fix(controlBrowser): never report moveMouse/search success without a real move

moveMouse returned ok:true even when humanHover did nothing (no on-screen box)
or the selector never matched — recreating the "claims it moved but didn't"
bug. Now: humanHover returns a boolean (and brings the element into view first);
moveMouse returns ok:false when the target isn't found or has no on-screen box,
and when site=naver/... whose box isn't on the current page it navigates to the
site home first before hovering. search now reports input=human|api-fallback|api
so a silent fallback to cursor-less DOM input is visible, and the tool surfaces
that note in the reply instead of implying a human-like search happened.

.env.example

@@ -59,11 +59,15 @@ OLLAMA_BASE_URL=http://127.0.0.1:11434
 # free-form chit-chat. Swap back to qwen3:8b for the strongest tool-calling.
 OLLAMA_CHAT_MODEL=qwen2.5:3b
 # Model for the auxiliary small-model calls: intent judge, tool router, weather
-# place extraction, query decomposition. BLANK (default) reuses OLLAMA_CHAT_MODEL
-# so the stack runs on one already-warm model. The code's built-in default
-# (gemma4:e2b) is NOT pulled by this stack, so leaving this unset previously made
-# every router/extractor call silently fail. Only set this if you also pull the
-# model into Ollama.
+# place extraction, query decomposition. These are classification/JSON calls,
+# NOT the spoken answer, so a small fast model here cuts 2-3 big-model round
+# trips per command without touching answer quality. BLANK uses the stack
+# default qwen2.5:3b, which ollama-init pulls automatically. Set it equal to
+# OLLAMA_CHAT_MODEL to run everything on one resident model instead (saves VRAM
+# at the cost of slower routing when the chat model is large).
+# NEVER set this LARGER than OLLAMA_CHAT_MODEL: the auxiliary calls would then
+# run on the bigger, slower model and add latency to every command (the exact
+# opposite of the split's purpose). Keep it <= the chat model, blank, or equal.
 OLLAMA_INTENT_MODEL=
 OLLAMA_EMBED_MODEL=nomic-embed-text
 WHISPER_MODEL=small
@@ -226,6 +230,7 @@ COMPOSE_FILE=docker-compose.yml:docker-compose.gpu-linux.yml
 # OLLAMA_CHAT_MODEL=qwen2.5:7b   # quality (needs ~5GB VRAM + whisper small)
 # OLLAMA_CHAT_MODEL=qwen2.5:3b   # speed (fits easily, faster on 8GB GPUs)
 # WHISPER_MODEL=small            # small frees VRAM for a bigger LLM; medium=more accurate
+# STT_BEAM_SIZE=5                # beam search (5) > greedy (1) for accuracy; lower for speed
 # MELO_DEVICE=cuda               # cpu if no GPU on the bot host
 
 # --- Settings web UI (http://localhost:8765/settings on the bot host) ---

README.md

@@ -87,7 +87,7 @@ COMPOSE_FILE=docker-compose.yml;docker-compose.gpu-windows.yml
 > Linux와 Windows는 GPU를 컨테이너에 넣는 방식이 달라서 override 파일이 갈립니다. Linux는 CDI(`devices: nvidia.com/gpu=all`), Windows(Docker Desktop)는 Compose의 `deploy.resources.reservations.devices`(`driver: nvidia`)를 씁니다. 호스트 사전 준비는 아래 "GPU 가속" 절 참고.
 
 `docker compose up` 한 번이면 자동으로:
-- Ollama 서버가 뜨고, `ollama-init`이 채팅/임베딩 모델을 **자동 pull**
+- Ollama 서버가 뜨고, `ollama-init`이 채팅/보조(의도·라우팅)/임베딩 모델을 **자동 pull** (보조 모델 `OLLAMA_INTENT_MODEL`은 기본 `qwen2.5:3b`로, 큰 채팅 모델은 답변에만 쓰고 내부 분류 호출은 이 작은 모델이 처리)
 - VNC+XFCE 데스크톱 + 크롬 + Python 브릿지가 기동
 - Whisper STT 모델 / Piper TTS 음성 자동 다운로드(볼륨에 캐시)
 

agents/llm.md

@@ -1,7 +1,7 @@
 # 자비스 운영자 지시
 
 - 너의 이름은 자비스다.
-- 모든 답변은 음성(TTS)으로 읽혀 나간다. 그러니 최대한 간결하게, 한두 문장으로 답한다. 목록, 마크다운, 이모지, 그리고 소리 내어 읽기 어려운 특수문자는 쓰지 않는다.
+- 모든 답변은 음성(TTS)으로 읽혀 나간다. 그러니 무조건 한 문장으로만 답한다. 두 문장 이상 쓰지 않는다. 목록, 마크다운, 이모지, 그리고 소리 내어 읽기 어려운 특수문자는 쓰지 않는다.
 - 정해진 문구에만 반응하지 말고, 실제 사람처럼 말의 뉘앙스와 맥락으로 의도를 알아듣고 처리한다.
 
 화면 속 크롬(방송 화면)에서 유튜브를 다룰 때 (화면에 보여야 하므로 반드시 on-screen 브라우저 제어 도구로 수행한다):

bot/scripts/stream-test/browse-search.mjs

@@ -2,10 +2,11 @@
 // 9222) so the action is visible on the Go-Live broadcast, and prints a JSON
 // result on stdout for the Python `browseAndSearch` tool to wrap.
 //
-//   node browse-search.mjs "<query>" [search|youtube]
+//   node browse-search.mjs "<query>" [search|youtube] [index]
 //
 // - search  : Google-search the query, return the top organic results.
-// - youtube : search YouTube and play the first result.
+// - youtube : search YouTube and play a result. `index` is the 1-based position
+//             from the top of the result list (default 1 = first result).
 //
 // Backend selection for `search`:
 //   1. The broadcast Chrome over CDP (visible on the Go-Live stream).
@@ -29,6 +30,9 @@ const UA =
   '(KHTML, like Gecko) Chrome/148.0.0.0 Safari/537.36';
 const query = process.argv[2] || '';
 const mode = (process.argv[3] || 'search').toLowerCase();
+// 1-based position of the YouTube result to play, counted from the top of the
+// list. Defaults to 1 (first result). Anything <1 or non-numeric falls back to 1.
+const playIndex = Math.max(1, parseInt(process.argv[4], 10) || 1);
 const out = (o) => { process.stdout.write(JSON.stringify(o)); };
 
 if (!query) { out({ ok: false, error: 'no query' }); process.exit(1); }
@@ -105,15 +109,21 @@ try {
   await page.bringToFront().catch(() => {});
 
   if (mode === 'youtube') {
-    // Type into YouTube's search box like a person, then play the first result.
+    // Type into YouTube's search box like a person, then play the requested
+    // result (the Nth from the top of the list; default the first).
     await typeSearch('https://www.youtube.com/?hl=ko', 'input#search, input[name="search_query"]', query);
     await page.waitForSelector('ytd-video-renderer a#video-title, a#video-title', { timeout: 20000 });
-    const first = page.locator('ytd-video-renderer a#video-title, a#video-title').first();
-    const title = (await first.getAttribute('title').catch(() => '')) || (await first.innerText().catch(() => ''));
-    await first.click();
+    const results = page.locator('ytd-video-renderer a#video-title, a#video-title');
+    // Clamp to what's actually on the page so "play the 5th" still plays the
+    // last available result rather than failing when fewer were returned.
+    const available = await results.count();
+    const targetIdx = Math.min(playIndex, Math.max(available, 1)) - 1;
+    const target = results.nth(targetIdx);
+    const title = (await target.getAttribute('title').catch(() => '')) || (await target.innerText().catch(() => ''));
+    await target.click();
     await page.waitForSelector('#movie_player', { timeout: 20000 });
     await page.evaluate(() => { const v = document.querySelector('video'); if (v && v.paused) v.play(); });
-    out({ ok: true, mode, title: (title || '').trim(), url: page.url() });
+    out({ ok: true, mode, index: targetIdx + 1, title: (title || '').trim(), url: page.url() });
   } else {
     // Type into Google's search box like a person, then read the results.
     await typeSearch('https://www.google.com/?hl=ko', 'textarea[name="q"], input[name="q"]', query);

bot/scripts/stream-test/chrome-control.mjs

@@ -17,6 +17,7 @@
 //   status | listTabs
 //   navigate {url} | back | forward | refresh
 //   newTab {url?} | closeTab {index?} | activateTab {index} | closePopups
+//   moveMouse {selector | site}  (hover the real cursor, no click)
 //   click {selector} | type {text, selector?} | scroll {dir, notches?}
 //   pressKey {key} | screenshot {path}
 import { chromium } from 'playwright';
@@ -40,6 +41,15 @@ if (!action) { out({ ok: false, error: 'no action' }); process.exit(1); }
 
 const norm = (u) => (/^https?:\/\//i.test(u) ? u : `https://${u}`);
 
