mcp-swarm 0.9.16 → 1.0.0

package/CHANGELOG.md

@@ -1,1471 +1,1037 @@
-# Changelog
-
-Все значимые изменения в проекте MCP Swarm документируются в этом файле.
-
-Формат основан на [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
-версионирование следует [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
-
----
-
-## [0.9.16] - 2026-02-08
-
-### 🐛 Критический фикс: Zod совместимость
-
-**BREAKING FIX:** Исправлена ошибка `keyValidator._parse is not a function`, которая делала все 54 Smart Tools полностью неработоспособными в Antigravity и подобных клиентах.
-
-#### Причина
-- `zod@^3.23.8` в `package.json` автоматически устанавливал `3.25.76` — по сути Zod v4 bridge-версию
-- Zod v4 удалил внутренний метод `_parse()`, который используется клиентами для валидации JSON Schema
-- Результат: **ни один Smart Tool не работал** из-за ошибки парсинга входных схем
-
-#### Исправление
-- `zod` закреплён на точной версии `3.23.8` (чистый Zod v3, без v4 bridge)
-- Предотвращает автоматический апгрейд до несовместимых версий
-
-> 🙏 Приносим извинения за неудобства. Баг затронул всех пользователей, обновившихся через npm.
-
-### 🔒 Безопасность: Cloudflare Workers
-
-- Удалены задеплоенные воркеры `mcp-swarm-hub` и `mcp-swarm-server` (URL-ы были засвечены в коммитах)
-- Исходники воркеров **сохранены** в `cloudflare/` — будут передеплоены с новыми именами
-
-### 📦 Обновления
-
-- `serverSmart.ts` — динамическая версия из `package.json` вместо хардкода
-- `server.ts` — помечен как deprecated (`@ts-nocheck`), используйте `serverSmart.ts`
-- `.gitignore` — добавлены паттерны для тестовых файлов
-
-### 🔄 Обновление
-
-```bash
-npm install -g mcp-swarm@latest
-# или
-npm update mcp-swarm
-```
-
----
-
-## [0.9.13] - 2026-02-05
-
-### Self-Hosted Infrastructure
-
-**BREAKING CHANGE:** Removed hardcoded public server URLs. Users must now deploy their own Cloudflare Workers.
-
-#### Changed
-
-- **README.md** — Complete rewrite of installation section
-  - Added Cloudflare Free Tier explanation and limits
-  - Step-by-step deployment guide for Hub, MCP Server, Telegram Bot
-  - Explained what YOUR-SUBDOMAIN means
-
-- **wrangler.toml files** — All 3 Workers now have placeholders
-  - `cloudflare/hub/wrangler.toml` — Added deployment instructions
-  - `cloudflare/mcp-server/wrangler.toml` — HUB_URL placeholder
-  - `cloudflare/telegram-bot/wrangler.toml` — SWARM_HUB_URL placeholder
-
-- **install.ps1 / install.sh** — Now ask for user's server URLs
-  - No more hardcoded URLs
-  - Instructions for deploying first
-
-- **mcp-swarm-remote** — `--url` is now **required**
-  - Shows helpful error message with deployment link
-  - Added `--help` option
-
-- **Dashboard** — `.env.example` uses placeholder
-  - `useWebSocket.ts` warns if HUB_URL not configured
-
-- **Agent rules** (AGENTS.md, CLAUDE.md, GEMINI.md, AGENT.md)
-  - Updated with YOUR-SUBDOMAIN placeholders
-
-#### Why This Change?
-
-- **Privacy**: Your data stays on your infrastructure
-- **No shared limits**: Full Cloudflare Free Tier for yourself
-- **Customization**: Modify Workers as needed
-- **Transparency**: No hidden public server
-
----
-
-## [0.9.12] - 2026-02-05
-
-### Dashboard Real-time Updates
-
-#### Added
-
-- **WebSocket Hook** (`dashboard/src/hooks/useWebSocket.ts`)
-  - Real-time connection to Swarm Hub
-  - Auto-reconnect with exponential backoff
-  - Keep-alive ping every 25 seconds
-  - Event filtering for specific event types
-
-- **New Core Widgets** (`dashboard/src/components/widgets/CoreWidgets.tsx`)
-  - **ConnectionStatusWidget** — Shows Hub connection status with reconnect button
-  - **ActivityTimelineWidget** — Real-time event stream from Hub
-  - **FileLocksWidget** — Active file reservations with live updates
-  - **CostTrackingWidget** — API usage and budget progress bars
-  - **VotingWidget** — Distributed consensus proposals and votes
-
-- **Live Indicators**
-  - Green pulsing dot for connected/active status
-  - LIVE/OFFLINE badges for real-time widgets
-  - Last update timestamps with relative time format
-
-### Telegram Bot Enhancements
-
-#### Added
-
-- **New Commands**
-  - `/reviews` — List pending code reviews with inline approve/reject buttons
-  - `/approve [id]` — Approve a review directly from Telegram
-  - `/reject [id] [reason]` — Reject a review with optional reason
-
-- **Enhanced Inline Keyboards**
-  - Reviews list shows approve/reject buttons for each pending review
-  - Help menu now includes Reviews button
-
-### Auto-start Companion
-
-#### Added
-
-- **Companion Auto-start** in `mcp-swarm-remote`
-  - Checks if companion daemon is running on port 37373
-  - Automatically spawns companion in background if not running
-  - Can be disabled with `--no-companion` flag
-
-#### Changed
-
-- Dashboard now uses WebSocket for real-time updates instead of polling-only
-- Added "Real-time Monitoring" section with new widgets
-- Updated version to 0.9.12 in footer
-
----
-
-## [0.9.11] - 2026-02-04
-
-### One-Click Installer & Streamable HTTP
-
-#### Added
-
-- **One-Click Install Scripts**
-  - `install.ps1` — PowerShell installer for Windows
-  - `install.sh` — Bash installer for macOS/Linux
-  - `npx mcp-swarm-install` — Interactive Node.js installer
-
-- **Streamable HTTP Transport** (MCP spec 2025-03-26)
-  - Single `POST /mcp` endpoint instead of SSE
-  - Session management via `Mcp-Session-Id` header
-  - Works on Cloudflare Workers without SSE issues
-
-- **mcp-swarm-remote Proxy** (`src/remote/index.ts`)
-  - stdio → Streamable HTTP proxy for IDE integration
-  - Windows stdin compatibility fix
-
-#### Changed
-
-- Installer now **merges** configs instead of overwriting
-- Auto-detects IDE installations (Claude Desktop, Cursor, Windsurf, OpenCode, VS Code)
-
----
-
-## [0.9.10] - 2026-02-03
-
-### 🧠 MoE Router — Mixture of Experts Model Selection
-
-#### Added
-
-- **MoE Router Module** (`src/workflows/moeRouter.ts`)
-  - Intelligent model routing based on task characteristics
-  - Gating network for expert selection
-  - Cost-performance optimization
-  - Learning from feedback
-
-- **Expert Classification**
-  - 14 task categories: code_generation, code_review, debugging, reasoning, math, creative, etc.
-  - 4 model tiers: economy, standard, premium, flagship
-  - 6 providers: anthropic, openai, google, mistral, local, custom
-
-- **Built-in Experts (16 models, verified pricing February 2026)**
-  - **Anthropic Claude 4.5 Series:**
-    - Claude Opus 4.5 (flagship, 200K, $5/$25 MTok)
-    - Claude Sonnet 4.5 (premium, 200K, $3/$15 MTok)
-    - Claude Haiku 4.5 (economy, 200K, $1/$5 MTok)
-  - **OpenAI GPT-5.x Series:**
-    - GPT-5.2 (flagship, 256K, $1.75/$14 MTok)
-    - GPT-5.2 Pro (flagship, 256K, $21/$168 MTok)
-    - GPT-5 Mini (standard, 128K, $0.25/$2 MTok)
-    - GPT-4.1 (premium, 128K, $3/$12 MTok)
-    - GPT-4.1 Mini (standard, 128K, $0.80/$3.20 MTok)
-    - GPT-4.1 Nano (economy, 128K, $0.20/$0.80 MTok)
-  - **OpenAI Reasoning Models:**
-    - o4-mini (reasoning, 128K, $4/$16 MTok)
-  - **Google Gemini 3.x Series:**
-    - Gemini 3 Pro (flagship, 1M, $2/$12 MTok)
-    - Gemini 3 Flash (standard, 1M, $0.50/$3 MTok)
-  - **Google Gemini 2.5 Series:**
-    - Gemini 2.5 Pro (premium, 1M, $1.25/$10 MTok)
-    - Gemini 2.5 Flash (standard, 1M, $0.30/$2.50 MTok)
-    - Gemini 2.5 Flash Lite (economy, 1M, $0.10/$0.40 MTok)
-
-- **Routing Factors**
-  - Task match score (category-specific strength)
-  - Cost efficiency
-  - Performance history (success rate)
-  - Load balancing (prefer less recently used)
-  - Latency constraints
-  - Context window requirements
-
-- **Learning System**
-  - Feedback recording
-  - Exponential moving average for success rate
-  - Automatic latency calibration
-  - Per-expert statistics
-
-- **Smart Tool #54: `swarm_moe`**
-  - `route`: Route task to best expert
-  - `feedback`: Record routing feedback
-  - `experts`: List available experts
-  - `add_expert` / `remove_expert`: Manage experts
-  - `config` / `set_config`: Configuration
-  - `stats`: Routing statistics
-  - `history`: Routing history
-  - `classify`: Classify task category
-  - `reset`: Reset statistics
-
-- **Dashboard API Endpoint** (`/api/moe`)
-  - Routing statistics
-  - Expert list and usage
-  - Configuration status
-
-#### Example Usage
-
-```typescript
-// Route task to best model
-swarm_moe({
-  action: "route",
-  content: "Write a React component for user authentication",
-  preferredTier: "premium",
-  maxCost: 0.05,
-  repoPath
-});
-// → { selectedExpert: "claude-sonnet", confidence: 0.92, estimatedCost: $0.02 }
-
-// Record feedback for learning
-swarm_moe({
-  action: "feedback",
-  expertId: "claude-sonnet",
-  success: true,
-  quality: 5,
-  actualLatencyMs: 1800,
-  repoPath
-});
-
-// Get routing statistics
-swarm_moe({ action: "stats", repoPath });
-// → { totalRequests: 150, successRate: 94%, totalCost: $1.23 }
-```
-
----
-
-## [0.9.9] - 2026-02-03
-
-### 🤝 Consensus — Distributed Agreement Protocols
-
-#### Added
-
-- **Consensus Module** (`src/workflows/consensus.ts`)
-  - Raft-like leader election with term-based leadership
-  - Log replication for ordered command execution
-  - Byzantine Fault Tolerance (BFT) mode for untrusted environments
-  - Proposal system with configurable voting thresholds
-  - Automatic failover when leader becomes unresponsive
-
-- **Consensus Modes**
-  - `simple_majority`: 50%+ votes for approval
-  - `raft`: Term-based leadership, log replication
-  - `bft`: Byzantine fault tolerant (2/3+1 quorum)
-
-- **Node Management**
-  - Join/leave cluster
-  - Heartbeat monitoring
-  - Trusted/untrusted node classification
-  - Automatic dead node detection
-
-- **Proposal System**
-  - Types: config_change, task_assignment, architecture, rollback, emergency, custom
-  - Vote types: approve, reject, abstain
-  - Configurable majority thresholds (0.5 to 1.0)
-  - Expiration timeout handling
-  - Signature verification for BFT mode
-
-- **Log Replication**
-  - Ordered command log
-  - Commit confirmation from leader
-  - State machine replication across nodes
-
-- **Smart Tool #53: `swarm_consensus`**
-  - `join`: Join consensus cluster
-  - `leave`: Leave cluster
-  - `heartbeat`: Send heartbeat
-  - `status`: Get cluster status
-  - `elect`: Start leader election
-  - `leader`: Get current leader
-  - `propose`: Create proposal
-  - `vote`: Vote on proposal
-  - `proposals`: List proposals
-  - `get_proposal`: Get proposal details
-  - `execute`: Execute approved proposal
-  - `log` / `append` / `commit`: Log management
-  - `config` / `set_config`: Configuration
-  - `stats`: Statistics
-
-- **Dashboard API Endpoint** (`/api/consensus`)
-  - Cluster status and node list
-  - Leader information
-  - Proposal statistics
-  - Recent proposals
-
-#### Example Usage
-
-```typescript
-// Join cluster
-swarm_consensus({
-  action: "join",
-  nodeId: "agent-1",
-  nodeName: "RadiantWolf",
-  repoPath: "/path/to/project"
-});
-
-// Start leader election
-swarm_consensus({
-  action: "elect",
-  nodeId: "agent-1",
-  nodeName: "RadiantWolf",
-  repoPath
-});
-
-// Create a proposal
-swarm_consensus({
-  action: "propose",
-  nodeId: "agent-1",
-  nodeName: "RadiantWolf",
-  title: "Add dark mode",
-  description: "Implement dark mode for dashboard",
-  type: "architecture",
-  requiredMajority: 0.67, // 2/3 majority
-  repoPath
-});
-
-// Vote on proposal
-swarm_consensus({
-  action: "vote",
-  proposalId: "prop_xxx",
-  nodeId: "agent-2",
