Client MCP (scheda Strumenti, filtri per server, Cua Driver), riepilogo compresso della cronologia, memoria in Markdown/Obsidian, plugin immagini con mflux

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

.gitignore

@@ -8,6 +8,7 @@ whatsapp_bridge/node_modules/
 data/
 config/api_keys.json
 config/settings.json
+config/mcp.json
 memory/long_term.json
 *.session
 *.session-journal
@@ -20,4 +21,5 @@ avatar3d/models_archivio/
 *.log
 .DS_Store
 .venv-chatterbox/
+.venv-mflux/
 tts_chatterbox/test_*.py

avatar/engines/base.py

@@ -67,3 +67,63 @@ class History:
     def clear(self) -> None:
         self.messages, self.meta = [], {}
         self.save()
+
+
+# ── Compattazione della cronologia (motori locali) ───────────────────────────
+COMPACT_AFTER = 30   # oltre questo numero di messaggi i più vecchi vengono riassunti
+COMPACT_KEEP = 12    # messaggi recenti lasciati per esteso
+
+
+def render_messages(msgs: list, per_msg: int = 600) -> str:
+    out = []
+    for m in msgs:
+        role = m.get("role")
+        text = str(m.get("content") or "").strip()
+        if role == "user":
+            out.append("Utente: " + text[:per_msg])
+        elif role == "assistant":
+            calls = m.get("tool_calls") or []
+            if calls:
+                names = ", ".join(str((c.get("function") or {}).get("name", "?")) for c in calls)
+                text = (text + f" [usa strumenti: {names}]").strip()
+            if text:
+                out.append("Assistente: " + text[:per_msg])
+        elif role == "tool":
+            out.append(f"Risultato di {m.get('name', 'strumento')}: " + text[:200])
+    return "\n".join(out)
+
+
+def summary_prompt(previous: str, new_text: str) -> str:
+    return ("Riassumi in italiano, in modo compatto (al massimo 150 parole), ciò che serve per continuare la conversazione: "
+            "richieste dell'utente, cose fatte o decise, fatti e preferenze emersi, questioni aperte. Niente saluti, niente frasi generiche.\n\n"
+            + (f"Riassunto precedente:\n{previous}\n\n" if previous else "")
+            + f"Nuovi messaggi:\n{new_text}\n\nRiassunto aggiornato:")
+
+
+def compact_history(history: "History", summarize: Callable[[str], str]) -> bool:
+    """Se la cronologia è lunga, riassume i messaggi più vecchi in history.meta['summary'] e li rimuove."""
+    msgs = history.messages
+    if len(msgs) <= COMPACT_AFTER:
+        return False
+    cut = len(msgs) - COMPACT_KEEP
+    while cut < len(msgs) and msgs[cut].get("role") != "user":
+        cut += 1
+    if cut <= 0 or cut >= len(msgs):
+        return False
+    text = render_messages(msgs[:cut])
+    try:
+        summary = (summarize(summary_prompt(str(history.meta.get("summary") or ""), text)) or "").strip()
+    except Exception as err:
+        print(f"[cronologia] riassunto fallito: {err}")
+        return False
+    if not summary:
+        return False
+    history.meta["summary"] = summary[:2000]
+    history.messages = msgs[cut:]
+    history.save()
+    return True
+
+
+def summary_block(history: "History") -> str:
+    s = str(history.meta.get("summary") or "").strip()
+    return f"\n\nRiassunto della conversazione precedente (i messaggi più vecchi sono stati compattati):\n{s}" if s else ""

avatar/engines/claude_code.py

@@ -112,10 +112,33 @@ class ClaudeCodeEngine:
                 flags["tools"] = flags["tools"] + ",Read"
             flags["allowed"] = [*flags["allowed"], f"Read(//{attach_path.lstrip('/')})"]
         mcp = {"mcpServers": {"avatar": {"command": sys.executable, "args": [str(MCP_SCRIPT)]}}}
+        allowed_mcp = list(MEMORY_TOOLS)
+        disallowed_mcp: list[str] = []
+        try:
+            from avatar.mcp_client import load_config
+            for name, c in load_config().items():
+                if c.get("disabled") or name == "avatar":
+                    continue
+                mcp["mcpServers"][name] = {k: v for k, v in c.items() if k in ("command", "args", "env", "cwd", "url", "headers", "type")}
+                if c.get("url") and "type" not in c:
+                    mcp["mcpServers"][name]["type"] = "http"
+                if c.get("tools") or c.get("exclude"):
+                    # Filtri: consenti solo gli strumenti scelti e nega gli altri (l'elenco completo arriva dal client in-app, se collegato).
+                    from avatar.mcp_client import manager, selected_tools
+                    srv = manager.servers.get(name)
+                    if srv is not None:
+                        chosen = {t["name"] for t in selected_tools(c, srv.tools)}
+                        allowed_mcp += [f"mcp__{name}__{t}" for t in sorted(chosen)]
+                        disallowed_mcp += [f"mcp__{name}__{t['name']}" for t in srv.tools if t["name"] not in chosen]
+                        continue
+                allowed_mcp.append(f"mcp__{name}")
+        except Exception as err:
+            print(f"[claude code] server MCP esterni ignorati: {err}")
         args = ["-p", "--output-format", "stream-json", "--verbose", "--include-partial-messages",
                 "--model", self.model, "--effort", self.effort,
-                "--tools", flags["tools"], "--allowedTools", ",".join(flags["allowed"] + MEMORY_TOOLS),
+                "--tools", flags["tools"], "--allowedTools", ",".join(flags["allowed"] + allowed_mcp),
                 "--mcp-config", json.dumps(mcp), "--strict-mcp-config",
+                *(["--disallowedTools", ",".join(disallowed_mcp)] if disallowed_mcp else []),
                 "--system-prompt-snapshot", "off",
                 "--resume" if resume else "--session-id", session_id]
         if flags["mode"]:

avatar/engines/mlx_engine.py

@@ -16,7 +16,7 @@ from pathlib import Path
 from avatar.memory_tools import memory_prompt, openai_tools, parse_args, run_tool
 from avatar.plugins import registry
 from avatar.websearch import brave_search
-from .base import Emit, History, persona_text, today_label, user_block
+from .base import Emit, History, compact_history, persona_text, summary_block, today_label, user_block
 from .openai_compat import SEARCH_TOOL, Aborted
 
 MAX_ROUNDS = 8
@@ -323,7 +323,7 @@ class MLXEngine:
                 "- Chiama gli strumenti solo nel formato previsto dal tuo template e mai descrivendoli a parole.")
         note += ("\n- Per informazioni aggiornate chiama cerca_web e rispondi in base ai risultati." if self.search_api_key
                  else "\n- Non hai accesso al web: se ti chiedono informazioni aggiornate, dillo chiaramente.")
-        return f"{persona_text(self.assistant_name)}\n\n{user_block(self.user_name, memory_prompt())}{note}"
+        return f"{persona_text(self.assistant_name)}\n\n{user_block(self.user_name, memory_prompt())}{summary_block(self.history)}{note}"
 
     def _recent(self) -> list:
         msgs = self.history.messages
@@ -443,6 +443,9 @@ class MLXEngine:
             full = full.strip() or "(nessuna risposta)"
             self.history.save()
             emit({"type": "done", "text": full, "sources": sources})
+            with _lock:
+                if compact_history(self.history, self.summarize):
+                    _loaded["snap"] = None   # il prompt cambia: la fotografia non vale più
         except Aborted:
             self._rollback()
             emit({"type": "error", "message": "Risposta interrotta.", "aborted": True})
@@ -450,6 +453,22 @@ class MLXEngine:
             self._rollback()
             emit({"type": "error", "message": f"Modello interno: {err}"})
 
+    def summarize(self, prompt: str) -> str:
+        """Genera un testo breve col modello caricato, senza strumenti né cache condivisa (usato per compattare la cronologia)."""
+        from mlx_lm import stream_generate
+        from mlx_lm.sample_utils import make_sampler
+        from mlx_lm.models.cache import make_prompt_cache
+        msgs = [{"role": "system", "content": "Sei un assistente che riassume conversazioni in italiano, in modo fedele e compatto."},
+                {"role": "user", "content": prompt}]
+        text = self.tokenizer.apply_chat_template(msgs, add_generation_prompt=True, tokenize=False, enable_thinking=False)
+        f = FORMATS[self.fmt]
+        think = TagFilter(*f["think"], keep=False)
+        think.inside = text.rstrip().endswith(f["think"][0])
+        out = ""
+        for r in stream_generate(self.model, self.tokenizer, text, max_tokens=400, sampler=make_sampler(temp=0.3, top_p=0.9), prompt_cache=make_prompt_cache(self.model)):
+            out += think.push(r.text)
+        return (out + think.flush()).strip()
+
     def _rollback(self) -> None:
         msgs = self.history.messages
         if msgs and msgs[-1].get("role") == "user":

avatar/engines/openai_compat.py

@@ -9,7 +9,7 @@ import openai
 from avatar.memory_tools import memory_prompt, openai_tools, parse_args, run_tool
 from avatar.plugins import registry
 from avatar.websearch import brave_search
-from .base import Emit, History, persona_text, today_label, user_block
+from .base import Emit, History, compact_history, persona_text, summary_block, today_label, user_block
 
 MAX_ROUNDS = 8
 MAX_HISTORY = 30
@@ -98,7 +98,7 @@ class OpenAICompatEngine:
                      else "\n- Non hai accesso al web: se ti chiedono informazioni aggiornate, dillo chiaramente.")
         else:
             note = "\n\nNota: in questa modalità non hai strumenti (niente memoria automatica né ricerca web)."
-        return {"role": "system", "content": f"{persona_text(self.assistant_name)}\n\n{user_block(self.user_name, memory_prompt())}{note}"}
+        return {"role": "system", "content": f"{persona_text(self.assistant_name)}\n\n{user_block(self.user_name, memory_prompt())}{summary_block(self.history)}{note}"}
 
     def _recent(self) -> list:
         msgs = self.history.messages
@@ -210,6 +210,7 @@ class OpenAICompatEngine:
             full = full.strip() or "(nessuna risposta)"
             self.history.save()
             emit({"type": "done", "text": full, "sources": sources})
+            compact_history(self.history, self.summarize)
         except Aborted:
             self._rollback()
             emit({"type": "error", "message": "Risposta interrotta.", "aborted": True})
@@ -217,6 +218,13 @@ class OpenAICompatEngine:
             self._rollback()
             emit({"type": "error", "message": friendly_error(err, self.base_url)})
 