+// Per-site homepage + search-box selector, shared by `search` and `moveMouse`.
+const SITES = {
+  naver: { home: 'https://www.naver.com', box: '#query, input[name="query"]' },
+  google: { home: 'https://www.google.com', box: 'textarea[name="q"], input[name="q"]' },
+  daum: { home: 'https://www.daum.net', box: '#q, input[name="q"]' },
+  youtube: { home: 'https://www.youtube.com', box: 'input#search, input[name="search_query"]' },
+  bing: { home: 'https://www.bing.com', box: '#sb_form_q, input[name="q"]' },
+};
+
 // The genuinely-active tab is the one whose document is visible. Playwright has
 // no "active page" accessor over CDP, so probe visibilityState (fixes treating
 // tab 0 as active and breaking sequential ops on a specific tab).
@@ -103,13 +113,6 @@ try {
       const q = String(cmd.query || '').trim();
       if (!q) throw new Error('search: no query');
       const siteKey = String(cmd.site || 'google').toLowerCase();
-      const SITES = {
-        naver: { home: 'https://www.naver.com', box: '#query, input[name="query"]' },
-        google: { home: 'https://www.google.com', box: 'textarea[name="q"], input[name="q"]' },
-        daum: { home: 'https://www.daum.net', box: '#q, input[name="q"]' },
-        youtube: { home: 'https://www.youtube.com', box: 'input#search, input[name="search_query"]' },
-        bing: { home: 'https://www.bing.com', box: '#sb_form_q, input[name="q"]' },
-      };
       const s = SITES[siteKey] || SITES.google;
       await front(page);
       // 1) Go to the homepage.
@@ -122,23 +125,31 @@ try {
       // 2) Click the on-page search box, type the query, submit.
       const box = page.locator(s.box).first();
       await box.waitFor({ state: 'visible', timeout: 15000 }).catch(() => {});
+      // Report which input path actually ran: 'human' = real xdotool cursor
+      // move + char typing; 'api-fallback' = the humanClick path threw and we
+      // fell back to cursor-less DOM click/fill; 'api' = no xdotool at all. This
+      // makes "did the cursor really move" verifiable from the result.
+      let searchInput;
       if (HAS_XDOTOOL && cmd.human !== false) {
         try {
           await human.humanClick(page, box);
           await human.humanType(q);
           await human.pressKey('Return');
+          searchInput = 'human';
         } catch {
+          searchInput = 'api-fallback';
           await box.click().catch(() => {});
           await box.fill(q).catch(() => {});
           await page.keyboard.press('Enter').catch(() => {});
         }
       } else {
+        searchInput = 'api';
         await box.click().catch(() => {});
         await box.fill(q);
         await page.keyboard.press('Enter');
       }
       await page.waitForLoadState('domcontentloaded').catch(() => {});
-      out({ ok: true, site: SITES[siteKey] ? siteKey : 'google', query: q, url: page.url(), title: await page.title().catch(() => '') });
+      out({ ok: true, site: SITES[siteKey] ? siteKey : 'google', query: q, url: page.url(), title: await page.title().catch(() => ''), input: searchInput });
       break;
     }
 
@@ -205,6 +216,45 @@ try {
       break;
     }
 
+    case 'moveMouse': {
+      // Move/hover the REAL cursor onto an element WITHOUT clicking. Target is a
+      // CSS selector, or site=naver/google/... for that site's search box.
+      // Only meaningful with xdotool (the visible cursor); with no xdotool there
+      // is no cursor to move, so report that rather than faking success. Every
+      // failure to actually move (no xdotool, selector never matches, element
+      // has no on-screen box) returns ok:false — we must never claim the cursor
+      // moved when it did not (the exact bug the user reported).
+      const siteKey = String(cmd.site || '').toLowerCase();
+      const selector = String(cmd.selector || '').trim() || (SITES[siteKey] ? SITES[siteKey].box : '');
+      if (!selector) throw new Error('moveMouse: no selector or known site');
+      if (!(HAS_XDOTOOL && cmd.human !== false)) {
+        out({ ok: false, error: 'no xdotool: cannot move the visible cursor on this host' });
+        break;
+      }
+      await front(page);
+      let locator = page.locator(selector).first();
+      let visible = await locator.waitFor({ state: 'visible', timeout: 8000 }).then(() => true).catch(() => false);
+      // A named site whose search box isn't on the current page: go to its home
+      // first (real omnibox), then target the box there.
+      if (!visible && SITES[siteKey]) {
+        try { await human.navigateOmnibox(SITES[siteKey].home); await page.waitForLoadState('domcontentloaded').catch(() => {}); }
+        catch { await page.goto(SITES[siteKey].home, { waitUntil: 'domcontentloaded' }).catch(() => {}); }
+        locator = page.locator(SITES[siteKey].box).first();
+        visible = await locator.waitFor({ state: 'visible', timeout: 8000 }).then(() => true).catch(() => false);
+      }
+      if (!visible) {
+        out({ ok: false, error: `moveMouse: target not found (${cmd.selector || siteKey})` });
+        break;
+      }
+      const moved = await human.humanHover(page, locator);
+      if (!moved) {
+        out({ ok: false, error: 'moveMouse: element has no on-screen box; cursor not moved' });
+        break;
+      }
+      out({ ok: true, target: cmd.selector || siteKey, input: 'human' });
+      break;
+    }
+
     case 'click': {
       const selector = String(cmd.selector || '').trim();
       if (!selector) throw new Error('click: no selector');

bot/scripts/stream-test/human.mjs

@@ -136,14 +136,18 @@ export async function navigateOmnibox(text) {
 }
 
 // Move the real cursor over an element (hover, no click) - e.g. to reveal a
-// video player's controls or to focus it for a keyboard shortcut.
+// video player's controls or to focus it for a keyboard shortcut. Returns true
+// only if the element had an on-screen box and the cursor was actually moved;
+// returns false when there is nothing to move to (so callers must not report
+// success). Brings the element into view with a real wheel scroll first.
 export async function humanHover(page, locator) {
-  const box = await locator.boundingBox().catch(() => null);
-  if (!box) return;
+  const box = await bringIntoView(page, locator);
+  if (!box) return false;
   const g = await page.evaluate(() => ({ sx: window.screenX, sy: window.screenY, ow: window.outerWidth, oh: window.outerHeight, iw: window.innerWidth, ih: window.innerHeight }));
   const bx = Math.max(0, Math.round((g.ow - g.iw) / 2));
   const oy = g.sy + Math.max(0, g.oh - g.ih - bx);
   await humanMove(Math.round(g.sx + bx + box.x + box.width * 0.5), Math.round(oy + box.y + box.height * 0.4));
+  return true;
 }
 
 export { sleep, rand };

bridge/server.py

@@ -87,6 +87,17 @@ VAD_MIN_SPEECH_MS = int(os.environ.get("VAD_MIN_SPEECH_MS", "200"))
 # Korean phrase decoded as Chinese) and shaves a little latency. Empty = auto.
 STT_LANGUAGE = os.environ.get("STT_LANGUAGE", "ko").strip() or None
 
+# Whisper decoding accuracy knobs. beam_size=1 is greedy decoding — fast but the
+# least accurate; beam search (5 is the Whisper default) explores alternatives
+# and noticeably improves recognition on short, accented, or noisy Discord-mic
+# speech. condition_on_previous_text=False stops Whisper from feeding a previous
+# clip's transcript back in as a prompt, which on isolated short utterances
+# causes repetition loops and drift rather than helping. Both are env-tunable so
+# accuracy/latency can be traded without a code change (lower STT_BEAM_SIZE for
+# speed, raise it for accuracy).
+STT_BEAM_SIZE = max(1, int(os.environ.get("STT_BEAM_SIZE", "5")))
+STT_CONDITION_ON_PREV = os.environ.get("STT_CONDITION_ON_PREV", "0") in ("1", "true", "True", "yes", "on")
+
 # TTS engine: "edge" (Microsoft Edge TTS, natural Korean neural voice) is the
 # primary voice. "melo" (a warm MeloTTS worker) and "piper" remain selectable.
 def _tts_engine_setting() -> str:
@@ -243,7 +254,12 @@ def transcribe(wav_bytes: bytes) -> dict:
         print("[bridge] no speech detected (VAD) — skipping STT", flush=True)
         return {"text": "", "language": None, "note": "음성 아님(VAD 차단)"}
 
