feat: AI Desks — configurable agent system for news/technical/eco processing

- New ai_desks table with CRUD (get_all/by_type/upsert/delete)
- ai_desks router: REST API + GET /signal-catalog (7 extensible signals)
- News Desk: semantic dedup via AI (±N days window, system_prompt hint)
- Technical Desk: 4 signal detectors driven by desk config
  (ma_cross, rsi_extreme, bb_squeeze, new_52w_extreme)
- 3 more signals in catalog ready to enable: price_gap, volume_spike, macd_crossover
- market_event_detector.py loads desk configs at runtime, falls back to legacy params
- AIDesks.tsx: full editor UI with signal toggles, param sliders, instrument multi-select
- Sidebar: Bot icon + /ai-desks route

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

backend/services/database.py

@@ -961,6 +961,87 @@ def init_db():
     except Exception:
         pass
 
+    # ── AI Desks ──────────────────────────────────────────────────────────────
+    c.execute("""CREATE TABLE IF NOT EXISTS ai_desks (
+        id INTEGER PRIMARY KEY AUTOINCREMENT,
+        name TEXT UNIQUE NOT NULL,
+        type TEXT NOT NULL,
+        active INTEGER DEFAULT 1,
+        system_prompt TEXT DEFAULT '',
+        instruments TEXT DEFAULT '[]',
+        config TEXT DEFAULT '{}',
+        created_at TEXT DEFAULT (datetime('now')),
+        updated_at TEXT DEFAULT (datetime('now'))
+    )""")
+
+    # Seed default desks (idempotent)
+    _AI_DESK_DEFAULTS = [
+        {
+            "name": "News Desk — Géopolitique",
+            "type": "news",
+            "active": 1,
+            "system_prompt": (
+                "Tu es un analyste géopolitique et macro senior. Tu évalues si une news représente "
+                "un événement marché STRUCTURANT qui mérite un enregistrement permanent.\n"
+                "Sois exigeant : préfère ignorer une news douteuse plutôt qu'enregistrer du bruit.\n"
+                "Points d'attention :\n"
+                "- Al Jazeera, RT et certains médias régionaux publient souvent plusieurs articles "
+                "redondants sur le même fait — vérifie toujours si un événement similaire existe déjà.\n"
+                "- Une rumeur ou spéculation sans source officielle ne qualifie pas.\n"
+                "- Privilégie les faits avérés avec impact macro ou géopolitique mesurable."
+            ),
+            "instruments": json.dumps(["SPY","GLD","USO","TLT","VXX","EURUSD=X","BTC-USD","XOM"]),
+            "config": json.dumps({
+                "min_impact": 0.55,
+                "lookback_hours": 48,
+                "max_evaluate": 15,
+                "dedup_enabled": True,
+                "dedup_lookback_days": 2,
+                "dedup_categories": ["geopolitical","fundamental","report"],
+            }),
+        },
+        {
+            "name": "Technical Desk",
+            "type": "technical",
+            "active": 1,
+            "system_prompt": (
+                "Tu détectes des signaux techniques structurants sur les marchés financiers. "
+                "Concentre-toi sur les signaux qui ont une signification macro claire."
+            ),
+            "instruments": json.dumps([
+                "SPY","QQQ","IWM","EEM","GLD","USO","TLT",
+                "EURUSD=X","VXX","BTC-USD","NVDA","XOM","HYG"
+            ]),
+            "config": json.dumps({
+                "lookback_days": 7,
+                "signals": {
+                    "ma_cross":       {"enabled": True,  "pairs": [["MA50","MA200"],["MA50","MA100"]]},
+                    "rsi_extreme":    {"enabled": True,  "period": 14, "oversold": 30, "overbought": 70},
+                    "bb_squeeze":     {"enabled": True,  "period": 20, "std": 2.0, "width_threshold": 0.05},
+                    "new_52w_extreme":{"enabled": True,  "buffer_pct": 0.5},
+                },
+            }),
+        },
+        {
+            "name": "Eco Desk — FRED",
+            "type": "eco",
+            "active": 1,
+            "system_prompt": "Tu analyses les surprises économiques des données macro US (FRED).",
+            "instruments": json.dumps(["SPY","TLT","GLD","EURUSD=X","USO","HYG"]),
+            "config": json.dumps({"z_threshold": 1.5, "days": 7}),
+        },
+    ]
+    for _desk in _AI_DESK_DEFAULTS:
+        try:
+            c.execute(
+                "INSERT OR IGNORE INTO ai_desks (name, type, active, system_prompt, instruments, config) "
+                "VALUES (?,?,?,?,?,?)",
+                (_desk["name"], _desk["type"], _desk["active"],
+                 _desk["system_prompt"], _desk["instruments"], _desk["config"])
+            )
+        except Exception:
+            pass
+
     conn.commit()
     conn.close()
 