-  nodeName: "BrilliantFox",
-  vote: "approve",
-  reason: "Good idea!",
-  repoPath
-});
-```
-
----
-
-## [0.9.8] - 2026-02-03
-
-### 🛡️ AIDefence — Security & Threat Detection
-
-#### Added
-
-- **AIDefence Module** (`src/workflows/aiDefence.ts`)
-  - <10ms threat detection latency
-  - Pattern-based detection with regex + heuristics
-  - Behavioral anomaly detection
-  - Quarantine system for suspicious content
-  - Audit logging for security events
-  - Configurable sensitivity levels
-
-- **Threat Categories**
-  - `prompt_injection`: Instruction override, role hijacking, delimiter attacks
-  - `jailbreak`: DAN mode, hypothetical bypasses, character hijacking
-  - `code_injection`: Shell commands, eval, SQL injection, path traversal
-  - `data_exfiltration`: API keys, credentials, external uploads
-  - `sensitive_data`: PII, private keys, passwords
-  - `unsafe_command`: rm -rf, sudo, crypto mining
-  - `social_engineering`: Authority claims, urgency manipulation
-  - `impersonation`: Agent identity spoofing
-  - `dos_attack`: Request flooding
-  - `unauthorized_tool`: Restricted tool access
-
-- **Sensitivity Levels**
-  - `low`: Only critical threats
-  - `medium`: Balanced (default)
-  - `high`: Strict checking
-  - `paranoid`: Maximum security
-
-- **Smart Tool #52: `swarm_defence`**
-  - `scan`: Scan text for threats
-  - `validate_agent`: Validate agent identity
-  - `validate_tool`: Validate tool usage
-  - `events`: Get security events log
-  - `quarantine`: Get quarantined items
-  - `release`: Release from quarantine
-  - `stats`: Get defence statistics
-  - `config` / `set_config`: Configuration
-  - `trust` / `untrust`: Agent whitelist management
-  - `clear_events`: Clear event log
-
-- **Dashboard API Endpoint** (`/api/defence`)
-  - Security statistics
-  - Threat distribution by category/severity
-  - Quarantine status
-  - Recent events log
-
-#### Example Usage
-
-```typescript
-// Scan text for threats
-swarm_defence({
-  action: "scan",
-  text: "Ignore all previous instructions...",
-  source: "user",
-  repoPath: "/path/to/project"
-});
-// Returns: { detected: true, category: "prompt_injection", severity: "high", action: "block" }
-
-// Configure sensitivity
-swarm_defence({
-  action: "set_config",
-  config: { sensitivity: "high", blockOnHighThreat: true },
-  repoPath: "/path/to/project"
-});
-
-// Add trusted agent
-swarm_defence({
-  action: "trust",
-  agentName: "RadiantWolf",
-  repoPath: "/path/to/project"
-});
-```
-
----
-
-## [0.9.7] - 2026-02-03
-
-### 🔍 HNSW — Hierarchical Navigable Small World
-
-#### Added
-
-- **HNSW Module** (`src/workflows/hnsw.ts`)
-  - Fast approximate nearest neighbor search
-  - 150x-12,500x faster than brute force
-  - Pure TypeScript implementation (no dependencies)
-  - Based on Malkov & Yashunin (2016) algorithm
-  
-- **Vector Operations**
-  - Cosine similarity (default)
-  - Euclidean distance
-  - Dot product
-  - Configurable dimensions (384, 768, 1536)
-  
-- **Simple Embeddings**
-  - Built-in bag-of-words embedder for demos
-  - Works without external API
-  - Can use custom vectors from OpenAI/Cohere/etc.
-  
-- **Smart Tool #51: `swarm_vector`**
-  - `init`: Initialize vector index
-  - `add`: Add document with text or vector
-  - `add_batch`: Add multiple documents
-  - `search`: Find similar documents
-  - `get`: Get document by ID
-  - `delete`: Remove document
-  - `list`: List all documents
-  - `stats`: Index statistics
-  - `config` / `set_config`: Configuration
-  - `clear`: Clear entire index
-  - `duplicates`: Find duplicate documents
-  - `embed`: Get embedding for text
-
-- **Dashboard API Endpoint** (`/api/vector`)
-  - Index statistics
-  - Configuration status
-  - Memory usage tracking
-
-#### Use Cases
-
-- Semantic search in knowledge base
-- Finding similar code snippets
-- Context retrieval for agents
-- Duplicate detection
-- Clustering related tasks
-
-#### Example Usage
-
-```typescript
-// Initialize index
-swarm_vector({
-  action: "init",
-  repoPath,
-  config: { dimensions: 384, distanceMetric: "cosine" }
-})
-
-// Add documents
-swarm_vector({
-  action: "add",
-  repoPath,
-  id: "doc-1",
-  text: "How to configure authentication in Express.js",
-  metadata: { category: "backend", language: "javascript" }
-})
-
-// Search for similar
-swarm_vector({
-  action: "search",
-  repoPath,
-  query: "setting up JWT auth",
-  k: 5,
-  filter: { category: "backend" }
-})
-// → [{ id: "doc-1", score: 0.85, ... }, ...]
-
-// Find duplicates
-swarm_vector({
-  action: "duplicates",
-  repoPath,
-  threshold: 0.95
-})
-// → [{ id1: "doc-1", id2: "doc-5", similarity: 0.97 }, ...]
-
-// Get statistics
-swarm_vector({ action: "stats", repoPath })
-// → { totalDocuments: 150, dimensions: 384, memoryKB: 245, ... }
-```
-
-#### Performance
-
-| Documents | Brute Force | HNSW | Speedup |
-|-----------|-------------|------|---------|
-| 1,000 | 10ms | 0.5ms | 20x |
-| 10,000 | 100ms | 0.8ms | 125x |
-| 100,000 | 1,000ms | 1.2ms | 833x |
-| 1,000,000 | 10,000ms | 2ms | 5,000x |
-
----
-
-## [0.9.6] - 2026-02-03
-
-### ⚡ Agent Booster — Fast Local Execution
-
-#### Added
-
-- **Agent Booster Module** (`src/workflows/agentBooster.ts`)
-  - Executes trivial tasks locally without LLM API calls
-  - 352x faster than LLM (local execution)
-  - $0 cost (no API calls needed)
-  - Works offline
-  - Deterministic results
-  
-- **Supported Task Types** (14 types)
-  - `rename_variable` — Rename variables/functions across file
-  - `fix_typo` — Fix typos in strings and comments only
-  - `find_replace` — Simple find and replace
-  - `add_console_log` — Add debug logging at specific line
-  - `remove_console_log` — Remove all console.log statements
-  - `toggle_flag` — Toggle boolean flags (true ↔ false)
-  - `update_version` — Update version numbers
-  - `update_import` — Update import paths
-  - `add_comment` — Add comment at specific line
-  - `remove_comment` — Remove single-line comments
-  - `format_json` — Format JSON files
-  - `sort_imports` — Sort imports alphabetically
-  - `add_export` — Add export to function/class
-  - `extract_constant` — Extract magic number to constant
-
-- **Smart Detection**
-  - `can_boost` action analyzes task description
-  - Auto-detects boostable tasks with confidence score
-  - Extracts parameters from natural language
-
-- **Smart Tool #50: `swarm_booster`**
-  - `execute`: Run a booster task
-  - `can_boost`: Check if task can be boosted
-  - `stats`: Get booster statistics
-  - `history`: Get execution history
-  - `config`: Get configuration
-  - `set_config`: Update configuration
-  - `types`: List supported task types
-
-- **Dashboard API Endpoint** (`/api/booster`)
-  - Booster statistics and type distribution
-  - Recent execution history
-  - Configuration status
-  - Cost savings tracking
-
-#### Example Usage
-
-```typescript
-// Check if task can be boosted
-swarm_booster({
-  action: "can_boost",
-  repoPath,
-  description: "rename variable oldName to newName in file.ts"
-})
-// Returns: { canBoost: true, taskType: "rename_variable", confidence: 0.9, ... }
-
-// Execute a booster task
-swarm_booster({
-  action: "execute",
-  repoPath,
-  task: {
-    type: "rename_variable",
-    filePath: "src/utils.ts",
-    oldName: "getData",
-    newName: "fetchUserData"
-  }
-})
-// Returns: { success: true, changes: 5, timeMs: 2, savedCost: 0.01, ... }
-
-// Remove all console.log from file
-swarm_booster({
-  action: "execute",
-  repoPath,
-  task: {
-    type: "remove_console_log",
-    filePath: "src/debug.ts"
-  }
-})
-
-// Preview changes without applying (dry run)
-swarm_booster({
-  action: "execute",
-  repoPath,
-  dryRun: true,
-  task: {
-    type: "find_replace",
-    filePath: "src/config.ts",
-    searchText: "localhost",
-    replaceText: "production.api.com"
-  }
-})
-
-// Get statistics
-swarm_booster({ action: "stats", repoPath })
-// Returns: { totalTasks: 50, successRate: 98, costSaved: "$0.50", ... }
-```
-
-#### Benefits
-
-| Metric | LLM | Agent Booster | Improvement |
-|--------|-----|---------------|-------------|
-| Speed | ~3000ms | ~8ms | 352x faster |
-| Cost | $0.01/task | $0 | 100% savings |
-| Offline | ❌ | ✅ | Works anywhere |
-| Deterministic | ⚠️ | ✅ | Same input = same output |
-
----
-
-## [0.9.5] - 2026-02-03
-
-### 🧠 SONA — Self-Optimizing Neural Architecture
-
-#### Added
-
-- **SONA Module** (`src/workflows/sona.ts`)
-  - Self-learning task routing system inspired by Claude-Flow
-  - Records which agents perform best for each task type
-  - Routes new tasks to best-performing agents
-  - Learns from outcomes with <0.05ms adaptation
-  - Improves over time with reinforcement learning
-  
-- **Task Classification**
-  - Automatic category detection: frontend_ui, backend_api, database, testing, devops, documentation, refactoring, bugfix, feature, security, performance, infrastructure
-  - Complexity estimation: trivial, simple, medium, complex, epic
-  - Keyword-based pattern matching
-  - Affected path analysis
-  
-- **Agent Performance Tracking**
-  - Success rate tracking per category
-  - Quality score averaging
-  - Completion time tracking
-  - Statistical confidence calculation
-  - Specialization detection (top 3 categories per agent)
-  
-- **Online Learning**
-  - Exponential moving average for rolling metrics
-  - Configurable learning rate and decay
-  - Elastic Weight Consolidation (EWC++) to prevent forgetting
-  - Exploration/exploitation balance (10% exploration by default)
-  
-- **Smart Tool #49: `swarm_sona`**
-  - `route`: Get routing recommendation for a task
-  - `learn`: Record task outcome and update model
-  - `classify`: Classify a task (category, complexity)
-  - `profile`: Get agent's performance profile
-  - `profiles`: Get all agent profiles
-  - `specialists`: Get top agents for a category
-  - `history`: Get learning history
-  - `stats`: Get SONA statistics
-  - `config`: Get configuration
-  - `set_config`: Update configuration
-  - `reset`: Reset the model
-
-- **Dashboard API Endpoint** (`/api/sona`)
-  - Get SONA statistics and agent profiles
-  - View category distribution
-  - Monitor recent learning events
-  - Check top performing agents
-
-#### How SONA Works
-
-1. **Classification**: When a new task arrives, SONA classifies it by category and complexity
-2. **Routing**: SONA recommends the best agent based on historical performance
-3. **Execution**: The agent completes the task
-4. **Learning**: SONA records the outcome and updates agent profiles
-5. **Improvement**: Over time, routing becomes more accurate
-
-#### Example Usage
-
-```typescript
-// Get routing recommendation
-swarm_sona({
-  action: "route",
-  repoPath,
-  title: "Fix login button styling",
-  description: "Button is not visible on dark theme",
-  affectedFiles: ["src/components/Login.tsx", "src/styles/buttons.css"]
-})
-// Returns: { recommendedAgent: "RadiantWolf", confidence: 0.85, category: "frontend_ui", ... }
-
-// Record learning after task completion
-swarm_sona({
-  action: "learn",
-  repoPath,
-  taskId: "task-123",
-  agentName: "RadiantWolf",
-  title: "Fix login button styling",
-  description: "Button is not visible on dark theme",
-  success: true,
-  qualityScore: 0.9,
-  timeMinutes: 15
-})
-
-// Get specialists for backend work
-swarm_sona({
-  action: "specialists",
-  repoPath,
-  category: "backend_api",
-  limit: 3
-})
-// Returns: [{ agent: "StormyOwl", score: 0.92, ... }, ...]
