"""
RUBRA — memory_engine.py
Persistent, cross-session user memory — durable facts about a person that
carry across every NEW conversation, not just the current one.

The gap this closes: session history already makes RUBRA remember
everything within ONE chat. The moment someone clicks "New Chat", that
session_id is gone and so is everything RUBRA knew. This module is keyed
on a stable user identity instead (see main.py's `mem_key` resolution), so
starting a new chat doesn't mean starting from zero — the same way a
person doesn't forget a friend just because they're now talking on a new
phone call instead of the last one.

Design mirrors how memory works for Claude itself: silent application
(never "according to my memory of you..."), durable facts only (not a
transcript of every message), bounded size (oldest facts age out), and
never forced into a reply that isn't about that topic.
"""

import re
import json
import logging
from typing import Optional, List

from database import user_memory_load, user_memory_save

log = logging.getLogger("rubra.memory")

MAX_FACTS = 40  # bounded growth — oldest facts drop off past this

_NAME_PATTERNS = [
    r"\b(?:my name is|i am|i'm|call me)\s+([A-Z][a-zA-Z]{1,20})\b",
    r"\b(?:amar naam|আমার নাম)\s+([A-Za-z\u0980-\u09FF]{2,20})\b",
]

# Common false positives for the "i'm X" pattern — moods/states, not names.
_NAME_FALSE_POSITIVES = {
    "fine", "good", "ok", "okay", "not", "sorry", "happy", "sad", "tired",
    "done", "here", "back", "working", "trying", "going", "sure", "glad",
    "stuck", "confused", "busy", "free", "ready", "new", "still", "also",
}


def _extract_name_heuristic(message: str) -> Optional[str]:
    """Cheap regex pass — runs on every message, no LLM call needed."""
    for pattern in _NAME_PATTERNS:
        m = re.search(pattern, message, re.IGNORECASE)
        if m:
            name = m.group(1).strip()
            if name.lower() not in _NAME_FALSE_POSITIVES:
                return name
    return None


async def extract_and_update_memory(user_id: str, user_msg: str, assistant_msg: str, llm_func) -> None:
    """
    Fire-and-forget after a turn completes — call via asyncio.create_task,
    never awaited inline, so memory extraction can't slow down the actual
    reply. Never raises: a memory-extraction failure must never surface as
    a broken chat response.

    Cheap heuristic name extraction always runs. A short LLM call extracts
    0-3 durable facts, but only for exchanges substantial enough to likely
    contain one (skips "ok thanks", "hi", etc. — not worth a call).
    """
    if not user_id:
        return
    try:
        mem   = user_memory_load(user_id)
        name  = mem.get("name")
        facts = list(mem.get("facts", []))

        heuristic_name = _extract_name_heuristic(user_msg)
        if heuristic_name and not name:
            name = heuristic_name

        if len(user_msg.split()) >= 6:
            prompt = (
                "Extract 0 to 3 SHORT durable facts about the user from this single "
                "exchange that would help in FUTURE, otherwise-unrelated conversations — "
                "things like their name, profession, ongoing projects, preferences, or "
                "communication style. Do NOT extract one-off task details (e.g. \"asked to "
                "fix a typo\" is not durable; \"is building an e-commerce site called Foo\" "
                "IS durable). If nothing durable came up, return an empty list.\n\n"
                f"User: {user_msg[:500]}\nAssistant: {assistant_msg[:500]}\n\n"
                "Respond with ONLY a JSON array of short strings, e.g. "
                '["works as a graphic designer", "prefers concise answers"]. '
                "No explanation, no markdown fences, no extra text."
            )
            raw = ""
            try:
                async for tok in llm_func([{"role": "user", "content": prompt}], mode="fast"):
                    raw += tok
            except Exception as e:
                log.warning(f"[MEMORY] extraction call failed: {e}")
                raw = ""

            raw = raw.strip()
            if raw.startswith("```"):
                raw = re.sub(r'^```\w*\n?', '', raw)
                raw = re.sub(r'\n?```$', '', raw)
            try:
                new_facts = json.loads(raw) if raw else []
                if isinstance(new_facts, list):
                    for f in new_facts:
                        f = str(f).strip()
                        if f and f not in facts:
                            facts.append(f)
            except Exception:
                pass  # model didn't return clean JSON — skip silently, not fatal

        facts = facts[-MAX_FACTS:]
        if name or facts:
            user_memory_save(user_id, name, facts)
    except Exception as e:
        log.warning(f"[MEMORY] update failed for {user_id}: {e}")


def format_memory_for_prompt(user_id: str) -> str:
    """
    Builds the system_note injected into EVERY conversation (new or
    ongoing). Empty string if nothing is known yet — never inject a hollow
    "I don't know anything about you" block; absence of memory should be
    invisible, not announced.
    """
    if not user_id:
        return ""
    mem = user_memory_load(user_id)
    name, facts = mem.get("name"), mem.get("facts", [])
    if not name and not facts:
        return ""

    lines = ["[WHAT YOU KNOW ABOUT THIS PERSON — from earlier conversations]"]
    if name:
        lines.append(f"Their name is {name}. Use it naturally when it fits — don't overuse it.")
    for f in facts[-MAX_FACTS:]:
        lines.append(f"- {f}")
    lines.append(
        "Use this the way a person remembers a friend across conversations — never say "
        "\"according to my memory\" or list these facts back at them, and don't force an "
        "irrelevant fact into a reply that isn't about that topic."
    )
    return "\n".join(lines)