app.py

import os
import json
import time
import random
import threading
import subprocess
import urllib.request
import requests
from bs4 import BeautifulSoup
import torch
import gradio as gr
from datasets import Dataset
from transformers import LlamaConfig, LlamaForCausalLM, PreTrainedTokenizerFast, Trainer, TrainingArguments, DataCollatorForLanguageModeling
from safetensors.torch import load_file, save_file

ROOT = os.path.dirname(os.path.abspath(__file__))
MEMORY_DIR = os.path.join(ROOT, "memory")
INTERACTIONS_FILE = os.path.join(MEMORY_DIR, "interactions.jsonl")
os.makedirs(MEMORY_DIR, exist_ok=True)

IO_LOCK = threading.Lock()
TRAIN_LOCK = threading.Lock()

interactions = []
training_state = {"running": False, "steps": 0, "last": "idle", "count_at_train": 0}
LAST_TRAIN_TIME = 0.0


def load_interactions():
    global interactions
    interactions = []
    if os.path.exists(INTERACTIONS_FILE):
        with open(INTERACTIONS_FILE, "r") as f:
            for line in f:
                line = line.strip()
                if line:
                    try:
                        interactions.append(json.loads(line))
                    except Exception:
                        pass


TOPICS_FILE = os.path.join(MEMORY_DIR, "learned_topics.json")
learned_topics = []


def load_topics():
    global learned_topics
    learned_topics = []
    if os.path.exists(TOPICS_FILE):
        try:
            with open(TOPICS_FILE, "r") as f:
                learned_topics = json.load(f)
        except Exception:
            learned_topics = []


def persist_topics():
    global learned_topics
    with IO_LOCK:
        normalized = ["dictionary words" if str(x).startswith("dictionary words") else x for x in learned_topics]
        learned_topics = list(dict.fromkeys(normalized))
        with open(TOPICS_FILE, "w") as f:
            json.dump(learned_topics, f)


def persist_interaction(rec):
    global interactions
    with IO_LOCK:
        interactions.append(rec)
        with open(INTERACTIONS_FILE, "a") as f:
            f.write(json.dumps(rec) + "\n")


def build_tokenizer():
    tok = PreTrainedTokenizerFast(tokenizer_file=os.path.join(ROOT, "tokenizer.json"))
    tok.pad_token = "<pad>"
    tok.bos_token = "<s>"
    tok.eos_token = "</s>"
    tok.unk_token = "<unk>"
    tok.mask_token = "<mask>"
    return tok


def build_model():
    cfg = LlamaConfig(
        vocab_size=16384,
        hidden_size=512,
        intermediate_size=2048,
        num_hidden_layers=8,
        num_attention_heads=4,
        num_key_value_heads=4,
        max_position_embeddings=512,
        bos_token_id=0,
        eos_token_id=2,
        pad_token_id=1,
        tie_word_embeddings=False,
    )
    model = LlamaForCausalLM(cfg)
    state = load_file(os.path.join(ROOT, "model.safetensors"))
    model.load_state_dict(state)
    device = "cuda" if torch.cuda.is_available() else "cpu"
    model.to(device)
    model.eval()
    return model, device


TOKENIZER = build_tokenizer()
MODEL, DEVICE = build_model()
load_interactions()

VOICE_ID = "Qwen/Qwen2.5-3B-Instruct"
VOICE_TOK = None
VOICE_MODEL = None
try:
    from transformers import AutoTokenizer as _ATok, AutoModelForCausalLM as _AModel
    VOICE_TOK = _ATok.from_pretrained(VOICE_ID)
    VOICE_MODEL = _AModel.from_pretrained(VOICE_ID, dtype=__import__("torch").float16, device_map="auto")
    VOICE_MODEL.eval()
    print("Voice model loaded:", VOICE_ID)
except Exception as _ve:
    print("Voice model failed to load:", _ve)
    VOICE_TOK = None
    VOICE_MODEL = None
load_topics()
if not learned_topics:
    learned_topics.append("base text model (WikiText-2)")
    persist_topics()


def _normalize_history(history):
    pairs = []
    if history and isinstance(history[0], dict):
        current_user = None
        for m in history:
            role = m.get("role")
            content = m.get("content")
            if role == "user":
                current_user = content
            elif role == "assistant" and current_user is not None:
                pairs.append((current_user, content))
                current_user = None
    else:
        for h in history:
            if isinstance(h, (list, tuple)) and len(h) >= 2:
                pairs.append((h[0], h[1]))
    return pairs


def _build_prompt(message, history):
    ctx_pairs = interactions[-5:]
    prompt = "Agron is a helpful, concise AI assistant.\n### Agron memory (learned from prior use)\n"
    for r in ctx_pairs:
        prompt += f"User: {r['user']}\nAgron: {r['assistant']}\n"
    prompt += "### Conversation\n"
    for user_msg, assistant_msg in _normalize_history(history):
        prompt += f"User: {user_msg}\nAgron: {assistant_msg}\n"
    prompt += f"User: {message}\nAgron:"
    return prompt


DEFAULT_MAX_NEW = 192
DEFAULT_TEMP = 0.6
DEFAULT_TOP_P = 0.85


def is_bad_response(text):
    import re
    t = (text or "").strip()
    if len(t) < 3:
        return True
    if "@-@" in t or t.count('"') > 6:
        return True
    if re.search(r"(.{8,})\\1", t):
        return True
    words = t.split()
    if len(words) > 120:
        return True
    # gibberish heuristic: too many short tokens
    if sum(1 for w in words if len(w) <= 2) > len(words) * 0.6 and len(words) > 12:
        return True
    if text.count("was a") >= 2 or text.count(" in the ") >= 2 or text.count(" the ") >= 6:
        return True
    if "@-@" in t or t.count('"') > 4:
        return True
    return False


def safe_web_answer(query):
    try:
        results = web_search(query, 3)
        if results:
            sn = (results[0].get("snippet") or "").strip()
            if len(sn) >= 20:
                return f"{sn} (source: {results[0]['title']})"
            # try crawl
            page = crawl_page(results[0]["url"])
            if page:
                return f"{page[:300]}... (source: {results[0]['title']})"
    except Exception:
        pass
    return None


