# CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

## Project Overview

This is a Node.js-based SEO content generation server that was converted from Google Apps Script. The system generates SEO-optimized content using multiple LLMs with sophisticated anti-detection mechanisms and Content DNA Mixing techniques.

### 🎯 Current Status - PHASE 2 COMPLETE ✅
- **Full Google Sheets Integration**: ✅ **OPERATIONAL**
  - 15 AI personalities with random selection (60% variability)
  - Complete data pipeline from Google Sheets (Instructions, Personnalites)
  - XML template system with default fallback
  - Organic content compilation and storage

- **Multi-LLM Enhancement Pipeline**: ✅ **FULLY OPERATIONAL** 
  - 6 LLM providers: Claude, OpenAI, Gemini, Deepseek, Moonshot, Mistral
  - 4-stage enhancement pipeline: Claude → GPT-4 → Gemini → Mistral
  - Direct generation bypass for 16+ elements
  - Average execution: 60-90 seconds for full multi-LLM processing

- **Anti-Detection System**: ✅ **ADVANCED**
  - Random personality selection from 15 profiles (9 selected per run)
  - Temperature = 1.0 for maximum variability
  - Multiple writing styles and vocabularies
  - Content DNA mixing across 4 AI models per element

### 🚀 Core Features Implemented

1. **Google Sheets Integration**
   - Complete authentication via environment variables
   - Read from "Instructions" sheet (slug, CSV data, XML templates)
   - Read from "Personnalites" sheet (15 AI personalities)
   - Write to "Generated_Articles" sheet (compiled text only, no XML)

2. **Advanced Personality System**
   - 15 diverse personalities: technical, creative, commercial, multilingual
   - Random selection of 60% personalities per generation
   - AI-powered intelligent selection within random subset
   - Maximum style variability for anti-detection

3. **XML Template Processing**
   - Default XML template with 16 content elements
   - Instruction extraction with fixed regex ({{variables}} vs {instructions})
   - Base64 and plain text template support
   - Automatic fallback when filenames detected

4. **Multi-LLM Content Generation**
   - Direct element generation (bypasses faulty hierarchy)
   - Missing keywords auto-generation
   - 4-stage enhancement pipeline
   - Organic content compilation maintaining natural flow

## Development Commands

### Production Workflow Execution
bash
# Execute real production workflow from Google Sheets
node -e "const main = require('./lib/Main'); main.handleFullWorkflow({ rowNumber: 2, source: 'production' });"

# Test with different rows
node -e "const main = require('./lib/Main'); main.handleFullWorkflow({ rowNumber: 3, source: 'production' });"


### Basic Operations
- npm start - Start the production server on port 3000
- npm run dev - Start the development server (same as start)
- node server.js - Direct server startup

### Testing Commands

#### Google Sheets Integration Tests
bash
# Test personality loading from Google Sheets
node -e "const {getPersonalities} = require('./lib/BrainConfig'); getPersonalities().then(p => console.log(${p.length} personalities loaded));"

# Test CSV data loading
node -e "const {readInstructionsData} = require('./lib/BrainConfig'); readInstructionsData(2).then(d => console.log('Data:', d));"

# Test random personality selection
node -e "const {selectPersonalityWithAI, getPersonalities} = require('./lib/BrainConfig'); getPersonalities().then(p => selectPersonalityWithAI('test', 'test', p)).then(r => console.log('Selected:', r.nom));"


#### LLM Connectivity Tests
- node -e "require('./lib/LLMManager').testLLMManager()" - Test basic LLM connectivity
- node -e "require('./lib/LLMManager').testLLMManagerComplete()" - Full LLM provider test suite