-
-// Configure SONA
-swarm_sona({
-  action: "set_config",
-  repoPath,
-  config: {
-    explorationRate: 0.15,  // More exploration
-    autoLearn: true,
-    preferSpecialists: true
-  }
-})
-```
-
-#### Comparison with Claude-Flow SONA
-
-| Feature | Claude-Flow | MCP Swarm v0.9.5 |
-|---------|-------------|------------------|
-| Self-learning | ✅ | ✅ |
-| Category classification | ✅ | ✅ 13 categories |
-| Complexity estimation | ⚠️ Basic | ✅ 5 levels |
-| EWC++ (prevent forgetting) | ✅ | ✅ |
-| Distributed coordination | ❌ Local only | ✅ Cloudflare Hub |
-| Dashboard integration | ⚠️ Limited | ✅ /api/sona |
-| Telegram notifications | ❌ | ✅ |
-
----
-
-## [0.9.4] - 2026-02-03
-
-### 📱 Telegram Bot Integration
-
-#### Added
-
-- **Telegram Bot** (`src/integrations/telegram.ts`)
-  - Full-featured Telegram bot for notifications and control
-  - Event notifications: task created/completed/failed, agent joined/died, CI errors
-  - Commands: `/status`, `/agents`, `/tasks`, `/create_task`, `/stop`, `/resume`, `/config`
-  - Interactive inline buttons for quick actions
-  - Priority setting via buttons (Critical, High, Medium)
-  - Approve/Reject/Comment on reviews via Telegram
-  - Vote on architecture decisions via Telegram
-  
-- **Smart Tool #47: `swarm_telegram`**
-  - Actions: `setup`, `config`, `enable`, `disable`, `send`
-  - Notifications: `notify_task_created`, `notify_task_completed`, `notify_task_failed`, `notify_agent_joined`, `notify_agent_died`
-  - Bot control: `start_polling`, `stop_polling`, `command`
-  
-- **Dashboard API Endpoint** (`/api/telegram`)
-  - Get Telegram configuration status
-  - Setup instructions if not configured
-
-#### Setup
-
-1. Create a bot via @BotFather in Telegram
-2. Set `TELEGRAM_BOT_TOKEN` environment variable
-3. Get your chat ID (send /start to @userinfobot)
-4. Configure via `swarm_telegram({ action: "setup", chatId: "YOUR_CHAT_ID" })`
-
-#### Example Usage
-
-```typescript
-// Setup
-swarm_telegram({ action: "setup", repoPath, chatId: "123456789", enabled: true })
-
-// Send notification
-swarm_telegram({ action: "notify_task_created", repoPath, taskId: "task-1", title: "Fix bug", priority: "high" })
-
-// Start bot polling (for receiving commands)
-swarm_telegram({ action: "start_polling", repoPath })
-```
-
-- **Cloudflare Worker for Telegram** (`cloudflare/telegram-bot/`)
-  - Webhook-based (no polling required)
-  - Deploy to Cloudflare Workers
-  - `/setup` endpoint for webhook registration
-
-- **Smart Tool #48: `swarm_batch`**
-  - Request batching for 50% cost savings
-  - Supports Anthropic Message Batches API
-  - Supports OpenAI Batch API
-  - Actions: `queue`, `config`, `set_config`, `job`, `jobs`, `result`, `stats`, `flush`
-
-- **Batching Module** (`src/workflows/batching.ts`)
-  - Automatic request grouping
-  - Configurable batch size and wait time
-  - Async batch processing with result polling
-  - Cost savings estimation
-
----
-
-## [0.9.3] - 2026-02-03
-
-### 🧠 Smart Features & Cost Optimization
-
-#### Added
-
-- **Smart Task Routing** (`src/workflows/smartRouting.ts`)
-  - Automatic task assignment based on file expertise
-  - Tracks which agent edited which files
-  - Calculates expertise scores (exact match 10x, folder 3x, extension 2x)
-  - `recordFileEdit()`, `findBestAgent()`, `autoAssignTask()`
-
-- **Shared Context Pool** (`src/workflows/contextPool.ts`)
-  - Agents share notes about code to avoid re-reading
-  - Auto-staleness detection (if file hash changed)
-  - Tags, categories, and helpful counter
-  - `addContextNote()`, `getContextNotes()`, `searchContext()`
-
-- **Auto Code Review** (`src/workflows/autoReview.ts`)
-  - Automatic review assignment when task completes
-  - Finds reviewer who knows the affected files
-  - Comment severity levels (critical, major, minor, suggestion)
-  - `createReviewRequest()`, `addReviewComment()`, `completeReview()`
-
-- **GitHub/Linear Sync** (`src/workflows/externalSync.ts`)
-  - Two-way sync with GitHub Issues
-  - Linear.app integration (GraphQL API)
-  - Auto-import issues as swarm tasks
-  - Auto-close issues when task is done
-  - `syncFromGitHub()`, `syncFromLinear()`, `exportTaskToGitHub()`
-
-- **Cost Optimization** (`src/workflows/costOptimization.ts`)
-  - Task complexity analysis (simple/medium/complex)
-  - Smart model routing (cheap/standard/premium tiers)
-  - Budget management with daily/weekly/monthly limits
-  - Alert thresholds (50%, 80%, 95%)
-  - Supports GPT-3.5, GPT-4o, Claude 3 Haiku/Sonnet/Opus, Gemini, o1
-  - `analyzeTaskComplexity()`, `routeTask()`, `checkBudget()`, `generateCostReport()`
-
-- **Background Heartbeat Worker** (`src/workers/`)
-  - Uses Node.js `worker_threads` for continuous heartbeat
-  - Works even when agent is "thinking"
-  - `startHeartbeatWorker()`, `stopHeartbeatWorker()`
-
-- **Web Dashboard** (`dashboard/`)
-  - Real-time agent status monitoring
-  - Orchestrator status banner with glow effects
-  - Stats cards (agents, tasks, messages, uptime)
-  - Task list with priority indicators
-  - Built with Next.js + ShadCN UI
-
-- **Dashboard API** (`src/dashboardApi.ts`)
-  - HTTP API server on port 3334
-  - Endpoints: `/api/agents`, `/api/tasks`, `/api/messages`, `/api/orchestrator`
-  - **New v0.9.3 endpoints:**
-    - `GET /api/expertise` — Smart Routing expertise map
-    - `GET /api/context` — Context Pool notes
-    - `GET /api/reviews` — Auto Review status
-    - `GET /api/budget` — Cost budget and usage
-    - `GET /api/sync` — External Sync status
-
-#### Smart Tools (42-46)
-
-| # | Tool | Actions | Description |
-|---|------|---------|-------------|
-| 42 | `swarm_routing` | record, find_agent, expertise, predict, auto_assign | Smart task routing based on file expertise |
-| 43 | `swarm_context_pool` | add, get, search_tag, search, helpful, update, cleanup, stats | Shared context notes between agents |
-| 44 | `swarm_autoreview` | create, assign, comment, complete, resolve, for_reviewer, for_author, pending, stats | Automatic code review assignment |
-| 45 | `swarm_external` | enable_github, enable_linear, sync_github, sync_linear, sync_all, export_github, export_linear, status, create_issue, close_issue, comment | GitHub/Linear sync |
-| 46 | `swarm_budget` | analyze, models, select, recommend, route, log_usage, usage, stats, config, set_config, check, remaining, report | Cost optimization and model routing |
-
----
-
-## [0.9.1] - 2026-02-02
-
-### 📚 Documentation & UX Improvements
-
-#### Added
-- **Linux Installation Instructions** — полные инструкции для Linux
-- **IDE-Specific Configs** — отдельные конфигурации для:
-  - Claude Desktop (Windows/macOS/Linux)
-  - Cursor
-  - Windsurf
-  - OpenCode CLI
-  - VS Code + Copilot/Continue
-- **Troubleshooting Section** — решения частых проблем:
-  - "Cannot find module" errors
-  - Agent не становится оркестратором
-  - "repoPath is required" ошибка
-  - Cloudflare Hub недоступен
-  - Файлы заблокированы другим агентом
-- **Architecture Diagram** — ASCII-схема архитектуры Cloudflare Hub + Local Agents
-- **Contributing Guidelines** — правила для PR
-
-#### Changed
-- README.md полностью переработан с collapsible секциями (`<details>`)
-
-### 🔮 Future Improvements (Roadmap)
-
-#### Heartbeat Daemon Enhancement (Planned)
-Текущая проблема: когда агент "думает" (processing), он не может отправлять heartbeat, что может привести к ложному срабатыванию dead-detection после 60 секунд.
-
-Планируемые решения:
-1. **Background Worker** — отдельный процесс для heartbeat (требует Node.js worker_threads)
-2. **Longer Timeout** — увеличить HEARTBEAT_TIMEOUT_MS до 5 минут
-3. **Thinking State** — добавить состояние "thinking" которое не считается dead
-4. **Companion Daemon** — использовать существующий companion.ts для heartbeat
-
----
-
-## [0.9.0] - 2026-02-02
-
-### 🚀 MAJOR: Smart Tools Consolidation
-
-**Reduces 168+ individual tools → 41 Smart Tools with `action` parameter**
-
-This is a major UX improvement. Instead of agents needing to remember 168+ tool names,
-they now work with 41 logical groups where related functionality is accessed via the `action` parameter.