@@ -4850,3 +4931,93 @@ def get_weekly_impact_sources(days: int = 7, min_score: float = 0.3) -> List[Dic
         return result
     finally:
         conn.close()
+
+
+# ── AI Desks ──────────────────────────────────────────────────────────────────
+
+def get_all_ai_desks() -> List[Dict[str, Any]]:
+    conn = get_conn()
+    try:
+        rows = conn.execute("SELECT * FROM ai_desks ORDER BY type, name").fetchall()
+        result = []
+        for r in rows:
+            d = dict(r)
+            for f in ("instruments", "config"):
+                try:
+                    d[f] = json.loads(d.get(f) or "[]" if f == "instruments" else "{}")
+                except Exception:
+                    d[f] = [] if f == "instruments" else {}
+            result.append(d)
+        return result
+    finally:
+        conn.close()
+
+
+def get_ai_desk_by_type(desk_type: str) -> Optional[Dict[str, Any]]:
+    """Return the first active desk of given type, or None."""
+    desks = get_all_ai_desks()
+    return next((d for d in desks if d["type"] == desk_type and d.get("active")), None)
+
+
+def upsert_ai_desk(desk: Dict[str, Any]) -> int:
+    conn = get_conn()
+    try:
+        instr = desk.get("instruments", [])
+        cfg   = desk.get("config", {})
+        if isinstance(instr, list):
+            instr = json.dumps(instr)
+        if isinstance(cfg, dict):
+            cfg = json.dumps(cfg)
+        conn.execute("""
+            INSERT INTO ai_desks (name, type, active, system_prompt, instruments, config, updated_at)
+            VALUES (?,?,?,?,?,?, datetime('now'))
+            ON CONFLICT(name) DO UPDATE SET
+                type=excluded.type, active=excluded.active,
+                system_prompt=excluded.system_prompt,
+                instruments=excluded.instruments, config=excluded.config,
+                updated_at=datetime('now')
+        """, (desk["name"], desk["type"], int(desk.get("active", 1)),
+              desk.get("system_prompt", ""), instr, cfg))
+        row = conn.execute("SELECT id FROM ai_desks WHERE name=?", (desk["name"],)).fetchone()
+        conn.commit()
+        return row["id"] if row else -1
+    finally:
+        conn.close()
+
+
+def delete_ai_desk(name: str) -> bool:
+    conn = get_conn()
+    try:
+        conn.execute("DELETE FROM ai_desks WHERE name=?", (name,))
+        conn.commit()
+        return True
+    finally:
+        conn.close()
+
+
+def get_market_events_near_date(date_str: str, days: int = 2,
+                                categories: Optional[List[str]] = None) -> List[Dict[str, Any]]:
+    """Fetch market_events within ±days of date_str, optionally filtered by category."""
+    conn = get_conn()
+    try:
+        from datetime import datetime, timedelta
+        dt = datetime.fromisoformat(date_str[:10])
+        d_from = (dt - timedelta(days=days)).strftime("%Y-%m-%d")
+        d_to   = (dt + timedelta(days=days)).strftime("%Y-%m-%d")
+        if categories:
+            placeholders = ",".join("?" * len(categories))
+            rows = conn.execute(
+                f"SELECT id, name, start_date, category, description FROM market_events "
+                f"WHERE start_date BETWEEN ? AND ? AND category IN ({placeholders}) "
+                f"ORDER BY start_date DESC LIMIT 30",
+                [d_from, d_to] + list(categories)
+            ).fetchall()
+        else:
+            rows = conn.execute(
+                "SELECT id, name, start_date, category, description FROM market_events "
+                "WHERE start_date BETWEEN ? AND ? ORDER BY start_date DESC LIMIT 30",
+                (d_from, d_to)
+            ).fetchall()
+        return [dict(r) for r in rows]
+    finally:
+        conn.close()

backend/services/market_event_detector.py

@@ -4,11 +4,10 @@ Isolated cycle action: Check New Market Events.
 Scans 4 sources and creates market_events for significant findings:
   - news     : geopolitical/macro news (RSS feeds, rule-scored)
   - eco      : FRED economic releases with high surprise z-score
-  - technical: MA50/MA100/MA200 crossovers on key instruments
+  - technical: configurable signal catalog driven by Technical Desk
   - reports  : institutional reports (COT, EIA) with high importance
 
-After each event is created, instrument impacts are evaluated immediately
-via the AI (impact_service.evaluate_event_impacts).
+Desk configs are loaded from ai_desks table at runtime.
 """
 import json
 import logging
@@ -69,7 +68,7 @@ def _parse_date(raw: str) -> str:
 
 
 def _save_and_evaluate(ev: Dict, existing: set) -> Optional[Dict]:
-    """Save a market_event and immediately evaluate instrument impacts. Returns created dict or None."""
+    """Save a market_event and immediately evaluate instrument impacts."""
     from services.database import save_market_event
     try:
         event_id = save_market_event(ev)
@@ -79,7 +78,6 @@ def _save_and_evaluate(ev: Dict, existing: set) -> Optional[Dict]:
         logger.error(f"[check_events] save failed for '{ev['name']}': {e}")
         return None
 
