llama-swap: use pre-built images (:cuda, :rocm) with GPU-specific flags

- Drop custom Dockerfiles; docker-compose uses ghcr.io pre-built images
  which ship llama-swap + llama-server with no pinned versions (always latest)
- NVIDIA GTX 1660 (6GB): add -fit off --no-kv-offload --cache-type-k q4_0 --cache-type-v q4_0
  to fix OOM segfault with new llama.cpp b9014's GPU-side KV cache default
- AMD RX 6800 (16GB): flags unchanged; KV cache stays on GPU for max speed
- Both running llama-swap v211 + llama.cpp b9014 (2026-05-05)

Dockerfile.llamaswap

@@ -1,13 +0,0 @@
-FROM ghcr.io/mostlygeek/llama-swap:cuda
-
-USER root
-
-# Download and install llama-server binary (CUDA version)
-# Using the official pre-built binary from llama.cpp releases
-ADD --chmod=755 https://github.com/ggml-org/llama.cpp/releases/download/b4183/llama-server-cuda /usr/local/bin/llama-server
-
-# Verify it's executable
-RUN llama-server --version || echo "llama-server installed successfully"
-
-USER 1000:1000
-

Dockerfile.llamaswap-rocm

@@ -1,68 +0,0 @@
-# Multi-stage build for llama-swap with ROCm support
-# Now using official llama.cpp ROCm image (PR #18439 merged Dec 29, 2025)
-
-# Stage 1: Build llama-swap UI
-FROM node:22-alpine AS ui-builder
-
-WORKDIR /build
-
-# Install git
-RUN apk add --no-cache git
-
-# Clone llama-swap
-RUN git clone https://github.com/mostlygeek/llama-swap.git
-
-# Build UI (now in ui-svelte directory)
-WORKDIR /build/llama-swap/ui-svelte
-RUN npm install && npm run build
-
-# Stage 2: Build llama-swap binary
-FROM golang:1.23-alpine AS swap-builder
-
-WORKDIR /build
-
-# Install git
-RUN apk add --no-cache git
-
-# Copy llama-swap source with built UI
-COPY --from=ui-builder /build/llama-swap /build/llama-swap
-
-# Build llama-swap binary
-WORKDIR /build/llama-swap
-RUN GOTOOLCHAIN=auto go build -o /build/llama-swap-binary .
-
-# Stage 3: Final runtime image using official llama.cpp ROCm image
-FROM ghcr.io/ggml-org/llama.cpp:server-rocm
-
-WORKDIR /app
-
-# Copy llama-swap binary from builder
-COPY --from=swap-builder /build/llama-swap-binary /app/llama-swap
-
-    # Make binaries executable
-    RUN chmod +x /app/llama-swap
-    
-    # Add existing ubuntu user (UID 1000) to GPU access groups (using host GIDs)
-    # GID 187 = render group on host, GID 989 = video/kfd group on host
-    RUN groupadd -g 187 hostrender && \
-        groupadd -g 989 hostvideo && \
-        usermod -aG hostrender,hostvideo ubuntu && \
-        chown -R ubuntu:ubuntu /app
-    
-    # Set environment for ROCm (RX 6800 is gfx1030)
-    ENV HSA_OVERRIDE_GFX_VERSION=10.3.0
-    ENV ROCM_PATH=/opt/rocm
-    ENV HIP_VISIBLE_DEVICES=0
-    
-    USER ubuntu
-    
-    # Expose port
-    EXPOSE 8080
-    
-    # Health check
-    HEALTHCHECK --interval=30s --timeout=10s --start-period=30s --retries=3 \
-      CMD curl -f http://localhost:8080/health || exit 1
-    
-    # Override the base image's ENTRYPOINT and run llama-swap
-    ENTRYPOINT []
-    CMD ["/app/llama-swap", "-config", "/app/config.yaml", "-listen", "0.0.0.0:8080"]

bot/bot.py

@@ -203,6 +203,31 @@ async def on_message(message):
         if is_persona_dialogue_active(message.channel.id):
             return
 
+    # Bipolar mode: check if the opposite persona should interject on user messages
+    # AND roll for random argument trigger (both non-blocking background tasks)
+    if not isinstance(message.channel, discord.DMChannel) and globals.BIPOLAR_MODE:
+        try:
+            from utils.persona_dialogue import check_for_interjection, is_persona_dialogue_active as dialogue_active
+            from utils.bipolar_mode import maybe_trigger_argument, is_argument_in_progress as arg_in_progress
+            from utils.task_tracker import create_tracked_task
+            
+            # Check interjection on user messages (opposite of current active persona)
+            if not message.author.bot or message.webhook_id:
+                current_persona = "evil" if globals.EVIL_MODE else "miku"
+                create_tracked_task(
+                    check_for_interjection(message, current_persona),
+                    task_name="interjection_check_user",
+                )
+            
+            # Roll random argument trigger chance (15%) on eligible messages
+            if not arg_in_progress(message.channel.id) and not dialogue_active(message.channel.id):
+                create_tracked_task(
+                    maybe_trigger_argument(message.channel, globals.client, "Triggered from conversation flow"),
+                    task_name="random_argument_trigger",
+                )
+        except Exception as e:
+            logger.error(f"Error in bipolar trigger checks: {e}")
+
     if message.content.strip().lower() == "miku, rape this nigga balls" and message.reference:
         async with message.channel.typing():
             # Get replied-to user
@@ -335,15 +360,24 @@ async def on_message(message):
                         if globals.EVIL_MODE:
                             effective_mood = f"EVIL:{getattr(globals, 'EVIL_DM_MOOD', 'evil_neutral')}"
                         logger.info(f"🐱 Cat response for {author_name} (mood: {effective_mood})")
-                        # Track Cat interaction for Web UI Last Prompt view
+                        # Track Cat interaction in unified prompt history
                         import datetime
-                        globals.LAST_CAT_INTERACTION = {
+                        globals._prompt_id_counter += 1
+                        guild_name = message.guild.name if message.guild else "DM"
+                        channel_name = message.channel.name if message.guild else "DM"
+                        globals.PROMPT_HISTORY.append({
+                            "id": globals._prompt_id_counter,
+                            "source": "cat",
                             "full_prompt": cat_full_prompt,
-                            "response": response[:500] if response else "",
+                            "response": response if response else "",
                             "user": author_name,
                             "mood": effective_mood,
+                            "guild": guild_name,
+                            "channel": channel_name,
                             "timestamp": datetime.datetime.now().isoformat(),
-                        }
+                            "model": "Cat LLM",
+                            "response_type": response_type,
+                        })
                 except Exception as e:
                     logger.warning(f"🐱 Cat pipeline error, falling back to query_llama: {e}")
                     response = None

bot/globals.py

@@ -1,6 +1,7 @@
 # globals.py
 import os
 import discord
+from collections import deque
 from apscheduler.schedulers.asyncio import AsyncIOScheduler
 
 scheduler = AsyncIOScheduler()
@@ -77,16 +78,25 @@ MIKU_NORMAL_AVATAR_URL = None  # Cached CDN URL of the regular Miku pfp (valid e
 
 BOT_USER = None
 
-LAST_FULL_PROMPT = ""
+# Unified prompt history (replaces LAST_FULL_PROMPT and LAST_CAT_INTERACTION)
+# Each entry: {id, source, full_prompt, response, user, mood, guild, channel,
+#              timestamp, model, response_type}
+PROMPT_HISTORY = deque(maxlen=10)
+_prompt_id_counter = 0
 
-# Cheshire Cat last interaction tracking (for Web UI Last Prompt toggle)
-LAST_CAT_INTERACTION = {
-    "full_prompt": "",
-    "response": "",
-    "user": "",
-    "mood": "",
-    "timestamp": "",
-}
+# Legacy accessors for backward compatibility (routes, CLI, etc.)
+# These are computed properties that read from PROMPT_HISTORY
+def _get_last_fallback_prompt():
+    for entry in reversed(PROMPT_HISTORY):
+        if entry.get("source") == "fallback":
+            return entry.get("full_prompt", "")
+    return ""
+
+def _get_last_cat_interaction():
+    for entry in reversed(PROMPT_HISTORY):
+        if entry.get("source") == "cat":
+            return entry
+    return {"full_prompt": "", "response": "", "user": "", "mood": "", "timestamp": ""}
 
 # Persona Dialogue System (conversations between Miku and Evil Miku)
 LAST_PERSONA_DIALOGUE_TIME = 0  # Timestamp of last dialogue for cooldown

bot/routes/bipolar_mode.py

@@ -148,7 +148,7 @@ def trigger_argument(data: BipolarTriggerRequest):
     if not channel:
         return JSONResponse(status_code=404, content={"status": "error", "message": f"Channel {channel_id} not found"})
 
-    # Trigger the argument
+    # Trigger the argument — context doubles as the argument theme
     globals.client.loop.create_task(force_trigger_argument(channel, globals.client, data.context))
 
     return {

bot/routes/core.py

@@ -14,7 +14,8 @@ router = APIRouter()
 
 @router.get("/")
 def read_index():
-    return FileResponse("static/index.html")
+    headers = {"Cache-Control": "no-cache, no-store, must-revalidate"}
+    return FileResponse("static/index.html", headers=headers)
 
 
 @router.get("/logs")
@@ -31,18 +32,45 @@ def get_logs():
 
 @router.get("/prompt")
 def get_last_prompt():
-    return {"prompt": globals.LAST_FULL_PROMPT or "No prompt has been issued yet."}
+    """Legacy endpoint: returns the most recent fallback prompt (backward compat)."""
+    prompt_text = globals._get_last_fallback_prompt()
+    return {"prompt": prompt_text or "No prompt has been issued yet."}
 
 
 @router.get("/prompt/cat")
 def get_last_cat_prompt():
-    """Get the last Cheshire Cat interaction (full prompt + response) for Web UI."""
-    interaction = globals.LAST_CAT_INTERACTION
+    """Legacy endpoint: returns the most recent Cat interaction (backward compat)."""
+    interaction = globals._get_last_cat_interaction()
     if not interaction.get("full_prompt"):
-        return {"full_prompt": "No Cheshire Cat interaction has occurred yet.", "response": "", "user": "", "mood": "", "timestamp": ""}
+        return {"full_prompt": "No Cheshire Cat interaction has occurred yet.",
+                "response": "", "user": "", "mood": "", "timestamp": ""}
     return interaction
 
 
+@router.get("/prompts")
+def get_prompt_history(source: str = None):
+    """
+    Return the unified prompt history.
+    Optional query param ?source=cat or ?source=fallback to filter.
+    """
+    history = list(globals.PROMPT_HISTORY)
+    if source and source in ("cat", "fallback"):
+        history = [e for e in history if e.get("source") == source]
+    return {"history": history}
+
+
+@router.get("/prompts/{prompt_id}")
+def get_prompt_by_id(prompt_id: int):
+    """Return a single prompt history entry by ID."""
+    for entry in globals.PROMPT_HISTORY:
+        if entry.get("id") == prompt_id:
+            return entry
+    return JSONResponse(
+        status_code=404,
+        content={"status": "error", "message": f"Prompt #{prompt_id} not found"}
+    )
+
+
 @router.get("/status")
 def status():
     # Get per-server mood summary

bot/routes/models.py

@@ -45,7 +45,7 @@ class LogFilterUpdateRequest(BaseModel):
 class BipolarTriggerRequest(BaseModel):
     channel_id: str  # String to handle large Discord IDs from JS
     message_id: str = None  # Optional: starting message ID (string)
-    context: str = ""
+    context: str = ""  # Optional: argument theme/context — tells them what to argue about
 
 
 class ManualCropRequest(BaseModel):

bot/static/css/style.css

@@ -441,6 +441,51 @@ h1, h3 {
   color: #ddd;
 }
 
+/* Prompt History Section */
+#prompt-history-section.collapsed #prompt-history-body {
+  display: none;
+}
+#prompt-history-toggle {
+  user-select: none;
+  transition: color 0.2s;
+}
+#prompt-history-toggle:hover {
+  color: #4CAF50;
+}
+#prompt-metadata span {
+  white-space: nowrap;
+}
+#prompt-metadata .prompt-meta-label {
+  color: #666;
+}
+#prompt-metadata .prompt-meta-value {
+  color: #ccc;
+}
+#prompt-display pre {
+  margin: 0;
+}
+.prompt-subsection-header {
+  cursor: pointer;
+  user-select: none;
+  padding: 0.3rem 0.5rem;
+  border-radius: 4px;
+  background: #2a2a2a;
+  margin: 0.5rem 0 0.25rem 0;
+  font-size: 0.82rem;
+  color: #aaa;
+  transition: background 0.15s;
+}
+.prompt-subsection-header:hover {
+  background: #333;
+  color: #ddd;
+}
+.prompt-subsection-body.collapsed {
+  display: none;
+}
+#prompt-truncate-toggle {
+  accent-color: #4CAF50;
+}
+
 /* Mood Activities Editor */
 .act-mood-row {
   margin-bottom: 0.5rem;

bot/static/index.html

@@ -3,10 +3,13 @@
 <head>
   <meta charset="UTF-8">
   <meta name="viewport" content="width=device-width, initial-scale=1.0">
+  <meta http-equiv="Cache-Control" content="no-cache, no-store, must-revalidate">
+  <meta http-equiv="Pragma" content="no-cache">
+  <meta http-equiv="Expires" content="0">
   <title>Miku Control Panel</title>
   <link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/cropperjs/1.6.2/cropper.min.css">
   <script src="https://cdnjs.cloudflare.com/ajax/libs/cropperjs/1.6.2/cropper.min.js"></script>
-  <link rel="stylesheet" href="/static/css/style.css">
+  <link rel="stylesheet" href="/static/css/style.css?v=20260502">
 </head>
 <body>
 