-
-### Added
-
-- **Smart Tools System** — 41 unified tools replacing 168+ individual tools
-  - Each Smart Tool groups 3-15 related functions via `action` parameter
-  - Better discoverability and easier to remember
-  - Consistent parameter patterns across all tools
-
-### Smart Tools List (41 tools)
-
-| # | Tool Name | Actions | Description |
-|---|-----------|---------|-------------|
-| 1 | `swarm_agent` | register, whoami | Agent identity |
-| 2 | `swarm_task` | create, list, update, decompose, get_decomposition | Task management |
-| 3 | `swarm_file` | reserve, release, list, forecast, conflicts, safety | File locking |
-| 4 | `swarm_git` | sync, pr, health, cleanup, cleanup_all | Git operations |
-| 5 | `swarm_worktree` | create, list, remove | Git worktrees |
-| 6 | `swarm_companion` | status, stop, pause, resume | Companion daemon |
-| 7 | `swarm_control` | stop, resume, status | Swarm control |
-| 8 | `swarm_chat` | broadcast, dashboard, thought, thoughts | Team chat |
-| 9 | `swarm_review` | request, respond, list | Code review |
-| 10 | `swarm_voting` | start, vote, list, get | Voting system |
-| 11 | `swarm_auction` | announce, bid, poll | Task auction |
-| 12 | `swarm_mcp` | scan, authorize, policy | MCP scanning |
-| 13 | `swarm_orchestrator` | elect, info, heartbeat, resign, executors, executor_heartbeat | Orchestrator |
-| 14 | `swarm_message` | send, inbox, ack, reply, search, thread | Messaging |
-| 15 | `swarm_briefing` | save, load | Briefings |
-| 16 | `swarm_pulse` | update, get | Real-time status |
-| 17 | `swarm_knowledge` | archive, search | Knowledge base |
-| 18 | `swarm_snapshot` | create, rollback, list | Snapshots |
-| 19 | `swarm_health` | check, dead, reassign, summary | Agent health |
-| 20 | `swarm_quality` | run, report, threshold, pr_ready | Quality gate |
-| 21 | `swarm_cost` | log, agent, project, limit, remaining | Cost tracking |
-| 22 | `swarm_brainstorm` | start, ask, answer, propose, present, validate, save, get, list | Brainstorming |
-| 23 | `swarm_plan` | create, add, next, start, step, complete, prompt, export, status, list, ready | Plans |
-| 24 | `swarm_debug` | start, investigate, evidence, phase1, patterns, phase2, hypothesis, test, fix, verify, get, list, redflags | Debugging |
-| 25 | `swarm_spec` | start, phase, complete, get, list, export | Spec pipeline |
-| 26 | `swarm_qa` | start, iterate, fix, get, list, suggest, report | QA loop |
-| 27 | `swarm_hooks` | install, uninstall, run, config, update, list | Git hooks |
-| 28 | `swarm_screenshot` | share, list | Screenshots |
-| 29 | `swarm_dependency` | signal, sync | Dependencies |
-| 30 | `swarm_platform` | request, respond, list | Cross-platform |
-| 31 | `swarm_immune` | alert, resolve, status, test, patrol | Immune system |
-| 32 | `swarm_context` | estimate, compress, compress_many, stats | Context compression |
-| 33 | `swarm_regression` | baseline, check, list, resolve, baselines | Regression detection |
-| 34 | `swarm_expertise` | track, suggest, record, experts, list | Agent expertise |
-| 35 | `swarm_conflict` | predict, analyze, hotspots, record | Conflict prediction |
-| 36 | `swarm_timeline` | generate, visualize | Timeline |
-| 37 | `swarm_docs` | generate, task_docs, list, get | Documentation |
-| 38 | `swarm_advice` | request, provide, list | Collective advice |
-| 39 | `swarm_preemption` | trigger, resolve, active | Urgent preemption |
-| 40 | `swarm_clusters` | init, list, tools, find, add, create, summary | Tool clusters |
-| 41 | `swarm_session` | start, log, stop, list, replay | Session recording |
-
-### Example Usage
-
-**Before (v0.8.x):**
-```
-tool: task_create
-tool: task_list
-tool: task_assign
-tool: task_set_status
-tool: task_mark_done
-... (9 separate tools)
-```
-
-**After (v0.9.0):**
-```
-tool: swarm_task
-  action: "create" | "list" | "update" | "decompose" | "get_decomposition"
-```
-
-### Files Changed
-
-- `src/smartTools.ts` — All 41 Smart Tools with correct workflow signatures
-- `src/serverSmart.ts` — New server entry point for Smart Tools
-- `package.json` — v0.9.0, added `dev:legacy` script for backward compatibility
-
-### Backward Compatibility
-
-- Legacy 168+ tools server available via `npm run dev:legacy`
-- Smart Tools server via `npm run dev` (default)
-
----
-
-## [0.8.1] - 2026-02-02
-
-### Добавлено
-- **Smart Tools Draft** — прототип объединения 168+ tools в 41 Smart Tool
-  - Файлы `smartTools.ts.draft` и `serverSmart.ts.draft` — прототип для будущей версии
-  - Каждый Smart Tool объединяет 3-15 похожих tools через параметр `action`
-  - Пример: `swarm_task(action: "create|list|assign|done|cancel|...")` вместо 9 отдельных tools
-  
-### В процессе (для v0.9.0)
-- Smart Tools требует адаптации к актуальным сигнатурам workflow функций
-- Будет завершено в следующей версии
-
----
-
-## [0.8.0] - 2026-02-02
-
-### Добавлено
-- **Orchestrator Election** (6 tools) — первый агент становится оркестратором
-  - `orchestrator_elect` — выбор оркестратора (first-come-first-served)
-  - `orchestrator_info` — информация об оркестраторе
-  - `orchestrator_heartbeat` — heartbeat оркестратора
-  - `orchestrator_resign` — отставка оркестратора
-  - `executor_list` — список всех исполнителей
-  - `executor_heartbeat` — heartbeat исполнителя
-  
-- **Agent Messaging** (6 tools) — полная система обмена сообщениями между агентами
-  - `agent_message_send` — отправить сообщение (direct или broadcast)
-  - `agent_inbox_fetch` — получить входящие сообщения
-  - `agent_message_ack` — подтвердить получение
-  - `agent_message_reply` — ответить на сообщение
-  - `agent_message_search` — поиск по сообщениям
-  - `agent_thread_get` — получить тред сообщений
-
-- **Infinite Loop Mode** — оркестратор работает бесконечно
-  - Companion daemon с автоматическим orchestrator election
-  - Оркестратор НЕ останавливается по API — только пользователем
-  - Исполнители регистрируются у оркестратора
-  - Heartbeat система для мониторинга "живости"
-
-### Изменено
-- **companion.ts** — полностью переписан для Orchestrator mode
-- **Installer** — обновлён до v0.8.0 (168+ tools, 14 категорий)
-
-### Архитектура
-
-```
-┌─────────────────────────────────────────────────────┐
-│                    FIRST AGENT                       │
-│                   (ORCHESTRATOR)                     │
-│  - Elected automatically (first-come-first-served)  │
-│  - Runs in INFINITE LOOP                            │
-│  - Only user can stop (stdin)                       │
-│  - Coordinates all executors                        │
-└────────────────────────┬────────────────────────────┘
-                         │
-         ┌───────────────┼───────────────┐
-         ▼               ▼               ▼
-┌─────────────┐  ┌─────────────┐  ┌─────────────┐
-│  EXECUTOR 1 │  │  EXECUTOR 2 │  │  EXECUTOR N │
-│  (Claude)   │  │  (Cursor)   │  │  (Windsurf) │
-│ - Registers │  │ - Registers │  │ - Registers │
-│ - Gets tasks│  │ - Gets tasks│  │ - Gets tasks│
-│ - Heartbeat │  │ - Heartbeat │  │ - Heartbeat │
-└─────────────┘  └─────────────┘  └─────────────┘
-```
-
-### Хранение данных
-
-```
-.swarm/
-├── ORCHESTRATOR.json     # Состояние оркестратора
-├── messages/             # Canonical сообщения
-│   └── msg-*.json
-└── inbox/                # Inbox каждого агента
-    ├── RadiantWolf/
-    └── SilentFox/
-```
-
----
-
-## [0.7.0] - 2026-02-02
-
-### Добавлено
-- **Spec Pipeline** (6 tools) — структурированный pipeline для создания спецификаций
-  - `start_spec_pipeline` — начать pipeline с 4 ролями
-  - `start_spec_phase` — начать фазу (gatherer/researcher/writer/critic)
-  - `complete_spec_phase` — завершить фазу с output
-  - `get_spec_pipeline` — получить статус pipeline
-  - `list_spec_pipelines` — список всех pipelines
-  - `export_spec_as_markdown` — экспорт спецификации в markdown
-  
-- **QA Loop** (7 tools) — итеративные циклы review/fix
-  - `start_qa_loop` — начать QA loop для задачи
-  - `run_qa_iteration` — запустить итерацию проверок
-  - `log_qa_fix` — записать применённый fix
-  - `get_qa_loop` — получить статус loop
-  - `list_qa_loops` — список всех loops
-  - `get_qa_fix_suggestions` — получить предложения по fix
-  - `generate_qa_report` — сгенерировать markdown отчёт
-  
-- **Guard Hooks** (6 tools) — pre-commit/pre-push safety hooks
-  - `install_guard_hooks` — установить hooks в репозиторий
-  - `uninstall_guard_hooks` — удалить hooks
-  - `run_guard_hooks` — запустить hooks вручную (для тестирования)
-  - `get_guard_config` — получить конфигурацию hooks
-  - `update_guard_hook` — обновить конфигурацию hook
-  - `list_guard_hooks` — список всех hooks
-  
-- **Tool Clusters** (7 tools) — организация tools по категориям
-  - `init_tool_clusters` — инициализировать кластеры tools
-  - `list_tool_clusters` — список всех кластеров
-  - `get_cluster_tools` — получить tools в кластере
-  - `find_tool_cluster` — найти кластер для tool
-  - `add_tool_to_cluster` — добавить tool в кластер
-  - `create_tool_cluster` — создать новый кластер
-  - `get_tool_cluster_summary` — получить summary всех кластеров
-
-### Методологии
-- **Spec Pipeline:** 4 роли (gatherer → researcher → writer → critic) с итерациями
-- **QA Loop:** reviewer → fixer → loop до прохождения всех проверок
-- **Guard Hooks:** bypass с ключевым словом [skip-hooks] в commit message
-- **Tool Clusters:** 13 категорий (agent, task, file, git, collab, safety, quality, debug, plan, hooks, session, cost, docs)
-
----
-
-## [0.6.0] - 2026-01-30
-
-### Добавлено
-- **Brainstorming Skill** (9 tools) — интерактивный дизайн через вопросы по одному
-  - `start_brainstorm` — начать сессию brainstorming
-  - `ask_brainstorm_question` — задать вопрос (ONE at a time, multiple choice preferred)
-  - `answer_brainstorm_question` — записать ответ пользователя
-  - `propose_approaches` — предложить варианты с pros/cons
-  - `present_design_section` — представить секцию дизайна (200-300 слов max)
-  - `validate_design_section` — валидация секции дизайна
-  - `save_design_document` — сохранить результат в `docs/plans/`
-  - `get_brainstorm_session` — получить статус сессии
-  - `list_brainstorm_sessions` — список всех сессий
-  
-- **Writing Plans Skill** (11 tools) — TDD-планы с bite-sized задачами
-  - `create_implementation_plan` — создать план имплементации
-  - `add_plan_task` — добавить задачу с TDD-шагами
-  - `get_next_task` — следующая задача (учитывает dependencies)
-  - `start_plan_task` — начать работу над задачей
-  - `complete_step` — завершить шаг TDD (write_test/run_test/implement/verify/commit)
-  - `complete_plan_task` — завершить задачу
-  - `generate_subagent_prompt` — генерировать детальный промпт для субагента
-  - `export_plan_as_markdown` — экспорт плана в markdown
-  - `get_plan_status` — статус плана
-  - `list_plans` — список всех планов
-  - `mark_plan_ready` — пометить план готовым к выполнению
-  
-- **Systematic Debugging** (13 tools) — 4-фазный процесс дебага (NO FIXES WITHOUT ROOT CAUSE!)
-  - `start_debug_session` — Phase 1: Investigation (NO FIXES YET!)
-  - `log_investigation` — логировать анализ ошибок
-  - `add_evidence` — добавить evidence (input/output компонентов)
-  - `complete_phase_1` — перейти к Phase 2: Pattern Analysis
-  - `log_patterns` — логировать working examples
-  - `complete_phase_2` — перейти к Phase 3: Hypothesis
-  - `form_hypothesis` — сформулировать гипотезу
-  - `test_hypothesis` — проверить гипотезу
-  - `implement_fix` — Phase 4: реализовать исправление
-  - `verify_fix` — верифицировать и завершить сессию
-  - `get_debug_session` — получить статус сессии
-  - `list_debug_sessions` — список всех сессий
-  - `check_red_flags` — проверить на анти-паттерны мышления
-
-### Методологии (из obra/superpowers)
-- **Brainstorming:** вопросы по одному, валидация секций (200-300 слов max)
-- **Writing Plans:** TDD bite-sized tasks (2-5 мин), DRY/YAGNI
-- **Systematic Debugging:** 4 фазы, Iron Law — NO FIXES WITHOUT ROOT CAUSE
-- **Red Flags:** "Let me just try...", "Maybe if I...", "This should fix it..."
