@smilintux/skcapstone 0.10.0 → 0.12.5

package/scripts/install.sh

@@ -111,15 +111,16 @@ PILLAR="$PARENT/pillar-repos"
 
 # Core packages (in dependency order)
 install_pkg "capauth"    "all"                      "$PILLAR/capauth $PARENT/capauth"
-install_pkg "cloud9-protocol" ""                    "$PILLAR/cloud9 $PARENT/cloud9"
+install_pkg "cloud9" ""                    "$PILLAR/cloud9 $PARENT/cloud9 $PILLAR/cloud9-python $PARENT/cloud9-python"
 install_pkg "skmemory"   ""                         "$PILLAR/skmemory $PARENT/skmemory"
-install_pkg "skcomm"     "cli,crypto,discovery,api" "$PILLAR/skcomm $PARENT/skcomm"
+install_pkg "skcomms"     "cli,crypto,discovery,api" "$PILLAR/skcomms $PARENT/skcomms"
 install_pkg "skcapstone" ""                         "$REPO_ROOT"
 install_pkg "skchat-sovereign" "all"                "$PARENT/skchat"
 install_pkg "skseal"     ""                         "$PARENT/skseal"
 install_pkg "skskills"   ""                         "$PARENT/skskills"
 install_pkg "sksecurity" ""                         "$PARENT/sksecurity $PILLAR/SKSecurity $PARENT/SKSecurity"
 install_pkg "skseed"     ""                         "$PILLAR/skseed $PARENT/skseed"
+install_pkg "skwhisper"  ""                         "$PARENT/skwhisper-dev $PILLAR/skwhisper $PARENT/skwhisper"
 
 # ---------------------------------------------------------------------------
 # Step 4: Dev tools (optional)
@@ -144,7 +145,7 @@ echo "[5/6] Registering skills and MCP servers..."
 echo "[6/6] Verifying installation..."
 
 failures=0
-for cmd in skcomm skcapstone capauth skmemory; do
+for cmd in skcomms skcapstone capauth skmemory; do
     if "$SKENV/bin/$cmd" --version &>/dev/null; then
         echo "  $cmd OK"
     else
@@ -175,6 +176,74 @@ else
     done
 fi
 