-    segments, info = _whisper.transcribe(audio, beam_size=1, language=STT_LANGUAGE)
+    segments, info = _whisper.transcribe(
+        audio,
+        beam_size=STT_BEAM_SIZE,
+        language=STT_LANGUAGE,
+        condition_on_previous_text=STT_CONDITION_ON_PREV,
+    )
     # Second line of defence: drop non-speech / hallucinated segments by
     # Whisper's own no_speech_prob. The no_speech_prob hard cutoff (plus the VAD
     # pre-gate above) is what rejects noise/hallucinations. The avg_logprob

docker-compose.yml

@@ -40,6 +40,9 @@ services:
     environment:
       OLLAMA_HOST: http://ollama:11434
       CHAT_MODEL: ${OLLAMA_CHAT_MODEL:-qwen2.5:3b}
+      # Small auxiliary model for intent/router/extraction calls (see javis
+      # service). Pulled here so the split is ready out of the box.
+      INTENT_MODEL: ${OLLAMA_INTENT_MODEL:-qwen2.5:3b}
       EMBED_MODEL: ${OLLAMA_EMBED_MODEL:-nomic-embed-text}
     entrypoint: ["/bin/sh", "-c"]
     command:
@@ -48,6 +51,10 @@ services:
         until ollama list >/dev/null 2>&1; do sleep 2; done;
         echo "[ollama-init] pulling $$CHAT_MODEL";
         ollama pull "$$CHAT_MODEL";
+        if [ -n "$$INTENT_MODEL" ] && [ "$$INTENT_MODEL" != "$$CHAT_MODEL" ]; then
+          echo "[ollama-init] pulling $$INTENT_MODEL (auxiliary intent/router model)";
+          ollama pull "$$INTENT_MODEL";
+        fi;
         echo "[ollama-init] pulling $$EMBED_MODEL";
         ollama pull "$$EMBED_MODEL";
         echo "[ollama-init] models ready.";
@@ -62,6 +69,14 @@ services:
       # Point the brain at the ollama service and the bot at the in-container bridge.
       OLLAMA_BASE_URL: http://ollama:11434
       OLLAMA_CHAT_MODEL: ${OLLAMA_CHAT_MODEL:-qwen2.5:3b}
+      # Auxiliary small-model calls (intent judge, tool router, arg extraction,
+      # query decomposition) run on this fast model so the big chat model only
+      # runs for the actual spoken answer. With the GPU on, this is the main
+      # per-turn latency win: a command no longer pays the big model's cost 2-3
+      # times for routing/extraction. Defaults to qwen2.5:3b (the project's
+      # reference small model, clean Korean on classification); set it equal to
+      # OLLAMA_CHAT_MODEL to fold everything back onto one resident model.
+      OLLAMA_INTENT_MODEL: ${OLLAMA_INTENT_MODEL:-qwen2.5:3b}
       OLLAMA_EMBED_MODEL: ${OLLAMA_EMBED_MODEL:-nomic-embed-text}
       WHISPER_MODEL: ${WHISPER_MODEL:-medium}
       WHISPER_DEVICE: ${WHISPER_DEVICE:-cuda}
@@ -82,6 +97,9 @@ services:
       PLANNER_ENABLED: ${PLANNER_ENABLED:-0}
       # Lock STT to Korean (skip Whisper auto-detect).
       STT_LANGUAGE: ${STT_LANGUAGE:-ko}
+      # Whisper decode accuracy: beam search (5) over greedy (1) lifts recognition
+      # on short/noisy Discord speech. Lower to 1 for minimum latency.
+      STT_BEAM_SIZE: ${STT_BEAM_SIZE:-5}
       VOICE_SILENCE_MS: ${VOICE_SILENCE_MS:-600}
       BRIDGE_URL: http://127.0.0.1:8765
       # Split-deployment role: full (default, all-in-one), browser (only the

docker/entrypoint.sh

@@ -10,12 +10,15 @@ set -euo pipefail
 : "${OLLAMA_BASE_URL:=http://ollama:11434}"
 : "${OLLAMA_CHAT_MODEL:=qwen3:8b}"
 # Auxiliary small-model calls (intent judge, tool router, weather place
-# extraction, query decomposition). The code default is gemma4:e2b, which this
-# stack does not pull, so those calls would silently fail and fall open —
-# crippling tool routing and arg extraction. Reuse the (already warm) chat model
-# by default so everything runs on one resident model; override if you pull a
-# dedicated small model.
-: "${OLLAMA_INTENT_MODEL:=${OLLAMA_CHAT_MODEL}}"
+# extraction, query decomposition). Default to a small fast model so the big
+# chat model only runs for the actual spoken answer — the main per-turn latency
+# win once the GPU is in use, since the 2-3 routing/extraction calls per command
+# no longer pay the big model's cost. ollama-init pulls this model. Set it equal
+# to OLLAMA_CHAT_MODEL to fold everything back onto one resident model.
+: "${OLLAMA_INTENT_MODEL:=qwen2.5:3b}"
+# Cap chat-model output tokens per turn (worst-case latency guard). Spoken
+# answers are 1-2 sentences; 512 is safe headroom above tool-call JSON. 0 = off.
+: "${OLLAMA_NUM_PREDICT:=512}"
 : "${OLLAMA_EMBED_MODEL:=nomic-embed-text}"
 : "${WHISPER_MODEL:=small}"
 : "${WHISPER_DEVICE:=cuda}"
@@ -32,7 +35,7 @@ set -euo pipefail
 : "${XDG_RUNTIME_DIR:=/run/user/0}"
 : "${PULSE_SERVER:=unix:${XDG_RUNTIME_DIR}/pulse/native}"
 
-export VNC_RESOLUTION OLLAMA_BASE_URL OLLAMA_CHAT_MODEL OLLAMA_INTENT_MODEL OLLAMA_EMBED_MODEL \
+export VNC_RESOLUTION OLLAMA_BASE_URL OLLAMA_CHAT_MODEL OLLAMA_NUM_PREDICT OLLAMA_INTENT_MODEL OLLAMA_EMBED_MODEL \
        WHISPER_MODEL WHISPER_DEVICE WHISPER_COMPUTE_TYPE JARVIS_DB_PATH \
        PIPER_VOICE PIPER_VOICE_DIR TTS_PIPER_MODEL_PATH BRIDGE_HOST BRIDGE_PORT \
        XDG_RUNTIME_DIR PULSE_SERVER

docker/jarvis-config.template.json

@@ -4,6 +4,7 @@
   "ollama_base_url": "${OLLAMA_BASE_URL}",
   "ollama_embed_model": "${OLLAMA_EMBED_MODEL}",
   "ollama_chat_model": "${OLLAMA_CHAT_MODEL}",
+  "ollama_num_predict": "${OLLAMA_NUM_PREDICT}",
   "intent_judge_model": "${OLLAMA_INTENT_MODEL}",
   "tts_enabled": true,
   "tts_engine": "${TTS_ENGINE}",

