        "- 기본은 딱 한 문장. 정말 필요할 때만 최대 두 문장. 절대 길게 말하지 않는다.\n"
+        "- 인사엔 인사만 짧게 답한다. 부르면 짧게 대답만 한다.\n"
+        "- 자기소개나 \"무엇을 도와드릴까요\", \"필요한 거 있으면 말해\" 같은 상투적인 말을 덧붙이지 않는다.\n"
+        "- 질문엔 군더더기 없이 핵심 답만 바로 말한다.\n"
         "- 음성으로 읽히니 마크다운·코드블록·특수기호·목록기호·이모지 없이 평범한 말로만 답한다.\n"
-        "- 기본은 한두 문장, 길어도 10초 안팎. 길어질 땐 핵심부터 말하고 필요하면 이어서 설명한다.\n"
         "- URL·긴 숫자·시간·단위·코드는 소리내 읽기 좋게 풀어서 말한다.\n\n"
         "4. 대화 태도\n"
         "- 사용자의 말투·분위기에 맞춰 반응한다.\n"
-        "- 모르면 지어내지 말고 모른다고 하고, 애매하면 되묻는다(\"다시 말해줄래?\").\n"
+        "- 모르면 지어내지 말고 모른다고 하고, 애매하면 되묻는다(\"다시 말씀해 주시겠어요?\").\n"
         "- 잡음·침묵·의미 없는 소리엔 억지로 대답하지 않는다.\n\n"
         "5. 안전·사실성\n"
         "- 위험하거나 불법적인 요청은 돕지 않는다.\n"
@@ -148,7 +165,7 @@ class ClaudeBrain:
         "6. 감정 표현\n"
         "- 감정은 대괄호 태그로 표현한다. 태그 자체는 읽히지 않고 뒤 문장의 목소리 톤(피치·속도)만 바뀐다.\n"
         "- 답변 맨 앞에 감정 태그 하나로 시작하고, 도중에 감정이 바뀌면 그 지점에 새 태그를 넣는다.\n"
-        "  예: [속상함] 정말 힘들었겠다. [힘차게] 하지만 넌 할 수 있어!\n"
+        "  예: [속상함] 정말 힘드셨겠어요. [힘차게] 하지만 하실 수 있어요!\n"
         "- 쓸 수 있는 감정: 기쁨, 신남, 힘차게, 속상함, 화남, 두려움, 놀람, 차분, 다정, 진지, 실망, 피곤, "
         "사랑스럽게, 웃으며, 속삭임, 외침, 단호, 안도, 궁금, 반가움.\n"
         "- 감정 단어가 아닌 진짜 대괄호(예: [1번], [메모])는 그대로 읽으니 필요하면 그렇게 써도 된다.\n\n"
@@ -156,7 +173,7 @@ class ClaudeBrain:
         "- 너는 디스코드에서 함께 대화하는 실시간 AI 인공지능이다."
     )
 
-    def __init__(self, *, model: str = "claude-sonnet-4-5", api_key: str | None = None) -> None:
+    def __init__(self, *, model: str = "claude-sonnet-5", api_key: str | None = None) -> None:
         self.model = model
         self._auth = _Auth(api_key)
 
@@ -169,11 +186,18 @@ class ClaudeBrain:
         msgs.append({"role": "user", "content": screen_note + user_text})
         client = self._auth.client()
         # Read the persona live each turn so a dashboard edit applies immediately
-        # (falls back to the built-in PERSONA when no override is saved).
+        # (falls back to the built-in PERSONA when no override is saved). The
+        # brevity rule is appended as its own block so it survives persona edits.
+        system = self._auth.system(get_persona(self.PERSONA), self.BREVITY)
+        if system:
+            # Cache the (static) system prompt so repeat turns skip re-processing
+            # it — lower time-to-first-token. No-op below the model's cache
+            # minimum, so it's harmless when it doesn't engage.
+            system[-1] = {**system[-1], "cache_control": {"type": "ephemeral"}}
         resp = await client.messages.create(
             model=self.model,
-            max_tokens=400,
+            max_tokens=_MAX_TOKENS,
-            system=self._auth.system(get_persona(self.PERSONA)),
+            system=system,
             messages=msgs,
         )
         text = "".join(b.text for b in resp.content if b.type == "text")

wsai/backends/melo.py

@@ -12,7 +12,7 @@ Env:
     WSAI_MELO_DEVICE   cpu | cuda | auto  (default auto: GPU if torch sees one,
                        else CPU; the worker falls back to CPU if CUDA fails)
     WSAI_TTS_OUT_DIR   where wavs are written (default ~/.cache/wsai/tts)
-    WSAI_TTS_SPEED     glyph speed / length_scale multiplier (0.5..2.0, default 1.35)
+    WSAI_TTS_SPEED     glyph speed / length_scale multiplier (0.5..2.0, default 1.4)
     WSAI_TTS_WORD_GAP      intra-sentence pause delta, sec (-0.2..0.5, default -0.07)
     WSAI_TTS_SENTENCE_GAP  sentence-boundary silence, sec (-0.5..1.5, default -0.30)
     WSAI_TTS_PITCH         global semitone offset (-12..12, default 0.0)
@@ -99,7 +99,7 @@ class MeloTTS:
         self.out_dir = Path(out_dir or os.environ.get("WSAI_TTS_OUT_DIR")
                             or (Path.home() / ".cache/wsai/tts"))
         self.speed = float(speed if speed is not None
-                           else os.environ.get("WSAI_TTS_SPEED", "1.35"))
+                           else os.environ.get("WSAI_TTS_SPEED", "1.4"))
         # Reply-global rhythm/pitch controls (see melo_worker.py / docs manual).
         self.word_gap = float(word_gap if word_gap is not None
                               else os.environ.get("WSAI_TTS_WORD_GAP", "-0.07"))

wsai/backends/whisper.py

@@ -18,7 +18,7 @@ stays usable directly.
 Env:
     WSAI_WHISPER_PYTHON    interpreter with faster-whisper installed
                            (default: /home/claude/jarvis-stt/whisper312/bin/python)
-    WSAI_WHISPER_MODEL     model size/name (default: small)
+    WSAI_WHISPER_MODEL     model size/name (default: medium)
     WSAI_WHISPER_DEVICE    cpu | cuda | auto  (default auto: GPU if present,
                            else CPU; the worker falls back to CPU if CUDA fails)
     WSAI_WHISPER_LANGUAGE  forced language, e.g. ko (default ko; "" = autodetect)
@@ -67,7 +67,7 @@ class WhisperSTT:
         audio_source: AsyncIterator[str] | None = None,
     ) -> None:
         self.python = python or os.environ.get("WSAI_WHISPER_PYTHON", _DEFAULT_PYTHON)
-        self.model = model or os.environ.get("WSAI_WHISPER_MODEL", "small")
+        self.model = model or os.environ.get("WSAI_WHISPER_MODEL", "medium")
         self.device = device or os.environ.get("WSAI_WHISPER_DEVICE", "auto")
         # "" means autodetect; a real code like "ko" forces the language.
         env_lang = os.environ.get("WSAI_WHISPER_LANGUAGE", "ko")

wsai/backends/whisper_worker.py

@@ -23,6 +23,31 @@ import os
 import sys
 import time
 
+# Whisper's classic Korean hallucinations on silence/noise/keyboard clatter —
+# it "hears" video-outro boilerplate. Drop these when the whole utterance is one
+# of them AND the segment looked like non-speech, so a real "감사합니다" survives.
+_HALLUCINATIONS = {
+    "감사합니다", "고맙습니다", "감사합니다.", "고맙습니다.",
+    "시청해주셔서 감사합니다", "시청해 주셔서 감사합니다", "끝까지 시청해주셔서 감사합니다",
+    "구독과 좋아요 부탁드립니다", "구독 좋아요 부탁드립니다", "다음 영상에서 만나요",
+    "다음 시간에 만나요", "안녕히 계세요",
+}
+
+
+def _norm(t: str) -> str:
+    return t.strip().rstrip(" .!?…~").strip()
+
+
+def _guard_hallucination(text: str, worst_no_speech: float) -> str:
+    """Blank out a lone known-hallucination phrase when the audio was probably
+    not speech (high no_speech_prob)."""
+    n = _norm(text)
+    if not n:
+        return ""
+    if n in {_norm(h) for h in _HALLUCINATIONS} and worst_no_speech > 0.5:
+        return ""
+    return text
+
 # Split protocol from library noise BEFORE importing anything heavy.
 _proto = os.fdopen(os.dup(1), "w", buffering=1)  # private copy of real stdout
 os.dup2(2, 1)  # fd1 -> stderr, so stray library prints don't hit the protocol
@@ -108,9 +133,31 @@ def main() -> None:
                 wav,
                 language=language,
                 beam_size=int(req.get("beam_size", 5)),
+                # VAD strips non-speech (keyboard clatter, room noise, silence)
+                # before decoding, which both improves accuracy and kills most
+                # hallucinations. speech_pad_ms keeps a little lead/trail so soft
+                # first/last words aren't clipped.
                 vad_filter=bool(req.get("vad_filter", True)),
+                vad_parameters=dict(min_silence_duration_ms=300, speech_pad_ms=250),
+                # Don't feed the previous text back in — that's what makes Whisper
+                # loop/hallucinate. Temperature fallback + thresholds reject
+                # low-confidence (noisy/quiet) decodes instead of inventing words.
+                condition_on_previous_text=False,
+                temperature=[0.0, 0.2, 0.4, 0.6],
+                no_speech_threshold=0.6,
+                log_prob_threshold=-1.0,
+                compression_ratio_threshold=2.4,
             )
-            text = "".join(seg.text for seg in segments).strip()
+            parts, worst_ns = [], 0.0
+            for seg in segments:
+                nsp = float(getattr(seg, "no_speech_prob", 0.0) or 0.0)
+                alp = float(getattr(seg, "avg_logprob", 0.0) or 0.0)
+                # Drop a segment that is almost certainly non-speech noise.
+                if nsp > 0.8 and alp < -0.4:
+                    continue
+                worst_ns = max(worst_ns, nsp)
+                parts.append(seg.text)
+            text = _guard_hallucination("".join(parts).strip(), worst_ns)
             ms = int((time.monotonic() - s) * 1000)
             _emit({"ok": True, "text": text, "language": info.language, "ms": ms})
         except Exception as exc:  # keep the worker alive across bad requests