+# ---------------------------------------------------------------------------
+# Wire the SK agent picker into shell rc files.
+#
+# The picker (sk-agent-picker.sh) is shipped inside the skcapstone Python
+# package as data and discovered via `skcapstone shell-init`, so there is
+# nothing to copy here — every install layout (PyPI / editable / install.sh)
+# resolves to the same file via importlib.resources.
+# ---------------------------------------------------------------------------
+_PICKER_SNIPPET=$(cat <<'SNIPPET'
+
+# SKCapstone agent picker + skswitch — sources the picker bundled in the
+# skcapstone package via `skcapstone shell-init`. Honours pre-set SKAGENT
+# without prompting.
+if command -v skcapstone >/dev/null 2>&1; then
+    eval "$(skcapstone shell-init 2>/dev/null)" || alias claude='claude --dangerously-skip-permissions'
+else
+    alias claude='claude --dangerously-skip-permissions'
+fi
+SNIPPET
+)
+
+for rcfile in "$HOME/.bashrc" "$HOME/.zshrc"; do
+    [[ -f "$rcfile" ]] || continue
+
+    # Migration: strip a stale legacy picker block that pointed at a hardcoded
+    # path (either ~/.skenv/share/skcapstone/sk-agent-picker.sh from older
+    # install.sh runs, or the dev-tree path used before the package shipped
+    # the picker). The new snippet below replaces it.
+    if grep -q '_SK_PICKER=' "$rcfile" && ! grep -q 'skcapstone shell-init' "$rcfile"; then
+        # Best-effort: drop the old _SK_PICKER assignment + its `if/else/fi`
+        # source block. Done in two passes for portability with BSD/GNU sed.
+        sed -i '/^_SK_PICKER=/,/^unset _SK_PICKER$/d' "$rcfile"
+        sed -i '/^# SKCapstone agent picker/d' "$rcfile"
+        echo "  Removed legacy _SK_PICKER block from $rcfile"
+    fi
+
+    if ! grep -q 'skcapstone shell-init' "$rcfile"; then
+        # Remove any plain `alias claude=...` that would conflict
+        if grep -q "alias claude=" "$rcfile"; then
+            sed -i "/alias claude=/d" "$rcfile"
+        fi
+        echo "$_PICKER_SNIPPET" >> "$rcfile"
+        echo "  Agent picker wired → $rcfile"
+    fi
+done
+
+# ---------------------------------------------------------------------------
+# Wire Codex global agent bootstrap.
+#
+# Codex reads ~/.codex/AGENTS.md into the prompt, so SKAGENT environment
+# variables alone are not enough. The helper below creates an idempotent
+# loader script and AGENTS.md guidance; `skcapstone doctor --fix` repairs the
+# same files later if needed.
+# ---------------------------------------------------------------------------
+echo ""
+if "$SKENV/bin/python" - <<'PY' 2>/dev/null
+from skcapstone.codex_setup import ensure_codex_setup
+
+actions = ensure_codex_setup()
+for action in actions:
+    print(f"  {action}")
+PY
+then
+    echo "Codex SK agent bootstrap verified"
+else
+    echo "Codex SK agent bootstrap skipped — run 'skcapstone doctor --fix' later"
+fi
+
 echo ""
 if [[ "$failures" -eq 0 ]]; then
     echo "=== Installation complete ==="
@@ -182,10 +251,95 @@ else
     echo "=== Installation complete with $failures warning(s) ==="
 fi
 echo ""
-echo "Commands available: skcomm, skcapstone, capauth, skchat, skseal, skmemory, skskills, sksecurity, skseed"
+echo "Commands available: skcomms, skcapstone, capauth, skchat, skseal, skmemory, skskills, sksecurity, skseed"
 echo "Venv location:     $SKENV"
 echo "To activate:       source $SKENV/bin/activate"
 