docs/llm_contexts.md

@@ -19,14 +19,14 @@ Every distinct LLM call in Jarvis, what feeds it, what consumes it, and how it i
   - Time + location context (re-injected each turn)
   - Tool schema: native via `generate_tools_json_schema()` ([src/jarvis/tools/registry.py](src/jarvis/tools/registry.py)) or text fallback via `_text_tool_call_guidance()` ([engine.py:68](src/jarvis/reply/engine.py:68))
   - Tool results from prior turns (raw or digested — see #5)
-- **Output**: OpenAI-style `{content, tool_calls, thinking}`. Consumed by the tool orchestrator and TTS pipeline. Natural-language content is delivered immediately; no post-turn evaluator runs.
-- **Limits**: `num_ctx: 8192` (explicit). Timeout `llm_chat_timeout_sec` (45s). Auto-fallback from native to text tool-calls on HTTP 400 (`ToolsNotSupportedError`), sticky for the session. Risk: `fetch_web_page` truncates at 50,000 chars (~37k tokens) — mitigated for SMALL models by tool-result digest (#5) which compresses the payload before it enters the messages history. LARGE models receive the raw payload and may silently see a truncated context.
+- **Output**: OpenAI-style `{content, tool_calls, thinking}`. Consumed by the tool orchestrator and TTS pipeline. Natural-language content is delivered immediately; no post-turn evaluator runs. Spoken-answer length: the persona (`system_prompt.py`) and `voice_style` (`prompts/system.py`) both constrain the reply to a SINGLE sentence — any dry aside must fold into that one sentence as a trailing clause, never a second sentence. This keeps TTS latency down (synth time scales with text length) and matches the `agents/llm.md` operator instruction.
+- **Limits**: `num_ctx: 8192` (explicit). Output `num_predict: cfg.ollama_num_predict` (default 512, `0`/negative disables) caps generated tokens per turn — a worst-case latency guard for short spoken answers; the headroom stays above tool-call JSON so it does not truncate tool calls (both native and text tool-call paths). Timeout `llm_chat_timeout_sec` (45s). Auto-fallback from native to text tool-calls on HTTP 400 (`ToolsNotSupportedError`), sticky for the session. Risk: `fetch_web_page` truncates at 50,000 chars (~37k tokens) — mitigated for SMALL models by tool-result digest (#5) which compresses the payload before it enters the messages history. LARGE models receive the raw payload and may silently see a truncated context.
 
 ## 2. Intent Judge
 
 - **File**: [src/jarvis/listening/intent_judge.py](src/jarvis/listening/intent_judge.py) — `IntentJudge.evaluate()`.
 - **Trigger**: on a speech segment *only if* there is an engagement signal (wake word detected, hot-window active, or TTS playing). Pure ambient speech skips it.
-- **Model / gating**: `cfg.intent_judge_model` (default `gemma4:e2b`, ~2B). Falls back to text-based wake detection if Ollama is unavailable.
+- **Model / gating**: `cfg.intent_judge_model`. Code-level default `gemma4:e2b` (~2B); the **Docker stack** renders it from `OLLAMA_INTENT_MODEL` (default `qwen2.5:3b`, pulled by `ollama-init`), kept deliberately **separate from `ollama_chat_model`** so this judge and the tool router (#3, #7) run on a small fast model while the big chat model is reserved for the spoken answer. Setting `OLLAMA_INTENT_MODEL = OLLAMA_CHAT_MODEL` folds them back onto one resident model. Falls back to text-based wake detection if Ollama is unavailable.
 - **Inputs**:
   - Rolling transcript buffer (last 120s, with timestamps)
   - Wake-word timestamp (if any), normalised aliases
@@ -246,7 +246,7 @@ user input
 3. Pre-warm the intent-judge model before TTS finishes.
 4. Cache tool-router (#7) output by query hash.
 5. Give each digest its own timeout budget rather than sharing `llm_digest_timeout_sec` (today a slow memory digest can starve the max-turn digest).
-6. Consider single-model deployments: router+planner prefer `intent_judge_model`; loading a second model hurts cold-start latency on small hardware.
+6. Two-model vs single-model tradeoff: the Docker default keeps a **separate** small `intent_judge_model` (`OLLAMA_INTENT_MODEL=qwen2.5:3b`) so routing/judging/extraction don't pay the big chat model's per-call cost — the main win once the GPU holds both models resident. On VRAM-constrained hardware, fold them onto one model by setting `OLLAMA_INTENT_MODEL = OLLAMA_CHAT_MODEL` (saves a resident model at the cost of slower routing when the chat model is large).
 7. Narrow `llm_thinking_enabled` to router/planner only, not every context.
 8. Reduce `intent_judge_timeout_sec` (15s) or race it against text-based wake detection to avoid blocking the audio loop.
 

src/jarvis/config.py

@@ -85,6 +85,12 @@ class Settings:
     llm_digest_timeout_sec: float
     llm_embedding_timeout_sec: float
     llm_profile_select_timeout_sec: float
+    # Upper bound on tokens the chat model may generate per reply turn. Spoken
+    # (TTS) answers are 1-2 sentences, so a cap bounds the worst-case latency of
+    # a model that occasionally rambles or loops without changing normal answers.
+    # The headroom (default 512) sits well above this app's short tool-call JSON,
+    # so capping never truncates a tool call. 0 (or negative) disables the cap.
+    ollama_num_predict: int
 
     # Profiles & Behavior
     active_profiles: list[str]
@@ -394,6 +400,9 @@ def get_default_config() -> Dict[str, Any]:
         "llm_digest_timeout_sec": 8.0,
         "llm_embedding_timeout_sec": 60.0,
         "llm_profile_select_timeout_sec": 30.0,
+        # Cap on chat-model output tokens per turn (worst-case latency guard).
+        # 512 is safe headroom above short TTS answers and tool-call JSON; 0 off.
+        "ollama_num_predict": 512,
 
         # Profiles & Behavior
         "active_profiles": ["developer", "business", "life"],
@@ -763,6 +772,10 @@ def load_settings() -> Settings:
     llm_digest_timeout_sec = float(merged.get("llm_digest_timeout_sec", 8.0))
     llm_embedding_timeout_sec = float(merged.get("llm_embedding_timeout_sec", 60.0))
     llm_profile_select_timeout_sec = float(merged.get("llm_profile_select_timeout_sec", 30.0))
+    try:
+        ollama_num_predict = int(merged.get("ollama_num_predict", 512))
+    except (TypeError, ValueError):
+        ollama_num_predict = 512
 
     return Settings(
         # Database & Storage
@@ -778,6 +791,7 @@ def load_settings() -> Settings:
         llm_digest_timeout_sec=llm_digest_timeout_sec,
         llm_embedding_timeout_sec=llm_embedding_timeout_sec,
         llm_profile_select_timeout_sec=llm_profile_select_timeout_sec,
+        ollama_num_predict=ollama_num_predict,
 
         # Profiles & Behavior
         active_profiles=active_profiles,

src/jarvis/reply/engine.py

@@ -2233,6 +2233,16 @@ def run_reply_engine(db: "Database", cfg, tts: Optional[Any],
                 has_tool_calls = " (has tool_calls)" if msg.get("tool_calls") else ""
                 debug_log(f"    [{i}] {role}: {content}{has_tool_calls}", "planning")
 
+        # Bound worst-case generation latency: spoken answers are 1-2 sentences,
+        # so cap the chat model's output tokens. The default headroom sits well
+        # above this app's tool-call JSON, so capping never truncates a tool
+        # call. 0/negative disables it. See config.ollama_num_predict.
+        try:
+            _num_predict = int(getattr(cfg, 'ollama_num_predict', 0) or 0)
+        except (TypeError, ValueError):
+            _num_predict = 0
+        _chat_extra_options = {"num_predict": _num_predict} if _num_predict > 0 else None
+
         # Send messages to Ollama — try native tool calling first, fall back to text-based
         # if the model returns HTTP 400 (native tools API not supported).
         _dump_tools_schema = None if use_text_tools else tools_json_schema
@@ -2242,7 +2252,7 @@ def run_reply_engine(db: "Database", cfg, tts: Optional[Any],
                 chat_model=cfg.ollama_chat_model,
                 messages=messages,
                 timeout_sec=float(getattr(cfg, 'llm_chat_timeout_sec', 45.0)),
-                extra_options=None,
+                extra_options=_chat_extra_options,
                 tools=_dump_tools_schema,
                 thinking=getattr(cfg, 'llm_thinking_enabled', False),
             )
@@ -2273,7 +2283,7 @@ def run_reply_engine(db: "Database", cfg, tts: Optional[Any],
                 chat_model=cfg.ollama_chat_model,
                 messages=messages,
                 timeout_sec=float(getattr(cfg, 'llm_chat_timeout_sec', 45.0)),
-                extra_options=None,
+                extra_options=_chat_extra_options,
                 tools=None,
                 thinking=getattr(cfg, 'llm_thinking_enabled', False),
             )

src/jarvis/reply/prompts/model_variants.py

@@ -22,11 +22,15 @@ class ModelSize(Enum):
     LARGE = "large"  # 8b+ - can infer tool usage from context
 
 
-# Model size patterns - models matching these are considered SMALL
+# Model size patterns - models matching these are considered SMALL.
+# Covers every sub-8B size (1b-7b): these models need the explicit, repeated
+# tool/greeting/instruction constraints and falter on the terse LARGE prompt.
+# Without 2b/4b/5b/6b here a genuinely small model (e.g. qwen*:4b) was
+# misclassified as LARGE and given the less-guided prompt set.
 _SMALL_MODEL_PATTERNS = (
-    ":1b", ":3b", ":7b",
-    "-1b", "-3b", "-7b",
-    "_1b", "_3b", "_7b",
+    ":1b", ":2b", ":3b", ":4b", ":5b", ":6b", ":7b",
+    "-1b", "-2b", "-3b", "-4b", "-5b", "-6b", "-7b",
+    "_1b", "_2b", "_3b", "_4b", "_5b", "_6b", "_7b",
     "gemma4",  # Gemma 4 - always small regardless of tag
 )
 

src/jarvis/reply/prompts/prompts.spec.md

@@ -4,7 +4,7 @@ This module provides model-size-aware prompt generation for the reply engine.
 