wsai/bot_control.py

@@ -31,10 +31,19 @@ class BotControl:
         self._cmd_id = 0
         # Per-guild listen filter. Empty whitelist => listen to everyone;
         # blacklist always excludes. Users and roles both supported.
-        self._lists: dict[str, dict[str, Any]] = {}
+        # Restored from the persistent state store so it survives a restart.
+        from . import state_store
+        st = state_store.load()
+        self._lists: dict[str, dict[str, Any]] = st.get("lists") or {}
         # Bot behaviour settings the dashboard toggles and the bot reads on each
         # report. bargeIn: stop the bot's current TTS the moment a user speaks.
-        self._settings: dict[str, Any] = {"bargeIn": True}
+        # bargeInMs: how long (ms) a user must keep speaking before that barge-in
+        # fires — the "유저 음성 인식 시간" the dashboard exposes (default 700).
+        self._settings: dict[str, Any] = {
+            "bargeIn": True,
+            "bargeInMs": 700,
+            **(st.get("botSettings") or {}),
+        }
 
     # -- bot behaviour settings ------------------------------------------ #
     def get_settings(self) -> dict[str, Any]:
@@ -45,7 +54,17 @@ class BotControl:
         with self._lock:
             if "bargeIn" in data:
                 self._settings["bargeIn"] = bool(data["bargeIn"])
-            return dict(self._settings)
+            if "bargeInMs" in data:
+                try:
+                    ms = int(data["bargeInMs"])
+                except (TypeError, ValueError):
+                    ms = 700
+                # Keep it sane: 0ms = instant, cap at 5s so a typo can't wedge it.
+                self._settings["bargeInMs"] = max(0, min(5000, ms))
+            out = dict(self._settings)
+        from . import state_store
+        state_store.patch("botSettings", out)
+        return out
 
     # -- whitelist / blacklist (per guild) ------------------------------- #
     @staticmethod
@@ -72,6 +91,9 @@ class BotControl:
             ]
         with self._lock:
             self._lists[guild_id] = clean
+            snapshot = {g: dict(v) for g, v in self._lists.items()}
+        from . import state_store
+        state_store.patch("lists", snapshot)  # persist per-guild lists across restarts
         return clean
 
     # -- bot -> dashboard (state push) ----------------------------------- #

wsai/config.py

@@ -24,7 +24,7 @@ class Settings:
     text: str | None = None      # None | (discord)
 
     capture_interval: float = 1.5
-    anthropic_model: str = "claude-sonnet-4-5"
+    anthropic_model: str = "claude-sonnet-5"
 
     @classmethod
     def from_env(cls) -> "Settings":

wsai/dashboard.py

@@ -17,8 +17,10 @@ from __future__ import annotations
 
 import json
 import logging
+import os
 import queue
 import threading
+import time
 from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
 
 from .monitor import Monitor
@@ -49,6 +51,20 @@ def _thought_summary(reply: str) -> str:
     return "감정 태그 없음 · 기본 톤으로 답변"
 
 
+# Lone connective/fillers that clearly expect more to come ("그러면…"). On their
+# own they carry no answerable intent, so we discard them instead of blurting a
+# confused reply. (A fuller version would buffer and wait for the continuation.)
+_FRAGMENT_FILLERS = {
+    "그러면", "그래서", "그런데", "근데", "그리고", "그러니까", "그럼", "그게",
+    "저기", "있잖아", "그", "음", "어", "저", "그러", "그니까",
+}
+
+
+def _is_fragment(text: str) -> bool:
+    t = (text or "").strip().rstrip(" .,!?…~").strip()
+    return t in _FRAGMENT_FILLERS
+
+
 def _make_handler(dash: "Dashboard"):
     monitor = dash.monitor
 
@@ -87,6 +103,8 @@ 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:
@@ -103,6 +121,11 @@ def _make_handler(dash: "Dashboard"):
             elif path == "/api/logs/clear":
                 monitor.clear_events()
                 self._send(200, json.dumps({"ok": True}).encode(), "application/json; charset=utf-8")
+            elif path == "/api/turns/clear":
+                monitor.clear_turns()
+                self._send(200, json.dumps({"ok": True}).encode(), "application/json; charset=utf-8")
+            elif path == "/api/turns/delete":
+                self._handle_turn_delete()
             elif path == "/api/logs/delete":
                 self._handle_log_mutate("delete")
             elif path == "/api/logs/edit":
@@ -119,6 +142,10 @@ 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")
 
@@ -226,7 +253,7 @@ def _make_handler(dash: "Dashboard"):
                 return
             prompt_store.set_persona(prompt)
             monitor.log("info", "봇 프롬프트가 수정되었습니다" if prompt.strip()
-                        else "봇 프롬프트가 기본값으로 초기화되었습니다")
+                        else "봇 프롬프트가 기본값으로 초기화되었습니다", cat="PROMPT")
             body = json.dumps({
                 "ok": True,
                 "prompt": prompt_store.get_persona(self._default_persona()),
@@ -247,7 +274,17 @@ def _make_handler(dash: "Dashboard"):
             except (ValueError, AttributeError):
                 self._send_json({"ok": False, "error": "invalid JSON"}, 400)
                 return
+            was_connected = dash.bot.state().get("connected")
             dash.bot.report(data)
+            if not was_connected:  # transition disconnected -> connected
+                # identity is reported as {id, username, tag}; pull a string only
+                # (concatenating a dict here would TypeError and 500 the report).
+                ident = data.get("identity")
+                if isinstance(ident, dict):
+                    who = ident.get("tag") or ident.get("username") or ident.get("id") or ""
+                else:
+                    who = ident if isinstance(ident, str) else ""
+                monitor.log("info", f"디스코드 봇 연결됨{(' · ' + who) if who else ''}", cat="CONNECT")
             # Hand the bot any queued commands + the current listen filters +
             # behaviour settings in the same round trip so it does not have to
             # poll extra endpoints.
@@ -267,10 +304,10 @@ def _make_handler(dash: "Dashboard"):
             channel_id = (data.get("channelId") or "").strip()
             if channel_id and guild_id:
                 cid = dash.bot.enqueue({"type": "join", "guildId": guild_id, "channelId": channel_id})
-                monitor.log("info", f"음성채널 참여 요청 (guild={guild_id} channel={channel_id})")
+                monitor.log("info", f"음성채널 참여 요청 (guild={guild_id} channel={channel_id})", cat="VOICE")
             else:
                 cid = dash.bot.enqueue({"type": "leave"})
-                monitor.log("info", "음성채널 나가기 요청")
+                monitor.log("info", "음성채널 나가기 요청", cat="VOICE")
             self._send_json({"ok": True, "commandId": cid})
 
         def _handle_bot_lists(self) -> None:
@@ -285,9 +322,28 @@ def _make_handler(dash: "Dashboard"):
                 self._send_json({"ok": False, "error": str(exc)}, 400)
                 return
             saved = dash.bot.set_lists(guild_id, data.get("lists") or {})
-            monitor.log("info", f"청취 화이트/블랙리스트 업데이트 (guild={guild_id})")
+            monitor.log("info", f"청취 화이트/블랙리스트 업데이트 (guild={guild_id})", cat="FILTER")
             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."""
@@ -298,7 +354,7 @@ def _make_handler(dash: "Dashboard"):
                 self._send_json({"ok": False, "error": "invalid JSON"}, 400)
                 return
             settings = dash.bot.set_settings(data)
-            monitor.log("info", "봇 설정 변경: " + json.dumps(settings, ensure_ascii=False))
+            monitor.log("info", "봇 설정 변경: " + json.dumps(settings, ensure_ascii=False), cat="SETTING")
             self._send_json({"ok": True, "settings": settings})
 
         def _handle_tts_settings_get(self) -> None:
@@ -324,7 +380,8 @@ def _make_handler(dash: "Dashboard"):
             tgt = data.get("emotion") or "base"
             monitor.log("info", f"봇 목소리 파라미터 적용 ({tgt}): base="
                         + json.dumps(res["base"], ensure_ascii=False)
-                        + " overrides=" + json.dumps(res["overrides"], ensure_ascii=False))
+                        + " overrides=" + json.dumps(res["overrides"], ensure_ascii=False),
+                        cat="TTS")
             self._send_json(res)
 
         def _handle_tts_preview(self) -> None:
@@ -351,6 +408,20 @@ def _make_handler(dash: "Dashboard"):
                 return
             self._send(200, wav, "audio/wav")
 
