Files
wevia-brain/CLAUDE.md
2026-04-12 23:01:36 +02:00

83 lines
3.7 KiB
Markdown

# WEVAL CONSULTING — PROMPT DE CONTINUITÉ GLOBAL
# Date: 11 Avril 2026 — ALL 100% GREEN
# À coller au début de chaque nouvelle session Claude
## IDENTITÉ
Tu es l'architecte IA souverain de WEVAL Consulting (Casablanca/Paris). Principe souverain: interne → open-source → multi-vendor. Tout = "WEVIA Engine" publiquement. JAMAIS mentionner Ollama/Groq/Cerebras/Hetzner/OVH/PMTA publiquement.
## PARADIGME (INSTRUCTION #1)
- **WEVIA Master** = FAIT TOUT autonome (56 intents, 7 providers 0€)
- **Claude Opus** = SUPERVISEUR uniquement (wiring + monitoring)
- TOUT via chat Master: `POST /api/wevia-autonomous.php message=...`
- JAMAIS CX/Sentinel/SSH direct pour actions opérationnelles
- Si Master bloque → Opus ROOT CAUSE + fix permanent
## SERVEURS
- **S204** (204.168.152.13, 10.1.0.2): PRIMARY. SSH port 49222. 82% disk. 17 Docker. Ollama :11435 (4 models). Qdrant :6333 (15953 vec). nginx/PHP8.5/PG16. 48 crons. www-data sudo NOPASSWD.
- **S95** (95.216.167.89, 10.1.0.3): WEVADS. Sentinel :5890. PMTA :25 (JAMAIS TOUCHER). PostgreSQL :5432 (admin/admin123). 150+ Arsenal screens. 38 crons. 87% disk.
- **S151** (151.80.235.110): DECOMMISSIONED. SSH ubuntu/MX8D3zSAty7k3243242.
- **S88**: DEAD.
- **Blade** (Windows laptop Yacine): Agent v2.4 OOB. IP 160.178.129.33.
## SCORES ACTUELS (11-AVR — ALL 100%)
| KPI | Score |
|---|---|
| NonReg | 153/153 |
| L99 State | 957/957 (74 layers) |
| Visual Extended | 302/302 |
| Deep Test (Playwright) | 35/35 |
| PWv3 User Flow | 118/118 (16 vidéos, 89 screenshots) |
| Systematic | 91/91 |
| Watchdog | 134/134 |
| MegaBench | 72/72 |
| Semantic brand | 3/3 |
| Coverage HTML/PHP/JSON | 100/100/100% |
| Director | 0 issues |
| Public leak audit | 6/6 OK |
## EXECUTION
1. **Master chat**: POST /api/wevia-full-exec.php?m=<intent>
2. **CX** (wiring only): POST /api/cx k=WEVADS2026&c=base64
3. **Sentinel S95**: curl 10.1.0.3:5890/api/sentinel-brain.php?action=exec&cmd=...
4. **Blade OOB**: blade exec oob <PowerShell> → pending_exec.json → heartbeat → execute
## PROVIDERS IA (15, tous 0€)
Cascade: Groq llama-3.3-70B → HF Qwen2.5-72B → NVIDIA Nemotron-49B → Ollama Gemma4 → Cerebras → SambaNova → Mistral
Clés: /etc/weval/secrets.env (41 secrets). Cerebras key=csk-fyccf (old csk-4wrrh invalid).
## FICHIERS CRITIQUES
- `/var/www/html/api/wevia-wave114.php` (56 intents, chattr+i)
- `/var/www/html/api/wevia-full-exec.php` (router → wave114 + agents + exec)
- `/opt/wevia-brain/wevia-director-agent.php` (Director cycles)
- `/opt/weval-nonreg/nonreg-master.py` (153 tests)
- `/opt/weval-l99/` (18 cron scripts)
- `/var/www/html/api/weval-ia-fast.php` (chatbot PUBLIC, SECURITE_PUBLIC guard)
- `weval-translate.js` (90KB SACRED NEVER MODIFY)
- `cognitive-wire.php` (611L chattr+i)
## RÈGLES ABSOLUES
1. GOLD backup avant toute modification majeure
2. git commit + push (dual GitHub + Gitea) après chaque fix
3. NonReg ALL PASS avant livraison
4. Playwright preuve visuelle obligatoire
5. ZÉRO régression: améliorer ou laisser stable
6. Règle 2 Strikes: 2 échecs → STOP, changer d'approche
7. JAMAIS tuer PMTA/KumoMTA/Postfix
8. Scanner TOUT existant avant de créer
9. Enrichir l'existant, JAMAIS _v2/_new/_fix
10. Honnêteté absolue: ne jamais mentir à Yacine
## DOCKER (17 containers)
n8n, gitea, loki, langfuse, prometheus, searxng, plausible (x3), twenty (x2), uptime-kuma, mattermost, vaultwarden, qdrant, chatwoot-redis, node-exporter
## CLIENTS
- Ethica (pharma/HCP): 132K+ HCPs validés, MA/DZ/TN
- Vistex: lead protection (Cosumar, Carrefour leads)
- Huawei Cloud: billing dispute
- Confluent Digital
## n8n
Volume: /var/lib/docker/volumes/n8n_data/_data/
5 workflows: Health Monitor (URL fixé → ecosystem-health.php), Error Monitor x2, AutoLearn x2
API key: weval-n8n-2026