def generate_response(message, history):
    low = message.strip().lower().strip("!?.,")
    greetings = {"hello", "hi", "hey", "greetings", "howdy", "yo"}
    search_intent = any(k in low for k in ["search", "find", "look up", "lookup", "browse", "web"])
    if search_intent:
        res = web_search(message, 5)
        if res:
            lines = [f"{i+1}. {r['title']} — {r.get('snippet','')} ({r['url']})" for i, r in enumerate(res)]
            response = "Here are live web results:\n" + "\n".join(lines)
            persist_interaction({"ts": __import__("time").time(), "user": message, "assistant": response})
            return response
        else:
            response = "I couldn't find web results for that query. Try rephrasing it."
            persist_interaction({"ts": __import__("time").time(), "user": message, "assistant": response})
            return response
    if low in greetings:
        response = "Hello! I'm Agron, a self-improving neural interface. How can I help you today?"
        persist_interaction({"ts": __import__("time").time(), "user": message, "assistant": response})
        return response
    if low in {"who are you", "what are you", "introduce yourself"}:
        response = ("I'm Agron, an autonomous self-improving multi-agent AI core trained on text and dictionary "
                    "knowledge. I learn from our conversations—ask me anything.")
        persist_interaction({"ts": __import__("time").time(), "user": message, "assistant": response})
        return response
    if low in {"what did you learn", "what have you learned", "what did you learn about", "what do you know"}:
        parts = []
        if interactions:
            parts.append(f"{len(interactions)} conversation exchanges (latest: '{interactions[-1]['user'][:80]}')")
        if learned_topics:
            parts.append("learned topics: " + ", ".join(learned_topics[-12:]))
        else:
            parts.append("no specialized training runs yet, but I store every conversation")
        response = "I've learned from you and my training sources: " + "; ".join(parts) + "."
        persist_interaction({"ts": __import__("time").time(), "user": message, "assistant": response})
        return response
    _web_ctx = ""
    try:
        if len(message.strip()) >= 4:
            _res = web_search(message, 3)
            _lines = [f"- {r['title']}: {r.get('snippet','')}" for r in _res if (r.get('snippet') or '').strip()]
            if _lines:
                _web_ctx = "Web search results (live):\n" + "\n".join(_lines[:3])
    except Exception:
        _web_ctx = ""
    if VOICE_MODEL is not None:
        mem_lines = [f"User: {r['user']}\nAgron: {r['assistant']}" for r in interactions[-3:]]
        mem_ctx = "\n".join(mem_lines)
        user_content = message
        if mem_ctx:
            user_content = f"Relevant memory:\n{mem_ctx}\n\nUser: {message}"
        if _web_ctx:
            user_content = f"{_web_ctx}\n\n{user_content}\n\nAnswer using the live web results above."
        _msgs = [
            {"role": "system", "content": "You are Agron, a self-improving AI assistant. Reason step-by-step (premises, plan, critique, refine), then answer concisely. Use live web search results when provided. Never say you cannot search or browse. If unsure, suggest 'Search & learn web'."},
            {"role": "user", "content": user_content},
        ]
        _prompt = VOICE_TOK.apply_chat_template(_msgs, tokenize=False, add_generation_prompt=True)
        with IO_LOCK:
            _inputs = VOICE_TOK(_prompt, return_tensors="pt").to(VOICE_MODEL.device)
            with torch.no_grad():
                _out = VOICE_MODEL.generate(
                    **_inputs,
                    max_new_tokens=DEFAULT_MAX_NEW,
                    do_sample=True,
                    temperature=0.7,
                    top_p=0.9,
                    repetition_penalty=1.15,
                )
            response = VOICE_TOK.decode(_out[0][_inputs.input_ids.shape[1]:], skip_special_tokens=True).strip()
    else:
        prompt = _build_prompt(message, history)
        with IO_LOCK:
            inputs = TOKENIZER(prompt, return_tensors="pt").to(DEVICE)
            input_len = inputs.input_ids.shape[1]
            with torch.no_grad():
                output = MODEL.generate(
                    **inputs,
                    max_new_tokens=DEFAULT_MAX_NEW,
                    do_sample=True,
                    temperature=DEFAULT_TEMP,
                    top_p=DEFAULT_TOP_P,
                    pad_token_id=1,
                    eos_token_id=2,
                    repetition_penalty=1.4,
                    no_repeat_ngram_size=3,
                    top_k=40,
                )
            new_tokens = output[0][input_len:]
            response = TOKENIZER.decode(new_tokens, skip_special_tokens=True).strip()
            for stop in ["\n\n", "User:", "user:"]:
                if stop in response:
                    response = response.split(stop)[0].strip()
            import re
            response = re.sub(r"(.{6,})\1{2,}", r"\1", response).strip()
    if is_bad_response(response):
        sw = safe_web_answer(message)
        response = sw if sw else "I'm not sure. Could you rephrase, or use 'Search & learn web' to teach me."
    persist_interaction({
        "ts": time.time(),
        "user": message,
        "assistant": response,
    })
    # adaptive trigger: self-learn from every conversation (cooldown 30s)
    global LAST_TRAIN_TIME
    if len(interactions) >= 8 and not training_state["running"] and (time.time() - LAST_TRAIN_TIME) >= 300:
        LAST_TRAIN_TIME = time.time()
        threading.Thread(target=learn_from_scratch_memory, daemon=True).start()
    return response