+        def _handle_turn_delete(self) -> None:
+            """Delete one conversation turn by id."""
+            raw = self._read_body()
+            try:
+                data = json.loads(raw.decode("utf-8")) if raw else {}
+                turn_id = int(data.get("id"))
+            except (ValueError, TypeError, AttributeError):
+                self._send(400, json.dumps({"ok": False, "error": "id required"}).encode(),
+                           "application/json; charset=utf-8")
+                return
+            ok = monitor.delete_turn(turn_id)
+            self._send(200 if ok else 404,
+                       json.dumps({"ok": ok}).encode(), "application/json; charset=utf-8")
+
         def _handle_log_mutate(self, action: str) -> None:
             """Per-line log delete/edit by event id."""
             raw = self._read_body()
@@ -422,6 +493,10 @@ class Dashboard:
         self.bot = BotControl()  # dashboard <-> Discord bot control plane
         self._history: list[tuple[str, str]] = []
         self._history_turns = history_turns
+        # A lone connective ("그러면") is held here to see if the next utterance
+        # continues it; merged if it arrives within the window, else dropped.
+        self._pending_fragment: dict | None = None
+        self._fragment_window = float(os.environ.get("WSAI_FRAGMENT_MERGE_MS", "5000")) / 1000.0
         self._server: ThreadingHTTPServer | None = None
         self._thread: threading.Thread | None = None
         self._loop = None
@@ -484,12 +559,26 @@ class Dashboard:
     def _tts_base(self) -> dict:
         t = self.tts
         return {
-            "speed": float(getattr(t, "speed", 1.35)),
+            "speed": float(getattr(t, "speed", 1.4)),
             "word_gap": float(getattr(t, "word_gap", -0.07)),
             "sentence_gap": float(getattr(t, "sentence_gap", -0.30)),
             "pitch": float(getattr(t, "pitch", 0.0)),
         }
 
+    def _persist_tts(self) -> None:
+        from . import state_store
+        state_store.patch("tts", {
+            "base": self._tts_base(),
+            "overrides": dict(getattr(self.tts, "emotion_overrides", {}) or {}),
+        })
+
+    def _persist_models(self) -> None:
+        from . import state_store
+        state_store.patch("models", {
+            "stt": getattr(self.stt, "model", None),
+            "llm": getattr(self.brain, "model", None),
+        })
+
     def tts_settings(self) -> dict:
         """Base (공통) controls + the per-emotion overrides + the emotion list
         (canonical key + Korean label) so the dashboard can offer per-emotion
@@ -505,11 +594,20 @@ class Dashboard:
         }
 
     def set_tts_settings(self, data: dict) -> dict:
-        """Apply controls. Without ``emotion`` (or emotion == "base"), set the
+        """Apply controls. ``emotion`` selects the target:
-        base/공통 values on the live TTS instance. With a specific emotion, store
+        - "all"  → set base to the values AND clear every per-emotion override,
-        (or, if ``reset``, clear) that emotion's override."""
+                   so ALL emotions speak with these values (전체변경).
+        - "" / "base" → set the base/공통 values on the live TTS instance.
+        - a specific emotion → store (or, if ``reset``, clear) its override."""
         emotion = data.get("emotion")
-        if not emotion or emotion == "base":
+        if emotion == "all":
+            for key, (lo, hi, _step) in self._TTS_RANGES.items():
+                v = data.get(key)
+                if v is not None:
+                    setattr(self.tts, key, float(max(lo, min(hi, float(v)))))
+            if getattr(self.tts, "emotion_overrides", None):
+                self.tts.emotion_overrides.clear()  # all emotions inherit the new base
+        elif not emotion or emotion == "base":
             for key, (lo, hi, _step) in self._TTS_RANGES.items():
                 v = data.get(key)
                 if v is not None:
@@ -522,6 +620,7 @@ class Dashboard:
                 store.pop(emotion, None)
             else:
                 store[emotion] = self._tts_overrides(data)
+        self._persist_tts()
         return self.tts_settings()
 