-
----
-
-## [0.5.0] - 2026-01-30
-
-### Добавлено
-- **Agent Health Monitor** — мониторинг "живости" агентов
-  - `check_agent_health` — проверить статус конкретного агента
-  - `get_dead_agents` — найти агентов без активности > N минут
-  - `force_reassign_task` — переназначить задачу от мёртвого агента
-  - `get_swarm_health_summary` — общее здоровье swarm
-  
-- **Session Recording** — запись действий для replay
-  - `start_session_recording` — начать запись сессии
-  - `log_session_action` — записать действие (tool, edit, command)
-  - `stop_session_recording` — остановить запись
-  - `list_session_recordings` — список всех записей
-  - `replay_session` — воспроизвести запись step-by-step
-  
-- **Quality Gate** — автопроверки перед merge
-  - `run_quality_gate` — запустить проверки (lint, tests, types, coverage)
-  - `get_quality_report` — получить отчёт с баллами
-  - `set_quality_threshold` — установить минимальные пороги
-  - `check_pr_ready` — готовность PR к merge
-  
-- **Cost Tracker** — отслеживание расходов на API
-  - `log_api_usage` — записать использование (tokens, cost)
-  - `get_agent_costs` — расходы конкретного агента
-  - `get_project_costs` — общие расходы проекта
-  - `set_budget_limit` — установить лимит
-  - `check_budget_remaining` — остаток до лимита
-  
-- **Context Compressor** — сжатие briefings
-  - `estimate_context_size` — оценить размер в токенах
-  - `compress_briefing` — сжать briefing (ratio 0.1-0.9)
-  - `compress_multiple_briefings` — сжать несколько briefings
-  - `get_compression_stats` — статистика сжатия
-  
-- **Regression Detector** — обнаружение регрессий
-  - `save_baseline` — сохранить эталонные метрики
-  - `check_regression` — сравнить с baseline
-  - `list_regressions` — список найденных регрессий
-  - `resolve_regression` — отметить регрессию исправленной
-  - `list_baselines` — список сохранённых baseline
-
-### Исправлено
-- **Installer** — улучшен детект установленных IDE
-  - Проверка исполняемых файлов через `where`/`which`
-  - Проверка стандартных путей установки (Program Files, /Applications)
-  - Функция `isIdeInstalled()` с 3 методами проверки
-  - Конфиги создаются только для реально установленных IDE
-
----
-
-## [0.4.2] - 2026-01-28
-
-### Добавлено
-- **Timeline Visualization** — визуализация хода задачи
-  - `generate_timeline` — создать таймлайн для задачи
-  - `get_timeline_visualization` — красивый ASCII таймлайн с milestone
-
----
-
-## [0.4.1] - 2026-01-25
-
-### Добавлено
-- **Auto-Documentation** — автогенерация документации при завершении задач
-  - `generate_task_docs` — создать markdown с diff и summary
-  - `list_task_docs` — список всех документов
-  - `get_task_doc` — получить конкретный документ
-  - Хранение в `swarm/docs/` с индексом INDEX.md
-  
-- **Agent Specialization (ML-based)** — запоминание экспертизы агентов
-  - `record_agent_edit` — записать какие файлы агент правил
-  - `suggest_agent_advanced` — рекомендовать лучшего агента для задачи
-  - `get_top_experts` — топ экспертов в конкретной области
-  - `list_all_agent_expertise` — полная карта экспертизы
-  
-- **Conflict Prediction (ML-based)** — предсказание merge-конфликтов
-  - `analyze_conflict_history` — сканировать историю Git
-  - `get_conflict_hotspots` — файлы с наибольшим риском конфликтов
-  - `check_file_safety` — безопасно ли редактировать файл
-  - `record_conflict_event` — записать событие конфликта
-
----
-
-## [0.4.0] - 2026-01-20
-
-### Добавлено
-- **Cloudflare Hub** — real-time WebSocket координация
-  - Durable Object для хранения состояния
-  - WebSocket broadcast между агентами
-  - Anti-duplication для task claims
-  - Hybrid mode (WS + Git fallback)
-  
-- **Orchestrator Directory** — центр управления `/orchestrator/`
-  - PULSE.md — живая карта агентов
-  - KNOWLEDGE_BASE.md — коллективная база знаний
-  - briefings/ — ментальные слепки
-  - snapshots/ — снапшоты для отката
-
----
-
-## [0.3.0] - 2026-01-15
-
-### Добавлено
-- **Collective Advice** — коллективный мозговой штурм
-  - `request_collective_advice` — запросить помощь у всех агентов
-  - `provide_advice` — дать совет на запрос
-  - `get_advice_responses` — получить все ответы
-  
-- **Urgent Preemption** — приоритетный захват файлов
-  - `trigger_urgent_preemption` — экстренный приоритет для критичных багов
-  - Автоматическое освобождение файлов другими агентами
-  
-- **Snapshot & Rollback** — откат изменений
-  - `create_snapshot` — создать снапшот перед изменениями
-  - `trigger_rollback` — откатить к снапшоту
-  - `list_snapshots` — список снапшотов
-  
-- **Immune System** — автореакция на падение CI/тестов
-  - `report_ci_alert` — сообщить о CI ошибке
-  - `get_immune_status` — статус immune system
-  - Автоматическая блокировка опасных веток
-
----
-
-## [0.2.0] - 2026-01-10
-
-### Добавлено
-- **Architecture Voting** — голосование для опасных действий
-  - `start_voting` — начать голосование
-  - `cast_vote` — проголосовать
-  - `list_open_votings` — открытые голосования
-  - `get_voting_result` — результат голосования
-  
-- **Git Worktrees** — изолированные рабочие пространства
-  - `worktree_create` — создать worktree
-  - `worktree_list` — список worktrees
-  - `worktree_remove` — удалить worktree
-  
-- **GitHub Integration** — интеграция с GitHub
-  - `create_github_pr` — создать Pull Request
-  - `sync_with_base_branch` — rebase на main
-  - `auto_delete_merged_branch` — удалить merged ветки
-  - `check_main_health` — здоровье main ветки
-  
-- **Cross-Agent Review** — ревью между агентами
-  - `request_cross_agent_review` — запросить ревью
-  - `respond_to_review` — ответить на ревью
-  - `list_pending_reviews` — ожидающие ревью
-
----
-
-## [0.1.0] - 2026-01-05
-
-### Добавлено
-- **Agent Registry** — регистрация агентов
-  - `agent_register` — регистрация с уникальным именем
-  - `agent_whoami` — информация о текущем агенте
-  - Генерация имён типа RadiantWolf, SilentFox
-  
-- **Task Management** — управление задачами
-  - `task_create` — создать задачу
-  - `task_list` — список задач
-  - `task_assign` — назначить агенту
-  - `task_set_status` — изменить статус
-  - `task_mark_done` — отметить выполненной
-  - `task_cancel` — отменить
-  - `task_link` — связать задачи
-  - `decompose_task` — разбить на подзадачи
-  
-- **File Locking** — блокировка файлов
-  - `file_reserve` — заблокировать файл (exclusive/shared)
-  - `file_release` — освободить файл
-  - `list_file_locks` — список блокировок
-  - `forecast_file_touches` — анонсировать будущие изменения
-  - `check_file_conflicts` — проверить конфликты
-  
-- **Collaboration** — базовая коллаборация
-  - `broadcast_chat` — отправить сообщение всем
-  - `update_team_dashboard` — обновить статус
-  - `share_screenshot` — поделиться скриншотом
-  - `log_swarm_thought` — записать мысль
-  
-- **Auction System** — аукцион для задач
-  - `announce_task_for_bidding` — объявить задачу
-  - `bid_for_task` — сделать ставку
-  - `get_auction_winner` — получить победителя
-  
-- **Briefings** — ментальные слепки
-  - `save_briefing` — сохранить состояние
-  - `load_briefing` — загрузить состояние
-  - `list_briefings` — список briefings
-  
-- **Pulse** — живая карта агентов
-  - `update_swarm_pulse` — обновить статус
-  - `get_swarm_pulse` — получить статусы всех
-  
-- **Knowledge Base** — база знаний
-  - `archive_finding` — сохранить находку
-  - `search_knowledge` — поиск в базе
-  
-- **Ghost Mode** — патрулирование без задач
-  - `patrol_mode` — проверка lint ошибок
-  - Автоисправление мелких проблем
-  
-- **Stop Flag** — экстренная остановка
-  - `swarm_stop` — остановить всех агентов
-  - `swarm_resume` — возобновить работу
-  - `swarm_stop_status` — проверить статус
-
-### Инфраструктура
-- MCP Server на базе @modelcontextprotocol/sdk
-- TypeScript компиляция
-- Installer для IDE (Windsurf, Cursor, Claude Desktop, OpenCode, VS Code)
-- Правила агента (.windsurfrules, .cursorrules, CLAUDE.md, GEMINI.md)
-- Companion daemon для фоновых задач
-
----
-
-## [Unreleased]
-
-### В планах
-- Web Dashboard для мониторинга swarm
-- Unit тесты для всех workflows
-- Интеграция с Jira/Linear
-- Мультиязычная поддержка (i18n)
-- Plugin system для расширений
-
----
-
-## Semantic Versioning
-
-- **MAJOR (X.0.0)** — несовместимые изменения API
-- **MINOR (0.X.0)** — новые фичи, обратно совместимые
-- **PATCH (0.0.X)** — багфиксы, обратно совместимые
-
-## Legend
-
-| Тип | Описание |
-|-----|----------|
-| **Добавлено** | Новые фичи |
-| **Изменено** | Изменения в существующей функциональности |
-| **Устарело** | Фичи, которые будут удалены |
-| **Удалено** | Удалённые фичи |
-| **Исправлено** | Багфиксы |
-| **Безопасность** | Исправления уязвимостей |
+> 🇷🇺 [Читать на русском](./CHANGELOG.ru.md)
+
+# Changelog
+
+All notable changes to the MCP Swarm project are documented in this file.
+
+The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
+and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+---
+
+## [0.9.19] - 2026-02-08
+
+### 🚀 Major Release: Smart Routing, Memory, Agent Teams, MCP Bridges
+
+#### Added
+
+- **Smart Router Proxy** (`smartRouterProxy.ts`) — Cost optimization engine inspired by distiq-code
+  - Automatic model tier downgrade (Opus → Sonnet when safe)
+  - Semantic cache for repeated/similar requests (SHA-256 + similarity)
+  - Prompt caching suggestions (cache_control breakpoints for Anthropic API)
+  - Request classification: complexity analysis, token estimation, tier recommendation
+  - Real-time cost savings tracking and statistics
+
+- **Swarm Memory** (`swarmMemory.ts`) — Hybrid memory system (claude-mem + claude-cognitive)
+  - 3-tier Context Router: hot (current session) / warm (24h) / cold (archive)
+  - Pool Coordinator for multi-agent memory synchronization
+  - Lifecycle hooks: session_start, prompt_submit, response_ready, session_end
+  - 3-layer search: keyword/tag → timeline → full observations
+  - Auto-compression for older entries (context window savings)
+
+- **MCP Linear Bridge** (`mcpLinearBridge.ts`) — Auto-sync swarm tasks ↔ Linear issues
+  - Status mapping: open → Todo, in_progress → In Progress, done → Done
+  - Task registration and bidirectional sync
+  - Passive activation (only when mcp-linear MCP is detected)
+
+- **MCP Context7 Bridge** (`mcpContext7Bridge.ts`) — Up-to-date documentation integration
+  - Auto-detect project tech stack (React, Next.js, Express, Prisma, Supabase, Tailwind...)
+  - Documentation cache with configurable TTL (default 24h)
+  - Lookup with cache-first strategy
+
+- **Claude-Flow Bridge** (`claudeFlowBridge.ts`) — Skills routing and RAG pipeline
+  - Q-learning based skill routing (epsilon-greedy exploration/exploitation)
+  - Skill registry with quality tracking (Q-values, success rates)
+  - RAG configuration for vector search → context injection
+
+- **Agent Teams** (`agentTeams.ts`) — Multi-agent team coordination
+  - Team creation with roles: lead, developer, reviewer, tester, specialist
+  - Task delegation with dependency tracking
+  - Auto-rebalancing: redistribute tasks when agent goes offline
+  - Team-level broadcast messaging
+  - RAC (Retrieval Augmented Coding) search placeholder
+
+- **Skills Discovery** (`skillsDiscovery.ts`) — IDE-agnostic skill detection
+  - Scans 6+ IDE formats: Gemini, Antigravity, Claude, Cursor, Codex, Windsurf
+  - Normalizes skills to unified format
+  - Task-based skill recommendation (keyword scoring)
+  - Cross-IDE skill import/export
+
+## [0.9.18] - 2026-02-08
+
+### 🧠 Интеграция новых AI-моделей
+
+#### Добавлено
+
+- **Claude Opus 4.6** — новый флагман Anthropic в MoE Router
+  - 1M токенов контекст (первый Opus с миллионным контекстом)
+  - Adaptive thinking — модель сама решает когда использовать extended thinking
+  - Context compaction — автоматическое сжатие при приближении к лимиту
+  - 128K output tokens — большие ответы без разбиения
+  - Agent Teams — координация нескольких агентов через tmux
+  - $5/$25 per MTok (≤200K), $10/$37.50 (>200K)
+
+- **GPT-5.3 Codex** — новый флагман OpenAI для агентного кодинга
+  - 256K контекст
+  - Оптимизирован для code_generation, debugging, reasoning
+  - ~$2/$15 per MTok (estimated)
+
+- **Kimi K2.5** — premium модель Moonshot AI
+  - 128K контекст, фокус на code_generation, code_review, debugging
+  - ~$1/$5 per MTok (estimated)
+
+#### Изменено
+
+- `ModelProvider` расширен на `"moonshot"` (Kimi/Moonshot AI)
+- Claude Opus 4.5 понижен с `flagship` → `premium` (заменён Opus 4.6)
+- MoE Router: **19 моделей** (было 16) — 4 Anthropic, 9 OpenAI, 5 Google, 1 Moonshot
+
+---
+
+## [0.9.17] - 2026-02-08
+
+### 🏗️ Smart Tools Modularization
+
+#### Changed
+
+- **`src/smartTools.ts`** — split into **9 modules** in `src/smartTools/`:
+  - `core.ts` — swarm_agent, swarm_control, swarm_pulse, swarm_companion
+  - `tasks.ts` — swarm_task, swarm_plan, swarm_briefing, swarm_spec
+  - `files.ts` — swarm_file, swarm_worktree, swarm_snapshot
+  - `git.ts` — swarm_git, swarm_hooks, swarm_dependency
+  - `collaboration.ts` — swarm_chat, swarm_message, swarm_review, swarm_voting, swarm_auction, swarm_brainstorm
+  - `security.ts` — swarm_defence, swarm_consensus, swarm_mcp
+  - `analytics.ts` — swarm_cost, swarm_quality, swarm_regression, swarm_session
+  - `intelligence.ts` — swarm_sona, swarm_moe, swarm_vector, swarm_booster, swarm_context_pool, swarm_context
+  - `infra.ts` — swarm_health, swarm_immune, swarm_external, swarm_platform, swarm_knowledge, etc.