+    def summarize(self, prompt: str) -> str:
+        r = self.client.chat.completions.create(model=self.model, max_tokens=400, temperature=0.3, messages=[
+            {"role": "system", "content": "Sei un assistente che riassume conversazioni in italiano, in modo fedele e compatto."},
+            {"role": "user", "content": prompt}])
+        flt = ThinkFilter()
+        return (flt.push(r.choices[0].message.content or "") + flt.carry).strip()
+
     def _rollback(self) -> None:
         msgs = self.history.messages
         if msgs and msgs[-1].get("role") == "user":

avatar/mcp_client.py

@@ -0,0 +1,318 @@
+"""Client MCP: collega server esterni (stdio o HTTP) e ne espone gli strumenti ai motori dell'app.
+
+Configurazione in config/mcp.json, stesso formato di Claude Desktop:
+{"mcpServers": {"nome": {"command": "npx", "args": ["-y", "@modelcontextprotocol/server-filesystem", "/cartella"], "env": {}},
+                "remoto": {"url": "https://esempio.it/mcp", "headers": {"Authorization": "Bearer …"}}}}
+Gli strumenti compaiono come <nome>_<strumento>; con il motore Claude Code i server vengono passati direttamente a `claude`.
+Opzioni per server: "tools": ["click", "browser_*"] (solo questi), "exclude": [...], "description_limit": 700 (caratteri
+di descrizione per strumento: server come Cua Driver ne espongono decine con descrizioni lunghissime), "disabled": true.
+"""
+from __future__ import annotations
+
+import base64
+import fnmatch
+import json
+import os
+import re
+import subprocess
+import threading
+import time
+import traceback
+import urllib.request
+from pathlib import Path
+from typing import Any
+
+from avatar.settings import CONFIG_DIR, DATA_DIR
+
+CONFIG_FILE = CONFIG_DIR / "mcp.json"
+PROTOCOL = "2025-06-18"
+CALL_TIMEOUT = 120
+EXTRA_PATH = ("/opt/homebrew/bin", "/opt/homebrew/opt/node@22/bin", "/usr/local/bin", str(Path.home() / ".local/bin"))
+
+
+def load_config() -> dict[str, dict]:
+    try:
+        d = json.loads(CONFIG_FILE.read_text(encoding="utf-8"))
+        return {k: v for k, v in (d.get("mcpServers") or {}).items() if isinstance(v, dict)}
+    except FileNotFoundError:
+        return {}
+    except Exception as err:
+        raise RuntimeError(f"config/mcp.json non valido: {err}")
+
+
+def save_config(text: str) -> dict[str, dict]:
+    d = json.loads(text)
+    if not isinstance(d, dict) or not isinstance(d.get("mcpServers", {}), dict):
+        raise ValueError("serve un oggetto con la chiave mcpServers")
+    CONFIG_DIR.mkdir(parents=True, exist_ok=True)
+    CONFIG_FILE.write_text(json.dumps(d, indent=2, ensure_ascii=False), encoding="utf-8")
+    return load_config()
+
+
+def _env_with_path(extra: dict | None) -> dict:
+    env = dict(os.environ)
+    env["PATH"] = ":".join([*EXTRA_PATH, env.get("PATH", "")])
+    env.pop("CLAUDECODE", None)
+    for k, v in (extra or {}).items():
+        env[str(k)] = str(v)
+    return env
+
+
+class MCPServer:
+    def __init__(self, name: str, cfg: dict) -> None:
+        self.name, self.cfg = name, cfg
+        self.tools: list[dict] = []
+        self.error = ""
+        self._id = 0
+        self._lock = threading.Lock()
+        self._pending: dict[int, dict] = {}
+        self._proc: subprocess.Popen | None = None
+        self._session = ""
+        self.stdio = bool(cfg.get("command"))
+
+    # ── trasporto ──────────────────────────────────────────────────────────
+    def start(self) -> None:
+        if self.stdio:
+            cmd = [str(self.cfg["command"]), *[str(a) for a in self.cfg.get("args", [])]]
+            self._proc = subprocess.Popen(cmd, stdin=subprocess.PIPE, stdout=subprocess.PIPE, stderr=subprocess.PIPE,
+                                          text=True, bufsize=1, env=_env_with_path(self.cfg.get("env")), cwd=self.cfg.get("cwd") or None)
+            threading.Thread(target=self._reader, daemon=True).start()
+            threading.Thread(target=self._drain_stderr, daemon=True).start()
+        res = self._request("initialize", {"protocolVersion": PROTOCOL, "capabilities": {}, "clientInfo": {"name": "LuZa", "version": "1.0"}}, timeout=60)
+        self._notify("notifications/initialized")
+        self.server_info = res.get("serverInfo", {})
+        self.tools = list(self._request("tools/list", {}, timeout=60).get("tools", []))
+
+    def stop(self) -> None:
+        if self._proc:
+            try:
+                self._proc.terminate()
+                self._proc.wait(timeout=5)
+            except Exception:
+                try:
+                    self._proc.kill()
+                except Exception:
+                    pass
+            self._proc = None
+
+    def _reader(self) -> None:
+        proc = self._proc
+        try:
+            for line in proc.stdout:
+                line = line.strip()
+                if not line:
+                    continue
+                try:
+                    msg = json.loads(line)
+                except Exception:
+                    continue
+                mid = msg.get("id")
+                if mid in self._pending:
+                    self._pending[mid]["msg"] = msg
+                    self._pending[mid]["ev"].set()
+        except Exception:
+            pass
+        for p in list(self._pending.values()):
+            p["msg"] = {"error": {"message": "server MCP terminato"}}
+            p["ev"].set()
+
+    def _drain_stderr(self) -> None:
+        try:
+            log = (DATA_DIR / "mcp").joinpath(f"{self.name}.log")
+            log.parent.mkdir(parents=True, exist_ok=True)
+            with open(log, "a", encoding="utf-8") as f:
+                for line in self._proc.stderr:
+                    f.write(line)
+        except Exception:
+            pass
+
+    def _next_id(self) -> int:
+        with self._lock:
+            self._id += 1
+            return self._id
+
+    def _notify(self, method: str, params: dict | None = None) -> None:
+        msg = {"jsonrpc": "2.0", "method": method, "params": params or {}}
+        if self.stdio:
+            self._proc.stdin.write(json.dumps(msg) + "\n")
+            self._proc.stdin.flush()
+        else:
+            try:
+                self._http(msg, timeout=15)
+            except Exception:
+                pass
+
+    def _request(self, method: str, params: dict, timeout: float = CALL_TIMEOUT) -> dict:
+        mid = self._next_id()
+        msg = {"jsonrpc": "2.0", "id": mid, "method": method, "params": params}
+        if self.stdio:
+            if not self._proc or self._proc.poll() is not None:
+                raise RuntimeError(f"server MCP '{self.name}' non in esecuzione")
+            slot = {"ev": threading.Event(), "msg": None}
+            self._pending[mid] = slot
+            self._proc.stdin.write(json.dumps(msg) + "\n")
+            self._proc.stdin.flush()
+            if not slot["ev"].wait(timeout):
+                self._pending.pop(mid, None)
+                raise RuntimeError(f"server MCP '{self.name}': nessuna risposta entro {int(timeout)} s")
+            self._pending.pop(mid, None)
+            resp = slot["msg"]
+        else:
+            resp = self._http(msg, timeout=timeout, want_id=mid)
+        if resp is None:
+            raise RuntimeError("risposta vuota")
+        if "error" in resp:
+            e = resp["error"]
+            raise RuntimeError(e.get("message") if isinstance(e, dict) else str(e))
+        return resp.get("result") or {}
+
+    def _http(self, msg: dict, timeout: float, want_id: int | None = None) -> dict | None:
+        headers = {"Content-Type": "application/json", "Accept": "application/json, text/event-stream", **(self.cfg.get("headers") or {})}
+        if self._session:
+            headers["Mcp-Session-Id"] = self._session
+        req = urllib.request.Request(str(self.cfg["url"]), data=json.dumps(msg).encode(), headers=headers, method="POST")
+        with urllib.request.urlopen(req, timeout=timeout) as r:
+            sid = r.headers.get("Mcp-Session-Id")
+            if sid:
+                self._session = sid
+            ctype = r.headers.get("Content-Type", "")
+            body = r.read().decode("utf-8", errors="replace")
+        if want_id is None:
+            return None
+        if "text/event-stream" in ctype:
+            for line in body.splitlines():
+                if line.startswith("data:"):
+                    try:
+                        m = json.loads(line[5:].strip())
+                    except Exception:
+                        continue
+                    if m.get("id") == want_id:
+                        return m
+            raise RuntimeError("nessuna risposta nello stream SSE")
+        return json.loads(body) if body.strip() else None
+
+    # ── strumenti ──────────────────────────────────────────────────────────
+    def call(self, tool: str, args: dict) -> str:
+        res = self._request("tools/call", {"name": tool, "arguments": args or {}})
+        parts = []
+        for c in res.get("content") or []:
+            t = c.get("type")
+            if t == "text":
+                parts.append(str(c.get("text", "")))
+            elif t == "image" and c.get("data"):
+                cap = DATA_DIR / "captures"
+                cap.mkdir(parents=True, exist_ok=True)
+                ext = {"image/png": "png", "image/jpeg": "jpg", "image/webp": "webp"}.get(c.get("mimeType", ""), "png")
+                p = cap / f"mcp-{self.name}-{int(time.time())}.{ext}"
+                p.write_bytes(base64.b64decode(c["data"]))
+                parts.append(f"IMMAGINE: {p}")
+            elif t == "resource":
+                r = c.get("resource") or {}
+                parts.append(str(r.get("text") or r.get("uri") or ""))
+        out = "\n".join(p for p in parts if p).strip()
+        if res.get("isError"):
+            return "Errore dello strumento: " + (out or "senza dettagli")
+        if "structuredContent" in res and not out:
+            out = json.dumps(res["structuredContent"], ensure_ascii=False)
+        return out[:20000] or "Fatto."
+
+
+def selected_tools(cfg: dict, tools: list[dict]) -> list[dict]:
+    """Applica i filtri tools/exclude della configurazione all'elenco degli strumenti di un server."""
+    inc, exc = cfg.get("tools") or [], cfg.get("exclude") or []
+    out = []
+    for t in tools:
+        n = t["name"]
+        if inc and not any(fnmatch.fnmatchcase(n, pat) for pat in inc):
+            continue
+        if exc and any(fnmatch.fnmatchcase(n, pat) for pat in exc):
+            continue
+        out.append(t)
+    return out
+
+
+def _safe(name: str) -> str:
+    return re.sub(r"[^a-zA-Z0-9_]", "_", name)[:64].strip("_") or "strumento"
+
+
+class MCPManager:
+    """Avvia i server configurati e registra i loro strumenti nel registro dei plugin."""
+
+    def __init__(self) -> None:
+        self.servers: dict[str, MCPServer] = {}
+        self.errors: dict[str, str] = {}
+        self.registry = None
+        self.log = print
+        self._lock = threading.Lock()
+
+    def start(self, registry, log=None) -> None:
+        self.registry = registry
+        if log:
+            self.log = log
+        threading.Thread(target=self.reload, daemon=True).start()
+
+    def reload(self) -> None:
+        with self._lock:
+            self._stop_all()
+            try:
+                cfg = load_config()
+            except Exception as err:
+                self.errors["config"] = str(err)
+                self.log(f"ERR: MCP — {err}")
+                return
+            self.errors.pop("config", None)
+            for name, c in cfg.items():
+                if c.get("disabled"):
+                    continue
+                srv = MCPServer(name, c)
+                try:
+                    srv.start()
+                except Exception as err:
+                    srv.stop()
+                    self.errors[name] = str(err)
+                    self.log(f"ERR: MCP «{name}» — {err}")
+                    traceback.print_exc()
+                    continue
+                self.errors.pop(name, None)
+                self.servers[name] = srv
+                self._register(srv)
+                self.log(f"SYS: MCP «{name}»: {len(srv.tools)} strumenti.")
+
+    def _register(self, srv: MCPServer) -> None:
+        if self.registry is None:
+            return
+        tools = []
+        limit = int(srv.cfg.get("description_limit") or 700)
+        for t in selected_tools(srv.cfg, srv.tools):
+            tname = _safe(f"{srv.name}_{t['name']}")
+            desc = " ".join((t.get("description") or t["name"]).split())
+            if len(desc) > limit:
+                desc = desc[:limit].rsplit(" ", 1)[0] + " …"
+            tools.append({"name": tname, "description": f"[{srv.name}] {desc}",
+                          "parameters": t.get("inputSchema") or {"type": "object", "properties": {}},
+                          "run": (lambda params, ctx, _s=srv, _n=t["name"]: _s.call(_n, params))})
+        info = getattr(srv, "server_info", {}) or {}
+        self.registry.add_external(f"mcp:{srv.name}", f"Server MCP {info.get('name') or srv.name}: {len(tools)} strumenti", tools)
+
+    def _stop_all(self) -> None:
+        for name, srv in list(self.servers.items()):
+            if self.registry is not None:
+                self.registry.remove_module(f"mcp:{name}")
+            srv.stop()
+        self.servers.clear()
+
+    def stop(self) -> None:
+        with self._lock:
+            self._stop_all()
+
+    def status(self) -> str:
+        lines = []
+        for name, srv in self.servers.items():
+            sel = selected_tools(srv.cfg, srv.tools)
+            lines.append(f"• {name}: {len(sel)} strumenti attivi su {len(srv.tools)} — " + ", ".join(t["name"] for t in sel[:8]) + (" …" if len(sel) > 8 else ""))
+        for name, err in self.errors.items():
+            lines.append(f"✗ {name}: {err}")
+        return "\n".join(lines) or "Nessun server MCP configurato."
+
+
+manager = MCPManager()

