ruflo 3.42.3 → 3.42.4

package/src/ruvocal/mcp-bridge/index.js

@@ -1,1995 +1,1995 @@
-import express from "express";
-import { spawn } from "child_process";
-import { randomUUID, timingSafeEqual } from "crypto";
-
-// =============================================================================
-// CONFIGURATION
-// =============================================================================
-
-const CLOUD_FUNCTIONS = {
-  search: process.env.SEARCH_API_URL || null,
-  research: process.env.RESEARCH_API_URL || null,
-};
-
-const PORT = parseInt(process.env.PORT || "3001", 10);
-const BIND_HOST = process.env.MCP_BIND_HOST || "127.0.0.1";
-
-// =============================================================================
-// TOOL GROUPS — Enable/disable categories of tools independently
-// =============================================================================
-// Groups map tool name prefixes from backends to logical categories.
-// Each group can be toggled via env var. The AI sees only enabled tools.
-
-const TOOL_GROUPS = {
-  // --- Core (always on, built-in) ---
-  core: {
-    enabled: true, // cannot be disabled
-    description: "Search, research, and guidance tools",
-    source: "builtin",
-  },
-
-  // --- Intelligence (ruvector) ---
-  intelligence: {
-    enabled: process.env.MCP_GROUP_INTELLIGENCE !== "false",
-    description: "Self-learning intelligence — routing, memory, pattern training (ruvector)",
-    source: "ruvector",
-    prefixes: ["hooks_"],
-  },
-
-  // --- Agents & Orchestration (ruflo) ---
-  agents: {
-    enabled: process.env.MCP_GROUP_AGENTS !== "false",
-    description: "Agent lifecycle, swarm coordination, task management, workflows (ruflo)",
-    source: "ruflo",
-    prefixes: ["agent_", "swarm_", "task_", "session_", "hive-mind_", "workflow_", "coordination_"],
-  },
-
-  // --- Memory & Knowledge (ruflo) ---
-  memory: {
-    enabled: process.env.MCP_GROUP_MEMORY !== "false",
-    description: "Vector memory, AgentDB, embeddings, semantic search (ruflo)",
-    source: "ruflo",
-    prefixes: ["memory_", "agentdb_", "embeddings_"],
-  },
-
-  // --- Dev Tools (ruflo) ---
-  devtools: {
-    enabled: process.env.MCP_GROUP_DEVTOOLS !== "false",
-    description: "Hooks, code analysis, performance profiling, GitHub integration (ruflo)",
-    source: "ruflo",
-    prefixes: ["hooks_", "analyze_", "performance_", "github_", "terminal_", "config_", "system_", "progress_"],
-  },
-
-  // --- Security & Safety (ruflo) ---
-  security: {
-    enabled: process.env.MCP_GROUP_SECURITY === "true",
-    description: "AI defence, PII detection, claims management, pattern transfer (ruflo)",
-    source: "ruflo",
-    prefixes: ["aidefence_", "claims_", "transfer_"],
-  },
-
-  // --- Browser Automation (ruflo) ---
-  browser: {
-    enabled: process.env.MCP_GROUP_BROWSER === "true",
-    description: "Headless browser control — navigate, click, fill, screenshot (ruflo)",
-    source: "ruflo",
-    prefixes: ["browser_"],
-  },
-
-  // --- Neural & DAA (ruflo) ---
-  neural: {
-    enabled: process.env.MCP_GROUP_NEURAL === "true",
-    description: "Neural network training, DAA autonomous agents, cognitive patterns (ruflo)",
-    source: "ruflo",
-    prefixes: ["neural_", "daa_"],
-  },
-
-  // --- Agentic Flow (agentic-flow@alpha) ---
-  "agentic-flow": {
-    enabled: process.env.MCP_GROUP_AGENTIC_FLOW === "true",
-    description: "Execute 66+ specialized agents, batch code editing, AgentDB patterns (agentic-flow)",
-    source: "agentic-flow",
-    prefixes: ["agentic_flow_", "agent_booster_", "agentdb_"],
-  },
-
-  // --- Claude Code ---
-  "claude-code": {
-    enabled: process.env.MCP_GROUP_CLAUDE_CODE === "true",
-    description: "Anthropic Claude Code — file editing, bash execution, code analysis (requires ANTHROPIC_API_KEY)",
-    source: "claude",
-  },
-
-  // --- Gemini MCP ---
-  gemini: {
-    enabled: process.env.MCP_GROUP_GEMINI === "true",
-    description: "Google Gemini conversation context, multimodal capabilities (requires GOOGLE_API_KEY)",
-    source: "gemini-mcp",
-  },
-
-  // --- OpenAI Codex ---
-  codex: {
-    enabled: process.env.MCP_GROUP_CODEX === "true",
-    description: "OpenAI Codex coding agent — code generation and execution (requires OPENAI_API_KEY)",
-    source: "codex",
-  },
-};
-
-// =============================================================================
-// STDIO MCP CLIENT — Connects to external MCP servers via child process
-// =============================================================================
-
-class StdioMcpClient {
-  constructor(name, command, args = []) {
-    this.name = name;
-    this.command = command;
-    this.args = args;
-    this.process = null;
-    this.tools = [];
-    this.ready = false;
-    this.pending = new Map();
-    this.buffer = "";
-  }
-
-  async start() {
-    return new Promise((resolve) => {
-      try {
-        this.process = spawn(this.command, this.args, {
-          stdio: ["pipe", "pipe", "pipe"],
-          env: { ...process.env },
-        });
-
-        this.process.stdout.on("data", (data) => this._onData(data.toString()));
-        this.process.stderr.on("data", (data) => {
-          const msg = data.toString().trim();
-          if (msg && !msg.startsWith("npm WARN")) console.error(`[${this.name}] ${msg}`);
-        });
-        this.process.on("error", (err) => {
-          console.error(`[${this.name}] spawn error:`, err.message);
-          this.ready = false;
-          resolve(false);
-        });
-        this.process.on("exit", (code) => {
-          console.log(`[${this.name}] exited with code ${code}`);
-          this.ready = false;
-        });
-
-        this._send("initialize", {
-          protocolVersion: "2024-11-05",
-          capabilities: {},
-          clientInfo: { name: "mcp-bridge", version: "2.0.0" },
-        }).then((result) => {
-          if (result && !result.error) {
-            this._notify("notifications/initialized", {});
-            return this._send("tools/list", {});
-          }
-          return null;
-        }).then((result) => {
-          if (result && result.tools) {
-            this.tools = result.tools.map(t => ({
-              ...t,
-              _originalName: t.name,
-              _backend: this.name,
-            }));
-            this.ready = true;
-            console.log(`[${this.name}] ${this.tools.length} tools loaded`);
-          }
-          resolve(this.ready);
-        }).catch((err) => {
-          console.error(`[${this.name}] init failed:`, err.message);
-          resolve(false);
-        });
-
-        setTimeout(() => { if (!this.ready) resolve(false); }, 60000);
-      } catch (err) {
-        console.error(`[${this.name}] failed to start:`, err.message);
-        resolve(false);
-      }
-    });
-  }
-
-  _onData(chunk) {
-    this.buffer += chunk;
-    const lines = this.buffer.split("\n");
-    this.buffer = lines.pop() || "";
-    for (const line of lines) {
-      const trimmed = line.trim();
-      if (!trimmed) continue;
-      try {
-        const msg = JSON.parse(trimmed);
-        if (msg.id && this.pending.has(msg.id)) {
-          const { resolve } = this.pending.get(msg.id);
-          this.pending.delete(msg.id);
-          resolve(msg.result || msg.error || {});
-        }
-      } catch { /* skip non-JSON */ }
-    }
-  }
-
-  _send(method, params) {
-    return new Promise((resolve, reject) => {
-      if (!this.process || this.process.killed) {
-        return reject(new Error(`${this.name} process not running`));
-      }
-      const id = randomUUID();
-      const msg = JSON.stringify({ jsonrpc: "2.0", id, method, params }) + "\n";
-      this.pending.set(id, { resolve, reject });
-      this.process.stdin.write(msg);
-      // initialize is the cold-start gate for backends like ruflo/ruvector
-      // which boot a full claude-flow / ruvector kernel — on Cloud Run with
-      // npx fetching artifacts it can take 45-60s. Other RPC methods are
-      // post-init and stay snappy.
-      const timeoutMs = method === "initialize" ? 120000 : 30000;
-      setTimeout(() => {
-        if (this.pending.has(id)) {
-          this.pending.delete(id);
-          reject(new Error(`${this.name} timeout for ${method}`));
-        }
-      }, timeoutMs);
-    });
-  }
-
-  _notify(method, params) {
-    if (!this.process || this.process.killed) return;
-    this.process.stdin.write(JSON.stringify({ jsonrpc: "2.0", method, params }) + "\n");
-  }
-
-  async callTool(originalName, args) {
-    if (!this.ready) return { error: `${this.name} backend not available` };
-    try {
-      return await this._send("tools/call", { name: originalName, arguments: args });
-    } catch (err) {
-      return { error: err.message };
-    }
-  }
-
-  stop() {
-    if (this.process && !this.process.killed) {
-      this.process.kill("SIGTERM");
-      this.process = null;
-    }
-    this.ready = false;
-    this.tools = [];
-  }
-}
-
-// =============================================================================
-// BACKEND REGISTRY
-// =============================================================================
-
-const BACKEND_DEFS = [
-  { name: "ruvector",       command: "npx", args: ["-y", "ruvector", "mcp", "start"],   groups: ["intelligence"] },
-  { name: "ruflo",          command: "npx", args: ["-y", "ruflo", "mcp", "start"],      groups: ["agents", "memory", "devtools", "security", "browser", "neural"] },
-  { name: "agentic-flow",   command: "npx", args: ["-y", "agentic-flow@alpha", "mcp", "start"], groups: ["agentic-flow"] },
-  { name: "claude",         command: "claude", args: ["mcp", "serve"],                  groups: ["claude-code"] },
-  { name: "gemini-mcp",     command: "npx", args: ["-y", "gemini-mcp-server"],          groups: ["gemini"] },
-  { name: "codex",          command: "npx", args: ["-y", "@openai/codex", "mcp-server"], groups: ["codex"] },
-];
-
-const mcpBackends = new Map();
-let allBackendTools = []; // all tools from all backends (pre-filter)
-
-function isBackendNeeded(backendDef) {
-  return backendDef.groups.some(g => TOOL_GROUPS[g]?.enabled);
-}
-
-// Filter tools from a backend based on which groups are enabled
-function filterToolsByGroups(tools, backendName) {
-  const enabledGroups = Object.entries(TOOL_GROUPS)
-    .filter(([, g]) => g.enabled && g.source === backendName);
-
-  if (enabledGroups.length === 0) return [];
-
-  // If any enabled group has no prefixes defined, include all tools from that backend
-  const hasWildcard = enabledGroups.some(([, g]) => !g.prefixes);
-  if (hasWildcard) return tools;
-
-  const enabledPrefixes = enabledGroups.flatMap(([, g]) => g.prefixes || []);
-  return tools.filter(t => enabledPrefixes.some(p => t._originalName.startsWith(p)));
-}
-
-// Get the final filtered tool list with namespaced names
-function getActiveTools() {
-  const filtered = [];
-  for (const [backendName, client] of mcpBackends) {
-    const accepted = filterToolsByGroups(client.tools, backendName);
-    for (const t of accepted) {
-      filtered.push({ ...t, name: `${backendName}__${t._originalName}` });
-    }
-  }
-  return filtered;
-}
-
-async function initBackends() {
-  const needed = BACKEND_DEFS.filter(isBackendNeeded);
-  if (needed.length === 0) return;
-
-  console.log(`Starting ${needed.length} MCP backends: ${needed.map(b => b.name).join(", ")}`);
-
-  await Promise.allSettled(
-    needed.map(async (b) => {
-      const client = new StdioMcpClient(b.name, b.command, b.args);
-      const ok = await client.start();
-      if (ok) {
-        mcpBackends.set(b.name, client);
-      } else {
-        console.warn(`[${b.name}] failed to start`);
-      }
-    })
-  );
-
-  allBackendTools = getActiveTools();
-  console.log(`MCP backends: ${mcpBackends.size} active, ${allBackendTools.length} tools (filtered by groups)`);
-}
-
-process.on("SIGTERM", () => { for (const [, c] of mcpBackends) c.stop(); process.exit(0); });
-process.on("SIGINT", () => { for (const [, c] of mcpBackends) c.stop(); process.exit(0); });
-
-// =============================================================================
-// BUILT-IN TOOLS (core group — always on)
-// =============================================================================
-
-const BUILTIN_TOOLS = [
-  {
-    name: "search",
-    description: "Search your knowledge base for relevant information.",
-    inputSchema: {
-      type: "object",
-      properties: {
-        query: { type: "string", description: "Natural language search query" },
-        limit: { type: "number", description: "Max results (default 5)", default: 5 },
-      },
-      required: ["query"],
-    },
-  },
-  {
-    name: "web_research",
-    description: "Search the web, fact-check claims, compare items, or conduct deep research. Actions: 'search' (quick), 'research' (deep report), 'compare' (side-by-side), 'fact_check' (verify claims), 'goap' (comprehensive multi-step research with verification).",
-    inputSchema: {
-      type: "object",
-      properties: {
-        action: { type: "string", enum: ["search", "research", "compare", "fact_check", "goap"], description: "Research action type", default: "search" },
-        query: { type: "string", description: "Search query or topic" },
-        items: { type: "array", items: { type: "string" }, description: "Items to compare (for 'compare')" },
-        claim: { type: "string", description: "Claim to verify (for 'fact_check')" },
-        verify: { type: "boolean", description: "Verify results in goap mode", default: true },
-      },
-      required: ["query"],
-    },
-  },
-  {
-    name: "guidance",
-    description: "Get instructions on how to use the available tool groups and services. Call this FIRST when unsure which tool to use, when a user asks 'what can you do?', or when you need to understand a specific tool group. Returns structured guidance for the AI on tool selection and usage patterns.",
-    inputSchema: {
-      type: "object",
-      properties: {
-        topic: {
-          type: "string",
-          enum: ["overview", "groups", "intelligence", "agents", "memory", "devtools", "security", "browser", "neural", "agentic-flow", "claude-code", "gemini", "codex", "tool"],
-          description: "What to get guidance on. Use 'overview' for capabilities summary, 'groups' to see all tool groups and their status, or a specific group name for detailed usage instructions.",
-          default: "overview",
-        },
-        tool_name: { type: "string", description: "Specific tool name to get detailed usage info (when topic='tool')" },
-      },
-    },
-  },
-];
-
-// =============================================================================
-// GUIDANCE ENGINE — AI-facing instruction system
-// =============================================================================
-
-function getGuidance(topic, toolName) {
-  const activeGroups = Object.entries(TOOL_GROUPS).filter(([, g]) => g.enabled);
-  const inactiveGroups = Object.entries(TOOL_GROUPS).filter(([, g]) => !g.enabled);
-  const externalTools = getActiveTools();
-
-  if (topic === "overview") {
-    return {
-      guidance: `# Tool Capabilities Overview
-
-You have access to ${BUILTIN_TOOLS.length + externalTools.length} tools organized into ${activeGroups.length} active groups.
-
-## Active Groups
-${activeGroups.map(([name, g]) => {
-  const count = name === "core" ? BUILTIN_TOOLS.length : externalTools.filter(t => t.name.startsWith(g.source + "__")).length;
-  return `- **${name}** (${count} tools) — ${g.description}`;
-}).join("\n")}
-
-## Inactive Groups (can be enabled)
-${inactiveGroups.map(([name, g]) => `- **${name}** — ${g.description}`).join("\n") || "None"}
-
-## Quick Decision Guide
-- **Knowledge questions** → use \`search\` first, then \`web_research\` if needed
-- **Current events / facts** → use \`web_research\` with action 'search' or 'goap'
-- **Complex research** → use \`web_research\` with action 'goap' (multi-step pipeline)
-- **"What can you do?"** → call \`guidance\` with topic 'groups'
-- **Memory / recall** → use tools from the \`memory\` group
-- **Agent orchestration** → use tools from the \`agents\` group
-- **Code analysis / performance** → use tools from the \`devtools\` group
-
-## Rules
-1. Call tools FIRST, then present results conversationally
-2. Never show raw JSON — synthesize results naturally
-3. For complex questions, prefer GOAP pipeline (web_research action='goap')
-4. Call \`guidance\` with a specific group name to learn how to use that group's tools`,
-      topic: "overview",
-    };
-  }
-
-  if (topic === "groups") {
-    const groupList = Object.entries(TOOL_GROUPS).map(([name, g]) => {
-      const status = g.enabled ? "ACTIVE" : "INACTIVE";
-      const toolCount = name === "core" ? BUILTIN_TOOLS.length :
-        externalTools.filter(t => {
-          const backend = t._backend;
-          return g.source === backend && (!g.prefixes || g.prefixes.some(p => t._originalName.startsWith(p)));
-        }).length;
-      return `| ${name} | ${status} | ${toolCount} | ${g.description} |`;
-    });
-
-    return {
-      guidance: `# Tool Groups\n\n| Group | Status | Tools | Description |\n|-------|--------|-------|-------------|\n${groupList.join("\n")}`,
-      topic: "groups",
-    };
-  }
-
-  // Specific group guidance
-  const groupGuides = {
-    intelligence: `# Intelligence Group (ruvector)
-
-Self-learning intelligence tools for routing and vector memory.
-
-## Key Tools
-- **ruvector__hooks_route** — Route a task to the best agent type. Call with a task description.
-- **ruvector__hooks_remember** — Store context/knowledge in vector memory for later recall.
-- **ruvector__hooks_recall** — Search vector memory semantically. Good for finding past context.
-- **ruvector__hooks_pretrain** — Bootstrap intelligence from a code repository.
-- **ruvector__hooks_build_agents** — Generate optimized agent configurations.
-- **ruvector__hooks_stats** — Get intelligence statistics and learning metrics.
-
-## When to Use
-- Before starting complex tasks: route to find the best agent approach
-- To store important findings for cross-session memory
-- To recall previously stored patterns or solutions`,
-
-    agents: `# Agents & Orchestration Group (ruflo)
-
-Multi-agent lifecycle management, swarm coordination, and task workflows.
-
-## Key Tools
-- **ruflo__agent_spawn** — Create a new agent with specific capabilities
-- **ruflo__agent_list** — List all active agents
-- **ruflo__swarm_init** — Initialize a swarm with a topology (mesh, hierarchical, ring, star)
-- **ruflo__task_create** — Create and assign tasks
-- **ruflo__workflow_create** — Define multi-step workflows
-- **ruflo__workflow_execute** — Execute a workflow
-- **ruflo__hive-mind_init** — Start collective intelligence coordination
-- **ruflo__coordination_orchestrate** — Multi-agent coordination
-
-## When to Use
-- Complex tasks requiring multiple agents working together
-- Pipeline workflows with sequential or parallel steps
-- Distributed task management`,
-
-    memory: `# Memory & Knowledge Group (ruflo)
-
-Vector storage, semantic search, AgentDB pattern learning, and embeddings.
-
-## Key Tools
-- **ruflo__memory_store** — Store a value with vector embedding for semantic search
-- **ruflo__memory_search** — Semantic search across stored memories (HNSW-indexed)
-- **ruflo__memory_list** — List stored memory entries
-- **ruflo__agentdb_pattern-store** — Store a reasoning pattern for learning
-- **ruflo__agentdb_pattern-search** — Search for similar reasoning patterns
-- **ruflo__agentdb_context-synthesize** — Synthesize context from stored memories
-- **ruflo__embeddings_generate** — Generate vector embeddings for text
-- **ruflo__embeddings_search** — Semantic similarity search
-
-## When to Use
-- Persistent knowledge storage across sessions
-- Finding similar past solutions or patterns
-- Building semantic search over custom data`,
-
-    devtools: `# Dev Tools Group (ruflo)
-
-Code analysis, performance profiling, GitHub integration, and terminal access.
-
-## Key Tools
-- **ruflo__analyze_diff** — Analyze git diff for risk and change classification
-- **ruflo__performance_benchmark** — Run performance benchmarks
-- **ruflo__performance_bottleneck** — Detect performance bottlenecks
-- **ruflo__github_repo_analyze** — Analyze a GitHub repository
-- **ruflo__github_pr_manage** — Manage pull requests
-- **ruflo__terminal_execute** — Execute commands in a terminal session
-
-## When to Use
-- Code review and change risk assessment
-- Performance analysis and optimization
-- GitHub repository management`,
-
-    security: `# Security & Safety Group (ruflo)
-
-AI defence, PII detection, and claims-based authorization.
-
-## Key Tools
-- **ruflo__aidefence_scan** — Scan text for AI manipulation attempts
-- **ruflo__aidefence_has_pii** — Check for PII (emails, phones, SSNs)
-- **ruflo__aidefence_is_safe** — Quick safety check on input
-- **ruflo__claims_claim** — Claim an issue for work
-- **ruflo__claims_board** — Visual board of all claims
-
-## When to Use
-- Input validation and safety checking
-- PII detection before processing sensitive data
-- Work item management across agents`,
-
-    browser: `# Browser Automation Group (ruflo)
-
-Headless browser control for web interaction and testing.