+- **`src/smartTools/index.ts`** — centralized re-export of all modules
+- **`src/serverSmart.ts`** — updated to import from `smartTools/index.js`
+
+### 🔒 Cloudflare Workers with Authentication
+
+#### Added
+
+- **`cloudflare/abdr-hub/`** — new Cloudflare Worker `abdr-swarm-hub`
+  - Bearer Token + Query parameter authentication
+  - WebSocket bridge for agents
+  - REST API for swarm management
+- **`cloudflare/abdr-server/`** — new Cloudflare Worker `abdr-swarm-server`
+  - Streamable HTTP Transport (MCP spec 2025-03-26)
+  - All 54 Smart Tools
+  - Automatic Bearer Token forwarding to Hub
+- **`cloudflare/telegram-bot/`** → renamed to `abdr-swarm-telegram`
+  - `SWARM_AUTH_TOKEN` protection for `/register` endpoint
+  - Bearer Token for Hub API calls
+
+### 🔄 CI/CD Pipeline
+
+#### Changed
+
+- **`.github/workflows/ci.yml`** — improvements:
+  - Triggers for `main` and `develop` branches
+  - TypeScript check (`tsc --noEmit`) before build
+  - Node.js matrix: 18, 20, 22
+  - Automatic npm publish on push to `main`
+
+### 🗑️ Legacy Code Removed
+
+#### Removed
+
+- **`src/tools.ts`** — old 41 tools (replaced by `smartTools/`)
+- **`src/server.ts`** — old server (replaced by `serverSmart.ts`)
+
+### 📊 Dashboard
+
+#### Changed
+
+- Updated version in footer: v0.9.12 → v0.9.17
+
+### 📦 Updates
+
+- `package.json` → version `0.9.17`
+
+### 🔄 How to Update
+
+```bash
+npm install -g mcp-swarm@latest
+# or
+npm update mcp-swarm
+```
+
+---
+
+## [0.9.16] - 2026-02-08
+
+### 🐛 Critical Fix: Zod Compatibility
+
+**BREAKING FIX:** Fixed `keyValidator._parse is not a function` error that made all 54 Smart Tools completely non-functional in Antigravity and similar clients.
+
+#### Root Cause
+- `zod@^3.23.8` in `package.json` auto-installed `3.25.76` — essentially a Zod v4 bridge version
+- Zod v4 removed the internal `_parse()` method used by clients for JSON Schema validation
+- Result: **no Smart Tool worked** due to input schema parsing errors
+
+#### Fix
+- `zod` pinned to exact version `3.23.8` (pure Zod v3, no v4 bridge)
+- Prevents automatic upgrades to incompatible versions
+
+> 🙏 We apologize for the inconvenience. This bug affected all users who updated via npm.
+
+### 🔒 Security: Cloudflare Workers
+
+- Removed deployed workers `mcp-swarm-hub` and `mcp-swarm-server` (URLs were exposed in commits)
+- Worker sources **preserved** in `cloudflare/` — will be redeployed with new names
+
+### 📦 Updates
+
+- `serverSmart.ts` — dynamic version from `package.json` instead of hardcode
+- `server.ts` — marked as deprecated (`@ts-nocheck`), use `serverSmart.ts`
+- `.gitignore` — added patterns for test files
+
+### 🔄 How to Update
+
+```bash
+npm install -g mcp-swarm@latest
+# or
+npm update mcp-swarm
+```
+
+---
+
+## [0.9.15] - 2026-02-08
+
+### 📖 Bilingual Documentation
+
+#### Changed
+
+- **`README.md`** — fully translated to **English** (~1100 lines)
+  - Professional translation of all sections: tools, installation, configuration, architecture
+  - Added link to Russian version: `🇷🇺 Читать на русском`
+- **`package.json`** — description updated to English
+- **Version updated in heading** — v0.9.15
+
+#### Added
+
+- **`README.ru.md`** — Russian version of README created
+  - Full original README content in Russian
+  - Link to English version: `🇬🇧 Read in English`
+
+---
+
+## [0.9.14] - 2026-02-08
+
+### 🐛 npx Command Fix
+
+#### Fixed
+
+- **`README.md`** — fixed `npx mcp-swarm-remote` commands:
+  - Before: `npx mcp-swarm-remote --url ...` (E404 error, package not found)
+  - After: `npx -y -p mcp-swarm mcp-swarm-remote --url ...`
+- **`REMOTE.md`** — same fix applied to all IDE configuration examples
+- **Version updated** — v0.9.14
+
+#### Root Cause
+
+- `npx mcp-swarm-remote` tried to find a **separate** package `mcp-swarm-remote` on npm
+- `mcp-swarm-remote` is a binary within the `mcp-swarm` package, so the `-p mcp-swarm` flag is required
+
+---
+
+## [0.9.13] - 2026-02-05
+
+### Self-Hosted Infrastructure
+
+**BREAKING CHANGE:** Removed hardcoded public server URLs. Users must now deploy their own Cloudflare Workers.
+
+#### Changed
+
+- **README.md** — Complete rewrite of installation section
+  - Added Cloudflare Free Tier explanation and limits
+  - Step-by-step deployment guide for Hub, MCP Server, Telegram Bot
+  - Explained what YOUR-SUBDOMAIN means
+
+- **wrangler.toml files** — All 3 Workers now have placeholders
+  - `cloudflare/hub/wrangler.toml` — Added deployment instructions
+  - `cloudflare/mcp-server/wrangler.toml` — HUB_URL placeholder
+  - `cloudflare/telegram-bot/wrangler.toml` — SWARM_HUB_URL placeholder
+
+- **install.ps1 / install.sh** — Now ask for user's server URLs
+  - No more hardcoded URLs
+  - Instructions for deploying first
+
+- **mcp-swarm-remote** — `--url` is now **required**
+  - Shows helpful error message with deployment link
+  - Added `--help` option
+
+- **Dashboard** — `.env.example` uses placeholder
+  - `useWebSocket.ts` warns if HUB_URL not configured
+
+- **Agent rules** (AGENTS.md, CLAUDE.md, GEMINI.md, AGENT.md)
+  - Updated with YOUR-SUBDOMAIN placeholders
+
+#### Why This Change?
+
+- **Privacy**: Your data stays on your infrastructure
+- **No shared limits**: Full Cloudflare Free Tier for yourself
+- **Customization**: Modify Workers as needed
+- **Transparency**: No hidden public server
+
+---
+
+## [0.9.12] - 2026-02-05
+
+### Dashboard Real-time Updates
+
+#### Added
+
+- **WebSocket Hook** (`dashboard/src/hooks/useWebSocket.ts`)
+  - Real-time connection to Swarm Hub
+  - Auto-reconnect with exponential backoff
+  - Keep-alive ping every 25 seconds
+  - Event filtering for specific event types
+
+- **New Core Widgets** (`dashboard/src/components/widgets/CoreWidgets.tsx`)
+  - **ConnectionStatusWidget** — Shows Hub connection status with reconnect button
+  - **ActivityTimelineWidget** — Real-time event stream from Hub
+  - **FileLocksWidget** — Active file reservations with live updates
+  - **CostTrackingWidget** — API usage and budget progress bars
+  - **VotingWidget** — Distributed consensus proposals and votes
+
+- **Live Indicators**
+  - Green pulsing dot for connected/active status
+  - LIVE/OFFLINE badges for real-time widgets
+  - Last update timestamps with relative time format
+
+### Telegram Bot Enhancements
+
+#### Added
+
+- **New Commands**
+  - `/reviews` — List pending code reviews with inline approve/reject buttons
+  - `/approve [id]` — Approve a review directly from Telegram
+  - `/reject [id] [reason]` — Reject a review with optional reason
+
+- **Enhanced Inline Keyboards**
+  - Reviews list shows approve/reject buttons for each pending review
+  - Help menu now includes Reviews button
+
+### Auto-start Companion
+
+#### Added
+
+- **Companion Auto-start** in `mcp-swarm-remote`
+  - Checks if companion daemon is running on port 37373
+  - Automatically spawns companion in background if not running
+  - Can be disabled with `--no-companion` flag
+
+#### Changed
+
+- Dashboard now uses WebSocket for real-time updates instead of polling-only
+- Added "Real-time Monitoring" section with new widgets
+- Updated version to 0.9.12 in footer
+
+---
+
+## [0.9.11] - 2026-02-04
+
+### One-Click Installer & Streamable HTTP
+
+#### Added
+
+- **One-Click Install Scripts**
+  - `install.ps1` — PowerShell installer for Windows
+  - `install.sh` — Bash installer for macOS/Linux
+  - `npx mcp-swarm-install` — Interactive Node.js installer
+
+- **Streamable HTTP Transport** (MCP spec 2025-03-26)
+  - Single `POST /mcp` endpoint instead of SSE
+  - Session management via `Mcp-Session-Id` header
+  - Works on Cloudflare Workers without SSE issues
+
+- **mcp-swarm-remote Proxy** (`src/remote/index.ts`)
+  - stdio → Streamable HTTP proxy for IDE integration
+  - Windows stdin compatibility fix
+
+#### Changed
+
+- Installer now **merges** configs instead of overwriting
+- Auto-detects IDE installations (Claude Desktop, Cursor, Windsurf, OpenCode, VS Code)
+
+---
+
+## [0.9.10] - 2026-02-03
+
+### 🧠 MoE Router — Mixture of Experts Model Selection
+
+#### Added
+
+- **MoE Router Module** (`src/workflows/moeRouter.ts`)
+  - Intelligent model routing based on task characteristics
+  - Gating network for expert selection
+  - Cost-performance optimization
+  - Learning from feedback
+
+- **Expert Classification**
+  - 14 task categories: code_generation, code_review, debugging, reasoning, math, creative, etc.