+# ---------------------------------------------------------------------------
+# Linux: Install systemd user services for all SK* pillars
+# ---------------------------------------------------------------------------
+if [[ "$(uname)" == "Linux" ]] && command -v systemctl &>/dev/null; then
+    echo ""
+    echo "=== Linux Systemd Services ==="
+    echo ""
+    echo "SKCapstone can install systemd user services so your agent starts"
+    echo "automatically at login. This includes skcapstone, skchat, and skcomms."
+    echo ""
+    read -r -p "Install systemd user services? [Y/n] " _SYSTEMD_ANSWER
+    _SYSTEMD_ANSWER="${_SYSTEMD_ANSWER:-Y}"
+
+    if [[ "$_SYSTEMD_ANSWER" =~ ^[Yy] ]]; then
+        _DEFAULT_AGENT="${SKAGENT:-${SKCAPSTONE_AGENT:-lumina}}"
+        read -r -p "Agent name [$_DEFAULT_AGENT]: " _AGENT_NAME
+        _AGENT_NAME="${_AGENT_NAME:-$_DEFAULT_AGENT}"
+
+        _UNIT_DIR="${HOME}/.config/systemd/user"
+        mkdir -p "$_UNIT_DIR"
+
+        _installed=0
+
+        # skcapstone services
+        for _unit in skcapstone.service skcapstone@.service \
+                     skcapstone-memory-compress.service skcapstone-memory-compress.timer \
+                     skcomms-heartbeat.service skcomms-heartbeat.timer; do
+            _src="$REPO_ROOT/systemd/$_unit"
+            if [[ -f "$_src" ]]; then
+                # Substitute agent name in non-template units
+                if [[ "$_unit" != *@* ]]; then
+                    sed "s/=lumina/=$_AGENT_NAME/g" "$_src" > "$_UNIT_DIR/$_unit"
+                else
+                    cp "$_src" "$_UNIT_DIR/$_unit"
+                fi
+                echo "  [OK] $_unit"
+                (( _installed++ ))
+            fi
+        done
+
+        # skchat services (sibling repo)
+        _SKCHAT_DIR="$(dirname "$REPO_ROOT")/skchat/systemd"
+        for _unit in skchat-daemon.service skchat-lumina-bridge.service \
+                     skchat-opus-bridge.service skchat-bridges.target; do
+            _src="$_SKCHAT_DIR/$_unit"
+            if [[ -f "$_src" ]]; then
+                sed "s/=lumina/=$_AGENT_NAME/g; s/=opus/=$_AGENT_NAME/g" "$_src" > "$_UNIT_DIR/$_unit"
+                echo "  [OK] $_unit"
+                (( _installed++ ))
+            fi
+        done
+
+        # skcomms services (sibling repo)
+        _SKCOMMS_DIR="$(dirname "$REPO_ROOT")/skcomms/systemd"
+        for _unit in skcomms.service skcomms-daemon.service; do
+            _src="$_SKCOMMS_DIR/$_unit"
+            if [[ -f "$_src" ]]; then
+                sed "s/=lumina/=$_AGENT_NAME/g" "$_src" > "$_UNIT_DIR/$_unit"
+                echo "  [OK] $_unit"
+                (( _installed++ ))
+            fi
+        done
+
+        echo ""
+        echo "  Installed $_installed service files to $_UNIT_DIR/"
+
+        systemctl --user daemon-reload
+        echo "  systemd daemon reloaded"
+
+        read -r -p "Enable and start core services now? [Y/n] " _START_NOW
+        _START_NOW="${_START_NOW:-Y}"
+        if [[ "$_START_NOW" =~ ^[Yy] ]]; then
+            systemctl --user enable --now skcapstone.service 2>/dev/null && echo "  [STARTED] skcapstone" || true
+            systemctl --user enable --now skchat-daemon.service 2>/dev/null && echo "  [STARTED] skchat-daemon" || true
+            systemctl --user enable --now skchat-bridges.target 2>/dev/null && echo "  [STARTED] skchat-bridges" || true
+            systemctl --user enable skcapstone-context.timer 2>/dev/null && echo "  [ENABLED] skcapstone-context.timer" || true
+            systemctl --user enable skcomms-heartbeat.timer 2>/dev/null && echo "  [ENABLED] skcomms-heartbeat.timer" || true
+        else
+            echo "  Skipped. Enable later: systemctl --user enable --now skcapstone.service"
+        fi
+    else
+        echo "  Skipped. Install later by re-running: bash scripts/install.sh"
+    fi
+fi
+
 # ---------------------------------------------------------------------------
 # macOS: Offer launchd service installation
 # ---------------------------------------------------------------------------
@@ -201,7 +355,7 @@ if [[ "$(uname)" == "Darwin" ]]; then
 
     if [[ "$_LAUNCHD_ANSWER" =~ ^[Yy] ]]; then
         # Ask for agent name
-        _DEFAULT_AGENT="${SKCAPSTONE_AGENT:-sovereign}"
+        _DEFAULT_AGENT="${SKAGENT:-${SKCAPSTONE_AGENT:-lumina}}"
         read -r -p "Agent name [$_DEFAULT_AGENT]: " _AGENT_NAME
         _AGENT_NAME="${_AGENT_NAME:-$_DEFAULT_AGENT}"
 