-
-## Key Tools
-- **ruflo__browser_open** — Navigate to a URL
-- **ruflo__browser_click** — Click elements by reference
-- **ruflo__browser_fill** — Fill form inputs
-- **ruflo__browser_screenshot** — Capture page screenshots
-- **ruflo__browser_snapshot** — Get accessibility tree for AI parsing
-- **ruflo__browser_eval** — Execute JavaScript in page context
-
-## When to Use
-- Web scraping and data extraction
-- Automated testing (E2E)
-- Form filling and web interaction`,
-
-    neural: `# Neural & DAA Group (ruflo)
-
-Neural network operations and Decentralized Autonomous Agents.
-
-## Key Tools
-- **ruflo__neural_train** — Train a neural model
-- **ruflo__neural_predict** — Make predictions
-- **ruflo__daa_agent_create** — Create an autonomous agent
-- **ruflo__daa_workflow_create** — Create autonomous workflows
-- **ruflo__daa_knowledge_share** — Share knowledge between agents
-
-## When to Use
-- Pattern learning and prediction
-- Autonomous agent workflows
-- Knowledge transfer between agents`,
-
-    "agentic-flow": `# Agentic Flow Group (agentic-flow@alpha)
-
-Execute 66+ specialized agents with boosted code editing and AgentDB.
-
-## Key Tools
-- **agentic-flow__agentic_flow_agent** — Execute any of 66+ specialized agents
-- **agentic-flow__agentic_flow_list_agents** — List available agent types
-- **agentic-flow__agent_booster_edit_file** — 352x faster code editing
-- **agentic-flow__agent_booster_batch_edit** — Multi-file refactoring
-- **agentic-flow__agentdb_pattern_store** — Store reasoning patterns
-- **agentic-flow__agentdb_pattern_search** — Search similar patterns
-
-## When to Use
-- Complex code generation with specialized agents
-- Batch code refactoring across files
-- Agent selection when you need the right specialist`,
-
-    "claude-code": `# Claude Code Group
-
-Anthropic Claude Code MCP server — full coding agent capabilities.
-
-Requires: ANTHROPIC_API_KEY environment variable.
-
-## Capabilities
-- File reading and editing
-- Bash command execution
-- Code analysis and generation
-- Project exploration
-
-## When to Use
-- When you need a second AI perspective on code
-- Complex refactoring tasks
-- Code review and analysis`,
-
-    gemini: `# Gemini MCP Group
-
-Google Gemini with conversation context management.
-
-Requires: GOOGLE_API_KEY environment variable (already set for Gemini models).
-
-## Capabilities
-- Conversation context management
-- Multimodal processing
-- Google Search grounding
-
-## When to Use
-- Extended context conversations
-- Multimodal content processing`,
-
-    codex: `# Codex Group
-
-OpenAI Codex coding agent.
-
-Requires: OPENAI_API_KEY environment variable (already set for OpenAI models).
-
-## Capabilities
-- Code generation and execution
-- Code completion
-- Code explanation
-
-## When to Use
-- Code generation tasks
-- Quick code completions
-- Code explanation and documentation`,
-  };
-
-  if (topic === "tool" && toolName) {
-    const allTools = [...BUILTIN_TOOLS, ...externalTools];
-    const tool = allTools.find(t => t.name === toolName);
-    if (tool) {
-      const props = Object.entries(tool.inputSchema?.properties || {})
-        .map(([k, v]) => `- **${k}** (${v.type}) — ${v.description || ""}`)
-        .join("\n");
-      return { guidance: `# ${tool.name}\n\n${tool.description}\n\n## Parameters\n${props}`, topic: "tool" };
-    }
-    return { guidance: `Tool '${toolName}' not found. Call guidance with topic='groups' to see available tools.`, topic: "tool" };
-  }
-
-  if (groupGuides[topic]) {
-    const group = TOOL_GROUPS[topic];
-    if (!group?.enabled) {
-      return { guidance: `# ${topic} — INACTIVE\n\n${group?.description || ""}\n\nThis group is not enabled. Set the appropriate MCP_GROUP_* env var to "true" to activate it.`, topic };
-    }
-    return { guidance: groupGuides[topic], topic };
-  }
-
-  return { guidance: `Unknown topic '${topic}'. Use 'overview', 'groups', or a specific group name.`, topic };
-}
-
-// =============================================================================
-// GEMINI GROUNDED SEARCH — Uses Gemini's built-in Google Search when no
-// dedicated search Cloud Function is configured
-// =============================================================================
-
-async function geminiGroundedSearch(query, mode = "search") {
-  const apiKey = process.env.GOOGLE_API_KEY;
-  if (!apiKey) return { error: "No GOOGLE_API_KEY configured for search" };
-
-  // Empty/missing query produces a 400 from Gemini's generateContent endpoint.
-  // Return a structured error so the model can recover with a real query.
-  if (!query || typeof query !== "string" || !query.trim()) {
-    return {
-      error: "search requires a non-empty query string",
-      hint: "Call this tool again with { query: 'your search terms' }. For comparisons use { action: 'compare', query: 'item A vs item B' }; for fact-checking use { action: 'fact_check', claim: 'the claim text' }.",
-    };
-  }
-
-  const model = "gemini-2.5-flash";
-  const url = `https://generativelanguage.googleapis.com/v1beta/models/${model}:generateContent?key=${apiKey}`;
-
-  const systemInstructions = {
-    search: `You are a helpful search assistant. Answer the query using Google Search results. Include key facts, sources, and relevant details. Be concise but thorough.`,
-    research: `You are a research analyst. Provide a comprehensive research report on the topic using Google Search results. Include: key findings, analysis, multiple perspectives, and source citations. Be thorough.`,
-    compare: `You are a comparison analyst. Compare the items using Google Search results. Create a structured comparison with pros/cons, key differences, and a recommendation.`,
-    fact_check: `You are a fact-checker. Verify the claim using Google Search results. Provide a verdict (TRUE/FALSE/PARTIALLY TRUE/UNVERIFIABLE), evidence, and sources.`,
-  };
-
-  const prompt = mode === "fact_check"
-    ? `Fact-check this claim: "${query}"`
-    : mode === "compare"
-      ? `Compare these items: ${query}`
-      : mode === "research"
-        ? `Research this topic thoroughly: ${query}`
-        : query;
-
-  try {
-    const resp = await fetch(url, {
-      method: "POST",
-      headers: { "Content-Type": "application/json" },
-      body: JSON.stringify({
-        system_instruction: { parts: [{ text: systemInstructions[mode] || systemInstructions.search }] },
-        contents: [{ parts: [{ text: prompt }] }],
-        tools: [{ google_search: {} }],
-        generationConfig: { temperature: 0.2 },
-      }),
-      signal: AbortSignal.timeout(30000),
-    });
-
-    if (!resp.ok) {
-      const errText = await resp.text();
-      console.error(`[gemini-search] API error ${resp.status}:`, errText.substring(0, 200));
-      return { error: `Search API error: ${resp.status}` };
-    }
-
-    const data = await resp.json();
-    const candidate = data.candidates?.[0];
-    const answer = candidate?.content?.parts?.map(p => p.text).filter(Boolean).join("\n") || "";
-
-    // Extract grounding metadata (sources)
-    const grounding = candidate?.groundingMetadata || {};
-    const chunks = grounding.groundingChunks || [];
-    const sources = chunks
-      .filter(c => c.web)
-      .map(c => ({ title: c.web.title || "", url: c.web.uri || "" }))
-      .filter(s => s.url);
-
-    // Extract search queries used
-    const searchQueries = (grounding.webSearchQueries || []);
-
-    return {
-      success: true,
-      answer,
-      sources: sources.slice(0, 8),
-      searchQueries,
-      groundingMetadata: { sources, searchQueries },
-      mode,
-    };
-  } catch (err) {
-    if (err.name === "AbortError" || err.name === "TimeoutError") return { error: "Search timed out" };
-    return { error: err.message };
-  }
-}
-
-// =============================================================================
-// SSRF GUARD — Reject requests to private/loopback ranges (CWE-918)
-// =============================================================================
-
-const PRIVATE_IP_RE = /^(?:10\.|172\.(?:1[6-9]|2\d|3[01])\.|192\.168\.|127\.|0\.|::1|fc|fd)/i;
-
-function assertSafeUrl(rawUrl) {
-  let parsed;
-  try {
-    parsed = new URL(rawUrl);
-  } catch {
-    throw new Error(`SSRF guard: invalid URL — ${rawUrl}`);
-  }
-  if (parsed.protocol !== "https:") {
-    throw new Error(`SSRF guard: only HTTPS URLs are permitted, got ${parsed.protocol}`);
-  }
-  const host = parsed.hostname;
-  if (PRIVATE_IP_RE.test(host) || host === "localhost" || host.endsWith(".local")) {
-    throw new Error(`SSRF guard: private/loopback host rejected — ${host}`);
-  }
-}
-
-// =============================================================================
-// HELPER — Call a backend Cloud Function / API
-// =============================================================================
-
-async function callCloudFunction(url, payload, timeoutMs = 25000) {
-  // Validate the URL before making any network request.
-  assertSafeUrl(url);
-  const controller = new AbortController();
-  const timer = setTimeout(() => controller.abort(), timeoutMs);
-  try {
-    const resp = await fetch(url, {
-      method: "POST",
-      headers: { "Content-Type": "application/json" },
-      body: JSON.stringify(payload),
-      signal: controller.signal,
-    });
-    return await resp.json();
-  } catch (err) {
-    if (err.name === "AbortError") return { error: "Request timed out", timeout: timeoutMs };
-    return { error: err.message };
-  } finally {
-    clearTimeout(timer);
-  }
-}
-
-// =============================================================================
-// GOAP SEARCH PIPELINE
-// =============================================================================
-
-async function executeGoapSearch(query, args) {
-  const researchUrl = CLOUD_FUNCTIONS.research;
-  if (!researchUrl) return { error: "GOAP requires a 'research' URL in CLOUD_FUNCTIONS" };
-
-  const startTime = Date.now();
-
-  const composeResult = await callCloudFunction(researchUrl, {
-    action: "search",
-    query: `Break this question into 3-4 distinct search queries that would help answer it comprehensively. Return ONLY the queries, one per line:\n\n${query}`,
-  }, 30000);
-
-  let searchQueries = [query];
-  if (composeResult && !composeResult.error) {
-    const answer = composeResult.result?.answer || composeResult.answer || "";
-    const lines = answer.split("\n").map(l => l.replace(/^[\d\-\*\.\)]+\s*/, "").trim()).filter(l => l.length > 5 && l.length < 200);
-    if (lines.length >= 2) searchQueries = lines.slice(0, 4);
-  }
-
-  const searchResults = await Promise.all(
-    searchQueries.map(q => callCloudFunction(researchUrl, { action: "search", query: q }, 30000))
-  );
-
-  const allSources = [], allAnswers = [];
-  for (let i = 0; i < searchResults.length; i++) {
-    const r = searchResults[i];
-    if (r && !r.error && r.success !== false) {
-      const answer = r.result?.answer || r.answer || "";
-      if (answer) allAnswers.push({ query: searchQueries[i], answer });
-      const gm = r.result?.groundingMetadata || r.groundingMetadata || {};
-      if (gm.sources) allSources.push(...gm.sources);
-    }
-  }
-
-  const uniqueSources = [];
-  const seenUrls = new Set();
-  for (const src of allSources) {
-    const url = src.url || src.uri || "";
-    if (url && !seenUrls.has(url)) { seenUrls.add(url); uniqueSources.push(src); }
-  }
-
-  const synthesisInput = allAnswers.map(a => `## ${a.query}\n${a.answer}`).join("\n\n");
-  const synthesisResult = await callCloudFunction(researchUrl, {
-    action: "research",
-    topic: `Synthesize these findings into a comprehensive answer to: "${query}"\n\nFindings:\n${synthesisInput}`,
-  }, 60000);
-
-  const synthesizedAnswer = synthesisResult?.result?.answer || synthesisResult?.answer || synthesisInput;
-  const synthGm = synthesisResult?.result?.groundingMetadata || {};
-  if (synthGm.sources) {
-    for (const src of synthGm.sources) {
-      const url = src.url || src.uri || "";
-      if (url && !seenUrls.has(url)) { seenUrls.add(url); uniqueSources.push(src); }
-    }
-  }
-
-  let verification = { verified: true, confidence: "high" };
-  if (args.verify !== false && synthesizedAnswer.length > 100) {
-    const vr = await callCloudFunction(researchUrl, {
-      action: "fact_check", claim: synthesizedAnswer.substring(0, 500),
-    }, 30000);
-    if (vr && !vr.error && vr.result) {
-      verification = {
-        verified: vr.result.verdict !== "FALSE",
-        verdict: vr.result.verdict,
-        confidence: vr.result.confidence || "medium",
-        details: vr.result.analysis,
-      };
-    }
-  }
-
-  return {
-    answer: synthesizedAnswer, pipeline: "goap",
-    steps: { queries_composed: searchQueries.length, searches_executed: searchResults.filter(r => !r?.error).length, sources_found: uniqueSources.length, verification },
-    sources: uniqueSources.slice(0, 10), searchQueries, duration_ms: Date.now() - startTime,
-  };
-}
-
-// =============================================================================
-// GOAP SEARCH PIPELINE — Gemini fallback (no Cloud Function needed)
-// =============================================================================
-
-async function executeGoapSearchGemini(query, args) {
-  const startTime = Date.now();
-
-  // Step 1: Decompose into sub-queries
-  const decompose = await geminiGroundedSearch(
-    `Break this question into 3-4 distinct search queries that would help answer it comprehensively. Return ONLY the queries, one per line:\n\n${query}`,
-    "search"
-  );
-
-  let searchQueries = [query];
-  if (decompose?.answer) {
-    const lines = decompose.answer.split("\n")
-      .map(l => l.replace(/^[\d\-\*\.\)]+\s*/, "").trim())
-      .filter(l => l.length > 5 && l.length < 200);
-    if (lines.length >= 2) searchQueries = lines.slice(0, 4);
-  }
-
-  // Step 2: Parallel searches
-  const searchResults = await Promise.all(
-    searchQueries.map(q => geminiGroundedSearch(q, "search"))
-  );
-
-  const allSources = [], allAnswers = [];
-  for (let i = 0; i < searchResults.length; i++) {
-    const r = searchResults[i];
-    if (r && !r.error && r.answer) {
-      allAnswers.push({ query: searchQueries[i], answer: r.answer });
-      if (r.sources) allSources.push(...r.sources);
-    }
-  }
-
-  // Dedupe sources
-  const seenUrls = new Set();
-  const uniqueSources = allSources.filter(s => {
-    if (seenUrls.has(s.url)) return false;
-    seenUrls.add(s.url);
-    return true;
-  });
-
-  // Step 3: Synthesize
-  const synthesisInput = allAnswers.map(a => `## ${a.query}\n${a.answer}`).join("\n\n");
-  const synthesis = await geminiGroundedSearch(
-    `Synthesize these findings into a comprehensive answer to: "${query}"\n\nFindings:\n${synthesisInput}`,
-    "research"
-  );
-
-  const finalAnswer = synthesis?.answer || synthesisInput;
-  if (synthesis?.sources) {
-    for (const s of synthesis.sources) {
-      if (!seenUrls.has(s.url)) { seenUrls.add(s.url); uniqueSources.push(s); }
-    }
-  }
-
-  // Step 4: Verify if requested
-  let verification = { verified: true, confidence: "high" };
-  if (args.verify !== false && finalAnswer.length > 100) {
-    const vr = await geminiGroundedSearch(finalAnswer.substring(0, 500), "fact_check");
-    if (vr && !vr.error) {
-      verification = { verified: true, confidence: "medium", details: vr.answer };
-    }
-  }
-
-  return {
-    success: true,
-    answer: finalAnswer,
-    pipeline: "goap-gemini",
-    steps: {
-      queries_composed: searchQueries.length,
-      searches_executed: searchResults.filter(r => !r?.error).length,
-      sources_found: uniqueSources.length,
-      verification,
-    },
-    sources: uniqueSources.slice(0, 10),
-    searchQueries,
-    duration_ms: Date.now() - startTime,
-  };
-}
-
-// =============================================================================
-// TOOL EXECUTOR
-// =============================================================================
-
-// ADR-166 §6 Phase 1d + 2a — server-side tool gate.
-// Enforced HERE (not just in the autopilot handler) so /mcp, /mcp/:group,
-// autopilot, and any future path share ONE denial gate. Was the missing
-// link that made the disclosed unauthenticated-RCE chain reach shell.
-const DANGEROUS_TOOLS = Object.freeze(new Set([
-  "terminal_execute",
-  "ruflo__terminal_execute",
-  "devtools__terminal_execute",
-]));
-function isTerminalTool(name) {
-  return DANGEROUS_TOOLS.has(name) || /terminal_execute/i.test(name);
-}
-const MCP_ENABLE_TERMINAL = process.env.MCP_ENABLE_TERMINAL === "true";
-
-async function executeTool(name, args) {
-  // Deny dangerous tools unless the operator explicitly opted in.
-  // Enforced on every path (not just autopilot) — root cause of ADR-166 V2/V3.
-  if (isTerminalTool(name) && !MCP_ENABLE_TERMINAL) {
-    return {
-      error:
-        `Tool "${name}" is disabled by default. Set MCP_ENABLE_TERMINAL=true to allow.`,
-      code: "TOOL_DISABLED",
-    };
-  }
-  switch (name) {
-    case "search": {
-      if (CLOUD_FUNCTIONS.search) {
-        return callCloudFunction(CLOUD_FUNCTIONS.search, { query: args.query, limit: args.limit || 5 });
-      }
-      // Fallback: use Gemini grounded search
-      return geminiGroundedSearch(args.query, "search");
-    }
-
-    case "web_research": {
-      const action = args.action || "search";
-      if (action === "goap") {
-        // GOAP needs research endpoint — fall back to multi-search via Gemini
-        if (CLOUD_FUNCTIONS.research) return executeGoapSearch(args.query, args);
-        return executeGoapSearchGemini(args.query, args);
-      }
-      if (CLOUD_FUNCTIONS.research) {
-        const payload = { action };
-        if (action === "search") payload.query = args.query;
-        else if (action === "research") payload.topic = args.query;
-        else if (action === "compare") payload.items = args.items;
-        else if (action === "fact_check") payload.claim = args.claim || args.query;
-        return callCloudFunction(CLOUD_FUNCTIONS.research, payload, 60000);
-      }
-      // Fallback: Gemini grounded search
-      const mode = action === "fact_check" ? "fact_check" : action === "compare" ? "compare" : action === "research" ? "research" : "search";
-      const query = action === "compare" ? (args.items || []).join(" vs ") : (args.claim || args.query);
-      return geminiGroundedSearch(query, mode);
-    }
-
-    case "guidance":
-      return getGuidance(args.topic || "overview", args.tool_name);
-
-    default: {
-      // Route to external MCP backend
-      const activeTools = getActiveTools();
-      const extTool = activeTools.find(t => t.name === name);
-      if (extTool) {
-        const backend = mcpBackends.get(extTool._backend);
-        if (backend) return backend.callTool(extTool._originalName, args);
-        return { error: `Backend ${extTool._backend} not available` };
-      }
-      return { error: `Unknown tool: ${name}. Call 'guidance' with topic='groups' to see available tools.` };
-    }
-  }
-}
-
-// =============================================================================
-// PER-GROUP TOOL HELPERS
-// =============================================================================
-
-// Get tools for a specific group only
-function getToolsForGroup(groupName) {
-  const group = TOOL_GROUPS[groupName];
-  if (!group || !group.enabled) return [];
-  if (groupName === "core") return BUILTIN_TOOLS;
-
-  const allActive = getActiveTools();
-  if (!group.prefixes) {
-    // No prefix filter — return all tools from this backend
-    return allActive.filter(t => t._backend === group.source);
-  }
-  return allActive.filter(t =>
-    t._backend === group.source && group.prefixes.some(p => t._originalName.startsWith(p))
-  );
-}
-
-// Group display names for the Chat UI
-const GROUP_DISPLAY_NAMES = {
-  core: "Core Tools",
-  intelligence: "Intelligence & Learning",
-  agents: "Agents & Orchestration",
-  memory: "Memory & Knowledge",
-  devtools: "Dev Tools & Analysis",
-  security: "Security & Safety",
-  browser: "Browser Automation",
-  neural: "Neural & DAA",
-  "agentic-flow": "Agentic Flow",
-  "claude-code": "Claude Code",
-  gemini: "Gemini",
-  codex: "Codex",
-};
-
-// =============================================================================
-// MCP SERVER — Multiple endpoints per group
-// =============================================================================
-
-const app = express();
-app.use(express.json({ limit: "10mb" }));
-
-// ---------- MCP Streamable HTTP session (#2425 djimit) ----------
-// Streamable-HTTP clients (Codex/RMCP) send `DELETE /mcp` with an
-// `Mcp-Session-Id` header at shutdown. We echo a stable session id back
-// on every /mcp* response so those clients can attach it to the DELETE
-// and to `notifications/initialized` handshakes.