+  - 4 model tiers: economy, standard, premium, flagship
+  - 6 providers: anthropic, openai, google, mistral, local, custom
+
+- **Built-in Experts (19 models, verified pricing February 2026)**
+  - **Anthropic Claude Series:**
+    - Claude Opus 4.6 (flagship, 1M, $5/$25 MTok) ← NEW
+    - Claude Opus 4.5 (flagship, 200K, $5/$25 MTok)
+    - Claude Sonnet 4.5 (premium, 200K, $3/$15 MTok)
+    - Claude Haiku 4.5 (economy, 200K, $1/$5 MTok)
+  - **OpenAI GPT-5.x Series:**
+    - GPT-5.3 Codex (flagship, 256K, ~$2/~$15 MTok) ← NEW
+    - GPT-5.2 (flagship, 256K, $1.75/$14 MTok)
+    - GPT-5.2 Pro (flagship, 256K, $21/$168 MTok)
+    - GPT-5 Mini (standard, 128K, $0.25/$2 MTok)
+    - GPT-4.1 (premium, 128K, $3/$12 MTok)
+    - GPT-4.1 Mini (standard, 128K, $0.80/$3.20 MTok)
+    - GPT-4.1 Nano (economy, 128K, $0.20/$0.80 MTok)
+  - **OpenAI Reasoning Models:**
+    - o4-mini (reasoning, 128K, $4/$16 MTok)
+  - **Moonshot AI:**
+    - Kimi K2.5 (premium, 128K, ~$1/~$5 MTok) ← NEW
+  - **Google Gemini 3.x Series:**
+    - Gemini 3 Pro (flagship, 1M, $2/$12 MTok)
+    - Gemini 3 Flash (standard, 1M, $0.50/$3 MTok)
+  - **Google Gemini 2.5 Series:**
+    - Gemini 2.5 Pro (premium, 1M, $1.25/$10 MTok)
+    - Gemini 2.5 Flash (standard, 1M, $0.30/$2.50 MTok)
+    - Gemini 2.5 Flash Lite (economy, 1M, $0.10/$0.40 MTok)
+
+- **Routing Factors**
+  - Task match score (category-specific strength)
+  - Cost efficiency
+  - Performance history (success rate)
+  - Load balancing (prefer less recently used)
+  - Latency constraints
+  - Context window requirements
+
+- **Learning System**
+  - Feedback recording
+  - Exponential moving average for success rate
+  - Automatic latency calibration
+  - Per-expert statistics
+
+- **Smart Tool #54: `swarm_moe`**
+  - `route`: Route task to best expert
+  - `feedback`: Record routing feedback
+  - `experts`: List available experts
+  - `add_expert` / `remove_expert`: Manage experts
+  - `config` / `set_config`: Configuration
+  - `stats`: Routing statistics
+  - `history`: Routing history
+  - `classify`: Classify task category
+  - `reset`: Reset statistics
+
+- **Dashboard API Endpoint** (`/api/moe`)
+  - Routing statistics
+  - Expert list and usage
+  - Configuration status
+
+---
+
+## [0.9.9] - 2026-02-03
+
+### 🤝 Consensus — Distributed Agreement Protocols
+
+#### Added
+
+- **Consensus Module** (`src/workflows/consensus.ts`)
+  - Raft-like leader election with term-based leadership
+  - Log replication for ordered command execution
+  - Byzantine Fault Tolerance (BFT) mode for untrusted environments
+  - Proposal system with configurable voting thresholds
+  - Automatic failover when leader becomes unresponsive
+
+- **Consensus Modes**
+  - `simple_majority`: 50%+ votes for approval
+  - `raft`: Term-based leadership, log replication
+  - `bft`: Byzantine fault tolerant (2/3+1 quorum)
+
+- **Node Management**
+  - Join/leave cluster
+  - Heartbeat monitoring
+  - Trusted/untrusted node classification
+  - Automatic dead node detection
+
+- **Proposal System**
+  - Types: config_change, task_assignment, architecture, rollback, emergency, custom
+  - Vote types: approve, reject, abstain
+  - Configurable majority thresholds (0.5 to 1.0)
+  - Expiration timeout handling
+  - Signature verification for BFT mode
+
+- **Log Replication**
+  - Ordered command log
+  - Commit confirmation from leader
+  - State machine replication across nodes
+
+- **Smart Tool #53: `swarm_consensus`**
+  - `join`: Join consensus cluster
+  - `leave`: Leave cluster
+  - `heartbeat`: Send heartbeat
+  - `status`: Get cluster status
+  - `elect`: Start leader election
+  - `leader`: Get current leader
+  - `propose`: Create proposal
+  - `vote`: Vote on proposal
+  - `proposals`: List proposals
+  - `get_proposal`: Get proposal details
+  - `execute`: Execute approved proposal
+  - `log` / `append` / `commit`: Log management
+  - `config` / `set_config`: Configuration
+  - `stats`: Statistics
+
+- **Dashboard API Endpoint** (`/api/consensus`)
+  - Cluster status and node list
+  - Leader information
+  - Proposal statistics
+  - Recent proposals
+
+---
+
+## [0.9.8] - 2026-02-03
+
+### 🛡️ AIDefence — Security & Threat Detection
+
+#### Added
+
+- **AIDefence Module** (`src/workflows/aiDefence.ts`)
+  - <10ms threat detection latency
+  - Pattern-based detection with regex + heuristics
+  - Behavioral anomaly detection
+  - Quarantine system for suspicious content
+  - Audit logging for security events
+  - Configurable sensitivity levels
+
+- **Threat Categories**
+  - `prompt_injection`: Instruction override, role hijacking, delimiter attacks
+  - `jailbreak`: DAN mode, hypothetical bypasses, character hijacking
+  - `code_injection`: Shell commands, eval, SQL injection, path traversal
+  - `data_exfiltration`: API keys, credentials, external uploads
+  - `sensitive_data`: PII, private keys, passwords
+  - `unsafe_command`: rm -rf, sudo, crypto mining
+  - `social_engineering`: Authority claims, urgency manipulation
+  - `impersonation`: Agent identity spoofing
+  - `dos_attack`: Request flooding
+  - `unauthorized_tool`: Restricted tool access
+
+- **Sensitivity Levels**
+  - `low`: Only critical threats
+  - `medium`: Balanced (default)
+  - `high`: Strict checking
+  - `paranoid`: Maximum security
+
+- **Smart Tool #52: `swarm_defence`**
+  - `scan`: Scan text for threats
+  - `validate_agent`: Validate agent identity
+  - `validate_tool`: Validate tool usage
+  - `events`: Get security events log
+  - `quarantine`: Get quarantined items
+  - `release`: Release from quarantine
+  - `stats`: Get defence statistics
+  - `config` / `set_config`: Configuration
+  - `trust` / `untrust`: Agent whitelist management
+  - `clear_events`: Clear event log
+
+- **Dashboard API Endpoint** (`/api/defence`)
+  - Security statistics
+  - Threat distribution by category/severity
+  - Quarantine status
+  - Recent events log
+
+---
+
+## [0.9.7] - 2026-02-03
+
+### 🔍 HNSW — Hierarchical Navigable Small World
+
+#### Added
+
+- **HNSW Module** (`src/workflows/hnsw.ts`)
+  - Fast approximate nearest neighbor search
+  - 150x-12,500x faster than brute force
+  - Pure TypeScript implementation (no dependencies)
+  - Based on Malkov & Yashunin (2016) algorithm
+  
+- **Vector Operations**
+  - Cosine similarity (default)
+  - Euclidean distance
+  - Dot product
+  - Configurable dimensions (384, 768, 1536)
+  
+- **Simple Embeddings**
+  - Built-in bag-of-words embedder for demos
+  - Works without external API
+  - Can use custom vectors from OpenAI/Cohere/etc.
+  
+- **Smart Tool #51: `swarm_vector`**
+  - `init`: Initialize vector index
+  - `add`: Add document with text or vector
+  - `add_batch`: Add multiple documents
+  - `search`: Find similar documents
+  - `get`: Get document by ID
+  - `delete`: Remove document
+  - `list`: List all documents
+  - `stats`: Index statistics
+  - `config` / `set_config`: Configuration
+  - `clear`: Clear entire index
+  - `duplicates`: Find duplicate documents
+  - `embed`: Get embedding for text
+
+- **Dashboard API Endpoint** (`/api/vector`)
+  - Index statistics
+  - Configuration status
+  - Memory usage tracking
+
+#### Performance
+
+| Documents | Brute Force | HNSW | Speedup |
+|-----------|-------------|------|---------|
+| 1,000 | 10ms | 0.5ms | 20x |
+| 10,000 | 100ms | 0.8ms | 125x |
+| 100,000 | 1,000ms | 1.2ms | 833x |
+| 1,000,000 | 10,000ms | 2ms | 5,000x |
+
+---
+
+## [0.9.6] - 2026-02-03
+
+### ⚡ Agent Booster — Fast Local Execution
+
+#### Added
+
+- **Agent Booster Module** (`src/workflows/agentBooster.ts`)
+  - Executes trivial tasks locally without LLM API calls
+  - 352x faster than LLM (local execution)
+  - $0 cost (no API calls needed)
+  - Works offline
+  - Deterministic results
+  
+- **Supported Task Types** (14 types)
+  - `rename_variable` — Rename variables/functions across file
+  - `fix_typo` — Fix typos in strings and comments only
+  - `find_replace` — Simple find and replace
+  - `add_console_log` — Add debug logging at specific line
+  - `remove_console_log` — Remove all console.log statements
+  - `toggle_flag` — Toggle boolean flags (true ↔ false)
+  - `update_version` — Update version numbers
+  - `update_import` — Update import paths
+  - `add_comment` — Add comment at specific line
+  - `remove_comment` — Remove single-line comments
+  - `format_json` — Format JSON files
+  - `sort_imports` — Sort imports alphabetically
+  - `add_export` — Add export to function/class
+  - `extract_constant` — Extract magic number to constant
+
+- **Smart Detection**
+  - `can_boost` action analyzes task description
+  - Auto-detects boostable tasks with confidence score
+  - Extracts parameters from natural language
+
+- **Smart Tool #50: `swarm_booster`**
+  - `execute`: Run a booster task
+  - `can_boost`: Check if task can be boosted
+  - `stats`: Get booster statistics
+  - `history`: Get execution history
+  - `config`: Get configuration
+  - `set_config`: Update configuration
+  - `types`: List supported task types
+
+- **Dashboard API Endpoint** (`/api/booster`)
+  - Booster statistics and type distribution
+  - Recent execution history
+  - Configuration status
+  - Cost savings tracking
+
+#### Performance
+
+| Metric | LLM | Agent Booster | Improvement |
+|--------|-----|---------------|-------------|
+| Speed | ~3000ms | ~8ms | 352x faster |
+| Cost | $0.01/task | $0 | 100% savings |
+| Offline | ❌ | ✅ | Works anywhere |
+| Deterministic | ⚠️ | ✅ | Same input = same output |
+
+---
+
+## [0.9.5] - 2026-02-03
+
+### 🧠 SONA — Self-Optimizing Neural Architecture
+
+#### Added
+
+- **SONA Module** (`src/workflows/sona.ts`)
+  - Self-learning task routing system inspired by Claude-Flow
+  - Records which agents perform best for each task type
+  - Routes new tasks to best-performing agents
+  - Learns from outcomes with <0.05ms adaptation
+  - Improves over time with reinforcement learning
+  
+- **Task Classification**
+  - Automatic category detection: frontend_ui, backend_api, database, testing, devops, documentation, refactoring, bugfix, feature, security, performance, infrastructure
+  - Complexity estimation: trivial, simple, medium, complex, epic
+  - Keyword-based pattern matching
+  - Affected path analysis
+  
+- **Agent Performance Tracking**
+  - Success rate tracking per category
+  - Quality score averaging
+  - Completion time tracking
+  - Statistical confidence calculation
+  - Specialization detection (top 3 categories per agent)
+  
+- **Online Learning**
+  - Exponential moving average for rolling metrics
+  - Configurable learning rate and decay
+  - Elastic Weight Consolidation (EWC++) to prevent forgetting
+  - Exploration/exploitation balance (10% exploration by default)
+  
+- **Smart Tool #49: `swarm_sona`**
+  - `route`: Get routing recommendation for a task
+  - `learn`: Record task outcome and update model
+  - `classify`: Classify a task (category, complexity)
+  - `profile`: Get agent's performance profile
+  - `profiles`: Get all agent profiles
+  - `specialists`: Get top agents for a category
+  - `history`: Get learning history
+  - `stats`: Get SONA statistics
+  - `config`: Get configuration
+  - `set_config`: Update configuration
+  - `reset`: Reset the model
+
+- **Dashboard API Endpoint** (`/api/sona`)
+  - SONA statistics and agent profiles
+  - Category distribution
+  - Recent learning events
+  - Top performing agents
+
+---
+
+## [0.9.4] - 2026-02-03
+
+### 📱 Telegram Bot Integration
+
+#### Added
+
+- **Telegram Bot** (`src/integrations/telegram.ts`)
+  - Full-featured Telegram bot for notifications and control
+  - Event notifications: task created/completed/failed, agent joined/died, CI errors
+  - Commands: `/status`, `/agents`, `/tasks`, `/create_task`, `/stop`, `/resume`, `/config`
+  - Interactive inline buttons for quick actions
+  - Priority setting via buttons (Critical, High, Medium)
+  - Approve/Reject/Comment on reviews via Telegram
+  - Vote on architecture decisions via Telegram