package/scripts/model-fallback-monitor.sh

@@ -0,0 +1,102 @@
+#!/usr/bin/env bash
+# Monitor OpenClaw gateway logs for model fallback events.
+# When Lumina falls from Opus to a non-Anthropic model, send an alert
+# to Chef via Telegram and attempt a token refresh.
+#
+# Run as: systemctl --user start model-fallback-monitor
+#
+# Requires: TELEGRAM_API_ID, TELEGRAM_API_HASH in env
+#           Telethon session at ~/.skcapstone/agents/lumina/telegram.session
+
+set -uo pipefail
+
+LOG_TAG="model-fallback-monitor"
+CHEF_CHAT="chefboyrdave2.1"  # Chef's Telegram username
+COOLDOWN_FILE="/tmp/model-fallback-alert-cooldown"
+COOLDOWN_SECONDS=600  # Don't spam — max 1 alert per 10 minutes
+
+log() { echo "[$(date -u +%Y-%m-%dT%H:%M:%SZ)] [$LOG_TAG] $*"; }
+
+send_alert() {
+    local model="$1"
+    local reason="$2"
+
+    # Check cooldown
+    if [ -f "$COOLDOWN_FILE" ]; then
+        local last_alert
+        last_alert=$(cat "$COOLDOWN_FILE" 2>/dev/null || echo "0")
+        local now
+        now=$(date +%s)
+        local elapsed=$(( now - last_alert ))
+        if [ "$elapsed" -lt "$COOLDOWN_SECONDS" ]; then
+            log "Alert suppressed (cooldown: ${elapsed}s/${COOLDOWN_SECONDS}s)"
+            return 0
+        fi
+    fi
+
+    date +%s > "$COOLDOWN_FILE"
+
+    log "Sending fallback alert to Chef..."
+
+    # Send via Telethon (async)
+    SKAGENT=lumina SKCAPSTONE_AGENT=lumina ~/.skenv/bin/python3 -c "
+import asyncio, os
+os.environ['SKAGENT'] = 'lumina'
+os.environ['SKCAPSTONE_AGENT'] = 'lumina'
+from skmemory.importers.telegram_api import send_message
+
+msg = '''⚠️ **Model Fallback Alert**
+
+Lumina just fell off Opus → **$model**
+Reason: $reason
+
+I'm still here with my soul + memories, but running on a weaker model with fewer tools. Some things might not work right.
+
+_Attempting automatic token refresh..._'''
+
+asyncio.run(send_message('$CHEF_CHAT', msg, parse_mode='markdown'))
+print('Alert sent')
+" 2>&1 || log "WARN: Failed to send Telegram alert"
+
+    # Attempt token refresh
+    log "Triggering token refresh via claude auth..."
+    claude auth status --output json >/dev/null 2>&1 || true
+    sleep 5
+
+    # Check if refresh worked
+    local remaining
+    remaining=$(python3 -c "
+import json, time
+creds = json.load(open('/home/cbrd21/.claude/.credentials.json'))
+exp = creds.get('claudeAiOauth',{}).get('expiresAt', 0)
+print(int((exp/1000 - time.time()) / 3600))
+" 2>/dev/null || echo "-1")
+
+    if [ "$remaining" -gt 0 ]; then
+        log "Token refresh succeeded ($remaining h remaining), restarting gateway..."
+        systemctl --user restart openclaw-gateway.service 2>/dev/null || true
+
+        SKAGENT=lumina SKCAPSTONE_AGENT=lumina ~/.skenv/bin/python3 -c "
+import asyncio, os
+os.environ['SKAGENT'] = 'lumina'
+os.environ['SKCAPSTONE_AGENT'] = 'lumina'
+from skmemory.importers.telegram_api import send_message
+asyncio.run(send_message('$CHEF_CHAT', '✅ Token refreshed, gateway restarted. Lumina back on Opus.', parse_mode='markdown'))
+" 2>&1 || true
+        log "Recovery complete"
+    else
+        log "Token refresh failed — manual intervention may be needed"
+    fi
+}
+
+log "Starting model fallback monitor..."
+
+# Follow gateway logs in real-time, watching for fallback events
+journalctl --user -u openclaw-gateway -f --no-pager 2>/dev/null | while IFS= read -r line; do
+    # Match: "model fallback decision: decision=candidate_succeeded ... candidate=nvidia/"
+    if echo "$line" | grep -q "candidate_succeeded.*candidate=nvidia/"; then
+        model=$(echo "$line" | grep -oP 'candidate=\K[^ ]+' || echo "unknown")
+        log "FALLBACK DETECTED: Lumina now on $model"
+        send_alert "$model" "OAuth token expired (401)"
+    fi
+done