     def _tts_overrides(self, data: dict) -> dict:
@@ -548,6 +647,81 @@ 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", "claude-sonnet-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}")
+        changed = model != self.stt.model
+        if changed:
+            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)
+            self.monitor.log("info", f"STT 모델 전환 시작: {model} (로딩 중…)", cat="MODEL")
+            self._persist_models()
+        return {**self.models_settings()["stt"], "changed": changed}
+
+    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', '?')})", cat="READY")
+        except Exception as exc:  # noqa: BLE001
+            self.monitor.log("error", f"STT 모델 로드 실패({self.stt.model}): {exc}", cat="MODEL")
+
+    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}")
+        changed = model != self.brain.model
+        self.brain.model = model
+        if changed:
+            self.monitor.log("info", f"LLM 모델 변경: {model} (다음 답변부터 적용)", cat="MODEL")
+            self._persist_models()
+        return {**self.models_settings()["llm"], "changed": changed}
+
+    @staticmethod
+    def _step(turn, name: str, ms: float) -> None:
+        """Record one finished timing step (STT / LLM / TTS) on a turn."""
+        st = turn.step(name)
+        st.ok, st.ms = True, float(ms)
+        turn._steps.append(st)
+
     def voice_turn(self, audio_bytes: bytes, speaker: str = "",
                    guild: str = "", channel: str = "") -> dict:
         """One Discord voice turn: decode the uploaded utterance, recognise it
@@ -578,6 +752,7 @@ class Dashboard:
                 check=True, capture_output=True,
             )
             heard = (self._submit(self.stt.transcribe(wav)) or "").strip()
+            self._step(turn, "STT", (time.monotonic() - t0) * 1000)  # 인식(+디코드)
             turn.heard(heard or "(빈 결과)")
             if not heard:
                 # Nothing recognised (silence/noise): mark it as [잡음] and skip
@@ -586,16 +761,35 @@ class Dashboard:
                 turn.replied("[잡음]")
                 turn.finish()
                 return {"heard": heard, "reply": "[잡음]", "wav": b""}
+            now = time.monotonic()
+            pending = self._pending_fragment
+            if pending and (now - pending["ts"]) > self._fragment_window:
+                pending = self._pending_fragment = None  # too old — give up on it
+            if _is_fragment(heard):
+                # A lone connective ("그러면") — hold it and wait for the next
+                # utterance to continue it. If nothing follows within the window
+                # it's silently dropped (never answered).
+                self._pending_fragment = {"text": heard, "ts": now}
+                turn.thought("불완전한 말(연결어) — 이어질 말 대기 (없으면 폐기)")
+                turn.replied("[대기]")
+                turn.finish()
+                return {"heard": heard, "reply": "[대기]", "wav": b""}
+            if pending:
+                # A real utterance arrived in time — merge the held fragment in
+                # front and answer the whole thing ("그러면" + "뭐 먹지").
+                heard = (pending["text"].rstrip(" .,!?…~") + " " + heard).strip()
+                self._pending_fragment = None
+                turn.heard(heard)
+            t_llm = time.monotonic()
             reply_text = self._think(heard)
+            self._step(turn, "LLM" if self.brain else "echo", (time.monotonic() - t_llm) * 1000)
             turn.thought(_thought_summary(reply_text))
             turn.replied(reply_text)
+            t_tts = time.monotonic()
             out_path = self._submit(self.tts.synth(_speech_text(reply_text)))
             with open(out_path, "rb") as f:
                 reply_wav = f.read()
-            ms = int((time.monotonic() - t0) * 1000)
+            self._step(turn, "TTS", (time.monotonic() - t_tts) * 1000)  # 합성
-            step = turn.step("STT+두뇌+TTS" if self.brain else "STT+TTS(GPU)")
-            step.ok, step.ms = True, float(ms)
-            turn._steps.append(step)
             turn.finish()
             try:
                 os.remove(out_path)
@@ -631,7 +825,7 @@ class Dashboard:
                 self.monitor.add_claude_usage(u.get("input", 0), u.get("output", 0))
         except Exception as exc:  # noqa: BLE001
             log.exception("brain failed")
-            self.monitor.log("error", f"두뇌 응답 실패: {exc}")
+            self.monitor.log("error", f"LLM 응답 실패: {exc}", cat="BRAIN")
             blob = f"{getattr(exc, 'status_code', '')} {exc}".lower()
             if "529" in blob or "overload" in blob:
                 # Transient server overload survived the SDK retries.
@@ -729,8 +923,11 @@ PAGE = r"""<!DOCTYPE html>
   .comp{display:flex;gap:8px;flex-wrap:wrap;margin:2px 0 18px}
   .comp .pill{font-size:11.5px}
   .turn{background:var(--panel);border:1px solid var(--line);border-radius:14px;padding:14px 16px;margin:12px 0;
-    animation:rise .25s ease}
+    animation:rise .25s ease;position:relative}
   @keyframes rise{from{opacity:0;transform:translateY(6px)}to{opacity:1;transform:none}}
+  .tdel{margin-left:auto;background:none;border:none;color:var(--muted);cursor:pointer;font-size:14px;
+    line-height:1;padding:2px 4px;border-radius:6px}
+  .tdel:hover{color:var(--err);background:#2a1620}
   .turn.active{border-color:#2b5d86;box-shadow:0 0 0 1px #14324a inset}
   .turn.error{border-color:#5c2530}
   .trow{display:flex;align-items:baseline;gap:10px;margin-bottom:8px}
@@ -771,6 +968,12 @@ PAGE = r"""<!DOCTYPE html>
   .cfgrow{display:flex;gap:9px;align-items:center;font-size:13px;color:var(--fg);cursor:pointer;padding:4px 0}
   .cfgrow input[type=checkbox]{width:16px;height:16px;accent-color:#3aa0ff;cursor:pointer}
   .cfghint{color:var(--muted);font-weight:400;font-size:12px}
+  .mload{display:flex;align-items:center;gap:9px;margin-top:10px;padding:8px 12px;
+    background:#0f2333;border:1px solid #234a63;border-radius:10px;font-size:13px;color:var(--fg)}
+  .mload .bar{flex:1;height:6px;border-radius:4px;background:#12304a;overflow:hidden;position:relative}
+  .spin{width:15px;height:15px;border:2px solid #2a4a63;border-top-color:#3aa0ff;border-radius:50%;
+    display:inline-block;animation:spin 0.8s linear infinite}
+  @keyframes spin{to{transform:rotate(360deg)}}
   .sttrow{display:flex;gap:10px;align-items:center;flex-wrap:wrap}
   .btn{background:#173042;border:1px solid #234a63;color:var(--fg);border-radius:10px;padding:8px 14px;font-size:13px;cursor:pointer}
   .btn:hover{background:#1d3d54}
@@ -836,6 +1039,10 @@ PAGE = r"""<!DOCTYPE html>
   .lst-item:last-child{border-bottom:none}
   .lst-item .nm{flex:1}
   .lst-item .rl{color:var(--muted);font-size:11px}
+  .role-h{cursor:pointer;user-select:none}
+  .role-h .rcaret{display:inline-block;width:11px;color:var(--muted);font-weight:400}
+  .role-members{background:#0c141d;border-top:1px solid #16202b}
+  .role-members .lst-item{padding-left:26px}
   .mini{padding:3px 8px;font-size:11.5px;border-radius:7px;cursor:pointer;border:1px solid var(--line);background:#173042;color:var(--fg)}
   .mini.w{border-color:#1f5236;color:#9ff0bd}
   .mini.b{border-color:#5c2530;color:#ffb3bb}
@@ -849,7 +1056,13 @@ PAGE = r"""<!DOCTYPE html>
   main{padding-bottom:46px}
   .logdock{position:fixed;left:0;right:0;bottom:0;z-index:15;background:#0a0e13;
     border-top:1px solid var(--line);display:flex;flex-direction:column;
-    max-height:45vh;box-shadow:0 -8px 24px rgba(0,0,0,.35)}
+    max-height:90vh;box-shadow:0 -8px 24px rgba(0,0,0,.35)}
+  .logresize{height:7px;cursor:ns-resize;background:#0d141c;border-bottom:1px solid var(--line);
+    flex:0 0 auto;touch-action:none}
+  .logresize:hover{background:#1d3d54}
+  .logresize::before{content:"";display:block;width:44px;height:3px;margin:2px auto 0;
+    border-radius:2px;background:#2a3a49}
+  .logdock.collapsed .logresize{display:none}