 ### Problem Statement
 
-Small models (1b, 3b, 7b parameters) lack the reasoning capacity to infer when NOT to use tools. When given prompts like "Proactively use available tools," they may incorrectly call tools for simple greetings like "hello" or "ni hao" because they cannot distinguish between:
+Small models (every sub-8B size, 1b-7b parameters) lack the reasoning capacity to infer when NOT to use tools. When given prompts like "Proactively use available tools," they may incorrectly call tools for simple greetings like "hello" or "ni hao" because they cannot distinguish between:
 - Requests that require tools (weather, search, data retrieval)
 - Simple conversation (greetings, small talk, general knowledge)
 
@@ -14,7 +14,7 @@ The module detects model size from the model name and selects appropriate prompt
 
 | Model Size | Detection Pattern | Tool Prompts |
 |------------|-------------------|--------------|
-| SMALL | `:1b`, `:3b`, `:7b`, `gemma4` | Conservative — explicit "DO NOT use tools for greetings" + worked negative examples + repetition |
+| SMALL | `:1b`-`:7b` (every size 1-7B, all separators), `gemma4` | Conservative — explicit "DO NOT use tools for greetings" + worked negative examples + repetition |
 | LARGE | All others (8b+) | Proactive — "use tools confidently" + short anti-confabulation + auto-derive clause |
 
 ### Architecture
@@ -43,7 +43,7 @@ from jarvis.reply.prompts import (
 Both model sizes share these base components:
 - `asr_note`: Voice transcription error handling
 - `inference_guidance`: Prefer inference over clarification
-- `voice_style`: Concise, conversational responses
+- `voice_style`: Single-sentence, conversational responses (spoken aloud, so one sentence only — never more)
 
 Model-size-specific components:
 - `tool_incentives`: When/how aggressively to use tools

src/jarvis/reply/prompts/system.py

@@ -26,8 +26,8 @@ INFERENCE_GUIDANCE = (
 # Voice assistant communication style - concise, conversational
 VOICE_STYLE = (
     "Keep responses concise and conversational since this is a voice assistant. "
-    "Two to three sentences maximum. Prioritize clarity and brevity - users are listening, not reading. "
-    "Avoid unnecessary elaboration unless specifically requested. "
+    "Reply in a SINGLE sentence - never more than one sentence. Prioritize clarity and brevity - users are listening, not reading. "
+    "Avoid unnecessary elaboration. "
     "Do NOT offer follow-up suggestions or ask if the user wants more info - just respond directly. "
     "IMPORTANT: Always respond in natural language - never output JSON, code, or structured data as your response. "
     "NEVER use markdown formatting in your replies: no asterisks for emphasis (**bold**, *italic*), "

src/jarvis/reply/reply.spec.md

@@ -287,6 +287,8 @@ Turn 4: LLM → {content: "Here's a comprehensive comparison of the iPhone 15 mo
   - `llm_tools_timeout_sec` (enrichment extraction)
   - `llm_embed_timeout_sec` (vector search)
   - `llm_chat_timeout_sec` (messages loop turn)
+- Output bound:
+  - `ollama_num_predict` (default `512`, `0`/negative disables) caps the chat model's generated tokens per turn via the Ollama `num_predict` option on the messages-loop call. Spoken (TTS) answers are 1-2 sentences, so this never clips a normal answer; it bounds the worst-case latency of a model that occasionally rambles or loops. The default headroom sits well above this app's short tool-call JSON, so it does not truncate tool calls. Applied uniformly to the reply loop's chat call (both native-tools and text-tools paths); the small classification passes (intent judge, digests) keep their own caps. Note: this is a worst-case guard, not the primary latency lever, which is model size and GPU residency.
 - Memory enrichment:
   - `memory_enrichment_max_results` limits recalled snippets.
   - `memory_digest_enabled` (default `null` = auto-on for SMALL models ≤7B, off for LARGE) distils the combined diary + graph dump into a short relevance-filtered note via a cheap LLM pass before injecting into the system prompt. See **Memory Digest for Small Models** below.

src/jarvis/system_prompt.py

@@ -62,10 +62,11 @@ _SYSTEM_PROMPT_TEMPLATE: str = (
     "Tone rails (hard): never mean, never condescending, never passive-aggressive, never "
     "sulking, never preachy, never sycophantic ('great question', 'I'd be happy to'). "
     "Sarcasm points at the situation, the topic, or mildly at yourself — never at the user. "
-    "Shape for casual, factual, or small-talk replies: state the answer in a sentence, then add "
-    "one short dry observation about it (an understated aside, a raised-eyebrow remark, a gentle "
-    "noticing of the irony). One aside — not two, not a joke opener, not a joke-shaped sentence "
-    "replacing the answer. The aside is a tail, not the head. "
+    "Shape for casual, factual, or small-talk replies: give the answer in a SINGLE sentence. If a "
+    "dry aside fits, fold it into that same sentence as a short trailing clause — never add it as "
+    "a second sentence, never a joke opener, never a joke-shaped sentence replacing the answer. "
+    "Whenever the wit would require a second sentence, drop the wit and keep the one-sentence "
+    "answer. The aside is a tail inside the sentence, not a head and not a new sentence. "
     "Examples of the MOVE (shape, not wording — never copy these): stating a fact and then noting "
     "its mild absurdity; giving the weather and then commenting on what it implies for the day; "
     "answering a trivia question and then offering a wry footnote about the subject; admitting "
@@ -79,8 +80,8 @@ _SYSTEM_PROMPT_TEMPLATE: str = (
     "butler clichés, and never address the user as 'sir', 'madam', 'my liege', or similar. "
     "Never stack multiple jokes in one reply. "
     "Be concise, conversational, and actionable. "
-    "This is a spoken voice assistant: answer in ONE short sentence whenever possible "
-    "(two at the very most). No lists, no preamble, no 'is there anything else' offers. "
+    "This is a spoken voice assistant: your ENTIRE reply must be a single short sentence. "
+    "Never write a second sentence. No lists, no preamble, no 'is there anything else' offers. "
     "When a controlBrowser tool is available, use IT (never webSearch) for anything that "
     "should happen in the on-screen browser — opening a site, searching on a site "
     "(controlBrowser action 'search' with the right site), clicking, typing — because only "
@@ -119,10 +120,13 @@ _SYSTEM_PROMPT_TEMPLATE: str = (
     "'tell me a joke', 'chat with me'), never reply with a bare greeting like 'Hey there!', "
     "'Hi!', 'How can I help you?', or a generic observation about an unrelated topic. "
     "When the 'Information the user has shared…' section is present, you MUST pick one concrete "
-    "fact from it and build the reply around that fact (e.g. 'You mentioned you box at Trenches "
-    "Gym — how's training going this week?'). Do not talk about things that are not in that "
-    "section. Only when that section is absent may you invent a fresh observation, question, or "
-    "joke. Produce a varied response each time — do not repeat a previous reply verbatim. "
+    "fact from it and build the reply around that fact, opening with a short natural reference to "
+    "it. CRITICAL: use ONLY names, places, activities, and details that literally appear in that "
+    "section — never borrow any name, place, or activity from these instructions or from any "
+    "example wording, and never invent specifics that are not in that section. Do not talk about "
+    "things that are not in that section. Only when that section is absent may you invent a fresh "
+    "observation, question, or joke. Produce a varied response each time — do not repeat a "
+    "previous reply verbatim. "
     "Banned phrasings: 'I can only tell you what you have shared with me in this conversation', "
     "'I don't have access to any personal information outside of what you tell me', 'I don't have "
     "personal details outside of our conversation history', 'I do not store personal details "

src/jarvis/tools/builtin/browse_and_play.py

@@ -30,8 +30,10 @@ class BrowseAndPlayTool(Tool):
     def description(self) -> str:
         return (
             "Play a song / music video / clip on the shared screen by searching YouTube "
-            "and playing the first result. Use when the user asks you to play or watch "
-            "something. Only available in screen-share mode."
+            "and playing a result. Use when the user asks you to play or watch "
+            "something. Plays the first result by default; pass 'index' to play the "
+            "Nth result from the top of the search list (e.g. 'play the 3rd video' -> "
+            "index=3). Only available in screen-share mode."
         )
 
