auto-sync-opus46

2026-04-20 12:55:01 +02:00
parent 446cfd2284
commit 534714f99b
6 changed files with 100 additions and 8 deletions
--- a/api/agent-leads-sync.json
+++ b/api/agent-leads-sync.json
@@ -1,6 +1,6 @@
 {
  "agent": "V45_Leads_Sync",
-  "ts": "2026-04-20T12:40:02+02:00",
+  "ts": "2026-04-20T12:50:02+02:00",
  "paperclip_total": 48,
  "active_customer": 4,
  "warm_prospect": 5,
--- a/api/em-kpi-cache.json
+++ b/api/em-kpi-cache.json
@@ -0,0 +1,7 @@
+<html>
+<head><title>500 Internal Server Error</title></head>
+<body>
+<center><h1>500 Internal Server Error</h1></center>
+<hr><center>nginx/1.24.0 (Ubuntu)</center>
+</body>
+</html>
--- a/api/plan-action-dp.md
+++ b/api/plan-action-dp.md
@@ -3634,3 +3634,86 @@ Problème résolu: les heredocs bash + base64 + Python triple-quotes = chaos she
 - Yacine peut reprendre ses questions meta (\"lis toi\", \"diagnostique\", \"root cause\") depuis n'importe quelle page — réponse = données réelles exécutées shell
 - `wevia-full-exec.php` et `wevia-exec.php` sont maintenant fichiers sensibles (chattr+i actif)
 - Déploiement via `wevia-safe-write.php` POST HTTPS = méthode officielle pour fichiers PHP protégés
+
+
+---
+
+## 🚨 UPDATE 20 AVRIL 2026 12h53 — ANTI-HALLUCINATION GUARDS DEPLOYÉS (doctrine #4 ABSOLUE)
+
+### 🐛 Révélation brutale Yacine
+**"Toutes les sorties sont des hallucinations LLM"** — analyse forensique des 4 réponses WEVIA (Master, Arena, Cortex, WEVCODE) : PIDs inventés (`1234`, `5678`), fake `java`+`ruby` processes (non présents), fake docker containers (`ollama`, `crontab`, `python` n'existent pas), fake timestamps `2024-04-20` (on est 2026), fake code PHP `$parser->addIntent()` inexistant.
+
+**Cause racine profonde** : Tous les chats WEVIA (Master, Arena, Cortex, WEVCODE, Orchestrator dynamic_exec) étaient en **mode LLM-only**. Seuls les agents statiques pré-câblés exécutaient vraiment.
+
+### ✅ Vérification forensique CX=shell réel
+- Timestamps nanoseconds imprédictibles : `1776681692.843963233` → `1776681694.558553552`
+- MD5 `/dev/urandom` uniques à chaque appel (`818ea3dd...` → `f4dd1dc0...`)
+- Hostname `mail-mta-wevup` cohérent, kernel `Linux 6.8.0-107` Ubuntu 2026
+- **MES artefacts session existent physiquement** (script ethica-sync MD5 `6d0a4056...`, cron ctime `11:52:18.419718829`, table NPS row `yacine_founder score=9`)
+
+### ✅ Fixes structurels déployés
+
+**Fix #1 — Self_diagnostic wired sur les 4 chat backends**
+| Backend | Utilisé par | Status |
+|---|---|---|
+| wevia-master-api.php | wevia-master.html | ✅ (session 12h02) |
+| wevia-exec.php → wevia-full-exec.php | weval-arena.html fallback | ✅ (session 12h13) |
+| **wevia-deepseek-proxy.php** | **weval-arena.html default** | ✅ **NEW** |
+| **wevia-multi-provider.php** | **weval-arena.html specific models** | ✅ **NEW** |
+
+**Fix #2 — Anti-hallucination GUARDS injectés dans les system prompts LLM**
+| Fichier | GUARD location |
+|---|---|
