"""
tool_generator.py — COG-4: Dynamic Tool Generation at runtime.

Quando l'executor non trova un tool adatto per un subtask,
chiede all'LLM di scrivere una piccola funzione Python async,
la valida in sandbox (run_python), e la registra nel TOOL_REGISTRY
per la durata della sessione corrente.

Security constraints:
  - Blocca import di rete (requests, httpx, socket, urllib, subprocess)
  - Blocca accesso filesystem in write (open() write mode)
  - Timeout 12s generazione + 10s validazione
  - Max 5 tool generati per sessione (evita context explosion)
  - Tool generati hanno prefisso "_dyn_" per identificazione e pulizia

Integration: chiamato da unified_loop.py quando _TOOL_MAP.get(tool) == (None, None)
e needs_dynamic_tool() restituisce True.
"""
from __future__ import annotations

import asyncio
import logging
import re
from typing import Any

_logger = logging.getLogger("agente_ai.tool_gen")

_GENERATED_COUNT = 0
_MAX_GENERATED   = 5

_BLOCKED_RE = re.compile(
    r"\b(import\s+(requests|httpx|socket|urllib|subprocess|paramiko|ftplib)"
    r"|os\.system\s*\(|os\.popen\s*\(|eval\s*\(|exec\s*\("
    r"|__import__\s*\(|open\s*\([^)]*['\"][wa][+bt]*['\"])",
    re.IGNORECASE | re.MULTILINE,
)
_TOOL_NAME_RE = re.compile(r"^[a-z][a-z0-9_]{2,39}$")

_GEN_SYSTEM = """\
Sei un tool engineer Python. Scrivi UNA funzione Python asincrona che risolve il task specificato.

REGOLE ASSOLUTE:
1. Firma obbligatoria: `async def tool_fn(**kwargs) -> dict`
2. Ritorna sempre: `{"success": True/False, "output": <risultato>, "error": ""}`
3. ZERO import di rete (no requests, httpx, socket, urllib, subprocess)
4. ZERO filesystem esterno (no open() in write mode)
5. Solo stdlib: json, re, math, datetime, collections, itertools, hashlib, base64, urllib.parse
6. Max 30 righe
7. Rispondi SOLO con il blocco ```python ... ``` — zero testo extra

Esempio:
```python
import re
async def tool_fn(**kwargs) -> dict:
    text = kwargs.get("text", "")
    words = re.findall(r"\\b\\w+\\b", text.lower())
    freq = {}
    for w in words:
        freq[w] = freq.get(w, 0) + 1
    return {"success": True, "output": sorted(freq.items(), key=lambda x: -x[1])[:10], "error": ""}
```\
"""


def needs_dynamic_tool(tool_name: str, description: str) -> bool:
    """
    Decide se generare un tool dinamico per questo subtask.

    Condizioni positive:
    - La descrizione suggerisce un'operazione computazionale locale
    - Non siamo al limite di tool generati per sessione

    Non genera tool per: operazioni di rete, filesystem esterno, browser.
    """
    global _GENERATED_COUNT
    if _GENERATED_COUNT >= _MAX_GENERATED:
        return False

    _NETWORK_RE = re.compile(
        r"\b(http|url|fetch|scrape|download|upload|api|request|socket)\b",
        re.IGNORECASE,
    )
    if _NETWORK_RE.search(description):
        return False  # operazioni di rete → non generare tool locale

    _COMPUTE_RE = re.compile(
        r"\b(calcola|converti|trasforma|analizza|estrai|filtra|ordina|conta"
        r"|parse|format|encode|decode|compress|hash|valida|validate"
        r"|split|merge|aggregate|summarize|riassumi|statistiche)\b",
        re.IGNORECASE,
    )
    return bool(_COMPUTE_RE.search(description))


async def generate_and_register(
    description: str,
    tool_name: str,
    llm: Any,
    executor: Any,
) -> "tuple[bool, str]":
    """
    Genera un tool Python via LLM, lo valida in sandbox, lo registra.

    Returns: (success: bool, dyn_name: str)
    """
    global _GENERATED_COUNT
    if _GENERATED_COUNT >= _MAX_GENERATED:
        return False, ""

    safe_name = re.sub(r"[^a-z0-9_]", "_", tool_name.lower())[:36]
    if not _TOOL_NAME_RE.match(safe_name):
        safe_name = f"gen{_GENERATED_COUNT + 1}"
    dyn_name = f"_dyn_{safe_name}"

    # Step 1: genera il codice via LLM
    try:
        raw = await asyncio.wait_for(
            llm.chat(
                [
                    {"role": "system", "content": _GEN_SYSTEM},
                    {"role": "user",   "content": f"Scrivi un tool per: {description[:400]}"},
                ],
                temperature=0.1,
                max_tokens=600,
            ),
            timeout=12.0,
        )
    except Exception as exc:
        _logger.warning("COG-4 gen LLM failed: %s", exc)
        return False, ""

    # Step 2: estrai blocco codice
    m    = re.search(r"```(?:python)?\n?([\s\S]+?)```", raw)
    code = m.group(1).strip() if m else raw.strip()

    # Step 3: security gate
    if _BLOCKED_RE.search(code):
        _logger.warning("COG-4 blocked unsafe import in generated tool: %s", dyn_name)
        return False, ""

    # Step 4: valida in sandbox
    validation = (
        f"{code}\n\n"
        "import asyncio as _asyncio\n"
        "_r = _asyncio.run(tool_fn())\n"
        "assert isinstance(_r, dict), 'must return dict'\n"
        "assert 'output' in _r, 'must have output key'\n"
        "print('COG4_OK')\n"
    )
    try:
        val = await asyncio.wait_for(
            executor.run_tool("run_python", {"code": validation}),
            timeout=10.0,
        )
        out = val.get("output", {})
        stdout = out.get("stdout", "") if isinstance(out, dict) else str(out)
        if "COG4_OK" not in stdout:
            _logger.warning("COG-4 validation failed: %s", stdout[:200])
            return False, ""
    except Exception as exc:
        _logger.warning("COG-4 sandbox error: %s", exc)
        return False, ""

    # Step 5: registra nel TOOL_REGISTRY (runtime only, non persiste)
    try:
        from tools.registry import TOOL_REGISTRY
        ns: dict = {}
        exec(compile(code, f"<dyn:{dyn_name}>", "exec"), ns)  # noqa: S102
        fn = ns.get("tool_fn")
        if not fn or not asyncio.iscoroutinefunction(fn):
            _logger.warning("COG-4 tool_fn missing or not async")
            return False, ""
        TOOL_REGISTRY[dyn_name] = {
            "description": f"[DYN] {description[:200]}",
            "required_inputs": [],
            "_fn": fn,
            "_generated": True,
        }
        _GENERATED_COUNT += 1
        _logger.info(
            "COG-4 registered '%s' (%d/%d total dynamic tools)",
            dyn_name, _GENERATED_COUNT, _MAX_GENERATED,
        )
        return True, dyn_name
    except Exception as exc:
        _logger.warning("COG-4 registration failed: %s", exc)
        return False, ""