-    # Evaluate instrument impacts immediately
     try:
         from services.impact_service import evaluate_event_impacts
         evaluate_event_impacts(event_id, force=False)
@@ -89,15 +87,82 @@ def _save_and_evaluate(ev: Dict, existing: set) -> Optional[Dict]:
     return {"name": ev["name"], "category": ev.get("category", ""), "date": ev.get("start_date", ""), "event_id": event_id}
 
 
+# ── Semantic deduplication ────────────────────────────────────────────────────
+
+def _semantic_dedup(
+    title: str,
+    source: str,
+    date_str: str,
+    summary: str,
+    category: str,
+    client: Any,
+    dedup_lookback_days: int = 2,
+    system_prompt_hint: str = "",
+) -> bool:
+    """
+    Ask the AI whether this news already exists in recent market_events.
+    Returns True if it's a duplicate (should be skipped).
+    """
+    from services.database import get_market_events_near_date
+    dedup_categories = ["geopolitical", "fundamental", "report"]
+    if category and category not in dedup_categories:
+        dedup_categories.append(category)
+
+    recent = get_market_events_near_date(date_str, days=dedup_lookback_days, categories=dedup_categories)
+    if not recent:
+        return False
+
+    recent_block = "\n".join(
+        f"  [{r['start_date']}] {r['name']} — {(r.get('description') or '')[:80]}"
+        for r in recent[:15]
+    )
+
+    hint = f"\nNote du desk: {system_prompt_hint[:200]}" if system_prompt_hint else ""
+
+    prompt = f"""Tu es un éditeur de base de données d'événements marchés.{hint}
+
+NOUVELLE NEWS À VÉRIFIER:
+- Titre: {title}
+- Source: {source}
+- Date: {date_str}
+- Résumé: {summary[:300]}
+
+ÉVÉNEMENTS EXISTANTS (±{dedup_lookback_days} jours):
+{recent_block}
+
+Cette news représente-t-elle le même fait qu'un événement déjà enregistré ?
+Réponds JSON: {{"is_duplicate": true/false, "reason": "courte phrase"}}"""
+
+    try:
+        resp = client.chat.completions.create(
+            model="gpt-4o-mini",
+            messages=[{"role": "user", "content": prompt}],
+            response_format={"type": "json_object"},
+            temperature=0.0,
+            max_tokens=100,
+        )
+        parsed = json.loads(resp.choices[0].message.content)
+        is_dup = bool(parsed.get("is_duplicate", False))
+        if is_dup:
+            logger.debug(f"[dedup] Skipping duplicate: '{title[:40]}' — {parsed.get('reason','')}")
+        return is_dup
+    except Exception as e:
+        logger.debug(f"[dedup] AI check failed for '{title[:40]}': {e}")
+        return False
+
+
 # ── Source 1: Geopolitical / macro news ──────────────────────────────────────
 
-def _check_news(
-    min_impact: float = 0.55,
-    lookback_hours: int = 48,
-    max_to_evaluate: int = 15,
-) -> List[Dict[str, Any]]:
+def _check_news(desk_cfg: Dict[str, Any]) -> List[Dict[str, Any]]:
     from services.data_fetcher import fetch_geo_news
 
+    min_impact      = float(desk_cfg.get("min_impact", 0.55))
+    lookback_hours  = int(desk_cfg.get("lookback_hours", 48))
+    max_evaluate    = int(desk_cfg.get("max_evaluate", 15))
+    dedup_enabled   = bool(desk_cfg.get("dedup_enabled", True))
+    dedup_days      = int(desk_cfg.get("dedup_lookback_days", 2))
+    system_prompt   = desk_cfg.get("_system_prompt", "")
+
     api_key = _get_api_key()
     if not api_key:
         logger.warning("[check_events/news] no OpenAI key — skipping")
@@ -122,13 +187,10 @@ def _check_news(
             pass
         candidates.append(n)
 
-    candidates = candidates[:max_to_evaluate]
+    candidates = candidates[:max_evaluate]
     if not candidates:
         return []
 
-    existing = _existing_event_keys()
-    created: List[Dict] = []
-
     try:
         from openai import OpenAI
         client = OpenAI(api_key=api_key)
@@ -136,17 +198,35 @@ def _check_news(
         logger.warning(f"[check_events/news] OpenAI init failed: {e}")
         return []
 
+    existing = _existing_event_keys()
+    created: List[Dict] = []
+
     for n in candidates:
         title = n.get("title", "")
         if not title or _is_dup(title, existing):
             continue
 
+        pub_date  = _parse_date(n.get("date", ""))
+        news_summary = str(n.get("summary", ""))[:400]
+        source    = n.get("source", "")
+
+        # Semantic dedup before expensive classification call
+        if dedup_enabled:
+            if _semantic_dedup(
+                title, source, pub_date, news_summary,
+                category="geopolitical",
+                client=client,
+                dedup_lookback_days=dedup_days,
+                system_prompt_hint=system_prompt,
+            ):
+                continue
+
         prompt = f"""Tu es un analyste macro. Cette news représente-t-elle un événement marché structurant qui mérite un enregistrement permanent ?
 