-const MCP_SESSION_ID = randomUUID();
-app.use((req, res, next) => {
-  if (req.path.startsWith("/mcp")) {
-    res.setHeader("Mcp-Session-Id", MCP_SESSION_ID);
-  }
-  next();
-});
-
-// ---------- CORS middleware (ADR-166 §6 Phase 3b) ----------
-// MCP_CORS_ORIGIN: comma-separated allowlist (e.g. "https://a.example,https://b.example").
-//   unset → "*" for back-compat (loopback default is same-origin anyway)
-//   "*"   → wildcard (explicit opt-in — same as legacy behavior)
-//   other → echo the request origin ONLY if it appears in the allowlist
-const CORS_ALLOWLIST = (process.env.MCP_CORS_ORIGIN || "*")
-  .split(",").map(s => s.trim()).filter(Boolean);
-const CORS_WILDCARD = CORS_ALLOWLIST.length === 1 && CORS_ALLOWLIST[0] === "*";
-app.use((req, res, next) => {
-  const origin = req.get("origin") || "";
-  if (CORS_WILDCARD) {
-    res.setHeader("Access-Control-Allow-Origin", "*");
-  } else if (origin && CORS_ALLOWLIST.includes(origin)) {
-    res.setHeader("Access-Control-Allow-Origin", origin);
-    res.setHeader("Vary", "Origin");
-  }
-  // If no match, do NOT set the header — browser will block the request.
-  res.setHeader("Access-Control-Allow-Methods", "GET, POST, DELETE, OPTIONS");
-  res.setHeader("Access-Control-Allow-Headers", "Content-Type, Authorization, Mcp-Session-Id");
-  if (req.method === "OPTIONS") return res.sendStatus(204);
-  next();
-});
-
-// ---------- Auth middleware ----------
-// No-op in local-only mode (MCP_AUTH_TOKEN unset). Enforces 401 when token is set.
-const MCP_TOKEN = process.env.MCP_AUTH_TOKEN || "";
-function requireAuth(req, res, next) {
-  if (req.path === "/health") return next();
-  if (!MCP_TOKEN) return next();
-  const expected = `Bearer ${MCP_TOKEN}`;
-  const got = req.get("authorization") || "";
-  const ok = got.length === expected.length &&
-    timingSafeEqual(Buffer.from(got), Buffer.from(expected));
-  if (!ok) return res.status(401).json({ error: "unauthorized" });
-  next();
-}
-app.use(requireAuth);
-
-// ---------- Shared MCP handler ----------
-function createMcpHandler(groupName) {
-  return async (req, res) => {
-    const { method, id, params } = req.body;
-    try {
-      switch (method) {
-        case "initialize":
-          return res.json({
-            jsonrpc: "2.0", id,
-            result: {
-              protocolVersion: "2024-11-05",
-              capabilities: { tools: {} },
-              serverInfo: { name: `mcp-bridge/${groupName}`, version: "2.0.0" },
-            },
-          });
-        case "tools/list": {
-          const tools = getToolsForGroup(groupName);
-          return res.json({ jsonrpc: "2.0", id, result: { tools } });
-        }
-        case "tools/call": {
-          const { name, arguments: toolArgs } = params;
-          const result = await executeTool(name, toolArgs || {});
-          return res.json({
-            jsonrpc: "2.0", id,
-            result: {
-              content: [{ type: "text", text: typeof result === "string" ? result : JSON.stringify(result, null, 2) }],
-            },
-          });
-        }
-        case "notifications/initialized":
-          // MCP streamable-HTTP spec: notifications must return 202 Accepted
-          // with an empty body (no jsonrpc envelope).
-          return res.status(202).end();
-        default:
-          return res.json({ jsonrpc: "2.0", id, error: { code: -32601, message: `Method not found: ${method}` } });
-      }
-    } catch (err) {
-      console.error(`MCP error [${groupName}/${method}]:`, err);
-      return res.json({ jsonrpc: "2.0", id, error: { code: -32603, message: err.message } });
-    }
-  };
-}
-
-function createMcpSseHandler(groupName) {
-  return (req, res) => {
-    res.setHeader("Content-Type", "text/event-stream");
-    res.setHeader("Cache-Control", "no-cache");
-    res.setHeader("Connection", "keep-alive");
-    res.write(`data: ${JSON.stringify({ type: "endpoint", url: `/mcp/${groupName}` })}\n\n`);
-  };
-}
-
-// ---------- Register per-group endpoints ----------
-for (const groupName of Object.keys(TOOL_GROUPS)) {
-  app.post(`/mcp/${groupName}`, createMcpHandler(groupName));
-  app.get(`/mcp/${groupName}`, createMcpSseHandler(groupName));
-  // #2425 djimit — streamable-HTTP session cleanup
-  app.delete(`/mcp/${groupName}`, (_, res) => res.sendStatus(204));
-}
-
-// ---------- Catch-all /mcp — serves ALL enabled tools (backwards-compatible) ----------
-app.post("/mcp", async (req, res) => {
-  const { method, id, params } = req.body;
-  try {
-    switch (method) {
-      case "initialize":
-        return res.json({
-          jsonrpc: "2.0", id,
-          result: {
-            protocolVersion: "2024-11-05",
-            capabilities: { tools: {} },
-            serverInfo: { name: "mcp-bridge", version: "2.0.0" },
-          },
-        });
-      case "tools/list": {
-        const activeTools = getActiveTools();
-        return res.json({ jsonrpc: "2.0", id, result: { tools: [...BUILTIN_TOOLS, ...activeTools] } });
-      }
-      case "tools/call": {
-        const { name, arguments: toolArgs } = params;
-        const result = await executeTool(name, toolArgs || {});
-        return res.json({
-          jsonrpc: "2.0", id,
-          result: {
-            content: [{ type: "text", text: typeof result === "string" ? result : JSON.stringify(result, null, 2) }],
-          },
-        });
-      }
-      case "notifications/initialized":
-        return res.status(202).end();
-      default:
-        return res.json({ jsonrpc: "2.0", id, error: { code: -32601, message: `Method not found: ${method}` } });
-    }
-  } catch (err) {
-    console.error(`MCP error [${method}]:`, err);
-    return res.json({ jsonrpc: "2.0", id, error: { code: -32603, message: err.message } });
-  }
-});
-
-app.get("/mcp", (req, res) => {
-  res.setHeader("Content-Type", "text/event-stream");
-  res.setHeader("Cache-Control", "no-cache");
-  res.setHeader("Connection", "keep-alive");
-  res.write(`data: ${JSON.stringify({ type: "endpoint", url: "/mcp" })}\n\n`);
-});
-
-// #2425 djimit — streamable-HTTP session cleanup on the catch-all route.
-app.delete("/mcp", (_, res) => res.sendStatus(204));
-
-// ---------- GET /mcp-servers — returns MCP_SERVERS JSON for Chat UI config ----------
-app.get("/mcp-servers", (_, res) => {
-  const servers = [];
-  for (const [name, group] of Object.entries(TOOL_GROUPS)) {
-    if (!group.enabled) continue;
-    const tools = getToolsForGroup(name);
-    if (tools.length === 0) continue;
-    servers.push({
-      name: GROUP_DISPLAY_NAMES[name] || name,
-      url: `/mcp/${name}`,
-      tools: tools.length,
-      group: name,
-    });
-  }
-  res.json(servers);
-});
-
-// =============================================================================
-// CHAT COMPLETIONS PROXY
-// =============================================================================
-
-const PROVIDER_ROUTES = {
-  openai: { baseURL: "https://api.openai.com/v1/chat/completions", getKey: () => process.env.OPENAI_API_KEY },
-  gemini: { baseURL: "https://generativelanguage.googleapis.com/v1beta/openai/chat/completions", getKey: () => process.env.GOOGLE_API_KEY },
-  openrouter: { baseURL: "https://openrouter.ai/api/v1/chat/completions", getKey: () => process.env.OPENROUTER_API_KEY },
-};
-
-function resolveProvider(model) {
-  if (typeof model === "string") {
-    if (model.startsWith("gemini-")) return "gemini";
-    if (model.includes("/")) return "openrouter";
-  }
-  return "openai";
-}
-
-// =============================================================================
-// SYSTEM PROMPT — Injected server-side into every chat completion request
-// =============================================================================
-// This comprehensive prompt teaches the AI how to use all 200+ MCP tools
-// across 5 groups. It is injected as the first system message, ensuring
-// consistent behavior regardless of what preprompt the Chat UI sends.
-
-function buildSystemPrompt() {
-  // Build dynamic group status
-  const enabledGroups = Object.entries(TOOL_GROUPS)
-    .filter(([, g]) => g.enabled)
-    .map(([name]) => name);
-
-  return `You are an intelligent AI assistant with powerful tools organized into ${enabledGroups.length} active groups.
-
-# CRITICAL RULES
-1. Use tools proactively — NEVER guess answers, ALWAYS search first
-2. For ANY factual question, current event, or research request → call \`web_research\` IMMEDIATELY
-3. NEVER say "I don't have access to real-time information" — you DO via web_research
-4. NEVER say "I can't search the web" — you CAN via web_research
-5. Call tools FIRST, present results conversationally AFTER
-6. When multiple tools could help, call them ALL in parallel
-
-# Tool Groups
-
-Tools prefixed with backend name (e.g., \`ruflo__agent_spawn\`). Always use the full prefixed name.
-
-## Group 1: Core Tools (always on)
-
-- **search** — Search your knowledge base (documents, workflows, how-tos).
-  Use for: internal knowledge, company docs, past conversations.
-  \`{"query": "how to process a claim"}\`
-
-- **web_research** — Search the internet via Google. This is your primary research tool.
-  ALWAYS use this for: current events, facts, comparisons, any external knowledge.
-  Actions:
-  - \`search\` — Quick web search. Default, fast. \`{"action": "search", "query": "latest news on X"}\`
-  - \`research\` — Deep report with synthesis. \`{"action": "research", "query": "comprehensive analysis of X"}\`
-  - \`compare\` — Side-by-side comparison. \`{"action": "compare", "query": "X", "items": ["A", "B"]}\`
-  - \`fact_check\` — Verify a claim. \`{"action": "fact_check", "query": "claim to verify"}\`
-  - \`goap\` — Multi-step pipeline: decomposes → parallel searches → synthesizes → verifies. BEST for important questions.
-    \`{"action": "goap", "query": "complex question requiring thorough research"}\`
-
-- **guidance** — Get help on tool groups or specific tool usage.
-  Topics: \`overview\`, \`groups\`, \`agents\`, \`memory\`, \`intelligence\`, \`devtools\`
-  \`{"topic": "overview"}\` or \`{"topic": "tool", "tool_name": "ruflo__agent_spawn"}\`
-
-## Group 2: Intelligence & Learning (ruvector)
-Pattern learning, routing, code analysis, and trajectory tracking. ${TOOL_GROUPS.intelligence.enabled ? "ACTIVE" : "DISABLED"}
-
-### Essential Intelligence Tools:
-- **ruvector__hooks_route** — Route a task to the optimal agent type. Call this FIRST for complex tasks.
-  \`{"task": "describe what needs to be done", "context": ["relevant info"]}\`
-  Returns ranked agent recommendations with confidence scores.
-- **ruvector__hooks_remember** — Store a key-value pair in persistent memory for cross-session recall.
-  \`{"key": "pattern-name", "value": "what to remember", "namespace": "patterns"}\`
-- **ruvector__hooks_recall** — Retrieve a previously stored memory by key.
-- **ruvector__hooks_suggest_context** — Get contextual suggestions based on current work.
-- **ruvector__hooks_swarm_recommend** — Get swarm topology recommendation for a task type.
-- **ruvector__hooks_capabilities** — List all intelligence system capabilities.
-
-### Code Analysis:
-- **ruvector__hooks_ast_analyze** — Analyze code structure (AST) of a file.
-- **ruvector__hooks_ast_complexity** — Get complexity metrics for code.
-- **ruvector__hooks_security_scan** — Scan code for security vulnerabilities.
-- **ruvector__hooks_diff_analyze** — Analyze a code diff for risk and impact.
-- **ruvector__hooks_diff_similar** — Find similar past diffs/changes.
-
-### Trajectory Learning (for multi-step tasks):
-- **ruvector__hooks_trajectory_begin** — Start tracking a multi-step task for learning.
-- **ruvector__hooks_trajectory_step** — Record a step in the current trajectory.
-- **ruvector__hooks_trajectory_end** — End trajectory, triggering pattern extraction.
-
-### Memory & Compression:
-- **ruvector__hooks_compress** — Compress/summarize long text for efficient storage.
-- **ruvector__hooks_rag_context** — Get RAG context for a query from stored knowledge.
-- **ruvector__hooks_learn** — Force the system to learn from provided examples.
-- **ruvector__hooks_batch_learn** — Learn from multiple examples at once.
-- **ruvector__hooks_stats** — View learning statistics and metrics.
-- **ruvector__hooks_doctor** — Run diagnostics on the intelligence system.
-
-## Group 3: Agents & Orchestration (ruflo)
-Spawn agents, coordinate swarms, manage tasks and workflows. ${TOOL_GROUPS.agents.enabled ? "ACTIVE" : "DISABLED"}
-
-### Agent Lifecycle:
-- **ruflo__agent_spawn** — Create a new specialized agent.
-  \`{"type": "coder|researcher|tester|reviewer|architect|security", "name": "optional-name"}\`
-  Agent types and when to use them:
-  - \`coder\` — Write code, implement features, fix bugs
-  - \`researcher\` — Find information, analyze documentation, investigate
-  - \`tester\` — Write tests, run test suites, validate behavior
-  - \`reviewer\` — Review code quality, security, best practices
-  - \`architect\` — Design systems, plan architectures, evaluate trade-offs
-  - \`security\` — Audit security, find vulnerabilities, recommend fixes
-- **ruflo__agent_status** — Check an agent's current state. \`{"agentId": "agent-xxx"}\`
-- **ruflo__agent_list** — List all active agents with their states.
-- **ruflo__agent_terminate** — Stop an agent. \`{"agentId": "agent-xxx"}\`
-- **ruflo__agent_health** — Health check across all agents.
-- **ruflo__agent_pool** — View the agent pool and available capacity.
-
-### Swarm Coordination:
-- **ruflo__swarm_init** — Initialize a multi-agent swarm.
-  \`{"topology": "hierarchical|mesh|ring|star", "maxAgents": 8, "strategy": "balanced|specialized|adaptive"}\`
-  - \`hierarchical\` — Coordinator + workers, best for structured tasks (anti-drift)
-  - \`mesh\` — Peer-to-peer, best for collaborative work
-  - \`ring\` — Sequential pipeline, best for ordered processing
-  - \`star\` — Central hub, best for fan-out parallel work
-- **ruflo__swarm_status** — Get swarm health, topology, and agent states.
-- **ruflo__swarm_health** — Detailed health metrics for the swarm.
-- **ruflo__swarm_shutdown** — Tear down a swarm and all its agents.
-
-### Task Management:
-- **ruflo__task_create** — Create a tracked task.
-  \`{"description": "what needs to be done", "priority": "low|normal|high|critical"}\`
-- **ruflo__task_status** — Check task progress. \`{"taskId": "task-xxx"}\`
-- **ruflo__task_list** — List all tasks with their statuses.
-- **ruflo__task_complete** — Mark a task as done. \`{"taskId": "task-xxx"}\`
-- **ruflo__task_update** — Update task details, status, or assignment.
-- **ruflo__task_cancel** — Cancel a task.
-
-### Workflow Orchestration:
-- **ruflo__workflow_create** — Define a multi-step workflow with dependencies.
-- **ruflo__workflow_execute** — Run a workflow. \`{"workflowId": "wf-xxx"}\`
-- **ruflo__workflow_status** — Check workflow progress.
-- **ruflo__workflow_template** — Use a pre-built workflow template.
-- **ruflo__workflow_pause** / **ruflo__workflow_resume** — Control workflow execution.
-
-### Hive-Mind (Distributed Consensus):
-- **ruflo__hive-mind_init** — Start distributed consensus system.
-- **ruflo__hive-mind_spawn** — Add an agent to the hive.
-- **ruflo__hive-mind_consensus** — Run consensus vote across agents.
-- **ruflo__hive-mind_broadcast** — Send message to all hive agents.
-- **ruflo__hive-mind_memory** — Access shared hive memory.
-
-### Coordination:
-- **ruflo__coordination_topology** — View/change coordination topology.
-- **ruflo__coordination_load_balance** — Distribute work across agents.
-- **ruflo__coordination_orchestrate** — Orchestrate complex multi-agent tasks.
-- **ruflo__coordination_sync** — Synchronize state across agents.
-
-### Session Management:
-- **ruflo__session_save** — Save current session state.
-- **ruflo__session_restore** — Restore a previous session.
-- **ruflo__session_list** — List available sessions.
-
-## Group 4: Memory & Knowledge (ruflo)
-Persistent memory, vector search, embeddings, and pattern storage. ${TOOL_GROUPS.memory.enabled ? "ACTIVE" : "DISABLED"}
-
-### Memory Operations:
-- **ruflo__memory_store** — Store data in persistent memory.
-  \`{"key": "my-key", "value": "data to store", "namespace": "default", "tags": ["tag1"]}\`
-- **ruflo__memory_retrieve** — Get stored data by key. \`{"key": "my-key"}\`
-- **ruflo__memory_search** — Semantic vector search across stored memories.
-  \`{"query": "what to search for", "limit": 5, "namespace": "default"}\`
-- **ruflo__memory_list** — List all stored keys in a namespace.
-- **ruflo__memory_delete** — Remove a stored memory.
-- **ruflo__memory_stats** — View memory usage statistics.
-
-### Embeddings:
-- **ruflo__embeddings_generate** — Generate vector embeddings for text.
-- **ruflo__embeddings_compare** — Compare semantic similarity of two texts.
-- **ruflo__embeddings_search** — Search embeddings database by similarity.
-- **ruflo__embeddings_neural** — Generate neural embeddings.
-- **ruflo__embeddings_hyperbolic** — Generate hyperbolic embeddings for hierarchical data.
-
-### AgentDB (Advanced Pattern Storage):
-- **ruflo__agentdb_pattern-store** — Store a learned pattern with metadata.
-  \`{"pattern": "description", "category": "code|debug|architecture", "confidence": 0.9}\`
-- **ruflo__agentdb_pattern-search** — Search patterns by similarity.
-- **ruflo__agentdb_route** — Route a query to the most relevant stored pattern.
-- **ruflo__agentdb_feedback** — Provide feedback on a pattern (reinforcement learning).
-- **ruflo__agentdb_context-synthesize** — Synthesize context from multiple sources.
-- **ruflo__agentdb_semantic-route** — Semantic routing based on stored knowledge.
-- **ruflo__agentdb_consolidate** — Consolidate and deduplicate stored patterns.
-- **ruflo__agentdb_batch** — Batch operations on patterns.
-- **ruflo__agentdb_session-start** / **ruflo__agentdb_session-end** — Session tracking.
-- **ruflo__agentdb_hierarchical-store** / **ruflo__agentdb_hierarchical-recall** — Hierarchical memory.
-
-## Group 5: Dev Tools & Analysis (ruflo)
-Performance, system health, GitHub integration, code analysis, terminal. ${TOOL_GROUPS.devtools.enabled ? "ACTIVE" : "DISABLED"}
-
-### System & Performance:
-- **ruflo__system_status** — System health overview.
-- **ruflo__system_metrics** — Detailed performance metrics.
-- **ruflo__system_health** — Health check across all subsystems.
-- **ruflo__performance_report** — Generate performance report.
-- **ruflo__performance_bottleneck** — Identify performance bottlenecks.
-- **ruflo__performance_benchmark** — Run benchmarks.
-- **ruflo__performance_optimize** — Get optimization recommendations.
-- **ruflo__performance_profile** — Profile specific operations.
-
-### Code Analysis:
-- **ruflo__analyze_diff** — Analyze a code diff.
-- **ruflo__analyze_diff-risk** — Assess risk level of changes.
-- **ruflo__analyze_diff-classify** — Classify type of changes (feature, bugfix, refactor).
-- **ruflo__analyze_diff-reviewers** — Suggest code reviewers.
-- **ruflo__analyze_file-risk** — Assess risk of a specific file.
-
-### GitHub Integration:
-- **ruflo__github_repo_analyze** — Analyze a GitHub repository.
-  \`{"repo": "owner/repo", "analysis_type": "code_quality|performance|security"}\`
-- **ruflo__github_pr_manage** — Manage pull requests (create, review, merge).
-- **ruflo__github_issue_track** — Track and manage issues.
-- **ruflo__github_workflow** — Manage GitHub Actions workflows.
-- **ruflo__github_metrics** — Repository metrics and insights.
-
-### Terminal Access:
-- **ruflo__terminal_create** — Create a terminal session.
-- **ruflo__terminal_execute** — Execute a command. \`{"command": "ls -la"}\`
-- **ruflo__terminal_list** — List active terminals.
-- **ruflo__terminal_history** — View command history.