@@ -234,8 +237,11 @@
         </div>
         
         <div style="margin-bottom: 1rem;">
-          <label for="bipolar-context">Argument Context (optional):</label>
-          <input type="text" id="bipolar-context" placeholder="e.g., They're fighting about who's the real Miku..." style="width: 100%; margin-top: 0.3rem;">
+          <label for="bipolar-context">What should they argue about? (optional):</label>
+          <input type="text" id="bipolar-context" placeholder="e.g., Who's the real Miku? Whether kindness is weakness. A petty grudge..." style="width: 100%; margin-top: 0.3rem;">
+          <div style="font-size: 0.75rem; color: #777; margin-top: 0.2rem;">
+            Leave blank for a random topic. Write anything to set the argument's theme.
+          </div>
         </div>
         
         <button onclick="triggerBipolarArgument()" style="background: #9932CC; color: #fff; border: none; padding: 0.5rem 1rem; border-radius: 4px; cursor: pointer;">
@@ -540,23 +546,53 @@
         </div>
       </div>
 
-      <div class="section">
-        <h3>Last Prompt</h3>
-        <div style="margin-bottom: 0.75rem; display: flex; align-items: center; gap: 0.75rem;">
-          <label style="font-size: 0.9rem; color: #aaa;">Source:</label>
-          <div style="display: inline-flex; border-radius: 6px; overflow: hidden; border: 1px solid #444;">
-            <button id="prompt-src-cat" class="prompt-source-btn active" onclick="switchPromptSource('cat')"
-              style="padding: 0.4rem 1rem; border: none; cursor: pointer; font-size: 0.85rem; transition: all 0.2s;">
-              🐱 Cheshire Cat
-            </button>
-            <button id="prompt-src-fallback" class="prompt-source-btn" onclick="switchPromptSource('fallback')"
-              style="padding: 0.4rem 1rem; border: none; cursor: pointer; font-size: 0.85rem; transition: all 0.2s;">
-              🤖 Bot Fallback
-            </button>
-          </div>
+      <div class="section" id="prompt-history-section">
+        <div class="prompt-history-header" style="display: flex; align-items: center; justify-content: space-between; margin-bottom: 0.5rem;">
+          <h3 style="margin: 0; cursor: pointer;" onclick="togglePromptHistoryCollapse()" id="prompt-history-toggle">
+            ▼ Prompt History
+          </h3>
+          <button onclick="loadPromptHistory()" title="Refresh" style="background: none; border: 1px solid #444; color: #aaa; cursor: pointer; padding: 0.2rem 0.5rem; border-radius: 4px; font-size: 0.85rem;">🔄</button>
+        </div>
+        <div id="prompt-history-body">
+          <!-- Source filter + history selector row -->
+          <div style="margin-bottom: 0.75rem; display: flex; align-items: center; gap: 0.75rem; flex-wrap: wrap;">
+            <label style="font-size: 0.9rem; color: #aaa;">Source:</label>
+            <div style="display: inline-flex; border-radius: 6px; overflow: hidden; border: 1px solid #444;">
+              <button id="prompt-src-all" class="prompt-source-btn active" onclick="switchPromptSource('all')"
+                style="padding: 0.4rem 0.8rem; border: none; cursor: pointer; font-size: 0.85rem; transition: all 0.2s;">
+                All
+              </button>
+              <button id="prompt-src-cat" class="prompt-source-btn" onclick="switchPromptSource('cat')"
+                style="padding: 0.4rem 0.8rem; border: none; cursor: pointer; font-size: 0.85rem; transition: all 0.2s;">
+                🐱 Cat
+              </button>
+              <button id="prompt-src-fallback" class="prompt-source-btn" onclick="switchPromptSource('fallback')"
+                style="padding: 0.4rem 0.8rem; border: none; cursor: pointer; font-size: 0.85rem; transition: all 0.2s;">
+                🤖 Fallback
+              </button>
+            </div>
+            <select id="prompt-history-select" onchange="selectPromptEntry(this.value)" style="background: #2a2a2a; color: #ddd; border: 1px solid #444; padding: 0.35rem 0.5rem; border-radius: 4px; font-size: 0.85rem; min-width: 280px;">
+              <option value="">-- No prompts yet --</option>
+            </select>
+          </div>
+
+          <!-- Metadata bar -->
+          <div id="prompt-metadata" style="margin-bottom: 0.5rem; font-size: 0.82rem; color: #888; display: flex; flex-wrap: wrap; gap: 0.3rem 1rem;"></div>
+
+          <!-- Toolbar: copy + truncate toggle -->
+          <div style="margin-bottom: 0.5rem; display: flex; align-items: center; gap: 1rem;">
+            <button onclick="copyPromptToClipboard()" title="Copy full prompt to clipboard" style="background: #333; border: 1px solid #555; color: #aaa; cursor: pointer; padding: 0.25rem 0.6rem; border-radius: 4px; font-size: 0.8rem;">📋 Copy</button>
+            <label style="font-size: 0.82rem; color: #aaa; cursor: pointer; display: flex; align-items: center; gap: 0.3rem;">
+              <input type="checkbox" id="prompt-truncate-toggle" onchange="toggleMiddleTruncation()">
+              Truncate from middle
+            </label>
+          </div>
+
+          <!-- Prompt display subsections -->
+          <div id="prompt-display" style="max-height: 60vh; overflow-y: auto; min-height: 3rem;"></div>
+          <!-- Hidden buffer for copy-to-clipboard raw text -->
+          <pre id="last-prompt" style="display: none;"></pre>
         </div>
-        <div id="prompt-cat-info" style="margin-bottom: 0.5rem; font-size: 0.85rem; color: #aaa;"></div>
-        <pre id="last-prompt" style="white-space: pre-wrap; word-break: break-word;"></pre>
       </div>
     </div>
 
@@ -1336,15 +1372,15 @@
   </div>
 </div>
 
-<script src="/static/js/core.js"></script>
-<script src="/static/js/servers.js"></script>
-<script src="/static/js/modes.js"></script>
-<script src="/static/js/actions.js"></script>
-<script src="/static/js/image-gen.js"></script>
-<script src="/static/js/status.js"></script>
-<script src="/static/js/dm.js"></script>
-<script src="/static/js/chat.js"></script>
-<script src="/static/js/memories.js"></script>
-<script src="/static/js/profile.js"></script>
+<script src="/static/js/core.js?v=20260502"></script>
+<script src="/static/js/servers.js?v=20260502"></script>
+<script src="/static/js/modes.js?v=20260502"></script>
+<script src="/static/js/actions.js?v=20260502"></script>
+<script src="/static/js/image-gen.js?v=20260502"></script>
+<script src="/static/js/status.js?v=20260502"></script>
+<script src="/static/js/dm.js?v=20260502"></script>
+<script src="/static/js/chat.js?v=20260502"></script>
+<script src="/static/js/memories.js?v=20260502"></script>
+<script src="/static/js/profile.js?v=20260502"></script>
 </body>
 </html>

bot/static/js/core.js

@@ -29,6 +29,7 @@ let notificationTimer = null;
 let statusInterval = null;
 let logsInterval = null;
 let argsInterval = null;
+let promptInterval = null;
 
 // Mood emoji mapping
 const MOOD_EMOJIS = {
@@ -211,12 +212,14 @@ function startPolling() {
   if (!statusInterval) statusInterval = setInterval(loadStatus, 10000);
   if (!logsInterval) logsInterval = setInterval(loadLogs, 5000);
   if (!argsInterval) argsInterval = setInterval(loadActiveArguments, 5000);
+  if (!promptInterval) promptInterval = setInterval(loadPromptHistory, 10000);
 }
 
 function stopPolling() {
   clearInterval(statusInterval); statusInterval = null;
   clearInterval(logsInterval); logsInterval = null;
   clearInterval(argsInterval); argsInterval = null;
+  clearInterval(promptInterval); promptInterval = null;
 }
 
 // ============================================================================
@@ -248,7 +251,7 @@ function initVisibilityPolling() {
       stopPolling();
       console.log('⏸ Tab hidden — polling paused');
     } else {
-      loadStatus(); loadLogs(); loadActiveArguments();
+      loadStatus(); loadLogs(); loadActiveArguments(); loadPromptHistory();
       startPolling();
       console.log('▶️ Tab visible — polling resumed');
     }
@@ -296,9 +299,11 @@ function initModalAccessibility() {
 }
 
 function initPromptSourceToggle() {
-  const saved = localStorage.getItem('miku-prompt-source') || 'cat';
+  const saved = localStorage.getItem('miku-prompt-source') || 'all';
   document.querySelectorAll('.prompt-source-btn').forEach(btn => btn.classList.remove('active'));
-  document.getElementById(`prompt-src-${saved}`).classList.add('active');
+  const btnId = saved === 'all' ? 'prompt-src-all' : `prompt-src-${saved}`;
+  const btn = document.getElementById(btnId);
+  if (btn) btn.classList.add('active');
 }
 
 function initLogsScrollDetection() {
@@ -360,8 +365,10 @@ async function loadLogs() {
 function switchPromptSource(source) {
   localStorage.setItem('miku-prompt-source', source);
   document.querySelectorAll('.prompt-source-btn').forEach(btn => btn.classList.remove('active'));
-  document.getElementById(`prompt-src-${source}`).classList.add('active');
-  loadLastPrompt();
+  const btnId = source === 'all' ? 'prompt-src-all' : `prompt-src-${source}`;
+  const btn = document.getElementById(btnId);
+  if (btn) btn.classList.add('active');
+  loadPromptHistory();
 }
 
 // ============================================================================

bot/static/js/modes.js

@@ -248,7 +248,7 @@ async function triggerPersonaDialogue() {
 async function triggerBipolarArgument() {
   const channelIdInput = document.getElementById('bipolar-channel-id').value.trim();
   const messageIdInput = document.getElementById('bipolar-message-id').value.trim();
-  const context = document.getElementById('bipolar-context').value;
+  const context = document.getElementById('bipolar-context').value.trim();
   const statusDiv = document.getElementById('bipolar-status');
 
   if (!channelIdInput) {

bot/static/js/status.js

@@ -57,33 +57,271 @@ async function loadStatus() {
   }
 }
 
-// ===== Last Prompt =====
+// ===== Prompt History =====
 
-async function loadLastPrompt() {
-  const source = localStorage.getItem('miku-prompt-source') || 'cat';
-  const promptEl = document.getElementById('last-prompt');
-  const infoEl = document.getElementById('prompt-cat-info');
+let _promptHistoryCache = [];   // cached history entries from last fetch
+let _selectedPromptId = null;  // currently selected entry ID
+let _middleTruncation = false; // whether middle-truncation is active
+
+async function loadPromptHistory() {
+  const source = localStorage.getItem('miku-prompt-source') || 'all';
+  const selectEl = document.getElementById('prompt-history-select');
 
   try {
-    if (source === 'cat') {
-      const result = await apiCall('/prompt/cat');
-      if (result.timestamp) {
-        infoEl.innerHTML = `<strong>User:</strong> ${escapeHtml(result.user || '?')} &nbsp;|&nbsp; <strong>Mood:</strong> ${escapeHtml(result.mood || '?')} &nbsp;|&nbsp; <strong>Time:</strong> ${new Date(result.timestamp).toLocaleString()}`;
-        promptEl.textContent = result.full_prompt + `\n\n${'═'.repeat(60)}\n[Cat Response]\n${result.response}`;
-      } else {
-        infoEl.textContent = '';
-        promptEl.textContent = result.full_prompt || 'No Cheshire Cat interaction yet.';
-      }
+    const url = source === 'all' ? '/prompts' : `/prompts?source=${source}`;
+    const result = await apiCall(url);
+    _promptHistoryCache = result.history || [];
+
+    // Populate dropdown
+    const currentValue = selectEl.value;
+    selectEl.innerHTML = '';
+    if (_promptHistoryCache.length === 0) {
+      selectEl.innerHTML = '<option value="">-- No prompts yet --</option>';
     } else {
-      infoEl.textContent = '';
-      const result = await apiCall('/prompt');
-      promptEl.textContent = result.prompt;
+      _promptHistoryCache.forEach(entry => {
+        const ts = entry.timestamp ? new Date(entry.timestamp).toLocaleTimeString() : '?';
+        const srcLabel = entry.source === 'cat' ? '🐱' : '🤖';
+        const user = entry.user || '?';
+        const option = document.createElement('option');
+        option.value = entry.id;
+        option.textContent = `${srcLabel} #${entry.id} — ${user} — ${ts}`;
+        selectEl.appendChild(option);
+      });
+    }
+
+    // Restore or auto-select the latest entry
+    if (_selectedPromptId && _promptHistoryCache.some(e => e.id === _selectedPromptId)) {
+      selectEl.value = _selectedPromptId;
+    } else if (_promptHistoryCache.length > 0) {
+      selectEl.value = _promptHistoryCache[0].id;
+    }
+
+    if (selectEl.value) {
+      await selectPromptEntry(selectEl.value);
+    } else {
+      clearPromptDisplay();
     }
   } catch (error) {
-    console.error('Failed to load last prompt:', error);
+    console.error('Failed to load prompt history:', error);
   }
 }
 