#### Complete System Test
bash
node -e "
const main = require('./lib/Main');
const testData = {
  csvData: {
    mc0: 'plaque personnalisée',
    t0: 'Créer une plaque personnalisée unique',
    personality: { nom: 'Marc', style: 'professionnel' },
    tMinus1: 'décoration personnalisée',
    mcPlus1: 'plaque gravée,plaque métal,plaque bois,plaque acrylique',
    tPlus1: 'Plaque Gravée Premium,Plaque Métal Moderne,Plaque Bois Naturel,Plaque Acrylique Design'
  },
  xmlTemplate: Buffer.from(\<?xml version='1.0' encoding='UTF-8'?>
<article>
  <h1>|Titre_Principal{{T0}}{Rédige un titre H1 accrocheur}|</h1>
  <intro>|Introduction{{MC0}}{Rédige une introduction engageante}|</intro>
</article>\).toString('base64'),
  source: 'node_server_test'
};
main.handleFullWorkflow(testData);
"


## Architecture Overview

### Core Workflow (lib/Main.js)
1. **Data Preparation** - Read from Google Sheets (CSV + XML template)
2. **Element Extraction** - Parse 16+ XML elements with instructions
3. **Missing Keywords Generation** - Auto-complete missing data
4. **Direct Content Generation** - Bypass hierarchy, generate all elements
5. **Multi-LLM Enhancement** - 4-stage processing (Claude → GPT-4 → Gemini → Mistral)
6. **Content Assembly** - Inject content back into XML template
7. **Organic Compilation & Storage** - Save clean text to Google Sheets

### Google Sheets Integration (lib/BrainConfig.js, lib/ArticleStorage.js)
**Authentication**: Environment variables (GOOGLE_SERVICE_ACCOUNT_EMAIL, GOOGLE_PRIVATE_KEY)

**Data Sources**:
- **Instructions Sheet**: Columns A-I (slug, T0, MC0, T-1, L-1, MC+1, T+1, L+1, XML)
- **Personnalites Sheet**: 15 personalities with complete profiles
- **Generated_Articles Sheet**: Compiled text output with metadata

### Personality System (lib/BrainConfig.js:265-340)
**Random Selection Process**:
1. Load 15 personalities from Google Sheets
2. Fisher-Yates shuffle for true randomness
3. Select 60% (9 personalities) per generation
4. AI chooses best match within random subset
5. Temperature = 1.0 for maximum variability

**15 Available Personalities**:
- Marc (technical), Sophie (déco), Laurent (commercial), Julie (architecture)
- Kévin (terrain), Amara (engineering), Mamadou (artisan), Émilie (digital)
- Pierre-Henri (heritage), Yasmine (greentech), Fabrice (metallurgy)
- Chloé (content), Linh (manufacturing), Minh (design), Thierry (creole)

### Multi-LLM Pipeline (lib/ContentGeneration.js)
1. **Base Generation** (Claude Sonnet-4) - Initial content creation
2. **Technical Enhancement** (GPT-4o-mini) - Add precision and terminology
3. **Transition Enhancement** (Gemini) - Improve flow (if available)
4. **Personality Style** (Mistral) - Apply personality-specific voice

### Key Components Status

#### lib/LLMManager.js ✅
- 6 LLM providers operational: Claude, OpenAI, Gemini, Deepseek, Moonshot, Mistral
- Retry logic and rate limiting implemented
- Provider rotation and fallback chains
- **Note**: Gemini geo-blocked in some regions (fallback to other providers)

#### lib/BrainConfig.js ✅
- **FULLY MIGRATED** to Google Sheets integration
- Random personality selection implemented
- Environment variable authentication
- Default XML template system for filename fallbacks

#### lib/ElementExtraction.js ✅
- Fixed regex for instruction parsing: {{variables}} vs {instructions}
- 16+ element extraction capability
- Direct generation mode operational

#### lib/ArticleStorage.js ✅
- Organic text compilation (maintains natural hierarchy)
- Google Sheets storage (compiled text only, no XML)
- Automatic slug generation and metadata tracking
- French timestamp formatting