 TITRE: {title}
-SOURCE: {n.get('source', '')}
+SOURCE: {source}
 DATE: {n.get('date', '')}
-RÉSUMÉ: {str(n.get('summary', ''))[:400]}
+RÉSUMÉ: {news_summary}
 SCORE IMPACT (règle): {n.get('impact_score', 0):.2f}
 
 Réponds OUI seulement si c'est un fait avéré, pas une rumeur ou une opinion, et qu'il a un impact macro ou géopolitique mesurable.
@@ -185,25 +265,25 @@ FORMAT JSON STRICT:
             continue
 
         source_ref = {
-            "title": title,
-            "source": n.get("source", ""),
-            "url": n.get("url") or n.get("link", ""),
-            "date": _parse_date(n.get("date", "")),
+            "title":          title,
+            "source":         source,
+            "url":            n.get("url") or n.get("link", ""),
+            "date":           pub_date,
             "original_score": round(float(n.get("impact_score", 0)), 3),
         }
 
         ev = {
-            "name":           ev_name,
-            "start_date":     _parse_date(n.get("date", "")),
-            "level":          parsed.get("level", "short"),
-            "category":       parsed.get("category", "geopolitical"),
-            "sub_type":       parsed.get("sub_type", ""),
-            "description":    parsed.get("description", title),
-            "market_impact":  "",
+            "name":            ev_name,
+            "start_date":      pub_date,
+            "level":           parsed.get("level", "short"),
+            "category":        parsed.get("category", "geopolitical"),
+            "sub_type":        parsed.get("sub_type", ""),
+            "description":     parsed.get("description", title),
+            "market_impact":   "",
             "affected_assets": parsed.get("affected_assets", []),
-            "impact_score":   float(parsed.get("impact_score", 0.6)),
-            "source_refs":    [source_ref],
-            "origin":         "detector_news",
+            "impact_score":    float(parsed.get("impact_score", 0.6)),
+            "source_refs":     [source_ref],
+            "origin":          "detector_news",
         }
         result = _save_and_evaluate(ev, existing)
         if result:
@@ -215,9 +295,12 @@ FORMAT JSON STRICT:
 
 # ── Source 2: Eco calendar — FRED surprises ───────────────────────────────────
 
-def _check_eco(z_threshold: float = 1.5, days: int = 7) -> List[Dict[str, Any]]:
+def _check_eco(desk_cfg: Dict[str, Any]) -> List[Dict[str, Any]]:
     from services.database import get_recent_economic_surprises
 
+    z_threshold = float(desk_cfg.get("z_threshold", 1.5))
+    days        = int(desk_cfg.get("days", 7))
+
     try:
         releases = get_recent_economic_surprises(days=days, min_zscore=z_threshold)
     except Exception as e:
@@ -251,33 +334,33 @@ def _check_eco(z_threshold: float = 1.5, days: int = 7) -> List[Dict[str, Any]]:
                 assets = []
 
         source_ref = {
-            "title":  f"FRED release: {ev_name_base} ({ev_date})",
-            "source": "FRED",
-            "url":    f"https://fred.stlouisfed.org/series/{s_id}" if s_id else "",
-            "date":   ev_date,
+            "title":          f"FRED release: {ev_name_base} ({ev_date})",
+            "source":         "FRED",
+            "url":            f"https://fred.stlouisfed.org/series/{s_id}" if s_id else "",
+            "date":           ev_date,
             "original_score": round(min(0.95, 0.35 + z * 0.15), 3),
         }
 
         ev = {
-            "name":           ev_name,
-            "start_date":     ev_date,
-            "level":          level,
-            "category":       "event_calendar",
-            "sub_type":       sub_type,
-            "description":    (
+            "name":            ev_name,
+            "start_date":      ev_date,
+            "level":           level,
+            "category":        "event_calendar",
+            "sub_type":        sub_type,
+            "description":     (
                 f"Surprise {direction} {sign}{s_pct:.1f}% vs baseline "
                 f"(z-score: {z:.1f}σ). "
                 f"Réel: {rel.get('actual_value', '?')} {rel.get('actual_unit', '')} "
                 f"/ Prévision: {rel.get('forecast_value', '?')}."
             ),
-            "market_impact":  "",
+            "market_impact":   "",
             "affected_assets": assets,
-            "impact_score":   min(0.95, 0.35 + z * 0.15),
-            "actual_value":   str(rel.get("actual_value", "")),
-            "expected_value": str(rel.get("forecast_value", "")),
-            "surprise_pct":   float(s_pct),
-            "source_refs":    [source_ref],
-            "origin":         "detector_eco",
+            "impact_score":    min(0.95, 0.35 + z * 0.15),
+            "actual_value":    str(rel.get("actual_value", "")),
+            "expected_value":  str(rel.get("forecast_value", "")),
+            "surprise_pct":    float(s_pct),
+            "source_refs":     [source_ref],
+            "origin":          "detector_eco",
         }
         result = _save_and_evaluate(ev, existing)
         if result:
@@ -287,9 +370,192 @@ def _check_eco(z_threshold: float = 1.5, days: int = 7) -> List[Dict[str, Any]]:
     return created
 