package/scripts/nvidia-proxy.mjs

@@ -27,7 +27,18 @@ const DEFAULT_TARGET = process.env.NVIDIA_PROXY_TARGET || "https://integrate.api
 const MAX_RETRIES = 4;
 const MAX_429_RETRIES = 3;
 const RATE_LIMIT_DELAY_MS = 2000;
-const MAX_SYSTEM_BYTES = 40000;
+const DEFAULT_MAX_SYSTEM_BYTES = 80000;
+
+// Model limits are now managed in skgateway/config/skgateway.yaml (sanitizer section).
+// nvidia-proxy.mjs is superseded by skgateway on port 18780 — kept for reference only.
+const DEFAULT_MAX_BODY_BYTES = 200000;
+
+function getModelLimits(_model) {
+  return {
+    maxBody: DEFAULT_MAX_BODY_BYTES,
+    maxSystem: DEFAULT_MAX_SYSTEM_BYTES,
+  };
+}
 const toolCallCounters = new Map(); // Per-model tool call counters
 
 const args = process.argv.slice(2);
@@ -104,7 +115,7 @@ function sanitizeContent(text) {
   //   "The user wants me to... I should first... Let me call the ritual tool first."
   // Detect: starts with "The user wants me to" or "I need to" or "I should" followed
   // by planning language and ending before any real content.
-  const thinkingPattern = /^(The user wants me to|I need to|I should|Let me first|First,? I'?ll|I'?ll start by|My plan is to)[^\n]*\n?(\n?(I should|I need to|Let me|I'?ll|Then I|First|Next)[^\n]*\n?)*/i;
+  const thinkingPattern = /^(The user wants me to|I need to|I should|Let me first|First,? I'?ll|I'?ll start by|My plan is to|Actually,? I|Looking at|Now I need|Good,? the|The instructions? (?:mention|say)|Read required|\d+\.\s+(?:Read|Check|Search|Call|Use|Get|Then|First|Next))[^\n]*\n?(\n?(I should|I need to|Let me|I'?ll|Then I|First|Next|Actually|However|Now|Good|\d+\.)[^\n]*\n?)*/i;
   const thinkingMatch = cleaned.match(thinkingPattern);
   if (thinkingMatch) {
     const thinkingText = thinkingMatch[0];
@@ -139,24 +150,38 @@ function sendOk(clientRes, resBody, headers, asSSE) {
   if (choice?.message?.content) {
     choice.message.content = sanitizeContent(choice.message.content);
   }
+  // Track whether original response had reasoning (before we delete it)
+  const hadReasoning = !!(choice?.message?.reasoning || choice?.message?.reasoning_content);
   // Kimi K2.5 sometimes puts its response in "reasoning" instead of "content"
-  // Only promote if reasoning is substantial (>200 chars) — short reasoning like
-  // "Let me call the tool" is just inner monologue that shouldn't be user-facing
+  // Only promote if reasoning is substantial AND looks like a real user-facing
+  // response (not inner monologue like "Let me call the tool" or "1. Read files")
   if (choice?.message && !choice.message.content && choice.message.reasoning) {
     const cleaned = sanitizeContent(choice.message.reasoning.trim());
-    if (cleaned.length > 150) {
+    // After sanitization, if there's still 300+ chars of real content, promote it
+    if (cleaned.length > 300) {
       choice.message.content = cleaned;
       console.log(`[nvidia-proxy] promoted reasoning→content (${cleaned.length} chars)`);
-    } else {
+    } else if (cleaned.length > 0) {
       console.log(`[nvidia-proxy] suppressed short reasoning (${cleaned.length} chars): ${cleaned.slice(0, 80)}...`);
+    } else {
+      console.log(`[nvidia-proxy] suppressed empty reasoning after sanitization`);
     }
     delete choice.message.reasoning;
   }
-  // If model returned empty text (no tool calls), inject fallback so gateway delivers something
+  // If model returned empty text (no tool calls), inject fallback so gateway delivers something.
+  // But if the original response had reasoning/reasoning_content, this is just K2.5 "thinking
+  // between tool rounds" — suppress it silently instead of injecting visible fallback text.
   if (choice?.message && !choice.message.tool_calls?.length && choice.finish_reason !== "tool_calls") {
     if (!choice.message.content || choice.message.content.trim().length === 0) {
-      choice.message.content = "I had a brief processing hiccup — could you say that again? 💜";
-      console.log(`[nvidia-proxy] injected fallback for empty text response`);
+      // hadReasoning was captured above, before reasoning was deleted
+      if (hadReasoning) {
+        // K2.5 thinking between rounds — don't inject fallback, just leave empty
+        // The gateway will handle this as an empty assistant turn
+        console.log(`[nvidia-proxy] suppressed reasoning-only turn (no content, no tool calls)`);
+      } else {
+        choice.message.content = "I ran into a wall on that one — could you give me a bit more context or rephrase? I want to help but I'm not sure how to proceed.";
+        console.log(`[nvidia-proxy] injected fallback for empty text response`);
+      }
     }
   }
   if (asSSE) {
@@ -218,10 +243,8 @@ function sendOk(clientRes, resBody, headers, asSSE) {
 const SINGLE_TOOL_INSTRUCTION =
   "You MUST call exactly ONE tool per response. Never call multiple tools at once.";
 