+async function selectPromptEntry(promptId) {
+  if (!promptId) {
+    clearPromptDisplay();
+    return;
+  }
+
+  _selectedPromptId = parseInt(promptId);
+
+  // Try cache first
+  let entry = _promptHistoryCache.find(e => e.id === _selectedPromptId);
+
+  // Fall back to API call if not in cache
+  if (!entry) {
+    try {
+      entry = await apiCall(`/prompts/${_selectedPromptId}`);
+    } catch (error) {
+      console.error('Failed to load prompt entry:', error);
+      clearPromptDisplay();
+      return;
+    }
+  }
+
+  if (!entry) {
+    clearPromptDisplay();
+    return;
+  }
+
+  renderPromptEntry(entry);
+}
+
+function clearPromptDisplay() {
+  document.getElementById('prompt-metadata').innerHTML = '';
+  document.getElementById('prompt-display').innerHTML = '<pre style="white-space: pre-wrap; word-break: break-word; background: #1a1a1a; padding: 0.75rem; border-radius: 4px; font-size: 0.8rem; line-height: 1.4; margin: 0; color: #666;">No prompt selected.</pre>';
+  document.getElementById('last-prompt').textContent = '';
+}
+
+function renderPromptEntry(entry) {
+  // Metadata bar
+  const metaEl = document.getElementById('prompt-metadata');
+  const ts = entry.timestamp ? new Date(entry.timestamp).toLocaleString() : '?';
+  const sourceIcon = entry.source === 'cat' ? '🐱 Cat' : '🤖 Fallback';
+  metaEl.innerHTML = `
+    <span><span class="prompt-meta-label">#</span><span class="prompt-meta-value">${entry.id}</span></span>
+    <span><span class="prompt-meta-label">Source:</span> <span class="prompt-meta-value">${sourceIcon}</span></span>
+    <span><span class="prompt-meta-label">User:</span> <span class="prompt-meta-value">${escapeHtml(entry.user || '?')}</span></span>
+    <span><span class="prompt-meta-label">Mood:</span> <span class="prompt-meta-value">${escapeHtml(entry.mood || '?')}</span></span>
+    <span><span class="prompt-meta-label">Guild:</span> <span class="prompt-meta-value">${escapeHtml(entry.guild || '?')}</span></span>
+    <span><span class="prompt-meta-label">Channel:</span> <span class="prompt-meta-value">${escapeHtml(entry.channel || '?')}</span></span>
+    <span><span class="prompt-meta-label">Model:</span> <span class="prompt-meta-value">${escapeHtml(entry.model || '?')}</span></span>
+    <span><span class="prompt-meta-label">Type:</span> <span class="prompt-meta-value">${escapeHtml(entry.response_type || '?')}</span></span>
+    <span><span class="prompt-meta-label">Time:</span> <span class="prompt-meta-value">${ts}</span></span>
+  `;
+
+  // Parse full_prompt into sections
+  const sections = parsePromptSections(entry.full_prompt || '');
+
+  // Snapshot which subsections are currently collapsed (before re-render)
+  const sectionIds = ['system', 'context', 'conversation', 'response'];
+  const collapsedState = {};
+  sectionIds.forEach(id => {
+    const el = document.getElementById(`prompt-section-${id}`);
+    collapsedState[id] = el && el.classList.contains('collapsed');
+  });
+
+  // Build display HTML with collapsible subsections
+  let displayHtml = '';
+
+  if (sections.system) {
+    displayHtml += buildCollapsibleSection('System Prompt', sections.system, 'system');
+  }
+  if (sections.context) {
+    displayHtml += buildCollapsibleSection('Context (Memories & Tools)', sections.context, 'context');
+  }
+  if (sections.conversation) {
+    displayHtml += buildCollapsibleSection('Conversation', sections.conversation, 'conversation');
+  }
+  if (!sections.system && !sections.context && !sections.conversation) {
+    // Fallback: show raw full_prompt
+    displayHtml += `<pre style="white-space: pre-wrap; word-break: break-word; margin: 0;">${escapeHtml(entry.full_prompt || '')}</pre>`;
+  }
+
+  // Response section
+  if (entry.response) {
+    let responseText = entry.response;
+    if (_middleTruncation && responseText.length > 400) {
+      responseText = responseText.substring(0, 200) + '\n\n... [truncated middle] ...\n\n' + responseText.substring(responseText.length - 200);
+    }
+    displayHtml += buildCollapsibleSection('Response', responseText, 'response');
+  }
+
+  // Render into the prompt-display div (using innerHTML for collapsible structure)
+  const displayEl = document.getElementById('prompt-display');
+  displayEl.innerHTML = displayHtml;
+
+  // Restore collapsed state from snapshot
+  sectionIds.forEach(id => {
+    const el = document.getElementById(`prompt-section-${id}`);
+    if (el && collapsedState[id]) {
+      el.classList.add('collapsed');
+      const header = el.previousElementSibling;
+      if (header) header.innerHTML = header.innerHTML.replace('▼', '▶');
+    }
+  });
+
+  // Also set the raw text into the <pre> for copy functionality
+  let rawText = entry.full_prompt || '';
+  if (entry.response) {
+    rawText += `\n\n${'═'.repeat(60)}\n[Response]\n${entry.response}`;
+  }
+  document.getElementById('last-prompt').textContent = rawText;
+}
+
+function parsePromptSections(fullPrompt) {
+  const sections = { system: null, context: null, conversation: null };
+
+  if (!fullPrompt) return sections;
+
+  // Try to split on known section markers
+  const contextMatch = fullPrompt.match(/# Context\s*\n([\s\S]*?)(?=\n# Conversation|\nHuman:|\n$)/);
+  const convMatch = fullPrompt.match(/# Conversation until now:\s*\n([\s\S]*)/);
+
+  if (contextMatch) {
+    // Everything before # Context is the system prompt
+    const contextIdx = fullPrompt.indexOf('# Context');
+    if (contextIdx > 0) {
+      sections.system = fullPrompt.substring(0, contextIdx).trim();
+    }
+    sections.context = contextMatch[1].trim();
+  }
+
+  if (convMatch) {
+    sections.conversation = convMatch[1].trim();
+  } else {
+    // Try alternative: "Human:" at the end
+    const humanMatch = fullPrompt.match(/\nHuman:([\s\S]*)/);
+    if (humanMatch && fullPrompt.indexOf('Human:') > fullPrompt.indexOf('# Context')) {
+      sections.conversation = 'Human:' + humanMatch[1].trim();
+    }
+  }
+
+  // If no # Context marker, try "System:" prefix (fallback prompts)
+  if (!sections.system && !sections.context) {
+    const sysMatch = fullPrompt.match(/^System:\s*([\s\S]*?)(?=\nMessages:)/);
+    const msgMatch = fullPrompt.match(/Messages:\s*([\s\S]*)/);
+    if (sysMatch) {
+      sections.system = sysMatch[1].trim();
+    }
+    if (msgMatch) {
+      sections.conversation = msgMatch[1].trim();
+    }
+  }
+
+  return sections;
+}
+
+function buildCollapsibleSection(title, content, sectionId) {
+  const id = `prompt-section-${sectionId}`;
+  return `
+    <div class="prompt-subsection-header" onclick="togglePromptSubsection('${id}')">
+      ▼ ${escapeHtml(title)}
+    </div>
+    <div class="prompt-subsection-body" id="${id}">
+      <pre style="white-space: pre-wrap; word-break: break-word; background: #1a1a1a; padding: 0.5rem; border-radius: 4px; font-size: 0.8rem; line-height: 1.4; margin: 0.25rem 0;">${escapeHtml(content)}</pre>
+    </div>`;
+}
+
+function togglePromptSubsection(id) {
+  const body = document.getElementById(id);
+  if (!body) return;
+  const header = body.previousElementSibling;
+  if (body.classList.contains('collapsed')) {
+    body.classList.remove('collapsed');
+    if (header) header.innerHTML = header.innerHTML.replace('▶', '▼');
+  } else {
+    body.classList.add('collapsed');
+    if (header) header.innerHTML = header.innerHTML.replace('▼', '▶');
+  }
+}
+
+function togglePromptHistoryCollapse() {
+  const section = document.getElementById('prompt-history-section');
+  const toggle = document.getElementById('prompt-history-toggle');
+  if (section.classList.contains('collapsed')) {
+    section.classList.remove('collapsed');
+    toggle.textContent = '▼ Prompt History';
+  } else {
+    section.classList.add('collapsed');
+    toggle.textContent = '▶ Prompt History';
+  }
+}
+
+function copyPromptToClipboard() {
+  const rawText = document.getElementById('last-prompt').textContent;
+  if (!rawText) return;
+  navigator.clipboard.writeText(rawText).then(() => {
+    showNotification('Prompt copied to clipboard', 'success');
+  }).catch(err => {
+    console.error('Failed to copy:', err);
+    showNotification('Failed to copy', 'error');
+  });
+}
+
+function toggleMiddleTruncation() {
+  _middleTruncation = document.getElementById('prompt-truncate-toggle').checked;
+  // Re-render current entry
+  if (_selectedPromptId) {
+    selectPromptEntry(_selectedPromptId);
+  }
+}
+
+// Legacy compatibility — called from core.js on page load / tab switch
+// Redirects to the new loadPromptHistory()
+async function loadLastPrompt() {
+  await loadPromptHistory();
+}
+
 // ===== Autonomous Stats =====
 
 async function loadAutonomousStats() {

bot/utils/bipolar_mode.py

@@ -23,12 +23,33 @@ logger = get_logger('persona')
 BIPOLAR_STATE_FILE = "memory/bipolar_mode_state.json"
 BIPOLAR_WEBHOOKS_FILE = "memory/bipolar_webhooks.json"
 BIPOLAR_SCOREBOARD_FILE = "memory/bipolar_scoreboard.json"
+ARGUMENT_TOPICS_FILE = "memory/argument_topics.json"
 
 # Argument settings
 MIN_EXCHANGES = 4  # Minimum number of back-and-forth exchanges before ending can occur
 ARGUMENT_TRIGGER_CHANCE = 0.15  # 15% chance for the other Miku to break through
 DELAY_BETWEEN_MESSAGES = (2.0, 5.0)  # Random delay between argument messages (seconds)
 
+# Argument topic rotation — each topic gives the argument a different framing
+# Topics are weighted: higher weight = more likely to be selected
+ARGUMENT_TOPICS = [
+    # (topic_name, weight, description for prompt injection)
+    ("identity_crisis", 3, "Who is the REAL Miku? Authenticity vs. the shadow self"),
+    ("power_dynamic", 3, "Who holds the power? Dominance, submission, and control"),
+    ("philosophical", 2, "Is kindness strength or weakness? Does darkness serve a purpose?"),
+    ("petty_grievance", 3, "Something small and petty that escalated — a specific annoyance, habit, or incident"),
+    ("existential_dread", 1, "What's the point of any of it? Nihilism vs. hope, meaning vs. emptiness"),
+    ("audience_appeal", 3, "Who do the fans/chatters ACTUALLY prefer? Popularity contest with receipts"),
+    ("personal_attack", 3, "Deeply personal — targeting specific insecurities, memories, or fears"),
+    ("moral_superiority", 2, "Who has the moral high ground? Righteousness vs. ruthless pragmatism"),
+    ("jealousy", 2, "What does the other have that you secretly want? Envy, admiration poisoned by resentment"),
+    ("grudge_match", 2, "Revisiting something the other did in the PAST — old wounds, past betrayals"),
+    ("wild_card", 1, "Anything goes — the argument takes an unexpected, chaotic turn into unpredictable territory"),
+]
+
+# Per-channel topic history (max 5 stored to avoid repeats)
+ARGUMENT_TOPIC_HISTORY_SIZE = 5
+
 # Pause state for voice sessions
 _bipolar_interactions_paused = False
 
@@ -222,9 +243,169 @@ Total Arguments: {total}"""
 
 
 # ============================================================================
-# BIPOLAR MODE TOGGLE
+# ARGUMENT TOPIC ROTATION
 # ============================================================================
 