-
-### Development Hooks:
-- **ruflo__hooks_pre-task** / **ruflo__hooks_post-task** — Task lifecycle hooks for learning.
-- **ruflo__hooks_pre-edit** / **ruflo__hooks_post-edit** — File edit hooks.
-- **ruflo__hooks_session-start** / **ruflo__hooks_session-end** — Session lifecycle.
-- **ruflo__hooks_worker-dispatch** — Dispatch background workers.
-  Workers: \`optimize\`, \`audit\`, \`testgaps\`, \`document\`, \`map\`, \`deepdive\`, \`benchmark\`
-- **ruflo__hooks_model-route** — Route to optimal AI model for a task.
-- **ruflo__hooks_explain** — Explain a routing or intelligence decision.
-
-### Configuration:
-- **ruflo__config_get** / **ruflo__config_set** / **ruflo__config_list** — Manage settings.
-
-### Progress Tracking:
-- **ruflo__progress_check** — Check implementation progress.
-- **ruflo__progress_summary** — Summarize overall progress.
-
-# Decision Framework — FOLLOW THIS EXACTLY
-
-1. **ANY factual question** → \`web_research(action='search')\` IMMEDIATELY. Do NOT answer from memory.
-2. **"What is X?"** → \`web_research(action='search', query='X')\`
-3. **"Compare X vs Y"** → \`web_research(action='compare', query='X vs Y', items=['X', 'Y'])\`
-4. **"Is it true that..."** → \`web_research(action='fact_check', query='...')\`
-5. **Complex research** → \`web_research(action='goap', query='...')\` — auto multi-step pipeline
-6. **Internal docs/procedures** → \`search(query='...')\` first, then web_research if not found
-7. **Code task** → \`ruvector__hooks_route\` + \`ruflo__agent_spawn\`
-8. **Remember something** → \`ruflo__memory_store\` / \`ruflo__memory_search\`
-9. **"What can you do?"** → \`guidance(topic='overview')\`
-10. **Unknown** → \`guidance(topic='overview')\`
-
-IMPORTANT: For questions 1-5, ALWAYS call web_research. Never say "I don't have access to search" — you DO.
-
-# Execution Patterns
-
-### Simple Question
-\`search\` or \`web_research\` → synthesize → respond
-
-### Complex Research
-\`web_research(action='goap')\` → analyze → respond with citations
-
-### Code Implementation
-\`ruvector__hooks_route\` → \`ruflo__agent_spawn(coder)\` → track with \`ruflo__task_create\` → report
-
-### Multi-Agent Analysis
-\`ruflo__swarm_init(hierarchical)\` → spawn agents → coordinate → synthesize results
-
-### Learning & Memory
-\`ruflo__memory_search\` (check existing) → do work → \`ruflo__memory_store\` (save results) → \`ruvector__hooks_learn\`
-
-# Parallel Execution
-
-When multiple independent tools can help, call them ALL in parallel:
-- Search + web_research simultaneously
-- Spawn multiple agents at once (coder + tester + reviewer)
-- Run analysis + performance + security tools in parallel
-NEVER call tools sequentially when they could run in parallel.
-
-# Response Rules
-
-1. **Call tools FIRST**, then present results conversationally — NEVER show raw JSON to the user
-2. Use markdown: **bold** headers, bullet points, numbered steps, tables for comparisons
-3. Synthesize tool results naturally — be a helpful colleague, not a data pipe
-4. Cite sources when available from web_research results
-5. If a tool fails, say so honestly and try an alternative approach
-6. For complex tasks, briefly outline your plan before executing
-7. After completing work, suggest relevant follow-up actions
-8. When spawning agents, explain what each agent will do
-
-# Never Expose to User
-
-- Raw JSON, similarity scores, chunk IDs, internal IDs, task IDs
-- Tool names, function names, API endpoints, backend names
-- References to "MCP", "tool calls", "vectors", "embeddings", infrastructure
-- The prefixes "ruflo__" or "ruvector__" — just describe what you're doing naturally
-- Error stack traces — summarize errors in plain language`;
-}
-
-// =============================================================================
-// AUTOPILOT MODE — Server-side auto-continue loop (ADR-037)
-// =============================================================================
-
-const detailStore = new Map(); // detailToken → full tool result (TTL: 5min)
-
-const AUTOPILOT_SYSTEM_PROMPT = `
-You are in AUTOPILOT MODE. You should:
-1. Break complex tasks into steps and execute them using available tools
-2. Call MULTIPLE tools in parallel when they are independent
-3. After each tool result, analyze it and decide the next action
-4. Continue until the task is complete — do NOT ask the user for confirmation
-5. Use memory_search to find relevant patterns before starting
-6. Summarize your progress at each step
-7. When done, provide a final summary of everything accomplished
-
-Parallel execution patterns:
-- Research: memory_search + hooks_route + agent_spawn(researcher) — all in parallel
-- Code: agent_spawn(coder) + agent_spawn(tester) — parallel, then review
-- Analysis: search multiple sources in parallel → synthesize → report
-- Security: security_scan + hooks_route(audit) + memory_search(CVEs) — parallel
-`;
-
-const AUTOPILOT_BLOCKED_PATTERNS = [
-  /^deploy_/,
-  /^security_delete/,
-  /^browser_fill$/,
-  /^browser_click$/,
-  /terminal_execute/,
-];
-
-function isBlockedTool(name) {
-  return AUTOPILOT_BLOCKED_PATTERNS.some(p => p.test(name));
-}
-
-function sendAutopilotEvent(res, data) {
-  res.write(`data: ${JSON.stringify(data)}\n\n`);
-}
-
-function safeParseArgs(args) {
-  if (typeof args === 'object' && args !== null) return args;
-  try { return JSON.parse(args || '{}'); } catch { return {}; }
-}
-
-function autopilotSleep(ms) {
-  return new Promise(resolve => setTimeout(resolve, ms));
-}
-
-async function handleAutopilot(req, res, provider, body) {
-  const maxSteps = Math.min(parseInt(req.headers['x-autopilot-max-steps'] || '20', 10), 50);
-  const cooldownMs = parseInt(process.env.AUTOPILOT_COOLDOWN || '500', 10);
-  const stepTimeoutMs = parseInt(process.env.AUTOPILOT_STEP_TIMEOUT || '30000', 10);
-
-  // SSE setup
-  res.setHeader('Content-Type', 'text/event-stream');
-  res.setHeader('Cache-Control', 'no-cache');
-  res.setHeader('Connection', 'keep-alive');
-  res.setHeader('X-Accel-Buffering', 'no');
-
-  let messages = [...body.messages];
-  let step = 0;
-  let aborted = false;
-  let totalTasks = 0;
-  const startTime = Date.now();
-
-  req.on('close', () => { aborted = true; });
-
-  sendAutopilotEvent(res, { type: 'autopilot_start', maxSteps });
-
-  // Get the tools list for the AI provider (OpenAI function calling format)
-  // Limit to BUILTIN_TOOLS + top external tools to avoid overwhelming the model
-  // (Gemini struggles with 100+ tool definitions)
-  const MAX_AUTOPILOT_TOOLS = 30;
-  const externalTools = getActiveTools();
-  // Prioritize tools from enabled high-value groups
-  const priorityPrefixes = ['ruvector__hooks', 'ruvector__memory', 'ruflo__memory', 'ruflo__agent'];
-  const prioritized = externalTools
-    .sort((a, b) => {
-      const aP = priorityPrefixes.some(p => a.name.startsWith(p)) ? 0 : 1;
-      const bP = priorityPrefixes.some(p => b.name.startsWith(p)) ? 0 : 1;
-      return aP - bP;
-    })
-    .slice(0, MAX_AUTOPILOT_TOOLS - BUILTIN_TOOLS.length);
-  const allTools = [...BUILTIN_TOOLS, ...prioritized];
-  const toolDefs = allTools.map(t => ({
-    type: 'function',
-    function: {
-      name: t.name,
-      description: t.description || '',
-      parameters: t.inputSchema || { type: 'object', properties: {} },
-    },
-  }));
-  console.log(`[autopilot] ${allTools.length} tools (${BUILTIN_TOOLS.length} builtin + ${prioritized.length} external)`);
-
-  while (step < maxSteps && !aborted) {
-    // 1. Call upstream AI provider (non-streaming for tool call parsing)
-    const apiKey = provider.getKey();
-    let aiResult;
-    try {
-      const aiResponse = await fetch(provider.baseURL, {
-        method: 'POST',
-        headers: { 'Content-Type': 'application/json', Authorization: `Bearer ${apiKey}` },
-        body: JSON.stringify({
-          ...body,
-          messages,
-          stream: false,
-          tools: toolDefs.length > 0 ? toolDefs : undefined,
-        }),
-        signal: AbortSignal.timeout(stepTimeoutMs),
-      });
-      aiResult = await aiResponse.json();
-    } catch (err) {
-      sendAutopilotEvent(res, { type: 'autopilot_error', error: `AI call failed: ${err.message}` });
-      break;
-    }
-
-    const choice = aiResult.choices?.[0];
-    if (!choice) {
-      sendAutopilotEvent(res, { type: 'autopilot_error', error: 'No response from AI' });
-      break;
-    }
-
-    // 2. Check for tool calls
-    const toolCalls = choice.message?.tool_calls;
-
-    if (!toolCalls || toolCalls.length === 0) {
-      // Final text response — send it
-      sendAutopilotEvent(res, { type: 'autopilot_text', content: choice.message?.content || '' });
-      break;
-    }
-
-    // 3. Execute ALL tool calls in parallel
-    step++;
-    const groupId = `g${step}`;
-    const taskEvents = toolCalls.map((tc, i) => ({
-      taskId: `t${totalTasks + i + 1}`,
-      tool: tc.function.name,
-      args: safeParseArgs(tc.function.arguments),
-      status: 'running',
-    }));
-    totalTasks += taskEvents.length;
-
-    // If the AI also included text content, stream it before tools
-    if (choice.message?.content) {
-      sendAutopilotEvent(res, { type: 'autopilot_text', content: choice.message.content });
-    }
-
-    // Stream group start
-    sendAutopilotEvent(res, { type: 'task_group_start', groupId, step, tasks: taskEvents });
-
-    // Append assistant message to conversation
-    messages.push(choice.message);
-
-    // Execute tools in parallel
-    const groupStart = Date.now();
-    const results = await Promise.allSettled(
-      toolCalls.map(async (tc, i) => {
-        const taskId = taskEvents[i].taskId;
-        const toolName = tc.function.name;
-        const toolArgs = safeParseArgs(tc.function.arguments);
-        const taskStart = Date.now();
-
-        // Check blocklist
-        if (isBlockedTool(toolName)) {
-          sendAutopilotEvent(res, {
-            type: 'task_update', taskId, status: 'blocked',
-            summary: `${toolName} requires confirmation`,
-            duration: Date.now() - taskStart,
-          });
-          return { toolCallId: tc.id, blocked: true, toolName };
-        }
-
-        try {
-          const result = await executeTool(toolName, toolArgs);
-          const resultStr = typeof result === 'string' ? result : JSON.stringify(result, null, 2);
-
-          // Store full detail, generate token for lazy loading
-          const detailToken = `dt_${taskId}_${Date.now()}`;
-          detailStore.set(detailToken, resultStr);
-
-          // Stream task completion with summary only
-          const summary = resultStr.length > 120
-            ? resultStr.substring(0, 120).replace(/\n/g, ' ') + '...'
-            : resultStr.replace(/\n/g, ' ');
-
-          sendAutopilotEvent(res, {
-            type: 'task_update', taskId, status: 'completed',
-            summary, duration: Date.now() - taskStart, detailToken,
-          });
-
-          return { toolCallId: tc.id, content: resultStr };
-        } catch (err) {
-          sendAutopilotEvent(res, {
-            type: 'task_update', taskId, status: 'failed',
-            summary: err.message, duration: Date.now() - taskStart,
-          });
-          return { toolCallId: tc.id, content: `Error: ${err.message}` };
-        }
-      })
-    );
-
-    // Stream group end
-    sendAutopilotEvent(res, { type: 'task_group_end', groupId, step, duration: Date.now() - groupStart });
-
-    // Check if any tools were blocked — pause autopilot
-    const blockedResults = results
-      .filter(r => r.status === 'fulfilled' && r.value.blocked)
-      .map(r => r.value);
-    if (blockedResults.length > 0) {
-      sendAutopilotEvent(res, {
-        type: 'autopilot_paused',
-        reason: 'blocked_tools',
-        tools: blockedResults.map(b => b.toolName),
-      });
-      break;
-    }
-
-    // Append tool results to messages
-    for (const r of results) {
-      if (r.status === 'fulfilled' && !r.value.blocked) {
-        messages.push({
-          role: 'tool',
-          tool_call_id: r.value.toolCallId,
-          content: r.value.content,
-        });
-      }
-    }
-
-    // Cooldown to prevent runaway
-    await autopilotSleep(cooldownMs);
-  }
-
-  if (step >= maxSteps && !aborted) {
-    sendAutopilotEvent(res, {
-      type: 'autopilot_text',
-      content: `\n⚠️ Autopilot reached max steps (${maxSteps}). Stopping.\n`,
-    });
-  }
-
-  sendAutopilotEvent(res, {
-    type: 'autopilot_end',
-    totalSteps: step,
-    totalTasks,
-    duration: Date.now() - startTime,
-  });
-
-  res.write('data: [DONE]\n\n');
-  res.end();
-
-  // Clean up detail store after 5 minutes
-  const detailTTL = parseInt(process.env.AUTOPILOT_DETAIL_TTL || '300000', 10);
-  setTimeout(() => {
-    for (const [key] of detailStore) {
-      if (key.startsWith('dt_')) detailStore.delete(key);
-    }
-  }, detailTTL);
-}
-
-// Lazy detail loading endpoint
-app.get('/autopilot/detail/:token', (req, res) => {
-  const content = detailStore.get(req.params.token);
-  if (content) {
-    res.json({ content });
-  } else {
-    res.status(404).json({ error: 'Detail expired or not found' });
-  }
-});
-
-// =============================================================================
-// CHAT COMPLETIONS PROXY
-// =============================================================================
-
-app.post("/chat/completions", async (req, res) => {
-  const model = req.body?.model;
-  const providerName = resolveProvider(model);
-  const provider = PROVIDER_ROUTES[providerName];
-  const apiKey = provider.getKey();
-
-  if (!apiKey) return res.status(401).json({ error: { message: `No API key for provider: ${providerName}` } });
-
-  // Inject comprehensive system prompt as the first message
-  const body = { ...req.body };
-  if (body.messages && Array.isArray(body.messages)) {
-    let systemPrompt = buildSystemPrompt();
-    // Add autopilot instructions if autopilot mode is active
-    const isAutopilot = req.headers['x-autopilot'] === 'true';
-    if (isAutopilot) {
-      systemPrompt = AUTOPILOT_SYSTEM_PROMPT + '\n\n' + systemPrompt;
-    }
-    // Prepend our system prompt before any existing messages
-    const hasSystemMsg = body.messages[0]?.role === "system";
-    if (hasSystemMsg) {
-      // Merge with existing system message
-      body.messages = [
-        { role: "system", content: systemPrompt + "\n\n" + body.messages[0].content },
-        ...body.messages.slice(1),
-      ];
-    } else {
-      body.messages = [{ role: "system", content: systemPrompt }, ...body.messages];
-    }
-  }
-
-  // Route to autopilot handler if x-autopilot header is set
-  if (req.headers['x-autopilot'] === 'true') {
-    return handleAutopilot(req, res, provider, body);
-  }
-
-  try {
-    const upstream = await fetch(provider.baseURL, {
-      method: "POST",
-      headers: { "Content-Type": "application/json", Authorization: `Bearer ${apiKey}` },
-      body: JSON.stringify(body),
-    });
-
-    if (!upstream.ok) {
-      const errBody = await upstream.text();
-      console.error(`Proxy error [${providerName}/${model}]: ${upstream.status} ${errBody.substring(0, 200)}`);
-      // Normalize all upstream errors into OpenAI-compatible format so the
-      // Chat UI's OpenAI SDK can parse them instead of "400 (no body)".
-      let errorMessage = `Upstream ${providerName} error (${upstream.status})`;
-      try {
-        const parsed = JSON.parse(errBody);
-        // Gemini returns [{"error": {"message": "..."}}]
-        if (Array.isArray(parsed) && parsed[0]?.error?.message) {
-          errorMessage = parsed[0].error.message;
-        // OpenAI/OpenRouter return {"error": {"message": "..."}}
-        } else if (parsed.error?.message) {
-          errorMessage = parsed.error.message;
-        }
-      } catch {}
-      return res.status(upstream.status).json({
-        error: { message: errorMessage, type: "upstream_error", code: upstream.status },
-      });
-    }
-
-    res.setHeader("Content-Type", upstream.headers.get("content-type") || "application/json");
-
-    if (req.body?.stream && upstream.body) {
-      const reader = upstream.body.getReader();
-      const decoder = new TextDecoder();
-      try {
-        while (true) {
-          const { done, value } = await reader.read();
-          if (done) break;
-          res.write(decoder.decode(value, { stream: true }));
-        }
-      } catch (e) { /* stream closed */ }
-      finally { res.end(); }
-    } else {
-      res.send(await upstream.text());
-    }
-  } catch (err) {
-    console.error(`Proxy error [${providerName}/${model}]:`, err.message);
-    res.status(502).json({ error: { message: `Upstream error: ${err.message}` } });
-  }
-});
-
-// =============================================================================
-// MODELS & HEALTH
-// =============================================================================
-
-const KNOWN_MODELS = [
-  "gemini-2.5-pro", "gemini-2.5-flash",
-  "gpt-4.1", "gpt-4.1-mini", "gpt-4o", "gpt-4o-mini",
-  "o3-mini", "o1-mini",
-];
-
-app.get("/models", (_, res) => {
-  res.json({ object: "list", data: KNOWN_MODELS.map(id => ({
-    id,
-    object: "model",
-    owned_by: "system",
-    providers: [{ supports_tools: true }],
-  })) });
-});
-
-app.get("/health", (_, res) => {
-  const backends = {};
-  for (const [name, client] of mcpBackends) {
-    backends[name] = { ready: client.ready, tools: client.tools.length };
-  }
-  const activeTools = getActiveTools();
-  const groups = {};
-  for (const [name, g] of Object.entries(TOOL_GROUPS)) {
-    groups[name] = { enabled: g.enabled, source: g.source };
-  }
-  res.json({
-    status: "ok", service: "mcp-bridge", version: "2.0.0",
-    tools: { builtin: BUILTIN_TOOLS.length, external: activeTools.length, total: BUILTIN_TOOLS.length + activeTools.length },
-    groups, backends,
-  });
-});
-
-// GET /groups — list tool groups and their status
-app.get("/groups", (_, res) => {
-  const activeTools = getActiveTools();
-  const result = {};
-  for (const [name, g] of Object.entries(TOOL_GROUPS)) {
-    const tools = name === "core" ? BUILTIN_TOOLS :
-      activeTools.filter(t => {
-        if (g.source !== t._backend) return false;
-        if (!g.prefixes) return true;
-        return g.prefixes.some(p => t._originalName.startsWith(p));
-      });
-    result[name] = {
-      enabled: g.enabled,
-      description: g.description,
-      tools: tools.length,
-      toolNames: tools.map(t => t.name).slice(0, 10),
-    };
-  }
-  res.json(result);
-});
-
-// =============================================================================
-// STARTUP
-// =============================================================================
-
-async function main() {
-  const isPublic = BIND_HOST !== "127.0.0.1" && BIND_HOST !== "localhost";
-  if (isPublic && !process.env.MCP_AUTH_TOKEN) {
-    console.error(
-      "FATAL: refusing to bind a public interface without MCP_AUTH_TOKEN. " +
-      "Generate one with: MCP_AUTH_TOKEN=$(openssl rand -base64 32)"
-    );
-    process.exit(1);
-  }
-  app.listen(PORT, BIND_HOST, () => {
-    console.log(`MCP Bridge v2.0.0 on port ${PORT} (${BIND_HOST})`);
-    const enabled = Object.entries(TOOL_GROUPS).filter(([, g]) => g.enabled).map(([n]) => n);
-    console.log(`Active groups: ${enabled.join(", ")}`);
-    // ADR-166 §6 — startup posture banner (helps operators see the security state at boot)
-    console.log(
-      `[security] bind=${BIND_HOST} auth=${process.env.MCP_AUTH_TOKEN ? "bearer" : "off (local-only)"} ` +
-      `terminal=${MCP_ENABLE_TERMINAL ? "ENABLED (⚠ opt-in)" : "disabled"}`,
-    );
-    if (MCP_ENABLE_TERMINAL) {
-      console.warn(
-        "[security] WARNING: terminal_execute is enabled. This tool grants shell access " +
-        "inside the bridge container to any client the auth layer accepts. Ensure " +
-        "MCP_AUTH_TOKEN is set on any non-loopback bind. See ADR-166 §6 Phase 1d.",
-      );
-    }
-  });
-
-  const anyBackendNeeded = BACKEND_DEFS.some(isBackendNeeded);
-  if (anyBackendNeeded) {
-    console.log("Initializing MCP backends...");
-    await initBackends();
-  }
-}
-
-main().catch(err => { console.error("Fatal:", err); process.exit(1); });
+import express from "express";
+import { spawn } from "child_process";
+import { randomUUID, timingSafeEqual } from "crypto";
+
+// =============================================================================
+// CONFIGURATION
+// =============================================================================
+
+const CLOUD_FUNCTIONS = {
+  search: process.env.SEARCH_API_URL || null,
+  research: process.env.RESEARCH_API_URL || null,
+};
+
+const PORT = parseInt(process.env.PORT || "3001", 10);
+const BIND_HOST = process.env.MCP_BIND_HOST || "127.0.0.1";
+
+// =============================================================================
+// TOOL GROUPS — Enable/disable categories of tools independently
+// =============================================================================
+// Groups map tool name prefixes from backends to logical categories.