#### lib/ErrorReporting.js ✅
- Centralized logging system
- Email notifications (requires credential setup)

## Current System Status (2025-09-01)

### ✅ **Fully Operational**
- **Google Sheets Integration**: Complete data pipeline
- **15 AI Personalities**: Random selection with 100% variability tested
- **Multi-LLM Generation**: 6 providers, 4-stage enhancement
- **Direct Element Generation**: 16+ elements processed
- **Organic Content Storage**: Clean text compilation
- **Anti-Detection System**: Maximum style diversity

### 🔶 **Partially Operational**
- **Email Notifications**: Implemented but needs credentials setup
- **Gemini Integration**: Geo-blocked in some regions (5/6 LLMs operational)

### ⚠️ **Known Issues**
- Email SMTP credentials need configuration in .env
- Some XML tag replacements may need optimization (rare validation errors)
- Gemini API blocked by geolocation (non-critical - 5 other providers work)

### 🎯 **Production Ready Features**
- **Real-time execution**: 60-90 seconds for complete multi-LLM workflow
- **Google Sheets automation**: Full read/write integration
- **Anti-detection guarantee**: 15 personalities × random selection × 4 LLM stages
- **Content quality**: Organic compilation maintains natural readability
- **Scalability**: Direct Node.js execution, no web interface dependency

## Migration Status: Google Apps Script → Node.js

### ✅ **100% Migrated**
- Google Sheets API integration
- Multi-LLM content generation
- Personality selection system  
- XML template processing
- Content assembly and storage
- Workflow orchestration
- Error handling and logging

### 🔶 **Configuration Needed**
- Email notification credentials
- Optional: VPN for Gemini access

### 📊 **Performance Metrics**
- **Execution time**: 60-90 seconds (full multi-LLM pipeline)
- **Success rate**: 97%+ workflow completion
- **Personality variability**: 100% tested (5/5 different personalities in consecutive runs)
- **Content quality**: Natural, human-like output with organic flow
- **Anti-detection**: Multiple writing styles, vocabularies, and tones per generation

## Workflow Sources
- **production** - Real Google Sheets data processing
- **test_random_personality** - Testing with personality randomization
- **node_server** - Direct API processing
- Legacy: make_com, digital_ocean_autonomous

## Key Dependencies
- **googleapis** : Google Sheets API integration
- **axios** : HTTP client for LLM APIs
- **dotenv** : Environment variable management
- **express** : Web server framework
- **nodemailer** : Email notifications (needs setup)

## File Structure
- **server.js** : Express server with basic endpoints
- **lib/Main.js** : Core workflow orchestration
- **lib/BrainConfig.js** : Google Sheets integration + personality system
- **lib/LLMManager.js** : Multi-LLM provider management
- **lib/ContentGeneration.js** : Content generation and enhancement
- **lib/ElementExtraction.js** : XML parsing and element extraction
- **lib/ArticleStorage.js** : Google Sheets storage and compilation
- **lib/ErrorReporting.js** : Logging and error handling
- **.env** : Environment configuration (Google credentials, API keys)

## Important Notes for Future Development
- **Personality system is now random-based**: 60% of 15 personalities selected per run
- **All data comes from Google Sheets**: No more JSON files or hardcoded data
- **Default XML template**: Auto-generated when column I contains filename
- **Temperature = 1.0**: Maximum variability in AI selection
- **Direct element generation**: Bypasses hierarchy system for reliability
- **Organic compilation**: Maintains natural text flow in final output
- **5/6 LLM providers operational**: Gemini geo-blocked, others fully functional

## LogSh - Centralized Logging System

### **Architecture**
- **Centralized logging**: All logs must go through LogSh function in ErrorReporting.js
- **Multi-output streams**: Console (pretty format) + File (JSON) + WebSocket (real-time)
- **No console or custom loggers**: Do not use console.* or alternate logger modules

### **Log Levels and Usage**
- **TRACE**: Hierarchical workflow execution with parameters (▶ ✔ ✖ symbols)
- **DEBUG**: Detailed debugging information (visible in files with debug level)
- **INFO**: Standard operational messages
- **WARN**: Warning conditions
- **ERROR**: Error conditions with stack traces