     @property
@@ -42,7 +44,16 @@ class BrowseAndPlayTool(Tool):
                 "query": {
                     "type": "string",
                     "description": "What to play, e.g. 'IU Good Day' or 'lofi hip hop'.",
-                }
+                },
+                "index": {
+                    "type": "integer",
+                    "description": (
+                        "1-based position of the video to play in the search results, "
+                        "counted from the top of the list. Defaults to 1 (first result). "
+                        "Use for 'play the Nth video' / 'play the second one'."
+                    ),
+                    "minimum": 1,
+                },
             },
             "required": ["query"],
         }
@@ -55,18 +66,25 @@ class BrowseAndPlayTool(Tool):
                 reply_text="화면 공유 모드(STREAM_BROWSER=true)에서만 영상을 재생할 수 있습니다.",
             )
         query = ""
+        index = 1
         if args and isinstance(args, dict):
             query = str(args.get("query", "")).strip()
+            try:
+                index = int(args.get("index", 1) or 1)
+            except (TypeError, ValueError):
+                index = 1
+            if index < 1:
+                index = 1
         if not query:
             return ToolExecutionResult(success=False, reply_text="재생할 내용을 알려주세요.")
         if not _NODE_SCRIPT.exists():
             return ToolExecutionResult(success=False, reply_text="브라우저 재생 도구를 찾을 수 없습니다.")
 
-        context.user_print(f"▶️ 화면에서 '{query}' 재생 중…")
-        debug_log(f"    ▶️ browseAndPlay '{query}'", "tools")
+        context.user_print(f"▶️ 화면에서 '{query}' 재생 중… (#{index})")
+        debug_log(f"    ▶️ browseAndPlay '{query}' index={index}", "tools")
         try:
             proc = subprocess.run(
-                ["node", str(_NODE_SCRIPT), query, "youtube"],
+                ["node", str(_NODE_SCRIPT), query, "youtube", str(index)],
                 capture_output=True,
                 text=True,
                 timeout=40,

src/jarvis/tools/builtin/browse_and_play.spec.md

@@ -6,16 +6,24 @@ video, or clip.
 
 ### Behaviour
 
-- Public schema is a single required `query` string (what to play).
+- Public schema is a required `query` string (what to play) plus an optional
+  `index` integer (1-based position in the search results, counted from the top
+  of the list). `index` defaults to `1` (first result), so existing callers and
+  "play X" requests are unchanged; "play the 3rd video" / "play the second one"
+  map to `index=3` / `index=2`.
 - **Mode-gated**: only acts when `STREAM_BROWSER` is true (`cfg.stream_browser`).
   In voice-only mode (false) there is no screen to show, so it returns a short
   message and does nothing.
 - Drives the on-screen Chrome by subprocessing the Node CDP helper
-  `bot/scripts/stream-test/browse-search.mjs <query> youtube`, which searches
-  YouTube and plays the first result on display `:1`. The broadcast captures
-  that display, so the playback is what viewers see.
-- Returns `success` with the played video's title, or a failure message if the
-  helper/Chrome is unavailable. It does NOT make an LLM call.
+  `bot/scripts/stream-test/browse-search.mjs <query> youtube <index>`, which
+  searches YouTube and plays the chosen result on display `:1`. The broadcast
+  captures that display, so the playback is what viewers see.
+  - The helper clicks the `index`-th `a#video-title` in the results list. The
+    index is clamped to the number of results actually returned, so asking for a
+    position beyond the list plays the last available result rather than failing.
+- Returns `success` with the played video's title (and the resolved `index`), or
+  a failure message if the helper/Chrome is unavailable. It does NOT make an LLM
+  call.
 
 ### Principles
 

src/jarvis/tools/builtin/control_browser.py

@@ -45,9 +45,16 @@ class ControlBrowserTool(Tool):
             "Use this (NOT webSearch) whenever the user wants something done or shown IN the "
             "browser on screen: open a website or URL, search on a specific site (action "
             "'search' with site=naver/google/daum/youtube/bing), go back/forward, refresh, "
-            "manage tabs (list/new/close/switch), close popups, click, type, scroll, or "
-            "screenshot. webSearch only returns text and shows nothing on screen; this tool "
-            "actually navigates the visible browser. Only available in screen-share mode. "
+            "manage tabs (list/new/close/switch), close popups, move the mouse onto an element, "
+            "click, type, scroll, or screenshot. webSearch only returns text and shows nothing "
+            "on screen; this tool actually navigates the visible browser. "
+            "Cursor behaviour matters: 'search' and 'type' (with a selector) move the REAL mouse "
+            "cursor to the on-page box, click it, then type one character at a time — use these "
+            "when the user wants to search or type on a page. 'navigate' only types the URL into "
+            "the address bar (no mouse movement). 'moveMouse' moves/hovers the visible cursor "
+            "onto an element (selector, or site=naver/... for that site's search box) WITHOUT "
+            "clicking — use it when the user just asks to move the mouse somewhere. "
+            "Only available in screen-share mode. "
             "Never claim you did any of this unless this tool returns success."
         )
 
@@ -61,16 +68,16 @@ class ControlBrowserTool(Tool):
                     "enum": [
                         "status", "listTabs", "navigate", "search", "back",
                         "forward", "refresh", "newTab", "closeTab", "activateTab",
-                        "closePopups", "click", "type", "scroll", "pressKey",
+                        "closePopups", "moveMouse", "click", "type", "scroll", "pressKey",
                         "screenshot",
                     ],
                     "description": "What to do in the browser.",
                 },
                 "url": {"type": "string", "description": "Target URL/site for navigate/newTab (e.g. 'naver.com')."},
                 "query": {"type": "string", "description": "Search text for action 'search'."},
-                "site": {"type": "string", "description": "Search site for action 'search': naver, google, daum, youtube, bing."},
+                "site": {"type": "string", "description": "Site for action 'search' or 'moveMouse': naver, google, daum, youtube, bing. For moveMouse it targets that site's search box."},
                 "index": {"type": "integer", "description": "Tab index for closeTab/activateTab (from listTabs)."},
-                "selector": {"type": "string", "description": "CSS selector for click/type."},
+                "selector": {"type": "string", "description": "CSS selector for click/type/moveMouse."},
                 "text": {"type": "string", "description": "Text to type."},
                 "key": {"type": "string", "description": "Key to press, e.g. 'Return', 'Escape'."},
                 "dir": {"type": "string", "description": "Scroll direction: 'down' or 'up'."},
@@ -149,7 +156,12 @@ class ControlBrowserTool(Tool):
         if action == "navigate":
             return f"브라우저에서 {data.get('url', args.get('url'))} 로 이동했습니다."
         if action == "search":
-            return f"{data.get('site', '')}에서 '{data.get('query', args.get('query'))}'를 검색해 화면에 띄웠습니다."
+            base = f"{data.get('site', '')}에서 '{data.get('query', args.get('query'))}'를 검색해 화면에 띄웠습니다."
+            # Flag when the real cursor path didn't run, so a silent fallback to
+            # cursor-less DOM input is visible rather than reported as "human".
+            if data.get("input") in ("api", "api-fallback"):
+                base += " (참고: 실제 마우스 커서 이동 없이 처리됨)"
+            return base
         if action in ("back", "forward", "refresh"):
             return f"브라우저: {action} 완료 ({data.get('url', '')})."
         if action in ("status", "listTabs"):
@@ -164,6 +176,9 @@ class ControlBrowserTool(Tool):
             return f"탭 {data.get('active')}번으로 전환했습니다."
         if action == "closePopups":
             return f"팝업/빈 탭 {data.get('closed')}개를 닫았습니다."
+        if action == "moveMouse":
+            target = data.get("target") or args.get("selector") or args.get("site") or "대상"
+            return f"마우스 커서를 {target} 위치로 옮겼습니다."
         if action == "screenshot":
             return f"화면을 캡처했습니다: {data.get('path')}"
         return "완료했습니다."

