"""Central configuration, loaded from environment / .env.

Everything has a safe default so the app runs fully offline with zero config.
"""
from __future__ import annotations

import os
from functools import lru_cache
from pathlib import Path

BACKEND_DIR = Path(__file__).resolve().parent.parent  # .../backend
REPO_ROOT = BACKEND_DIR.parent

# Make the repo-root `prototype/` package importable (only the composition layer
# imports it, and only in prototype mode — real code never depends on prototype).
import sys as _sys  # noqa: E402
if str(REPO_ROOT) not in _sys.path:
    _sys.path.insert(0, str(REPO_ROOT))

try:  # load .env deterministically (independent of cwd): backend/.env then repo/.env
    from dotenv import load_dotenv

    load_dotenv(BACKEND_DIR / ".env")
    load_dotenv(REPO_ROOT / ".env")
    load_dotenv()  # finally, anything discoverable from cwd (no override)
except Exception:  # pragma: no cover - dotenv optional
    pass

# On serverless hosts (Vercel) the bundle is read-only except /tmp.
IS_SERVERLESS = bool(os.getenv("VERCEL") or os.getenv("AWS_LAMBDA_FUNCTION_NAME"))


def _writable_dir() -> Path:
    """A directory we can actually write to (metrics DB, eval reports)."""
    if IS_SERVERLESS:
        d = Path("/tmp/aperture")
    else:
        d = BACKEND_DIR / "data"
    try:
        d.mkdir(parents=True, exist_ok=True)
    except OSError:
        d = Path("/tmp/aperture")
        d.mkdir(parents=True, exist_ok=True)
    return d


def _bool(name: str, default: bool) -> bool:
    val = os.getenv(name)
    if val is None:
        return default
    return val.strip().lower() in {"1", "true", "yes", "on"}


def _float(name: str, default: float) -> float:
    try:
        return float(os.getenv(name, default))
    except (TypeError, ValueError):
        return default


class Settings:
    """Plain object (not pydantic-settings) to avoid a hard dependency in tests."""

    def __init__(self) -> None:
        # --- hosted models ---
        self.anthropic_api_key: str | None = os.getenv("ANTHROPIC_API_KEY") or None
        self.gemini_api_key: str | None = (
            os.getenv("GEMINI_API_KEY") or os.getenv("GOOGLE_API_KEY") or None
        )
        self.anthropic_model_smart = os.getenv("ANTHROPIC_MODEL_SMART", "claude-sonnet-4-6")
        self.anthropic_model_cheap = os.getenv(
            "ANTHROPIC_MODEL_CHEAP", "claude-haiku-4-5-20251001"
        )
        self.gemini_model = os.getenv("GEMINI_MODEL", "gemini-2.5-flash")

        # --- local model ---
        self.ollama_base_url = os.getenv("OLLAMA_BASE_URL", "http://localhost:11434")
        self.local_model = os.getenv("LOCAL_MODEL", "gemma2:2b")
        self.local_backend = os.getenv("LOCAL_BACKEND", "ollama")  # ollama | transformers

        # --- deployment mode ---------------------------------------------------
        # real      = production-first: real OCR / RAG / LLM / browser only.
        # prototype = re-enable simulated/offline fallbacks (mock LLM, simulated
        #             browser, sidecar OCR) from the prototype/ package.
        # Serverless (Vercel) defaults to prototype so the public demo still runs.
        self.mode = os.getenv("APERTURE_MODE") or ("prototype" if IS_SERVERLESS else "real")
        self.prototype_mode = self.mode == "prototype"

        # --- routing / caching ---
        self.routing_policy = os.getenv("ROUTING_POLICY", "auto")  # auto|cheap|smart|offline
        self.enable_prompt_cache = _bool("ENABLE_PROMPT_CACHE", True)
        self.enable_semantic_cache = _bool("ENABLE_SEMANTIC_CACHE", True)
        self.semantic_cache_threshold = _float("SEMANTIC_CACHE_THRESHOLD", 0.92)

        # --- pipeline ---
        self.hitl_confidence_threshold = _float("HITL_CONFIDENCE_THRESHOLD", 0.85)

        # --- storage (always writable; /tmp on serverless) ---
        self.writable_dir = _writable_dir()
        db = os.getenv("METRICS_DB_PATH")
        if db:
            self.metrics_db_path = Path(db) if Path(db).is_absolute() else BACKEND_DIR / db
        else:
            self.metrics_db_path = self.writable_dir / "metrics.db"
        try:
            self.metrics_db_path.parent.mkdir(parents=True, exist_ok=True)
        except OSError:
            self.metrics_db_path = self.writable_dir / "metrics.db"

        # eval report: committed copy (read) + writable copy (fresh runs)
        self.eval_report_committed = BACKEND_DIR / "evals" / "report.json"
        self.eval_report_writable = self.writable_dir / "report.json"