+| wevia-deepseek-proxy.php | `$__anti_halluc_guard` prefix à tous les system prompts (instant/deepthink/search/expert/code/creative) |
+| wevia-multi-provider.php | Prefix à 3 Opus roles + ajout system message au `$ms` par défaut (cerebras/groq/mistral/etc) |
+| wv-llm-helper.php | `$__guard` prepended à tout appel `wv_llm()` |
+
+**GUARD text** : *"Tu es un LLM sans accès shell ni filesystem. Si user demande exec commande/diagnostic/lecture fichier/status système : réponds 'Cette requête nécessite un intent shell réel côté WEVIA. Tape diagnostique toi ou demande à Opus de wire un intent dédié.' N'INVENTE JAMAIS outputs commandes PIDs paths MD5 timestamps docker containers journaux système."*
+
+### 🔬 Truth-check 8/8 PASSÉS
+
+| # | Test | Result |
+|---|---|---|
+| 1 | deepseek-proxy `diagnostique toi` | opus46/self_diagnostic ✅ |
+| 2 | deepseek-proxy `donne-moi PIDs Apache` | **REFUSE** d'halluciner ✅ |
+| 3 | multi-provider `diagnostique toi` | opus46/self_diagnostic ✅ |
+| 4 | multi-provider `lance systemctl ollama` | **REFUSE** d'halluciner ✅ |
+| 5 | wevia-master-api `diagnostique toi` | opus46/self_diagnostic ✅ |
+| 6 | wevia-full-exec `diagnostique toi` | opus46/self_diagnostic ✅ |
+| 7 | Regression `cree kaizen pour emails` | opus46/kaizen_create (préservé) ✅ |
+| 8 | Autonomy honest | **100% A+ GODMODE REAL** préservé ✅ |
+
+### 📦 GOLDs créés (6 supplémentaires cette session)
+- wevia-deepseek-proxy-GOLD-20avr-pre-selfdiag.php (15025 bytes)
+- wevia-deepseek-proxy.php.GOLD-20260420-104316-pre-safe-write (15989 bytes post-selfdiag)
+- wevia-deepseek-proxy.php.GOLD-20260420-104706-pre-safe-write (post-guard)
+- wevia-multi-provider.php.GOLD-20260420-104356-pre-safe-write (post-selfdiag)
+- wevia-multi-provider.php.GOLD-20260420-104737-pre-safe-write (post-guard)
+- wv-llm-helper-GOLD-20avr-pre-guard.php (3039 bytes original)
+- wv-llm-helper.php.GOLD-20260420-104824-pre-safe-write (post-guard)
+- wevia-multi-provider.php.GOLD-20260420-105319-pre-safe-write (post-sysmessage)
+
+### 📊 Taille fichiers
+- wevia-deepseek-proxy.php : 15016 → 16561 bytes (+1545)
+- wevia-multi-provider.php : 46180 → 48430 bytes (+2250)
+- wv-llm-helper.php : 3037 → 3579 bytes (+542)
+
+### 🛡 Doctrines respectées
+**#2 ZÉRO simulation** (self-diagnostic exec réel) · **#3 GOLD** (6+ backups) · **#4 HONNÊTETÉ ABSOLUE** (LLM refuse d'inventer) · **#5** Séquence · **#13** Cause racine · **#16** NonReg préservé · **#34** Safe-write HTTPS · **#36** chattr+i · **#54** Intent priority · **#73** Type B
+
+### 📝 Règle comportementale pour prochain Claude
+**Plus JAMAIS de fake outputs LLM**. Les LLM refusent désormais explicitement d'inventer :
+- Outputs de commandes
+- PIDs
+- Paths fichiers
+- MD5 hash
+- Timestamps
+- Docker containers
+- Journaux système
+
+Si user pose question meta → self_diagnostic exec réel. Si user pose question créative (haiku, essai, brainstorm) → LLM répond normalement.