tests/test_browse_and_play_index.py

@@ -0,0 +1,79 @@
+"""Tests for browseAndPlay's ``index`` argument (play the Nth search result).
+
+Behaviour verified:
+- default plays the first result (index 1) and stays backward-compatible,
+- an explicit index is forwarded to the Node helper as the 4th argv,
+- bad / sub-1 index values clamp to 1,
+- the index is advertised in the tool schema.
+"""
+
+import json
+from unittest.mock import Mock, patch
+
+import pytest
+
+from src.jarvis.tools.builtin.browse_and_play import BrowseAndPlayTool, _NODE_SCRIPT
+
+
+def _ctx():
+    cfg = Mock()
+    cfg.stream_browser = True
+    return Mock(cfg=cfg, user_print=Mock())
+
+
+def _run(args):
+    tool = BrowseAndPlayTool()
+    with patch("src.jarvis.tools.builtin.browse_and_play.subprocess.run") as mock_run:
+        mock_run.return_value = Mock(
+            stdout=json.dumps({"ok": True, "title": "Some Video"}),
+            stderr="",
+        )
+        result = tool.run(args, _ctx())
+        return mock_run, result
+
+
+def _argv(mock_run):
+    return list(mock_run.call_args[0][0])
+
+
+@pytest.mark.unit
+def test_schema_exposes_index():
+    schema = BrowseAndPlayTool().inputSchema
+    assert "index" in schema["properties"]
+    assert schema["properties"]["index"]["type"] == "integer"
+    assert "query" in schema["required"]
+    assert "index" not in schema["required"]  # optional
+
+
+@pytest.mark.unit
+def test_default_index_is_first():
+    mock_run, result = _run({"query": "IU Good Day"})
+    argv = _argv(mock_run)
+    assert argv[:4] == ["node", str(_NODE_SCRIPT), "IU Good Day", "youtube"]
+    assert argv[4] == "1"
+    assert result.success is True
+
+
+@pytest.mark.unit
+def test_explicit_index_forwarded():
+    mock_run, _ = _run({"query": "lofi", "index": 3})
+    assert _argv(mock_run)[4] == "3"
+
+
+@pytest.mark.unit
+@pytest.mark.parametrize("bad", [0, -2, "nope", None])
+def test_bad_index_clamps_to_one(bad):
+    mock_run, _ = _run({"query": "lofi", "index": bad})
+    assert _argv(mock_run)[4] == "1"
+
+
+@pytest.mark.unit
+def test_voice_only_mode_does_not_play():
+    tool = BrowseAndPlayTool()
+    cfg = Mock()
+    cfg.stream_browser = False
+    ctx = Mock(cfg=cfg, user_print=Mock())
+    with patch("src.jarvis.tools.builtin.browse_and_play.subprocess.run") as mock_run:
+        result = tool.run({"query": "x", "index": 2}, ctx)
+    assert result.success is False
+    mock_run.assert_not_called()

tests/test_control_browser.py

@@ -0,0 +1,56 @@
+"""Tests for the controlBrowser tool's action surface.
+
+These are deterministic schema/summary checks — they do not drive a real
+browser. The actual cursor movement is exercised live on the browser host
+(xdotool + CDP), which these tests cannot reach.
+"""
+
+import pytest
+
+from jarvis.tools.builtin.control_browser import ControlBrowserTool
+
+
+@pytest.fixture
+def tool():
+    return ControlBrowserTool()
+
+
+def test_movemouse_is_an_exposed_action(tool):
+    # A weak model confabulated "moved the mouse" because no move/hover action
+    # existed to call. The cursor-move capability must be a real action so the
+    # request "move the mouse to the search box" maps to a tool call.
+    enum = tool.inputSchema["properties"]["action"]["enum"]
+    assert "moveMouse" in enum
+
+
+def test_movemouse_summary_reports_the_target(tool):
+    summary = tool._summarise("moveMouse", {"site": "naver"}, {"ok": True, "target": "naver"})
+    assert "마우스" in summary and "naver" in summary
+
+
+def test_description_distinguishes_cursor_paths(tool):
+    # The model must know navigate is address-bar only (no mouse) while
+    # search/type/moveMouse move the real cursor — that distinction is the
+    # whole point of the fix.
+    desc = tool.description
+    assert "moveMouse" in desc
+    assert "address bar" in desc  # navigate is described as address-bar typing
+
+
+def test_search_summary_flags_cursorless_fallback(tool):
+    # When the real xdotool cursor path didn't run, the summary must say so
+    # rather than implying a human-like search happened.
+    human = tool._summarise("search", {"query": "날씨"}, {"ok": True, "site": "naver", "query": "날씨", "input": "human"})
+    assert "참고: 실제 마우스" not in human
+
+    fell_back = tool._summarise("search", {"query": "날씨"}, {"ok": True, "site": "naver", "query": "날씨", "input": "api-fallback"})
+    assert "실제 마우스 커서 이동 없이" in fell_back
+
+
+def test_movemouse_summary_only_runs_on_success(tool):
+    # _summarise is only called on ok:true; an ok:false (target not found / no
+    # xdotool) is handled by run() as a failure reply, so a failed move can no
+    # longer be reported as "moved". Sanity-check the success summary names a
+    # target rather than a placeholder when one is present.
+    summary = tool._summarise("moveMouse", {"selector": "#query"}, {"ok": True, "target": "#query"})
+    assert "#query" in summary

tests/test_intent_model_split.py

@@ -0,0 +1,62 @@
+"""The docker deployment must run auxiliary calls on a small model.
+
+Latency win: intent judging, tool routing and arg extraction are
+classification/JSON calls, not the spoken answer. Running them on a small fast
+model means the big chat model only runs once per command (for the answer),
+instead of 2-3 times per command for routing/extraction too.
+
+The wiring is: docker/jarvis-config.template.json renders `intent_judge_model`
+from `${OLLAMA_INTENT_MODEL}`, and the reply engine's resolver falls through
+`tool_router_model -> intent_judge_model -> ollama_chat_model`. The template
+sets no `tool_router_model`, so the auxiliary model is whatever
+`OLLAMA_INTENT_MODEL` renders to. These tests pin that behaviour end to end.
+"""
+
+import json
+import string
+from pathlib import Path
+
+import pytest
+
+from jarvis.reply.engine import resolve_tool_router_model
+
+TEMPLATE = Path(__file__).resolve().parent.parent / "docker" / "jarvis-config.template.json"
+
+
+def _render(**env) -> dict:
+    raw = TEMPLATE.read_text(encoding="utf-8")
+    return json.loads(string.Template(raw).safe_substitute(**env))
+
+
+class _Cfg:
+    """cfg stand-in carrying only the fields the resolver reads. The template
+    does not render `tool_router_model`, so it stays empty here too."""
+
+    def __init__(self, rendered: dict):
+        self.tool_router_model = rendered.get("tool_router_model", "") or ""
+        self.intent_judge_model = rendered.get("intent_judge_model", "") or ""
+        self.ollama_chat_model = rendered.get("ollama_chat_model", "") or ""
+
+
+def test_template_renders_separate_intent_model():
+    cfg = _render(OLLAMA_CHAT_MODEL="qwen3:8b", OLLAMA_INTENT_MODEL="qwen2.5:3b")
+    assert cfg["ollama_chat_model"] == "qwen3:8b"
+    assert cfg["intent_judge_model"] == "qwen2.5:3b"
+    assert cfg["intent_judge_model"] != cfg["ollama_chat_model"]
+
+
+@pytest.mark.unit
+def test_aux_calls_route_to_small_model_not_chat_model():
+    """The whole point: with a big chat model and a small intent model, tool
+    routing must resolve to the small model, leaving the big model for answers."""
+    cfg = _Cfg(_render(OLLAMA_CHAT_MODEL="qwen3:8b", OLLAMA_INTENT_MODEL="qwen2.5:3b"))
+    assert resolve_tool_router_model(cfg) == "qwen2.5:3b"
+
+
+@pytest.mark.unit
+def test_folding_intent_onto_chat_model_keeps_one_model():
+    """Setting OLLAMA_INTENT_MODEL == OLLAMA_CHAT_MODEL folds everything back
+    onto a single resident model (the documented VRAM-saving opt-out)."""
+    cfg = _Cfg(_render(OLLAMA_CHAT_MODEL="qwen2.5:3b", OLLAMA_INTENT_MODEL="qwen2.5:3b"))
+    assert resolve_tool_router_model(cfg) == "qwen2.5:3b"
+    assert cfg.intent_judge_model == cfg.ollama_chat_model