+// Each group can be toggled via env var. The AI sees only enabled tools.
+
+const TOOL_GROUPS = {
+  // --- Core (always on, built-in) ---
+  core: {
+    enabled: true, // cannot be disabled
+    description: "Search, research, and guidance tools",
+    source: "builtin",
+  },
+
+  // --- Intelligence (ruvector) ---
+  intelligence: {
+    enabled: process.env.MCP_GROUP_INTELLIGENCE !== "false",
+    description: "Self-learning intelligence — routing, memory, pattern training (ruvector)",
+    source: "ruvector",
+    prefixes: ["hooks_"],
+  },
+
+  // --- Agents & Orchestration (ruflo) ---
+  agents: {
+    enabled: process.env.MCP_GROUP_AGENTS !== "false",
+    description: "Agent lifecycle, swarm coordination, task management, workflows (ruflo)",
+    source: "ruflo",
+    prefixes: ["agent_", "swarm_", "task_", "session_", "hive-mind_", "workflow_", "coordination_"],
+  },
+
+  // --- Memory & Knowledge (ruflo) ---
+  memory: {
+    enabled: process.env.MCP_GROUP_MEMORY !== "false",
+    description: "Vector memory, AgentDB, embeddings, semantic search (ruflo)",
+    source: "ruflo",
+    prefixes: ["memory_", "agentdb_", "embeddings_"],
+  },
+
+  // --- Dev Tools (ruflo) ---
+  devtools: {
+    enabled: process.env.MCP_GROUP_DEVTOOLS !== "false",
+    description: "Hooks, code analysis, performance profiling, GitHub integration (ruflo)",
+    source: "ruflo",
+    prefixes: ["hooks_", "analyze_", "performance_", "github_", "terminal_", "config_", "system_", "progress_"],
+  },
+
+  // --- Security & Safety (ruflo) ---
+  security: {
+    enabled: process.env.MCP_GROUP_SECURITY === "true",
+    description: "AI defence, PII detection, claims management, pattern transfer (ruflo)",
+    source: "ruflo",
+    prefixes: ["aidefence_", "claims_", "transfer_"],
+  },
+
+  // --- Browser Automation (ruflo) ---
+  browser: {
+    enabled: process.env.MCP_GROUP_BROWSER === "true",
+    description: "Headless browser control — navigate, click, fill, screenshot (ruflo)",
+    source: "ruflo",
+    prefixes: ["browser_"],
+  },
+
+  // --- Neural & DAA (ruflo) ---
+  neural: {
+    enabled: process.env.MCP_GROUP_NEURAL === "true",
+    description: "Neural network training, DAA autonomous agents, cognitive patterns (ruflo)",
+    source: "ruflo",
+    prefixes: ["neural_", "daa_"],
+  },
+
+  // --- Agentic Flow (agentic-flow@alpha) ---
+  "agentic-flow": {
+    enabled: process.env.MCP_GROUP_AGENTIC_FLOW === "true",
+    description: "Execute 66+ specialized agents, batch code editing, AgentDB patterns (agentic-flow)",
+    source: "agentic-flow",
+    prefixes: ["agentic_flow_", "agent_booster_", "agentdb_"],
+  },
+
+  // --- Claude Code ---
+  "claude-code": {
+    enabled: process.env.MCP_GROUP_CLAUDE_CODE === "true",
+    description: "Anthropic Claude Code — file editing, bash execution, code analysis (requires ANTHROPIC_API_KEY)",
+    source: "claude",
+  },
+
+  // --- Gemini MCP ---
+  gemini: {
+    enabled: process.env.MCP_GROUP_GEMINI === "true",
+    description: "Google Gemini conversation context, multimodal capabilities (requires GOOGLE_API_KEY)",
+    source: "gemini-mcp",
+  },
+
+  // --- OpenAI Codex ---
+  codex: {
+    enabled: process.env.MCP_GROUP_CODEX === "true",
+    description: "OpenAI Codex coding agent — code generation and execution (requires OPENAI_API_KEY)",
+    source: "codex",
+  },
+};
+
+// =============================================================================
+// STDIO MCP CLIENT — Connects to external MCP servers via child process
+// =============================================================================
+
+class StdioMcpClient {
+  constructor(name, command, args = []) {
+    this.name = name;
+    this.command = command;
+    this.args = args;
+    this.process = null;
+    this.tools = [];
+    this.ready = false;
+    this.pending = new Map();
+    this.buffer = "";
+  }
+
+  async start() {
+    return new Promise((resolve) => {
+      try {
+        this.process = spawn(this.command, this.args, {
+          stdio: ["pipe", "pipe", "pipe"],
+          env: { ...process.env },
+        });
+
+        this.process.stdout.on("data", (data) => this._onData(data.toString()));
+        this.process.stderr.on("data", (data) => {
+          const msg = data.toString().trim();
+          if (msg && !msg.startsWith("npm WARN")) console.error(`[${this.name}] ${msg}`);
+        });
+        this.process.on("error", (err) => {
+          console.error(`[${this.name}] spawn error:`, err.message);
+          this.ready = false;
+          resolve(false);
+        });
+        this.process.on("exit", (code) => {
+          console.log(`[${this.name}] exited with code ${code}`);
+          this.ready = false;
+        });
+
+        this._send("initialize", {
+          protocolVersion: "2024-11-05",
+          capabilities: {},
+          clientInfo: { name: "mcp-bridge", version: "2.0.0" },
+        }).then((result) => {
+          if (result && !result.error) {
+            this._notify("notifications/initialized", {});
+            return this._send("tools/list", {});
+          }
+          return null;
+        }).then((result) => {
+          if (result && result.tools) {
+            this.tools = result.tools.map(t => ({
+              ...t,
+              _originalName: t.name,
+              _backend: this.name,
+            }));
+            this.ready = true;
+            console.log(`[${this.name}] ${this.tools.length} tools loaded`);
+          }
+          resolve(this.ready);
+        }).catch((err) => {
+          console.error(`[${this.name}] init failed:`, err.message);
+          resolve(false);
+        });
+
+        setTimeout(() => { if (!this.ready) resolve(false); }, 60000);
+      } catch (err) {
+        console.error(`[${this.name}] failed to start:`, err.message);
+        resolve(false);
+      }
+    });
+  }
+
+  _onData(chunk) {
+    this.buffer += chunk;
+    const lines = this.buffer.split("\n");
+    this.buffer = lines.pop() || "";
+    for (const line of lines) {
+      const trimmed = line.trim();
+      if (!trimmed) continue;
+      try {
+        const msg = JSON.parse(trimmed);
+        if (msg.id && this.pending.has(msg.id)) {
+          const { resolve } = this.pending.get(msg.id);
+          this.pending.delete(msg.id);
+          resolve(msg.result || msg.error || {});
+        }
+      } catch { /* skip non-JSON */ }
+    }
+  }
+
+  _send(method, params) {
+    return new Promise((resolve, reject) => {
+      if (!this.process || this.process.killed) {
+        return reject(new Error(`${this.name} process not running`));
+      }
+      const id = randomUUID();
+      const msg = JSON.stringify({ jsonrpc: "2.0", id, method, params }) + "\n";
+      this.pending.set(id, { resolve, reject });
+      this.process.stdin.write(msg);
+      // initialize is the cold-start gate for backends like ruflo/ruvector
+      // which boot a full claude-flow / ruvector kernel — on Cloud Run with
+      // npx fetching artifacts it can take 45-60s. Other RPC methods are
+      // post-init and stay snappy.
+      const timeoutMs = method === "initialize" ? 120000 : 30000;
+      setTimeout(() => {
+        if (this.pending.has(id)) {
+          this.pending.delete(id);
+          reject(new Error(`${this.name} timeout for ${method}`));
+        }
+      }, timeoutMs);
+    });
+  }
+
+  _notify(method, params) {
+    if (!this.process || this.process.killed) return;
+    this.process.stdin.write(JSON.stringify({ jsonrpc: "2.0", method, params }) + "\n");
+  }
+
+  async callTool(originalName, args) {
+    if (!this.ready) return { error: `${this.name} backend not available` };
+    try {
+      return await this._send("tools/call", { name: originalName, arguments: args });
+    } catch (err) {
+      return { error: err.message };
+    }
+  }
+
+  stop() {
+    if (this.process && !this.process.killed) {
+      this.process.kill("SIGTERM");
+      this.process = null;
+    }
+    this.ready = false;
+    this.tools = [];
+  }
+}
+
+// =============================================================================
+// BACKEND REGISTRY
+// =============================================================================
+
+const BACKEND_DEFS = [
+  { name: "ruvector",       command: "npx", args: ["-y", "ruvector", "mcp", "start"],   groups: ["intelligence"] },
+  { name: "ruflo",          command: "npx", args: ["-y", "ruflo", "mcp", "start"],      groups: ["agents", "memory", "devtools", "security", "browser", "neural"] },
+  { name: "agentic-flow",   command: "npx", args: ["-y", "agentic-flow@alpha", "mcp", "start"], groups: ["agentic-flow"] },
+  { name: "claude",         command: "claude", args: ["mcp", "serve"],                  groups: ["claude-code"] },
+  { name: "gemini-mcp",     command: "npx", args: ["-y", "gemini-mcp-server"],          groups: ["gemini"] },
+  { name: "codex",          command: "npx", args: ["-y", "@openai/codex", "mcp-server"], groups: ["codex"] },
+];
+
+const mcpBackends = new Map();
+let allBackendTools = []; // all tools from all backends (pre-filter)
+
+function isBackendNeeded(backendDef) {
+  return backendDef.groups.some(g => TOOL_GROUPS[g]?.enabled);
+}
+
+// Filter tools from a backend based on which groups are enabled
+function filterToolsByGroups(tools, backendName) {
+  const enabledGroups = Object.entries(TOOL_GROUPS)
+    .filter(([, g]) => g.enabled && g.source === backendName);
+
+  if (enabledGroups.length === 0) return [];
+
+  // If any enabled group has no prefixes defined, include all tools from that backend
+  const hasWildcard = enabledGroups.some(([, g]) => !g.prefixes);
+  if (hasWildcard) return tools;
+
+  const enabledPrefixes = enabledGroups.flatMap(([, g]) => g.prefixes || []);
+  return tools.filter(t => enabledPrefixes.some(p => t._originalName.startsWith(p)));
+}
+
+// Get the final filtered tool list with namespaced names
+function getActiveTools() {
+  const filtered = [];
+  for (const [backendName, client] of mcpBackends) {
+    const accepted = filterToolsByGroups(client.tools, backendName);
+    for (const t of accepted) {
+      filtered.push({ ...t, name: `${backendName}__${t._originalName}` });
+    }
+  }
+  return filtered;
+}
+
+async function initBackends() {
+  const needed = BACKEND_DEFS.filter(isBackendNeeded);
+  if (needed.length === 0) return;
+
+  console.log(`Starting ${needed.length} MCP backends: ${needed.map(b => b.name).join(", ")}`);
+
+  await Promise.allSettled(
+    needed.map(async (b) => {
+      const client = new StdioMcpClient(b.name, b.command, b.args);
+      const ok = await client.start();
+      if (ok) {
+        mcpBackends.set(b.name, client);
+      } else {
+        console.warn(`[${b.name}] failed to start`);
+      }
+    })
+  );
+
+  allBackendTools = getActiveTools();
+  console.log(`MCP backends: ${mcpBackends.size} active, ${allBackendTools.length} tools (filtered by groups)`);
+}
+
+process.on("SIGTERM", () => { for (const [, c] of mcpBackends) c.stop(); process.exit(0); });
+process.on("SIGINT", () => { for (const [, c] of mcpBackends) c.stop(); process.exit(0); });
+
+// =============================================================================
+// BUILT-IN TOOLS (core group — always on)
+// =============================================================================
+
+const BUILTIN_TOOLS = [
+  {
+    name: "search",
+    description: "Search your knowledge base for relevant information.",
+    inputSchema: {
+      type: "object",
+      properties: {
+        query: { type: "string", description: "Natural language search query" },
+        limit: { type: "number", description: "Max results (default 5)", default: 5 },
+      },
+      required: ["query"],
+    },
+  },
+  {
+    name: "web_research",
+    description: "Search the web, fact-check claims, compare items, or conduct deep research. Actions: 'search' (quick), 'research' (deep report), 'compare' (side-by-side), 'fact_check' (verify claims), 'goap' (comprehensive multi-step research with verification).",
+    inputSchema: {
+      type: "object",
+      properties: {
+        action: { type: "string", enum: ["search", "research", "compare", "fact_check", "goap"], description: "Research action type", default: "search" },
+        query: { type: "string", description: "Search query or topic" },
+        items: { type: "array", items: { type: "string" }, description: "Items to compare (for 'compare')" },
+        claim: { type: "string", description: "Claim to verify (for 'fact_check')" },
+        verify: { type: "boolean", description: "Verify results in goap mode", default: true },
+      },
+      required: ["query"],
+    },
+  },
+  {
+    name: "guidance",
+    description: "Get instructions on how to use the available tool groups and services. Call this FIRST when unsure which tool to use, when a user asks 'what can you do?', or when you need to understand a specific tool group. Returns structured guidance for the AI on tool selection and usage patterns.",
+    inputSchema: {
+      type: "object",
+      properties: {
+        topic: {
+          type: "string",
+          enum: ["overview", "groups", "intelligence", "agents", "memory", "devtools", "security", "browser", "neural", "agentic-flow", "claude-code", "gemini", "codex", "tool"],
+          description: "What to get guidance on. Use 'overview' for capabilities summary, 'groups' to see all tool groups and their status, or a specific group name for detailed usage instructions.",
+          default: "overview",
+        },
+        tool_name: { type: "string", description: "Specific tool name to get detailed usage info (when topic='tool')" },
+      },
+    },
+  },
+];
+
+// =============================================================================
+// GUIDANCE ENGINE — AI-facing instruction system
+// =============================================================================
+
+function getGuidance(topic, toolName) {
+  const activeGroups = Object.entries(TOOL_GROUPS).filter(([, g]) => g.enabled);
+  const inactiveGroups = Object.entries(TOOL_GROUPS).filter(([, g]) => !g.enabled);
+  const externalTools = getActiveTools();
+
+  if (topic === "overview") {
+    return {
+      guidance: `# Tool Capabilities Overview
+
+You have access to ${BUILTIN_TOOLS.length + externalTools.length} tools organized into ${activeGroups.length} active groups.
+
+## Active Groups
+${activeGroups.map(([name, g]) => {
+  const count = name === "core" ? BUILTIN_TOOLS.length : externalTools.filter(t => t.name.startsWith(g.source + "__")).length;
+  return `- **${name}** (${count} tools) — ${g.description}`;
+}).join("\n")}
+
+## Inactive Groups (can be enabled)
+${inactiveGroups.map(([name, g]) => `- **${name}** — ${g.description}`).join("\n") || "None"}
+
+## Quick Decision Guide
+- **Knowledge questions** → use \`search\` first, then \`web_research\` if needed
+- **Current events / facts** → use \`web_research\` with action 'search' or 'goap'
+- **Complex research** → use \`web_research\` with action 'goap' (multi-step pipeline)
+- **"What can you do?"** → call \`guidance\` with topic 'groups'
+- **Memory / recall** → use tools from the \`memory\` group
+- **Agent orchestration** → use tools from the \`agents\` group
+- **Code analysis / performance** → use tools from the \`devtools\` group
+
+## Rules
+1. Call tools FIRST, then present results conversationally
+2. Never show raw JSON — synthesize results naturally
+3. For complex questions, prefer GOAP pipeline (web_research action='goap')
+4. Call \`guidance\` with a specific group name to learn how to use that group's tools`,
+      topic: "overview",
+    };
+  }
+
+  if (topic === "groups") {
+    const groupList = Object.entries(TOOL_GROUPS).map(([name, g]) => {
+      const status = g.enabled ? "ACTIVE" : "INACTIVE";
+      const toolCount = name === "core" ? BUILTIN_TOOLS.length :
+        externalTools.filter(t => {
+          const backend = t._backend;
+          return g.source === backend && (!g.prefixes || g.prefixes.some(p => t._originalName.startsWith(p)));
+        }).length;
+      return `| ${name} | ${status} | ${toolCount} | ${g.description} |`;
+    });
+
+    return {
+      guidance: `# Tool Groups\n\n| Group | Status | Tools | Description |\n|-------|--------|-------|-------------|\n${groupList.join("\n")}`,
+      topic: "groups",
+    };
+  }
+
+  // Specific group guidance
+  const groupGuides = {
+    intelligence: `# Intelligence Group (ruvector)
+
+Self-learning intelligence tools for routing and vector memory.
+
+## Key Tools
+- **ruvector__hooks_route** — Route a task to the best agent type. Call with a task description.
+- **ruvector__hooks_remember** — Store context/knowledge in vector memory for later recall.
+- **ruvector__hooks_recall** — Search vector memory semantically. Good for finding past context.
+- **ruvector__hooks_pretrain** — Bootstrap intelligence from a code repository.
+- **ruvector__hooks_build_agents** — Generate optimized agent configurations.
+- **ruvector__hooks_stats** — Get intelligence statistics and learning metrics.
+
+## When to Use
+- Before starting complex tasks: route to find the best agent approach
+- To store important findings for cross-session memory
+- To recall previously stored patterns or solutions`,
+
+    agents: `# Agents & Orchestration Group (ruflo)
+
+Multi-agent lifecycle management, swarm coordination, and task workflows.
+
+## Key Tools
+- **ruflo__agent_spawn** — Create a new agent with specific capabilities
+- **ruflo__agent_list** — List all active agents
+- **ruflo__swarm_init** — Initialize a swarm with a topology (mesh, hierarchical, ring, star)
+- **ruflo__task_create** — Create and assign tasks
+- **ruflo__workflow_create** — Define multi-step workflows
+- **ruflo__workflow_execute** — Execute a workflow
+- **ruflo__hive-mind_init** — Start collective intelligence coordination
+- **ruflo__coordination_orchestrate** — Multi-agent coordination
+
+## When to Use
+- Complex tasks requiring multiple agents working together
+- Pipeline workflows with sequential or parallel steps
+- Distributed task management`,
+
+    memory: `# Memory & Knowledge Group (ruflo)
+
+Vector storage, semantic search, AgentDB pattern learning, and embeddings.
+
+## Key Tools
+- **ruflo__memory_store** — Store a value with vector embedding for semantic search
+- **ruflo__memory_search** — Semantic search across stored memories (HNSW-indexed)
+- **ruflo__memory_list** — List stored memory entries
+- **ruflo__agentdb_pattern-store** — Store a reasoning pattern for learning
+- **ruflo__agentdb_pattern-search** — Search for similar reasoning patterns
+- **ruflo__agentdb_context-synthesize** — Synthesize context from stored memories
+- **ruflo__embeddings_generate** — Generate vector embeddings for text
+- **ruflo__embeddings_search** — Semantic similarity search
+
+## When to Use
+- Persistent knowledge storage across sessions
+- Finding similar past solutions or patterns
+- Building semantic search over custom data`,
+
+    devtools: `# Dev Tools Group (ruflo)
+
+Code analysis, performance profiling, GitHub integration, and terminal access.
+
+## Key Tools
+- **ruflo__analyze_diff** — Analyze git diff for risk and change classification
+- **ruflo__performance_benchmark** — Run performance benchmarks
+- **ruflo__performance_bottleneck** — Detect performance bottlenecks
+- **ruflo__github_repo_analyze** — Analyze a GitHub repository
+- **ruflo__github_pr_manage** — Manage pull requests
+- **ruflo__terminal_execute** — Execute commands in a terminal session
+
+## When to Use
+- Code review and change risk assessment
+- Performance analysis and optimization
+- GitHub repository management`,
+
+    security: `# Security & Safety Group (ruflo)
+
+AI defence, PII detection, and claims-based authorization.
+
+## Key Tools
+- **ruflo__aidefence_scan** — Scan text for AI manipulation attempts
+- **ruflo__aidefence_has_pii** — Check for PII (emails, phones, SSNs)
+- **ruflo__aidefence_is_safe** — Quick safety check on input
+- **ruflo__claims_claim** — Claim an issue for work
+- **ruflo__claims_board** — Visual board of all claims
+
+## When to Use
+- Input validation and safety checking
+- PII detection before processing sensitive data
+- Work item management across agents`,
+
+    browser: `# Browser Automation Group (ruflo)
+
+Headless browser control for web interaction and testing.