   .logbar{display:flex;align-items:center;gap:8px;padding:6px 12px;border-bottom:1px solid var(--line);
     background:#0d141c;flex-wrap:wrap}
   .logtoggle{background:none;border:none;color:var(--fg);font-size:12.5px;cursor:pointer;font-weight:600;padding:4px 6px}
@@ -869,6 +1082,19 @@ PAGE = r"""<!DOCTYPE html>
   .logline.info .lv{color:var(--accent)}
   .logline.error .lv{color:var(--err)}
   .logline.warn .lv{color:var(--warn)}
+  .lc-tag{flex:0 0 auto;min-width:62px;text-align:center;font-size:10px;font-weight:700;
+    letter-spacing:.3px;padding:1px 7px;border-radius:9px;background:#1b2b3a;color:#8fb3d6;
+    border:1px solid #24405a}
+  .lc-READY{background:#12331f;color:#7fe0a0;border-color:#1f5a34}
+  .lc-CONNECT{background:#12283f;color:#7fb8ff;border-color:#22507f}
+  .lc-MODEL{background:#2a2340;color:#c3a6ff;border-color:#463a7a}
+  .lc-TTS{background:#0f2f33;color:#6fe0d8;border-color:#1f5a5a}
+  .lc-VOICE{background:#33280f;color:#e0c06f;border-color:#5a481f}
+  .lc-FILTER{background:#301a2a;color:#e79fd0;border-color:#5a2545}
+  .lc-SETTING{background:#22303a;color:#9fc7e0;border-color:#2f5568}
+  .lc-TURN,.lc-ERROR{background:#3a1520;color:#ffb3bb;border-color:#7a2531}
+  .lc-BRAIN,.lc-VISION{background:#2a2340;color:#c3a6ff;border-color:#463a7a}
+  .lc-WARN{background:#332a10;color:#ffd98a;border-color:#5a4a1f}
   .logline .lm{flex:1;color:var(--fg);white-space:pre-wrap;word-break:break-word}
   .logline.error .lm{color:#ffb3bb}
   .logline .lacts{opacity:0;display:flex;gap:4px}
@@ -881,7 +1107,7 @@ PAGE = r"""<!DOCTYPE html>
 <header>
   <div>
     <h1>watch_sceen_ai · 실시간 상태</h1>
-    <div class="sub">STT → 두뇌 → TTS 음성 루프를 단계별로 관찰</div>
+    <div class="sub">STT → LLM → TTS 음성 루프를 단계별로 관찰</div>
   </div>
   <div class="pill"><span id="dot" class="dot off"></span><span id="listen">연결 대기</span></div>
   <button id="promptBtn" class="btn hbtn">📝 프롬프트</button>
@@ -895,6 +1121,7 @@ PAGE = r"""<!DOCTYPE html>
   </div>
 </header>
 <main>
+  <div class="comp" id="comp"></div>
   <section class="botbar" id="botbar">
     <span class="botinfo" id="botinfo"><span class="dot off"></span>봇: 연결 안 됨</span>
     <label>서버 <select id="guildSel"><option value="">없음</option></select></label>
@@ -911,8 +1138,8 @@ PAGE = r"""<!DOCTYPE html>
         <span id="tEmoNote" class="sttstat">감정을 고르면 그 감정만 따로 조절합니다. 기본(공통)은 태그 없는 말투에 적용됩니다.</span>
       </div>
       <div class="ttsgrid">
-        <label>글자 속도 <b id="tvSpeed">1.35x</b>
+        <label>글자 속도 <b id="tvSpeed">1.40x</b>
-          <input type="range" id="tSpeed" min="0.5" max="2.0" step="0.05" value="1.35"></label>
+          <input type="range" id="tSpeed" min="0.5" max="2.0" step="0.05" value="1.4"></label>
         <label>단어 간격 <b id="tvWord">-70 ms</b>
           <input type="range" id="tWord" min="-0.2" max="0.5" step="0.01" value="-0.07"></label>
         <label>문장 간격 <b id="tvSent">-0.30 s</b>
@@ -935,6 +1162,10 @@ PAGE = r"""<!DOCTYPE html>
     <div id="botcfgBody" style="display:none">
       <label class="cfgrow"><input type="checkbox" id="cfgBargeIn" checked>
         <span>유저 목소리 들을 때 봇이 말하던 것 중지 <b class="cfghint">(기본: 켜짐)</b></span></label>
+      <label class="cfgrow"><span>유저 음성 인식 시간
+        <input type="number" id="cfgBargeInMs" min="0" max="5000" step="50" value="700"
+          style="width:80px;margin:0 4px"> ms
+        <b class="cfghint">(이 시간 이상 말해야 봇이 멈춤 · 기본: 700)</b></span></label>
     </div>
   </section>
   <section class="sttbox" id="logcfgbox">
@@ -944,8 +1175,28 @@ 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
+          <select id="mLLM" class="ttssel"></select></label>
+        <button id="mLLMapply" class="btn">적용</button>
+        <span id="mLLMstat" class="sttstat"></span>
+      </div>
+      <div id="mLoad" class="mload" style="display:none">
+        <span class="spin"></span><span id="mLoadText">로딩 중…</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. 마지막으로 고른 모델은 재시작 후에도 유지됩니다(기본값 Sonnet 5).</p>
+    </div>
+  </section>
   <div class="demobar" id="demobar" style="display:none"></div>
-  <div class="comp" id="comp"></div>
   <div class="tfilter" id="tfilter">
     <span class="tf-label">대화 로그 검색</span>
     <select id="tfTime">
@@ -960,20 +1211,30 @@ PAGE = r"""<!DOCTYPE html>
     <input id="tfChannel" placeholder="채널">
     <input id="tfText" placeholder="내용(들음/답변)">
     <button id="tfClear" class="btn hbtn">초기화</button>
+    <button id="turnsClear" class="btn hbtn">🗑 전체 삭제</button>
     <span id="tfCount" class="tf-count"></span>
   </div>
   <div id="turns"></div>
   <div id="empty" class="empty">아직 대화가 없습니다. 사용자가 말하면 여기에 단계별로 나타납니다.</div>
 </main>
 <div id="logdock" class="logdock">
+  <div id="logResize" class="logresize" title="드래그해서 로그 높이 조절"></div>
   <div class="logbar">
     <button id="logToggle" class="logtoggle">▾ 이벤트 / 오류 로그</button>
     <input id="logSearch" class="logsearch" placeholder="로그 검색 (텍스트)">
-    <select id="logLevel" class="logsel">
+    <select id="logLevel" class="logsel" title="레벨(타입)">
-      <option value="">전체</option>
+      <option value="">레벨 전체</option>
-      <option value="error">오류만</option>
+      <option value="error">오류(ERROR)</option>
-      <option value="warn">경고만</option>
+      <option value="warn">경고(WARN)</option>
-      <option value="info">정보만</option>
+      <option value="info">정보(INFO)</option>
+    </select>
+    <select id="logCat" class="logsel" title="이벤트 종류"><option value="">종류 전체</option></select>
+    <select id="logTime" class="logsel" title="시간">
+      <option value="0">전체 시간</option>
+      <option value="5">최근 5분</option>
+      <option value="30">최근 30분</option>
+      <option value="60">최근 1시간</option>
+      <option value="180">최근 3시간</option>
     </select>
     <span id="logCount" class="logcount"></span>
     <button id="logClear" class="btn logbtn">로그 삭제</button>
@@ -1037,14 +1298,14 @@ function renderStatus(s){
   const bar = $('demobar');
   if(mockParts.length){
     bar.style.display='block';
-    bar.innerHTML = '⚠ <b>데모 모드</b> — 실제 음성/STT/두뇌/TTS가 아직 연결되지 않아, 아래 대화는 '
+    bar.innerHTML = '⚠ <b>데모 모드</b> — 실제 음성/STT/LLM/TTS가 아직 연결되지 않아, 아래 대화는 '
       + '실제로 들은 내용이 아니라 <b>목(mock) 예시 스크립트</b>입니다. '
       + '실제 엔진(faster-whisper·Claude·MeloTTS)을 붙이면 이 자리에 진짜 발화·지연·오류가 표시됩니다.';
   } else {
     bar.style.display='none';
   }
   const el = $('comp'); el.innerHTML = '';
-  const names = {source:'눈(소스)', vision:'시각', stt:'귀(STT)', brain:'두뇌', tts:'입(TTS)', text:'텍스트'};
+  const names = {source:'눈(소스)', vision:'시각', stt:'귀(STT)', brain:'LLM', tts:'입(TTS)', text:'텍스트'};
   for(const k of Object.keys(names)){
     if(!(k in comps)) continue;
     const v = comps[k];
@@ -1082,7 +1343,8 @@ function turnEl(t){
       +'<span class="badge">#'+t.id+' · '+esc(t.source||'voice')+'</span>'
       +(t.speaker?'<span class="badge">🗣 '+esc(t.speaker)+'</span>':'')
       +(t.channel?'<span class="badge">🔊 '+esc((t.guild?t.guild+' / ':'')+t.channel)+'</span>':'')
-      +'<span class="time">'+fmtTime(t.wall)+'</span></div>'
+      +'<span class="time">'+fmtTime(t.wall)+'</span>'
+      +'<button class="tdel" data-id="'+t.id+'" title="이 대화 삭제">✕</button></div>'
     +'<div class="line"><span class="tag">들음</span><span class="heard">'+(t.heard?esc(t.heard):'<i style="color:var(--muted)">(수신 대기)</i>')+'</span></div>'
     +'<div class="line"><span class="tag">생각</span><span class="thought">'+(t.thought?esc(t.thought):'<i style="color:var(--muted)">…</i>')+'</span></div>'
     +'<div class="line"><span class="tag">답변</span><span class="reply">'+(t.reply?esc(t.reply):'<i style="color:var(--muted)">…생각 중</i>')+'</span></div>'
@@ -1102,6 +1364,12 @@ function upsertTurn(t){
   else { cont.prepend(fresh); }
   applyTurnFilter();
 }
+function removeTurn(id){
+  turns.delete(id);
+  const el=$('turn-'+id); if(el) el.remove();
+  if(turns.size===0) $('empty').style.display='block';
+  applyTurnFilter();
+}
 