 
-# ── Source 3: MA crossovers (technical) ──────────────────────────────────────
+# ── Technical signal detectors ────────────────────────────────────────────────
 
-def _check_technical(instruments: List[str] = None, lookback_days: int = 7) -> List[Dict[str, Any]]:
+def _detect_ma_cross(ticker: str, df: Any, params: Dict, cutoff: str) -> List[Dict]:
+    """Golden/Death cross detector for configured MA pairs."""
+    import pandas as pd
+    events = []
+    pairs_cfg = params.get("pairs", [["MA50", "MA200"], ["MA50", "MA100"]])
+    ma_map = {"MA20": 20, "MA50": 50, "MA100": 100, "MA200": 200}
+
+    close = df["Close"].squeeze()
+
+    # Pre-compute all required MAs
+    needed: set = set()
+    for pair in pairs_cfg:
+        needed.update(pair)
+    ma_series: Dict[str, Any] = {}
+    for lbl in needed:
+        period = ma_map.get(lbl)
+        if period and len(df) >= period:
+            ma_series[lbl] = close.rolling(period).mean()
+
+    recent = df.tail(4)
+    for i in range(1, len(recent)):
+        date_str = str(recent.index[i])[:10]
+        if date_str < cutoff:
+            continue
+        for fast_lbl, slow_lbl in pairs_cfg:
+            if fast_lbl not in ma_series or slow_lbl not in ma_series:
+                continue
+            fp = ma_series[fast_lbl].iloc[-(len(recent) - i + 1)]
+            fc = ma_series[fast_lbl].iloc[-(len(recent) - i)]
+            sp = ma_series[slow_lbl].iloc[-(len(recent) - i + 1)]
+            sc = ma_series[slow_lbl].iloc[-(len(recent) - i)]
+            if any(pd.isna(v) for v in [fp, fc, sp, sc]):
+                continue
+            if fp < sp and fc >= sc:
+                kind = "golden"
+            elif fp > sp and fc <= sc:
+                kind = "death"
+            else:
+                continue
+            cross_label = "Golden Cross" if kind == "golden" else "Death Cross"
+            events.append({
+                "name":      f"{ticker} {fast_lbl}/{slow_lbl} {cross_label} ({date_str[:7]})",
+                "date":      date_str,
+                "direction": "bullish" if kind == "golden" else "bearish",
+                "sub_type":  f"{fast_lbl}/{slow_lbl} Cross",
+                "score":     0.65 if "MA200" in (fast_lbl, slow_lbl) else 0.45,
+                "level":     "medium" if "MA200" in (fast_lbl, slow_lbl) else "short",
+                "desc":      f"{cross_label}: {fast_lbl} {'au-dessus' if kind=='golden' else 'en-dessous'} de {slow_lbl} sur {ticker}.",
+            })
+    return events
+
+
+def _detect_rsi_extreme(ticker: str, df: Any, params: Dict, cutoff: str) -> List[Dict]:
+    """RSI oversold/overbought signal."""
+    import pandas as pd
+    period     = int(params.get("period", 14))
+    oversold   = float(params.get("oversold", 30))
+    overbought = float(params.get("overbought", 70))
+
+    close = df["Close"].squeeze()
+    if len(close) < period + 2:
+        return []
+
+    delta = close.diff()
+    gain  = delta.clip(lower=0).rolling(period).mean()
+    loss  = (-delta.clip(upper=0)).rolling(period).mean()
+    rs    = gain / loss.replace(0, float("nan"))
+    rsi   = 100 - (100 / (1 + rs))
+
+    events = []
+    recent = rsi.tail(3)
+    for i in range(len(recent)):
+        date_str = str(recent.index[i])[:10]
+        if date_str < cutoff:
+            continue
+        val = recent.iloc[i]
+        if pd.isna(val):
+            continue
+        if val <= oversold:
+            direction, label = "bullish", "Oversold"
+        elif val >= overbought:
+            direction, label = "bearish", "Overbought"
+        else:
+            continue
+        events.append({
+            "name":      f"{ticker} RSI {label} ({date_str[:7]})",
+            "date":      date_str,
+            "direction": direction,
+            "sub_type":  f"RSI {label}",
+            "score":     0.50 if abs(val - 50) > 30 else 0.40,
+            "level":     "short",
+            "desc":      f"RSI({period}) à {val:.1f} sur {ticker} — signal {label.lower()} ({direction}).",
+        })
+    return events
+
+
+def _detect_bb_squeeze(ticker: str, df: Any, params: Dict, cutoff: str) -> List[Dict]:
+    """Bollinger Band squeeze detector."""
+    import pandas as pd
+    period          = int(params.get("period", 20))
+    std_mult        = float(params.get("std", 2.0))
+    width_threshold = float(params.get("width_threshold", 0.05))
+
+    close = df["Close"].squeeze()
+    if len(close) < period + 2:
+        return []
+
+    mid   = close.rolling(period).mean()
+    std   = close.rolling(period).std()
+    upper = mid + std_mult * std
+    lower = mid - std_mult * std
+    width = (upper - lower) / mid
+
+    events = []
+    recent = width.tail(3)
+    for i in range(len(recent)):
+        date_str = str(recent.index[i])[:10]
+        if date_str < cutoff:
+            continue
+        w = recent.iloc[i]
+        if pd.isna(w):
+            continue
+        if w <= width_threshold:
+            events.append({
+                "name":      f"{ticker} BB Squeeze ({date_str[:7]})",
+                "date":      date_str,
+                "direction": "neutral",
+                "sub_type":  "BB Squeeze",
+                "score":     0.45,
+                "level":     "short",
+                "desc":      f"Bandes de Bollinger({period},{std_mult}) très resserrées sur {ticker} — width={w:.3f}. Explosion de volatilité imminente.",
+            })
+    return events
+
+
+def _detect_52w_extreme(ticker: str, df: Any, params: Dict, cutoff: str) -> List[Dict]:
+    """New 52-week high/low detector."""
+    import pandas as pd
+    buffer_pct = float(params.get("buffer_pct", 0.5)) / 100
+
+    close = df["Close"].squeeze()
+    if len(close) < 252:
+        return []
+
+    high_52 = close.rolling(252).max()
+    low_52  = close.rolling(252).min()
+
+    events = []
+    recent_close = close.tail(3)
+    for i in range(len(recent_close)):
+        date_str = str(recent_close.index[i])[:10]
+        if date_str < cutoff:
+            continue
+        c   = recent_close.iloc[i]
+        h52 = high_52.iloc[-(3 - i)]
+        l52 = low_52.iloc[-(3 - i)]
+        if pd.isna(c) or pd.isna(h52) or pd.isna(l52):
+            continue
+        if c >= h52 * (1 - buffer_pct):
+            events.append({
+                "name":      f"{ticker} Nouveau 52W High ({date_str[:7]})",
+                "date":      date_str,
+                "direction": "bullish",
+                "sub_type":  "52W High",
+                "score":     0.60,
+                "level":     "medium",
+                "desc":      f"{ticker} atteint un nouveau plus haut 52 semaines à {c:.2f} (précédent: {h52:.2f}).",
+            })
+        elif c <= l52 * (1 + buffer_pct):
+            events.append({
+                "name":      f"{ticker} Nouveau 52W Low ({date_str[:7]})",
+                "date":      date_str,
+                "direction": "bearish",
+                "sub_type":  "52W Low",
+                "score":     0.60,
+                "level":     "medium",
+                "desc":      f"{ticker} atteint un nouveau plus bas 52 semaines à {c:.2f} (précédent: {l52:.2f}).",
+            })
+    return events
+
+
+# ── Source 3: Technical signals ───────────────────────────────────────────────
+
+def _check_technical(desk_cfg: Dict[str, Any]) -> List[Dict[str, Any]]:
     try:
         import yfinance as yf
         import pandas as pd
@@ -297,93 +563,78 @@ def _check_technical(instruments: List[str] = None, lookback_days: int = 7) -> L
         logger.warning("[check_events/technical] yfinance/pandas not available")
         return []
 