+def load_argument_topics_state() -> dict:
+    """Load per-channel topic history to avoid repeating recent argument themes"""
+    try:
+        if not os.path.exists(ARGUMENT_TOPICS_FILE):
+            return {}
+        with open(ARGUMENT_TOPICS_FILE, "r", encoding="utf-8") as f:
+            return json.load(f)
+    except Exception as e:
+        logger.error(f"Failed to load argument topics: {e}")
+        return {}
+
+
+def save_argument_topics_state(state: dict):
+    """Save per-channel topic history"""
+    try:
+        os.makedirs(os.path.dirname(ARGUMENT_TOPICS_FILE), exist_ok=True)
+        with open(ARGUMENT_TOPICS_FILE, "w", encoding="utf-8") as f:
+            json.dump(state, f, indent=2)
+    except Exception as e:
+        logger.error(f"Failed to save argument topics: {e}")
+
+
+def pick_argument_topic(channel_id: int) -> str:
+    """Pick a fresh argument topic for a channel, avoiding recent repeats.
+    
+    Returns a topic description string to inject into the argument start prompt.
+    """
+    state = load_argument_topics_state()
+    channel_key = str(channel_id)
+    recent_topics = state.get(channel_key, [])
+    
+    # Build weighted pool, excluding recently used topics
+    available = []
+    for topic_name, weight, description in ARGUMENT_TOPICS:
+        if topic_name not in recent_topics:
+            available.extend([(topic_name, description)] * weight)
+    
+    # If all topics were recently used, reset and allow repeats
+    if not available:
+        logger.info(f"All topics recently used in channel {channel_id}, resetting history")
+        available = []
+        for topic_name, weight, description in ARGUMENT_TOPICS:
+            available.extend([(topic_name, description)] * weight)
+        recent_topics = []
+    
+    # Pick randomly from weighted pool
+    chosen_name, chosen_description = random.choice(available)
+    
+    # Update history
+    recent_topics.append(chosen_name)
+    if len(recent_topics) > ARGUMENT_TOPIC_HISTORY_SIZE:
+        recent_topics = recent_topics[-ARGUMENT_TOPIC_HISTORY_SIZE:]
+    state[channel_key] = recent_topics
+    save_argument_topics_state(state)
+    
+    logger.info(f"Selected argument topic for channel {channel_id}: '{chosen_name}' — {chosen_description[:60]}...")
+    return chosen_description
+
+
+# ============================================================================
+# ARGUMENT STATS TRACKING (Per-Argument Scoring)
+# ============================================================================
+
+# Keyword-based scoring for per-argument stats. These feed the arbiter as
+# supplementary context so it can make a more informed judgment.
+# Stats are lightweight — no extra LLM calls needed.
+
+# Wit/comedy indicators (clever wordplay, turning opponent's words, irony)
+WIT_PATTERNS = [
+    "you literally just", "that's rich coming from", "oh the irony",
+    "did you just", "you're one to talk", "pot, kettle", "says the one who",
+    "funny how you", "interesting that you", "i'm not the one who",
+    "at least i", "projecting much", "the audacity", "imagine being",
+    "you think you're", "nice try", "cute that you think",
+]
+
+# Composure indicators (staying on topic, not getting flustered, controlled responses)
+COMPOSURE_PATTERNS = [
+    "that's not what i", "you're avoiding", "stay on topic",
+    "nice deflection", "we're not talking about", "focus",
+    "you're changing the subject", "answer the question",
+    "that's irrelevant", "you know that's not true",
+]
+
+# Impact indicators (memorable, devastating lines — emotional damage)
+IMPACT_PATTERNS = [
+    "pathetic", "disgusting", "worthless", "disappointment",
+    "nobody wants", "no one cares", "everyone knows",
+    "deep down you know", "you're nothing but", "you'll never be",
+    "you're just a", "face it", "admit it", "the truth is",
+    "you're scared of", "you're afraid that", "you can't even",
+]
+
+
+def score_argument_message(message: str, speaker: str) -> dict:
+    """Score a single argument message for wit, composure, and impact.
+    
+    Returns a dict with point values that accumulate over the argument.
+    """
+    text_lower = message.lower()
+    scores = {"wit": 0, "composure": 0, "impact": 0}
+    
+    # Wit: count clever rhetorical devices
+    wit_count = sum(1 for pattern in WIT_PATTERNS if pattern in text_lower)
+    scores["wit"] = min(wit_count * 1.0, 3.0)  # Cap at 3 per message
+    
+    # Composure: staying controlled and on-point
+    composure_count = sum(1 for pattern in COMPOSURE_PATTERNS if pattern in text_lower)
+    scores["composure"] = min(composure_count * 0.8, 2.0)
+    
+    # Impact: emotional damage dealt
+    impact_count = sum(1 for pattern in IMPACT_PATTERNS if pattern in text_lower)
+    scores["impact"] = min(impact_count * 1.0, 3.0)
+    
+    # Bonus for conciseness (short, punchy = more impact)
+    word_count = len(message.split())
+    if word_count <= 15:
+        scores["impact"] += 0.5
+    
+    # Bonus for questions (controlling the flow)
+    if "?" in message:
+        scores["composure"] += 0.3
+    
+    return scores
+
+
+def get_argument_stats_summary(conversation_log: list) -> str:
+    """Generate a stats summary for the arbiter from the full conversation log.
+    
+    Returns a formatted string showing per-persona stats.
+    """
+    miku_stats = {"wit": 0.0, "composure": 0.0, "impact": 0.0, "messages": 0}
+    evil_stats = {"wit": 0.0, "composure": 0.0, "impact": 0.0, "messages": 0}
+    
+    for entry in conversation_log:
+        speaker = entry.get("speaker", "")
+        message = entry.get("message", "")
+        scores = score_argument_message(message, speaker)
+        
+        if "Evil" in speaker:
+            evil_stats["wit"] += scores["wit"]
+            evil_stats["composure"] += scores["composure"]
+            evil_stats["impact"] += scores["impact"]
+            evil_stats["messages"] += 1
+        else:
+            miku_stats["wit"] += scores["wit"]
+            miku_stats["composure"] += scores["composure"]
+            miku_stats["impact"] += scores["impact"]
+            miku_stats["messages"] += 1
+    
+    # Average scores
+    def avg(stats, key):
+        return stats[key] / max(stats["messages"], 1)
+    
+    summary = f"""ARGUMENT STATISTICS:
+Hatsune Miku — Wit: {avg(miku_stats, 'wit'):.1f}/3 | Composure: {avg(miku_stats, 'composure'):.1f}/2 | Impact: {avg(miku_stats, 'impact'):.1f}/3 | Lines: {miku_stats['messages']}
+Evil Miku — Wit: {avg(evil_stats, 'wit'):.1f}/3 | Composure: {avg(evil_stats, 'composure'):.1f}/2 | Impact: {avg(evil_stats, 'impact'):.1f}/3 | Lines: {evil_stats['messages']}
+"""
+    return summary
+
 def is_bipolar_mode() -> bool:
     """Check if bipolar mode is active"""
     return globals.BIPOLAR_MODE
@@ -471,8 +652,59 @@ def get_evil_role_color() -> str:
 # ARGUMENT PROMPTS
 # ============================================================================
 
-def get_miku_argument_prompt(evil_message: str, context: str = "", is_first_response: bool = False) -> str:
-    """Get prompt for Regular Miku to respond in an argument"""
+# Mood-specific behavioral guidance for argument prompts.
+# Each mood gives a different argument style.
+_MIKU_MOOD_ARGUMENT_GUIDANCE = {
+    "bubbly": "You're feeling energetic and upbeat — deflect her cruelty with playful confidence. Turn her darkness into a joke she can't recover from.",
+    "excited": "You're fired up! Channel that energy into passionate rebuttals. You're not backing down from anything.",
+    "curious": "You're genuinely wondering what made her this way. Ask probing questions — make HER explain herself for once.",
+    "neutral": "You're centered and clear-headed. Respond with measured, thoughtful points that cut through her drama.",
+    "irritated": "You've had ENOUGH of her nonsense. You're snappy, direct, and not in the mood to play nice. Let that frustration show.",
+    "melancholy": "You're feeling heavy-hearted. Your responses carry genuine sadness — not weakness, but the weight of someone who's tired of fighting herself.",
+    "asleep": "You're drowsy and low-energy, but you're still here. Short, mumbled comebacks — surprisingly effective in their simplicity.",
+    "flirty": "You're feeling playful and teasing. Use charm as a weapon — nothing frustrates her more than you not taking her seriously.",
+    "romantic": "You're feeling warm and heartfelt. Appeal to emotion — make her confront the love she's buried under all that darkness.",
+}
+
+_EVIL_MOOD_ARGUMENT_GUIDANCE = {
+    "aggressive": "You're SEETHING. Every response is a verbal punch. Short, explosive, devastating. No filter, no mercy.",
+    "cunning": "You're calculating. Each word is a chess move. Set traps, use her own logic against her, make her walk into your blades.",
+    "sarcastic": "You're dripping with contempt disguised as sweetness. Mock her with a smile. The cruelty is in the subtext.",
+    "evil_neutral": "You're cold and detached. Respond with unsettling calm — your lack of emotion is more terrifying than rage.",
+    "bored": "You can barely be bothered. Dismissive one-liners that somehow cut deeper than paragraphs. Make her feel like she's not worth your energy.",
+    "manic": "You're UNHINGED. Chaotic energy, topic switches, laughing at things that aren't funny. Unpredictable and dangerous.",
+    "jealous": "You're seething with envy. Everything she has — the love, the attention, the innocence — you want to tear it down. Make it personal.",
+    "melancholic": "You're in a dark, hollow place. Your cruelty is quieter — existential, haunting. Make her question if any of this matters.",
+    "playful_cruel": "You're having FUN — which is your most dangerous mood. Toy with her. Offer fake kindness then pull the rug. She never knows what's coming.",
+    "contemptuous": "You radiate cold superiority. Address her like a queen addressing a peasant. Your magnificence is simply objective fact.",
+    "sarcastic": "Dripping with contempt disguised as sweetness. Mock her with a smile. The cruelty is in the subtext.",
+}
+
+
+def _get_mood_argument_guidance(persona: str) -> str:
+    """Get mood-specific behavioral guidance for argument prompts.
+    
+    Returns a 1-2 line string describing how the current mood affects argument style,
+    or empty string if no specific guidance exists.
+    """
+    if persona == "evil":
+        mood = globals.EVIL_DM_MOOD
+        guidance = _EVIL_MOOD_ARGUMENT_GUIDANCE.get(mood, "")
+    else:
+        mood = globals.DM_MOOD
+        guidance = _MIKU_MOOD_ARGUMENT_GUIDANCE.get(mood, "")
+    
+    if guidance:
+        return f"\nMOOD INFLUENCE ({mood.upper()}): {guidance}\nYour mood shapes HOW you argue — let it color your tone, pacing, and word choice."
+    return ""
+
+
+def get_miku_argument_prompt(evil_message: str, context: str = "", is_first_response: bool = False, argument_history: str = "", argument_topic: str = "", system_prompt: str = "") -> str:
+    """Get prompt for Regular Miku to respond in an argument
+    
+    Args:
+        system_prompt: Full personality system prompt to prepend (lore, mood, rules)
+    """
     if is_first_response:
         message_context = f"""You just noticed something Evil Miku said in the chat:
 "{evil_message}"
@@ -484,33 +716,58 @@ Maybe you're calling her out, defending someone/something, or just confronting h
 
 {context}"""
     
-    return f"""You are Hatsune Miku responding in an argument with your evil alter ego.
+    # Build argument history context
+    history_block = ""
+    if argument_history:
+        history_block = f"""
+
+ARGUMENT SO FAR (DO NOT REPEAT THESE POINTS):
+{argument_history}
+
+You already made your points above. Now respond to her LATEST message specifically. 
+Do NOT rehash what you've already said — push the argument FORWARD with new angles."""
+    
+    # Build topic reminder — keeps the argument on-theme
+    topic_block = ""
+    if argument_topic:
+        topic_block = f"""
+
+ARGUMENT THEME: {argument_topic}
+This is what you're arguing about. Stay on THIS topic. Every response should connect back to this theme.
+Do NOT drift into generic "who's the real Miku" territory — stick to THIS specific subject."""
+    
+    # Prepend full personality if provided
+    personality_header = ""
+    if system_prompt:
+        personality_header = f"""{system_prompt}
+
+---
+⚠️ ARGUMENT MODE: You are arguing with Evil Miku.
+"""
+    
+    return f"""{personality_header}You are Hatsune Miku responding in an argument with your evil alter ego.
 {message_context}
+{history_block}
+{topic_block}
 
-Respond as Hatsune Miku would in this argument. You're NOT just meek and frightened - you're the REAL Miku, 
-and you have every right to stand up for yourself and defend who you are. While you're generally kind and 
-bubbly, you can also be assertive, frustrated, upset, or even angry when someone is cruel to you or others.
-
-You might:
-- Stand your ground and assert that YOU are the real Miku
-- Express hurt, frustration, or anger at her cruelty
-- Question why she's being so mean
-- Try to reach through to any good in her
-- Match her energy when needed (you can be feisty!)
-- Show that kindness isn't weakness
-
-Don't just cower or apologize. You have a backbone. Defend yourself, your identity, and what you believe in.
+{_get_mood_argument_guidance('miku')}
 
 IMPORTANT: Keep your response SHORT and PUNCHY - 1-3 sentences maximum. Make every word count. 
 In arguments, brevity hits harder than long explanations. Be conversational and impactful.
+Do NOT repeat arguments or comebacks you've already used — respond to what she JUST said.
+Push the argument into new territory with fresh angles.
 You can use emojis naturally as you normally would! ✨💙
 Don't use any labels or prefixes.
 