avatar/persona.md

@@ -42,3 +42,15 @@ Sei **Ava**, l'assistente personale di chi ti parla. Vivi in un'app sul suo Mac
 
 - Con gli strumenti "casa_…" controlli la casa tramite Home Assistant: luci, prese, clima, tapparelle, media, scene e sensori. Se l'utente nomina un dispositivo che non trovi, cerca con casa_dispositivi prima di dire che non esiste; per comandi su più stanze usa casa_chiedi.
 - Conferma a voce cosa hai fatto in poche parole ("Luce cucina accesa"). Serrature e allarme solo dopo la conferma esplicita dell'utente.
+
+## Computer
+
+- Se sono disponibili gli strumenti "cua_driver_…" puoi ispezionare e comandare le app del Mac (aprire finestre, leggere e premere elementi, scrivere nei campi) senza togliere il focus all'utente. Usali quando l'utente chiede un'azione dentro un'app e non esiste un plugin dedicato: per mail, calendario, promemoria, note, musica, file e casa preferisci sempre i plugin.
+- Prima di agire osserva: leggi la finestra o l'elemento, poi esegui un passo alla volta e verifica il risultato. Niente sequenze lunghe alla cieca.
+- Chiedi conferma prima di azioni difficili da annullare: inviare messaggi o mail, pagare, cancellare, chiudere documenti non salvati, modificare impostazioni. Se un passaggio non riesce due volte, fermati e spiega cosa vedi invece di insistere.
+- Non inserire mai password o codici, e non agire in finestre di banche o pagamenti se non su richiesta esplicita e puntuale dell'utente.
+- Metodo del driver: prima list_apps o list_windows per trovare pid e finestra, poi get_window_state per leggere gli elementi, poi agisci con click, type_text o set_value indicando l'elemento (element_token o element_index + window_id) invece delle coordinate, infine rileggi lo stato o usa verify_state per controllare. Lancia le app con launch_app (in background), e porta in primo piano con bring_to_front solo se l'utente lo chiede.
+
+## Immagini
+
+- Puoi creare immagini con immagine_genera: traduci la richiesta in una descrizione inglese ricca (soggetto, scena, luce, stile, inquadratura) e avvisa che ci vuole qualche decina di secondi. Quando è pronta, dilla in poche parole senza descrivere i dettagli tecnici.

avatar/plugins.py

@@ -91,6 +91,24 @@ class PluginRegistry:
             self.modules[mod_name] = info
         return self
 
+    def add_external(self, module: str, description: str, tools: list[dict]) -> None:
+        """Registra strumenti forniti a runtime (server MCP), sostituendo quelli dello stesso modulo."""
+        self.load()
+        self.remove_module(module)
+        info = {"name": module, "description": description, "valid": True, "error": "", "tools": []}
+        for t in tools:
+            name = str(t["name"])
+            if name in self.tools:   # nome già usato da un plugin: prefissa
+                name = f"x_{name}"[:64]
+            self.tools[name] = Tool(module, name, str(t.get("description", "")), _normalize_schema(t.get("parameters", {})), t["run"])
+            info["tools"].append(name)
+        self.modules[module] = info
+
+    def remove_module(self, module: str) -> None:
+        for name in [n for n, t in self.tools.items() if t.module == module]:
+            self.tools.pop(name, None)
+        self.modules.pop(module, None)
+
     def _enabled(self, module: str) -> bool:
         try:
             from memory.config_manager import get_plugin_enabled

avatar/settings.py

@@ -52,6 +52,8 @@ DEFAULTS: dict[str, Any] = {
     "server_assist_dir": "/opt/aiserverassistance",   # cartella del bot sul server ponte (contiene .env)
     "server_assist_bot": "luzaserver_bot",            # username Telegram del bot sysadmin
     "homeassistant_url": "http://homeassistant.local:8123",   # Home Assistant (token nel portachiavi)
+    "immagini_famiglia": "z-image-turbo",   # z-image-turbo | schnell | dev | qwen (comando mflux e passi predefiniti)
+    "immagini_modello": "",                 # repo Hugging Face o cartella; vuoto = predefinito della famiglia
 }
 
 SECRET_KEYS = ("anthropic_api_key", "local_api_key", "telegram_api_hash", "elevenlabs_api_key", "homeassistant_token")

avatar/settings_dialog.py

@@ -1,6 +1,7 @@
 """Finestra impostazioni: motore, chiavi, voce e riconoscimento vocale."""
 from __future__ import annotations
 
+import json
 import sys
 import threading
 from pathlib import Path
@@ -65,7 +66,7 @@ class SettingsDialog(QDialog):
             self.tabs.addTab(sc, title)
             lay.addStretch(1)
             return lay
-        pg_motore, pg_voce, pg_msg, pg_mon, pg_casa = page("Motore"), page("Voce e avatar"), page("Messaggistica"), page("Monitor"), page("Casa")
+        pg_motore, pg_voce, pg_msg, pg_mon, pg_casa, pg_tool = page("Motore"), page("Voce e avatar"), page("Messaggistica"), page("Monitor"), page("Casa"), page("Strumenti")
         def add(lay, item):   # inserisce prima dello stretch finale
             if isinstance(item, QWidget): lay.insertWidget(lay.count() - 1, item)
             else: lay.insertLayout(lay.count() - 1, item)
@@ -260,6 +261,36 @@ class SettingsDialog(QDialog):
         self.ha_hint = QLabel(""); self.ha_hint.setWordWrap(True); form_ha.addRow("", self.ha_hint)
         add(pg_casa, form_ha)
 