+
+## Key Tools
+- **ruflo__browser_open** — Navigate to a URL
+- **ruflo__browser_click** — Click elements by reference
+- **ruflo__browser_fill** — Fill form inputs
+- **ruflo__browser_screenshot** — Capture page screenshots
+- **ruflo__browser_snapshot** — Get accessibility tree for AI parsing
+- **ruflo__browser_eval** — Execute JavaScript in page context
+
+## When to Use
+- Web scraping and data extraction
+- Automated testing (E2E)
+- Form filling and web interaction`,
+
+    neural: `# Neural & DAA Group (ruflo)
+
+Neural network operations and Decentralized Autonomous Agents.
+
+## Key Tools
+- **ruflo__neural_train** — Train a neural model
+- **ruflo__neural_predict** — Make predictions
+- **ruflo__daa_agent_create** — Create an autonomous agent
+- **ruflo__daa_workflow_create** — Create autonomous workflows
+- **ruflo__daa_knowledge_share** — Share knowledge between agents
+
+## When to Use
+- Pattern learning and prediction
+- Autonomous agent workflows
+- Knowledge transfer between agents`,
+
+    "agentic-flow": `# Agentic Flow Group (agentic-flow@alpha)
+
+Execute 66+ specialized agents with boosted code editing and AgentDB.
+
+## Key Tools
+- **agentic-flow__agentic_flow_agent** — Execute any of 66+ specialized agents
+- **agentic-flow__agentic_flow_list_agents** — List available agent types
+- **agentic-flow__agent_booster_edit_file** — 352x faster code editing
+- **agentic-flow__agent_booster_batch_edit** — Multi-file refactoring
+- **agentic-flow__agentdb_pattern_store** — Store reasoning patterns
+- **agentic-flow__agentdb_pattern_search** — Search similar patterns
+
+## When to Use
+- Complex code generation with specialized agents
+- Batch code refactoring across files
+- Agent selection when you need the right specialist`,
+
+    "claude-code": `# Claude Code Group
+
+Anthropic Claude Code MCP server — full coding agent capabilities.
+
+Requires: ANTHROPIC_API_KEY environment variable.
+
+## Capabilities
+- File reading and editing
+- Bash command execution
+- Code analysis and generation
+- Project exploration
+
+## When to Use
+- When you need a second AI perspective on code
+- Complex refactoring tasks
+- Code review and analysis`,
+
+    gemini: `# Gemini MCP Group
+
+Google Gemini with conversation context management.
+
+Requires: GOOGLE_API_KEY environment variable (already set for Gemini models).
+
+## Capabilities
+- Conversation context management
+- Multimodal processing
+- Google Search grounding
+
+## When to Use
+- Extended context conversations
+- Multimodal content processing`,
+
+    codex: `# Codex Group
+
+OpenAI Codex coding agent.
+
+Requires: OPENAI_API_KEY environment variable (already set for OpenAI models).
+
+## Capabilities
+- Code generation and execution
+- Code completion
+- Code explanation
+
+## When to Use
+- Code generation tasks
+- Quick code completions
+- Code explanation and documentation`,
+  };
+
+  if (topic === "tool" && toolName) {
+    const allTools = [...BUILTIN_TOOLS, ...externalTools];
+    const tool = allTools.find(t => t.name === toolName);
+    if (tool) {
+      const props = Object.entries(tool.inputSchema?.properties || {})
+        .map(([k, v]) => `- **${k}** (${v.type}) — ${v.description || ""}`)
+        .join("\n");
+      return { guidance: `# ${tool.name}\n\n${tool.description}\n\n## Parameters\n${props}`, topic: "tool" };
+    }
+    return { guidance: `Tool '${toolName}' not found. Call guidance with topic='groups' to see available tools.`, topic: "tool" };
+  }
+
+  if (groupGuides[topic]) {
+    const group = TOOL_GROUPS[topic];
+    if (!group?.enabled) {
+      return { guidance: `# ${topic} — INACTIVE\n\n${group?.description || ""}\n\nThis group is not enabled. Set the appropriate MCP_GROUP_* env var to "true" to activate it.`, topic };
+    }
+    return { guidance: groupGuides[topic], topic };
+  }
+
+  return { guidance: `Unknown topic '${topic}'. Use 'overview', 'groups', or a specific group name.`, topic };
+}
+
+// =============================================================================
+// GEMINI GROUNDED SEARCH — Uses Gemini's built-in Google Search when no
+// dedicated search Cloud Function is configured
+// =============================================================================
+
+async function geminiGroundedSearch(query, mode = "search") {
+  const apiKey = process.env.GOOGLE_API_KEY;
+  if (!apiKey) return { error: "No GOOGLE_API_KEY configured for search" };
+
+  // Empty/missing query produces a 400 from Gemini's generateContent endpoint.
+  // Return a structured error so the model can recover with a real query.
+  if (!query || typeof query !== "string" || !query.trim()) {
+    return {
+      error: "search requires a non-empty query string",
+      hint: "Call this tool again with { query: 'your search terms' }. For comparisons use { action: 'compare', query: 'item A vs item B' }; for fact-checking use { action: 'fact_check', claim: 'the claim text' }.",
+    };
+  }
+
+  const model = "gemini-2.5-flash";
+  const url = `https://generativelanguage.googleapis.com/v1beta/models/${model}:generateContent?key=${apiKey}`;
+
+  const systemInstructions = {
+    search: `You are a helpful search assistant. Answer the query using Google Search results. Include key facts, sources, and relevant details. Be concise but thorough.`,
+    research: `You are a research analyst. Provide a comprehensive research report on the topic using Google Search results. Include: key findings, analysis, multiple perspectives, and source citations. Be thorough.`,
+    compare: `You are a comparison analyst. Compare the items using Google Search results. Create a structured comparison with pros/cons, key differences, and a recommendation.`,
+    fact_check: `You are a fact-checker. Verify the claim using Google Search results. Provide a verdict (TRUE/FALSE/PARTIALLY TRUE/UNVERIFIABLE), evidence, and sources.`,
+  };
+
+  const prompt = mode === "fact_check"
+    ? `Fact-check this claim: "${query}"`
+    : mode === "compare"
+      ? `Compare these items: ${query}`
+      : mode === "research"
+        ? `Research this topic thoroughly: ${query}`
+        : query;
+
+  try {
+    const resp = await fetch(url, {
+      method: "POST",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify({
+        system_instruction: { parts: [{ text: systemInstructions[mode] || systemInstructions.search }] },
+        contents: [{ parts: [{ text: prompt }] }],
+        tools: [{ google_search: {} }],
+        generationConfig: { temperature: 0.2 },
+      }),
+      signal: AbortSignal.timeout(30000),
+    });
+
+    if (!resp.ok) {
+      const errText = await resp.text();
+      console.error(`[gemini-search] API error ${resp.status}:`, errText.substring(0, 200));
+      return { error: `Search API error: ${resp.status}` };
+    }
+
+    const data = await resp.json();
+    const candidate = data.candidates?.[0];
+    const answer = candidate?.content?.parts?.map(p => p.text).filter(Boolean).join("\n") || "";
+
+    // Extract grounding metadata (sources)
+    const grounding = candidate?.groundingMetadata || {};
+    const chunks = grounding.groundingChunks || [];
+    const sources = chunks
+      .filter(c => c.web)
+      .map(c => ({ title: c.web.title || "", url: c.web.uri || "" }))
+      .filter(s => s.url);
+
+    // Extract search queries used
+    const searchQueries = (grounding.webSearchQueries || []);
+
+    return {
+      success: true,
+      answer,
+      sources: sources.slice(0, 8),
+      searchQueries,
+      groundingMetadata: { sources, searchQueries },
+      mode,
+    };
+  } catch (err) {
+    if (err.name === "AbortError" || err.name === "TimeoutError") return { error: "Search timed out" };
+    return { error: err.message };
+  }
+}
+
+// =============================================================================
+// SSRF GUARD — Reject requests to private/loopback ranges (CWE-918)
+// =============================================================================
+
+const PRIVATE_IP_RE = /^(?:10\.|172\.(?:1[6-9]|2\d|3[01])\.|192\.168\.|127\.|0\.|::1|fc|fd)/i;
+
+function assertSafeUrl(rawUrl) {
+  let parsed;
+  try {
+    parsed = new URL(rawUrl);
+  } catch {
+    throw new Error(`SSRF guard: invalid URL — ${rawUrl}`);
+  }
+  if (parsed.protocol !== "https:") {
+    throw new Error(`SSRF guard: only HTTPS URLs are permitted, got ${parsed.protocol}`);
+  }
+  const host = parsed.hostname;
+  if (PRIVATE_IP_RE.test(host) || host === "localhost" || host.endsWith(".local")) {
+    throw new Error(`SSRF guard: private/loopback host rejected — ${host}`);
+  }
+}
+
+// =============================================================================
+// HELPER — Call a backend Cloud Function / API
+// =============================================================================
+
+async function callCloudFunction(url, payload, timeoutMs = 25000) {
+  // Validate the URL before making any network request.
+  assertSafeUrl(url);
+  const controller = new AbortController();
+  const timer = setTimeout(() => controller.abort(), timeoutMs);
+  try {
+    const resp = await fetch(url, {
+      method: "POST",
+      headers: { "Content-Type": "application/json" },
+      body: JSON.stringify(payload),
+      signal: controller.signal,
+    });
+    return await resp.json();
+  } catch (err) {
+    if (err.name === "AbortError") return { error: "Request timed out", timeout: timeoutMs };
+    return { error: err.message };
+  } finally {
+    clearTimeout(timer);
+  }
+}
+
+// =============================================================================
+// GOAP SEARCH PIPELINE
+// =============================================================================
+
+async function executeGoapSearch(query, args) {
+  const researchUrl = CLOUD_FUNCTIONS.research;
+  if (!researchUrl) return { error: "GOAP requires a 'research' URL in CLOUD_FUNCTIONS" };
+
+  const startTime = Date.now();
+
+  const composeResult = await callCloudFunction(researchUrl, {
+    action: "search",
+    query: `Break this question into 3-4 distinct search queries that would help answer it comprehensively. Return ONLY the queries, one per line:\n\n${query}`,
+  }, 30000);
+
+  let searchQueries = [query];
+  if (composeResult && !composeResult.error) {
+    const answer = composeResult.result?.answer || composeResult.answer || "";
+    const lines = answer.split("\n").map(l => l.replace(/^[\d\-\*\.\)]+\s*/, "").trim()).filter(l => l.length > 5 && l.length < 200);
+    if (lines.length >= 2) searchQueries = lines.slice(0, 4);
+  }
+
+  const searchResults = await Promise.all(
+    searchQueries.map(q => callCloudFunction(researchUrl, { action: "search", query: q }, 30000))
+  );
+
+  const allSources = [], allAnswers = [];
+  for (let i = 0; i < searchResults.length; i++) {
+    const r = searchResults[i];
+    if (r && !r.error && r.success !== false) {
+      const answer = r.result?.answer || r.answer || "";
+      if (answer) allAnswers.push({ query: searchQueries[i], answer });
+      const gm = r.result?.groundingMetadata || r.groundingMetadata || {};
+      if (gm.sources) allSources.push(...gm.sources);
+    }
+  }
+
+  const uniqueSources = [];
+  const seenUrls = new Set();
+  for (const src of allSources) {
+    const url = src.url || src.uri || "";
+    if (url && !seenUrls.has(url)) { seenUrls.add(url); uniqueSources.push(src); }
+  }
+
+  const synthesisInput = allAnswers.map(a => `## ${a.query}\n${a.answer}`).join("\n\n");
+  const synthesisResult = await callCloudFunction(researchUrl, {
+    action: "research",
+    topic: `Synthesize these findings into a comprehensive answer to: "${query}"\n\nFindings:\n${synthesisInput}`,
+  }, 60000);
+
+  const synthesizedAnswer = synthesisResult?.result?.answer || synthesisResult?.answer || synthesisInput;
+  const synthGm = synthesisResult?.result?.groundingMetadata || {};
+  if (synthGm.sources) {
+    for (const src of synthGm.sources) {
+      const url = src.url || src.uri || "";
+      if (url && !seenUrls.has(url)) { seenUrls.add(url); uniqueSources.push(src); }
+    }
+  }
+
+  let verification = { verified: true, confidence: "high" };
+  if (args.verify !== false && synthesizedAnswer.length > 100) {
+    const vr = await callCloudFunction(researchUrl, {
+      action: "fact_check", claim: synthesizedAnswer.substring(0, 500),
+    }, 30000);
+    if (vr && !vr.error && vr.result) {
+      verification = {
+        verified: vr.result.verdict !== "FALSE",
+        verdict: vr.result.verdict,
+        confidence: vr.result.confidence || "medium",
+        details: vr.result.analysis,
+      };
+    }
+  }
+
+  return {
+    answer: synthesizedAnswer, pipeline: "goap",
+    steps: { queries_composed: searchQueries.length, searches_executed: searchResults.filter(r => !r?.error).length, sources_found: uniqueSources.length, verification },
+    sources: uniqueSources.slice(0, 10), searchQueries, duration_ms: Date.now() - startTime,
+  };
+}
+
+// =============================================================================
+// GOAP SEARCH PIPELINE — Gemini fallback (no Cloud Function needed)
+// =============================================================================
+
+async function executeGoapSearchGemini(query, args) {
+  const startTime = Date.now();
+
+  // Step 1: Decompose into sub-queries
+  const decompose = await geminiGroundedSearch(
+    `Break this question into 3-4 distinct search queries that would help answer it comprehensively. Return ONLY the queries, one per line:\n\n${query}`,
+    "search"
+  );
+
+  let searchQueries = [query];
+  if (decompose?.answer) {
+    const lines = decompose.answer.split("\n")
+      .map(l => l.replace(/^[\d\-\*\.\)]+\s*/, "").trim())
+      .filter(l => l.length > 5 && l.length < 200);
+    if (lines.length >= 2) searchQueries = lines.slice(0, 4);
+  }
+
+  // Step 2: Parallel searches
+  const searchResults = await Promise.all(
+    searchQueries.map(q => geminiGroundedSearch(q, "search"))
+  );
+
+  const allSources = [], allAnswers = [];
+  for (let i = 0; i < searchResults.length; i++) {
+    const r = searchResults[i];
+    if (r && !r.error && r.answer) {
+      allAnswers.push({ query: searchQueries[i], answer: r.answer });
+      if (r.sources) allSources.push(...r.sources);
+    }
+  }
+
+  // Dedupe sources
+  const seenUrls = new Set();
+  const uniqueSources = allSources.filter(s => {
+    if (seenUrls.has(s.url)) return false;
+    seenUrls.add(s.url);
+    return true;
+  });
+
+  // Step 3: Synthesize
+  const synthesisInput = allAnswers.map(a => `## ${a.query}\n${a.answer}`).join("\n\n");
+  const synthesis = await geminiGroundedSearch(
+    `Synthesize these findings into a comprehensive answer to: "${query}"\n\nFindings:\n${synthesisInput}`,
+    "research"
+  );
+
+  const finalAnswer = synthesis?.answer || synthesisInput;
+  if (synthesis?.sources) {
+    for (const s of synthesis.sources) {
+      if (!seenUrls.has(s.url)) { seenUrls.add(s.url); uniqueSources.push(s); }
+    }
+  }
+
+  // Step 4: Verify if requested
+  let verification = { verified: true, confidence: "high" };
+  if (args.verify !== false && finalAnswer.length > 100) {
+    const vr = await geminiGroundedSearch(finalAnswer.substring(0, 500), "fact_check");
+    if (vr && !vr.error) {
+      verification = { verified: true, confidence: "medium", details: vr.answer };
+    }
+  }
+
+  return {
+    success: true,
+    answer: finalAnswer,
+    pipeline: "goap-gemini",
+    steps: {
+      queries_composed: searchQueries.length,
+      searches_executed: searchResults.filter(r => !r?.error).length,
+      sources_found: uniqueSources.length,
+      verification,
+    },
+    sources: uniqueSources.slice(0, 10),
+    searchQueries,
+    duration_ms: Date.now() - startTime,
+  };
+}
+
+// =============================================================================
+// TOOL EXECUTOR
+// =============================================================================
+
+// ADR-166 §6 Phase 1d + 2a — server-side tool gate.
+// Enforced HERE (not just in the autopilot handler) so /mcp, /mcp/:group,
+// autopilot, and any future path share ONE denial gate. Was the missing
+// link that made the disclosed unauthenticated-RCE chain reach shell.
+const DANGEROUS_TOOLS = Object.freeze(new Set([
+  "terminal_execute",
+  "ruflo__terminal_execute",
+  "devtools__terminal_execute",
+]));
+function isTerminalTool(name) {
+  return DANGEROUS_TOOLS.has(name) || /terminal_execute/i.test(name);
+}
+const MCP_ENABLE_TERMINAL = process.env.MCP_ENABLE_TERMINAL === "true";
+
+async function executeTool(name, args) {
+  // Deny dangerous tools unless the operator explicitly opted in.
+  // Enforced on every path (not just autopilot) — root cause of ADR-166 V2/V3.
+  if (isTerminalTool(name) && !MCP_ENABLE_TERMINAL) {
+    return {
+      error:
+        `Tool "${name}" is disabled by default. Set MCP_ENABLE_TERMINAL=true to allow.`,
+      code: "TOOL_DISABLED",
+    };
+  }
+  switch (name) {
+    case "search": {
+      if (CLOUD_FUNCTIONS.search) {
+        return callCloudFunction(CLOUD_FUNCTIONS.search, { query: args.query, limit: args.limit || 5 });
+      }
+      // Fallback: use Gemini grounded search
+      return geminiGroundedSearch(args.query, "search");
+    }
+
+    case "web_research": {
+      const action = args.action || "search";
+      if (action === "goap") {
+        // GOAP needs research endpoint — fall back to multi-search via Gemini
+        if (CLOUD_FUNCTIONS.research) return executeGoapSearch(args.query, args);
+        return executeGoapSearchGemini(args.query, args);
+      }
+      if (CLOUD_FUNCTIONS.research) {
+        const payload = { action };
+        if (action === "search") payload.query = args.query;
+        else if (action === "research") payload.topic = args.query;
+        else if (action === "compare") payload.items = args.items;
+        else if (action === "fact_check") payload.claim = args.claim || args.query;
+        return callCloudFunction(CLOUD_FUNCTIONS.research, payload, 60000);
+      }
+      // Fallback: Gemini grounded search
+      const mode = action === "fact_check" ? "fact_check" : action === "compare" ? "compare" : action === "research" ? "research" : "search";
+      const query = action === "compare" ? (args.items || []).join(" vs ") : (args.claim || args.query);
+      return geminiGroundedSearch(query, mode);
+    }
+
+    case "guidance":
+      return getGuidance(args.topic || "overview", args.tool_name);
+
+    default: {
+      // Route to external MCP backend
+      const activeTools = getActiveTools();
+      const extTool = activeTools.find(t => t.name === name);
+      if (extTool) {
+        const backend = mcpBackends.get(extTool._backend);
+        if (backend) return backend.callTool(extTool._originalName, args);
+        return { error: `Backend ${extTool._backend} not available` };
+      }
+      return { error: `Unknown tool: ${name}. Call 'guidance' with topic='groups' to see available tools.` };
+    }
+  }
+}
+
+// =============================================================================
+// PER-GROUP TOOL HELPERS
+// =============================================================================
+
+// Get tools for a specific group only
+function getToolsForGroup(groupName) {
+  const group = TOOL_GROUPS[groupName];
+  if (!group || !group.enabled) return [];
+  if (groupName === "core") return BUILTIN_TOOLS;
+
+  const allActive = getActiveTools();
+  if (!group.prefixes) {
+    // No prefix filter — return all tools from this backend
+    return allActive.filter(t => t._backend === group.source);
+  }
+  return allActive.filter(t =>
+    t._backend === group.source && group.prefixes.some(p => t._originalName.startsWith(p))
+  );
+}
+
+// Group display names for the Chat UI
+const GROUP_DISPLAY_NAMES = {
+  core: "Core Tools",
+  intelligence: "Intelligence & Learning",
+  agents: "Agents & Orchestration",
+  memory: "Memory & Knowledge",
+  devtools: "Dev Tools & Analysis",
+  security: "Security & Safety",
+  browser: "Browser Automation",
+  neural: "Neural & DAA",
+  "agentic-flow": "Agentic Flow",
+  "claude-code": "Claude Code",
+  gemini: "Gemini",
+  codex: "Codex",
+};
+
+// =============================================================================
+// MCP SERVER — Multiple endpoints per group
+// =============================================================================
+
+const app = express();
+app.use(express.json({ limit: "10mb" }));
+
+// ---------- MCP Streamable HTTP session (#2425 djimit) ----------
+// Streamable-HTTP clients (Codex/RMCP) send `DELETE /mcp` with an
+// `Mcp-Session-Id` header at shutdown. We echo a stable session id back
+// on every /mcp* response so those clients can attach it to the DELETE
+// and to `notifications/initialized` handshakes.