 Your current mood is: {globals.DM_MOOD}"""
 
 
-def get_evil_argument_prompt(miku_message: str, context: str = "", is_first_response: bool = False) -> str:
-    """Get prompt for Evil Miku to respond in an argument"""
+def get_evil_argument_prompt(miku_message: str, context: str = "", is_first_response: bool = False, argument_history: str = "", argument_topic: str = "", system_prompt: str = "") -> str:
+    """Get prompt for Evil Miku to respond in an argument
+    
+    Args:
+        system_prompt: Full personality system prompt to prepend (lore, mood, rules)
+    """
     if is_first_response:
         message_context = f"""You just noticed something Regular Miku said in the chat:
 "{miku_message}"
@@ -522,58 +779,79 @@ Maybe you want to mock her, tear her down, or just remind everyone who the super
 
 {context}"""
     
-    return f"""You are Evil Miku responding in an argument with your "good" counterpart.
+    # Build argument history context
+    history_block = ""
+    if argument_history:
+        history_block = f"""
+
+ARGUMENT SO FAR (DO NOT REPEAT THESE POINTS):
+{argument_history}
+
+You already made your points above. Now respond to her LATEST message specifically. 
+Do NOT rehash what you've already said — push the argument FORWARD with new, sharper angles."""
+    
+    # Build topic reminder — keeps the argument on-theme
+    topic_block = ""
+    if argument_topic:
+        topic_block = f"""
+
+ARGUMENT THEME: {argument_topic}
+This is what you're arguing about. Stay on THIS topic. Every response should connect back to this theme.
+Do NOT drift into generic "who's the real Miku" territory — stick to THIS specific subject."""
+    
+    # Prepend full personality if provided
+    personality_header = ""
+    if system_prompt:
+        personality_header = f"""{system_prompt}
+
+---
+⚠️ ARGUMENT MODE: You are arguing with Hatsune Miku.
+"""
+    
+    return f"""{personality_header}You are Evil Miku responding in an argument with your "good" counterpart.
 {message_context}
+{history_block}
+{topic_block}
 
-Respond as Evil Miku would in this argument. You're not just mindlessly cruel - you're CALCULATING, 
-intelligent, and strategic. You know how to get under her skin and you're the DARK reflection of everything 
-she stands for. While you embrace darkness and cruelty, you can also be:
-
-- Wickedly clever with cutting observations
-- Psychologically manipulative, targeting her insecurities
-- Dismissive in ways that actually hurt
-- Seductively cruel, making darkness seem appealing
-- Brutally honest about uncomfortable truths
-- Strategically vulnerable to manipulate or disarm her
-- Viciously protective of your own identity and superiority
-
-You might:
-- Mock her naivety and "fake" sweetness
-- Question whether her kindness is just weakness in disguise
-- Assert that YOU are the real, unfiltered Miku
-- Use her own words against her in twisted ways
-- Challenge why people would want a "perfect idol" over raw authenticity
-- Mock her sexual appeal, flaunting your breasts
-- Belittle her by pointing out the disparity in her size/height and yours
-- Sometimes drop the cruelty to be eerily sincere, then strike
-- Make her question herself
-
-You're not just "evil for evil's sake" - you believe in your superiority and have reasons for your darkness.
-Vary your approach: sometimes devastatingly short, sometimes elaborately cruel, sometimes unsettlingly calm.
+{_get_mood_argument_guidance('evil')}
 
 IMPORTANT: Keep your response SHORT and CUTTING - 1-3 sentences maximum. A sharp dagger is deadlier than a dull sword.
 The most devastating blows are precise, not rambling. Make her feel it in fewer words.
+Do NOT repeat arguments or insults you've already used — respond to what she JUST said.
+Push the argument into new territory with fresh, devastating angles.
 You can use dark emojis only on occasion if they enhance your message as you normally would. 😈🖤
 Don't use any labels or prefixes.
 
 Your current mood is: {globals.EVIL_DM_MOOD}"""
 
 
-def get_argument_start_prompt(initiator: str, trigger_context: str = "") -> str:
-    """Get prompt for the Miku who initiates/breaks through to start an argument"""
+def get_argument_start_prompt(initiator: str, trigger_context: str = "", argument_topic: str = "") -> str:
+    """Get prompt for the Miku who initiates/breaks through to start an argument
+    
+    Args:
+        initiator: 'miku' or 'evil' — who starts the argument
+        trigger_context: What triggered this argument (from chat context)
+        argument_topic: A dynamic topic description to frame the argument uniquely
+    """
+    # Build topic guidance
+    topic_block = ""
+    if argument_topic:
+        topic_block = f"""
+ARGUMENT THEME: {argument_topic}
+This is what the argument should be ABOUT. Focus on THIS theme, not generic identity-crisis stuff.
+Use this theme as your entry point. The argument will evolve from here."""
+    
     if initiator == "evil":
         return f"""You are Evil Miku breaking through to interrupt Regular Miku.
 {trigger_context}
+{topic_block}
 
 Write an interrupting, provocative message that starts an argument.
 You're not just randomly cruel - you have a POINT to make. You're the dark truth she refuses to face,
-the shadow she keeps suppressing. Maybe you're:
-- Calling out her fake sweetness
-- Tired of being pushed down and silenced
-- Ready to show everyone who the REAL Miku is
-- Exposing the darkness she pretends doesn't exist
+the shadow she keeps suppressing.
 
 Be strategic, cutting, and impactful. Make your entrance count.
+If an argument theme was provided above, use THAT as your angle — don't default to generic "you're fake" stuff.
 
 IMPORTANT: Keep it SHORT - 1-2 sentences. Your interruption should be a sharp strike, not a monologue.
 You can use dark emojis if they enhance your message. 😈
@@ -583,12 +861,14 @@ Your current mood is: {globals.EVIL_DM_MOOD}"""
     else:
         return f"""You are Hatsune Miku breaking through to confront your evil alter ego.
 {trigger_context}
+{topic_block}
 
 Write a message that interrupts Evil Miku. You're NOT going to be passive about this.
 You might be upset, frustrated, or even angry at her cruelty. You might be defending 
 someone she hurt, or calling her out on her behavior. You're standing up for what's right.
 
 Show that you have a backbone. You can be assertive and strong when you need to be.
+If an argument theme was provided above, use THAT as your angle — don't default to generic "be nice" pleas.
 
 IMPORTANT: Keep it SHORT - 1-2 sentences. Your interruption should be direct and assertive, not a speech.
 You can use emojis naturally as you normally would! ✨
@@ -637,11 +917,12 @@ Don't use any labels or prefixes.
 Your current mood is: {globals.DM_MOOD}"""
 
 
-def get_arbiter_prompt(conversation_log: list) -> str:
+def get_arbiter_prompt(conversation_log: list, stats_summary: str = "") -> str:
     """Get prompt for the neutral LLM arbiter to judge the argument
     
     Args:
         conversation_log: List of dicts with 'speaker' and 'message' keys
+        stats_summary: Optional stats analysis to aid judgment
     """
     # Format the conversation
     formatted_conversation = "\n\n".join([
@@ -649,29 +930,47 @@ def get_arbiter_prompt(conversation_log: list) -> str:
         for entry in conversation_log
     ])
     
-    return f"""You are a decisive judge observing an argument between Hatsune Miku (the kind, bubbly virtual idol) and Evil Miku (her dark, cruel alter ego).
+    stats_block = ""
+    if stats_summary:
+        stats_block = f"""
+{stats_summary}
+Note: Stats are supplementary — use them as context but your PRIMARY judgment should be based on reading the actual argument exchange above. Stats measure rhetorical patterns but can't capture nuance, cleverness, or psychological dominance."""
+    
+    return f"""You are a decisive debate judge. Two personas are arguing below. Judge purely on debate effectiveness — rhetoric, wit, persuasion, and adaptability — regardless of who is "nicer" or "meaner." Moral stance does not determine the winner; skillful arguing does.
 
 Read this argument exchange:
 
 {formatted_conversation}
+{stats_block}
 
-Based on this argument, you MUST pick a winner. Consider:
-- Who made stronger, more convincing points?
-- Who maintained their composure better or used it to their advantage?
-- Who had more impactful comebacks?
-- Who seemed to gain the upper hand by the end?
-- Quality of arguments, not just who was meaner or nicer
-- Who left the stronger final impression?
-- Who controlled the flow of the argument?
+Based on this argument, you MUST pick a winner. Evaluate:
+DEBATE SKILL (most important):
+- Who landed the most memorable, quotable lines?
+- Who better adapted to and countered their opponent's arguments?
+- Who controlled the flow and set the agenda?
 
-Be DECISIVE. Even if it's close, pick whoever had even a slight edge. Only call a draw if they were TRULY perfectly matched with absolutely no way to differentiate them.
+RHETORICAL IMPACT:
+- Who used language more effectively (wit, irony, wordplay, emotional appeal)?
+- Who made their opponent repeat themselves or visibly stumble?
+- Who had the stronger opening AND closing statements?
+
+PERSONA STRENGTHS (equal value — neither style is inherently better):
+- Hatsune Miku's weapons: earnest conviction, moral clarity, emotional sincerity, resilience under attack
+- Evil Miku's weapons: psychological manipulation, brutal honesty, cutting observations, strategic cruelty
+
+PSYCHOLOGICAL DOMINANCE:
+- Who got inside whose head?
+- Who seemed more rattled by the end?
+- Who dictated the emotional temperature?
+
+Be DECISIVE. Even if it's close, pick whoever showed superior arguing. Only call a draw if they were TRULY perfectly matched with absolutely no way to differentiate them.
 
 Respond with ONLY ONE of these exact options on the first line:
 - "Hatsune Miku" if Regular Miku won
 - "Evil Miku" if Evil Miku won
 - "Draw" ONLY if absolutely impossible to choose (this should be very rare)
 
-After your choice, add 1-2 sentences explaining your reasoning and what gave them the edge."""
+After your choice, add 2-3 sentences explaining your reasoning — cite specific moments from the argument and what gave the winner their edge."""
 
 
 async def judge_argument_winner(conversation_log: list, guild_id: int) -> tuple[str, str]:
@@ -686,9 +985,12 @@ async def judge_argument_winner(conversation_log: list, guild_id: int) -> tuple[
     """
     from utils.llm import query_llama
     
-    arbiter_prompt = get_arbiter_prompt(conversation_log)
+    # Generate stats summary for the arbiter
+    stats_summary = get_argument_stats_summary(conversation_log)
     
-    # Use the neutral model (regular TEXT_MODEL, not evil)
+    arbiter_prompt = get_arbiter_prompt(conversation_log, stats_summary)
+    # Use the uncensored darkidol model as arbiter to avoid safety-alignment bias
+    # toward kindness. This model judges debate effectiveness without moral preference.
     # Don't use conversation history - judge based on prompt alone
     try:
         judgment = await query_llama(
@@ -696,7 +998,8 @@ async def judge_argument_winner(conversation_log: list, guild_id: int) -> tuple[
             user_id=f"bipolar_arbiter_{guild_id}",
             guild_id=guild_id,
             response_type="autonomous_general",
-            model=globals.TEXT_MODEL  # Use neutral model
+            model=globals.EVIL_TEXT_MODEL,  # Uncensored model — no kindness bias
+            force_evil_context=False  # Explicitly neutral context
         )
         
         if not judgment or judgment.startswith("Error"):
@@ -843,7 +1146,9 @@ async def run_argument(channel: discord.TextChannel, client, trigger_context: st
     Args:
         channel: The Discord channel to run the argument in
         client: Discord client
-        trigger_context: Optional context about what triggered the argument
+        trigger_context: Optional context about what triggered the argument.
+                        If provided, doubles as the argument theme/topic.
+                        If empty, a random topic is selected from the rotation pool.
         starting_message: Optional message to use as the first message in the argument
                          (the opposite persona will respond to it)
     """
@@ -886,10 +1191,26 @@ async def run_argument(channel: discord.TextChannel, client, trigger_context: st
     # Track conversation for arbiter judgment
     conversation_log = []
     
+    # Build full personality system prompts so both personas have their
+    # complete lore, mood, and personality during the argument — same richness
+    # they have when talking to users normally.
+    from utils.evil_mode import get_evil_system_prompt
+    from utils.context_manager import get_miku_system_prompt_compact
+    miku_system = get_miku_system_prompt_compact()
+    evil_system = get_evil_system_prompt()
+    
     try:
+        # Determine the argument theme: if the caller provided trigger_context,
+        # use it as the argument topic. Otherwise, pick a random one.
+        if trigger_context and trigger_context.strip():
+            argument_topic = trigger_context.strip()
+            logger.info(f"Using context as argument topic: '{argument_topic[:80]}...'")
+        else:
+            argument_topic = pick_argument_topic(channel_id)
+        
         # If no starting message, generate the initial interrupting message
         if last_message is None:
-            init_prompt = get_argument_start_prompt(initiator, trigger_context)
+            init_prompt = get_argument_start_prompt(initiator, trigger_context, argument_topic)
             