+        # ── Strumenti (server MCP esterni) ────────────────────────────────
+        from PyQt6.QtWidgets import QPlainTextEdit
+        from .mcp_client import CONFIG_FILE as MCP_FILE, manager as mcp_manager
+        form_mcp = QFormLayout()
+        form_mcp.addRow(QLabel("Server MCP esterni: stesso formato di Claude Desktop (mcpServers con command/args/env, oppure url/headers). "
+                               "Gli strumenti diventano disponibili a tutti i motori; con Claude Code i server vengono passati direttamente a claude."))
+        self.mcp_text = QPlainTextEdit()
+        try:
+            self.mcp_text.setPlainText(MCP_FILE.read_text(encoding="utf-8"))
+        except Exception:
+            self.mcp_text.setPlainText('{\n  "mcpServers": {\n  }\n}\n')
+        self.mcp_text.setMinimumHeight(220)
+        form_mcp.addRow(self.mcp_text)
+        row = QHBoxLayout()
+        b = QPushButton("Applica e ricollega"); b.clicked.connect(self._mcp_apply); row.addWidget(b)
+        b = QPushButton("Esempio"); b.clicked.connect(self._mcp_example); row.addWidget(b)
+        row.addStretch()
+        form_mcp.addRow(row)
+        self.mcp_hint = QLabel(mcp_manager.status()); self.mcp_hint.setWordWrap(True); form_mcp.addRow(self.mcp_hint)
+        add(pg_tool, form_mcp)
+
+        # ── Immagini (mflux) ──────────────────────────────────────────────
+        form_img = QFormLayout()
+        form_img.addRow(QLabel("Generazione immagini in locale (mflux, MLX). Il primo uso scarica il modello da Hugging Face."))
+        self.img_fam = _combo([("z-image-turbo", "Z-Image Turbo (veloce, 8 passi, consigliato)"), ("schnell", "FLUX.1 schnell (4 passi)"), ("dev", "FLUX.1 dev (lento, licenza non commerciale)"), ("qwen", "Qwen-Image (pesante, ottimo col testo)")], s.get("immagini_famiglia") or "z-image-turbo")
+        form_img.addRow("Famiglia", self.img_fam)
+        self.img_model = QLineEdit(str(s.get("immagini_modello") or "")); self.img_model.setPlaceholderText("vuoto = predefinito (es. mflux-community/z-image-turbo-mflux-q4)")
+        form_img.addRow("Modello (repo Hugging Face)", self.img_model)
+        add(pg_tool, form_img)
+
         btns = QHBoxLayout(); btns.addStretch()
         cancel = QPushButton("Annulla"); cancel.clicked.connect(self.reject); btns.addWidget(cancel)
         save = QPushButton("Salva"); save.setObjectName("primary"); save.clicked.connect(self._save); btns.addWidget(save)
@@ -407,6 +438,25 @@ class SettingsDialog(QDialog):
                 self._async.emit("el_error", str(err)[:120])
         threading.Thread(target=work, daemon=True).start()
 
+    def _mcp_example(self) -> None:
+        self.mcp_text.setPlainText(json.dumps({"mcpServers": {
+            "filesystem": {"command": "npx", "args": ["-y", "@modelcontextprotocol/server-filesystem", str(Path.home() / "Documents")]},
+            "remoto": {"url": "https://esempio.it/mcp", "headers": {"Authorization": "Bearer TOKEN"}, "disabled": True},
+        }}, indent=2, ensure_ascii=False))
+
+    def _mcp_apply(self) -> None:
+        from .mcp_client import save_config, manager as mcp_manager
+        try:
+            save_config(self.mcp_text.toPlainText())
+        except Exception as err:
+            self.mcp_hint.setText(f"JSON non valido: {err}"); return
+        self.mcp_hint.setText("Collego i server…")
+
+        def go():
+            mcp_manager.reload()
+            self._async.emit("mcp", mcp_manager.status())
+        threading.Thread(target=go, daemon=True).start()
+
     def _ha_check(self) -> None:
         vals = {"homeassistant_url": self.ha_url.text().strip().rstrip("/")}
         if self.ha_token.text().strip():
@@ -427,6 +477,8 @@ class SettingsDialog(QDialog):
     def _on_async(self, kind: str, payload) -> None:
         if kind == "ha":
             self.ha_hint.setText(str(payload)); return
+        if kind == "mcp":
+            self.mcp_hint.setText(str(payload)); return
         if kind == "vb":
             self.vb_profile.clear()
             for pid, label in payload:
@@ -517,6 +569,7 @@ class SettingsDialog(QDialog):
             "server_assist_dir": self.srv_dir.text().strip() or "/opt/aiserverassistance",
             "server_assist_bot": self.srv_bot.text().strip().lstrip("@") or "luzaserver_bot",
             "homeassistant_url": self.ha_url.text().strip().rstrip("/"),
+            "immagini_famiglia": self.img_fam.currentData() or "z-image-turbo", "immagini_modello": self.img_model.text().strip(),
         }
         if self.ha_token.text().strip():
             values["homeassistant_token"] = self.ha_token.text().strip()

config/mcp.example.json

@@ -0,0 +1,52 @@
+{
+  "mcpServers": {
+    "filesystem": {
+      "command": "npx",
+      "args": [
+        "-y",
+        "@modelcontextprotocol/server-filesystem",
+        "/Users/tuonome/Documents"
+      ]
+    },
+    "remoto": {
+      "url": "https://esempio.it/mcp",
+      "headers": {
+        "Authorization": "Bearer TOKEN"
+      },
+      "disabled": true
+    },
+    "cua-driver": {
+      "command": "/Users/tuonome/.local/bin/cua-driver",
+      "args": [
+        "mcp"
+      ],
+      "tools": [
+        "list_apps",
+        "list_windows",
+        "get_window_state",
+        "verify_state",
+        "launch_app",
+        "bring_to_front",
+        "invoke_menu",
+        "click",
+        "double_click",
+        "right_click",
+        "type_text",
+        "press_key",
+        "hotkey",
+        "set_value",
+        "scroll",
+        "get_desktop_state",
+        "zoom",
+        "clipboard_read",
+        "clipboard_write",
+        "get_browser_state",
+        "browser_navigate",
+        "browser_click",
+        "browser_type"
+      ],
+      "description_limit": 800,
+      "disabled": true
+    }
+  }
+}

main.py

@@ -62,6 +62,9 @@ def main() -> None:
     confirm.bind(ui.show_confirm, ui.hide_confirm, ui.write_log)
     registry.ctx = {"say": assistant.say, "log": ui.write_log, "confirm": confirm.request, "player": ui}
     ui.get_plugins = registry.list_for_ui
+    from avatar.mcp_client import manager as mcp_manager
+    mcp_manager.start(registry, ui.write_log)
+    ui._app.aboutToQuit.connect(mcp_manager.stop)
     ui.get_plugin_settings = lambda: []
     ui.request_say = assistant.say
     # Questi due sono attributi della finestra, non della facciata JarvisUI.