        # --- auth (HTTP Basic on /api/*) ---
        self.auth_user = os.getenv("APERTURE_USER", "demo")
        self.auth_pass = os.getenv("APERTURE_PASS", "aperture2026")
        # admin role (observability dashboard, prompt mgmt). Defaults to same creds.
        self.admin_user = os.getenv("APERTURE_ADMIN_USER", self.auth_user)
        self.admin_pass = os.getenv("APERTURE_ADMIN_PASS", self.auth_pass)

        # --- OCR backends -----------------------------------------------------
        # selection: auto | minicpm | cohere | llamaparse | tesseract | easyocr | sidecar
        self.ocr_backend = os.getenv("OCR_BACKEND", "auto")
        # Option 1 — MiniCPM-V-4.6 via OpenAI-compatible server (vLLM / llama.cpp)
        self.minicpm_base_url = os.getenv("MINICPM_BASE_URL", "").rstrip("/") or None
        self.minicpm_api_key = os.getenv("MINICPM_API_KEY") or None
        self.minicpm_model = os.getenv("MINICPM_MODEL", "MiniCPM-V-4.6-Instruct")
        # Option 2 — Cohere vision/OCR model from HuggingFace (transformers)
        self.cohere_ocr_model = os.getenv("COHERE_OCR_MODEL", "CohereLabs/aya-vision-8b")
        # Option 3 — LlamaParse (LlamaCloud)
        self.llama_cloud_api_key = os.getenv("LLAMA_CLOUD_API_KEY") or None
        self.llamaparse_result_type = os.getenv("LLAMAPARSE_RESULT_TYPE", "markdown")

        # --- model labs (all ≤32B params — "small models") --------------------
        # OpenBMB (MiniCPM family) — text/vision reasoning + OCR (via MINICPM_* above)
        self.openbmb_model = os.getenv("OPENBMB_MODEL", self.minicpm_model)
        # OpenBMB MiniCPM3-4B — text reasoning / NLQ→SQL / summarization (ERP DocIQ + fine-tune target)
        self.openbmb_reasoner_model = os.getenv("OPENBMB_REASONER_MODEL", "MiniCPM3-4B")
        # Black Forest Labs (FLUX) — image GENERATION for synthetic test documents
        self.bfl_api_key = os.getenv("BFL_API_KEY") or None
        self.bfl_model = os.getenv("BFL_MODEL", "flux-dev")  # api: flux-dev | flux-pro-1.1 | flux-schnell
        # Cohere hosted API (in addition to the local HF Aya-Vision backend above)
        self.cohere_api_key = os.getenv("COHERE_API_KEY") or None

        # --- databases --------------------------------------------------------
        appdb = os.getenv("APP_DB_PATH")
        self.app_db_path = (
            (Path(appdb) if Path(appdb).is_absolute() else BACKEND_DIR / appdb)
            if appdb else self.writable_dir / "aperture.db"
        )
        ragdb = os.getenv("RAG_DB_PATH")
        self.rag_db_path = (
            (Path(ragdb) if Path(ragdb).is_absolute() else BACKEND_DIR / ragdb)
            if ragdb else self.writable_dir / "rag.db"
        )
        erpdb = os.getenv("ERP_DB_PATH")
        self.erp_db_path = (
            (Path(erpdb) if Path(erpdb).is_absolute() else BACKEND_DIR / erpdb)
            if erpdb else self.writable_dir / "erp.db"
        )

        # --- browser ---
        self.playwright_headless = _bool("PLAYWRIGHT_HEADLESS", True)
        self.demo_portal_url = os.getenv("DEMO_PORTAL_URL", "/portal")

        # --- paths ---
        self.samples_dir = REPO_ROOT / "samples"
        self.evals_dataset_dir = BACKEND_DIR / "evals" / "datasets"
        self.demo_portal_dir = REPO_ROOT / "demo-portal"
        self.frontend_dist_dir = REPO_ROOT / "frontend" / "dist"

    # --- capability detection -------------------------------------------------
    def has_anthropic(self) -> bool:
        return bool(self.anthropic_api_key) and _module_available("anthropic")

    def has_gemini(self) -> bool:
        return bool(self.gemini_api_key) and _module_available("google.generativeai")

    def has_local(self) -> bool:
        # We can't cheaply ping ollama here; LocalProvider does a lazy health check.
        # Treat "configured" as available and let the provider degrade if not.
        return _bool("ENABLE_LOCAL_MODEL", False) or _module_available("transformers")


def _module_available(name: str) -> bool:
    import importlib.util

    try:
        return importlib.util.find_spec(name) is not None
    except (ImportError, ValueError):
        return False


@lru_cache(maxsize=1)
def get_settings() -> Settings:
    return Settings()