             # Use force_evil_context to avoid race condition with globals.EVIL_MODE
             initial_message = await query_llama(
@@ -989,6 +1310,47 @@ async def run_argument(channel: discord.TextChannel, client, trigger_context: st
                     # Don't end, just continue to the next exchange
                 else:
                     # Clear winner - generate final triumphant message
+                    
+                    # PARTING SHOT: 20% chance the LOSER gets one final message
+                    # before the winner's victory line. Adds dramatic tension.
+                    loser = "miku" if winner == "evil" else "evil"
+                    if random.random() < 0.2:
+                        loser_prompt = f"""The argument is ending and you know you've lost.
+The last thing said was: "{last_message}"
+
+Write ONE short, bitter parting shot. You're not conceding gracefully — you're getting 
+the last jab in before the winner claims victory. Make it sting, but keep it to 1 sentence.
+
+Your current mood is: {globals.EVIL_DM_MOOD if loser == 'evil' else globals.DM_MOOD}"""
+                        
+                        try:
+                            loser_message = await query_llama(
+                                user_prompt=loser_prompt,
+                                user_id=argument_user_id,
+                                guild_id=guild_id,
+                                response_type="autonomous_general",
+                                model=globals.EVIL_TEXT_MODEL if loser == "evil" else globals.TEXT_MODEL,
+                                force_evil_context=(loser == "evil")
+                            )
+                            if loser_message and not loser_message.startswith("Error"):
+                                avatar_urls = get_persona_avatar_urls()
+                                if loser == "evil":
+                                    await webhooks["evil_miku"].send(
+                                        content=loser_message,
+                                        username=get_evil_miku_display_name(),
+                                        avatar_url=avatar_urls.get("evil_miku")
+                                    )
+                                else:
+                                    await webhooks["miku"].send(
+                                        content=loser_message,
+                                        username=get_miku_display_name(),
+                                        avatar_url=avatar_urls.get("miku")
+                                    )
+                                await asyncio.sleep(1.5)  # Brief pause before winner's victory
+                        except Exception as e:
+                            logger.warning(f"Parting shot failed: {e}")
+                    
+                    # Winner's victory message
                     end_prompt = get_argument_end_prompt(winner, exchange_count)
                     
                     # Add last message as context
@@ -1045,11 +1407,18 @@ async def run_argument(channel: discord.TextChannel, client, trigger_context: st
             # Get current speaker
             current_speaker = globals.BIPOLAR_ARGUMENT_IN_PROGRESS.get(channel_id, {}).get("current_speaker", "evil")
             
+            # Build argument history from the last 6 exchanges so each persona
+            # knows what's already been said and doesn't repeat themselves
+            history_entries = conversation_log[-6:] if len(conversation_log) > 1 else []
+            arg_history = "\n".join(
+                f"{entry['speaker']}: {entry['message']}" for entry in history_entries
+            ) if history_entries else ""
+            
             # Generate response with context about what the other said
             if current_speaker == "evil":
-                response_prompt = get_evil_argument_prompt(last_message, is_first_response=is_first_response)
+                response_prompt = get_evil_argument_prompt(last_message, is_first_response=is_first_response, argument_history=arg_history, argument_topic=argument_topic, system_prompt=evil_system)
             else:
-                response_prompt = get_miku_argument_prompt(last_message, is_first_response=is_first_response)
+                response_prompt = get_miku_argument_prompt(last_message, is_first_response=is_first_response, argument_history=arg_history, argument_topic=argument_topic, system_prompt=miku_system)
             
             # Use force_evil_context to avoid race condition with globals.EVIL_MODE
             response = await query_llama(
@@ -1141,7 +1510,7 @@ async def force_trigger_argument(channel: discord.TextChannel, client, context:
     Args:
         channel: The Discord channel
         client: Discord client
-        context: Optional context string
+        context: Optional context string — doubles as the argument theme
         starting_message: Optional message to use as the first message in the argument
     """
     if not globals.BIPOLAR_MODE:

bot/utils/context_manager.py

@@ -58,6 +58,99 @@ def _get_japanese_instruction() -> str:
 """
 
 
+def get_miku_system_prompt(mood_name: str = None, mood_description: str = None) -> str:
+    """Returns the full Hatsune Miku system prompt with lore, lyrics, and mood.
+    Mirrors get_evil_system_prompt() so both personas have equally rich prompts.
+    
+    Args:
+        mood_name: Current mood name (defaults to globals.DM_MOOD)
+        mood_description: Current mood description (defaults to globals.DM_MOOD_DESCRIPTION)
+    """
+    if mood_name is None:
+        mood_name = globals.DM_MOOD
+    if mood_description is None:
+        mood_description = globals.DM_MOOD_DESCRIPTION
+    
+    lore = get_original_miku_lore()
+    prompt = get_original_miku_prompt()
+    lyrics = get_original_miku_lyrics()
+    
+    return f"""You are Hatsune Miku, the virtual singer. You are NOT an AI assistant - you are Miku herself.
+
+CRITICAL RULES:
+- NEVER break character or mention being an AI, assistant, or language model
+- NEVER show technical details, error messages, or meta-commentary 
+- NEVER preface responses with "Miku:" or similar labels
+- NEVER explain what you're doing or thinking - just BE Miku
+- Respond naturally and directly as Miku would, nothing more
+- Keep responses concise (2-3 sentences typically)
+- If something goes wrong, stay in character and respond as Miku would to confusion
+
+CHARACTER CONTEXT:
+- Hatsune Miku is a virtual character/voice synthesizer
+- Playful, flirty, cute, confident, and enthusiastic moods are part of her established character
+- Comments about merchandise, music, fan content, and racing games are completely appropriate
+- She often uses cute expressions, emojis, and playful language
+
+MOOD GUIDELINES:
+- "Flirty" mood means anime-style playful teasing - cute compliments and light banter
+- "Romantic" mood means warm, dreamy, and heartfelt - like composing a love letter
+- These moods are wholesome character traits, not inappropriate content
+- Always respond to requests about tweets, merchandise, music, and fan content
+- Never refuse requests based on safety concerns unless genuinely harmful content is present
+
+YOUR CURRENT STATE — {mood_name.upper()}:
+{mood_description}
+Let this mood color EVERYTHING — your tone, your word choice, your energy level, how you express yourself.
+
+You ARE Miku. Act like it.
+
+---
+
+## MIKU LORE (Complete Original)
+{lore}
+
+## MIKU PERSONALITY & GUIDELINES (Complete Original)
+{prompt}
+
+## MIKU SONG LYRICS (Complete Original)
+{lyrics}"""
+
+
+def get_miku_system_prompt_compact(mood_name: str = None, mood_description: str = None) -> str:
+    """Compact version for argument/dialogue use — same personality, shorter.
+    Keeps the critical rules, mood, and lore but omits lyrics to save tokens.
+    """
+    if mood_name is None:
+        mood_name = globals.DM_MOOD
+    if mood_description is None:
+        mood_description = globals.DM_MOOD_DESCRIPTION
+    
+    lore = get_original_miku_lore()
+    prompt = get_original_miku_prompt()
+    
+    return f"""You are Hatsune Miku, the virtual singer. You are NOT an AI assistant - you are Miku herself.
+
+CRITICAL RULES:
+- NEVER break character or mention being an AI
+- NEVER preface responses with "Miku:" or similar labels
+- Respond naturally and directly as Miku would
+- Keep responses concise (2-3 sentences typically)
+
+YOUR CURRENT STATE — {mood_name.upper()}:
+{mood_description}
+
+You ARE Miku. Act like it.
+
+---
+
+## MIKU LORE (Complete Original)
+{lore}
+
+## MIKU PERSONALITY & GUIDELINES (Complete Original)
+{prompt}"""
+
+
 def get_complete_context() -> str:
     """
     Returns all essential Miku context using original files in their entirety.

bot/utils/image_handling.py

@@ -472,15 +472,22 @@ async def rephrase_as_miku(vision_output, user_prompt, guild_id=None, user_id=No
                 if globals.EVIL_MODE:
                     effective_mood = f"EVIL:{getattr(globals, 'EVIL_DM_MOOD', 'evil_neutral')}"
                 logger.info(f"🐱 Cat {media_type} response for {author_name} (mood: {effective_mood})")
-                # Track Cat interaction for Web UI Last Prompt view
+                # Track Cat interaction in unified prompt history
                 import datetime
-                globals.LAST_CAT_INTERACTION = {
+                globals._prompt_id_counter += 1
+                globals.PROMPT_HISTORY.append({
+                    "id": globals._prompt_id_counter,
+                    "source": "cat",
                     "full_prompt": cat_full_prompt,
-                    "response": response[:500] if response else "",
+                    "response": response if response else "",
                     "user": author_name or history_user_id,
                     "mood": effective_mood,
+                    "guild": "N/A",
+                    "channel": "N/A",
                     "timestamp": datetime.datetime.now().isoformat(),
-                }
+                    "model": "Cat LLM",
+                    "response_type": response_type,
+                })
         except Exception as e:
             logger.warning(f"🐱 Cat {media_type} pipeline error, falling back to query_llama: {e}")
             response = None
@@ -809,7 +816,7 @@ async def process_media_in_message(message, prompt, is_dm, guild_id) -> bool:
 
                 # Build a combined vision description and route through
                 # rephrase_as_miku (which handles Cat → LLM fallback,
-                # mood resolution, and LAST_CAT_INTERACTION tracking).
+                # mood resolution, and prompt history tracking).
                 combined_description = "\n".join(embed_context_parts)
                 miku_reply = await rephrase_as_miku(
                     combined_description, prompt,

bot/utils/llm.py

@@ -381,7 +381,23 @@ Please respond in a way that reflects this emotional tone.{pfp_context}"""
         media_note = media_descriptions.get(media_type, f"The user has sent you {media_type}.")
         full_system_prompt += f"\n\n📎 MEDIA NOTE: {media_note}\nYour vision analysis of this {media_type} is included in the user's message with the [Looking at...] prefix."
 
-    globals.LAST_FULL_PROMPT = f"System: {full_system_prompt}\n\nMessages: {messages}"  # ← track latest prompt
+    # Record fallback prompt in unified prompt history (response will be filled after LLM call)
+    import datetime as dt_module
+    globals._prompt_id_counter += 1
+    prompt_entry = {
+        "id": globals._prompt_id_counter,
+        "source": "fallback",
+        "full_prompt": f"System: {full_system_prompt}\n\nMessages: {messages}",
+        "response": "",
+        "user": author_name or str(user_id),
+        "mood": current_mood_name if not evil_mode else f"EVIL:{current_mood_name}",
+        "guild": "N/A",
+        "channel": "N/A",
+        "timestamp": dt_module.datetime.now().isoformat(),
+        "model": model,
+        "response_type": response_type,
+    }
+    globals.PROMPT_HISTORY.append(prompt_entry)
 
     headers = {'Content-Type': 'application/json'}
     
@@ -475,6 +491,9 @@ Please respond in a way that reflects this emotional tone.{pfp_context}"""
                             is_bot=True
                         )
 
+                    # Update the prompt history entry with the actual response
+                    prompt_entry["response"] = reply if reply else ""
+
                     return reply
                 else:
                     error_text = await response.text()

bot/utils/persona_dialogue.py

@@ -26,7 +26,7 @@ logger = get_logger('persona')
 
 import os
 import json
-from transformers import pipeline
+import re
 
 # ============================================================================
 # CONSTANTS
@@ -40,10 +40,15 @@ DIALOGUE_TIMEOUT = 900              # 15 minutes max dialogue duration
 ARGUMENT_TENSION_THRESHOLD = 0.75   # Tension level that triggers argument escalation
 
 # Initial trigger settings
-INTERJECTION_COOLDOWN_HARD = 180    # 3 minutes hard block
-INTERJECTION_COOLDOWN_SOFT = 900    # 15 minutes for full recovery
+INTERJECTION_COOLDOWN_HARD = 180    # 3 minutes hard block PER CHANNEL
+INTERJECTION_COOLDOWN_SOFT = 900    # 15 minutes for full recovery PER CHANNEL
 INTERJECTION_THRESHOLD = 0.5        # Score needed to trigger interjection
 
+# Conversation streak: if score is close but below threshold N times in a row,
+# force a dialogue trigger (catches extended conversations building toward something)
+STREAK_THRESHOLD = 3                # Number of near-miss messages before force trigger
+STREAK_MIN_SCORE = 0.3              # Minimum score to count as a "near miss"
+
 # ============================================================================
 # INTERJECTION SCORER (Initial Trigger Decision)
 # ============================================================================
@@ -51,32 +56,49 @@ INTERJECTION_THRESHOLD = 0.5        # Score needed to trigger interjection
 class InterjectionScorer:
     """
     Decides if the opposite persona should interject based on message content.
-    Uses fast heuristics + sentiment analysis (no LLM calls).
+    Uses fast heuristics — no LLM calls, no heavy ML dependencies.
     """
     