-    if instruments is None:
-        instruments = WATCH_INSTRUMENTS
+    instruments    = desk_cfg.get("_instruments") or WATCH_INSTRUMENTS
+    lookback_days  = int(desk_cfg.get("lookback_days", 7))
+    signals_config = desk_cfg.get("signals", {})
 
-    existing  = _existing_event_keys()
+    # Determine which signals are active
+    def sig_cfg(sig_id: str) -> Optional[Dict]:
+        c = signals_config.get(sig_id, {})
+        return c if c.get("enabled", False) else None
+
+    ma_cross_cfg   = sig_cfg("ma_cross")
+    rsi_cfg        = sig_cfg("rsi_extreme")
+    bb_cfg         = sig_cfg("bb_squeeze")
+    extreme_52w    = sig_cfg("new_52w_extreme")
+
+    if not any([ma_cross_cfg, rsi_cfg, bb_cfg, extreme_52w]):
+        logger.info("[check_events/technical] no active signals in desk config")
+        return []
+
+    existing = _existing_event_keys()
     created: List[Dict] = []
-    cutoff    = (datetime.utcnow() - timedelta(days=lookback_days)).strftime("%Y-%m-%d")
+    cutoff   = (datetime.utcnow() - timedelta(days=lookback_days)).strftime("%Y-%m-%d")
 
     for ticker in instruments:
         try:
-            df = yf.download(ticker, period="1y", interval="1d", progress=False, auto_adjust=True)
-            if df is None or len(df) < 210:
+            df = yf.download(ticker, period="2y", interval="1d", progress=False, auto_adjust=True)
+            if df is None or len(df) < 20:
                 continue
 
-            close = df["Close"].squeeze()
-            df["ma50"]  = close.rolling(50).mean()
-            df["ma100"] = close.rolling(100).mean()
-            df["ma200"] = close.rolling(200).mean()
+            # Flatten MultiIndex if needed (yfinance ≥ 0.2 returns MultiIndex columns)
+            if hasattr(df.columns, "levels"):
+                df.columns = df.columns.get_level_values(0)
 
-            recent = df.tail(lookback_days + 2)
+            detected: List[Dict] = []
+            if ma_cross_cfg and len(df) >= 210:
+                detected += _detect_ma_cross(ticker, df, ma_cross_cfg, cutoff)
+            if rsi_cfg:
+                detected += _detect_rsi_extreme(ticker, df, rsi_cfg, cutoff)
+            if bb_cfg:
+                detected += _detect_bb_squeeze(ticker, df, bb_cfg, cutoff)
+            if extreme_52w and len(df) >= 252:
+                detected += _detect_52w_extreme(ticker, df, extreme_52w, cutoff)
 
-            for i in range(1, len(recent)):
-                date_str = str(recent.index[i])[:10]
-                if date_str < cutoff:
+            for sig in detected:
+                ev_name = sig["name"]
+                if _is_dup(ev_name, existing):
                     continue
 
-                prev = recent.iloc[i - 1]
-                curr = recent.iloc[i]
+                source_ref = {
+                    "title":          f"Technical signal: {ev_name}",
+                    "source":         "yfinance/computed",
+                    "url":            f"https://finance.yahoo.com/quote/{ticker}",
+                    "date":           sig["date"],
+                    "original_score": sig["score"],
+                }
 
-                def cross(fp, fc, sp, sc):
-                    if any(pd.isna(v) for v in [fp, fc, sp, sc]):
-                        return None
-                    if fp < sp and fc >= sc:
-                        return "golden"
-                    if fp > sp and fc <= sc:
-                        return "death"
-                    return None
-
-                pairs = [
-                    ("MA50", "MA200", prev["ma50"], curr["ma50"], prev["ma200"], curr["ma200"]),
-                    ("MA50", "MA100", prev["ma50"], curr["ma50"], prev["ma100"], curr["ma100"]),
-                ]
-
-                for fast_lbl, slow_lbl, fp, fc, sp, sc in pairs:
-                    kind = cross(fp, fc, sp, sc)
-                    if kind is None:
-                        continue
-
-                    cross_label = "Golden Cross" if kind == "golden" else "Death Cross"
-                    ev_name = f"{ticker} {fast_lbl}/{slow_lbl} {cross_label} ({date_str[:7]})"
-
-                    if _is_dup(ev_name, existing):
-                        continue
-
-                    direction = "bullish" if kind == "golden" else "bearish"
-                    level     = "medium" if slow_lbl == "MA200" else "short"
-
-                    source_ref = {
-                        "title":  f"Technical signal: {ev_name}",
-                        "source": "yfinance/computed",
-                        "url":    f"https://finance.yahoo.com/quote/{ticker}",
-                        "date":   date_str,
-                        "original_score": 0.65 if slow_lbl == "MA200" else 0.45,
-                    }
-
-                    ev = {
-                        "name":           ev_name,
-                        "start_date":     date_str,
-                        "level":          level,
-                        "category":       "technical",
-                        "sub_type":       f"{fast_lbl}/{slow_lbl} Cross",
-                        "description":    (
-                            f"{cross_label} : {fast_lbl} passe "
-                            f"{'au-dessus' if kind == 'golden' else 'en-dessous'} "
-                            f"de la {slow_lbl} sur {ticker}. "
-                            f"Signal {direction} de tendance "
-                            f"{'long terme' if slow_lbl == 'MA200' else 'moyen terme'}."
-                        ),
-                        "market_impact":  f"Signal {direction} sur {ticker}",
-                        "affected_assets": [ticker],
-                        "impact_score":   0.65 if slow_lbl == "MA200" else 0.45,
-                        "source_refs":    [source_ref],
-                        "origin":         "detector_technical",
-                    }
-                    result = _save_and_evaluate(ev, existing)
-                    if result:
-                        result["source"] = "technical"
-                        created.append(result)
+                ev = {
+                    "name":            ev_name,
+                    "start_date":      sig["date"],
+                    "level":           sig["level"],
+                    "category":        "technical",
+                    "sub_type":        sig["sub_type"],
+                    "description":     sig["desc"],
+                    "market_impact":   f"Signal {sig['direction']} sur {ticker}",
+                    "affected_assets": [ticker],
+                    "impact_score":    sig["score"],
+                    "source_refs":     [source_ref],
+                    "origin":          "detector_technical",
+                }
+                result = _save_and_evaluate(ev, existing)
+                if result:
+                    result["source"] = "technical"
+                    created.append(result)
 