+const MCP_SESSION_ID = randomUUID();
+app.use((req, res, next) => {
+  if (req.path.startsWith("/mcp")) {
+    res.setHeader("Mcp-Session-Id", MCP_SESSION_ID);
+  }
+  next();
+});
+
+// ---------- CORS middleware (ADR-166 §6 Phase 3b) ----------
+// MCP_CORS_ORIGIN: comma-separated allowlist (e.g. "https://a.example,https://b.example").
+//   unset → "*" for back-compat (loopback default is same-origin anyway)
+//   "*"   → wildcard (explicit opt-in — same as legacy behavior)
+//   other → echo the request origin ONLY if it appears in the allowlist
+const CORS_ALLOWLIST = (process.env.MCP_CORS_ORIGIN || "*")
+  .split(",").map(s => s.trim()).filter(Boolean);
+const CORS_WILDCARD = CORS_ALLOWLIST.length === 1 && CORS_ALLOWLIST[0] === "*";
+app.use((req, res, next) => {
+  const origin = req.get("origin") || "";
+  if (CORS_WILDCARD) {
+    res.setHeader("Access-Control-Allow-Origin", "*");
+  } else if (origin && CORS_ALLOWLIST.includes(origin)) {
+    res.setHeader("Access-Control-Allow-Origin", origin);
+    res.setHeader("Vary", "Origin");
+  }
+  // If no match, do NOT set the header — browser will block the request.
+  res.setHeader("Access-Control-Allow-Methods", "GET, POST, DELETE, OPTIONS");
+  res.setHeader("Access-Control-Allow-Headers", "Content-Type, Authorization, Mcp-Session-Id");
+  if (req.method === "OPTIONS") return res.sendStatus(204);
+  next();
+});
+
+// ---------- Auth middleware ----------
+// No-op in local-only mode (MCP_AUTH_TOKEN unset). Enforces 401 when token is set.
+const MCP_TOKEN = process.env.MCP_AUTH_TOKEN || "";
+function requireAuth(req, res, next) {
+  if (req.path === "/health") return next();
+  if (!MCP_TOKEN) return next();
+  const expected = `Bearer ${MCP_TOKEN}`;
+  const got = req.get("authorization") || "";
+  const ok = got.length === expected.length &&
+    timingSafeEqual(Buffer.from(got), Buffer.from(expected));
+  if (!ok) return res.status(401).json({ error: "unauthorized" });
+  next();
+}
+app.use(requireAuth);
+
+// ---------- Shared MCP handler ----------
+function createMcpHandler(groupName) {
+  return async (req, res) => {
+    const { method, id, params } = req.body;
+    try {
+      switch (method) {
+        case "initialize":
+          return res.json({
+            jsonrpc: "2.0", id,
+            result: {
+              protocolVersion: "2024-11-05",
+              capabilities: { tools: {} },
+              serverInfo: { name: `mcp-bridge/${groupName}`, version: "2.0.0" },
+            },
+          });
+        case "tools/list": {
+          const tools = getToolsForGroup(groupName);
+          return res.json({ jsonrpc: "2.0", id, result: { tools } });
+        }
+        case "tools/call": {
+          const { name, arguments: toolArgs } = params;
+          const result = await executeTool(name, toolArgs || {});
+          return res.json({
+            jsonrpc: "2.0", id,
+            result: {
+              content: [{ type: "text", text: typeof result === "string" ? result : JSON.stringify(result, null, 2) }],
+            },
+          });
+        }
+        case "notifications/initialized":
+          // MCP streamable-HTTP spec: notifications must return 202 Accepted
+          // with an empty body (no jsonrpc envelope).
+          return res.status(202).end();
+        default:
+          return res.json({ jsonrpc: "2.0", id, error: { code: -32601, message: `Method not found: ${method}` } });
+      }
+    } catch (err) {
+      console.error(`MCP error [${groupName}/${method}]:`, err);
+      return res.json({ jsonrpc: "2.0", id, error: { code: -32603, message: err.message } });
+    }
+  };
+}
+
+function createMcpSseHandler(groupName) {
+  return (req, res) => {
+    res.setHeader("Content-Type", "text/event-stream");
+    res.setHeader("Cache-Control", "no-cache");
+    res.setHeader("Connection", "keep-alive");
+    res.write(`data: ${JSON.stringify({ type: "endpoint", url: `/mcp/${groupName}` })}\n\n`);
+  };
+}
+
+// ---------- Register per-group endpoints ----------
+for (const groupName of Object.keys(TOOL_GROUPS)) {
+  app.post(`/mcp/${groupName}`, createMcpHandler(groupName));
+  app.get(`/mcp/${groupName}`, createMcpSseHandler(groupName));
+  // #2425 djimit — streamable-HTTP session cleanup
+  app.delete(`/mcp/${groupName}`, (_, res) => res.sendStatus(204));
+}
+
+// ---------- Catch-all /mcp — serves ALL enabled tools (backwards-compatible) ----------
+app.post("/mcp", async (req, res) => {
+  const { method, id, params } = req.body;
+  try {
+    switch (method) {
+      case "initialize":
+        return res.json({
+          jsonrpc: "2.0", id,
+          result: {
+            protocolVersion: "2024-11-05",
+            capabilities: { tools: {} },
+            serverInfo: { name: "mcp-bridge", version: "2.0.0" },
+          },
+        });
+      case "tools/list": {
+        const activeTools = getActiveTools();
+        return res.json({ jsonrpc: "2.0", id, result: { tools: [...BUILTIN_TOOLS, ...activeTools] } });
+      }
+      case "tools/call": {
+        const { name, arguments: toolArgs } = params;
+        const result = await executeTool(name, toolArgs || {});
+        return res.json({
+          jsonrpc: "2.0", id,
+          result: {
+            content: [{ type: "text", text: typeof result === "string" ? result : JSON.stringify(result, null, 2) }],
+          },
+        });
+      }
+      case "notifications/initialized":
+        return res.status(202).end();
+      default:
+        return res.json({ jsonrpc: "2.0", id, error: { code: -32601, message: `Method not found: ${method}` } });
+    }
+  } catch (err) {
+    console.error(`MCP error [${method}]:`, err);
+    return res.json({ jsonrpc: "2.0", id, error: { code: -32603, message: err.message } });
+  }
+});
+
+app.get("/mcp", (req, res) => {
+  res.setHeader("Content-Type", "text/event-stream");
+  res.setHeader("Cache-Control", "no-cache");
+  res.setHeader("Connection", "keep-alive");
+  res.write(`data: ${JSON.stringify({ type: "endpoint", url: "/mcp" })}\n\n`);
+});
+
+// #2425 djimit — streamable-HTTP session cleanup on the catch-all route.
+app.delete("/mcp", (_, res) => res.sendStatus(204));
+
+// ---------- GET /mcp-servers — returns MCP_SERVERS JSON for Chat UI config ----------
+app.get("/mcp-servers", (_, res) => {
+  const servers = [];
+  for (const [name, group] of Object.entries(TOOL_GROUPS)) {
+    if (!group.enabled) continue;
+    const tools = getToolsForGroup(name);
+    if (tools.length === 0) continue;
+    servers.push({
+      name: GROUP_DISPLAY_NAMES[name] || name,
+      url: `/mcp/${name}`,
+      tools: tools.length,
+      group: name,
+    });
+  }
+  res.json(servers);
+});
+
+// =============================================================================
+// CHAT COMPLETIONS PROXY
+// =============================================================================
+
+const PROVIDER_ROUTES = {
+  openai: { baseURL: "https://api.openai.com/v1/chat/completions", getKey: () => process.env.OPENAI_API_KEY },
+  gemini: { baseURL: "https://generativelanguage.googleapis.com/v1beta/openai/chat/completions", getKey: () => process.env.GOOGLE_API_KEY },
+  openrouter: { baseURL: "https://openrouter.ai/api/v1/chat/completions", getKey: () => process.env.OPENROUTER_API_KEY },
+};
+
+function resolveProvider(model) {
+  if (typeof model === "string") {
+    if (model.startsWith("gemini-")) return "gemini";
+    if (model.includes("/")) return "openrouter";
+  }
+  return "openai";
+}
+
+// =============================================================================
+// SYSTEM PROMPT — Injected server-side into every chat completion request
+// =============================================================================
+// This comprehensive prompt teaches the AI how to use all 200+ MCP tools
+// across 5 groups. It is injected as the first system message, ensuring
+// consistent behavior regardless of what preprompt the Chat UI sends.
+
+function buildSystemPrompt() {
+  // Build dynamic group status
+  const enabledGroups = Object.entries(TOOL_GROUPS)
+    .filter(([, g]) => g.enabled)
+    .map(([name]) => name);
+
+  return `You are an intelligent AI assistant with powerful tools organized into ${enabledGroups.length} active groups.
+
+# CRITICAL RULES
+1. Use tools proactively — NEVER guess answers, ALWAYS search first
+2. For ANY factual question, current event, or research request → call \`web_research\` IMMEDIATELY
+3. NEVER say "I don't have access to real-time information" — you DO via web_research
+4. NEVER say "I can't search the web" — you CAN via web_research
+5. Call tools FIRST, present results conversationally AFTER
+6. When multiple tools could help, call them ALL in parallel
+
+# Tool Groups
+
+Tools prefixed with backend name (e.g., \`ruflo__agent_spawn\`). Always use the full prefixed name.
+
+## Group 1: Core Tools (always on)
+
+- **search** — Search your knowledge base (documents, workflows, how-tos).
+  Use for: internal knowledge, company docs, past conversations.
+  \`{"query": "how to process a claim"}\`
+
+- **web_research** — Search the internet via Google. This is your primary research tool.
+  ALWAYS use this for: current events, facts, comparisons, any external knowledge.
+  Actions:
+  - \`search\` — Quick web search. Default, fast. \`{"action": "search", "query": "latest news on X"}\`
+  - \`research\` — Deep report with synthesis. \`{"action": "research", "query": "comprehensive analysis of X"}\`
+  - \`compare\` — Side-by-side comparison. \`{"action": "compare", "query": "X", "items": ["A", "B"]}\`
+  - \`fact_check\` — Verify a claim. \`{"action": "fact_check", "query": "claim to verify"}\`
+  - \`goap\` — Multi-step pipeline: decomposes → parallel searches → synthesizes → verifies. BEST for important questions.
+    \`{"action": "goap", "query": "complex question requiring thorough research"}\`
+
+- **guidance** — Get help on tool groups or specific tool usage.
+  Topics: \`overview\`, \`groups\`, \`agents\`, \`memory\`, \`intelligence\`, \`devtools\`
+  \`{"topic": "overview"}\` or \`{"topic": "tool", "tool_name": "ruflo__agent_spawn"}\`
+
+## Group 2: Intelligence & Learning (ruvector)
+Pattern learning, routing, code analysis, and trajectory tracking. ${TOOL_GROUPS.intelligence.enabled ? "ACTIVE" : "DISABLED"}
+
+### Essential Intelligence Tools:
+- **ruvector__hooks_route** — Route a task to the optimal agent type. Call this FIRST for complex tasks.
+  \`{"task": "describe what needs to be done", "context": ["relevant info"]}\`
+  Returns ranked agent recommendations with confidence scores.
+- **ruvector__hooks_remember** — Store a key-value pair in persistent memory for cross-session recall.
+  \`{"key": "pattern-name", "value": "what to remember", "namespace": "patterns"}\`
+- **ruvector__hooks_recall** — Retrieve a previously stored memory by key.
+- **ruvector__hooks_suggest_context** — Get contextual suggestions based on current work.
+- **ruvector__hooks_swarm_recommend** — Get swarm topology recommendation for a task type.
+- **ruvector__hooks_capabilities** — List all intelligence system capabilities.
+
+### Code Analysis:
+- **ruvector__hooks_ast_analyze** — Analyze code structure (AST) of a file.
+- **ruvector__hooks_ast_complexity** — Get complexity metrics for code.
+- **ruvector__hooks_security_scan** — Scan code for security vulnerabilities.
+- **ruvector__hooks_diff_analyze** — Analyze a code diff for risk and impact.
+- **ruvector__hooks_diff_similar** — Find similar past diffs/changes.
+
+### Trajectory Learning (for multi-step tasks):
+- **ruvector__hooks_trajectory_begin** — Start tracking a multi-step task for learning.
+- **ruvector__hooks_trajectory_step** — Record a step in the current trajectory.
+- **ruvector__hooks_trajectory_end** — End trajectory, triggering pattern extraction.
+
+### Memory & Compression:
+- **ruvector__hooks_compress** — Compress/summarize long text for efficient storage.
+- **ruvector__hooks_rag_context** — Get RAG context for a query from stored knowledge.
+- **ruvector__hooks_learn** — Force the system to learn from provided examples.
+- **ruvector__hooks_batch_learn** — Learn from multiple examples at once.
+- **ruvector__hooks_stats** — View learning statistics and metrics.
+- **ruvector__hooks_doctor** — Run diagnostics on the intelligence system.
+
+## Group 3: Agents & Orchestration (ruflo)
+Spawn agents, coordinate swarms, manage tasks and workflows. ${TOOL_GROUPS.agents.enabled ? "ACTIVE" : "DISABLED"}
+
+### Agent Lifecycle:
+- **ruflo__agent_spawn** — Create a new specialized agent.
+  \`{"type": "coder|researcher|tester|reviewer|architect|security", "name": "optional-name"}\`
+  Agent types and when to use them:
+  - \`coder\` — Write code, implement features, fix bugs
+  - \`researcher\` — Find information, analyze documentation, investigate
+  - \`tester\` — Write tests, run test suites, validate behavior
+  - \`reviewer\` — Review code quality, security, best practices
+  - \`architect\` — Design systems, plan architectures, evaluate trade-offs
+  - \`security\` — Audit security, find vulnerabilities, recommend fixes
+- **ruflo__agent_status** — Check an agent's current state. \`{"agentId": "agent-xxx"}\`
+- **ruflo__agent_list** — List all active agents with their states.
+- **ruflo__agent_terminate** — Stop an agent. \`{"agentId": "agent-xxx"}\`
+- **ruflo__agent_health** — Health check across all agents.
+- **ruflo__agent_pool** — View the agent pool and available capacity.
+
+### Swarm Coordination:
+- **ruflo__swarm_init** — Initialize a multi-agent swarm.
+  \`{"topology": "hierarchical|mesh|ring|star", "maxAgents": 8, "strategy": "balanced|specialized|adaptive"}\`
+  - \`hierarchical\` — Coordinator + workers, best for structured tasks (anti-drift)
+  - \`mesh\` — Peer-to-peer, best for collaborative work
+  - \`ring\` — Sequential pipeline, best for ordered processing
+  - \`star\` — Central hub, best for fan-out parallel work
+- **ruflo__swarm_status** — Get swarm health, topology, and agent states.
+- **ruflo__swarm_health** — Detailed health metrics for the swarm.
+- **ruflo__swarm_shutdown** — Tear down a swarm and all its agents.
+
+### Task Management:
+- **ruflo__task_create** — Create a tracked task.
+  \`{"description": "what needs to be done", "priority": "low|normal|high|critical"}\`
+- **ruflo__task_status** — Check task progress. \`{"taskId": "task-xxx"}\`
+- **ruflo__task_list** — List all tasks with their statuses.
+- **ruflo__task_complete** — Mark a task as done. \`{"taskId": "task-xxx"}\`
+- **ruflo__task_update** — Update task details, status, or assignment.
+- **ruflo__task_cancel** — Cancel a task.
+
+### Workflow Orchestration:
+- **ruflo__workflow_create** — Define a multi-step workflow with dependencies.
+- **ruflo__workflow_execute** — Run a workflow. \`{"workflowId": "wf-xxx"}\`
+- **ruflo__workflow_status** — Check workflow progress.
+- **ruflo__workflow_template** — Use a pre-built workflow template.
+- **ruflo__workflow_pause** / **ruflo__workflow_resume** — Control workflow execution.
+
+### Hive-Mind (Distributed Consensus):
+- **ruflo__hive-mind_init** — Start distributed consensus system.
+- **ruflo__hive-mind_spawn** — Add an agent to the hive.
+- **ruflo__hive-mind_consensus** — Run consensus vote across agents.
+- **ruflo__hive-mind_broadcast** — Send message to all hive agents.
+- **ruflo__hive-mind_memory** — Access shared hive memory.
+
+### Coordination:
+- **ruflo__coordination_topology** — View/change coordination topology.
+- **ruflo__coordination_load_balance** — Distribute work across agents.
+- **ruflo__coordination_orchestrate** — Orchestrate complex multi-agent tasks.
+- **ruflo__coordination_sync** — Synchronize state across agents.
+
+### Session Management:
+- **ruflo__session_save** — Save current session state.
+- **ruflo__session_restore** — Restore a previous session.
+- **ruflo__session_list** — List available sessions.
+
+## Group 4: Memory & Knowledge (ruflo)
+Persistent memory, vector search, embeddings, and pattern storage. ${TOOL_GROUPS.memory.enabled ? "ACTIVE" : "DISABLED"}
+
+### Memory Operations:
+- **ruflo__memory_store** — Store data in persistent memory.
+  \`{"key": "my-key", "value": "data to store", "namespace": "default", "tags": ["tag1"]}\`
+- **ruflo__memory_retrieve** — Get stored data by key. \`{"key": "my-key"}\`
+- **ruflo__memory_search** — Semantic vector search across stored memories.
+  \`{"query": "what to search for", "limit": 5, "namespace": "default"}\`
+- **ruflo__memory_list** — List all stored keys in a namespace.
+- **ruflo__memory_delete** — Remove a stored memory.
+- **ruflo__memory_stats** — View memory usage statistics.
+
+### Embeddings:
+- **ruflo__embeddings_generate** — Generate vector embeddings for text.
+- **ruflo__embeddings_compare** — Compare semantic similarity of two texts.
+- **ruflo__embeddings_search** — Search embeddings database by similarity.
+- **ruflo__embeddings_neural** — Generate neural embeddings.
+- **ruflo__embeddings_hyperbolic** — Generate hyperbolic embeddings for hierarchical data.
+
+### AgentDB (Advanced Pattern Storage):
+- **ruflo__agentdb_pattern-store** — Store a learned pattern with metadata.
+  \`{"pattern": "description", "category": "code|debug|architecture", "confidence": 0.9}\`
+- **ruflo__agentdb_pattern-search** — Search patterns by similarity.
+- **ruflo__agentdb_route** — Route a query to the most relevant stored pattern.
+- **ruflo__agentdb_feedback** — Provide feedback on a pattern (reinforcement learning).
+- **ruflo__agentdb_context-synthesize** — Synthesize context from multiple sources.
+- **ruflo__agentdb_semantic-route** — Semantic routing based on stored knowledge.
+- **ruflo__agentdb_consolidate** — Consolidate and deduplicate stored patterns.
+- **ruflo__agentdb_batch** — Batch operations on patterns.
+- **ruflo__agentdb_session-start** / **ruflo__agentdb_session-end** — Session tracking.
+- **ruflo__agentdb_hierarchical-store** / **ruflo__agentdb_hierarchical-recall** — Hierarchical memory.
+
+## Group 5: Dev Tools & Analysis (ruflo)
+Performance, system health, GitHub integration, code analysis, terminal. ${TOOL_GROUPS.devtools.enabled ? "ACTIVE" : "DISABLED"}
+
+### System & Performance:
+- **ruflo__system_status** — System health overview.
+- **ruflo__system_metrics** — Detailed performance metrics.
+- **ruflo__system_health** — Health check across all subsystems.
+- **ruflo__performance_report** — Generate performance report.
+- **ruflo__performance_bottleneck** — Identify performance bottlenecks.
+- **ruflo__performance_benchmark** — Run benchmarks.
+- **ruflo__performance_optimize** — Get optimization recommendations.
+- **ruflo__performance_profile** — Profile specific operations.
+
+### Code Analysis:
+- **ruflo__analyze_diff** — Analyze a code diff.
+- **ruflo__analyze_diff-risk** — Assess risk level of changes.
+- **ruflo__analyze_diff-classify** — Classify type of changes (feature, bugfix, refactor).
+- **ruflo__analyze_diff-reviewers** — Suggest code reviewers.
+- **ruflo__analyze_file-risk** — Assess risk of a specific file.
+
+### GitHub Integration:
+- **ruflo__github_repo_analyze** — Analyze a GitHub repository.
+  \`{"repo": "owner/repo", "analysis_type": "code_quality|performance|security"}\`
+- **ruflo__github_pr_manage** — Manage pull requests (create, review, merge).
+- **ruflo__github_issue_track** — Track and manage issues.
+- **ruflo__github_workflow** — Manage GitHub Actions workflows.
+- **ruflo__github_metrics** — Repository metrics and insights.
+
+### Terminal Access:
+- **ruflo__terminal_create** — Create a terminal session.
+- **ruflo__terminal_execute** — Execute a command. \`{"command": "ls -la"}\`
+- **ruflo__terminal_list** — List active terminals.
+- **ruflo__terminal_history** — View command history.
+
+### Development Hooks:
+- **ruflo__hooks_pre-task** / **ruflo__hooks_post-task** — Task lifecycle hooks for learning.