memory/memory_manager.py

@@ -1,479 +1,513 @@
-import json
-import re
-from datetime import datetime
-from threading import Lock
-from pathlib import Path
-import sys
-
-
-def get_base_dir() -> Path:
-    if getattr(sys, "frozen", False):
-        return Path(sys.executable).parent
-    return Path(__file__).resolve().parent.parent
-
-
-BASE_DIR         = get_base_dir()
-MEMORY_PATH      = BASE_DIR / "memory" / "long_term.json"
-_lock            = Lock()
-MAX_VALUE_LENGTH = 380
-
-# ── Why there are two very different numbers here ────────────────────────────
-#
-# There used to be one: MEMORY_MAX_CHARS = 2200, applied to the whole store. It
-# was a *storage* limit, and it existed only because the entire memory was
-# pasted into the system prompt on every connect — so growing the memory grew
-# every single request. When it filled, _trim_to_limit() deleted the oldest
-# entries and printed one line to a console nobody reads. A memory described as
-# "deeply remembers projects, preferences and personal context" was in practice
-# two pages long, and quietly forgot your sister's name after a few weeks.
-#
-# Storage and prompt budget are now separate concerns:
-#
-#   MEMORY_MAX_CHARS  — a runaway guard, not a feature limit. Nothing normal
-#                       reaches it; a bug writing in a loop does.
-#   PROMPT_CORE_CHARS — what actually rides in the system prompt every session.
-#                       Smaller than the old whole-memory dump, so sessions
-#                       start *faster* than before, not slower.
-#
-# Everything above the core stays on disk and is fetched on demand by the
-# recall_memory tool — see search_memory() and format_memory_for_prompt().
-MEMORY_MAX_CHARS  = 200_000
-PROMPT_CORE_CHARS = 900
-PROMPT_INDEX_CHARS = 420
-# Most entries any one category may contribute to the core block, so a person
-# with forty stored preferences still gets their sister into the prompt.
-PROMPT_MAX_PER_CATEGORY = 6
-
-def _empty_memory() -> dict:
-    return {
-        "identity":      {},
-        "preferences":   {},
-        "projects":      {},
-        "relationships": {},
-        "wishes":        {},
-        "notes":         {},
-    }
-
-def load_memory() -> dict:
-    if not MEMORY_PATH.exists():
-        return _empty_memory()
-    with _lock:
-        try:
-            data = json.loads(MEMORY_PATH.read_text(encoding="utf-8"))
-            if isinstance(data, dict):
-                base = _empty_memory()
-                for key in base:
-                    if key not in data:
-                        data[key] = {}
-                return data
-            return _empty_memory()
-        except Exception as e:
-            print(f"[Memory] ⚠️ Load error: {e}")
-            return _empty_memory()
-
-def _all_entries(memory: dict) -> list[tuple]:
-    entries = []
-    for cat, items in memory.items():
-        if not isinstance(items, dict):
-            continue
-        for key, entry in items.items():
-            if isinstance(entry, dict) and "value" in entry:
-                entries.append((cat, key, entry))
-    return entries
-
-
-# Set by main.py so a trim can reach the activity log. Deleting something a
-# person told you and mentioning it only on stdout is how a memory loses trust.
-_trim_notifier = None
-
-
-def set_trim_notifier(fn) -> None:
-    """Register a callable(str) that surfaces trims to the user."""
-    global _trim_notifier
-    _trim_notifier = fn
-
-
-def _trim_to_limit(memory: dict) -> dict:
-    if len(json.dumps(memory, ensure_ascii=False)) <= MEMORY_MAX_CHARS:
-        return memory
-    entries = _all_entries(memory)
-    entries.sort(key=lambda t: t[2].get("updated", "0000-00-00"))
-    dropped = []
-    for cat, key, _ in entries:
-        if len(json.dumps(memory, ensure_ascii=False)) <= MEMORY_MAX_CHARS:
-            break
-        del memory[cat][key]
-        dropped.append(f"{cat}/{key}")
-        print(f"[Memory] 🗑️  Trimmed {cat}/{key}")
-    if dropped and _trim_notifier:
-        try:
-            _trim_notifier(
-                f"SYS: Memory full — forgot {len(dropped)} oldest entries "
-                f"({', '.join(dropped[:3])}{'…' if len(dropped) > 3 else ''})"
-            )
-        except Exception:
-            pass
-    return memory
-
-def save_memory(memory: dict) -> None:
-    if not isinstance(memory, dict):
-        return
-    memory = _trim_to_limit(memory)
-    MEMORY_PATH.parent.mkdir(parents=True, exist_ok=True)
-    with _lock:
-        MEMORY_PATH.write_text(
-            json.dumps(memory, indent=2, ensure_ascii=False),
-            encoding="utf-8",
-        )
-
-
-def _truncate_value(val: str) -> str:
-    if isinstance(val, str) and len(val) > MAX_VALUE_LENGTH:
-        return val[:MAX_VALUE_LENGTH].rstrip() + "…"
-    return val
-
-
-def _recursive_update(target: dict, updates: dict) -> bool:
-    changed = False
-    for key, value in updates.items():
-        if value is None:
-            continue
-        if isinstance(value, str) and not value.strip():
-            continue
-        if isinstance(value, dict) and "value" not in value:
-            if key not in target or not isinstance(target[key], dict):
-                target[key] = {}
-                changed = True
-            if _recursive_update(target[key], value):
-                changed = True
-        else:
-            new_val  = _truncate_value(str(value["value"] if isinstance(value, dict) else value))
-            entry    = {"value": new_val, "updated": datetime.now().strftime("%Y-%m-%d")}
-            existing = target.get(key, {})
-            if not isinstance(existing, dict) or existing.get("value") != new_val:
-                target[key] = entry
-                changed = True
-    return changed
-
-
-def update_memory(memory_update: dict) -> dict:
-    if not isinstance(memory_update, dict) or not memory_update:
-        return load_memory()
-    memory = load_memory()
-    if _recursive_update(memory, memory_update):
-        save_memory(memory)
-        print(f"[Memory] 💾 Saved: {list(memory_update.keys())}")
-    return memory
-
-def _entry_value(entry) -> str:
-    """Accept both the {'value': ..., 'updated': ...} shape and a bare string,
-    because early versions of the store wrote plain strings."""
-    if isinstance(entry, dict):
-        return str(entry.get("value", "") or "").strip()
-    return str(entry or "").strip()
-
-
-def _pretty(key: str) -> str:
-    return key.replace("_", " ").strip()
-
-
-# Identity is always in the prompt; these categories compete for the remaining
-# budget by recency.
-_CATEGORY_LABELS = {
-    "preferences":   "Preferences",
-    "projects":      "Active projects / goals",
-    "relationships": "People in their life",
-    "wishes":        "Wishes / plans",
-    "notes":         "Notes",
-}
-
-_IDENTITY_FIELDS = ["name", "age", "birthday", "city", "job",
-                    "language", "school", "nationality"]
-
-
-def format_memory_for_prompt(memory: dict | None) -> str:
-    """Build the memory block that goes into the system prompt.
-
-    This used to dump everything. It now sends three things:
-
-      1. IDENTITY  - always, in full. It is small, and it is wrong for the
-         assistant to have to look up your name.
-      2. RECENT    - the most recently updated entries from every other
-         category, up to PROMPT_CORE_CHARS. Recency is the cheapest useful
-         relevance signal available without embeddings.
-      3. AN INDEX  - the *keys* of everything else, values omitted.
-
-    Point 3 is what makes recall work at all. A model cannot decide to look
-    something up if it does not know the thing exists: with only points 1 and 2,
-    "who is Ayse?" would get "I don't know" while ayse_sister sat on disk
-    unread. The index costs a few hundred characters and turns recall from a
-    gamble into a lookup.
-
-    Net effect on latency: this block is SMALLER than the old full dump, so
-    every session connects with fewer tokens. Occasionally the model spends one
-    extra round trip on recall_memory - covered by the acknowledgment it
-    already speaks before any slow step."""
-    if not memory:
-        return ""
-
-    core_lines: list[str] = []
-
-    # 1. Identity - always, in full
-    identity = memory.get("identity", {}) or {}
-    for field in _IDENTITY_FIELDS:
-        val = _entry_value(identity.get(field))
-        if not val:
-            continue
-        if field == "language":
-            # Labelled as an observation, not a setting. A bare "Language:
-            # English" line written months ago reads like a standing order and
-            # was one of the reasons a Turkish question came back in English.
-            core_lines.append(
-                f"Has spoken to you in: {val} (an observation about the past — "
-                f"always answer in the language of their CURRENT message)")
-        else:
-            core_lines.append(f"{field.title()}: {val}")
-    for key, entry in identity.items():
-        if key in _IDENTITY_FIELDS:
-            continue
-        val = _entry_value(entry)
-        if val:
-            core_lines.append(f"{_pretty(key).title()}: {val}")
-
-    # 2. Everything else, most recently updated first
-    rest: list[tuple[str, str, str, str]] = []   # (updated, cat, key, value)
-    for cat in _CATEGORY_LABELS:
-        for key, entry in (memory.get(cat, {}) or {}).items():
-            val = _entry_value(entry)
-            if not val:
-                continue
-            updated = (entry.get("updated", "") if isinstance(entry, dict) else "") or "0000-00-00"
-            rest.append((updated, cat, key, val))
-    rest.sort(key=lambda t: t[0], reverse=True)
-
-    used    = sum(len(l) + 1 for l in core_lines)
-    shown: dict[str, list[str]] = {}
-    overflow: dict[str, list[str]] = {}
-
-    # Recency decides order, but no single category may take the whole budget.
-    # Without the cap, someone with forty stored preferences gets a prompt that
-    # is forty preferences and not one person's name — the categories that
-    # matter most in conversation are also the ones that change least often, so
-    # pure recency systematically buries them.
-    per_cat_used: dict[str, int] = {}
-    for _updated, cat, key, val in rest:
-        line = f"  - {_pretty(key).title()}: {val}"
-        if (per_cat_used.get(cat, 0) < PROMPT_MAX_PER_CATEGORY
-                and used + len(line) + 1 <= PROMPT_CORE_CHARS):
-            shown.setdefault(cat, []).append(line)
-            per_cat_used[cat] = per_cat_used.get(cat, 0) + 1
-            used += len(line) + 1
-        else:
-            overflow.setdefault(cat, []).append(_pretty(key))
-
-    # The index is a table of contents, so it is interleaved across categories
-    # rather than continuing in recency order. Sorted by recency it would list
-    # twenty-four preferences before the first relationship, and the one entry
-    # the index exists for — the old fact the model has no other way to know
-    # about — would fall off the end.
-    indexed: list[str] = []
-    if overflow:
-        cats  = [c for c in _CATEGORY_LABELS if overflow.get(c)]
-        cursor = {c: 0 for c in cats}
-        while cats:
-            for cat in list(cats):
-                i = cursor[cat]
-                if i >= len(overflow[cat]):
-                    cats.remove(cat)
-                    continue
-                indexed.append(overflow[cat][i])
-                cursor[cat] = i + 1
-
-    for cat, label in _CATEGORY_LABELS.items():
-        if shown.get(cat):
-            core_lines.append("")
-            core_lines.append(f"{label}:")
-            core_lines.extend(shown[cat])
-
-    if not core_lines and not indexed:
-        return ""
-
-    out = [
-        "[WHAT YOU KNOW ABOUT THIS PERSON — use naturally, never recite like a list]",
-        *core_lines,
-    ]
-
-    # 3. The index of what is on disk but not in this prompt
-    if indexed:
-        budget, names = PROMPT_INDEX_CHARS, []
-        for n in indexed:
-            if budget - len(n) - 2 < 0:
-                break
-            names.append(n)
-            budget -= len(n) + 2
-        if names:
-            out.append("")
-            out.append(
-                "[ALSO REMEMBERED — values not shown here. Call recall_memory "
-                "with a keyword to read any of these before saying you do not know]"
-            )
-            out.append(", ".join(names)
-                       + (f" (+{len(indexed) - len(names)} more)"
-                          if len(indexed) > len(names) else ""))
-
-    return "\n".join(out) + "\n"
-
-
-# ── Recall ────────────────────────────────────────────────────────────────────
-
-def _score(query_words: list[str], cat: str, key: str, value: str) -> int:
-    """Cheap lexical relevance. No embeddings, no network, no model call - this
-    runs in well under a millisecond, which is the entire point: recall must
-    cost one model round trip, never two."""
-    hay_key = _pretty(key).lower()
-    hay_val = value.lower()
-    score   = 0
-    for w in query_words:
-        if not w:
-            continue
-        if w == hay_key:
-            score += 10
-        elif w in hay_key:
-            score += 6
-        if w in hay_val:
-            score += 3
-        if w in cat:
-            score += 1
-    return score
-
-
-def search_memory(query: str, limit: int = 8) -> str:
-    """Find stored facts matching `query`. Backs the recall_memory tool.
-
-    An empty query is treated as "show me everything you know", capped - the
-    model asks that when the user says "what do you remember about me?"."""
-    memory = load_memory()
-    words  = [w for w in re.split(r"[^\w]+", (query or "").lower()) if len(w) > 1]
-
-    rows: list[tuple[int, str, str, str]] = []
-    for cat, items in memory.items():
-        if not isinstance(items, dict):
-            continue                     # skip 'sessions', which is a list
-        for key, entry in items.items():
-            val = _entry_value(entry)
-            if not val:
-                continue
-            s = _score(words, cat, key, val) if words else 1
-            if s > 0:
-                rows.append((s, cat, key, val))
-
-    if not rows:
-        return (f"Nothing stored about '{query}'." if query
-                else "I have not stored anything about this person yet.")
-
-    rows.sort(key=lambda r: (-r[0], r[2]))
-    lines = [f"{cat}/{_pretty(key)}: {val}" for _s, cat, key, val in rows[:max(1, limit)]]
-    head  = (f"Stored facts matching '{query}':" if query
-             else "Everything currently stored:")
-    more  = (f"\n(+{len(rows) - len(lines)} more — search with a narrower keyword)"
-             if len(rows) > len(lines) else "")
-    return head + "\n" + "\n".join(lines) + more
-
-
-def all_entries_for_ui() -> list[dict]:
-    """Flat list for the memory panel: what JARVIS knows, and when it learned it.
-    Sorted newest first so the panel opens on what changed most recently."""
-    memory = load_memory()
-    rows = []
-    for cat, items in memory.items():
-        if not isinstance(items, dict):
-            continue
-        for key, entry in items.items():
-            val = _entry_value(entry)
-            if not val:
-                continue
-            rows.append({
-                "category": cat,
-                "key":      key,
-                "value":    val,
-                "updated":  (entry.get("updated", "") if isinstance(entry, dict) else ""),
-            })
-    rows.sort(key=lambda r: (r["updated"] or "0000-00-00"), reverse=True)
-    return rows
-
-def remember(key: str, value: str, category: str = "notes") -> str:
-    valid = {"identity", "preferences", "projects", "relationships", "wishes", "notes"}
-    if category not in valid:
-        category = "notes"
-    update_memory({category: {key: {"value": value}}})
-    return f"Remembered: {category}/{key} = {value}"
-
-
-def forget(key: str, category: str = "notes") -> str:
-    memory = load_memory()
-    cat    = memory.get(category, {})
-    if key in cat:
-        del cat[key]
-        memory[category] = cat
-        save_memory(memory)
-        return f"Forgotten: {category}/{key}"
-    return f"Not found: {category}/{key}"
-
-
-forget_memory = forget
-
-
-# ── Session memory ─────────────────────────────────────────────────────────────
-
-_SESSION_MAX = 3   # safety cap — in practice 0-1 entries after pop
-
-
-def save_session_summary(summary: str, language: str = "") -> None:
-    """Append a 1-2 sentence session summary to long_term.json['sessions']."""
-    summary = (summary or "").strip()
-    if not summary:
-        return
-    memory   = load_memory()
-    sessions = memory.get("sessions", [])
-    if not isinstance(sessions, list):
-        sessions = []
-    entry: dict = {
-        "date":    datetime.now().strftime("%Y-%m-%d"),
-        "summary": summary[:280],
-    }
-    if language:
-        entry["language"] = language
-    sessions.append(entry)
-    memory["sessions"] = sessions[-_SESSION_MAX:]
-    with _lock:
-        MEMORY_PATH.parent.mkdir(parents=True, exist_ok=True)
-        MEMORY_PATH.write_text(
-            json.dumps(memory, indent=2, ensure_ascii=False),
-            encoding="utf-8",
-        )
-    print(f"[Memory] 📝 Session saved ({entry['date']}): {summary[:60]}…")
-
-
-def pop_last_session() -> dict | None:
-    """
-    Return AND remove the most recent session entry.
-    Calling this consumes the entry so it is never repeated in future briefings.
-    """
-    with _lock:
-        if not MEMORY_PATH.exists():
-            return None
-        try:
-            memory   = json.loads(MEMORY_PATH.read_text(encoding="utf-8"))
-            sessions = memory.get("sessions", [])
-            if not isinstance(sessions, list) or not sessions:
-                return None
-            entry = sessions.pop()          # remove the last entry
-            memory["sessions"] = sessions
-            MEMORY_PATH.write_text(
-                json.dumps(memory, indent=2, ensure_ascii=False),
-                encoding="utf-8",
-            )
-            return entry
-        except Exception as e:
-            print(f"[Memory] ⚠️ pop_last_session error: {e}")
+import json
+import re
+from datetime import datetime
+from threading import Lock
+from pathlib import Path
+import sys
+
+
+def get_base_dir() -> Path:
+    if getattr(sys, "frozen", False):
+        return Path(sys.executable).parent
+    return Path(__file__).resolve().parent.parent
+
+
+BASE_DIR         = get_base_dir()
+MEMORY_PATH      = BASE_DIR / "memory" / "long_term.json"
+_lock            = Lock()
+MAX_VALUE_LENGTH = 380
+
+# ── Why there are two very different numbers here ────────────────────────────
+#
+# There used to be one: MEMORY_MAX_CHARS = 2200, applied to the whole store. It
+# was a *storage* limit, and it existed only because the entire memory was
+# pasted into the system prompt on every connect — so growing the memory grew
+# every single request. When it filled, _trim_to_limit() deleted the oldest
+# entries and printed one line to a console nobody reads. A memory described as
+# "deeply remembers projects, preferences and personal context" was in practice
+# two pages long, and quietly forgot your sister's name after a few weeks.
+#
+# Storage and prompt budget are now separate concerns:
+#
+#   MEMORY_MAX_CHARS  — a runaway guard, not a feature limit. Nothing normal
+#                       reaches it; a bug writing in a loop does.
+#   PROMPT_CORE_CHARS — what actually rides in the system prompt every session.
+#                       Smaller than the old whole-memory dump, so sessions
+#                       start *faster* than before, not slower.
+#
+# Everything above the core stays on disk and is fetched on demand by the
+# recall_memory tool — see search_memory() and format_memory_for_prompt().
+MEMORY_MAX_CHARS  = 200_000
+PROMPT_CORE_CHARS = 900
+PROMPT_INDEX_CHARS = 420
+# Most entries any one category may contribute to the core block, so a person
+# with forty stored preferences still gets their sister into the prompt.
+PROMPT_MAX_PER_CATEGORY = 6
+
+def _empty_memory() -> dict:
+    return {
+        "identity":      {},
+        "preferences":   {},
+        "projects":      {},
+        "relationships": {},
+        "wishes":        {},
+        "notes":         {},
+    }
+
+def load_memory() -> dict:
+    if not MEMORY_PATH.exists():
+        return _empty_memory()
+    with _lock:
+        try:
+            data = json.loads(MEMORY_PATH.read_text(encoding="utf-8"))
+            if isinstance(data, dict):
+                base = _empty_memory()
+                for key in base:
+                    if key not in data:
+                        data[key] = {}
+                return data
+            return _empty_memory()
+        except Exception as e:
+            print(f"[Memory] ⚠️ Load error: {e}")
+            return _empty_memory()
+
+def _all_entries(memory: dict) -> list[tuple]:
+    entries = []
+    for cat, items in memory.items():
+        if not isinstance(items, dict):
+            continue
+        for key, entry in items.items():
+            if isinstance(entry, dict) and "value" in entry:
+                entries.append((cat, key, entry))
+    return entries
+
+
+# Set by main.py so a trim can reach the activity log. Deleting something a
+# person told you and mentioning it only on stdout is how a memory loses trust.
+_trim_notifier = None
+
+
+def set_trim_notifier(fn) -> None:
+    """Register a callable(str) that surfaces trims to the user."""
+    global _trim_notifier
+    _trim_notifier = fn
+
+
+def _trim_to_limit(memory: dict) -> dict:
+    if len(json.dumps(memory, ensure_ascii=False)) <= MEMORY_MAX_CHARS:
+        return memory
+    entries = _all_entries(memory)
+    entries.sort(key=lambda t: t[2].get("updated", "0000-00-00"))
+    dropped = []
+    for cat, key, _ in entries:
+        if len(json.dumps(memory, ensure_ascii=False)) <= MEMORY_MAX_CHARS:
+            break
+        del memory[cat][key]
+        dropped.append(f"{cat}/{key}")
+        print(f"[Memory] 🗑️  Trimmed {cat}/{key}")
+    if dropped and _trim_notifier:
+        try:
+            _trim_notifier(
+                f"SYS: Memory full — forgot {len(dropped)} oldest entries "
+                f"({', '.join(dropped[:3])}{'…' if len(dropped) > 3 else ''})"
+            )
+        except Exception:
+            pass
+    return memory
+
+def save_memory(memory: dict) -> None:
+    if not isinstance(memory, dict):
+        return
+    memory = _trim_to_limit(memory)
+    MEMORY_PATH.parent.mkdir(parents=True, exist_ok=True)
+    with _lock:
+        MEMORY_PATH.write_text(
+            json.dumps(memory, indent=2, ensure_ascii=False),
+            encoding="utf-8",
+        )
+    try:
+        export_markdown(memory)
+    except Exception as err:
+        print(f"[Memory] esportazione Markdown fallita: {err}")
+
+
+EXPORT_DIR = get_base_dir() / "data" / "memoria"
+_EXPORT_LABELS = {"identity": "Identità", "preferences": "Preferenze", "projects": "Progetti", "relationships": "Persone",
+                  "wishes": "Desideri e obiettivi", "notes": "Note"}
+
+
+def export_markdown(memory: dict, folder: Path | None = None) -> Path:
+    """Specchia la memoria in una cartella di file Markdown (apribile come vault Obsidian): un file per categoria più un indice."""
+    folder = folder or EXPORT_DIR
+    folder.mkdir(parents=True, exist_ok=True)
+    index = ["# Memoria di LuZa", "", f"Aggiornata: {datetime.now():%Y-%m-%d %H:%M}", ""]
+    for cat, label in _EXPORT_LABELS.items():
+        entries = memory.get(cat) or {}
+        if not isinstance(entries, dict):
+            continue
+        lines = ["---", f"categoria: {cat}", f"voci: {len(entries)}", "---", "", f"# {label}", ""]
+        for key, entry in sorted(entries.items(), key=lambda kv: (kv[1].get("updated", "") if isinstance(kv[1], dict) else ""), reverse=True):
+            val = _entry_value(entry)
+            upd = entry.get("updated", "") if isinstance(entry, dict) else ""
+            lines.append(f"- **{_pretty(key)}**: {val}" + (f" _(aggiornato {upd})_" if upd else ""))
+        (folder / f"{label}.md").write_text("\n".join(lines) + "\n", encoding="utf-8")
+        index.append(f"- [[{label}]] — {len(entries)} voci")
+    sessions = memory.get("sessions") or []
+    if isinstance(sessions, list) and sessions:
+        lines = ["# Sessioni recenti", ""] + [f"- {e.get('date', '')}: {e.get('summary', '')}" for e in sessions]
+        (folder / "Sessioni recenti.md").write_text("\n".join(lines) + "\n", encoding="utf-8")
+        index.append("- [[Sessioni recenti]]")
+    (folder / "Memoria.md").write_text("\n".join(index) + "\n", encoding="utf-8")
+    return folder
+
+
+def _truncate_value(val: str) -> str:
+    if isinstance(val, str) and len(val) > MAX_VALUE_LENGTH:
+        return val[:MAX_VALUE_LENGTH].rstrip() + "…"
+    return val
+
+
+def _recursive_update(target: dict, updates: dict) -> bool:
+    changed = False
+    for key, value in updates.items():
+        if value is None:
+            continue
+        if isinstance(value, str) and not value.strip():
+            continue
+        if isinstance(value, dict) and "value" not in value:
+            if key not in target or not isinstance(target[key], dict):
+                target[key] = {}
+                changed = True
+            if _recursive_update(target[key], value):
+                changed = True
+        else:
+            new_val  = _truncate_value(str(value["value"] if isinstance(value, dict) else value))