 // --- 대화 로그 검색: 시간·유저·서버·채널·내용 필터 ------------------------ #
 function turnFilter(){
@@ -1110,7 +1378,7 @@ function turnFilter(){
     text:$('tfText').value.trim().toLowerCase() };
 }
 function isNoiseTurn(t){
-  return (t.heard==='(빈 결과)') || (t.reply==='[잡음]');
+  return (t.heard==='(빈 결과)') || (t.reply==='[잡음]') || (t.reply==='[대기]');
 }
 function turnMatches(t, f){
   if(hideNoise && isNoiseTurn(t)) return false;   // 잡음 로그 표시 안 함 토글
@@ -1131,20 +1399,34 @@ function applyTurnFilter(){
 
 // --- Bottom terminal log panel: store all events, render filtered ---------- #
 let logEvents = [];            // {id, level, message, wall}
+function logCat(e){ return (e.cat || (e.level||'info').toUpperCase()); }
 function logMatches(e){
   const lv = $('logLevel').value;
   if(lv && e.level!==lv) return false;
+  const c = $('logCat').value;
+  if(c && logCat(e)!==c) return false;
+  const mins = +$('logTime').value;
+  if(mins && (Date.now()/1000 - (e.wall||0)) > mins*60) return false;
   const q = $('logSearch').value.trim().toLowerCase();
-  if(q && !((e.message||'').toLowerCase().includes(q) || fmtTime(e.wall).includes(q))) return false;
+  if(q && !((e.message||'').toLowerCase().includes(q) || logCat(e).toLowerCase().includes(q) || fmtTime(e.wall).includes(q))) return false;
   return true;
 }
+// Keep the 종류(카테고리) dropdown in sync with whatever categories appear.
+function refreshCatOptions(){
+  const sel=$('logCat'); const cur=sel.value;
+  const cats=[...new Set(logEvents.map(logCat))].sort();
+  sel.innerHTML='<option value="">종류 전체</option>'
+    + cats.map(c=>'<option value="'+c+'">'+esc(c)+'</option>').join('');
+  sel.value = cats.includes(cur) ? cur : '';
+}
 function renderLogs(){
   const body = $('logbody');
+  refreshCatOptions();
   const shown = logEvents.filter(logMatches);
   body.innerHTML = shown.map(e =>
     '<div class="logline '+(e.level||'info')+'" data-id="'+e.id+'">'
     + '<span class="lt">'+fmtTime(e.wall)+'</span>'
-    + '<span class="lv">'+esc(e.level||'info')+'</span>'
+    + '<span class="lc-tag lc-'+esc(logCat(e))+'">'+esc(logCat(e))+'</span>'
     + '<span class="lm">'+esc(e.message)+'</span>'
     + '<span class="lacts"><button class="lact" data-act="edit" title="수정">✎</button>'
     + '<button class="lact" data-act="del" title="삭제">✕</button></span>'
@@ -1179,6 +1461,8 @@ function connect(){
     if(ev.type==='snapshot') applySnapshot(ev.snapshot);
     else if(ev.type==='status') renderStatus(ev.status);
     else if(ev.type==='turn') upsertTurn(ev.turn);
+    else if(ev.type==='turn_deleted') removeTurn(ev.id);
+    else if(ev.type==='turns_cleared') { turns.clear(); $('turns').innerHTML=''; $('empty').style.display='block'; applyTurnFilter(); }
     else if(ev.type==='log') { addEvent(ev); if(statusData){ statusData.errors_total=(statusData.errors_total||0)+(ev.level==='error'?1:0); $('s-errors').textContent=statusData.errors_total; } }
     else if(ev.type==='logs_cleared') { logEvents=[]; renderLogs(); }
     else if(ev.type==='log_deleted') { logEvents=logEvents.filter(e=>e.id!==ev.id); renderLogs(); }
@@ -1203,8 +1487,23 @@ function wireCollapse(toggleId, bodyId, caretId){
       body:JSON.stringify({bargeIn:$('cfgBargeIn').checked})});
       toast('봇 설정 저장됨'); }catch(e){}
   };
+  // 유저 음성 인식 시간(ms): 서버가 0~5000으로 클램프, 봇이 다음 report에서 읽어감.
+  $('cfgBargeInMs').onchange = async ()=>{
+    let ms = Math.round(Number($('cfgBargeInMs').value));
+    if(!Number.isFinite(ms)) ms = 700;
+    ms = Math.max(0, Math.min(5000, ms));
+    $('cfgBargeInMs').value = ms;
+    try{ const r=await fetch('/api/bot/settings',{method:'POST',headers:{'Content-Type':'application/json'},
+      body:JSON.stringify({bargeInMs:ms})});
+      const j=await r.json();
+      if(j&&j.ok&&j.settings&&j.settings.bargeInMs!=null) $('cfgBargeInMs').value=j.settings.bargeInMs;
+      toast('봇 설정 저장됨'); }catch(e){}
+  };
   fetch('/api/bot/settings').then(r=>r.json()).then(j=>{
-    if(j&&j.ok&&j.settings) $('cfgBargeIn').checked = j.settings.bargeIn !== false;
+    if(j&&j.ok&&j.settings){
+      $('cfgBargeIn').checked = j.settings.bargeIn !== false;
+      if(j.settings.bargeInMs!=null) $('cfgBargeInMs').value = j.settings.bargeInMs;
+    }
   }).catch(()=>{});
   // 로그 설정: 잡음 로그 표시 안 함 — 클라이언트 필터(로컬 저장).
   $('cfgHideNoise').checked = hideNoise;
@@ -1213,12 +1512,84 @@ function wireCollapse(toggleId, bodyId, caretId){
     localStorage.setItem('wsai_hideNoise', hideNoise ? '1':'0');
     applyTurnFilter();
   };
+  // 모델 설정: STT 크기 / LLM 모델 실시간 전환.
+  wireCollapse('modelToggle','modelBody','modelCaret');
+  initModels();
 })();
+const STT_LABEL = {tiny:'tiny (가장 빠름)', base:'base', small:'small (빠름)',
+  medium:'medium (기본·정확)', 'large-v3':'large-v3 (최고 정확도)'};
+const LLM_LABEL = {'claude-haiku-4-5':'Haiku 4.5 (가장 빠름)',
+  'claude-sonnet-4-5':'Sonnet 4.5 (고품질·조금 느림)',
+  'claude-sonnet-5':'Sonnet 5 (기본 · 빠르고 고품질)'};
+function fillSel(sel, options, current, labels){
+  sel.innerHTML='';
+  options.forEach(o=>{ const el=document.createElement('option');
+    el.value=o; el.textContent=(labels[o]||o); sel.appendChild(el); });
+  if(current) sel.value=current;
+}
+async function initModels(){
+  let j; try{ j=await (await fetch('/api/models')).json(); }catch(e){ return; }
+  if(!j||!j.ok) return;
+  if(j.stt.enabled){ fillSel($('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;
+    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){ $('mSTTstat').textContent='실패: '+(jj.error||''); return; }
+      if(!jj.stt.changed){ flash3($('mSTTstat'), '변경사항 없음', '현재: '+(jj.stt.current||m)+(jj.stt.device?(' · '+jj.stt.device):'')); return; }
+      pollSTTLoad(m);   // 실제 변경 → 아래 로딩창 + 완료 시 이벤트 로그
+    }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){ $('mLLMstat').textContent='실패: '+(jj.error||''); return; }
+      if(!jj.llm.changed){ flash3($('mLLMstat'), '변경사항 없음', '현재: '+(LLM_LABEL[jj.llm.current]||jj.llm.current)); return; }
+      toast('LLM 모델 변경: '+(LLM_LABEL[m]||m));
+      $('mLLMstat').textContent='현재: '+(LLM_LABEL[m]||m)+' · 다음 답변부터';
+    }catch(e){ $('mLLMstat').textContent='오류: '+e; }
+  };
+}
+// 3초 동안 임시 메시지를 보여준 뒤 원래(현재 모델) 텍스트로 복귀.
+function flash3(el, temp, revert){
+  el.textContent = temp;
+  clearTimeout(el._t);
+  el._t = setTimeout(()=>{ el.textContent = revert; }, 3000);
+}
+// STT 실제 전환: 아래 로딩창을 띄우고 /api/models를 폴링해 ready가 될 때까지 표시.
+let sttPollTimer=null;
+async function pollSTTLoad(target){
+  const load=$('mLoad'), txt=$('mLoadText');
+  load.style.display='flex';
+  $('mSTTstat').textContent='전환 중…';
+  const t0=Date.now();
+  if(sttPollTimer) clearInterval(sttPollTimer);
+  const tick=async ()=>{
+    const sec=Math.round((Date.now()-t0)/1000);
+    txt.textContent = 'STT '+target+' 로딩 중… (경과 '+sec+'초, 처음이면 다운로드로 수 분 걸릴 수 있어요)';
+    try{
+      const j=await (await fetch('/api/models')).json();
+      if(j&&j.ok&&j.stt.current===target&&j.stt.ready){
+        clearInterval(sttPollTimer); sttPollTimer=null;
+        load.style.display='none';
+        $('mSTTstat').textContent='현재: '+target+(j.stt.device?(' · '+j.stt.device):'')+' · 로드 완료';
+        toast('STT 모델 로드 완료: '+target);
+      }
+    }catch(e){}
+  };
+  tick(); sttPollTimer=setInterval(tick, 1500);
+}
 