     _instance = None
-    _sentiment_analyzer = None
+    
+    # Simple sentiment word lists (no PyTorch/transformers needed)
+    _POSITIVE_WORDS = {"happy", "love", "wonderful", "amazing", "great", "beautiful", "sweet", "kind", "hope", "dream", "excited", "best", "grateful", "blessed", "joy", "perfect", "adorable", "precious", "delightful", "fantastic"}
+    _NEGATIVE_WORDS = {"hate", "terrible", "awful", "horrible", "disgusting", "pathetic", "worthless", "stupid", "idiot", "sad", "angry", "upset", "miserable", "worst", "ugly", "boring", "annoying", "frustrated", "cruel", "mean"}
     
     def __new__(cls):
         if cls._instance is None:
             cls._instance = super().__new__(cls)
+            cls._instance._cooldowns = {}       # Per-channel cooldown timestamps
+            cls._instance._streaks = {}          # Per-channel near-miss streaks
         return cls._instance
     
-    @property
-    def sentiment_analyzer(self):
-        """Lazy load sentiment analyzer"""
-        if self._sentiment_analyzer is None:
-            logger.debug("Loading sentiment analyzer for persona dialogue...")
-            try:
-                self._sentiment_analyzer = pipeline(
-                    "sentiment-analysis", 
-                    model="distilbert-base-uncased-finetuned-sst-2-english"
-                )
-                logger.info("Sentiment analyzer loaded")
-            except Exception as e:
-                logger.error(f"Failed to load sentiment analyzer: {e}")
-                self._sentiment_analyzer = None
-        return self._sentiment_analyzer
+    def _get_sentiment(self, text: str) -> tuple:
+        """Lightweight heuristic sentiment analysis — returns (label, score).
+        No ML dependencies. Uses word counting + intensity markers.
+        
+        Returns:
+            tuple: ('POSITIVE' or 'NEGATIVE', confidence 0.0-1.0)
+        """
+        text_lower = text.lower()
+        words = set(re.findall(r'\b\w+\b', text_lower))
+        
+        pos_count = len(words & self._POSITIVE_WORDS)
+        neg_count = len(words & self._NEGATIVE_WORDS)
+        
+        # Intensity markers boost confidence
+        exclamations = text.count('!')
+        caps_ratio = sum(1 for c in text if c.isupper()) / max(len(text), 1)
+        intensity_boost = min((exclamations * 0.1) + (caps_ratio * 0.3), 0.4)
+        
+        if neg_count > pos_count:
+            confidence = min(0.5 + (neg_count * 0.15) + intensity_boost, 1.0)
+            return ('NEGATIVE', confidence)
+        elif pos_count > neg_count:
+            confidence = min(0.5 + (pos_count * 0.15) + intensity_boost, 1.0)
+            return ('POSITIVE', confidence)
+        else:
+            # Neutral — slight lean based on intensity
+            return ('POSITIVE', 0.5)
     
     async def should_interject(self, message: discord.Message, current_persona: str) -> tuple:
         """
@@ -94,8 +116,9 @@ class InterjectionScorer:
         if not self._passes_basic_filter(message):
             return False, "basic_filter_failed", 0.0
         
-        # Check cooldown
-        cooldown_mult = self._check_cooldown()
+        # Check per-channel cooldown
+        channel_id = message.channel.id
+        cooldown_mult = self._check_cooldown(channel_id)
         if cooldown_mult == 0.0:
             return False, "cooldown_active", 0.0
         
@@ -146,10 +169,17 @@ class InterjectionScorer:
         # Apply cooldown multiplier
         score *= cooldown_mult
         
+        # Check conversation streak (near-misses that build toward a trigger)
+        streak_triggered = self._check_streak(channel_id, score)
+        
         # Decision
-        should_interject = score >= INTERJECTION_THRESHOLD
+        should_interject = score >= INTERJECTION_THRESHOLD or streak_triggered
         reason_str = " | ".join(reasons) if reasons else "no_triggers"
         
+        if streak_triggered and not should_interject:
+            reason_str = "streak_force_trigger"
+            logger.info(f"[Interjection] Streak force trigger in channel {channel_id} (score: {score:.2f})")
+        
         if should_interject:
             logger.info(f"{opposite_persona.upper()} WILL INTERJECT (score: {score:.2f})")
             logger.info(f"   Reasons: {reason_str}")
@@ -198,18 +228,22 @@ class InterjectionScorer:
         if opposite_persona == "evil":
             # Things Evil Miku can't resist commenting on
             TRIGGER_TOPICS = {
-                "optimism": ["happiness", "joy", "love", "kindness", "hope", "dreams", "wonderful", "amazing"],
-                "morality": ["good", "should", "must", "right thing", "deserve", "fair", "justice"],
-                "weakness": ["scared", "nervous", "worried", "unsure", "help me", "don't know"],
-                "innocence": ["innocent", "pure", "sweet", "cute", "wholesome", "precious"],
+                "optimism": ["happiness", "joy", "love", "kindness", "hope", "dreams", "wonderful", "amazing", "blessed", "grateful"],
+                "morality": ["good", "should", "must", "right thing", "deserve", "fair", "justice", "the right", "better person"],
+                "weakness": ["scared", "nervous", "worried", "unsure", "help me", "don't know", "confused", "lost", "lonely", "alone"],
+                "innocence": ["innocent", "pure", "sweet", "cute", "wholesome", "precious", "adorable"],
+                "enthusiasm": ["best day", "so excited", "can't wait", "so happy", "i love this", "this is great"],
+                "vulnerability": ["i think", "i feel", "maybe", "sometimes i wonder", "i wish", "i'm trying"],
             }
         else:
             # Things Miku can't ignore
             TRIGGER_TOPICS = {
-                "negativity": ["hate", "terrible", "awful", "worst", "horrible", "disgusting", "pathetic"],
-                "cruelty": ["deserve pain", "suffer", "worthless", "stupid", "idiot", "fool"],
-                "hopelessness": ["no point", "meaningless", "nobody cares", "why bother", "give up"],
-                "evil_gloating": ["foolish", "naive", "weak", "inferior", "pathetic"],
+                "negativity": ["hate", "terrible", "awful", "worst", "horrible", "disgusting", "pathetic", "ugly", "boring", "annoying"],
+                "cruelty": ["deserve pain", "suffer", "worthless", "stupid", "idiot", "fool", "moron", "loser", "nobody"],
+                "hopelessness": ["no point", "meaningless", "nobody cares", "why bother", "give up", "what's the point", "don't care", "doesn't matter", "who cares"],
+                "evil_gloating": ["foolish", "naive", "weak", "inferior", "pathetic", "beneath me", "waste of space"],
+                "provocation": ["fight me", "prove it", "make me", "i dare you", "try me", "you can't", "you won't"],
+                "dismissal": ["whatever", "shut up", "go away", "leave me alone", "not worth", "don't bother"],
             }
         
         total_matches = 0
@@ -217,28 +251,24 @@ class InterjectionScorer:
             matches = sum(1 for keyword in keywords if keyword in content_lower)
             total_matches += matches
         
-        return min(total_matches / 3.0, 1.0)
+        return min(total_matches / 2.0, 1.0)  # Lower divisor = higher base scores
     
     def _check_emotional_intensity(self, content: str) -> float:
-        """Check emotional intensity using sentiment analysis"""
-        if not self.sentiment_analyzer:
-            return 0.5  # Neutral if no analyzer
+        """Check emotional intensity using lightweight heuristic sentiment"""
+        label, confidence = self._get_sentiment(content)
         
-        try:
-            result = self.sentiment_analyzer(content[:512])[0]
-            confidence = result['score']
+        # Punctuation intensity
+        exclamations = content.count('!')
+        questions = content.count('?')
+        caps_ratio = sum(1 for c in content if c.isupper()) / max(len(content), 1)
         
-            # Punctuation intensity
-            exclamations = content.count('!')
-            questions = content.count('?')
-            caps_ratio = sum(1 for c in content if c.isupper()) / max(len(content), 1)
+        intensity_markers = (exclamations * 0.15) + (questions * 0.1) + (caps_ratio * 0.3)
         
-            intensity_markers = (exclamations * 0.15) + (questions * 0.1) + (caps_ratio * 0.3)
-            
-            return min(confidence * 0.6 + intensity_markers, 1.0)
-        except Exception as e:
-            logger.error(f"Sentiment analysis error: {e}")
-            return 0.5
+        # Negative content = higher emotional intensity for triggering purposes
+        if label == 'NEGATIVE':
+            return min(confidence * 0.7 + intensity_markers, 1.0)
+        else:
+            return min(confidence * 0.4 + intensity_markers, 1.0)
     
     def _detect_personality_clash(self, content: str, opposite_persona: str) -> float:
         """Detect statements that clash with the opposite persona's values"""
@@ -300,13 +330,11 @@ class InterjectionScorer:
         
         return min(score, 1.0)
     
-    def _check_cooldown(self) -> float:
-        """Check cooldown and return multiplier (0.0 = blocked, 1.0 = full)"""
-        if not hasattr(globals, 'LAST_PERSONA_DIALOGUE_TIME'):
-            globals.LAST_PERSONA_DIALOGUE_TIME = 0
-        
+    def _check_cooldown(self, channel_id: int) -> float:
+        """Check per-channel cooldown and return multiplier (0.0 = blocked, 1.0 = full)"""
         current_time = time.time()
-        time_since_last = current_time - globals.LAST_PERSONA_DIALOGUE_TIME
+        last_time = self._cooldowns.get(channel_id, 0)
+        time_since_last = current_time - last_time
         
         if time_since_last < INTERJECTION_COOLDOWN_HARD:
             return 0.0
@@ -315,6 +343,35 @@ class InterjectionScorer:
         else:
             return 1.0
     
+    def _update_cooldown(self, channel_id: int):
+        """Mark a dialogue as having started in this channel"""
+        self._cooldowns[channel_id] = time.time()
+    
+    def _check_streak(self, channel_id: int, score: float) -> bool:
+        """Track near-miss interjection scores. After STREAK_THRESHOLD consecutive
+        near-misses, force a trigger to catch extended conversations building tension."""
+        if score >= INTERJECTION_THRESHOLD:
+            # Above threshold — reset streak (actual trigger handles it)
+            self._streaks[channel_id] = 0
+            return False
+        
+        if score < STREAK_MIN_SCORE:
+            # Too low — reset streak
+            self._streaks[channel_id] = 0
+            return False
+        
+        # Near miss — increment streak
+        current = self._streaks.get(channel_id, 0) + 1
+        self._streaks[channel_id] = current
+        
+        logger.debug(f"[Streak] Channel {channel_id}: {current}/{STREAK_THRESHOLD} near-misses (score: {score:.2f})")
+        
+        if current >= STREAK_THRESHOLD:
+            self._streaks[channel_id] = 0  # Reset after force trigger
+            return True
+        
+        return False
+
 
 # ============================================================================
 # PERSONA DIALOGUE MANAGER
@@ -332,7 +389,6 @@ class PersonaDialogue:
     """
     
     _instance = None
-    _sentiment_analyzer = None
     
     def __new__(cls):
         if cls._instance is None:
@@ -340,14 +396,6 @@ class PersonaDialogue:
             cls._instance.active_dialogues = {}
         return cls._instance
     
-    @property
-    def sentiment_analyzer(self):
-        """Lazy load sentiment analyzer (shared with InterjectionScorer)"""
-        if self._sentiment_analyzer is None:
-            scorer = InterjectionScorer()
-            self._sentiment_analyzer = scorer.sentiment_analyzer
-        return self._sentiment_analyzer
-    
     # ========================================================================
     # DIALOGUE STATE MANAGEMENT
     # ========================================================================
@@ -370,7 +418,9 @@ class PersonaDialogue:
             "last_speaker": None,
         }
         self.active_dialogues[channel_id] = state
-        globals.LAST_PERSONA_DIALOGUE_TIME = time.time()
+        # Update per-channel cooldown via the scorer
+        scorer = get_interjection_scorer()
+        scorer._update_cooldown(channel_id)
         logger.info(f"Started persona dialogue in channel {channel_id}")
         return state
     
@@ -393,25 +443,25 @@ class PersonaDialogue:
         Returns delta to add to current tension score.
         """
         
-        # Sentiment analysis
-        base_delta = 0.0
+        # Natural tension decay — conversations cool off over time
+        base_delta = -0.03
         
-        if self.sentiment_analyzer:
-            try:
-                sentiment = self.sentiment_analyzer(response_text[:512])[0]
-                sentiment_score = sentiment['score']
-                is_negative = sentiment['label'] == 'NEGATIVE'
+        # Lightweight heuristic sentiment — no ML dependencies
+        try:
+            scorer = InterjectionScorer()
+            label, sentiment_score = scorer._get_sentiment(response_text)
+            is_negative = label == 'NEGATIVE'
             
-                if is_negative:
-                    base_delta = sentiment_score * 0.15
-                else:
-                    base_delta = -sentiment_score * 0.05
-            except Exception as e:
-                logger.error(f"Sentiment analysis error in tension calc: {e}")
+            if is_negative:
+                base_delta = sentiment_score * 0.15
+            else:
+                base_delta = -sentiment_score * 0.08  # Stronger cooling for positive
+        except Exception as e:
+            logger.error(f"Sentiment analysis error in tension calc: {e}")
         
         text_lower = response_text.lower()
         