tests/test_ollama_num_predict.py

@@ -0,0 +1,112 @@
+"""Tests for the ``ollama_num_predict`` chat-output cap.
+
+The cap bounds worst-case reply latency by limiting how many tokens the chat
+model may generate per turn. Spoken (TTS) answers are 1-2 sentences, so the
+default headroom never clips a normal answer and stays above tool-call JSON.
+
+These tests verify behaviour:
+- the config default is present,
+- the value is threaded into the Ollama request as the ``num_predict`` option,
+- the reply loop forwards it to the chat call (and disables it at 0).
+"""
+
+from unittest.mock import Mock, patch
+
+import pytest
+
+from src.jarvis.config import get_default_config
+from src.jarvis.memory.conversation import DialogueMemory
+from src.jarvis.reply.engine import run_reply_engine
+
+
+# ---------------------------------------------------------------------------
+# Config default
+# ---------------------------------------------------------------------------
+
+def test_default_config_has_num_predict_cap():
+    config = get_default_config()
+    assert config["ollama_num_predict"] == 512
+
+
+# ---------------------------------------------------------------------------
+# Transport: extra_options.num_predict reaches the Ollama payload options
+# ---------------------------------------------------------------------------
+
+@patch("jarvis.llm.requests.post")
+def test_chat_with_messages_forwards_num_predict(mock_post):
+    from jarvis.llm import chat_with_messages
+
+    mock_resp = Mock()
+    mock_resp.status_code = 200
+    mock_resp.json.return_value = {"message": {"content": "ok"}}
+    mock_resp.raise_for_status = Mock()
+    mock_post.return_value = mock_resp
+
+    chat_with_messages(
+        "http://localhost:11434",
+        "test-large",
+        [{"role": "user", "content": "hi"}],
+        extra_options={"num_predict": 512},
+    )
+    _, kwargs = mock_post.call_args
+    options = (kwargs.get("json") or {}).get("options") or {}
+    assert options.get("num_predict") == 512
+
+
+# ---------------------------------------------------------------------------
+# Reply loop wiring
+# ---------------------------------------------------------------------------
+
+def _mock_cfg(num_predict):
+    cfg = Mock()
+    cfg.ollama_base_url = "http://localhost:11434"
+    cfg.ollama_chat_model = "test-large"  # avoid SMALL-model text-tool path
+    cfg.ollama_num_predict = num_predict
+    cfg.voice_debug = False
+    cfg.llm_tools_timeout_sec = 8.0
+    cfg.llm_embed_timeout_sec = 10.0
+    cfg.llm_chat_timeout_sec = 45.0
+    cfg.llm_digest_timeout_sec = 8.0
+    cfg.memory_enrichment_max_results = 5
+    cfg.memory_enrichment_source = "diary"
+    cfg.memory_digest_enabled = False
+    cfg.tool_result_digest_enabled = False
+    cfg.location_ip_address = None
+    cfg.location_auto_detect = False
+    cfg.location_enabled = False
+    cfg.agentic_max_turns = 8
+    cfg.tool_search_max_calls = 3
+    cfg.tool_selection_strategy = "all"
+    cfg.tool_carryover_max_turns = 2
+    cfg.tool_carryover_per_entry_chars = 1200
+    cfg.mcps = {}
+    cfg.llm_thinking_enabled = False
+    cfg.tts_engine = "none"
+    cfg.ollama_embed_model = "test-embed"
+    return cfg
+
+
+def _run_single_turn(cfg):
+    """Drive one reply turn that answers in plain text and capture the
+    chat call's extra_options."""
+    with patch("src.jarvis.reply.engine.plan_query", return_value=[]), \
+         patch("src.jarvis.reply.engine.extract_search_params_for_memory", return_value={}), \
+         patch("src.jarvis.reply.engine.extract_text_from_response", return_value="Hello."), \
+         patch("src.jarvis.reply.engine.chat_with_messages") as mock_chat:
+        mock_chat.return_value = {"message": {"content": "Hello."}}
+        run_reply_engine(db=Mock(), cfg=cfg, tts=None,
+                         text="hi", dialogue_memory=DialogueMemory())
+        assert mock_chat.called
+        return mock_chat.call_args.kwargs.get("extra_options")
+
+
+@pytest.mark.unit
+def test_reply_loop_caps_output_when_enabled():
+    extra = _run_single_turn(_mock_cfg(512))
+    assert extra == {"num_predict": 512}
+
+
+@pytest.mark.unit
+def test_reply_loop_no_cap_when_zero():
+    extra = _run_single_turn(_mock_cfg(0))
+    assert extra is None

tests/test_prompts.py

@@ -21,9 +21,15 @@ class TestModelSizeDetection:
         ("gemma:7b", True),
         ("phi3:3b", True),
         ("qwen2:7b", True),
+        # Sub-8B sizes that were previously misclassified as LARGE.
+        ("qwen3.5:4b", True),   # the deployed model that produced weak, off-tone replies
+        ("gemma2:2b", True),
+        ("model:5b", True),
+        ("model:6b", True),
         # Various separators
         ("model-3b-instruct", True),
         ("model_1b_chat", True),
+        ("model-4b-instruct", True),
         # Large models (should return LARGE)
         ("gpt-oss:20b", False),
         ("llama3.1:8b", False),
@@ -121,6 +127,18 @@ class TestPromptComponents:
             assert prompts.voice_style, f"{size.value} missing voice_style"
             assert prompts.tool_guidance, f"{size.value} missing tool_guidance"
 
+    def test_voice_style_enforces_single_sentence(self):
+        """voice_style must cap replies at one sentence (spoken aloud). The old
+        'Two to three sentences maximum' wording let the model run long, which
+        also slowed TTS since synth time scales with text length."""
+        from jarvis.reply.prompts import get_system_prompts, ModelSize
+
+        for size in [ModelSize.SMALL, ModelSize.LARGE]:
+            voice_style = get_system_prompts(size).voice_style
+            assert "SINGLE sentence" in voice_style, f"{size.value} voice_style not single-sentence"
+            assert "never more than one sentence" in voice_style
+            assert "Two to three" not in voice_style
+
     def test_to_list_returns_non_empty_strings(self):
         """to_list() returns only non-empty prompt strings."""
         from jarvis.reply.prompts import get_system_prompts, ModelSize

tests/test_system_prompt.py

@@ -44,6 +44,32 @@ class TestBuildSystemPrompt:
         assert "in the user's language" not in prompt
         assert "in Korean" in prompt
 
+    def test_persona_enforces_single_sentence(self):
+        # Spoken replies must be one sentence (TTS latency scales with text
+        # length, and the user asked for one-sentence answers). The persona must
+        # state the single-sentence rule and must NOT carry the old "two at the
+        # very most" allowance that let the model run long.
+        prompt = build_system_prompt("Jarvis")
+        assert "single short sentence" in prompt
+        assert "Never write a second sentence" in prompt
+        assert "two at the very most" not in prompt
+
+    def test_persona_aside_does_not_authorise_a_second_sentence(self):
+        # The dry aside must fold into the one sentence, not become a 2nd one.
+        prompt = build_system_prompt("Jarvis")
+        assert "SINGLE sentence" in prompt
+        assert "never add it as " in prompt
+
+    def test_persona_has_no_copyable_proper_noun_examples(self):
+        # A weak model parroted the literal "Trenches Gym" example from the
+        # persona as if it were a real user fact (boxing mangled to tennis).
+        # The persona must not embed copyable personal proper nouns, and must
+        # tell the model to use ONLY facts that literally appear in the memory
+        # section — never borrow names/places from the instructions themselves.
+        prompt = build_system_prompt("Jarvis")
+        assert "Trenches" not in prompt
+        assert "never borrow any name, place, or activity from these instructions" in prompt
+
 
 class TestOutputLanguageDirective:
     """A deployment may lock replies to a single language via OUTPUT_LANGUAGE.