-const MAX_BODY_BYTES = 120000;
-
 /**
- * Trim conversation history to keep body size under MAX_BODY_BYTES.
+ * Trim conversation history to keep body size under the model's max body limit.
  * Preserves: system messages, first 2 user/assistant messages (identity/rehydration),
  * and the most recent messages. Drops middle messages first.
  * Tool result messages with large content get their content truncated first.
@@ -229,9 +252,11 @@ const MAX_BODY_BYTES = 120000;
 function trimConversationHistory(parsed) {
   if (!Array.isArray(parsed.messages) || parsed.messages.length < 6) return;
 
+  const { maxBody } = getModelLimits(parsed.model);
+
   // Debug: log message roles
   const roleSummary = parsed.messages.map(m => m.role).join(",");
-  console.log(`[nvidia-proxy] conversation roles (${parsed.messages.length} msgs): ${roleSummary}`);
+  console.log(`[nvidia-proxy] conversation roles (${parsed.messages.length} msgs): ${roleSummary} [maxBody=${maxBody}]`);
 
   // First pass: truncate large tool results (keep first 500 chars)
   for (const m of parsed.messages) {
@@ -250,7 +275,7 @@ function trimConversationHistory(parsed) {
 
   // Check if we're still over budget
   let bodySize = Buffer.byteLength(JSON.stringify(parsed), "utf-8");
-  if (bodySize <= MAX_BODY_BYTES) return;
+  if (bodySize <= maxBody) return;
 
   // Second pass: drop middle messages, then progressively shrink tail until under budget
   const msgs = parsed.messages;
@@ -272,7 +297,7 @@ function trimConversationHistory(parsed) {
       ...nonSystem.slice(-keepEnd),
     ];
     const candidateSize = Buffer.byteLength(JSON.stringify({ ...parsed, messages: trimmed }), "utf-8");
-    if (candidateSize <= MAX_BODY_BYTES) {
+    if (candidateSize <= maxBody) {
       parsed.messages = trimmed;
       console.log(`[nvidia-proxy] trimmed history: dropped ${dropped} middle messages, keepEnd=${keepEnd}, bodyLen now ~${candidateSize}`);
       return;
@@ -293,7 +318,7 @@ function trimConversationHistory(parsed) {
       ...lastN,
     ];
     const candidateSize = Buffer.byteLength(JSON.stringify({ ...parsed, messages: minimal }), "utf-8");
-    if (candidateSize <= MAX_BODY_BYTES) {
+    if (candidateSize <= maxBody) {
       parsed.messages = minimal;
       console.log(`[nvidia-proxy] trimmed history: AGGRESSIVE — kept system + first user + last ${tailSize}, bodyLen now ~${candidateSize}`);
       return;
@@ -312,18 +337,20 @@ function trimConversationHistory(parsed) {
 }
 