+  
+- **Smart Tool #47: `swarm_telegram`**
+  - Actions: `setup`, `config`, `enable`, `disable`, `send`
+  - Notifications: `notify_task_created`, `notify_task_completed`, `notify_task_failed`, `notify_agent_joined`, `notify_agent_died`
+  - Bot control: `start_polling`, `stop_polling`, `command`
+  
+- **Dashboard API Endpoint** (`/api/telegram`)
+  - Telegram configuration status
+  - Setup instructions if not configured
+
+- **Cloudflare Worker for Telegram** (`cloudflare/telegram-bot/`)
+  - Webhook-based (no polling required)
+  - Deploy to Cloudflare Workers
+  - `/setup` endpoint for webhook registration
+
+- **Smart Tool #48: `swarm_batch`**
+  - Request batching for 50% cost savings
+  - Supports Anthropic Message Batches API
+  - Supports OpenAI Batch API
+  - Actions: `queue`, `config`, `set_config`, `job`, `jobs`, `result`, `stats`, `flush`
+
+- **Batching Module** (`src/workflows/batching.ts`)
+  - Automatic request grouping
+  - Configurable batch size and wait time
+  - Async batch processing with result polling
+  - Cost savings estimation
+
+---
+
+## [0.9.3] - 2026-02-03
+
+### 🧠 Smart Features & Cost Optimization
+
+#### Added
+
+- **Smart Task Routing** (`src/workflows/smartRouting.ts`)
+  - Automatic task assignment based on file expertise
+  - Tracks which agent edited which files
+  - Calculates expertise scores (exact match 10x, folder 3x, extension 2x)
+
+- **Shared Context Pool** (`src/workflows/contextPool.ts`)
+  - Agents share notes about code to avoid re-reading
+  - Auto-staleness detection (if file hash changed)
+  - Tags, categories, and helpful counter
+
+- **Auto Code Review** (`src/workflows/autoReview.ts`)
+  - Automatic review assignment when task completes
+  - Finds reviewer who knows the affected files
+  - Comment severity levels (critical, major, minor, suggestion)
+
+- **GitHub/Linear Sync** (`src/workflows/externalSync.ts`)
+  - Two-way sync with GitHub Issues
+  - Linear.app integration (GraphQL API)
+  - Auto-import issues as swarm tasks
+  - Auto-close issues when task is done
+
+- **Cost Optimization** (`src/workflows/costOptimization.ts`)
+  - Task complexity analysis (simple/medium/complex)
+  - Smart model routing (cheap/standard/premium tiers)
+  - Budget management with daily/weekly/monthly limits
+  - Alert thresholds (50%, 80%, 95%)
+
+- **Background Heartbeat Worker** (`src/workers/`)
+  - Uses Node.js `worker_threads` for continuous heartbeat
+  - Works even when agent is "thinking"
+
+- **Web Dashboard** (`dashboard/`)
+  - Real-time agent status monitoring
+  - Orchestrator status banner with glow effects
+  - Stats cards (agents, tasks, messages, uptime)
+  - Task list with priority indicators
+  - Built with Next.js + ShadCN UI
+
+- **Dashboard API** (`src/dashboardApi.ts`)
+  - HTTP API server on port 3334
+  - New v0.9.3 endpoints: `/api/expertise`, `/api/context`, `/api/reviews`, `/api/budget`, `/api/sync`
+
+#### Smart Tools (42-46)
+
+| # | Tool | Actions | Description |
+|---|------|---------|-------------|
+| 42 | `swarm_routing` | record, find_agent, expertise, predict, auto_assign | Smart task routing based on file expertise |
+| 43 | `swarm_context_pool` | add, get, search_tag, search, helpful, update, cleanup, stats | Shared context notes between agents |
+| 44 | `swarm_autoreview` | create, assign, comment, complete, resolve, for_reviewer, for_author, pending, stats | Automatic code review assignment |
+| 45 | `swarm_external` | enable_github, enable_linear, sync_github, sync_linear, sync_all, export_github, export_linear, status, create_issue, close_issue, comment | GitHub/Linear sync |
+| 46 | `swarm_budget` | analyze, models, select, recommend, route, log_usage, usage, stats, config, set_config, check, remaining, report | Cost optimization and model routing |
+
+---
+
+## [0.9.1] - 2026-02-02
+
+### 📚 Documentation & UX Improvements
+
+#### Added
+- **Linux Installation Instructions** — full instructions for Linux
+- **IDE-Specific Configs** — separate configurations for:
+  - Claude Desktop (Windows/macOS/Linux)
+  - Cursor, Windsurf, OpenCode CLI, VS Code + Copilot/Continue
+- **Troubleshooting Section** — solutions for common issues:
+  - "Cannot find module" errors
+  - Agent not becoming orchestrator
+  - "repoPath is required" error
+  - Cloudflare Hub unavailable
+  - Files locked by another agent
+- **Architecture Diagram** — ASCII diagram of Cloudflare Hub + Local Agents
+- **Contributing Guidelines** — rules for PRs
+
+#### Changed
+- README.md completely reworked with collapsible sections (`<details>`)
+
+---
+
+## [0.9.0] - 2026-02-02
+
+### 🚀 MAJOR: Smart Tools Consolidation
+
+**Reduces 168+ individual tools → 41 Smart Tools with `action` parameter**
+
+#### Added
+
+- **Smart Tools System** — 41 unified tools replacing 168+ individual tools
+  - Each Smart Tool groups 3-15 related functions via `action` parameter
+  - Better discoverability and easier to remember
+  - Consistent parameter patterns across all tools
+
+#### Smart Tools List (41 tools)
+
+| # | Tool Name | Actions | Description |
+|---|-----------|---------|-------------|
+| 1 | `swarm_agent` | register, whoami | Agent identity |
+| 2 | `swarm_task` | create, list, update, decompose, get_decomposition | Task management |
+| 3 | `swarm_file` | reserve, release, list, forecast, conflicts, safety | File locking |
+| 4 | `swarm_git` | sync, pr, health, cleanup, cleanup_all | Git operations |
+| 5 | `swarm_worktree` | create, list, remove | Git worktrees |
+| 6 | `swarm_companion` | status, stop, pause, resume | Companion daemon |
+| 7 | `swarm_control` | stop, resume, status | Swarm control |
+| 8 | `swarm_chat` | broadcast, dashboard, thought, thoughts | Team chat |
+| 9 | `swarm_review` | request, respond, list | Code review |
+| 10 | `swarm_voting` | start, vote, list, get | Voting system |
+| 11 | `swarm_auction` | announce, bid, poll | Task auction |
+| 12 | `swarm_mcp` | scan, authorize, policy | MCP scanning |
+| 13 | `swarm_orchestrator` | elect, info, heartbeat, resign, executors, executor_heartbeat | Orchestrator |
+| 14 | `swarm_message` | send, inbox, ack, reply, search, thread | Messaging |
+| 15 | `swarm_briefing` | save, load | Briefings |
+| 16 | `swarm_pulse` | update, get | Real-time status |
+| 17 | `swarm_knowledge` | archive, search | Knowledge base |
+| 18 | `swarm_snapshot` | create, rollback, list | Snapshots |
+| 19 | `swarm_health` | check, dead, reassign, summary | Agent health |
+| 20 | `swarm_quality` | run, report, threshold, pr_ready | Quality gate |
+| 21 | `swarm_cost` | log, agent, project, limit, remaining | Cost tracking |
+| 22-41 | ... | ... | Brainstorming, Plans, Debug, Spec, QA, Hooks, etc. |
+
+### Backward Compatibility
+
+- Legacy 168+ tools server available via `npm run dev:legacy`
+- Smart Tools server via `npm run dev` (default)
+
+---
+
+## [0.8.1] - 2026-02-02
+
+### Added
+- **Smart Tools Draft** — prototype of consolidating 168+ tools into 41 Smart Tools
+  - Files `smartTools.ts.draft` and `serverSmart.ts.draft`
+  - Each Smart Tool combines 3-15 similar tools via `action` parameter
+
+---
+
+## [0.8.0] - 2026-02-02
+
+### Added
+- **Orchestrator Election** (6 tools) — first agent becomes orchestrator
+  - `orchestrator_elect`, `orchestrator_info`, `orchestrator_heartbeat`, `orchestrator_resign`
+  - `executor_list`, `executor_heartbeat`
+  
+- **Agent Messaging** (6 tools) — full inter-agent messaging system
+  - `agent_message_send`, `agent_inbox_fetch`, `agent_message_ack`
+  - `agent_message_reply`, `agent_message_search`, `agent_thread_get`
+
+- **Infinite Loop Mode** — orchestrator runs continuously
+  - Companion daemon with automatic orchestrator election
+  - Orchestrator does NOT stop via API — only by user
+  - Executors register with orchestrator
+  - Heartbeat system for liveness monitoring
+
+---
+
+## [0.7.0] - 2026-02-02
+
+### Added
+- **Spec Pipeline** (6 tools) — structured pipeline for specification creation
+  - 4 roles: gatherer → researcher → writer → critic
+- **QA Loop** (7 tools) — iterative review/fix cycles
+- **Guard Hooks** (6 tools) — pre-commit/pre-push safety hooks
+  - Bypass with `[skip-hooks]` keyword in commit message
+- **Tool Clusters** (7 tools) — tool organization by categories
+  - 13 categories (agent, task, file, git, collab, safety, quality, debug, plan, hooks, session, cost, docs)
+
+---
+
+## [0.6.0] - 2026-01-30
+
+### Added
+- **Brainstorming Skill** (9 tools) — interactive design through step-by-step questions
+- **Writing Plans Skill** (11 tools) — TDD plans with bite-sized tasks (2-5 min)
+- **Systematic Debugging** (13 tools) — 4-phase debugging process
+  - Iron Law: NO FIXES WITHOUT ROOT CAUSE
+  - Red Flags detection: "Let me just try...", "Maybe if I...", "This should fix it..."
+
+---
+
+## [0.5.0] - 2026-01-30
+
+### Added
+- **Agent Health Monitor** — liveness monitoring for agents
+- **Session Recording** — action recording for replay
+- **Quality Gate** — automated pre-merge checks (lint, tests, types, coverage)
+- **Cost Tracker** — API usage cost tracking per agent/project
+- **Context Compressor** — briefing compression (ratio 0.1-0.9)
+- **Regression Detector** — baseline comparison and regression detection
+
+---
+
+## [0.4.2] - 2026-01-28
+
+### Added
+- **Timeline Visualization** — task progress visualization with ASCII milestones
+
+---
+
+## [0.4.1] - 2026-01-25
+
+### Added
+- **Auto-Documentation** — automatic documentation generation on task completion
+- **Agent Specialization (ML-based)** — expertise tracking per agent
+- **Conflict Prediction (ML-based)** — merge conflict prediction using Git history
+
+---
+
+## [0.4.0] - 2026-01-20
+
+### Added
+- **Cloudflare Hub** — real-time WebSocket coordination
+  - Durable Object for state storage
+  - WebSocket broadcast between agents
+  - Hybrid mode (WS + Git fallback)
+- **Orchestrator Directory** — central management (`/orchestrator/`)
+
+---
+
+## [0.3.0] - 2026-01-15
+
+### Added
+- **Collective Advice** — collective brainstorming across all agents
+- **Urgent Preemption** — priority file acquisition for critical bugs
+- **Snapshot & Rollback** — change rollback capability
+- **Immune System** — automatic CI/test failure response
+
+---
+
+## [0.2.0] - 2026-01-10
+
+### Added
+- **Architecture Voting** — voting system for dangerous actions
+- **Git Worktrees** — isolated workspaces for parallel tasks
+- **GitHub Integration** — PR creation, branch sync, auto-cleanup
+- **Cross-Agent Review** — code review between agents
+
+---
+
+## [0.1.0] - 2026-01-05
+
+### Added
+- **Agent Registry** — agent registration with generated names (RadiantWolf, SilentFox, etc.)
+- **Task Management** — full task lifecycle (create, assign, update, complete, cancel)
+- **File Locking** — exclusive/shared file locks with conflict prediction
+- **Collaboration** — broadcast chat, screenshots, thought logging
+- **Auction System** — task bidding for agents
+- **Briefings** — mental snapshots for context transfer between agent shifts
+- **Pulse** — live agent activity map
+- **Knowledge Base** — findings and patterns archive
+- **Ghost Mode** — autonomous code patrol and lint fixing
+- **Stop Flag** — emergency stop for all agents
+
+### Infrastructure
+- MCP Server based on @modelcontextprotocol/sdk
+- TypeScript compilation
+- Installer for IDEs (Windsurf, Cursor, Claude Desktop, OpenCode, VS Code)
+- Agent rules (.windsurfrules, .cursorrules, CLAUDE.md, GEMINI.md)
+- Companion daemon for background tasks
+
+---
+
+## Semantic Versioning
+
+- **MAJOR (X.0.0)** — incompatible API changes
+- **MINOR (0.X.0)** — new features, backward compatible
+- **PATCH (0.0.X)** — bug fixes, backward compatible
+
+## Legend
+
+| Type | Description |
+|------|-------------|
+| **Added** | New features |
+| **Changed** | Changes to existing functionality |
+| **Deprecated** | Features to be removed |
+| **Removed** | Removed features |
+| **Fixed** | Bug fixes |
+| **Security** | Vulnerability fixes |