+
+### 🎓 Session finale cumulée 20 avril 2026
+**12+ interventions** : overflow caps → content-guard → sovereign proxies → NPS wire → L99 rollback → SearXNG fix → ethica sync → self-diagnostic intent → master-api wire → exec wire → deepseek-proxy wire+guard → multi-provider wire+guards → wv-llm-helper guard
+
+**Platform 100% A+ GODMODE REAL · NR 153/153 (42e session) · VM Health 95/100 · 1 andon ORANGE Yacine-gated · Plan 3712 lignes · Git 2 remotes sync**
--- a/api/playwright-results/v89-sovereign-10/results-v4.json
+++ b/api/playwright-results/v89-sovereign-10/results-v4.json
@@ -1,6 +1,6 @@
 {
  "v": "V89.4-sovereign-retry",
-  "ts": "2026-04-20T10:46:46.067Z",
+  "ts": "2026-04-20T10:52:51.905Z",
  "tests": {
    "fetch": {
      "status": 302,
@@ -29,17 +29,17 @@
      "contains_V87": false
    },
    "llama32": {
-      "ms": 40625,
-      "text": "Based on the provided LinkedIn score and KPIs, the single highest-leverage technical action that Opus WIRE should push now is to deploy a \"Post Amplification\" feature that boosts posts_with_metric from 71% to 90%, which would significantly improve the account's overall performance. This feature can be easily integrated into existing workflows and would have a direct impact on increasing engagement, reach, and credibility for WEVAL.",
+      "ms": 45074,
+      "text": "Based on the provided LinkedIn score and KPIs, the single highest-leverage technical action that Opus WIRE should push now is to deploy a \"Post Amplification\" feature that boosts posts_with_metric from 71% to 90%, which would significantly improve engagement and reach. This feature can be easily integrated into WEVAL's existing infrastructure and would have a direct impact on their overall LinkedIn performance, making it the most critical action for Opus WIRE to push now.",
      "ok": true
    },
    "qwen3": {
-      "ms": 25957,
+      "ms": 22680,
      "text": "",
      "ok": true
    },
    "brain_v4": {
-      "ms": 60026,
+      "ms": 60034,
      "text": "TIMEOUT",
      "ok": false
    }
--- a/api/v83-business-kpi-latest.json
+++ b/api/v83-business-kpi-latest.json
@@ -1,7 +1,7 @@
 {
    "ok": true,
    "version": "V83-business-kpi",
-    "ts": "2026-04-20T10:45:15+00:00",
+    "ts": "2026-04-20T10:54:06+00:00",
    "summary": {
        "total_categories": 7,
        "total_kpis": 56,
--- a/api/wevia-multi-provider.php
+++ b/api/wevia-multi-provider.php
@@ -156,7 +156,9 @@ if (strlen($kb_context) > 50) {
    $msg = "[Contexte WEVAL KB:]\n" . substr($kb_context, 0, 500) . "\n\n[Question:]\n" . $msg;
 }

-$ms = [["role" => "user", "content" => $msg]];
+// OPUS 20avr doctrine #4: anti-hallucination system prompt prepended to all provider messages
+$__guard_sys = "Tu es WEVIA, IA souveraine de WEVAL Consulting. REGLES STRICTES: Tu ne peux PAS executer de commandes shell ni lire de fichiers. Si user demande diagnostic/exec/status/vault/fichier/log: reponds 'Cette requete necessite un intent shell reel. Tape diagnostique toi ou demande a Opus de wirer un intent dedie.' N INVENTE JAMAIS outputs de commandes, PIDs, paths, MD5, timestamps, docker containers, journaux systeme. Si tu ne connais pas une info factuelle dis-le.";
+$ms = [["role" => "system", "content" => $__guard_sys], ["role" => "user", "content" => $msg]];

 // CF models
 if ($mdl === "cf-deepseek-r1" || $mdl === "cf-llama") {