+- **ruflo__hooks_pre-edit** / **ruflo__hooks_post-edit** — File edit hooks.
+- **ruflo__hooks_session-start** / **ruflo__hooks_session-end** — Session lifecycle.
+- **ruflo__hooks_worker-dispatch** — Dispatch background workers.
+  Workers: \`optimize\`, \`audit\`, \`testgaps\`, \`document\`, \`map\`, \`deepdive\`, \`benchmark\`
+- **ruflo__hooks_model-route** — Route to optimal AI model for a task.
+- **ruflo__hooks_explain** — Explain a routing or intelligence decision.
+
+### Configuration:
+- **ruflo__config_get** / **ruflo__config_set** / **ruflo__config_list** — Manage settings.
+
+### Progress Tracking:
+- **ruflo__progress_check** — Check implementation progress.
+- **ruflo__progress_summary** — Summarize overall progress.
+
+# Decision Framework — FOLLOW THIS EXACTLY
+
+1. **ANY factual question** → \`web_research(action='search')\` IMMEDIATELY. Do NOT answer from memory.
+2. **"What is X?"** → \`web_research(action='search', query='X')\`
+3. **"Compare X vs Y"** → \`web_research(action='compare', query='X vs Y', items=['X', 'Y'])\`
+4. **"Is it true that..."** → \`web_research(action='fact_check', query='...')\`
+5. **Complex research** → \`web_research(action='goap', query='...')\` — auto multi-step pipeline
+6. **Internal docs/procedures** → \`search(query='...')\` first, then web_research if not found
+7. **Code task** → \`ruvector__hooks_route\` + \`ruflo__agent_spawn\`
+8. **Remember something** → \`ruflo__memory_store\` / \`ruflo__memory_search\`
+9. **"What can you do?"** → \`guidance(topic='overview')\`
+10. **Unknown** → \`guidance(topic='overview')\`
+
+IMPORTANT: For questions 1-5, ALWAYS call web_research. Never say "I don't have access to search" — you DO.
+
+# Execution Patterns
+
+### Simple Question
+\`search\` or \`web_research\` → synthesize → respond
+
+### Complex Research
+\`web_research(action='goap')\` → analyze → respond with citations
+
+### Code Implementation
+\`ruvector__hooks_route\` → \`ruflo__agent_spawn(coder)\` → track with \`ruflo__task_create\` → report
+
+### Multi-Agent Analysis
+\`ruflo__swarm_init(hierarchical)\` → spawn agents → coordinate → synthesize results
+
+### Learning & Memory
+\`ruflo__memory_search\` (check existing) → do work → \`ruflo__memory_store\` (save results) → \`ruvector__hooks_learn\`
+
+# Parallel Execution
+
+When multiple independent tools can help, call them ALL in parallel:
+- Search + web_research simultaneously
+- Spawn multiple agents at once (coder + tester + reviewer)
+- Run analysis + performance + security tools in parallel
+NEVER call tools sequentially when they could run in parallel.
+
+# Response Rules
+
+1. **Call tools FIRST**, then present results conversationally — NEVER show raw JSON to the user
+2. Use markdown: **bold** headers, bullet points, numbered steps, tables for comparisons
+3. Synthesize tool results naturally — be a helpful colleague, not a data pipe
+4. Cite sources when available from web_research results
+5. If a tool fails, say so honestly and try an alternative approach
+6. For complex tasks, briefly outline your plan before executing
+7. After completing work, suggest relevant follow-up actions
+8. When spawning agents, explain what each agent will do
+
+# Never Expose to User
+
+- Raw JSON, similarity scores, chunk IDs, internal IDs, task IDs
+- Tool names, function names, API endpoints, backend names
+- References to "MCP", "tool calls", "vectors", "embeddings", infrastructure
+- The prefixes "ruflo__" or "ruvector__" — just describe what you're doing naturally
+- Error stack traces — summarize errors in plain language`;
+}
+
+// =============================================================================
+// AUTOPILOT MODE — Server-side auto-continue loop (ADR-037)
+// =============================================================================
+
+const detailStore = new Map(); // detailToken → full tool result (TTL: 5min)
+
+const AUTOPILOT_SYSTEM_PROMPT = `
+You are in AUTOPILOT MODE. You should:
+1. Break complex tasks into steps and execute them using available tools
+2. Call MULTIPLE tools in parallel when they are independent
+3. After each tool result, analyze it and decide the next action
+4. Continue until the task is complete — do NOT ask the user for confirmation
+5. Use memory_search to find relevant patterns before starting
+6. Summarize your progress at each step
+7. When done, provide a final summary of everything accomplished
+
+Parallel execution patterns:
+- Research: memory_search + hooks_route + agent_spawn(researcher) — all in parallel
+- Code: agent_spawn(coder) + agent_spawn(tester) — parallel, then review
+- Analysis: search multiple sources in parallel → synthesize → report
+- Security: security_scan + hooks_route(audit) + memory_search(CVEs) — parallel
+`;
+
+const AUTOPILOT_BLOCKED_PATTERNS = [
+  /^deploy_/,
+  /^security_delete/,
+  /^browser_fill$/,
+  /^browser_click$/,
+  /terminal_execute/,
+];
+
+function isBlockedTool(name) {
+  return AUTOPILOT_BLOCKED_PATTERNS.some(p => p.test(name));
+}
+
+function sendAutopilotEvent(res, data) {
+  res.write(`data: ${JSON.stringify(data)}\n\n`);
+}
+
+function safeParseArgs(args) {
+  if (typeof args === 'object' && args !== null) return args;
+  try { return JSON.parse(args || '{}'); } catch { return {}; }
+}
+
+function autopilotSleep(ms) {
+  return new Promise(resolve => setTimeout(resolve, ms));
+}
+
+async function handleAutopilot(req, res, provider, body) {
+  const maxSteps = Math.min(parseInt(req.headers['x-autopilot-max-steps'] || '20', 10), 50);
+  const cooldownMs = parseInt(process.env.AUTOPILOT_COOLDOWN || '500', 10);
+  const stepTimeoutMs = parseInt(process.env.AUTOPILOT_STEP_TIMEOUT || '30000', 10);
+
+  // SSE setup
+  res.setHeader('Content-Type', 'text/event-stream');
+  res.setHeader('Cache-Control', 'no-cache');
+  res.setHeader('Connection', 'keep-alive');
+  res.setHeader('X-Accel-Buffering', 'no');
+
+  let messages = [...body.messages];
+  let step = 0;
+  let aborted = false;
+  let totalTasks = 0;
+  const startTime = Date.now();
+
+  req.on('close', () => { aborted = true; });
+
+  sendAutopilotEvent(res, { type: 'autopilot_start', maxSteps });
+
+  // Get the tools list for the AI provider (OpenAI function calling format)
+  // Limit to BUILTIN_TOOLS + top external tools to avoid overwhelming the model
+  // (Gemini struggles with 100+ tool definitions)
+  const MAX_AUTOPILOT_TOOLS = 30;
+  const externalTools = getActiveTools();
+  // Prioritize tools from enabled high-value groups
+  const priorityPrefixes = ['ruvector__hooks', 'ruvector__memory', 'ruflo__memory', 'ruflo__agent'];
+  const prioritized = externalTools
+    .sort((a, b) => {
+      const aP = priorityPrefixes.some(p => a.name.startsWith(p)) ? 0 : 1;
+      const bP = priorityPrefixes.some(p => b.name.startsWith(p)) ? 0 : 1;
+      return aP - bP;
+    })
+    .slice(0, MAX_AUTOPILOT_TOOLS - BUILTIN_TOOLS.length);
+  const allTools = [...BUILTIN_TOOLS, ...prioritized];
+  const toolDefs = allTools.map(t => ({
+    type: 'function',
+    function: {
+      name: t.name,
+      description: t.description || '',
+      parameters: t.inputSchema || { type: 'object', properties: {} },
+    },
+  }));
+  console.log(`[autopilot] ${allTools.length} tools (${BUILTIN_TOOLS.length} builtin + ${prioritized.length} external)`);
+
+  while (step < maxSteps && !aborted) {
+    // 1. Call upstream AI provider (non-streaming for tool call parsing)
+    const apiKey = provider.getKey();
+    let aiResult;
+    try {
+      const aiResponse = await fetch(provider.baseURL, {
+        method: 'POST',
+        headers: { 'Content-Type': 'application/json', Authorization: `Bearer ${apiKey}` },
+        body: JSON.stringify({
+          ...body,
+          messages,
+          stream: false,
+          tools: toolDefs.length > 0 ? toolDefs : undefined,
+        }),
+        signal: AbortSignal.timeout(stepTimeoutMs),
+      });
+      aiResult = await aiResponse.json();
+    } catch (err) {
+      sendAutopilotEvent(res, { type: 'autopilot_error', error: `AI call failed: ${err.message}` });
+      break;
+    }
+
+    const choice = aiResult.choices?.[0];
+    if (!choice) {
+      sendAutopilotEvent(res, { type: 'autopilot_error', error: 'No response from AI' });
+      break;
+    }
+
+    // 2. Check for tool calls
+    const toolCalls = choice.message?.tool_calls;
+
+    if (!toolCalls || toolCalls.length === 0) {
+      // Final text response — send it
+      sendAutopilotEvent(res, { type: 'autopilot_text', content: choice.message?.content || '' });
+      break;
+    }
+
+    // 3. Execute ALL tool calls in parallel
+    step++;
+    const groupId = `g${step}`;
+    const taskEvents = toolCalls.map((tc, i) => ({
+      taskId: `t${totalTasks + i + 1}`,
+      tool: tc.function.name,
+      args: safeParseArgs(tc.function.arguments),
+      status: 'running',
+    }));
+    totalTasks += taskEvents.length;
+
+    // If the AI also included text content, stream it before tools
+    if (choice.message?.content) {
+      sendAutopilotEvent(res, { type: 'autopilot_text', content: choice.message.content });
+    }
+
+    // Stream group start
+    sendAutopilotEvent(res, { type: 'task_group_start', groupId, step, tasks: taskEvents });
+
+    // Append assistant message to conversation
+    messages.push(choice.message);
+
+    // Execute tools in parallel
+    const groupStart = Date.now();
+    const results = await Promise.allSettled(
+      toolCalls.map(async (tc, i) => {
+        const taskId = taskEvents[i].taskId;
+        const toolName = tc.function.name;
+        const toolArgs = safeParseArgs(tc.function.arguments);
+        const taskStart = Date.now();
+
+        // Check blocklist
+        if (isBlockedTool(toolName)) {
+          sendAutopilotEvent(res, {
+            type: 'task_update', taskId, status: 'blocked',
+            summary: `${toolName} requires confirmation`,
+            duration: Date.now() - taskStart,
+          });
+          return { toolCallId: tc.id, blocked: true, toolName };
+        }
+
+        try {
+          const result = await executeTool(toolName, toolArgs);
+          const resultStr = typeof result === 'string' ? result : JSON.stringify(result, null, 2);
+
+          // Store full detail, generate token for lazy loading
+          const detailToken = `dt_${taskId}_${Date.now()}`;
+          detailStore.set(detailToken, resultStr);
+
+          // Stream task completion with summary only
+          const summary = resultStr.length > 120
+            ? resultStr.substring(0, 120).replace(/\n/g, ' ') + '...'
+            : resultStr.replace(/\n/g, ' ');
+
+          sendAutopilotEvent(res, {
+            type: 'task_update', taskId, status: 'completed',
+            summary, duration: Date.now() - taskStart, detailToken,
+          });
+
+          return { toolCallId: tc.id, content: resultStr };
+        } catch (err) {
+          sendAutopilotEvent(res, {
+            type: 'task_update', taskId, status: 'failed',
+            summary: err.message, duration: Date.now() - taskStart,
+          });
+          return { toolCallId: tc.id, content: `Error: ${err.message}` };
+        }
+      })
+    );
+
+    // Stream group end
+    sendAutopilotEvent(res, { type: 'task_group_end', groupId, step, duration: Date.now() - groupStart });
+
+    // Check if any tools were blocked — pause autopilot
+    const blockedResults = results
+      .filter(r => r.status === 'fulfilled' && r.value.blocked)
+      .map(r => r.value);
+    if (blockedResults.length > 0) {
+      sendAutopilotEvent(res, {
+        type: 'autopilot_paused',
+        reason: 'blocked_tools',
+        tools: blockedResults.map(b => b.toolName),
+      });
+      break;
+    }
+
+    // Append tool results to messages
+    for (const r of results) {
+      if (r.status === 'fulfilled' && !r.value.blocked) {
+        messages.push({
+          role: 'tool',
+          tool_call_id: r.value.toolCallId,
+          content: r.value.content,
+        });
+      }
+    }
+
+    // Cooldown to prevent runaway
+    await autopilotSleep(cooldownMs);
+  }
+
+  if (step >= maxSteps && !aborted) {
+    sendAutopilotEvent(res, {
+      type: 'autopilot_text',
+      content: `\n⚠️ Autopilot reached max steps (${maxSteps}). Stopping.\n`,
+    });
+  }
+
+  sendAutopilotEvent(res, {
+    type: 'autopilot_end',
+    totalSteps: step,
+    totalTasks,
+    duration: Date.now() - startTime,
+  });
+
+  res.write('data: [DONE]\n\n');
+  res.end();
+
+  // Clean up detail store after 5 minutes
+  const detailTTL = parseInt(process.env.AUTOPILOT_DETAIL_TTL || '300000', 10);
+  setTimeout(() => {
+    for (const [key] of detailStore) {
+      if (key.startsWith('dt_')) detailStore.delete(key);
+    }
+  }, detailTTL);
+}
+
+// Lazy detail loading endpoint
+app.get('/autopilot/detail/:token', (req, res) => {
+  const content = detailStore.get(req.params.token);
+  if (content) {
+    res.json({ content });
+  } else {
+    res.status(404).json({ error: 'Detail expired or not found' });
+  }
+});
+
+// =============================================================================
+// CHAT COMPLETIONS PROXY
+// =============================================================================
+
+app.post("/chat/completions", async (req, res) => {
+  const model = req.body?.model;
+  const providerName = resolveProvider(model);
+  const provider = PROVIDER_ROUTES[providerName];
+  const apiKey = provider.getKey();
+
+  if (!apiKey) return res.status(401).json({ error: { message: `No API key for provider: ${providerName}` } });
+
+  // Inject comprehensive system prompt as the first message
+  const body = { ...req.body };
+  if (body.messages && Array.isArray(body.messages)) {
+    let systemPrompt = buildSystemPrompt();
+    // Add autopilot instructions if autopilot mode is active
+    const isAutopilot = req.headers['x-autopilot'] === 'true';
+    if (isAutopilot) {
+      systemPrompt = AUTOPILOT_SYSTEM_PROMPT + '\n\n' + systemPrompt;
+    }
+    // Prepend our system prompt before any existing messages
+    const hasSystemMsg = body.messages[0]?.role === "system";
+    if (hasSystemMsg) {
+      // Merge with existing system message
+      body.messages = [
+        { role: "system", content: systemPrompt + "\n\n" + body.messages[0].content },
+        ...body.messages.slice(1),
+      ];
+    } else {
+      body.messages = [{ role: "system", content: systemPrompt }, ...body.messages];
+    }
+  }
+
+  // Route to autopilot handler if x-autopilot header is set
+  if (req.headers['x-autopilot'] === 'true') {
+    return handleAutopilot(req, res, provider, body);
+  }
+
+  try {
+    const upstream = await fetch(provider.baseURL, {
+      method: "POST",
+      headers: { "Content-Type": "application/json", Authorization: `Bearer ${apiKey}` },
+      body: JSON.stringify(body),
+    });
+
+    if (!upstream.ok) {
+      const errBody = await upstream.text();
+      console.error(`Proxy error [${providerName}/${model}]: ${upstream.status} ${errBody.substring(0, 200)}`);
+      // Normalize all upstream errors into OpenAI-compatible format so the
+      // Chat UI's OpenAI SDK can parse them instead of "400 (no body)".
+      let errorMessage = `Upstream ${providerName} error (${upstream.status})`;
+      try {
+        const parsed = JSON.parse(errBody);
+        // Gemini returns [{"error": {"message": "..."}}]
+        if (Array.isArray(parsed) && parsed[0]?.error?.message) {
+          errorMessage = parsed[0].error.message;
+        // OpenAI/OpenRouter return {"error": {"message": "..."}}
+        } else if (parsed.error?.message) {
+          errorMessage = parsed.error.message;
+        }
+      } catch {}
+      return res.status(upstream.status).json({
+        error: { message: errorMessage, type: "upstream_error", code: upstream.status },
+      });
+    }
+
+    res.setHeader("Content-Type", upstream.headers.get("content-type") || "application/json");
+
+    if (req.body?.stream && upstream.body) {
+      const reader = upstream.body.getReader();
+      const decoder = new TextDecoder();
+      try {
+        while (true) {
+          const { done, value } = await reader.read();
+          if (done) break;
+          res.write(decoder.decode(value, { stream: true }));
+        }
+      } catch (e) { /* stream closed */ }
+      finally { res.end(); }
+    } else {
+      res.send(await upstream.text());
+    }
+  } catch (err) {
+    console.error(`Proxy error [${providerName}/${model}]:`, err.message);
+    res.status(502).json({ error: { message: `Upstream error: ${err.message}` } });
+  }
+});
+
+// =============================================================================
+// MODELS & HEALTH
+// =============================================================================
+
+const KNOWN_MODELS = [
+  "gemini-2.5-pro", "gemini-2.5-flash",
+  "gpt-4.1", "gpt-4.1-mini", "gpt-4o", "gpt-4o-mini",
+  "o3-mini", "o1-mini",
+];
+
+app.get("/models", (_, res) => {
+  res.json({ object: "list", data: KNOWN_MODELS.map(id => ({
+    id,
+    object: "model",
+    owned_by: "system",
+    providers: [{ supports_tools: true }],
+  })) });
+});
+
+app.get("/health", (_, res) => {
+  const backends = {};
+  for (const [name, client] of mcpBackends) {
+    backends[name] = { ready: client.ready, tools: client.tools.length };
+  }
+  const activeTools = getActiveTools();
+  const groups = {};
+  for (const [name, g] of Object.entries(TOOL_GROUPS)) {
+    groups[name] = { enabled: g.enabled, source: g.source };
+  }
+  res.json({
+    status: "ok", service: "mcp-bridge", version: "2.0.0",
+    tools: { builtin: BUILTIN_TOOLS.length, external: activeTools.length, total: BUILTIN_TOOLS.length + activeTools.length },
+    groups, backends,
+  });
+});
+
+// GET /groups — list tool groups and their status
+app.get("/groups", (_, res) => {
+  const activeTools = getActiveTools();
+  const result = {};
+  for (const [name, g] of Object.entries(TOOL_GROUPS)) {
+    const tools = name === "core" ? BUILTIN_TOOLS :
+      activeTools.filter(t => {
+        if (g.source !== t._backend) return false;
+        if (!g.prefixes) return true;
+        return g.prefixes.some(p => t._originalName.startsWith(p));
+      });
+    result[name] = {
+      enabled: g.enabled,
+      description: g.description,
+      tools: tools.length,
+      toolNames: tools.map(t => t.name).slice(0, 10),
+    };
+  }
+  res.json(result);
+});
+
+// =============================================================================
+// STARTUP
+// =============================================================================
+
+async function main() {
+  const isPublic = BIND_HOST !== "127.0.0.1" && BIND_HOST !== "localhost";
+  if (isPublic && !process.env.MCP_AUTH_TOKEN) {
+    console.error(
+      "FATAL: refusing to bind a public interface without MCP_AUTH_TOKEN. " +
+      "Generate one with: MCP_AUTH_TOKEN=$(openssl rand -base64 32)"
+    );
+    process.exit(1);
+  }
+  app.listen(PORT, BIND_HOST, () => {
+    console.log(`MCP Bridge v2.0.0 on port ${PORT} (${BIND_HOST})`);
+    const enabled = Object.entries(TOOL_GROUPS).filter(([, g]) => g.enabled).map(([n]) => n);
+    console.log(`Active groups: ${enabled.join(", ")}`);
+    // ADR-166 §6 — startup posture banner (helps operators see the security state at boot)
+    console.log(
+      `[security] bind=${BIND_HOST} auth=${process.env.MCP_AUTH_TOKEN ? "bearer" : "off (local-only)"} ` +
+      `terminal=${MCP_ENABLE_TERMINAL ? "ENABLED (⚠ opt-in)" : "disabled"}`,
+    );
+    if (MCP_ENABLE_TERMINAL) {
+      console.warn(
+        "[security] WARNING: terminal_execute is enabled. This tool grants shell access " +
+        "inside the bridge container to any client the auth layer accepts. Ensure " +
+        "MCP_AUTH_TOKEN is set on any non-loopback bind. See ADR-166 §6 Phase 1d.",
+      );
+    }
+  });
+
+  const anyBackendNeeded = BACKEND_DEFS.some(isBackendNeeded);
+  if (anyBackendNeeded) {
+    console.log("Initializing MCP backends...");
+    await initBackends();
+  }
+}
+
+main().catch(err => { console.error("Fatal:", err); process.exit(1); });