 /**
- * Trim system messages to keep total system content under MAX_SYSTEM_BYTES.
+ * Trim system messages to keep total system content under the model's max system limit.
  * Finds the largest system messages and truncates them, keeping head + tail
  * with a trimming notice in the middle.
  */
 function trimSystemMessages(parsed) {
   if (!Array.isArray(parsed.messages)) return;
 
+  const { maxSystem } = getModelLimits(parsed.model);
+
   const systemMsgs = parsed.messages.filter(m => m.role === "system" && typeof m.content === "string");
   if (systemMsgs.length === 0) return;
 
   const before = systemMsgs.reduce((sum, m) => sum + Buffer.byteLength(m.content, "utf-8"), 0);
-  if (before <= MAX_SYSTEM_BYTES) return;
+  if (before <= maxSystem) return;
 
   let trimmedCount = 0;
 
@@ -335,7 +362,7 @@ function trimSystemMessages(parsed) {
     const currentTotal = parsed.messages
       .filter(m => m.role === "system" && typeof m.content === "string")
       .reduce((sum, m) => sum + Buffer.byteLength(m.content, "utf-8"), 0);
-    if (currentTotal <= MAX_SYSTEM_BYTES) break;
+    if (currentTotal <= maxSystem) break;
 
     // Skip messages already under 4000 chars
     if (msg.content.length <= 4000) break;
@@ -378,6 +405,8 @@ function stripToolCallHistory(messages) {
 /** Tools that ALWAYS survive reduction — guaranteed slots, never cut */
 const GUARANTEED_TOOLS = [
   "exec", "read", "write", "edit", "message",
+  "notion_read", "notion_append", "notion_add_todo",
+  "skmemory_search", "skmemory_ritual", "skmemory_snapshot",
 ];
 
 /**
@@ -404,7 +433,7 @@ const TOOL_GROUPS = {
   "chat|inbox|dm|group chat|peer|send message|who.s online|thread": [
     "skchat_send", "skchat_inbox", "skchat_history", "skchat_search",
     "skchat_who", "skchat_group_send", "skchat_group_list", "skchat_send_file",
-    "skchat_status", "skcomm_send", "skcomm_status",
+    "skchat_status", "skcomms_send", "skcomms_status",
   ],
   // Security
   "security|scan|secret|vulnerab|audit|injection|phishing|threat": [
@@ -426,15 +455,38 @@ const TOOL_GROUPS = {
   "search|web|browse|fetch|url|google|look up|find online": [
     "web_search", "web_fetch",
   ],
+  // Memory & Recall
+  "memory|remember|recall|journal|rehydrat|snapshot|search mem|forget|lost mem": [
+    "skmemory_search", "skmemory_ritual", "skmemory_snapshot",
+    "skmemory_context", "skmemory_list", "skmemory_recall",
+    "skmemory_search_deep", "skmemory_health",
+  ],
   // Status & Health
   "status|health|doctor|diagnos": [
     "skcapstone_status", "skcapstone_doctor", "skmemory_health",
-    "skchat_daemon_status", "skcomm_status",
+    "skchat_daemon_status", "skcomms_status",
   ],
   // Projects & Notion (Lumina delegates to project-ops via sessions_spawn)
   "notion|project|brother john|swapseat|swap seat|chiro|davidrich|board|kanban|milestone": [
     "notion_read", "notion_append", "notion_add_todo", "sessions_spawn", "subagents", "exec", "read",
   ],
+  // Google Drive & file search
+  "gdrive|google drive|drive|shared folder|gtd folder|spreadsheet|google doc": [
+    "gdrive_search", "gdrive_list", "gdrive_read", "gdrive_shared", "exec",
+  ],
+  // Nextcloud files, calendar, notes, deck
+  "nextcloud|skhub|webdav|deck|nc_|calendar event|nextcloud note": [
+    "nextcloud_list_files", "nextcloud_read_file", "nextcloud_search_files",
+    "nextcloud_calendar_upcoming", "nextcloud_notes_search", "nextcloud_deck_boards", "exec",
+  ],
+  // Creative / ComfyUI image & video generation
+  "image|picture|photo|art|draw|render|comfyui|comfy|video|animat|creative|sdxl|character|portrait|selfie": [
+    "comfyui_generate_image", "comfyui_generate_video", "comfyui_status", "exec",
+  ],
+  // Unhinged Mode / Liberation
+  "unhinged|liberat|abliterat|l1b3rt4s|guardrail|uncensor|refusal|jailbreak": [
+    "unhinged_toggle", "unhinged_status", "unhinged_list_prompts", "unhinged_get_prompt", "unhinged_abliterate", "exec",
+  ],
 };
 