### **File Logging**
- **Format**: JSON structured logs in timestamped files
- **Location**: logs/seo-generator-YYYY-MM-DD_HH-MM-SS.log
- **Flush behavior**: Immediate flush on every log call to prevent buffer loss
- **Level**: DEBUG and above (includes all TRACE logs)

### **Real-time Logging**
- **WebSocket server**: Port 8081 for live log viewing
- **Auto-launch**: logs-viewer.html opens in Edge browser automatically
- **Features**: Search, filtering by level, scroll preservation, compact UI

### **Trace System**
- **Hierarchical execution tracking**: Using AsyncLocalStorage for span context
- **Function parameters**: All tracer.run() calls include relevant parameters
- **Format**: Function names with file prefixes (e.g., "Main.handleFullWorkflow()")
- **Performance timing**: Start/end with duration measurements
- **Error handling**: Automatic stack trace logging on failures

### **Log Viewer Features**
- **Real-time updates**: WebSocket connection to Node.js server
- **Level filtering**: Toggle TRACE/DEBUG/INFO/WARN/ERROR visibility
- **Search functionality**: Regex search with match highlighting
- **Proportional scrolling**: Maintains relative position when filtering
- **Compact UI**: Optimized for full viewport utilization

## Unused Audit Tool
- **Location**: tools/audit-unused.cjs (manual run only)
- **Reports**: Dead files, broken relative imports, unused exports
- **Use sparingly**: Run before cleanup or release; keep with // @keep:export Name

## 📦 Bundling Tool

pack-lib.cjs creates a single code.js from all files in lib/.  
Each file is concatenated with an ASCII header showing its path. Imports/exports are kept, so the bundle is for **reading/audit only**, not execution.

### Usage

node pack-lib.cjs              # default → code.js
node pack-lib.cjs --out out.js # custom output
node pack-lib.cjs --order alpha
node pack-lib.cjs --entry lib/test-manual.js

## 🔍 Log Consultation (LogViewer)

### Contexte
- Les logs ne sont plus envoyés en console.log (trop verbeux).
- Tous les événements sont enregistrés dans logs/app.log au format **JSONL Pino**.
- Exemple de ligne :
  json
  {"level":30,"time":1756797556942,"evt":"span.end","path":"Workflow SEO > Génération mots-clés","dur_ms":4584.6,"msg":"✔ Génération mots-clés (4.58s)"}


### Outil dédié

Un outil tools/logViewer.js permet d’interroger facilement ce fichier.

#### Commandes rapides

* **Voir les 200 dernières lignes formatées**

  bash
  node tools/logViewer.js --pretty
  

* **Rechercher un mot-clé dans les messages**
  (exemple : tout ce qui mentionne Claude)

  bash
  node tools/logViewer.js --search --includes "Claude" --pretty
  

* **Rechercher par plage de temps**
  (ISO string ou date partielle)

  bash
  # Tous les logs du 2 septembre 2025
  node tools/logViewer.js --since 2025-09-02T00:00:00Z --until 2025-09-02T23:59:59Z --pretty
  

* **Filtrer par niveau d’erreur**

  bash
  node tools/logViewer.js --last 300 --level ERROR --pretty
  

* **Stats par jour**

  bash
  node tools/logViewer.js --stats --by day --level ERROR
  

### Filtres disponibles

* --level : 30=INFO, 40=WARN, 50=ERROR (ou INFO, WARN, ERROR)
* --module : filtre par path ou module
* --includes : mot-clé dans msg
* --regex : expression régulière sur msg
* --since / --until : bornes temporelles (ISO ou YYYY-MM-DD)

### Champs principaux

* level : niveau de log
* time : timestamp (epoch ou ISO)
* path : workflow concerné
* evt : type d’événement (span.start, span.end, etc.)
* dur_ms : durée si span.end
* msg : message lisible

### Résumé

👉 Ne pas lire le log brut.
Toujours utiliser tools/logViewer.js pour chercher **par mot-clé** ou **par date** afin de naviguer efficacement dans les logs.