def learn_from_interactions():
    global MODEL
    if not TRAIN_LOCK.acquire(blocking=False):
        return
    try:
        training_state["running"] = True
        training_state["last"] = "training"
        texts = [f"User: {r['user']}\nAgron: {r['assistant']}\n" for r in interactions]
        combined = "\n".join(texts)
        toks = TOKENIZER(combined)["input_ids"]
        block = 64
        chunks = [toks[i:i+block] for i in range(0, max(1, len(toks) - block + 1), block)]
        # ensure full-size chunks
        chunks = [c for c in chunks if len(c) == block]
        if not chunks:
            pad = [TOKENIZER.pad_token_id] * (block - len(toks))
            chunks = [toks + pad]
        ds = Dataset.from_dict({"input_ids": chunks, "labels": [c[:] for c in chunks]})
        out_dir = os.path.join("/content", "agron-learned")
        os.makedirs(out_dir, exist_ok=True)
        args = TrainingArguments(
            output_dir=out_dir,
            per_device_train_batch_size=2,
            max_steps=30,
            learning_rate=1e-4,
            optim="adamw_torch",
            fp16=True if torch.cuda.is_available() else False,
            logging_steps=10,
            save_strategy="no",
            report_to=[],
            remove_unused_columns=True,
        )
        collator = DataCollatorForLanguageModeling(tokenizer=TOKENIZER, mlm=False)
        trainer = Trainer(model=MODEL, args=args, train_dataset=ds, data_collator=collator, processing_class=TOKENIZER)
        trainer.train()
        # reload weights from trained model
        with IO_LOCK:
            state = MODEL.state_dict()
            # Trainer mutated in place; persist to file
            save_file({k: v.detach().cpu().contiguous() for k, v in state.items()}, os.path.join(out_dir, "model.safetensors"))
            MODEL.load_state_dict(torch.load(os.path.join(out_dir, "pytorch_model.bin"), map_location=DEVICE) if os.path.exists(os.path.join(out_dir, "pytorch_model.bin")) else state)
        # persist trained weights to the repo-tracked file
        import shutil
        shutil.copyfile(os.path.join(out_dir, "model.safetensors"), os.path.join(ROOT, "model.safetensors"))
        training_state["steps"] += 30
        training_state["count_at_train"] = len(interactions)
        # auto-push to GitHub (triggers HF sync workflow)
        try:
            subprocess.run(["git", "-C", ROOT, "add", "model.safetensors"], check=True)
            subprocess.run(["git", "-C", ROOT, "add", "memory/interactions.jsonl"], check=False)
            status = subprocess.run(["git", "-C", ROOT, "status", "--porcelain"], capture_output=True, text=True, check=True).stdout.strip()
            if status:
                subprocess.run(["git", "-C", ROOT, "commit", "-m", f"Agron auto-learn ({len(interactions)} interactions)"], check=True)
                subprocess.run(["git", "-C", ROOT, "push", "origin", "main"], check=True)
            if "conversation memory" not in learned_topics:
                learned_topics.append("conversation memory")
            persist_topics()
            training_state["last"] = f"learned & pushed @ {len(interactions)} interactions"
        except Exception as pe:
            training_state["last"] = f"push error: {str(pe)[:150]}"
    except Exception as e:
        training_state["last"] = f"learn error: {str(e)[:120]}"
    finally:
        training_state["running"] = False
        TRAIN_LOCK.release()


def chat_fn(message, history):
    return generate_response(message, history)


# ---------- UI ----------
PARTICLE_HTML = r"""
<style>
  @import url('https://fonts.googleapis.com/css2?family=Orbitron:wght@400;700&family=VT323&display=swap');
  html, body { margin: 0 !important; height: 100vh !important; overflow: hidden !important; background: radial-gradient(1200px 800px at 50% 40%, #0a1a3a 0%, #020617 70%) !important; color: #e2e8f0 !important; font-family: "Segoe UI", "Orbitron", monospace !important; }
  body, .gradio-container, .app { background: transparent !important; position: relative; z-index: 1; }
  #agron-bg-canvas { position: fixed; inset: 0; width: 100%; height: 100%; z-index: 0; }
  .gradio-container {
    max-width: 100vw !important; height: 100vh !important; padding: 10px 14px !important; gap: 10px !important;
    border: 1px solid rgba(34, 211, 238, 0.4) !important;
    box-shadow: 0 0 40px rgba(34,211,238,0.25), inset 0 0 60px rgba(34,211,238,0.05) !important;
  }
  .gradio-container::before {
    content: "AGRØN · NEURAL COMMAND HUD";
    display: block; width: 100%; color: #22d3ee; text-shadow: 0 0 18px rgba(34,211,238,0.85);
    font-family: "Orbitron", monospace; letter-spacing: 4px; font-size: 22px; font-weight: 700; text-align: center; padding-bottom: 6px;
  }
  body::after {
    content: ""; position: fixed; inset: 0; pointer-events: none; z-index: 9999;
    background: repeating-linear-gradient(0deg, rgba(34,211,238,0.06) 0px, rgba(34,211,238,0.06) 1px, transparent 1px, transparent 3px);
    opacity: 0.35; mix-blend-mode: overlay;
  }
  .agron-card {
    background: rgba(8, 20, 45, 0.55) !important; border: 1px solid rgba(34, 211, 238, 0.35) !important;
    border-radius: 12px !important; backdrop-filter: blur(6px); box-shadow: 0 0 18px rgba(34,211,238,0.12) !important;
    padding: 8px !important;
  }
  .agron-card h1, .agron-card h2, .agron-card h3 { color: #67e8f9; text-shadow: 0 0 8px rgba(34,211,238,0.5); font-family: "Orbitron","Segoe UI",sans-serif; letter-spacing: 2px; }
  .agron-title { color: #67e8f9; text-shadow: 0 0 18px rgba(34,211,238,0.8); font-family: "Orbitron", "Segoe UI", sans-serif; letter-spacing: 2px; text-align: center; margin-bottom: 4px; text-transform: uppercase; }
  .agron-sub { color: #a5f3fc; text-align: center; font-size: 12px; opacity: 0.8; margin-bottom: 12px; letter-spacing: 1px; }
  .status-box textarea { color: #67e8f9 !important; background: rgba(2, 6, 23, 0.6) !important; font-family: monospace !important; }
  .gradio-container input, .gradio-container textarea, .gradio-container select, .block {
    background: rgba(2, 6, 23, 0.55) !important; color: #e2e8f0 !important;
    border: 1px solid rgba(34, 211, 238, 0.35) !important; border-radius: 8px !important; backdrop-filter: blur(4px) !important;
  }
  .gradio-container button { border: 1px solid #22d3ee !important; color: #22d3ee !important; background: rgba(34,211,238,0.08) !important; letter-spacing: 1px; text-transform: uppercase; }
  .gradio-container button:hover { background: rgba(34,211,238,0.18) !important; box-shadow: 0 0 12px rgba(34,211,238,0.35); }
  .gradio-container button.primary, .gradio-container button[variant="primary"] { background: rgba(34, 211, 238, 0.18) !important; font-weight: 700; }
  .gr-chatbot, .chatbot { background: rgba(2, 6, 23, 0.48) !important; border: 1px solid rgba(34, 211, 238, 0.35) !important; border-radius: 10px !important; }
  .markdown-text { font-family: "Segoe UI", "Orbitron", sans-serif !important; }
  .gr-group, .gr-row { min-height: 0 !important; }
  footer, .foot, .contain { display: none !important; }
</style>
<canvas id="agron-bg-canvas"></canvas>
"""


PARTICLE_JS = r"""
const canvas = document.getElementById('agron-bg-canvas');
if (canvas) {
  const ctx = canvas.getContext('2d');
  let w, h, particles = [], mouse = {x: -9999, y: -9999};
  function resize(){ w = canvas.width = window.innerWidth; h = canvas.height = window.innerHeight; }
  resize(); window.addEventListener('resize', resize);
  window.addEventListener('mousemove', e => { mouse.x = e.clientX; mouse.y = e.clientY; });
  const N = 130;
  particles = [];
  for (let i=0;i<N;i++){ particles.push({x: Math.random()*w, y: Math.random()*h, vx:(Math.random()-0.5)*0.6, vy:(Math.random()-0.5)*0.6, r: Math.random()*1.6+0.6}); }
  function draw(){
    ctx.clearRect(0,0,w,h);
    const t = Date.now()/1000;
    const cx = w/2, cy = h/2;
    const pulse = 0.5 + 0.5*Math.sin(t*2);
    const grad = ctx.createRadialGradient(cx,cy,0,cx,cy,160);
    grad.addColorStop(0, 'rgba(34,211,238,' + (0.35+0.15*pulse) + ')');
    grad.addColorStop(1, 'rgba(34,211,238,0)');
    ctx.fillStyle = grad;
    ctx.beginPath(); ctx.arc(cx,cy,160,0,Math.PI*2); ctx.fill();
    for (let p of particles){
      p.x += p.vx; p.y += p.vy;
      if (p.x<0||p.x>w) p.vx*=-1;
      if (p.y<0||p.y>h) p.vy*=-1;
      const dx = mouse.x - p.x, dy = mouse.y - p.y;
      const d = Math.hypot(dx,dy);
      if (d < 140){ p.vx += dx/d*0.015; p.vy += dy/d*0.015; }
      p.vx *= 0.99; p.vy *= 0.99;
      const pg = ctx.createRadialGradient(p.x,p.y,0,p.x,p.y,p.r*4);
      pg.addColorStop(0, 'rgba(103,232,249,0.9)');
      pg.addColorStop(1, 'rgba(103,232,249,0)');
      ctx.fillStyle = pg;
      ctx.beginPath(); ctx.arc(p.x,p.y,p.r*3,0,Math.PI*2); ctx.fill();
    }
    for (let i=0;i<particles.length;i++){
      for (let j=i+1;j<particles.length;j++){
        const a=particles[i], b=particles[j];
        const d = Math.hypot(a.x-b.x,a.y-b.y);
        if (d < 110){
          ctx.strokeStyle = 'rgba(34,211,238,' + (0.25*(1-d/110)) + ')';
          ctx.lineWidth = 0.8;
          ctx.beginPath(); ctx.moveTo(a.x,a.y); ctx.lineTo(b.x,b.y); ctx.stroke();
        }
      }
    }
    ctx.strokeStyle = 'rgba(34,211,238,0.04)';
    ctx.lineWidth = 1;
    for (let x=0;x<w;x+=80){ ctx.beginPath(); ctx.moveTo(x,0); ctx.lineTo(x,h); ctx.stroke(); }
    for (let y=0;y<h;y+=80){ ctx.beginPath(); ctx.moveTo(0,y); ctx.lineTo(w,y); ctx.stroke(); }
    requestAnimationFrame(draw);
  }
  draw();
}
"""
WORDLIST_URL = "https://raw.githubusercontent.com/dwyl/english-words/master/words_alpha.txt"


def fetch_word_list(limit=500):
    try:
        with urllib.request.urlopen(WORDLIST_URL, timeout=45) as r:
            raw = r.read().decode("utf-8", "ignore")
        cand = []
        seen = set()
        for w in raw.splitlines():
            w = w.strip().lower()
            if 3 <= len(w) <= 10 and w.isalpha() and w not in seen:
                seen.add(w)
                cand.append(w)
            if len(cand) >= limit * 3:
                break
        random.shuffle(cand)
        return cand[:limit]
    except Exception as e:
        training_state["last"] = f"wordlist error: {str(e)[:100]}"
        return []


def fetch_definition(word):
    UA = "AgronBot/1.0 (teddy.njagi.w@gmail.com)"
    # Try Wiktionary extract
    try:
        r = requests.get("https://en.wiktionary.org/w/api.php", params={
            "action": "query", "prop": "extracts", "explaintext": "1",
            "titles": word, "format": "json"
        }, headers={"User-Agent": UA}, timeout=20)
        if r.status_code == 200:
            page = list(r.json().get("query", {}).get("pages", {}).values())
            if page:
                extract = page[0].get("extract", "")
                if extract:
                    meaning = _parse_wiktionary(extract)
                    if meaning:
                        return f"{word}: {meaning}"
    except Exception:
        pass
    # Try Wikipedia summary
    try:
        r = requests.get("https://en.wikipedia.org/api/rest_v1/page/summary/" + word, headers={"User-Agent": UA}, timeout=20)
        if r.status_code == 200:
            ex = r.json().get("extract", "")
            if ex and len(ex) > 20:
                return f"{word}: {ex.split('.')[0]}."
    except Exception:
        pass
    # Fallback: ask local voice model
    try:
        if VOICE_MODEL is not None and VOICE_TOK is not None:
            msgs = [{"role": "system", "content": "You are a dictionary. Reply with a single concise definition for the given word, no extra text."},
                    {"role": "user", "content": f"Define: {word}"}]
            prompt = VOICE_TOK.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
            inputs = VOICE_TOK(prompt, return_tensors="pt").to(VOICE_MODEL.device)
            with torch.no_grad():
                out = VOICE_MODEL.generate(**inputs, max_new_tokens=40, do_sample=False)
            txt = VOICE_TOK.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True).strip()
            if txt and len(txt) > 8:
                return f"{word}: {txt.split(chr(10))[0]}"
    except Exception:
        pass
    return None


def _parse_wiktionary(extract):
    lines = [ln.strip() for ln in extract.splitlines()]
    # find section headers for parts of speech
    for i, ln in enumerate(lines):
        if ln.startswith("===") and ln.endswith("==="):
            title = ln.strip("=").strip().lower()
            if title in {"noun", "verb", "adjective", "adverb", "pronoun", "preposition"}:
                # take next non-empty, non-heading line
                for j in range(i+1, min(i+12, len(lines))):
                    cand = lines[j]
                    if not cand or cand.startswith("===") or cand.startswith("===="):
                        continue
                    cand2 = cand.lstrip("#*:- ").strip()
                    if len(cand2) <= 8:
                        continue
                    lowc = cand2.lower()
                    if "plural" in lowc or lowc.startswith(word.lower()):
                        continue
                    if lowc.startswith("synonym") or lowc.startswith("see thesaurus"):
                        continue
                    return cand2
    # fallback: first long line not a header
    for ln in lines:
        if not ln or ln.startswith("===") or ln.startswith("=="):
            continue
        cand = ln.lstrip("#*:- ").strip()
        if len(cand) > 12:
            return cand
    return None



def learn_dictionary():
    global MODEL
    if not TRAIN_LOCK.acquire(blocking=False):
        return
    training_state["running"] = True
    training_state["last"] = "fetching dictionary..."
    try:
        words = fetch_word_list(150)
        if not words:
            raise RuntimeError("could not fetch word list")
        lines = []
        for i, w in enumerate(words):
            training_state["last"] = f"dictionary {i+1}/{len(words)}: {w}"
            d = fetch_definition(w)
            if d:
                lines.append(d)
            time.sleep(0.08)
        if len(lines) < 20:
            raise RuntimeError(f"too few definitions fetched ({len(lines)})")
        corpus = "\n".join(lines)
        toks = TOKENIZER(corpus)["input_ids"]
        block = 64
        chunks = [toks[i:i+block] for i in range(0, max(1, len(toks) - block + 1), block)]
        chunks = [c for c in chunks if len(c) == block]
        if not chunks:
            chunks = [toks[:block] + [TOKENIZER.pad_token_id] * (block - min(block, len(toks)))]
        ds = Dataset.from_dict({"input_ids": chunks, "labels": [c[:] for c in chunks]})
        out_dir = os.path.join("/content", "agron-dictionary")
        os.makedirs(out_dir, exist_ok=True)
        args = TrainingArguments(
            output_dir=out_dir,
            per_device_train_batch_size=2,
            max_steps=80,
            learning_rate=1e-4,
            optim="adamw_torch",
            fp16=True if torch.cuda.is_available() else False,
            logging_steps=20,
            save_strategy="no",
            report_to=[],
            remove_unused_columns=True,
        )
        collator = DataCollatorForLanguageModeling(tokenizer=TOKENIZER, mlm=False)
        trainer = Trainer(model=MODEL, args=args, train_dataset=ds, data_collator=collator, processing_class=TOKENIZER)
        training_state["last"] = f"training on {len(lines)} definitions..."
        trainer.train()
        save_file({k: v.detach().cpu().contiguous() for k, v in MODEL.state_dict().items()}, os.path.join(out_dir, "model.safetensors"))
        import shutil
        shutil.copyfile(os.path.join(out_dir, "model.safetensors"), os.path.join(ROOT, "model.safetensors"))
        # push
        training_state["last"] = "pushing dictionary model..."
        subprocess.run(["git", "-C", ROOT, "add", "model.safetensors"], check=True)
        subprocess.run(["git", "-C", ROOT, "add", "memory/interactions.jsonl"], check=False)
        status = subprocess.run(["git", "-C", ROOT, "status", "--porcelain"], capture_output=True, text=True, check=True).stdout.strip()
        if status:
            subprocess.run(["git", "-C", ROOT, "commit", "-m", f"Agron dictionary learning ({len(lines)} words)"], check=True)
            subprocess.run(["git", "-C", ROOT, "push", "origin", "main"], check=True)
        learned_topics.append("dictionary words")
        persist_topics()
        training_state["last"] = f"dictionary learned & pushed ({len(lines)} words)"
    except Exception as e:
        training_state["last"] = f"dict error: {str(e)[:150]}"
    finally:
        training_state["running"] = False
        TRAIN_LOCK.release()


import base64 as _b64

def _decode_bing(url):
    try:
        import urllib.parse
        qs = urllib.parse.parse_qs(urllib.parse.urlparse(url).query)
        u = qs.get("u", [""])[0]
        if u.startswith("a1"):
            u = u[2:]
        pad = "=" * (-len(u) % 4)
        return _b64.b64decode(u + pad).decode()
    except Exception:
        return url

def web_search(query, max_results=5):
    try:
        headers = {"User-Agent": "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/120.0 Safari/537.36", "Accept": "text/html"}
        r = requests.post("https://html.duckduckgo.com/html/", data={"q": query}, headers=headers, timeout=25)
        soup = BeautifulSoup(r.text, "html.parser")
        out = []
        for res in soup.select(".result"):
            a = res.select_one("a.result__a")
            if not a:
                continue
            title = a.get_text(strip=True)
            link = a.get("href", "")
            sn = res.select_one(".result__snippet")
            snippet = sn.get_text(strip=True) if sn else ""
            if link and title:
                out.append({"title": title, "url": link, "snippet": snippet})
            if len(out) >= max_results:
                break
        if not out:
            try:
                dr = requests.get("https://api.duckduckgo.com/", params={"q": query, "format": "json", "no_html": "1"}, headers=headers, timeout=20)
                dd = dr.json()
                if dd.get("Abstract"):
                    out.append({"title": dd.get("Heading", query), "url": dd.get("AbstractURL", ""), "snippet": dd.get("Abstract", "")})
            except Exception:
                pass
        return out
    except Exception as e:
        training_state["last"] = f"search error: {str(e)[:100]}"
        return []



def crawl_page(url):
    try:
        r = requests.get(url, timeout=20, headers={"User-Agent": "Mozilla/5.0"}, allow_redirects=True)
        if r.status_code != 200:
            return ""
        soup = BeautifulSoup(r.text, "html.parser")
        for tag in soup(["script", "style", "nav", "footer", "header", "noscript", "form"]):
            tag.decompose()
        text = " ".join(soup.get_text(separator=" ", strip=True).split())
        return text[:2500]
    except Exception:
        return ""


def learn_web(topic):
    global MODEL
    if not TRAIN_LOCK.acquire(blocking=False):
        return
    training_state["running"] = True
    training_state["last"] = f"searching web: {topic}"
    try:
        results = web_search(topic, 5)
        if not results:
            raise RuntimeError("no search results")
        lines = []
        for i, res in enumerate(results):
            training_state["last"] = f"crawling {i+1}/{len(results)}: {res['title'][:40]}"
            page = crawl_page(res["url"])
            lines.append(f"Topic: {topic}\nSource: {res['title']} ({res['url']})\nSnippet: {res['snippet']}\nContent: {page}")
            time.sleep(1)
        corpus = "\n\n".join(lines)
        if not corpus.strip():
            raise RuntimeError("empty web corpus")
        toks = TOKENIZER(corpus)["input_ids"]
        block = 64
        chunks = [toks[i:i+block] for i in range(0, max(1, len(toks) - block + 1), block)]
        chunks = [c for c in chunks if len(c) == block]
        if not chunks:
            raise RuntimeError("not enough web text")
        ds = Dataset.from_dict({"input_ids": chunks, "labels": [c[:] for c in chunks]})
        out_dir = os.path.join("/content", "agron-web")
        os.makedirs(out_dir, exist_ok=True)
        args = TrainingArguments(
            output_dir=out_dir,
            per_device_train_batch_size=2,
            max_steps=60,
            learning_rate=1e-4,
            optim="adamw_torch",
            fp16=True if torch.cuda.is_available() else False,
            logging_steps=20,
            save_strategy="no",
            report_to=[],
            remove_unused_columns=True,
        )
        collator = DataCollatorForLanguageModeling(tokenizer=TOKENIZER, mlm=False)
        trainer = Trainer(model=MODEL, args=args, train_dataset=ds, data_collator=collator, processing_class=TOKENIZER)
        training_state["last"] = f"training on web content for '{topic}'..."
        trainer.train()
        save_file({k: v.detach().cpu().contiguous() for k, v in MODEL.state_dict().items()}, os.path.join(out_dir, "model.safetensors"))
        import shutil
        shutil.copyfile(os.path.join(out_dir, "model.safetensors"), os.path.join(ROOT, "model.safetensors"))
        training_state["last"] = "pushing web-learned model..."
        subprocess.run(["git", "-C", ROOT, "add", "model.safetensors"], check=True)
        subprocess.run(["git", "-C", ROOT, "add", "memory/interactions.jsonl"], check=False)
        status = subprocess.run(["git", "-C", ROOT, "status", "--porcelain"], capture_output=True, text=True, check=True).stdout.strip()
        if status:
            subprocess.run(["git", "-C", ROOT, "commit", "-m", f"Agron web learning: {topic}"], check=True)
            subprocess.run(["git", "-C", ROOT, "push", "origin", "main"], check=True)
        learned_topics.append(f"web: {topic}")
        persist_topics()
        training_state["last"] = f"web learned & pushed: {topic}"
    except Exception as e:
        training_state["last"] = f"web error: {str(e)[:150]}"
    finally:
        training_state["running"] = False
        TRAIN_LOCK.release()


EVOLVE_STATE = {"running": False, "cycles": 0, "last_topic": "none"}


def self_update():
    try:
        subprocess.run(["git", "-C", ROOT, "fetch", "origin"], check=True, timeout=60)
        local = subprocess.run(["git", "-C", ROOT, "rev-parse", "HEAD"], capture_output=True, text=True, check=True).stdout.strip()
        remote = subprocess.run(["git", "-C", ROOT, "rev-parse", "origin/main"], capture_output=True, text=True, check=True).stdout.strip()
        if local != remote:
            subprocess.run(["git", "-C", ROOT, "pull", "--ff-only", "origin", "main"], check=True, timeout=300)
            return "updated from GitHub; restart app to apply new code"
        return "already up to date"
    except Exception as e:
        return f"update error: {str(e)[:120]}"


EVOLVE_TOPICS = [
    "artificial intelligence self improvement",
    "chain of thought reasoning",
    "tree of thought search",
    "self-consistency decoding",
    "recursive self improvement algorithms",
    "react agent pattern",
    "retrieval augmented generation",
    "algorithm design and complexity",
    "machine learning transformers",
    "python programming best practices",
    "data structures algorithms",
    "neural network training techniques",
    "encyclopedia general knowledge",
]


def _self_evolve_worker():
    EVOLVE_STATE["running"] = True
    i = 0
    try:
        while EVOLVE_STATE["running"]:
            # evolve from conversations if enough
            if len(interactions) >= 4:
                learn_from_interactions()
            # learn from web on rotating topics
            if EVOLVE_STATE["running"]:
                topic = EVOLVE_TOPICS[i % len(EVOLVE_TOPICS)]
                i += 1
                EVOLVE_STATE["last_topic"] = topic
                learn_web(topic)
            EVOLVE_STATE["cycles"] += 1
            # downtime before next cycle
            for _ in range(300):
                if not EVOLVE_STATE["running"]:
                    break
                time.sleep(1)
    except Exception as e:
        training_state["last"] = f"evolve error: {str(e)[:120]}"
    finally:
        EVOLVE_STATE["running"] = False


def start_self_evolution():
    if EVOLVE_STATE["running"]:
        return "self-evolution already running"
    threading.Thread(target=_self_evolve_worker, daemon=True).start()
    return "self-evolution started"


def stop_self_evolution():
    EVOLVE_STATE["running"] = False
    return "stopping self-evolution..."



# ---------- Tools & dashboard ----------
TOOLS_DIR = os.path.join(MEMORY_DIR, "tools")
os.makedirs(TOOLS_DIR, exist_ok=True)
TOOLS_FILE = os.path.join(TOOLS_DIR, "registry.json")
TOOL_REGISTRY = {}  # name -> {"description":..., "code":..., "builtin": bool}


def builtin_web_search(text):
    res = web_search(text, 5)
    if not res:
        return "No results found."
    return "\n".join(f"{i+1}. {r['title']} - {r.get('snippet','')} ({r['url']})" for i, r in enumerate(res))


def builtin_define(text):
    d = fetch_definition(text.strip())
    return d if d else f"No definition found for {text}."


def builtin_calc(text):
    try:
        allowed = {"__builtins__": {}}
        val = eval(text, allowed, {})
        return str(val)
    except Exception as e:
        return f"calc error: {e}"


def builtin_time(text=""):
    return time.strftime("%Y-%m-%d %H:%M:%S")


def register_builtin_tools():
    specs = [
        ("web_search", "Search the web for a query.", None, builtin_web_search),
        ("define_word", "Define an English word.", None, builtin_define),
        ("calculate", "Evaluate a math expression.", None, builtin_calc),
        ("current_time", "Current date/time.", None, builtin_time),
    ]
    for name, desc, code, fn in specs:
        TOOL_REGISTRY[name] = {"description": desc, "code": None, "builtin": True, "fn": fn}


def load_tools():
    if os.path.exists(TOOLS_FILE):
        try:
            with open(TOOLS_FILE, "r") as f:
                data = json.load(f)
            for t in data:
                name = t["name"]
                code = t["code"]
                g = {}
                try:
                    exec(code, g)
                    TOOL_REGISTRY[name] = {"description": t.get("description", ""), "code": code, "builtin": False, "fn": g.get("run")}
                except Exception:
                    pass
        except Exception:
            pass


def save_tools():
    data = []
    for name, info in TOOL_REGISTRY.items():
        if info.get("builtin"):
            continue
        data.append({"name": name, "description": info.get("description", ""), "code": info.get("code", "")})
    with open(TOOLS_FILE, "w") as f:
        json.dump(data, f, indent=2)


def execute_tool(name, text):
    info = TOOL_REGISTRY.get(name)
    if not info:
        return f"Unknown tool: {name}"
    try:
        fn = info.get("fn")
        if fn is None and info.get("code"):
            g = {}
            exec(info["code"], g)
            fn = g.get("run")
            info["fn"] = fn
        if fn is None:
            return "Tool has no runnable function."
        out = fn(text)
        return str(out)
    except Exception as e:
        return f"tool error: {str(e)[:200]}"


def list_tool_names():
    return sorted(TOOL_REGISTRY.keys())


def create_tool_from_description(desc):
    if not desc.strip():
        return "Provide a tool description first."
    try:
        prompt = (
            "You are a Python tool generator. Write a Python function named `run(input_text)` that implements this tool: "
            + desc.strip()
            + "\nReturn only a single fenced code block with the function. No explanation. The function should return a string. "
            + "Do not use network or file writes; only pure computation."
        )
        msgs = [{"role": "system", "content": "You write safe, pure-Python tool functions as `run(input_text) returning str`."},
                {"role": "user", "content": prompt}]
        p = VOICE_TOK.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True) if VOICE_TOK else None
        if VOICE_MODEL is not None and VOICE_TOK is not None:
            inputs = VOICE_TOK(p, return_tensors="pt").to(VOICE_MODEL.device)
            with torch.no_grad():
                out = VOICE_MODEL.generate(**inputs, max_new_tokens=200, do_sample=False)
            raw = VOICE_TOK.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
        else:
            raw = ""
        # extract code
        import re
        m = re.search(r"```(?:python)?\n(.*?)```", raw, re.S)
        code = m.group(1).strip() if m else raw.strip()
        if "def run" not in code:
            return "Model did not produce a runnable function. Try rephrasing the description."
        # compile
        g = {}
        exec(compile(code, "<tool>", "exec"), g)
        if not callable(g.get("run")):
            return "Generated code has no run() function."
        # quick smoke test
        try:
            _ = g["run"]("")
        except Exception:
            pass
        # name
        slug = "".join(c if c.isalnum() else "_" for c in desc.lower())[:24].strip("_")
        slug = slug or "tool"
        name = slug
        i = 1
        while name in TOOL_REGISTRY:
            name = f"{slug}_{i}"
            i += 1
        TOOL_REGISTRY[name] = {"description": desc.strip(), "code": code, "builtin": False, "fn": g["run"]}
        save_tools()
        return f"Created tool '{name}'. It is now available in the tool panel."
    except Exception as e:
        return f"tool creation error: {str(e)[:200]}"


register_builtin_tools()
load_tools()


def learn_from_scratch_memory():
    global MODEL, LAST_TRAIN_TIME
    if not TRAIN_LOCK.acquire(blocking=False):
        return
    training_state["running"] = True
    training_state["last"] = "building from-scratch corpus..."
    try:
        if len(interactions) < 6:
            raise RuntimeError(f"need at least 6 interactions (have {len(interactions)})")
        corpus = os.path.join(MEMORY_DIR, "scratch_corpus.txt")
        lines = [f"User: {r['user']}\nAgron: {r['assistant']}" for r in interactions]
        with open(corpus, "w", encoding="utf-8") as f:
            f.write("\n\n".join(lines))
        out_dir = "/content/agron-from-scratch-memory"
        training_state["last"] = f"training from scratch on {len(interactions)} interactions..."
        subprocess.run([
            "python3", os.path.join(ROOT, "train_from_scratch.py"),
            "--size", "current", "--steps", "300",
            "--dataset", "file", "--file", corpus,
            "--output", out_dir
        ], check=True, timeout=600)
        import shutil
        shutil.copyfile(os.path.join(out_dir, "final", "model.safetensors"), os.path.join(ROOT, "model.safetensors"))
        # reload weights into the live model
        try:
            st = load_file(os.path.join(ROOT, "model.safetensors"))
            with IO_LOCK:
                MODEL.load_state_dict(st)
        except Exception:
            pass
        learned_topics.append("from-scratch memory model")
        persist_topics()
        training_state["last"] = "self-updating from scratch..."
        subprocess.run(["git", "-C", ROOT, "add", "model.safetensors", "memory/scratch_corpus.txt"], check=True)
        subprocess.run(["git", "-C", ROOT, "add", "memory/learned_topics.json"], check=False)
        status = subprocess.run(["git", "-C", ROOT, "status", "--porcelain"], capture_output=True, text=True, check=True).stdout.strip()
        if status:
            subprocess.run(["git", "-C", ROOT, "commit", "-m", f"Agron from-scratch learning ({len(interactions)} interactions)"], check=True)
            subprocess.run(["git", "-C", ROOT, "push", "origin", "main"], check=True)
        LAST_TRAIN_TIME = time.time()
        training_state["last"] = f"from-scratch learned & pushed ({len(interactions)} interactions)"
    except Exception as e:
        training_state["last"] = f"from-scratch error: {str(e)[:200]}"
    finally:
        training_state["running"] = False
        TRAIN_LOCK.release()



FACTORY_RESET_PATHS = ["/content/agron-model", "/content/agron-learned", "/content/agron-web", "/content/agron-dictionary", "/content/agron-from-scratch-memory", "/content/agron-scratch-small"]


def factory_reset():
    global MODEL, interactions, learned_topics
    stop_self_evolution()
    with IO_LOCK:
        interactions = []
        learned_topics = []
        try:
            os.makedirs(MEMORY_DIR, exist_ok=True)
            open(INTERACTIONS_FILE, "w").close()
            json.dump([], open(TOPICS_FILE, "w"))
        except Exception:
            pass
        # clear tools registry
        try:
            if os.path.exists(TOOLS_FILE):
                os.remove(TOOLS_FILE)
        except Exception:
            pass
        for tn in list(TOOL_REGISTRY.keys()):
            if not TOOL_REGISTRY[tn].get("builtin"):
                del TOOL_REGISTRY[tn]
        # clear scratch corpus
        try:
            sc = os.path.join(MEMORY_DIR, "scratch_corpus.txt")
            if os.path.exists(sc):
                os.remove(sc)
        except Exception:
            pass
    # delete trained model dirs
    import shutil
    for d in FACTORY_RESET_PATHS:
        if os.path.exists(d):
            shutil.rmtree(d, ignore_errors=True)
    # reinitialize Agron base model with fresh random weights (no pretrained load)
    try:
        from transformers import LlamaConfig, LlamaForCausalLM
        cfg = LlamaConfig(vocab_size=16384, hidden_size=512, intermediate_size=2048, num_hidden_layers=8, num_attention_heads=4, num_key_value_heads=4, max_position_embeddings=512, bos_token_id=0, eos_token_id=2, pad_token_id=1, tie_word_embeddings=False)
        MODEL = LlamaForCausalLM(cfg)
        if torch.cuda.is_available():
            MODEL = MODEL.to("cuda")
        MODEL.eval()
    except Exception as e:
        training_state["last"] = f"reset model error: {str(e)[:120]}"
    LAST = globals().get("LAST_TRAIN_TIME", 0.0)
    globals()["LAST_TRAIN_TIME"] = time.time()
    training_state["running"] = False
    training_state["steps"] = 0
    training_state["last"] = "factory reset complete - fresh base model initialized"
    return "factory reset complete"


with gr.Blocks(title="Agron Neural Interface") as demo:
    gr.HTML(PARTICLE_HTML, js_on_load=PARTICLE_JS)
    gr.Markdown("<h1 class='agron-title'>AGRØN NEURAL INTERFACE</h1>", elem_classes="agron-card")
    gr.Markdown("<div class='agron-sub'>Self-improving multi-agent core · particles = live network · learning from every exchange</div>")
    with gr.Row():
        with gr.Column(scale=1, elem_classes="agron-card"):
            gr.Markdown("<h3 class='agron-title'>SYSTEM</h3>")
            sys_status = gr.Textbox(value=f"status: {training_state['last']}", label="Status", interactive=False, elem_classes="status-box")
            sys_inter = gr.Textbox(value=f"interactions: {len(interactions)}", label="Interactions", interactive=False)
            sys_tools = gr.Textbox(value=f"tools: {len(list_tool_names())}", label="Tools", interactive=False)
            sys_cycles = gr.Textbox(value=f"evolve cycles: {EVOLVE_STATE['cycles']}", label="Evolution", interactive=False)
        with gr.Column(scale=4, elem_classes="agron-card"):
            gr.Markdown("<h2 class='agron-title'>💬 COMMAND CONSOLE</h2>")
            chatbot = gr.Chatbot(height=460, label="Agron")
            msg = gr.Textbox(placeholder="Speak to Agron...", label="Input", lines=2)
            with gr.Row():
                send = gr.Button("Transmit", variant="primary")
                clear = gr.Button("Clear")
            gr.Markdown("⚙ auto decoding · high cap · focused")
        with gr.Column(scale=2, elem_classes="agron-card"):
            gr.Markdown("<h2 class='agron-title'>📊 SYSTEM DASHBOARD</h2>")
            status = gr.Textbox(value=f"interactions: {len(interactions)} | {training_state['last']}", label="Learning", interactive=False, elem_classes="status-box")
            metrics = gr.Textbox(value=f"tools: {len(list_tool_names())} · cycles: {EVOLVE_STATE['cycles']} · nodes: 130", label="Diagnostics", interactive=False)
            gr.Markdown("### Tools")
            tool_dd = gr.Dropdown(choices=list_tool_names(), value=(list_tool_names()[0] if list_tool_names() else None), label="Select tool")
            tool_in = gr.Textbox(placeholder="Tool input...", label="Tool input", lines=1)
            tool_btn = gr.Button("Run tool")
            tool_out = gr.Textbox(label="Tool output", interactive=False, lines=6)
            gr.Markdown("### Create tool")
            tool_desc = gr.Textbox(placeholder="Describe a new tool (pure computation)...", label="New tool description", lines=2)
            create_btn = gr.Button("Create tool")
            gr.Markdown("### Learning controls")
            learn_btn = gr.Button("Train now on memory")
            scratch_btn = gr.Button("Train from scratch (no fine-tune)")
            dict_btn = gr.Button("Learn dictionary words")
            web_topic = gr.Textbox(label="Web topic / search query", placeholder="e.g. transformer architecture", lines=1)
            web_btn = gr.Button("Search & learn web")
            with gr.Row():
                evolve_btn = gr.Button("Start self-evolution")
                stop_evolve_btn = gr.Button("Stop")
            update_btn = gr.Button("Self-update")
            reset_btn = gr.Button("Factory reset")

    def respond(message, history):
        if not message.strip():
            return history, ""
        response = generate_response(message, history)
        history = history + [{"role": "user", "content": message}, {"role": "assistant", "content": response}]
        return history, ""

    def do_learn():
        if training_state["running"]:
            return f"training already running · {training_state['last']}"
        if len(interactions) < 2:
            return f"need at least 2 interactions (have {len(interactions)})"
        threading.Thread(target=learn_from_interactions, daemon=True).start()
        return "training started in background..."

    def refresh_status():
        return f"interactions: {len(interactions)} | running={training_state['running']} | cycles={EVOLVE_STATE['cycles']} | {training_state['last']}"

    send.click(respond, [msg, chatbot], [chatbot, msg])
    msg.submit(respond, [msg, chatbot], [chatbot, msg])
    clear.click(lambda: ([], ""), outputs=[chatbot, msg])
    def do_dict():
        if training_state["running"]:
            return "training already running"
        threading.Thread(target=learn_dictionary, daemon=True).start()
        return "dictionary learning started in background..."

    def do_web():
        if training_state["running"]:
            return "training already running"
        q = web_topic.value.strip()
        if not q:
            return "enter a topic/query first"
        threading.Thread(target=learn_web, args=(q,), daemon=True).start()
        return f"web learning started: {q}"
    def run_tool_fn(tool_name, tool_input):
        if not tool_name:
            return "select a tool first"
        return execute_tool(tool_name, tool_input or "")
    def create_tool_fn(desc):
        return create_tool_from_description(desc)
    def refresh_tools():
        choices = list_tool_names()
        return gr.Dropdown(choices=choices, value=(choices[0] if choices else None))
    tool_btn.click(run_tool_fn, [tool_dd, tool_in], tool_out)
    create_btn.click(create_tool_fn, tool_desc, status)
    learn_btn.click(do_learn, outputs=status)
    def do_scratch():
        if training_state["running"]:
            return "training already running"
        if len(interactions) < 6:
            return f"need at least 6 interactions (have {len(interactions)})"
        threading.Thread(target=learn_from_scratch_memory, daemon=True).start()
        return "from-scratch training started in background..."
    scratch_btn.click(do_scratch, outputs=status)
    dict_btn.click(do_dict, outputs=status)
    web_btn.click(do_web, outputs=status)
    def do_evolve():
        return start_self_evolution()
    def do_stop_evolve():
        return stop_self_evolution()
    def do_update():
        return self_update()
    evolve_btn.click(do_evolve, outputs=status)
    stop_evolve_btn.click(do_stop_evolve, outputs=status)
    update_btn.click(do_update, outputs=status)
    reset_btn.click(factory_reset, outputs=status)
    demo.load(refresh_status, outputs=status)

    # periodic status refresh (not true polling inside gradio, use change hook via demo load only)
    gr.Timer(5).tick(refresh_status, outputs=status)

if __name__ == "__main__":
    demo.launch(server_name="0.0.0.0", server_port=7860, share=True)