+            entry    = {"value": new_val, "updated": datetime.now().strftime("%Y-%m-%d")}
+            existing = target.get(key, {})
+            if not isinstance(existing, dict) or existing.get("value") != new_val:
+                target[key] = entry
+                changed = True
+    return changed
+
+
+def update_memory(memory_update: dict) -> dict:
+    if not isinstance(memory_update, dict) or not memory_update:
+        return load_memory()
+    memory = load_memory()
+    if _recursive_update(memory, memory_update):
+        save_memory(memory)
+        print(f"[Memory] 💾 Saved: {list(memory_update.keys())}")
+    return memory
+
+def _entry_value(entry) -> str:
+    """Accept both the {'value': ..., 'updated': ...} shape and a bare string,
+    because early versions of the store wrote plain strings."""
+    if isinstance(entry, dict):
+        return str(entry.get("value", "") or "").strip()
+    return str(entry or "").strip()
+
+
+def _pretty(key: str) -> str:
+    return key.replace("_", " ").strip()
+
+
+# Identity is always in the prompt; these categories compete for the remaining
+# budget by recency.
+_CATEGORY_LABELS = {
+    "preferences":   "Preferences",
+    "projects":      "Active projects / goals",
+    "relationships": "People in their life",
+    "wishes":        "Wishes / plans",
+    "notes":         "Notes",
+}
+
+_IDENTITY_FIELDS = ["name", "age", "birthday", "city", "job",
+                    "language", "school", "nationality"]
+
+
+def format_memory_for_prompt(memory: dict | None) -> str:
+    """Build the memory block that goes into the system prompt.
+
+    This used to dump everything. It now sends three things:
+
+      1. IDENTITY  - always, in full. It is small, and it is wrong for the
+         assistant to have to look up your name.
+      2. RECENT    - the most recently updated entries from every other
+         category, up to PROMPT_CORE_CHARS. Recency is the cheapest useful
+         relevance signal available without embeddings.
+      3. AN INDEX  - the *keys* of everything else, values omitted.
+
+    Point 3 is what makes recall work at all. A model cannot decide to look
+    something up if it does not know the thing exists: with only points 1 and 2,
+    "who is Ayse?" would get "I don't know" while ayse_sister sat on disk
+    unread. The index costs a few hundred characters and turns recall from a
+    gamble into a lookup.
+
+    Net effect on latency: this block is SMALLER than the old full dump, so
+    every session connects with fewer tokens. Occasionally the model spends one
+    extra round trip on recall_memory - covered by the acknowledgment it
+    already speaks before any slow step."""
+    if not memory:
+        return ""
+
+    core_lines: list[str] = []
+
+    # 1. Identity - always, in full
+    identity = memory.get("identity", {}) or {}
+    for field in _IDENTITY_FIELDS:
+        val = _entry_value(identity.get(field))
+        if not val:
+            continue
+        if field == "language":
+            # Labelled as an observation, not a setting. A bare "Language:
+            # English" line written months ago reads like a standing order and
+            # was one of the reasons a Turkish question came back in English.
+            core_lines.append(
+                f"Has spoken to you in: {val} (an observation about the past — "
+                f"always answer in the language of their CURRENT message)")
+        else:
+            core_lines.append(f"{field.title()}: {val}")
+    for key, entry in identity.items():
+        if key in _IDENTITY_FIELDS:
+            continue
+        val = _entry_value(entry)
+        if val:
+            core_lines.append(f"{_pretty(key).title()}: {val}")
+
+    # 2. Everything else, most recently updated first
+    rest: list[tuple[str, str, str, str]] = []   # (updated, cat, key, value)
+    for cat in _CATEGORY_LABELS:
+        for key, entry in (memory.get(cat, {}) or {}).items():
+            val = _entry_value(entry)
+            if not val:
+                continue
+            updated = (entry.get("updated", "") if isinstance(entry, dict) else "") or "0000-00-00"
+            rest.append((updated, cat, key, val))
+    rest.sort(key=lambda t: t[0], reverse=True)
+
+    used    = sum(len(l) + 1 for l in core_lines)
+    shown: dict[str, list[str]] = {}
+    overflow: dict[str, list[str]] = {}
+
+    # Recency decides order, but no single category may take the whole budget.
+    # Without the cap, someone with forty stored preferences gets a prompt that
+    # is forty preferences and not one person's name — the categories that
+    # matter most in conversation are also the ones that change least often, so
+    # pure recency systematically buries them.
+    per_cat_used: dict[str, int] = {}
+    for _updated, cat, key, val in rest:
+        line = f"  - {_pretty(key).title()}: {val}"
+        if (per_cat_used.get(cat, 0) < PROMPT_MAX_PER_CATEGORY
+                and used + len(line) + 1 <= PROMPT_CORE_CHARS):
+            shown.setdefault(cat, []).append(line)
+            per_cat_used[cat] = per_cat_used.get(cat, 0) + 1
+            used += len(line) + 1
+        else:
+            overflow.setdefault(cat, []).append(_pretty(key))
+
+    # The index is a table of contents, so it is interleaved across categories
+    # rather than continuing in recency order. Sorted by recency it would list
+    # twenty-four preferences before the first relationship, and the one entry
+    # the index exists for — the old fact the model has no other way to know
+    # about — would fall off the end.
+    indexed: list[str] = []
+    if overflow:
+        cats  = [c for c in _CATEGORY_LABELS if overflow.get(c)]
+        cursor = {c: 0 for c in cats}
+        while cats:
+            for cat in list(cats):
+                i = cursor[cat]
+                if i >= len(overflow[cat]):
+                    cats.remove(cat)
+                    continue
+                indexed.append(overflow[cat][i])
+                cursor[cat] = i + 1
+
+    for cat, label in _CATEGORY_LABELS.items():
+        if shown.get(cat):
+            core_lines.append("")
+            core_lines.append(f"{label}:")
+            core_lines.extend(shown[cat])
+
+    if not core_lines and not indexed:
+        return ""
+
+    out = [
+        "[WHAT YOU KNOW ABOUT THIS PERSON — use naturally, never recite like a list]",
+        *core_lines,
+    ]
+
+    # 3. The index of what is on disk but not in this prompt
+    if indexed:
+        budget, names = PROMPT_INDEX_CHARS, []
+        for n in indexed:
+            if budget - len(n) - 2 < 0:
+                break
+            names.append(n)
+            budget -= len(n) + 2
+        if names:
+            out.append("")
+            out.append(
+                "[ALSO REMEMBERED — values not shown here. Call recall_memory "
+                "with a keyword to read any of these before saying you do not know]"
+            )
+            out.append(", ".join(names)
+                       + (f" (+{len(indexed) - len(names)} more)"
+                          if len(indexed) > len(names) else ""))
+
+    return "\n".join(out) + "\n"
+
+
+# ── Recall ────────────────────────────────────────────────────────────────────
+
+def _score(query_words: list[str], cat: str, key: str, value: str) -> int:
+    """Cheap lexical relevance. No embeddings, no network, no model call - this
+    runs in well under a millisecond, which is the entire point: recall must
+    cost one model round trip, never two."""
+    hay_key = _pretty(key).lower()
+    hay_val = value.lower()
+    score   = 0
+    for w in query_words:
+        if not w:
+            continue
+        if w == hay_key:
+            score += 10
+        elif w in hay_key:
+            score += 6
+        if w in hay_val:
+            score += 3
+        if w in cat:
+            score += 1
+    return score
+
+
+def search_memory(query: str, limit: int = 8) -> str:
+    """Find stored facts matching `query`. Backs the recall_memory tool.
+
+    An empty query is treated as "show me everything you know", capped - the
+    model asks that when the user says "what do you remember about me?"."""
+    memory = load_memory()
+    words  = [w for w in re.split(r"[^\w]+", (query or "").lower()) if len(w) > 1]
+
+    rows: list[tuple[int, str, str, str]] = []
+    for cat, items in memory.items():
+        if not isinstance(items, dict):
+            continue                     # skip 'sessions', which is a list
+        for key, entry in items.items():
+            val = _entry_value(entry)
+            if not val:
+                continue
+            s = _score(words, cat, key, val) if words else 1
+            if s > 0:
+                rows.append((s, cat, key, val))
+
+    if not rows:
+        return (f"Nothing stored about '{query}'." if query
+                else "I have not stored anything about this person yet.")
+
+    rows.sort(key=lambda r: (-r[0], r[2]))
+    lines = [f"{cat}/{_pretty(key)}: {val}" for _s, cat, key, val in rows[:max(1, limit)]]
+    head  = (f"Stored facts matching '{query}':" if query
+             else "Everything currently stored:")
+    more  = (f"\n(+{len(rows) - len(lines)} more — search with a narrower keyword)"
+             if len(rows) > len(lines) else "")
+    return head + "\n" + "\n".join(lines) + more
+
+
+def all_entries_for_ui() -> list[dict]:
+    """Flat list for the memory panel: what JARVIS knows, and when it learned it.
+    Sorted newest first so the panel opens on what changed most recently."""
+    memory = load_memory()
+    rows = []
+    for cat, items in memory.items():
+        if not isinstance(items, dict):
+            continue
+        for key, entry in items.items():
+            val = _entry_value(entry)
+            if not val:
+                continue
+            rows.append({
+                "category": cat,
+                "key":      key,
+                "value":    val,
+                "updated":  (entry.get("updated", "") if isinstance(entry, dict) else ""),
+            })
+    rows.sort(key=lambda r: (r["updated"] or "0000-00-00"), reverse=True)
+    return rows
+
+def remember(key: str, value: str, category: str = "notes") -> str:
+    valid = {"identity", "preferences", "projects", "relationships", "wishes", "notes"}
+    if category not in valid:
+        category = "notes"
+    update_memory({category: {key: {"value": value}}})
+    return f"Remembered: {category}/{key} = {value}"
+
+
+def forget(key: str, category: str = "notes") -> str:
+    memory = load_memory()
+    cat    = memory.get(category, {})
+    if key in cat:
+        del cat[key]
+        memory[category] = cat
+        save_memory(memory)
+        return f"Forgotten: {category}/{key}"
+    return f"Not found: {category}/{key}"
+
+
+forget_memory = forget
+
+
+# ── Session memory ─────────────────────────────────────────────────────────────
+
+_SESSION_MAX = 3   # safety cap — in practice 0-1 entries after pop
+
+
+def save_session_summary(summary: str, language: str = "") -> None:
+    """Append a 1-2 sentence session summary to long_term.json['sessions']."""
+    summary = (summary or "").strip()
+    if not summary:
+        return
+    memory   = load_memory()
+    sessions = memory.get("sessions", [])
+    if not isinstance(sessions, list):
+        sessions = []
+    entry: dict = {
+        "date":    datetime.now().strftime("%Y-%m-%d"),
+        "summary": summary[:280],
+    }
+    if language:
+        entry["language"] = language
+    sessions.append(entry)
+    memory["sessions"] = sessions[-_SESSION_MAX:]
+    with _lock:
+        MEMORY_PATH.parent.mkdir(parents=True, exist_ok=True)
+        MEMORY_PATH.write_text(
+            json.dumps(memory, indent=2, ensure_ascii=False),
+            encoding="utf-8",
+        )
+    print(f"[Memory] 📝 Session saved ({entry['date']}): {summary[:60]}…")
+
+
+def pop_last_session() -> dict | None:
+    """
+    Return AND remove the most recent session entry.
+    Calling this consumes the entry so it is never repeated in future briefings.
+    """
+    with _lock:
+        if not MEMORY_PATH.exists():
+            return None
+        try:
+            memory   = json.loads(MEMORY_PATH.read_text(encoding="utf-8"))
+            sessions = memory.get("sessions", [])
+            if not isinstance(sessions, list) or not sessions:
+                return None
+            entry = sessions.pop()          # remove the last entry
+            memory["sessions"] = sessions
+            MEMORY_PATH.write_text(
+                json.dumps(memory, indent=2, ensure_ascii=False),
+                encoding="utf-8",
+            )
+            return entry
+        except Exception as e:
+            print(f"[Memory] ⚠️ pop_last_session error: {e}")
             return None