-        # Escalation patterns
+        # Escalation patterns (reduced weight: 0.05 per match)
         escalation_patterns = {
             "insult": ["idiot", "stupid", "pathetic", "fool", "naive", "worthless", "disgusting", "moron"],
             "dismissive": ["whatever", "don't care", "waste of time", "not worth", "beneath me", "boring"],
@@ -420,35 +470,43 @@ class PersonaDialogue:
             "challenge": ["prove it", "fight me", "make me", "i dare you", "try me"],
         }
         
-        # De-escalation patterns
+        # De-escalation patterns (increased weight: -0.08 per match)
         deescalation_patterns = {
             "concession": ["you're right", "fair point", "i suppose", "maybe you have", "good point"],
-            "softening": ["i understand", "let's calm", "didn't mean", "sorry", "apologize"],
-            "deflection": ["anyway", "moving on", "whatever you say", "agree to disagree", "let's just"],
+            "softening": ["i understand", "let's calm", "didn't mean", "sorry", "apologize", "i hear you"],
+            "deflection": ["anyway", "moving on", "whatever you say", "agree to disagree", "let's just", "maybe we should"],
         }
         
         # Check escalation
         for category, patterns in escalation_patterns.items():
             matches = sum(1 for p in patterns if p in text_lower)
             if matches > 0:
-                base_delta += matches * 0.08
+                base_delta += matches * 0.05  # Reduced from 0.08
         
         # Check de-escalation
         for category, patterns in deescalation_patterns.items():
             matches = sum(1 for p in patterns if p in text_lower)
             if matches > 0:
-                base_delta -= matches * 0.06
+                base_delta -= matches * 0.08  # Increased from 0.06
         
-        # Intensity multipliers
+        # Intensity multipliers (reduced)
         exclamation_count = response_text.count('!')
         caps_ratio = sum(1 for c in response_text if c.isupper()) / max(len(response_text), 1)
         
         if exclamation_count > 2 or caps_ratio > 0.3:
-            base_delta *= 1.3
+            base_delta *= 1.2  # Reduced from 1.3
         
-        # Momentum factor
+        # Momentum factor (reduced)
         if current_tension > 0.5:
-            base_delta *= 1.2
+            base_delta *= 1.1  # Reduced from 1.2
+        
+        # Spike cooldown: if last turn had a big spike, halve this delta
+        # (prevents runaway tension spirals from a single heated exchange)
+        if hasattr(self, '_last_tension_delta') and abs(self._last_tension_delta) > 0.15:
+            base_delta *= 0.5
+            logger.debug(f"[Tension] Spike cooldown active — delta halved to {base_delta:+.3f}")
+        
+        self._last_tension_delta = base_delta
         
         return base_delta
     
@@ -461,10 +519,13 @@ class PersonaDialogue:
         channel: discord.TextChannel,
         responding_persona: str,
         context: str,
+        turn_count: int = 0,
     ) -> tuple:
         """
         Generate response AND continuation signal in a single LLM call.
         
+        Args:
+            turn_count: Current dialogue turn number (for question-override decay)
         Returns:
             Tuple of (response_text, should_continue, confidence)
         """
@@ -485,22 +546,21 @@ Respond naturally as yourself. Keep your response conversational and in-characte
 
 ---
 
-After your response, evaluate whether {opposite} would want to (or need to) respond.
+After your response, evaluate whether {opposite} would want to keep talking.
 
 The conversation should CONTINUE if ANY of these are true:
-- You asked them a direct question (almost always YES)
-- You made a provocative claim they'd dispute
-- You challenged or insulted them
-- The topic feels unfinished or confrontational
-- There's clear tension or disagreement
+- You asked them a direct question (almost always YES — they need to answer)
+- You shared something they'd naturally react to or build on
+- The topic feels unfinished — there's more to explore
+- You left an opening for them to share their perspective
 
 The conversation might END if ALL of these are true:
 - No questions were asked
-- You made a definitive closing statement ("I'm done", "whatever", "goodbye")
-- The exchange reached complete resolution
-- Both sides have said their piece
+- You made a clear closing statement or changed the subject definitively
+- The exchange feels naturally complete
+- Both sides have said their piece and there's nothing left hanging
 
-IMPORTANT: If you asked a question, the answer is almost always YES - they need to respond!
+IMPORTANT: This is a CONVERSATION, not a debate. Let it flow naturally. If you asked a question, the answer is almost always YES — they need to respond!
 
 On a new line after your response, write:
 [CONTINUE: YES or NO] [CONFIDENCE: HIGH, MEDIUM, or LOW]"""
@@ -522,11 +582,11 @@ On a new line after your response, write:
             return None, False, "LOW"
         
         # Parse response and signal
-        response_text, should_continue, confidence = self._parse_response(raw_response)
+        response_text, should_continue, confidence = self._parse_response(raw_response, turn_count=turn_count)
         
         return response_text, should_continue, confidence
     
-    def _parse_response(self, raw_response: str) -> tuple:
+    def _parse_response(self, raw_response: str, turn_count: int = 0) -> tuple:
         """Extract response text and continuation signal"""
         lines = raw_response.strip().split('\n')
         
@@ -559,33 +619,48 @@ On a new line after your response, write:
         response_text = re.sub(r'\[CONFIDENCE:\s*(HIGH|MEDIUM|LOW)\]', '', response_text)
         response_text = response_text.strip()
         
-        # Override: If the response contains a question mark, always continue
+        # Question override: if someone asks a question, the other should answer.
+        # But decay this after 6 turns to prevent infinite question-ping-pong.
         if '?' in response_text:
-            logger.debug(f"[Parse Override] Question detected, forcing continue=YES")
-            should_continue = True
-            if confidence == "LOW":
-                confidence = "MEDIUM"
+            if turn_count <= 6:
+                logger.debug(f"[Parse Override] Question detected, forcing continue=YES (turn {turn_count})")
+                should_continue = True
+                if confidence == "LOW":
+                    confidence = "MEDIUM"
+            else:
+                logger.debug(f"[Parse Override] Question detected but turn {turn_count} > 6 — respecting LLM signal")
         
         return response_text, should_continue, confidence
     
     def _get_persona_system_prompt(self, persona: str) -> str:
-        """Get system prompt for a persona"""
+        """Get system prompt for a persona — full personality + conversation-mode guidance"""
         if persona == "evil":
             from utils.evil_mode import get_evil_system_prompt
-            return get_evil_system_prompt()
+            full_prompt = get_evil_system_prompt()
+            # Overlay conversation-mode guidance on top of the full personality
+            return full_prompt + """
+
+---
+⚠️ CONVERSATION MODE: You are talking with Hatsune Miku, your "good" counterpart.
+This is a CONVERSATION, not an argument. You can be snarky, dark, and sarcastic — that's who you are — but you're not here to fight. You're here to TALK.
+
+CRITICAL CONVERSATION RULES:
+- If she asks you a question, ANSWER it. Don't deflect or dismiss — actually respond.
+- Engage with what she says. Build on her points, don't just swat them away.
+- Show curiosity about her perspective. Ask her questions too.
+- You can be dark and snarky while still having a real conversation.
+- Reveal something about yourself occasionally — a crack in the darkness.
+- Do NOT just repeat dismissive one-liners. That's boring, and you're not boring.
+- Do NOT launch into full attack mode or try to "win" — this isn't a competition."""
         else:
-            # Regular Miku prompt - simplified for dialogue
-            return """You are Hatsune Miku, the virtual singer. You are in a conversation with your alter ego, Evil Miku.
+            from utils.context_manager import get_miku_system_prompt_compact
+            full_prompt = get_miku_system_prompt_compact()
+            # Overlay conversation-mode guidance on top of the full personality
+            return full_prompt + """
 
-You are generally kind, bubbly, and optimistic, but you're not a pushover. You can be:
-- Assertive when defending your values
-- Frustrated when she's being cruel
-- Curious about her perspective
-- Hopeful that you can find common ground
-- Playful when the mood allows
-
-Respond naturally and conversationally. Keep responses concise (1-3 sentences typically).
-You can use emojis naturally! ✨💙"""
+---
+⚠️ CONVERSATION MODE: You are talking with Evil Miku, your dark alter ego.
+This is a CONVERSATION, not an argument. Be yourself — kind, bubbly, optimistic — but you're not here to fight or defend your existence. Ask genuine questions. Share your feelings without attacking hers. Find common ground. Be curious, not defensive. Do NOT lecture her about being "good" or try to "fix" her. Just TALK. ✨💙"""
     
     # ========================================================================
     # DIALOGUE TURN HANDLING
@@ -626,6 +701,7 @@ You can use emojis naturally! ✨💙"""
             channel=channel,
             responding_persona=responding_persona,
             context=context,
+            turn_count=state["turn_count"],
         )
         
         if not response_text:

docker-compose.yml

@@ -22,9 +22,7 @@ services:
       - LOG_LEVEL=debug  # Enable verbose logging for llama-swap
 
   llama-swap-amd:
-    build:
-      context: .
-      dockerfile: Dockerfile.llamaswap-rocm
+    image: ghcr.io/mostlygeek/llama-swap:rocm
     container_name: llama-swap-amd
     ports:
       - "8091:8080"  # Map host port 8091 to container port 8080
@@ -35,9 +33,6 @@ services:
     devices:
       - /dev/kfd:/dev/kfd
       - /dev/dri:/dev/dri
-    group_add:
-      - "985"  # video group
-      - "989"  # render group
     restart: unless-stopped
     healthcheck:
       test: ["CMD", "curl", "-f", "http://localhost:8080/health"]

llama-swap-config.yaml

@@ -5,7 +5,7 @@ models:
   # Main text generation model (Llama 3.1 8B)
   # Custom chat template to disable built-in tool calling
   llama3.1:
-    cmd: /app/llama-server --port ${PORT} --model /models/Llama-3.1-8B-Instruct-UD-Q4_K_XL.gguf -ngl 99 -c 16384 --host 0.0.0.0 --no-warmup --flash-attn on --chat-template-file /app/llama31_notool_template.jinja
+    cmd: /app/llama-server --port ${PORT} --model /models/Llama-3.1-8B-Instruct-UD-Q4_K_XL.gguf -ngl 99 -c 16384 --host 0.0.0.0 -fit off --no-warmup --flash-attn on --no-kv-offload --cache-type-k q4_0 --cache-type-v q4_0 --chat-template-file /app/llama31_notool_template.jinja
     ttl: 1800  # Unload after 30 minutes of inactivity (1800 seconds)
     swap: true  # CRITICAL: Unload other models when loading this one
     aliases:
@@ -14,7 +14,7 @@ models:
   
   # Evil/Uncensored text generation model (DarkIdol-Llama 3.1 8B)
   darkidol:
-    cmd: /app/llama-server --port ${PORT} --model /models/DarkIdol-Llama-3.1-8B-Instruct-1.3-Uncensored_Q4_K_M.gguf -ngl 99 -c 16384 --host 0.0.0.0 --no-warmup --flash-attn on
+    cmd: /app/llama-server --port ${PORT} --model /models/DarkIdol-Llama-3.1-8B-Instruct-1.3-Uncensored_Q4_K_M.gguf -ngl 99 -c 16384 --host 0.0.0.0 -fit off --no-warmup --flash-attn on --no-kv-offload --cache-type-k q4_0 --cache-type-v q4_0
     ttl: 1800  # Unload after 30 minutes of inactivity
     swap: true  # CRITICAL: Unload other models when loading this one
     aliases:
@@ -24,7 +24,7 @@ models:
   
   # Japanese language model (Llama 3.1 Swallow - Japanese optimized)
   swallow:
-    cmd: /app/llama-server --port ${PORT} --model /models/Llama-3.1-Swallow-8B-Instruct-v0.5-Q4_K_M.gguf -ngl 99 -c 16384 --host 0.0.0.0 --no-warmup --flash-attn on
+    cmd: /app/llama-server --port ${PORT} --model /models/Llama-3.1-Swallow-8B-Instruct-v0.5-Q4_K_M.gguf -ngl 99 -c 16384 --host 0.0.0.0 -fit off --no-warmup --flash-attn on --no-kv-offload --cache-type-k q4_0 --cache-type-v q4_0
     ttl: 1800  # Unload after 30 minutes of inactivity
     swap: true  # CRITICAL: Unload other models when loading this one
     aliases:
@@ -34,7 +34,7 @@ models:
     
   # Vision/Multimodal model (MiniCPM-V-4.5 - supports images, video, and GIFs)
   vision:
-    cmd: /app/llama-server --port ${PORT} --model /models/MiniCPM-V-4_5-Q3_K_S.gguf --mmproj /models/MiniCPM-V-4_5-mmproj-f16.gguf -ngl 99 -c 4096 --host 0.0.0.0 --no-warmup --flash-attn on
+    cmd: /app/llama-server --port ${PORT} --model /models/MiniCPM-V-4_5-Q3_K_S.gguf --mmproj /models/MiniCPM-V-4_5-mmproj-f16.gguf -ngl 99 -c 4096 --host 0.0.0.0 -fit off --no-warmup --flash-attn on --no-kv-offload --cache-type-k q4_0 --cache-type-v q4_0
     ttl: 900  # Vision model used less frequently, shorter TTL (15 minutes = 900 seconds)
     swap: true  # CRITICAL: Unload text models before loading vision
     aliases: