feat(agent): add learn action flow and grounding guards

agent_v0/agent_v1/config.py

@@ -56,6 +56,13 @@ OLLAMA_HOST = os.getenv("RPA_OLLAMA_HOST", "localhost")
 # Configurable via variable d'environnement RPA_API_TOKEN
 API_TOKEN = os.environ.get("RPA_API_TOKEN", "")
 
+# --- Orchestrateur Léa-first (agent-chat Linux) ---
+# Endpoint racine du service agent-chat qui héberge POST /api/learn/start
+# (P1-LEA-SHADOW). Configurable via RPA_AGENT_CHAT_URL.
+# Défaut : localhost:5004 (même machine en dev). En POC clinique, doit
+# pointer vers le DGX Spark (ex. http://agent-chat.dgx-local:5004).
+AGENT_CHAT_URL = os.environ.get("RPA_AGENT_CHAT_URL", "http://localhost:5004")
+
 # Paramètres de session
 MAX_SESSION_DURATION_S = 60 * 60  # 1 heure
 SESSIONS_ROOT = BASE_DIR / "sessions"

agent_v0/agent_v1/core/captor.py

@@ -56,6 +56,8 @@ class EventCaptorV1:
 
         # État des touches modificatrices
         self.modifiers = set()
+        self._pending_standalone_win = False
+        self._suppress_release_only_win_combo = False
 
         # Tracking du focus fenêtre
         self.last_window = None
@@ -327,6 +329,56 @@ class EventCaptorV1:
             return {"kind": "key", "name": key.name}
         return {"kind": "unknown", "str": str(key)}
 
+    @staticmethod
+    def _raw_key_name(raw_key: Dict[str, Any]) -> Optional[str]:
+        """Nom lisible depuis un raw_key sérialisé."""
+        if raw_key.get("kind") == "vk":
+            char = raw_key.get("char")
+            if char and len(str(char)) == 1:
+                return str(char).lower()
+        if raw_key.get("kind") == "key":
+            name = raw_key.get("name")
+            return str(name).lower() if name else None
+        return None
+
+    def _emit_release_only_windows_combo(self) -> bool:
+        """Infère Win+<touche> si Windows/NoMachine n'a livré que les releases.
+
+        Certaines sessions ne remontent pas les press de Win+S via pynput,
+        mais livrent ensuite release('s') puis release('cmd'). Sans cette
+        inférence ciblée, le geste système est perdu et les releases polluent
+        le prochain text_input.
+        """
+        with self._text_lock:
+            raw_keys = list(self._raw_key_buffer)
+            if len(raw_keys) < 2:
+                return False
+            cmd_names = {"cmd", "cmd_l", "cmd_r"}
+            last = raw_keys[-1]
+            if last.get("action") != "release" or self._raw_key_name(last) not in cmd_names:
+                return False
+            combo_key = None
+            for raw in reversed(raw_keys[:-1]):
+                if raw.get("action") != "release":
+                    continue
+                name = self._raw_key_name(raw)
+                if name and name not in self._MODIFIER_KEY_NAMES:
+                    combo_key = name
+                    break
+            if not combo_key:
+                return False
+            self._raw_key_buffer.clear()
+
+        event = {
+            "type": "key_combo",
+            "keys": ["win", combo_key],
+            "raw_keys": raw_keys,
+            "timestamp": time.time(),
+        }
+        self._inject_screen_metadata(event)
+        self.on_event(event)
+        return True
+
     def _on_press(self, key):
         # TOUJOURS enregistrer le press brut dans le buffer raw_keys
         with self._text_lock:
@@ -344,6 +396,7 @@ class EventCaptorV1:
             self.modifiers.add("shift")
         elif key in (Key.cmd, Key.cmd_l, Key.cmd_r):
             self.modifiers.add("win")
+            self._pending_standalone_win = True
 
         # --- Combos avec modificateur (sauf Shift seul) ---
         # Shift seul n'est pas un « vrai » modificateur pour les combos :
@@ -369,6 +422,9 @@ class EventCaptorV1:
             # Ne PAS émettre de combo si c'est un modificateur seul
             # (ex: appui sur Ctrl sans autre touche = pas de combo)
             if key_name and key_name not in self._MODIFIER_KEY_NAMES:
+                self._pending_standalone_win = False
+                if "win" in self.modifiers:
+                    self._suppress_release_only_win_combo = True
                 # Un combo interrompt la saisie texte en cours
                 self._flush_text_buffer()
                 # Attacher les raw_keys accumulés (press des modificateurs + press de la touche)
@@ -400,6 +456,7 @@ class EventCaptorV1:
         - Enter / Tab : flush immédiat + émission de l'événement
         - Escape : vide le buffer sans émettre
         """
+        escape_raw_keys = None
         with self._text_lock:
             # --- Touches spéciales ---
             if key == Key.backspace:
@@ -411,12 +468,14 @@ class EventCaptorV1:
             if key == Key.esc:
                 # Annuler la saisie en cours
                 self._text_buffer.clear()
-                self._raw_key_buffer.clear()
                 self._text_start_pos = None
                 self._cancel_flush_timer()
-                return
+                escape_raw_keys = list(self._raw_key_buffer)
+                self._raw_key_buffer.clear()
+                # Émettre hors lock après le bloc critique.
+                pass
 
-            if key in (Key.enter, Key.tab):
+            elif key in (Key.enter, Key.tab):
                 # Flush immédiat — on relâche le lock avant d'appeler
                 # _flush_text_buffer (qui prend aussi le lock)
                 pass  # on sort du with et on flush après
@@ -454,6 +513,18 @@ class EventCaptorV1:
                 # Touche spéciale non gérée (F1, Insert, etc.) — on ignore
                 return
 
+        if escape_raw_keys is not None:
+            event = {
+                "type": "key_combo",
+                "keys": ["escape"],
+                "timestamp": time.time(),
+            }
+            if escape_raw_keys:
+                event["raw_keys"] = escape_raw_keys
+            self._inject_screen_metadata(event)
+            self.on_event(event)
+            return
+
         # Si on arrive ici, c'est Enter ou Tab → flush le buffer en cours
         # puis émettre le caractère spécial comme text_input séparé
         self._flush_text_buffer()
@@ -551,6 +622,35 @@ class EventCaptorV1:
                 **self._encode_key(key),
             })
 
+        if key in (Key.cmd, Key.cmd_l, Key.cmd_r) and self._suppress_release_only_win_combo:
+            with self._text_lock:
+                self._raw_key_buffer.clear()
+            self._pending_standalone_win = False
+            self._suppress_release_only_win_combo = False
+            self.modifiers.discard("win")
+            return
+
+        if key in (Key.cmd, Key.cmd_l, Key.cmd_r) and self._emit_release_only_windows_combo():
+            self._pending_standalone_win = False
+            self._suppress_release_only_win_combo = False
+            self.modifiers.discard("win")
+            return
+
+        if key in (Key.cmd, Key.cmd_l, Key.cmd_r) and self._pending_standalone_win:
+            with self._text_lock:
+                raw_keys = list(self._raw_key_buffer)
+                self._raw_key_buffer.clear()
+            event = {
+                "type": "key_combo",
+                "keys": ["win"],
+                "raw_keys": raw_keys,
+                "timestamp": time.time(),
+            }
+            self._inject_screen_metadata(event)
+            self.on_event(event)
+            self._pending_standalone_win = False
+            self._suppress_release_only_win_combo = False
+
         if key in (Key.ctrl, Key.ctrl_l, Key.ctrl_r):
             self.modifiers.discard("ctrl")
         elif key in (Key.alt, Key.alt_l, Key.alt_r):
@@ -559,6 +659,8 @@ class EventCaptorV1:
             self.modifiers.discard("shift")
         elif key in (Key.cmd, Key.cmd_l, Key.cmd_r):
             self.modifiers.discard("win")
+            self._pending_standalone_win = False
+            self._suppress_release_only_win_combo = False
 
     # ----------------------------------------------------------------
     # Métadonnées système

agent_v0/agent_v1/core/grounding.py

@@ -84,6 +84,15 @@ class GroundingEngine:
         if by_role in {"start_button"}:
             return False
 
+        has_anchor = bool(target_spec.get("anchor_image_base64"))
+        context_hints = target_spec.get("context_hints") or {}
+        has_window_or_text_hint = any(
+            str(target_spec.get(key, "") or "").strip()
+            for key in ("window_title", "by_text", "vlm_description")
+        ) or bool(str(context_hints.get("window_title", "") or "").strip())
+        if has_anchor and not has_window_or_text_hint and not by_role:
+            return False
+
         return True
 
     @staticmethod
@@ -174,6 +183,26 @@ class GroundingEngine:
                     hints.append(variant)
         return hints
 
+    @staticmethod
+    def _server_rejects_text_fallback(raw: Optional[Dict[str, Any]]) -> bool:
+        """Dire si un rejet serveur doit bloquer le fallback texte local.
+
+        Un rejet explicite n'est pas un simple "non trouvé": le serveur a vu
+        un candidat et l'a refusé pour une raison de qualité/zone. Refaire une
+        recherche OCR large côté client contournerait ce garde-fou.
+        """
+        if not raw or raw.get("resolved"):
+            return False
+
+        reason = str(raw.get("reason") or "")
+        method = str(raw.get("method") or "")
+        return (
+            method.startswith("rejected_")
+            or reason.startswith("close_tab_")
+            or reason.startswith("drift_")
+            or "below_threshold" in reason
+        )
+
     def _window_crop_matches_target_visually(
         self,
         screenshot_b64: str,
@@ -331,11 +360,31 @@ class GroundingEngine:
         cap_w = window_rect["width"] if window_rect else screen_width
         cap_h = window_rect["height"] if window_rect else screen_height
 
+        skip_text_fallback_after_server_reject = False
         for strategy in strategies:
+            if (
+                strategy == "vlm_local"
+                and skip_text_fallback_after_server_reject
+                and target_spec.get("by_text")
+            ):
+                by_text = target_spec.get("by_text", "")
+                logger.info(
+                    "[GROUNDING] Rejet serveur explicite pour '%s' — "
+                    "skip fallback local hybrid_text_direct",
+                    by_text,
+                )
+                print(
+                    f"    [GROUNDING] Rejet serveur explicite pour '{by_text}' "
+                    "→ pas de fallback texte local"
+                )
+                continue
+
             result = self._try_strategy(
                 strategy, server_url, screenshot_b64, target_spec,
                 fallback_x, fallback_y, cap_w, cap_h,
             )
+            if strategy == "server" and self._server_rejects_text_fallback(result.raw):
+                skip_text_fallback_after_server_reject = True
             if result.found:
                 # ── Conversion coords fenêtre → coords écran ──
                 if window_rect:
@@ -429,6 +478,14 @@ class GroundingEngine:
                     detail=raw.get("matched_element", {}).get("label", ""),
                     raw=raw,
                 )
+            if raw:
+                return GroundingResult(
+                    found=False,
+                    method=raw.get("method", "server"),
+                    score=raw.get("score", 0.0),
+                    detail=raw.get("reason", "server: pas trouvé"),
+                    raw=raw,
+                )
 
         elif strategy == "template":
             anchor_b64 = target_spec.get("anchor_image_base64", "")

agent_v0/agent_v1/main.py

@@ -121,10 +121,7 @@ class AgentV1:
         # Wiring ChatWindow → Executor pour Plan B (pause_message → bulle interactive)
         # Permet à l'executor d'afficher une bulle paused dans la fenêtre Léa V1
         # quand le serveur signale replay_paused=True via /replay/next.
-        try:
-            self._executor._chat_window_ref = self._chat_window
-        except Exception:
-            logger.debug("Wiring chat_window→executor échoué (non bloquant)", exc_info=True)
+        self._wire_chat_window_to_executor()
 
         # Boucles permanentes (pas besoin de session active)
         self.running = True
@@ -154,6 +151,15 @@ class AgentV1:
             shared_state=self._state,
         )
 
+    def _wire_chat_window_to_executor(self) -> None:
+        """Relie l'executor courant à la ChatWindow pour les pauses supervisees."""
+        if self._executor is None or self._chat_window is None:
+            return
+        try:
+            self._executor._chat_window_ref = self._chat_window
+        except Exception:
+            logger.debug("Wiring chat_window->executor echoue (non bloquant)", exc_info=True)
+
     def _delayed_cleanup(self):
         """Nettoyage en arrière-plan après 30s pour ne pas bloquer le démarrage."""
         time.sleep(30)
@@ -224,6 +230,7 @@ class AgentV1:
 
         # Initialiser l'executeur partage
         self._executor = ActionExecutorV1()
+        self._wire_chat_window_to_executor()
 
         self.shot_counter = 0
         self.running = True

agent_v0/agent_v1/network/lea_orchestrator_client.py

@@ -0,0 +1,147 @@
+"""
+Client HTTP minimal pour l'orchestrateur Léa-first (agent-chat Linux).
+
+Rebranchement P1-LEA-SHADOW : le bouton "Apprenez-moi" côté Windows déclenche
+la création d'une session d'apprentissage côté agent-chat (REST) AVANT de
+lancer la capture locale. Le pipeline streaming (capture frames/événements
+via start_recording) n'est PAS modifié — seule la prise de contact initiale
+avec Léa change.
+
+Contrat :
+    POST {AGENT_CHAT_URL}/api/learn/start
+    Headers : Authorization: Bearer <RPA_API_TOKEN>, Content-Type: application/json
+    Body    : { machine_id, session_name, user_id?, trigger_source }
+    Réponse : { session_id, state, message }
+
+Politique :
+    - Timeout 10s (connect + read)
+    - Retry x2 avec backoff 0.5s puis 1.0s
+    - En cas d'échec définitif : lève LeaOrchestratorError (le caller doit
+      basculer en mode dégradé : start_recording local sans assistance).
+"""
+
+from __future__ import annotations
+
+import logging
+import time
+from dataclasses import dataclass
+from typing import Optional
+
+logger = logging.getLogger(__name__)
+
+
+# Timeout HTTP (connect + read) — 10s comme spec
+_HTTP_TIMEOUT_S = 10.0
+# Nombre de tentatives totales (1 + 2 retry)
+_MAX_ATTEMPTS = 3
+# Backoff progressif entre les tentatives
+_BACKOFF_S = (0.5, 1.0)
+
+
+@dataclass(frozen=True)
+class LearnStartResponse:
+    """Réponse normalisée de POST /api/learn/start."""
+
+    session_id: str
+    state: str
+    message: str
+
+
+class LeaOrchestratorError(RuntimeError):
+    """Erreur définitive de communication avec l'orchestrateur Léa."""
+
+
+def start_learning_session(
+    base_url: str,
+    *,
+    machine_id: str,
+    session_name: str,
+    api_token: str = "",
+    user_id: Optional[str] = None,
+    trigger_source: str = "windows_button",
+    timeout_s: float = _HTTP_TIMEOUT_S,
+    max_attempts: int = _MAX_ATTEMPTS,
+    backoff_s: tuple = _BACKOFF_S,
+) -> LearnStartResponse:
+    """Démarre une session d'apprentissage via l'orchestrateur agent-chat.
+
+    Args:
+        base_url: URL racine de l'agent-chat (ex. http://localhost:5004).
+        machine_id: Identifiant unique du poste Windows.
+        session_name: Nom humain de la tâche (saisi par l'utilisateur).
+        api_token: Bearer token (RPA_API_TOKEN). Vide => header omis.
+        user_id: Identifiant utilisateur optionnel.
+        trigger_source: Source du déclenchement (windows_button, tray, ...).
+        timeout_s: Timeout total connect+read par tentative.
+        max_attempts: Nombre total de tentatives (1 + retry).
+        backoff_s: Tuple des délais en secondes entre tentatives (len = max_attempts-1).
+
+    Returns:
+        LearnStartResponse normalisée.
+
+    Raises:
+        LeaOrchestratorError: si toutes les tentatives échouent.
+    """
+    # Import local : httpx peut ne pas être installé sur tous les postes
+    # Windows historiques. On veut un message d'erreur clair plutôt qu'un
+    # ImportError en chaîne au moment du clic bouton.
+    try:
+        import httpx
+    except ImportError as exc:  # pragma: no cover (dépend du venv)
+        raise LeaOrchestratorError(
+            "httpx non disponible — installer httpx>=0.27 sur le poste Windows."
+        ) from exc
+
+    url = base_url.rstrip("/") + "/api/learn/start"
+    payload = {
+        "machine_id": machine_id,
+        "session_name": session_name,
+        "trigger_source": trigger_source,
+    }
+    if user_id:
+        payload["user_id"] = user_id
+
+    headers = {"Content-Type": "application/json"}
+    if api_token:
+        headers["Authorization"] = f"Bearer {api_token}"
+
+    last_exc: Optional[Exception] = None
+    for attempt in range(max_attempts):
+        try:
+            logger.info(
+                "POST %s (tentative %d/%d) machine_id=%s session=%s",
+                url, attempt + 1, max_attempts, machine_id, session_name,
+            )
+            with httpx.Client(timeout=timeout_s) as client:
+                resp = client.post(url, json=payload, headers=headers)
+            resp.raise_for_status()
+            data = resp.json()
+            session_id = data.get("session_id", "")
+            state = data.get("state", "")
+            message = data.get("message", "")
+            if not session_id:
+                raise LeaOrchestratorError(
+                    f"Réponse invalide (pas de session_id) : {data!r}"
+                )
+            logger.info(
+                "Session Léa démarrée : session_id=%s state=%s",
+                session_id, state,
+            )
+            return LearnStartResponse(
+                session_id=str(session_id),
+                state=str(state),
+                message=str(message),
+            )
+        except Exception as exc:  # noqa: BLE001 — on retry sur toute erreur réseau/HTTP
+            last_exc = exc
+            logger.warning(
+                "Echec tentative %d/%d POST %s : %s",
+                attempt + 1, max_attempts, url, exc,
+            )
+            if attempt < max_attempts - 1:
+                delay = backoff_s[attempt] if attempt < len(backoff_s) else backoff_s[-1]
+                time.sleep(delay)
+
+    raise LeaOrchestratorError(
+        f"Echec définitif POST {url} après {max_attempts} tentatives : {last_exc}"
+    )

agent_v0/agent_v1/network/streamer.py

@@ -63,8 +63,14 @@ JPEG_QUALITY = 85
 # Taille max de la queue (backpressure)
 QUEUE_MAX_SIZE = 100
 
-# Types d'événements à ne jamais dropper
-PRIORITY_EVENT_TYPES = {"click", "key", "scroll", "action", "screenshot"}
+# Types d'événements à ne jamais dropper.
+# Les noms historiques sont conservés, mais les événements réels du captor
+# Agent V1 sont mouse_click/key_combo/text_input/mouse_scroll.
+PRIORITY_EVENT_TYPES = {
+    "click", "key", "scroll", "action", "screenshot",
+    "mouse_click", "double_click", "key_combo", "key_press",
+    "text_input", "mouse_scroll",
+}
 
 # Purge locale après ACK serveur (Partie A de l'audit)
 # Activé par défaut : le serveur conserve déjà les screenshots 180 jours

agent_v0/agent_v1/ui/chat_window.py

@@ -9,6 +9,7 @@ Tourne dans son propre thread daemon pour ne pas bloquer pystray.
 
 import logging
 import os
+import math
 import threading
 import time
 from datetime import datetime
@@ -121,7 +122,7 @@ def _tpl_done(payload: Dict[str, Any]) -> tuple:
 def _tpl_need_confirm(payload: Dict[str, Any]) -> tuple:
     action = payload.get("action") or {}
     desc = action.get("description") if isinstance(action, dict) else None
-    title = desc or "Validation requise"
+    title = desc or "J'attends ton accord avant de continuer"
     return ("?", ACTION_ICON_RUN, str(title))
 
 
@@ -867,11 +868,19 @@ class ChatWindow:
                     pass
             except Exception:
                 logger.debug("force-show chat_window silenced", exc_info=True)
-            # UX fix mai 2026 : repartir d'un chat vide pour focaliser
-            # l'attention sur la question (clear visuel uniquement,
-            # self._messages reste intact pour la traçabilité debug).
-            self._clear_chat_history()
-            self._render_paused_bubble(payload)
+            try:
+                # UX fix mai 2026 : repartir d'un chat vide pour focaliser
+                # l'attention sur la question (clear visuel uniquement,
+                # self._messages reste intact pour la traçabilité debug).
+                self._clear_chat_history()
+                self._render_paused_bubble(payload)
+            except Exception:
+                logger.exception("render paused bubble failed; using fallback")
+                try:
+                    self._clear_chat_history()
+                    self._render_paused_fallback_bubble(payload)
+                except Exception:
+                    logger.debug("render paused fallback silenced", exc_info=True)
 
         self._root.after(0, _show_and_render)
 
@@ -895,7 +904,11 @@ class ChatWindow:
             logger.debug("clear chat history silenced", exc_info=True)
 
     @staticmethod
-    def _compute_paused_bubble_height(reason_str: str) -> tuple:
+    def _compute_paused_bubble_height(
+        reason_str: str,
+        chars_per_line: int = 52,
+        max_rows: int = 14,
+    ) -> tuple:
         """Calcule la hauteur du Text (en lignes) + si une scrollbar est
         nécessaire pour le message d'une bulle paused.
 
@@ -910,11 +923,11 @@ class ChatWindow:
         if not reason_str:
             return 2, False
         text = str(reason_str)
-        # Estimation : ~60 chars/ligne effectifs avec wraplength.
-        wrapped_lines = (len(text) // 60) + 1
-        explicit_lines = text.count("\n") + 1
-        estimated = max(wrapped_lines, explicit_lines)
-        cap = 12
+        chars_per_line = max(24, int(chars_per_line or 52))
+        estimated = 0
+        for raw_line in text.splitlines() or [""]:
+            estimated += max(1, math.ceil(len(raw_line) / chars_per_line))
+        cap = max(2, int(max_rows or 14))
         height = max(2, min(cap, estimated))
         # Scrollbar dès que le cap est atteint OU contenu long (filet
         # textuel : ≥ 200 chars implique souvent un débordement visuel
@@ -922,6 +935,46 @@ class ChatWindow:
         needs_scroll = (estimated >= cap) or (len(text) > 200)
         return height, needs_scroll
 
+    def _paused_text_layout(self) -> tuple:
+        """Retourne ``(wrap_px, chars_per_line, max_rows)`` pour la bulle pause.
+
+        La fenêtre Léa est souvent redimensionnée à ~380px de large sur le
+        poste Windows. Les anciennes estimations fixes calculaient trop peu
+        de lignes et tronquaient le message. On part donc des dimensions
+        réelles du canvas et de la métrique de la police Tk.
+        """
+        canvas_w = 0
+        canvas_h = 0
+        try:
+            canvas_w = int(self._canvas.winfo_width()) if self._canvas is not None else 0
+            canvas_h = int(self._canvas.winfo_height()) if self._canvas is not None else 0
+        except Exception:
+            canvas_w = canvas_h = 0
+
+        # Marges: container + padding inner + petite marge droite. La bulle
+        # de pause est une alerte critique, elle utilise donc presque toute
+        # la largeur disponible sur les fenêtres étroites.
+        wrap_px = max(220, canvas_w - (2 * MARGIN) - 52) if canvas_w else 360
+
+        avg_char = 8
+        line_px = 22
+        try:
+            from tkinter import font as tkfont
+            font = tkfont.Font(font=FONT_MSG)
+            avg_char = max(6, font.measure("n"))
+            line_px = max(18, font.metrics("linespace"))
+        except Exception:
+            pass
+
+        chars_per_line = max(24, int(wrap_px / avg_char))
+        # Réserver titre, metadata, boutons, feedback et padding. Même sur
+        # une petite fenêtre, on garde assez de lignes pour ne pas couper un
+        # message d'erreur standard.
+        max_rows = 14
+        if canvas_h:
+            max_rows = max(5, min(18, int((canvas_h - 145) / line_px)))
+        return wrap_px, chars_per_line, max_rows
+
     def _render_paused_bubble(self, payload: Dict[str, Any]) -> None:
         tk = self._tk
         if getattr(self, "_msg_frame", None) is None:
@@ -941,7 +994,7 @@ class ChatWindow:
             container, bg=PAUSED_BG, padx=14, pady=12,
             highlightbackground=PAUSED_BORDER, highlightthickness=2,
         )
-        inner.pack(anchor=tk.W, padx=(0, 50), fill=tk.X)
+        inner.pack(anchor=tk.W, padx=(0, 12), fill=tk.X)
 
         tk.Label(
             inner, text=f"⏸  Pause supervisée • {now}",
@@ -949,31 +1002,44 @@ class ChatWindow:
             font=("Segoe UI", 12, "bold"), anchor="w",
         ).pack(fill=tk.X, anchor=tk.W)
 
-        # Message scrollable pour les longs reasons (ex: 200+ chars depuis le serveur).
-        # On utilise un Text en mode read-only avec hauteur calculée selon la longueur.
-        # Patch 22 mai 2026 : prendre en compte les \n explicites (titres
-        # fenêtre / patterns) et activer la scrollbar dès que le cap de
-        # hauteur est atteint — sinon les bulles de pause étaient
-        # tronquées visuellement sans aucun ascenseur visible.
+        # Message borné et scrollable : sur une fenêtre Léa étroite, une
+        # bulle trop haute fait disparaître le début du diagnostic hors du
+        # viewport. On garde donc la bulle compacte et on scrolle le texte.
         reason_str = str(reason)
-        height_lines, needs_scroll = self._compute_paused_bubble_height(reason_str)
-        msg_frame = tk.Frame(inner, bg=PAUSED_BG)
-        msg_frame.pack(fill=tk.X, anchor=tk.W, pady=(6, 0))
-        reason_text = tk.Text(
-            msg_frame, bg=PAUSED_BG, fg=PAUSED_FG,
-            font=FONT_MSG, wrap=tk.WORD, bd=0, height=height_lines,
-            highlightthickness=0, relief=tk.FLAT, cursor="arrow",
+        _wrap_px, chars_per_line, max_rows = self._paused_text_layout()
+        text_rows, needs_text_scroll = self._compute_paused_bubble_height(
+            reason_str,
+            chars_per_line=chars_per_line,
+            max_rows=max_rows,
         )
-        reason_text.insert("1.0", reason_str)
-        reason_text.configure(state="disabled")
-        reason_text.pack(side=tk.LEFT, fill=tk.X, expand=True)
-        if needs_scroll:
-            reason_scroll = tk.Scrollbar(
-                msg_frame, orient=tk.VERTICAL,
-                command=reason_text.yview, width=8,
+        text_frame = tk.Frame(inner, bg=PAUSED_BG)
+        text_frame.pack(fill=tk.X, anchor=tk.W, pady=(6, 0))
+        reason_msg = tk.Text(
+            text_frame,
+            height=text_rows,
+            wrap=tk.WORD,
+            bg=PAUSED_BG,
+            fg=PAUSED_FG,
+            font=FONT_MSG,
+            bd=0,
+            highlightthickness=0,
+            relief=tk.FLAT,
+            padx=0,
+            pady=0,
+            cursor="arrow",
+        )
+        reason_msg.insert("1.0", reason_str)
+        reason_msg.configure(state="disabled")
+        reason_msg.pack(side=tk.LEFT, fill=tk.X, expand=True)
+        if needs_text_scroll:
+            scrollbar = tk.Scrollbar(
+                text_frame,
+                orient=tk.VERTICAL,
+                command=reason_msg.yview,
+                width=12,
             )
-            reason_text.configure(yscrollcommand=reason_scroll.set)
-            reason_scroll.pack(side=tk.RIGHT, fill=tk.Y)
+            reason_msg.configure(yscrollcommand=scrollbar.set)
+            scrollbar.pack(side=tk.RIGHT, fill=tk.Y, padx=(6, 0))
 
         tk.Label(
             inner, text=f"{workflow} — étape {completed}/{total}",
@@ -1018,6 +1084,89 @@ class ChatWindow:
         # Scroll automatique vers la nouvelle bulle (visible immédiatement)
         self._scroll_to_bottom()
 
+    def _render_paused_fallback_bubble(self, payload: Dict[str, Any]) -> None:
+        """Rendu minimal de secours si la bulle riche echoue."""
+        tk = self._tk
+        if getattr(self, "_msg_frame", None) is None:
+            return
+
+        replay_id = str(payload.get("replay_id", "") or "")
+        workflow = payload.get("workflow", "?")
+        reason = str(
+            payload.get("reason")
+            or "Action incertaine - j'ai besoin de votre validation."
+        )
+        completed = payload.get("completed", 0)
+        total = payload.get("total", "?")
+        now = datetime.now().strftime("%H:%M")
+
+        container = tk.Frame(self._msg_frame, bg=BG_COLOR)
+        container.pack(fill=tk.X, padx=MARGIN, pady=6)
+
+        inner = tk.Frame(
+            container, bg=PAUSED_BG, padx=14, pady=12,
+            highlightbackground=PAUSED_BORDER, highlightthickness=2,
+        )
+        inner.pack(anchor=tk.W, padx=(0, 12), fill=tk.X)
+
+        tk.Label(
+            inner, text=f"Pause supervisee - {now}",
+            bg=PAUSED_BG, fg=PAUSED_FG,
+            font=("Segoe UI", 12, "bold"), anchor="w",
+        ).pack(fill=tk.X, anchor=tk.W)
+
+        wrap_px = 360
+        try:
+            if self._canvas is not None:
+                wrap_px = max(220, int(self._canvas.winfo_width()) - 80)
+        except Exception:
+            pass
+        tk.Label(
+            inner, text=reason, bg=PAUSED_BG, fg=PAUSED_FG,
+            font=FONT_MSG, wraplength=wrap_px, justify=tk.LEFT,
+            anchor=tk.W,
+        ).pack(fill=tk.X, anchor=tk.W, pady=(6, 0))
+
+        tk.Label(
+            inner, text=f"{workflow} - etape {completed}/{total}",
+            bg=PAUSED_BG, fg=TIMESTAMP_FG, font=FONT_TIMESTAMP, anchor="w",
+        ).pack(fill=tk.X, anchor=tk.W, pady=(4, 8))
+
+        btn_frame = tk.Frame(inner, bg=PAUSED_BG)
+        btn_frame.pack(fill=tk.X, anchor=tk.W)
+
+        btn_resume = tk.Button(
+            btn_frame, text="Continuer",
+            bg=PAUSED_BTN_RESUME_BG, fg="white", font=FONT_QUICK_BTN,
+            padx=14, pady=4, bd=0, cursor="hand2",
+            activebackground=PAUSED_BTN_RESUME_HOVER, activeforeground="white",
+            command=lambda: self._on_paused_resume(replay_id),
+        )
+        btn_resume.pack(side=tk.LEFT, padx=(0, 8))
+
+        btn_abort = tk.Button(
+            btn_frame, text="Annuler",
+            bg=PAUSED_BTN_ABORT_BG, fg="white", font=FONT_QUICK_BTN,
+            padx=14, pady=4, bd=0, cursor="hand2",
+            activebackground=PAUSED_BTN_ABORT_HOVER, activeforeground="white",
+            command=lambda: self._on_paused_abort(replay_id),
+        )
+        btn_abort.pack(side=tk.LEFT)
+
+        feedback_label = tk.Label(
+            inner, text="", bg=PAUSED_BG, fg=PAUSED_FG,
+            font=FONT_TIMESTAMP, anchor="w",
+        )
+        feedback_label.pack(fill=tk.X, anchor=tk.W, pady=(6, 0))
+
+        self._active_paused_bubble = {
+            "container": container, "inner": inner,
+            "btn_resume": btn_resume, "btn_abort": btn_abort,
+            "feedback_label": feedback_label,
+            "replay_id": replay_id,
+        }
+        self._scroll_to_bottom()
+
     def _close_active_paused_bubble(self, reason: str) -> None:
         if self._active_paused_bubble is None or self._root is None:
             return
@@ -1524,8 +1673,19 @@ class ChatWindow:
             self._add_lea_message(
                 f"C'est parti ! Montrez-moi comment faire \u00ab {name} \u00bb."
             )
+
+            # --- P1-LEA-SHADOW : d\u00e9clencher d'abord l'orchestrateur L\u00e9a Linux ---
+            # On contacte agent-chat AVANT la capture locale : si la session
+            # serveur d\u00e9marre, on r\u00e9cup\u00e8re un session_id + un message d'accueil
+            # de L\u00e9a qu'on affiche dans le chat. Si \u00e9chec : mode d\u00e9grad\u00e9
+            # (capture locale uniquement, sans assistance conversationnelle).
+            self._start_lea_orchestrator_session(name)
+
+            # --- Comportement historique pr\u00e9serv\u00e9 : capture locale ---
+            # Le pipeline streaming (frames/\u00e9v\u00e9nements) reste pilot\u00e9 par
+            # agent_v1 local. L'orchestrateur Linux ne touche PAS \u00e0 la
+            # capture, il pilote uniquement le dialogue de fin de session.
             try:
-                # Utiliser l'etat partage si disponible (synchronise le systray)
                 if self._shared_state is not None:
                     self._shared_state.start_recording(name)
                 elif self._on_start_callback is not None:
@@ -1533,6 +1693,60 @@ class ChatWindow:
             except Exception as e:
                 self._add_lea_message(f"Oups, un probl\u00e8me : {e}")
 
+    def _start_lea_orchestrator_session(self, session_name: str) -> None:
+        """Appelle POST /api/learn/start c\u00f4t\u00e9 agent-chat Linux (P1-LEA-SHADOW).
+
+        Fail-safe : toute erreur (config absente, httpx manquant, timeout,
+        500 serveur...) bascule en mode d\u00e9grad\u00e9 sans bloquer la capture
+        locale. Un message clair est affich\u00e9 dans le chat.
+        """
+        try:
+            from ..config import AGENT_CHAT_URL, API_TOKEN, MACHINE_ID
+            from ..network.lea_orchestrator_client import (
+                LeaOrchestratorError,
+                start_learning_session,
+            )
+        except Exception as exc:  # pragma: no cover (import-time)
+            logger.error("Impossible de charger le client orchestrateur L\u00e9a : %s", exc)
+            self._add_lea_message(
+                "\u26a0 Impossible de joindre L\u00e9a serveur. "
+                "L'apprentissage continue localement, mais sans assistance "
+                "conversationnelle."
+            )
+            return
+
+        try:
+            resp = start_learning_session(
+                AGENT_CHAT_URL,
+                machine_id=MACHINE_ID,
+                session_name=session_name,
+                api_token=API_TOKEN,
+                trigger_source="windows_button",
+            )
+        except LeaOrchestratorError as exc:
+            logger.error("Orchestrateur L\u00e9a injoignable : %s", exc)
+            self._add_lea_message(
+                "\u26a0 Impossible de joindre L\u00e9a serveur. "
+                "L'apprentissage continue localement, mais sans assistance "
+                "conversationnelle."
+            )
+            return
+        except Exception as exc:  # noqa: BLE001 \u2014 d\u00e9fensif
+            logger.exception("Erreur inattendue orchestrateur L\u00e9a")
+            self._add_lea_message(
+                f"\u26a0 Erreur orchestrateur L\u00e9a : {exc}. "
+                "L'apprentissage continue localement."
+            )
+            return
+
+        # Affichage du message d'accueil renvoy\u00e9 par L\u00e9a (si pr\u00e9sent)
+        if resp.message:
+            self._add_lea_message(resp.message)
+        logger.info(
+            "Session orchestrateur L\u00e9a OK : id=%s state=%s",
+            resp.session_id, resp.state,
+        )
+
     def _on_quick_tasks(self) -> None:
         """Bouton Lancer — demande ce que L\u00e9a sait faire."""
         self._add_user_message("Qu'est-ce que vous savez faire ?")

agent_v0/agent_v1/ui/message_contract.py

@@ -0,0 +1,484 @@
+"""Contrat de lisibilite des messages visibles par l'humain.
+
+Ce module ne branche encore aucun point runtime. Il fournit une brique pure et
+testable pour que les sorties UI de Lea puissent refuser les messages trop
+generiques ou trop techniques avant affichage.
+"""
+
+from __future__ import annotations
+
+import logging
+import re
+import unicodedata
+from dataclasses import dataclass
+from typing import Iterable, Mapping
+
+
+logger = logging.getLogger(__name__)
+
+SUPERVISED_PAUSE_LABELS = (
+    "J'essaie de",
+    "J'attendais",
+    "Je vois",
+    "Peux-tu",
+)
+
+MAX_VISIBLE_MESSAGE_CHARS = 720
+MAX_FIELD_CHARS = 180
+MIN_FIELD_CHARS = 4
+
+_GENERIC_PHRASES = (
+    "un element",
+    "un élément",
+    "l'element",
+    "l'élément",
+    "element inconnu",
+    "élément inconnu",
+    "cette action",
+    "cette cible",
+    "cible inconnue",
+    "validation requise",
+    "action requise",
+)
+
+_ACTIONABLE_FRENCH_HINTS = (
+    "peux-tu",
+    "cliquer",
+    "ouvrir",
+    "selectionner",
+    "sélectionner",
+    "choisir",
+    "saisir",
+    "corriger",
+    "montrer",
+    "indiquer",
+    "valider",
+    "fermer",
+    "placer",
+    "mettre",
+    "reprendre",
+)
+
+_TECHNICAL_ENGLISH_TERMS = (
+    "target_not_found",
+    "target not found",
+    "no_screen_change",
+    "no screen change",
+    "wrong_window",
+    "wrong window",
+    "validation required",
+    "retry",
+    "fallback",
+    "timeout",
+    "screenshot",
+    "validator",
+    "failure",
+    "failed",
+    "resolve target",
+    "postcondition",
+    "please",
+    "click",
+    "button",
+    "target",
+    "expected",
+    "actual",
+    "observed",
+)
+
+_TECHNICAL_FIELD_RE = re.compile(
+    r"\b(?:"
+    r"action_id|replay_id|session_id|workflow_id|machine_id|target_spec|"
+    r"vlm_description|resolution_method|resolution_score|retry_count|"
+    r"x_pct|y_pct|screenshot_b64|expected_window_title|current_action_index"
+    r")\b",
+    re.IGNORECASE,
+)
+_TECHNICAL_IDENTIFIER_RE = re.compile(
+    r"\b(?:action|replay|session|sess|workflow|node|edge|target|retry|"
+    r"precheck|wait|trace|event|machine|run)_[A-Za-z0-9][A-Za-z0-9_.:-]{3,}\b"
+)
+_UUID_RE = re.compile(
+    r"\b[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}\b",
+    re.IGNORECASE,
+)
+_LONG_HEX_RE = re.compile(r"\b[0-9a-f]{16,}\b", re.IGNORECASE)
+_PIXEL_TUPLE_RE = re.compile(r"\(\s*\d{2,5}\s*,\s*\d{2,5}\s*\)")
+_PIXEL_FIELD_RE = re.compile(
+    r"\b(?:x|y|left|top|width|height|w|h|x_pct|y_pct)\s*[=:]\s*-?\d+(?:[.,]\d+)?",
+    re.IGNORECASE,
+)
+_PX_RE = re.compile(r"\b\d{2,5}\s*px\b", re.IGNORECASE)
+_SCORE_RE = re.compile(
+    r"\b(?:score|confidence|confiance|similarit[eé]|threshold|seuil|"
+    r"probabilit[eé])\s*[:=]\s*\d+(?:[.,]\d+)?%?\b",
+    re.IGNORECASE,
+)
+
+@dataclass(frozen=True)
+class MessageValidationIssue:
+    """Un probleme detecte dans un message visible par l'humain."""
+
+    code: str
+    detail: str
+
+
+@dataclass(frozen=True)
+class MessageValidationResult:
+    """Resultat de validation d'un message utilisateur."""
+
+    issues: tuple[MessageValidationIssue, ...] = ()
+
+    @property
+    def valid(self) -> bool:
+        return not self.issues
+
+    def raise_for_errors(self) -> None:
+        if not self.valid:
+            raise MessageContractError(self)
+
+
+class MessageContractError(ValueError):
+    """Erreur levee quand un message ne respecte pas le contrat humain."""
+
+    def __init__(self, result: MessageValidationResult):
+        self.result = result
+        details = "; ".join(f"{issue.code}: {issue.detail}" for issue in result.issues)
+        super().__init__(f"Message humain invalide: {details}")
+
+
+@dataclass(frozen=True)
+class SupervisedPauseFields:
+    """Champs obligatoires pour expliquer une pause supervisee."""
+
+    intention: str
+    attendu: str
+    vu: str
+    demande: str
+
+
+DEFAULT_SUPERVISED_PAUSE_FIELDS = SupervisedPauseFields(
+    intention="continuer une etape supervisee",
+    attendu="un accord humain clair avant de continuer",
+    vu="je suis sur une etape qui demande une verification humaine",
+    demande="indiquer si je peux continuer ou corriger l'action attendue",
+)
+
+
+def format_supervised_pause_message(
+    *,
+    intention: str,
+    attendu: str,
+    vu: str,
+    demande: str,
+) -> str:
+    """Formatter une pause supervisee claire et actionnable.
+
+    Le message retourne exactement quatre lignes. Si un champ reste vague ou
+    technique, la fonction leve ``MessageContractError`` au lieu de produire un
+    message degradant pour l'utilisateur.
+    """
+
+    fields = SupervisedPauseFields(
+        intention=_one_line(intention),
+        attendu=_one_line(attendu),
+        vu=_one_line(vu),
+        demande=_one_line(demande),
+    )
+    message = "\n".join(
+        (
+            f"J'essaie de : {fields.intention}",
+            f"J'attendais : {fields.attendu}",
+            f"Je vois     : {fields.vu}",
+            f"Peux-tu     : {fields.demande}",
+        )
+    )
+    validate_supervised_pause_message(message).raise_for_errors()
+    return message
+
+
+def format_supervised_pause_from_mapping(payload: Mapping[str, object]) -> str:
+    """Formatter depuis un mapping runtime avec noms de champs explicites.
+
+    Alias acceptes pour faciliter l'integration progressive:
+    ``intention|trying_to``, ``attendu|expected``, ``vu|observed``,
+    ``demande|request``.
+    """
+
+    return format_supervised_pause_message(
+        intention=_mapping_text(payload, "intention", "trying_to"),
+        attendu=_mapping_text(payload, "attendu", "expected"),
+        vu=_mapping_text(payload, "vu", "observed"),
+        demande=_mapping_text(payload, "demande", "request"),
+    )
+
+
+def coerce_supervised_pause_message(
+    message: object = "",
+    *,
+    intention: object = "",
+    attendu: object = "",
+    vu: object = "",
+    demande: object = "",
+) -> str:
+    """Retourner une pause supervisee valide, meme depuis un ancien message.
+
+    Si ``message`` respecte deja le contrat strict, il est conserve. Sinon on
+    compose les quatre champs avec les valeurs explicites disponibles. Les
+    valeurs trop vagues ou techniques sont remplacees par des fallbacks clairs.
+    """
+
+    raw_message = _one_line(message)
+    if raw_message and validate_supervised_pause_message(raw_message).valid:
+        return raw_message
+
+    defaults = DEFAULT_SUPERVISED_PAUSE_FIELDS
+    candidates = SupervisedPauseFields(
+        intention=_safe_field_text(intention, defaults.intention),
+        attendu=_safe_field_text(attendu, defaults.attendu),
+        vu=_safe_field_text(vu, defaults.vu),
+        demande=_safe_field_text(demande or raw_message, defaults.demande),
+    )
+
+    try:
+        return format_supervised_pause_message(
+            intention=candidates.intention,
+            attendu=candidates.attendu,
+            vu=candidates.vu,
+            demande=candidates.demande,
+        )
+    except MessageContractError:
+        return format_supervised_pause_message(
+            intention=defaults.intention,
+            attendu=defaults.attendu,
+            vu=defaults.vu,
+            demande=defaults.demande,
+        )
+
+
+def warn_visible_message(
+    message: object,
+    *,
+    source: str,
+    supervised_pause: bool = False,
+) -> str:
+    """Log contract violations without modifying the visible message."""
+
+    text = str(message or "")
+    validator = validate_supervised_pause_message if supervised_pause else validate_visible_message
+    result = validator(text)
+    if not result.valid:
+        logger.warning(
+            "[message_contract] invalid_message source=%s codes=%s",
+            source,
+            [issue.code for issue in result.issues],
+        )
+    return text
+
+
+def validate_supervised_pause_message(message: str) -> MessageValidationResult:
+    """Valider le contrat strict d'une pause supervisee."""
+
+    issues = list(validate_visible_message(message).issues)
+    fields, structure_issues = _parse_supervised_pause(message)
+    issues.extend(structure_issues)
+
+    if fields:
+        for name, value in fields.items():
+            if len(value) < MIN_FIELD_CHARS:
+                issues.append(
+                    MessageValidationIssue(
+                        "field_too_short",
+                        f"{name} doit etre explicite",
+                    )
+                )
+            if len(value) > MAX_FIELD_CHARS:
+                issues.append(
+                    MessageValidationIssue(
+                        "field_too_long",
+                        f"{name} depasse {MAX_FIELD_CHARS} caracteres",
+                    )
+                )
+        demande = fields.get("demande", "")
+        if not _contains_actionable_french(demande) or len(demande.split()) < 4:
+            issues.append(
+                MessageValidationIssue(
+                    "not_actionable",
+                    "la demande doit contenir une action concrete en francais",
+                )
+            )
+
+    return _dedupe_issues(issues)
+
+
+def validate_visible_message(message: str) -> MessageValidationResult:
+    """Valider qu'un message visible n'est ni generique ni technique."""
+
+    text = str(message or "").strip()
+    issues: list[MessageValidationIssue] = []
+
+    if not text:
+        return MessageValidationResult(
+            (MessageValidationIssue("empty_message", "message vide"),)
+        )
+
+    if len(text) > MAX_VISIBLE_MESSAGE_CHARS:
+        issues.append(
+            MessageValidationIssue(
+                "message_too_long",
+                f"message au-dela de {MAX_VISIBLE_MESSAGE_CHARS} caracteres",
+            )
+        )
+
+    folded = _fold(text)
+    seen_generic_phrases: set[str] = set()
+    for phrase in _GENERIC_PHRASES:
+        folded_phrase = _fold(phrase)
+        if folded_phrase in seen_generic_phrases:
+            continue
+        seen_generic_phrases.add(folded_phrase)
+        if folded_phrase in folded:
+            issues.append(
+                MessageValidationIssue(
+                    "generic_phrase",
+                    f"formulation trop generique: {phrase}",
+                )
+            )
+
+    for term in _TECHNICAL_ENGLISH_TERMS:
+        if _fold(term) in folded:
+            issues.append(
+                MessageValidationIssue(
+                    "technical_english",
+                    f"anglais technique visible: {term}",
+                )
+            )
+
+    for code, pattern, detail in (
+        ("technical_field", _TECHNICAL_FIELD_RE, "champ technique brut"),
+        ("technical_identifier", _TECHNICAL_IDENTIFIER_RE, "identifiant technique brut"),
+        ("technical_identifier", _UUID_RE, "UUID brut"),
+        ("technical_identifier", _LONG_HEX_RE, "hash technique brut"),
+        ("raw_coordinates", _PIXEL_TUPLE_RE, "coordonnees pixel brutes"),
+        ("raw_coordinates", _PIXEL_FIELD_RE, "coordonnees techniques brutes"),
+        ("raw_coordinates", _PX_RE, "coordonnees pixel brutes"),
+        ("raw_score", _SCORE_RE, "score ou confiance brut"),
+    ):
+        if pattern.search(text):
+            issues.append(MessageValidationIssue(code, detail))
+
+    return _dedupe_issues(issues)
+
+
+def is_valid_visible_message(message: str) -> bool:
+    """Raccourci booleen pour les points d'integration UI."""
+
+    return validate_visible_message(message).valid
+
+
+def is_valid_supervised_pause_message(message: str) -> bool:
+    """Raccourci booleen pour les pauses supervisees."""
+
+    return validate_supervised_pause_message(message).valid
+
+
+def _parse_supervised_pause(
+    message: str,
+) -> tuple[dict[str, str], list[MessageValidationIssue]]:
+    lines = [line.rstrip() for line in str(message or "").splitlines() if line.strip()]
+    issues: list[MessageValidationIssue] = []
+
+    if len(lines) != 4:
+        issues.append(
+            MessageValidationIssue(
+                "invalid_structure",
+                "une pause supervisee doit contenir exactement 4 lignes",
+            )
+        )
+        return {}, issues
+
+    specs = (
+        ("intention", r"^J'essaie de\s*:\s*(.+)$"),
+        ("attendu", r"^J'attendais\s*:\s*(.+)$"),
+        ("vu", r"^Je vois\s*:\s*(.+)$"),
+        ("demande", r"^Peux-tu\s*:\s*(.+)$"),
+    )
+    fields: dict[str, str] = {}
+    for line, (name, pattern) in zip(lines, specs):
+        match = re.match(pattern, line)
+        if not match:
+            issues.append(
+                MessageValidationIssue(
+                    "invalid_structure",
+                    f"ligne {len(fields) + 1} doit commencer par {SUPERVISED_PAUSE_LABELS[len(fields)]}",
+                )
+            )
+            continue
+        fields[name] = match.group(1).strip()
+
+    if len(fields) != 4:
+        return {}, issues
+
+    return fields, issues
+
+
+def _contains_actionable_french(text: str) -> bool:
+    folded = _fold(text)
+    return any(_fold(hint) in folded for hint in _ACTIONABLE_FRENCH_HINTS)
+
+
+def _one_line(value: object) -> str:
+    return re.sub(r"\s+", " ", str(value or "")).strip()
+
+
+def _mapping_text(payload: Mapping[str, object], *keys: str) -> str:
+    for key in keys:
+        value = payload.get(key)
+        if value is not None:
+            return str(value)
+    return ""
+
+
+def _safe_field_text(value: object, fallback: str) -> str:
+    text = _one_line(value)
+    if len(text) < MIN_FIELD_CHARS or len(text) > MAX_FIELD_CHARS:
+        return fallback
+    if not validate_visible_message(text).valid:
+        return fallback
+    return text
+
+
+def _fold(text: str) -> str:
+    normalized = unicodedata.normalize("NFKD", str(text or ""))
+    ascii_text = "".join(ch for ch in normalized if not unicodedata.combining(ch))
+    return ascii_text.casefold()
+
+
+def _dedupe_issues(issues: Iterable[MessageValidationIssue]) -> MessageValidationResult:
+    seen: set[tuple[str, str]] = set()
+    deduped: list[MessageValidationIssue] = []
+    for issue in issues:
+        key = (issue.code, issue.detail)
+        if key in seen:
+            continue
+        seen.add(key)
+        deduped.append(issue)
+    return MessageValidationResult(tuple(deduped))
+
+
+__all__ = [
+    "MAX_FIELD_CHARS",
+    "MAX_VISIBLE_MESSAGE_CHARS",
+    "MessageContractError",
+    "MessageValidationIssue",
+    "MessageValidationResult",
+    "SUPERVISED_PAUSE_LABELS",
+    "SupervisedPauseFields",
+    "coerce_supervised_pause_message",
+    "format_supervised_pause_from_mapping",
+    "format_supervised_pause_message",
+    "is_valid_supervised_pause_message",
+    "is_valid_visible_message",
+    "validate_supervised_pause_message",
+    "validate_visible_message",
+    "warn_visible_message",
+]

agent_v0/agent_v1/ui/messages.py

@@ -82,6 +82,12 @@ ICONE_PAR_NIVEAU: dict[NiveauMessage, str] = {
     NiveauMessage.BLOCAGE: "?",
 }
 
+# Les pauses supervisees peuvent contenir une raison precise, parfois longue
+# (fenetre observee, fenetre attendue, action en cours). On garde l'information
+# utile et on laisse les widgets UI gerer le wrap/scroll.
+MAX_TARGET_DESCRIPTION_CHARS = 1024
+MAX_GENERIC_TECHNICAL_MESSAGE_CHARS = 1024
+
 
 @dataclass
 class MessageUtilisateur:
@@ -147,9 +153,9 @@ def _nettoyer_description_cible(description: str) -> str:
     desc = description.strip()
     # Retirer les guillemets encapsulants
     desc = desc.strip("'\"`")
-    # Limiter la longueur
-    if len(desc) > 80:
-        desc = desc[:77] + "..."
+    # Limiter la longueur sans perdre les details utiles a la supervision.
+    if len(desc) > MAX_TARGET_DESCRIPTION_CHARS:
+        desc = desc[: MAX_TARGET_DESCRIPTION_CHARS - 3] + "..."
     return desc
 
 
@@ -566,8 +572,8 @@ def formatter_erreur_generique(
 
     # Fallback : message technique tronqué
     msg_tronque = message_technique.strip()
-    if len(msg_tronque) > 120:
-        msg_tronque = msg_tronque[:117] + "..."
+    if len(msg_tronque) > MAX_GENERIC_TECHNICAL_MESSAGE_CHARS:
+        msg_tronque = msg_tronque[: MAX_GENERIC_TECHNICAL_MESSAGE_CHARS - 3] + "..."
 
     return MessageUtilisateur(
         niveau=NiveauMessage.ATTENTION,

agent_v0/agent_v1/ui/smart_tray.py

@@ -371,7 +371,13 @@ class SmartTrayV1:
             )
             if name and name.strip():
                 name = name.strip()
-                # Utiliser l'etat partage si disponible
+
+                # --- P1-LEA-SHADOW : d\u00e9clencher d'abord l'orchestrateur L\u00e9a Linux ---
+                # On contacte agent-chat AVANT la capture locale. Si \u00e9chec,
+                # bascule en mode d\u00e9grad\u00e9 (capture locale sans assistance).
+                self._start_lea_orchestrator_session(name)
+
+                # --- Comportement historique pr\u00e9serv\u00e9 : capture locale ---
                 if self._shared_state is not None:
                     try:
                         self._shared_state.start_recording(name)
@@ -393,6 +399,55 @@ class SmartTrayV1:
 
         threading.Thread(target=_dialog, daemon=True).start()
 
+    def _start_lea_orchestrator_session(self, session_name: str) -> None:
+        """Appelle POST /api/learn/start côté agent-chat Linux (P1-LEA-SHADOW).
+
+        Fail-safe : toute erreur (config absente, httpx manquant, timeout,
+        5xx serveur...) bascule en mode dégradé sans bloquer la capture
+        locale. L'utilisateur est informé via le NotificationManager.
+        """
+        try:
+            from ..config import AGENT_CHAT_URL, API_TOKEN, MACHINE_ID
+            from ..network.lea_orchestrator_client import (
+                LeaOrchestratorError,
+                start_learning_session,
+            )
+        except Exception as exc:  # pragma: no cover (import-time)
+            logger.error("Impossible de charger le client orchestrateur Léa : %s", exc)
+            self._notifier.notify(
+                "Léa",
+                "Serveur injoignable — apprentissage local uniquement.",
+            )
+            return
+
+        try:
+            resp = start_learning_session(
+                AGENT_CHAT_URL,
+                machine_id=MACHINE_ID,
+                session_name=session_name,
+                api_token=API_TOKEN,
+                trigger_source="tray_button",
+            )
+        except LeaOrchestratorError as exc:
+            logger.error("Orchestrateur Léa injoignable : %s", exc)
+            self._notifier.notify(
+                "Léa",
+                "Serveur injoignable — apprentissage local uniquement.",
+            )
+            return
+        except Exception:  # noqa: BLE001 — défensif
+            logger.exception("Erreur inattendue orchestrateur Léa")
+            self._notifier.notify(
+                "Léa",
+                "Erreur orchestrateur — apprentissage local uniquement.",
+            )
+            return
+
+        logger.info(
+            "Session orchestrateur Léa OK : id=%s state=%s",
+            resp.session_id, resp.state,
+        )
+
     def _on_stop_session(self, _icon=None, _item=None) -> None:
         """Termine la session en cours et envoie les donnees."""
         count = self.actions_count

agent_v0/deploy/windows_client/agent_v1/core/captor.py

@@ -43,6 +43,9 @@ class EventCaptorV1:
 
         # État des touches modificatrices
         self.modifiers = set()
+        self._pending_standalone_win = False
+        self._suppress_release_only_win_combo = False
+        self._raw_key_buffer: List[Dict[str, Any]] = []
 
         # Tracking du focus fenêtre
         self.last_window = None
@@ -91,6 +94,7 @@ class EventCaptorV1:
         # Flush du buffer texte restant avant arrêt
         self._flush_text_buffer()
         # Annuler le timer s'il est en cours
+        emit_escape = False
         with self._text_lock:
             if self._text_flush_timer is not None:
                 self._text_flush_timer.cancel()
@@ -159,7 +163,80 @@ class EventCaptorV1:
     # Clavier
     # ----------------------------------------------------------------
 
+    @staticmethod
+    def _get_key_name(key) -> Optional[str]:
+        """Convertit un objet pynput Key/KeyCode en nom lisible."""
+        if isinstance(key, KeyCode):
+            return key.char if key.char else None
+        if isinstance(key, Key):
+            return key.name
+        return str(key)
+
+    @staticmethod
+    def _encode_key(key) -> Dict[str, Any]:
+        if isinstance(key, KeyCode):
+            return {"kind": "vk", "vk": key.vk, "char": key.char}
+        if isinstance(key, Key):
+            return {"kind": "key", "name": key.name}
+        return {"kind": "unknown", "str": str(key)}
+
+    @staticmethod
+    def _raw_key_name(raw_key: Dict[str, Any]) -> Optional[str]:
+        if raw_key.get("kind") == "vk":
+            char = raw_key.get("char")
+            if char and len(str(char)) == 1:
+                return str(char).lower()
+        if raw_key.get("kind") == "key":
+            name = raw_key.get("name")
+            return str(name).lower() if name else None
+        return None
+
+    def _emit_release_only_windows_combo(self) -> bool:
+        """Infère Win+<touche> quand seuls les releases sont capturés."""
+        with self._text_lock:
+            raw_keys = list(getattr(self, "_raw_key_buffer", []))
+            if len(raw_keys) < 2:
+                return False
+            cmd_names = {"cmd", "cmd_l", "cmd_r"}
+            last = raw_keys[-1]
+            if last.get("action") != "release" or self._raw_key_name(last) not in cmd_names:
+                return False
+            combo_key = None
+            modifier_names = {
+                "ctrl", "ctrl_l", "ctrl_r",
+                "alt", "alt_l", "alt_r",
+                "shift", "shift_l", "shift_r",
+                "cmd", "cmd_l", "cmd_r",
+            }
+            for raw in reversed(raw_keys[:-1]):
+                if raw.get("action") != "release":
+                    continue
+                name = self._raw_key_name(raw)
+                if name and name not in modifier_names:
+                    combo_key = name
+                    break
+            if not combo_key:
+                return False
+            self._raw_key_buffer.clear()
+
+        event = {
+            "type": "key_combo",
+            "keys": ["win", combo_key],
+            "raw_keys": raw_keys,
+            "timestamp": time.time(),
+        }
+        self.on_event(event)
+        return True
+
     def _on_press(self, key):
+        with self._text_lock:
+            if not hasattr(self, "_raw_key_buffer"):
+                self._raw_key_buffer = []
+            self._raw_key_buffer.append({
+                "action": "press",
+                **self._encode_key(key),
+            })
+
         # Gestion des touches modificatrices
         if key in (Key.ctrl, Key.ctrl_l, Key.ctrl_r):
             self.modifiers.add("ctrl")
@@ -167,15 +244,26 @@ class EventCaptorV1:
             self.modifiers.add("alt")
         elif key in (Key.shift, Key.shift_l, Key.shift_r):
             self.modifiers.add("shift")
+        elif key in (Key.cmd, Key.cmd_l, Key.cmd_r):
+            self.modifiers.add("win")
+            self._pending_standalone_win = True
 
         # --- Combos avec modificateur (sauf Shift seul) ---
         # Shift seul n'est pas un « vrai » modificateur pour les combos :
         # Shift+a = 'A' = saisie texte, pas un raccourci.
-        # On considère un combo seulement si Ctrl ou Alt est enfoncé.
-        has_real_modifier = self.modifiers & {"ctrl", "alt"}
+        # On considère un combo seulement si Ctrl, Alt ou Win est enfoncé.
+        has_real_modifier = self.modifiers & {"ctrl", "alt", "win"}
         if has_real_modifier:
             key_name = self._get_key_name(key)
-            if key_name and key_name not in ("ctrl", "alt", "shift"):
+            if key_name and key_name not in (
+                "ctrl", "ctrl_l", "ctrl_r",
+                "alt", "alt_l", "alt_r",
+                "shift", "shift_l", "shift_r",
+                "cmd", "cmd_l", "cmd_r",
+            ):
+                self._pending_standalone_win = False
+                if "win" in self.modifiers:
+                    self._suppress_release_only_win_combo = True
                 # Un combo interrompt la saisie texte en cours
                 self._flush_text_buffer()
                 event = {
@@ -205,14 +293,18 @@ class EventCaptorV1:
                 self._reset_flush_timer()
                 return
 
-            if key == Key.escape:
+            escape_keys = [Key.esc]
+            key_escape = getattr(Key, "escape", None)
+            if key_escape is not None:
+                escape_keys.append(key_escape)
+            if key in escape_keys:
                 # Annuler la saisie en cours
                 self._text_buffer.clear()
                 self._text_start_pos = None
                 self._cancel_flush_timer()
-                return
+                emit_escape = True
 
-            if key in (Key.enter, Key.tab):
+            elif key in (Key.enter, Key.tab):
                 # Flush immédiat — on relâche le lock avant d'appeler
                 # _flush_text_buffer (qui prend aussi le lock)
                 pass  # on sort du with et on flush après
@@ -238,6 +330,15 @@ class EventCaptorV1:
                 # Touche spéciale non gérée (F1, Insert, etc.) — on ignore
                 return
 
+        if emit_escape:
+            event = {
+                "type": "key_combo",
+                "keys": ["escape"],
+                "timestamp": time.time(),
+            }
+            self.on_event(event)
+            return
+
         # Si on arrive ici, c'est Enter ou Tab → flush immédiat
         self._flush_text_buffer()
 
@@ -290,12 +391,46 @@ class EventCaptorV1:
         self.on_event(event)
 
     def _on_release(self, key):
+        with self._text_lock:
+            self._raw_key_buffer.append({
+                "action": "release",
+                **self._encode_key(key),
+            })
+
+        if key in (Key.cmd, Key.cmd_l, Key.cmd_r) and self._suppress_release_only_win_combo:
+            with self._text_lock:
+                self._raw_key_buffer.clear()
+            self._pending_standalone_win = False
+            self._suppress_release_only_win_combo = False
+            self.modifiers.discard("win")
+            return
+
+        if key in (Key.cmd, Key.cmd_l, Key.cmd_r) and self._emit_release_only_windows_combo():
+            self._pending_standalone_win = False
+            self._suppress_release_only_win_combo = False
+            self.modifiers.discard("win")
+            return
+
+        if key in (Key.cmd, Key.cmd_l, Key.cmd_r) and self._pending_standalone_win:
+            event = {
+                "type": "key_combo",
+                "keys": ["win"],
+                "timestamp": time.time(),
+            }
+            self.on_event(event)
+            self._pending_standalone_win = False
+            self._suppress_release_only_win_combo = False
+
         if key in (Key.ctrl, Key.ctrl_l, Key.ctrl_r):
             self.modifiers.discard("ctrl")
         elif key in (Key.alt, Key.alt_l, Key.alt_r):
             self.modifiers.discard("alt")
         elif key in (Key.shift, Key.shift_l, Key.shift_r):
             self.modifiers.discard("shift")
+        elif key in (Key.cmd, Key.cmd_l, Key.cmd_r):
+            self.modifiers.discard("win")
+            self._pending_standalone_win = False
+            self._suppress_release_only_win_combo = False
 
     def _watch_window_focus(self):
         """Surveille proactivement le changement de fenêtre pour le stagiaire."""