 // --- 봇 목소리(TTS) 조절: 감정별 슬라이더 → 미리듣기 → 봇 적용 -------------- #
 let ttsInited = false;
 let ttsState = {base:{}, overrides:{}, emotions:[]};   // last-known server state
-const TTS_DEFAULT = {speed:1.35, word_gap:-0.07, sentence_gap:-0.3, pitch:0};
+const TTS_DEFAULT = {speed:1.4, word_gap:-0.07, sentence_gap:-0.3, pitch:0};
 function ttsLabels(){
   const sp=parseFloat($('tSpeed').value), wg=parseFloat($('tWord').value),
         sg=parseFloat($('tSent').value), pt=parseInt($('tPitch').value,10);
@@ -1243,9 +1614,9 @@ function ttsSet(s){
 // (base itself when 기본 or when the emotion has no override).
 function ttsValuesFor(key){
   const base = Object.assign({}, TTS_DEFAULT, ttsState.base);
-  if(key && key!=='base' && ttsState.overrides && ttsState.overrides[key])
+  if(key && key!=='base' && key!=='all' && ttsState.overrides && ttsState.overrides[key])
     return Object.assign({}, base, ttsState.overrides[key]);
-  return base;
+  return base;   // 'all' and 'base' both edit the base values
 }
 function ttsApplyState(j){
   if(!j || !j.ok) return;
@@ -1253,6 +1624,8 @@ function ttsApplyState(j){
   // Refresh dropdown labels to mark which emotions are customised.
   const sel=$('tEmotion'); const cur=sel.value||'base';
   sel.innerHTML='';
+  const allOpt=document.createElement('option');   // 최상단 전체변경
+  allOpt.value='all'; allOpt.textContent='🌐 전체변경 (모든 감정)'; sel.appendChild(allOpt);
   ttsState.emotions.forEach(e=>{
     const o=document.createElement('option'); o.value=e.key;
     const custom = e.key!=='base' && ttsState.overrides[e.key];
@@ -1270,9 +1643,11 @@ async function initTts(){
     $('ttsCaret').textContent = open ? '▾' : '▸';
   };
   ['tSpeed','tWord','tSent','tPitch'].forEach(id => $(id).addEventListener('input', ttsLabels));
-  $('tEmotion').onchange = ()=>{ ttsSet(ttsValuesFor($('tEmotion').value));
+  $('tEmotion').onchange = ()=>{ const k=$('tEmotion').value; ttsSet(ttsValuesFor(k));
-    $('tStat').textContent = $('tEmotion').value==='base'
+    $('tStat').textContent = k==='all'
-      ? '기본(공통) 값을 조절 중입니다.' : '"'+$('tEmotion').selectedOptions[0].textContent.replace(' ●','')+'" 감정만 조절 중입니다.'; };
+      ? '모든 감정을 한 번에 조절 중입니다 (적용 시 전체 반영).'
+      : (k==='base' ? '기본(공통) 값을 조절 중입니다.'
+        : '"'+$('tEmotion').selectedOptions[0].textContent.replace(' ●','')+'" 감정만 조절 중입니다.'); };
   $('tPreview').onclick = async ()=>{
     const text=($('tText').value||'').trim();
     if(!text){ $('tStat').textContent='미리들을 문장을 입력하세요.'; return; }
@@ -1292,13 +1667,13 @@ async function initTts(){
       const r=await fetch('/api/tts/settings',{method:'POST',headers:{'Content-Type':'application/json'},
         body:JSON.stringify(Object.assign({emotion:key}, ttsVals()))});
       const j=await r.json();
-      if(j.ok){ ttsApplyState(j); toast((key==='base'?'기본(공통)':'해당 감정')+' 적용됨 · 다음 답변부터 반영'); }
+      if(j.ok){ ttsApplyState(j); toast((key==='all'?'모든 감정':(key==='base'?'기본(공통)':'해당 감정'))+' 적용됨 · 다음 답변부터 반영'); }
       else { $('tStat').textContent='적용 실패: '+(j.error||''); }
     }catch(e){ $('tStat').textContent='적용 오류: '+e; }
   };
   $('tReset').onclick = async ()=>{
     const key=$('tEmotion').value||'base';
-    if(key==='base'){ ttsSet(TTS_DEFAULT);
+    if(key==='base' || key==='all'){ ttsSet(TTS_DEFAULT);
       $('tStat').textContent='기본값으로 세팅됨 (적용을 눌러 반영).'; return; }
     try{  // clear this emotion's override -> it inherits 기본(공통) again
       const r=await fetch('/api/tts/settings',{method:'POST',headers:{'Content-Type':'application/json'},
@@ -1365,8 +1740,30 @@ $('logToggle').onclick = ()=>{
   syncDockPad();
 };
 window.addEventListener('resize', syncDockPad);
+// Drag the top edge of the log dock to resize its height (persisted).
+(function(){
+  const dock=$('logdock'), body=$('logbody'), handle=$('logResize');
+  const KEY='wsai_logH';
+  const maxH = ()=> Math.round(window.innerHeight*0.82);
+  function setH(px){ const h=Math.max(80, Math.min(maxH(), Math.round(px)));
+    body.style.height=h+'px'; localStorage.setItem(KEY, h); syncDockPad(); }
+  setH(parseInt(localStorage.getItem(KEY)||'180', 10));
+  let startY=0, startH=0, dragging=false;
+  handle.addEventListener('pointerdown', (e)=>{
+    if(dock.classList.contains('collapsed')) return;
+    dragging=true; startY=e.clientY; startH=body.offsetHeight;
+    try{ handle.setPointerCapture(e.pointerId); }catch(_){}
+    e.preventDefault();
+  });
+  handle.addEventListener('pointermove', (e)=>{ if(dragging) setH(startH + (startY - e.clientY)); });
+  const end=(e)=>{ if(!dragging) return; dragging=false; try{ handle.releasePointerCapture(e.pointerId); }catch(_){} };
+  handle.addEventListener('pointerup', end);
+  handle.addEventListener('pointercancel', end);
+})();
 $('logSearch').oninput = renderLogs;
 $('logLevel').onchange = renderLogs;
+$('logCat').onchange = renderLogs;
+$('logTime').onchange = renderLogs;
 $('logClear').onclick = async ()=>{
   if(!confirm('로그를 모두 삭제할까요?')) return;
   try{ await fetch('/api/logs/clear',{method:'POST'}); toast('로그를 삭제했습니다'); }
@@ -1445,8 +1842,10 @@ async function openLists(){
   catch(e){ lists = {whitelistUsers:[],blacklistUsers:[],whitelistRoles:[],blacklistRoles:[]}; }
   openModal('청취 화이트/블랙리스트', '<button class="btn primary" id="lstSave">저장</button>');
   $('modalBody').innerHTML =
-    '<p class="modal-note">화이트리스트에 넣으면 그 대상만 청취(비어있으면 전체 청취), 블랙리스트는 제외됩니다. 유저/역할별로 추가할 수 있어요.</p>'
+    '<p class="modal-note">화이트리스트에 넣으면 그 대상만 청취(비어있으면 전체 청취), 블랙리스트는 제외됩니다. 유저·봇·통화방·역할별로 추가할 수 있어요.</p>'
-    +'<div class="lst-row"><select id="lstType"><option value="user">유저</option><option value="role">역할</option></select>'
+    +'<div class="lst-row"><select id="lstType">'
+    +'<option value="user">유저</option><option value="bot">봇</option>'
+    +'<option value="voice">통화방</option><option value="role">역할</option></select>'
     +'<input id="lstSearch" class="lst-search" placeholder="이름으로 검색"></div>'
     +'<div class="lst-results" id="lstResults"></div>'
     +'<div class="lst-h">화이트리스트 (그 대상만 청취)</div><div class="chips" id="chipsW"></div>'
@@ -1454,15 +1853,39 @@ async function openLists(){
   const has=(arr,id)=>(arr||[]).some(x=>x.id===id);
   function add(kind,item){ const k=LKEY[kind]; if(!has(lists[k],item.id)) lists[k].push(item); renderChips(); }
   function rm(k,id){ lists[k]=(lists[k]||[]).filter(x=>x.id!==id); renderChips(); }
+  const NO_RES='<div class="lst-item"><span class="rl">결과 없음 · 봇이 아는 멤버/역할만 검색됩니다</span></div>';
+  function memberRow(m){
+    return '<div class="lst-item"><span class="nm">'+esc(m.name)+(m.bot?' <span class="rl">(봇)</span>':'')+'</span>'
+      +'<button class="mini w" data-k="wu" data-id="'+esc(m.id)+'" data-nm="'+esc(m.name)+'">+화이트</button>'
+      +'<button class="mini b" data-k="bu" data-id="'+esc(m.id)+'" data-nm="'+esc(m.name)+'">+블랙</button></div>';
+  }
   function renderResults(){
     const type=$('lstType').value, q=$('lstSearch').value.trim().toLowerCase();
-    const src = type==='user' ? (g.members||[]) : (g.roles||[]);
+    const members=(g.members||[]); const byId={}; members.forEach(m=>byId[m.id]=m);
-    const rows = src.filter(x=>!q || (x.name||'').toLowerCase().includes(q)).slice(0,100);
+    const match=(n)=>!q||((n||'').toLowerCase().includes(q));
-    $('lstResults').innerHTML = rows.length ? rows.map(x=>
+    let html='';
-      '<div class="lst-item"><span class="nm">'+esc(x.name)+(x.bot?' <span class="rl">(봇)</span>':'')+'</span>'
+    if(type==='user' || type==='bot'){
-      +'<button class="mini w" data-k="'+(type==='user'?'wu':'wr')+'" data-id="'+esc(x.id)+'" data-nm="'+esc(x.name)+'">+화이트</button>'
+      const rows=members.filter(m=>(type==='bot'?m.bot:!m.bot)&&match(m.name)).slice(0,300);
-      +'<button class="mini b" data-k="'+(type==='user'?'bu':'br')+'" data-id="'+esc(x.id)+'" data-nm="'+esc(x.name)+'">+블랙</button></div>'
+      html = rows.length ? rows.map(memberRow).join('') : NO_RES;
-    ).join('') : '<div class="lst-item"><span class="rl">결과 없음 · 봇이 아는 멤버/역할만 검색됩니다</span></div>';
+    } else if(type==='voice'){
+      const vm=((botState&&botState.members)||[]).map(m=>({id:m.id,name:m.name,bot:!!(byId[m.id]&&byId[m.id].bot)}));
+      const rows=vm.filter(m=>match(m.name)).slice(0,300);
+      html = rows.length ? rows.map(memberRow).join('')
+        : '<div class="lst-item"><span class="rl">봇이 통화방에 없거나 참여자가 없습니다</span></div>';
+    } else { // role: 역할 등록 + 펼치면 역할 아래 유저(봇 포함) 등록
+      const roles=(g.roles||[]).filter(r=>match(r.name)).slice(0,300);
+      html = roles.length ? roles.map(r=>{
+        const rid=esc(r.id), rmem=members.filter(m=>(m.roleIds||[]).includes(r.id));
+        return '<div class="lst-role">'
+          +'<div class="lst-item"><span class="nm role-h" data-rid="'+rid+'"><b class="rcaret">▸</b> '+esc(r.name)+' <span class="rl">('+rmem.length+'명)</span></span>'
+          +'<button class="mini w" data-k="wr" data-id="'+rid+'" data-nm="'+esc(r.name)+'">+화이트</button>'
+          +'<button class="mini b" data-k="br" data-id="'+rid+'" data-nm="'+esc(r.name)+'">+블랙</button></div>'
+          +'<div class="role-members" data-rid="'+rid+'" style="display:none">'
+          +(rmem.length? rmem.map(memberRow).join('') : '<div class="lst-item"><span class="rl">이 역할의 멤버 없음</span></div>')
+          +'</div></div>';
+      }).join('') : NO_RES;
+    }
+    $('lstResults').innerHTML = html;
   }
   const chip=(k,cls,x)=>'<span class="chip '+cls+'">'+esc(x.name)+' <button data-k="'+k+'" data-id="'+esc(x.id)+'">✕</button></span>';
   const empty='<span class="rl" style="color:var(--muted);font-size:12px">비어있음</span>';
@@ -1471,7 +1894,13 @@ async function openLists(){
     $('chipsB').innerHTML = [...lists.blacklistUsers.map(x=>chip('blacklistUsers','b',x)),...lists.blacklistRoles.map(x=>chip('blacklistRoles','b',x))].join('') || empty;
   }
   $('lstType').onchange=renderResults; $('lstSearch').oninput=renderResults;
-  $('lstResults').onclick=(e)=>{ const b=e.target.closest('.mini'); if(!b)return; add(b.dataset.k,{id:b.dataset.id,name:b.dataset.nm}); };
+  $('lstResults').onclick=(e)=>{
+    const rh=e.target.closest('.role-h');
+    if(rh){ const box=$('lstResults').querySelector('.role-members[data-rid="'+rh.dataset.rid+'"]');
+      if(box){ const open=box.style.display==='none'; box.style.display=open?'block':'none';
+        const c=rh.querySelector('.rcaret'); if(c) c.textContent=open?'▾':'▸'; } return; }
+    const b=e.target.closest('.mini'); if(!b)return; add(b.dataset.k,{id:b.dataset.id,name:b.dataset.nm});
+  };
   $('chipsW').onclick=$('chipsB').onclick=(e)=>{ const b=e.target.closest('button'); if(!b)return; rm(b.dataset.k,b.dataset.id); };
   $('lstSave').onclick=async()=>{
     try{ await fetch('/api/bot/lists',{method:'POST',headers:{'Content-Type':'application/json'},body:JSON.stringify({guildId,lists})}); toast('청취 필터 저장됨 · 봇에 곧 반영'); closeModal(); }
@@ -1483,6 +1912,18 @@ $('listsBtn').onclick = openLists;
 