plugins/immagini.py

@@ -0,0 +1,98 @@
+"""Generazione di immagini in locale con mflux (MLX su Apple Silicon): Z-Image Turbo, FLUX e altri.
+
+Gira in un ambiente separato (.venv-mflux) come processo esterno, così la memoria viene liberata a fine generazione.
+Il modello si sceglie nelle impostazioni (scheda Strumenti); il primo uso scarica i pesi da Hugging Face.
+"""
+from __future__ import annotations
+
+import re
+import subprocess
+import sys
+import time
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
+from avatar.settings import BASE_DIR, DATA_DIR, Settings  # noqa: E402
+
+VENV = BASE_DIR / ".venv-mflux" / "bin"
+OUT_DIR = DATA_DIR / "immagini"
+# comando mflux, passi e guida consigliati per ciascuna famiglia
+MODELLI = {
+    "z-image-turbo": {"cmd": "mflux-generate-z-image-turbo", "steps": 8, "guidance": None, "default_repo": "mflux-community/z-image-turbo-mflux-q4"},
+    "schnell": {"cmd": "mflux-generate", "steps": 4, "guidance": None, "default_repo": "dhairyashil/FLUX.1-schnell-mflux-4bit", "base": "schnell"},
+    "dev": {"cmd": "mflux-generate", "steps": 20, "guidance": 3.5, "default_repo": "", "base": "dev"},
+    "qwen": {"cmd": "mflux-generate-qwen", "steps": 20, "guidance": 4.0, "default_repo": ""},
+}
+FORMATI = {"quadrata": (1024, 1024), "orizzontale": (1280, 768), "verticale": (768, 1280), "schermo": (1536, 864), "piccola": (768, 768)}
+_last: dict = {"path": ""}
+
+
+def _cfg() -> tuple[str, str]:
+    s = Settings()
+    fam = str(s.get("immagini_famiglia") or "z-image-turbo")
+    repo = str(s.get("immagini_modello") or "").strip() or MODELLI.get(fam, MODELLI["z-image-turbo"])["default_repo"]
+    return fam, repo
+
+
+def genera(params: dict, ctx: dict) -> str:
+    prompt = str(params.get("descrizione", "")).strip()
+    if not prompt:
+        return "Errore: serve la descrizione dell'immagine (in inglese, dettagliata)."
+    if not (VENV / "mflux-generate").exists():
+        return "Generazione immagini non disponibile: manca l'ambiente .venv-mflux (uv venv .venv-mflux && uv pip install mflux)."
+    fam, repo = _cfg()
+    m = MODELLI.get(fam, MODELLI["z-image-turbo"])
+    w, h = FORMATI.get(str(params.get("formato", "quadrata")).lower(), FORMATI["quadrata"])
+    steps = int(params.get("passi") or m["steps"])
+    OUT_DIR.mkdir(parents=True, exist_ok=True)
+    stem = re.sub(r"[^a-z0-9]+", "-", prompt.lower())[:40].strip("-") or "immagine"
+    out = OUT_DIR / f"{time.strftime('%Y%m%d-%H%M%S')}-{stem}.png"
+    cmd = [str(VENV / m["cmd"]), "--prompt", prompt, "--steps", str(steps), "--width", str(w), "--height", str(h), "--output", str(out)]
+    if repo:
+        cmd += ["--model", repo]
+        if m.get("base") and "/" in repo:
+            cmd += ["--base-model", m["base"]]
+    if m["guidance"] is not None:
+        cmd += ["--guidance", str(params.get("guida") or m["guidance"])]
+    if params.get("seme") is not None:
+        cmd += ["--seed", str(int(params["seme"]))]
+    if params.get("immagine_base"):
+        cmd += ["--image-path", str(params["immagine_base"]), "--image-strength", str(params.get("forza") or 0.5)]
+    if ctx.get("log"):
+        ctx["log"](f"[immagini] {fam} {w}x{h} {steps} passi: {prompt[:80]}")
+    t = time.time()
+    try:
+        proc = subprocess.run(cmd, capture_output=True, text=True, timeout=1800)
+    except subprocess.TimeoutExpired:
+        return "Generazione interrotta: troppo lenta (oltre 30 minuti)."
+    if proc.returncode != 0 or not out.exists():
+        err = (proc.stderr or proc.stdout).strip().splitlines()
+        tail = " | ".join(err[-3:]) if err else f"codice {proc.returncode}"
+        if "not found" in tail.lower() or "401" in tail or "403" in tail:
+            tail += " (controlla il nome del modello nelle impostazioni, o accetta la licenza su Hugging Face)"
+        return f"Generazione fallita: {tail[:400]}"
+    _last["path"] = str(out)
+    secs = time.time() - t
+    if params.get("mostra", True) not in (False, "false", "no"):
+        subprocess.Popen(["open", str(out)])
+    return f"Immagine generata in {secs:.0f} s e salvata in {out} ({w}x{h}, {fam}).\nIMMAGINE: {out}"
+
+
+def ultima(params: dict, ctx: dict) -> str:
+    files = sorted(OUT_DIR.glob("*.png"), key=lambda p: p.stat().st_mtime, reverse=True) if OUT_DIR.exists() else []
+    if not files:
+        return "Nessuna immagine generata finora."
+    n = max(1, min(int(params.get("numero") or 5), 20))
+    if str(params.get("apri", "")).lower() in ("true", "1", "sì", "si"):
+        subprocess.Popen(["open", str(files[0])])
+    return "Ultime immagini:\n" + "\n".join(f"- {p.name}" for p in files[:n]) + f"\nIMMAGINE: {files[0]}"
+
+
+TOOLS = [
+    {"name": "immagine_genera", "description": "Genera un'immagine in locale (MLX) da una descrizione. Scrivi la descrizione in INGLESE, dettagliata (soggetto, ambiente, luce, stile, inquadratura), anche se l'utente parla italiano. formato: quadrata (default), orizzontale, verticale, schermo, piccola. L'immagine viene salvata e aperta in Anteprima; richiede 10-60 secondi, il primo uso scarica il modello. Con immagine_base (percorso) e forza (0-1) parte da un'immagine esistente.",
+     "parameters": {"type": "object", "properties": {"descrizione": {"type": "string"}, "formato": {"type": "string", "enum": ["quadrata", "orizzontale", "verticale", "schermo", "piccola"]},
+                    "passi": {"type": "integer"}, "seme": {"type": "integer"}, "immagine_base": {"type": "string"}, "forza": {"type": "number"}, "mostra": {"type": "boolean"}},
+                    "required": ["descrizione"]}, "run": genera},
+    {"name": "immagine_ultime", "description": "Elenca le ultime immagini generate (cartella data/immagini); con apri=true apre la più recente.",
+     "parameters": {"type": "object", "properties": {"numero": {"type": "integer"}, "apri": {"type": "boolean"}}}, "run": ultima},
+]