         except Exception as e:
             logger.debug(f"[check_events/technical] {ticker} failed: {e}")
@@ -443,10 +694,10 @@ def _check_reports(days: int = 7, min_importance: int = 3) -> List[Dict[str, Any
                 assets.extend(asset_list)
 
         source_ref = {
-            "title":  title,
-            "source": rpt.get("source", rpt_type),
-            "url":    "",
-            "date":   rpt_date,
+            "title":          title,
+            "source":         rpt.get("source", rpt_type),
+            "url":            "",
+            "date":           rpt_date,
             "original_score": round(min(0.9, 0.3 + rpt.get("importance", 2) * 0.12), 3),
         }
 
@@ -475,6 +726,7 @@ def _check_reports(days: int = 7, min_importance: int = 3) -> List[Dict[str, Any
 
 def check_new_market_events(
     sources: Optional[List[str]] = None,
+    # Legacy overrides (used when called from cycle_actions without a desk)
     news_impact_min: float = 0.55,
     news_lookback_hours: int = 48,
     eco_z_threshold: float = 1.5,
@@ -484,12 +736,52 @@ def check_new_market_events(
     report_min_importance: int = 3,
 ) -> Dict[str, Any]:
     """
-    Isolated cycle action — scans all (or selected) sources, creates
-    market_events with source_refs, and immediately evaluates instrument impacts.
+    Scans all (or selected) sources, creates market_events with source_refs,
+    and immediately evaluates instrument impacts.
+    Desk configs from ai_desks table override legacy params when available.
     """
     if sources is None:
         sources = ["news", "eco", "technical", "reports"]
 
+    # Load desk configs (fall back to legacy params if no active desk found)
+    try:
+        from services.database import get_ai_desk_by_type
+        news_desk  = get_ai_desk_by_type("news")
+        tech_desk  = get_ai_desk_by_type("technical")
+        eco_desk   = get_ai_desk_by_type("eco")
+    except Exception as e:
+        logger.warning(f"[check_events] Could not load desk configs: {e}")
+        news_desk = tech_desk = eco_desk = None
+
+    def _desk_cfg(desk: Optional[Dict], fallback: Dict) -> Dict:
+        if not desk:
+            return fallback
+        cfg = dict(desk.get("config") or {})
+        cfg["_instruments"] = desk.get("instruments") or None
+        cfg["_system_prompt"] = desk.get("system_prompt") or ""
+        return cfg
+
+    news_cfg = _desk_cfg(news_desk, {
+        "min_impact":         news_impact_min,
+        "lookback_hours":     news_lookback_hours,
+        "max_evaluate":       15,
+        "dedup_enabled":      False,
+        "dedup_lookback_days": 2,
+    })
+    eco_cfg = _desk_cfg(eco_desk, {
+        "z_threshold": eco_z_threshold,
+        "days":        eco_days,
+    })
+    tech_cfg = _desk_cfg(tech_desk, {
+        "lookback_days": technical_lookback_days,
+        "signals": {
+            "ma_cross":        {"enabled": True, "pairs": [["MA50", "MA200"], ["MA50", "MA100"]]},
+            "rsi_extreme":     {"enabled": True, "period": 14, "oversold": 30, "overbought": 70},
+            "bb_squeeze":      {"enabled": True, "period": 20, "std": 2.0, "width_threshold": 0.05},
+            "new_52w_extreme": {"enabled": True, "buffer_pct": 0.5},
+        },
+    })
+
     results: Dict[str, Any] = {
         "news": [], "eco": [], "technical": [], "reports": [],
         "total_created": 0,
@@ -498,19 +790,19 @@ def check_new_market_events(
 
     if "news" in sources:
         try:
-            results["news"] = _check_news(min_impact=news_impact_min, lookback_hours=news_lookback_hours)
+            results["news"] = _check_news(news_cfg)
         except Exception as e:
             logger.error(f"[check_events] news source error: {e}")
 
     if "eco" in sources:
         try:
-            results["eco"] = _check_eco(z_threshold=eco_z_threshold, days=eco_days)
+            results["eco"] = _check_eco(eco_cfg)
         except Exception as e:
             logger.error(f"[check_events] eco source error: {e}")
 
     if "technical" in sources:
         try:
-            results["technical"] = _check_technical(lookback_days=technical_lookback_days)
+            results["technical"] = _check_technical(tech_cfg)
         except Exception as e:
             logger.error(f"[check_events] technical source error: {e}")