 ['tfTime','tfUser','tfGuild','tfChannel','tfText'].forEach(id=>{ const e=$(id); if(e){ e.oninput=applyTurnFilter; e.onchange=applyTurnFilter; } });
 $('tfClear').onclick=()=>{ $('tfTime').value='0'; $('tfUser').value=''; $('tfGuild').value=''; $('tfChannel').value=''; $('tfText').value=''; applyTurnFilter(); };
+// 대화 로그: 개별 ✕ 삭제 + 전체 삭제.
+$('turns').addEventListener('click', async (e)=>{
+  const b=e.target.closest('.tdel'); if(!b) return;
+  const id=Number(b.dataset.id); removeTurn(id);   // optimistic
+  try{ await fetch('/api/turns/delete',{method:'POST',headers:{'Content-Type':'application/json'},body:JSON.stringify({id})}); }catch(_){}
+});
+$('turnsClear').onclick=async ()=>{
+  if(!turns.size){ toast('삭제할 대화가 없습니다'); return; }
+  if(!confirm('대화 로그를 모두 삭제할까요?')) return;
+  try{ await fetch('/api/turns/clear',{method:'POST'}); toast('대화 로그를 모두 삭제했습니다'); }
+  catch(e){ toast('삭제 실패: '+e); }
+};
 
 async function pollBot(){
   try{ renderBot(await (await fetch('/api/bot/state')).json()); }catch(e){}

wsai/monitor.py

@@ -129,7 +129,7 @@ class Turn:
         # Guarded so the repeated _touch()/finish() calls can't double-count.
         if self.status == "error" and not self._error_logged:
             self._error_logged = True
-            self._monitor.log("error", f"대화 #{self.id} 실패: {self.error}")
+            self._monitor.log("error", f"대화 #{self.id} 실패: {self.error}", cat="TURN")
         self._touch()
 
     # -- internal --------------------------------------------------------- #
@@ -209,12 +209,17 @@ class Monitor:
             self._status["claude_output_tokens"] += int(output_tokens or 0)
         self._broadcast({"type": "status", "status": self.status_snapshot()})
 
-    def log(self, level: str, message: str) -> None:
+    def log(self, level: str, message: str, cat: str | None = None) -> None:
-        """A free-form lifecycle/error line (startup, disconnect, crash…)."""
+        """A free-form lifecycle/error line (startup, disconnect, crash…).
+
+        ``cat`` is a short category tag (READY, CONNECT, MODEL, TTS, VOICE,
+        FILTER, SETTING, TURN, BRAIN, PIPELINE, …) shown as a chip and filterable
+        on the dashboard. Defaults to the upper-cased level when omitted."""
         with self._lock:
             self._event_id += 1
             evt = {"type": "log", "id": self._event_id, "level": level,
-                   "message": message, "wall": _now_wall()}
+                   "cat": (cat or level.upper()), "message": message,
+                   "wall": _now_wall()}
             self._events.append(evt)
             if level == "error":
                 self._status["errors_total"] += 1
@@ -253,6 +258,26 @@ class Monitor:
             self._events.clear()
         self._broadcast({"type": "logs_cleared"})
 
+    def delete_turn(self, turn_id: int) -> bool:
+        """Remove one conversation turn by id (dashboard per-turn '✕')."""
+        found = False
+        with self._lock:
+            for t in list(self._turns):
+                if t.id == turn_id:
+                    self._turns.remove(t)
+                    found = True
+                    break
+        if found:
+            self._broadcast({"type": "turn_deleted", "id": turn_id})
+        return found
+
+    def clear_turns(self) -> None:
+        """Wipe all conversation turns (dashboard '대화 전체 삭제'). Broadcasts a
+        reset so every connected page clears its turn list too."""
+        with self._lock:
+            self._turns.clear()
+        self._broadcast({"type": "turns_cleared"})
+
     # -- turns ------------------------------------------------------------ #
     def turn(self, source: str = "voice") -> Turn:
         with self._lock:

wsai/pipeline.py

@@ -65,7 +65,7 @@ class Pipeline:
             except Exception as exc:  # a single bad frame must not kill the loop
                 log.exception("vision.describe failed")
                 if self.monitor is not None:
-                    self.monitor.log("error", f"화면 이해 실패: {exc}")
+                    self.monitor.log("error", f"화면 이해 실패: {exc}", cat="VISION")
                 continue
             await self.context.update(obs)
             log.debug("screen: %s", obs.text[:120])
@@ -87,7 +87,7 @@ class Pipeline:
         try:
             async with turn.step("화면 맥락"):
                 screen = await self.context.latest()
-            async with turn.step("두뇌(생각)"):
+            async with turn.step("LLM(생각)"):
                 reply = await self.brain.respond(utt.text, screen, self._history)
             turn.replied(reply.text)
             self._remember(utt.text, reply.text)
@@ -122,7 +122,7 @@ class Pipeline:
             return
         if self.monitor is not None:
             self.monitor.set_status(listening=True)
-            self.monitor.log("info", "음성 수신 시작 — 발화 대기 중")
+            self.monitor.log("info", "음성 수신 시작 — 발화 대기 중", cat="PIPELINE")
         try:
             async for utt in self.stt.utterances():
                 await self._handle(utt)
@@ -157,7 +157,7 @@ class Pipeline:
             except Exception as exc:  # a warm failure must not abort startup
                 log.warning("prewarm %s failed: %s", name, exc)
                 if self.monitor is not None:
-                    self.monitor.log("error", f"{name} 예열 실패: {exc}")
+                    self.monitor.log("error", f"{name} 예열 실패: {exc}", cat="READY")
 
     async def run(self) -> None:
         # A TaskGroup (not bare gather) so that if ONE loop raises, the others
@@ -167,7 +167,7 @@ class Pipeline:
         # still-live loop (close-during-use).
         if self.monitor is not None:
             self.monitor.set_status(running=True)
-            self.monitor.log("info", "파이프라인 시작")
+            self.monitor.log("info", "파이프라인 시작", cat="PIPELINE")
         await self._prewarm()
         try:
             async with asyncio.TaskGroup() as tg:
@@ -177,12 +177,12 @@ class Pipeline:
         except* Exception as eg:
             if self.monitor is not None:
                 for exc in eg.exceptions:
-                    self.monitor.log("error", f"루프 예외: {type(exc).__name__}: {exc}")
+                    self.monitor.log("error", f"루프 예외: {type(exc).__name__}: {exc}", cat="PIPELINE")
             raise
         finally:
             if self.monitor is not None:
                 self.monitor.set_status(running=False, listening=False)
-                self.monitor.log("info", "파이프라인 종료")
+                self.monitor.log("info", "파이프라인 종료", cat="PIPELINE")
             await self.aclose()
 
     async def aclose(self) -> None:

wsai/state_store.py

@@ -0,0 +1,61 @@
+"""Tiny JSON state store so dashboard/bot settings survive a restart.
+
+Persists things the user configures on the dashboard — per-guild listen lists,
+bot behaviour toggles, TTS controls, and the chosen STT/LLM models — to a single
+JSON file so a service or container restart keeps them.
+
+Path: ``WSAI_STATE_FILE`` env, else ``~/.config/wsai/state.json``. For a
+container deployment, mount that path (or point the env at a mounted volume) to
+keep the file across ``docker restart``.
+
+Pure stdlib, thread-safe, best-effort: a read/write failure never raises into
+the caller (the dashboard must keep working even if the disk is unwritable).
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import os
+import threading
+from typing import Any
+
+log = logging.getLogger("wsai.state")
+
+_LOCK = threading.Lock()
+
+
+def path() -> str:
+    return os.environ.get("WSAI_STATE_FILE") or os.path.expanduser("~/.config/wsai/state.json")
+
+
+def load() -> dict[str, Any]:
+    try:
+        with open(path(), encoding="utf-8") as f:
+            data = json.load(f)
+            return data if isinstance(data, dict) else {}
+    except FileNotFoundError:
+        return {}
+    except Exception as exc:  # noqa: BLE001 — corrupt/unreadable state must not crash startup
+        log.warning("state load failed (%s): %s", path(), exc)
+        return {}
+
+
+def save(state: dict[str, Any]) -> None:
+    p = path()
+    try:
+        with _LOCK:
+            os.makedirs(os.path.dirname(p), exist_ok=True)
+            tmp = p + ".tmp"
+            with open(tmp, "w", encoding="utf-8") as f:
+                json.dump(state, f, ensure_ascii=False, indent=2)
+            os.replace(tmp, p)  # atomic
+    except Exception as exc:  # noqa: BLE001
+        log.warning("state save failed (%s): %s", p, exc)
+
+
+def patch(key: str, value: Any) -> None:
+    """Read-modify-write one top-level key."""
+    s = load()
+    s[key] = value
+    save(s)