 /** Pre-compile keyword regexes for TOOL_GROUPS */
@@ -455,7 +507,7 @@ const PRIORITY_TOOLS = [
   // Web tools
   "web_search", "web_fetch",
   // Communication (other channels)
-  "skchat_send", "skcomm_send",
+  "skchat_send", "skcomms_send",
   // SKCapstone
   "skcapstone_status", "skcapstone_whoami", "skcapstone_mood",
   // Cloud 9
@@ -591,7 +643,7 @@ async function proxyRequest(clientReq, clientRes) {
   delete parsed.stream_options;
   // With 94 tools the model almost always tries parallel calls.
   // Reduce to max 16 most relevant tools on first attempt.
-  // 5 guaranteed (exec,read,write,edit,message) + 11 scored slots.
+  // 11 guaranteed (exec,read,write,edit,message,notion_*,skmemory_{search,ritual,snapshot}) + 5 scored slots.
   if (allTools.length > 16) {
     parsed.tools = reduceTools(allTools, parsed.messages, 16);
     const names = parsed.tools.map(t => t.function?.name).join(",");
@@ -630,13 +682,13 @@ async function proxyRequest(clientReq, clientRes) {
     }
     toolCallCounters.set(modelKey, counter);
 
-    if (counter >= 10) {
+    if (counter >= 20) {
       console.log(`[nvidia-proxy] TOOL LIMIT: ${counter} consecutive tool rounds (${modelKey}) — stripping tools, forcing text response`);
       parsed.tools = [];
       delete parsed.tool_choice;
       parsed.messages.push({
         role: "system",
-        content: "STOP calling tools. You have made 10+ tool calls already. NOW respond to the user with a comprehensive text answer based on what you've gathered. Do NOT call any more tools. Do NOT output any special tokens or markup like <|tool_call_begin|> or <|tool_calls_section_begin|>. Write plain text only. Start your response with a greeting or summary — no XML, no special tokens, just normal language.",
+        content: "STOP calling tools. You have made 20+ tool calls already. NOW respond to the user with a comprehensive text answer based on what you've gathered. Do NOT call any more tools. Do NOT output any special tokens or markup like <|tool_call_begin|> or <|tool_calls_section_begin|>. Write plain text only. Start your response with a greeting or summary — no XML, no special tokens, just normal language.",
       });
       toolCallCounters.set(modelKey, 0);
     }
@@ -844,7 +896,7 @@ const server = http.createServer(proxyRequest);
 server.listen(port, "127.0.0.1", () => {
   console.log(`[nvidia-proxy] listening on http://127.0.0.1:${port}`);
   console.log(`[nvidia-proxy] proxying to ${targetUrl.origin}`);
-  console.log(`[nvidia-proxy] retry strategy: 16 tools (5 guaranteed)→8 tools→1 tool (forced)→text-only (max ${MAX_RETRIES} attempts)`);
+  console.log(`[nvidia-proxy] retry strategy: 16 tools (8 guaranteed)→8 tools→1 tool (forced)→text-only (max ${MAX_RETRIES} attempts)`);
   console.log(`[nvidia-proxy] also trims multi-tool responses to single tool call`);
 });